Datasets:

emozilla

/

dolma-v1_7-arxiv

like 1

\section{Introduction} \label{sec:introduction} Fission is a fundamental nuclear decay that is important in many areas of science, ranging from structure and stability of heavy and superheavy nuclei \cite{krappe2012,Schmidt2018,GiulianiRMP} to studies devoted to physics beyond the standard model of particle physics~\cite{vogel2015} and the synthesis of heavy elements~\cite{horowitz2018,vassh2019,Giuliani2020}. Theoretically, the nuclear fission process is an example of the nuclear large-amplitude collective motion originating from the single-particle motion of individual nucleons. Due to the complexity of this process, our understanding of nuclear fission is still incomplete. For the state of affairs in this field, we refer to the recent review \cite{schunck2016,Bender2020}. When it comes to realistic predictions, the self-consistent nuclear energy density functional (EDF) method \cite{BenderRMP,SchunckBook} has proven to be very successful in terms of quantitative reproduction of fission lifetimes and fragment yields. Unfortunately, realistic self-consistent fission calculations in a multidimensional collective space, based on the microscopic input, are computationally expensive when it comes to large-scale theoretical fission surveys. Given the computational cost of microscopic methods and the large number of fissioning nuclei that are, e.g., expected to contribute to the astrophysical r-process nucleosynthesis, calculations have mostly relied on simple parametrizations or highly phenomenological models. The new perspective is offered by state-of-the-art theoretical frameworks and modern computational techniques that promise to speed up the calculations to be able to carry out quantified global fission surveys for multiple inputs \cite{Bender2020}. This study is concerned with finding the optimal pathway during the tunneling motion phase of spontaneous fission (SF). Such a trajectory, dubbed the least-action path (LAP), is obtained by minimizing the collective action in a many-dimensional collective space \cite{Kapur1937,Brack1972}. A number of techniques have been proposed to deal with this challenging task. In the early application \cite{Ledergerber1973}, the trial pathways were assumed in a parametrized form and the LAP was obtained by minimizing the penetration integral with respect to the variational parameters. Grid-based techniques such as the dynamic-programming~\cite{Baran1981} and Ritz~\cite{Baran1978} methods have been used in numerous EDF calculations of LAPs \cite{Sadhukhan2013,Sadhukhan2014,Sadhukhan2016,Sadhukhan2017,Zhao2015,Zhao2016,Mercier2021}. In Refs.~\cite{Schmid1986,Eckern1992,Kindo1989,Iwamoto1992,Iwamoto1994,Scamps2015} LAPs were obtained by solving the eikonal equation by the method of characteristics. Effectively, this method can be related to a quantum mechanical propagation in imaginary time that amounts to solving the classical equations of motion in an inverted potential. Within this approach, only one trajectory, called the escape path, arrives at the outer turning surface with zero velocity. Other trajectories, corresponding to different initial conditions, cannot reach the outer turning surface. In this paper, we compare grid-based approaches to the LAPs with the nudged elastic band (NEB) method that was originally formulated in the context of molecular systems ~\cite{Garrett1983,Mills1994,Mills1995,Jonsson1998}. In NEB, the minimum action path can be obtained iteratively by continuously shifting the pathway to the nearest minimum action path \cite{Henkelman2000,Henkelman2000a,ASE}. A similar approach is a growing string method \cite{Peters2004}. To provide more insights, we also employ the Euler-Lagrange (EL) method to compute the stationary action path. In addition to the LAP, another characteristic trajectory in the collective space is the minimum-energy path (MEP), sometimes referred to as the static path. The MEP can serve as a first, rough approximation to the LAP. It is obtained by computing the steepest descent line on the potential energy surface, which passes through the local minima and saddle points. To find the MEP, a flooding, or watershed, algorithm has been applied \cite{Mamdouh1998,Moller2001,Iwamoto2002,Moller2004,Wang2019a}. The NEB approach can also be adopted to find the MEP and saddle points \cite{Asgeirsson2021}. (For a review of modern optimization methods for finding MEPs, see \cite{Sheppard2008,More2004}.) This paper is organized as follows. In Sec.~\ref{sec:EDF} we define the basics concepts of the nuclear EDF approach as applied to nuclear fission. Section~\ref{Sec: Methods algorithms} describes the path-optimization methods used. The results of our calculations and an analysis of trends are presented in Sec.~\ref{sec:results}. Finally, Sec.~\ref{sec:summary} contains the conclusions of this work. \section{Nuclear EDF approach to spontaneous fission}\label{sec:EDF} The main ingredients for a theoretical determination of SF lifetimes are the collective potential energy surface (PES) and the inertia tensor. To compute the PES, one solves the constrained Hartree-Fock-Bogoliubov (HFB) equations with the realistic energy density functional in the space of collective coordinates $\vec{q} \equiv \{q_i\}$. These are usually represented by the expectation values of the quadrupole moment operator $\hat{Q}_{20}$ (elongation), quadrupole moment operator $\hat{Q}_{22}$ (triaxiality), octupole moment operator $\hat{Q}_{30}$ (mass-asymmetry), and the particle-number dispersion term $\lambda_{2\tau}(\hat{ N}^2_\tau - \langle \hat{ N}_\tau\rangle^2)$ ($\tau=n,p)$ that controls dynamic pairing correlations \cite{Vaquero2011,Vaquero2013,Sadhukhan2014}. In some cases one also considers the hexadecapole moment $Q_{40}$ (necking coordinate) \cite{Warda2012}. That is, in practical applications, we consider 2-5 collective coordinates which describe the collective motion of the system. Figure~\ref{fig:PES} shows a representative PES of $^{256}$Fm in the space of $Q_{20}\equiv \langle \hat{Q}_{20}\rangle $ and $Q_{30}\equiv \langle \hat{Q}_{30}\rangle$. \begin{figure*}[!htb] \includegraphics[width=0.8\linewidth]{Figures/256Fm_PES.png} \caption{Potential energy surface of $^{256}$Fm calculated with nuclear EDF method using the D1S parametrization of the Gogny interaction \cite{berger1984microscopic} in the space of two collective coordinates: $Q_{20}$ (elongation) and $Q_{30}$ (mass asymmetry). The static fission pathways are marked by solid lines: red (symmetric pathway) and green (asymmetric pathway). The outer turning line (OTL) is indicated, together with the outer turning points associated with the static pathways. For simplicity, we assume that the inner turning point corresponds to the ground-state configuration (i.e., $E_0=0$). The high-energy region that is practically not accessible during collective motion is indicated in black. The intersections of fission pathways with outer turning points are indicated by dots; these are important for determining fission fragment yields \cite{Sadhukhan2020,Sadhukhan2022}. } \label{fig:PES} \end{figure*} The collective inertia (or mass) tensor $\mathcal{M}_{ij}(\vec{q})$ is obtained from the self-consistent densities by employing the the adiabatic time-dependent HFB approximation (ATDHFB) \cite{Baran2011,Giuliani2018b,Washiyama2021}. In this study, we use the non-perturbative cranking approximation~\cite{Baran2011}: \begin{equation} \label{cranking-mass} \mathcal{M}_{ij}(\vec{q})=\frac{\hbar^2}{2\dot{q}_i\dot{q}_j}\sum_{\alpha\beta}\frac{\left(F^{i*}_{\alpha\beta}F^{j}_{\alpha\beta}+ F^{i}_{\alpha\beta}F^{j*}_{\alpha\beta}\right)}{E_{\alpha}+E_{\beta}}, \end{equation} where $q_i$ is the collective coordinate, $\dot{q}_i$ represents the time derivative of $q_i$, and $E_{\alpha}$ are one-quasiparticle energies of HFB eigenstates $|\alpha\rangle$. The matrices $F^i$ are given by \begin{equation} \label{equation-F} \frac{F^{i*}}{\dot{q}_i}= A^T\frac{\partial\kappa^*}{\partial q_i}A +A^T\frac{\partial\rho^*}{\partial q_i}B -B^T\frac{\partial\rho}{\partial q_i}A - B^T\frac{\partial\kappa}{\partial q_i}B, \end{equation} where $A$ and $B$ are the matrices of the Bogoliubov transformation, and $\rho$ and $\kappa$ are particle and pairing density matrices, respectively, determined in terms of $A$ and $B$. Derivatives of the density matrices with respect to collective coordinates are calculated by employing the three-point Lagrange formula. It is important to remark that rapid variations in $\mathcal{M}_{ij}$ are expected in the regions of configuration changes (level crossings) due to strong variations of density derivatives in (\ref{equation-F}) associated with structural rearrangements \cite{Ledergerber1973,Sadhukhan2013}. Since SF is a quantum-mechanical tunneling process and the fission barriers are usually both high and wide, the SF lifetime is obtained semi-classically \cite{Brack1972} as $T_{1/2}=\ln2/(nP)$, where $n$ is the number of assaults on the fission barrier per unit time and $P$ is the penetration probability given by \begin{equation}\label{penertration} P=\left(1+\exp{[2S(L_{\rm min})]}\right)^{-1}, \end{equation} where $L_{\rm min}$ is the path that minimizes the fission action integral calculated along the one-dimensional trajectory $L(s)$ in the multidimensional collective space: \begin{equation} \label{action-integral} S(L)=\frac{1}{\hbar} \int_{s_{\rm in}}^{s_{\rm out}} {\cal S}(s)\,ds, \end{equation} where \begin{equation} \label{action-integrand} {\cal S}(s) =\sqrt{2\mathcal{M}_{\text{eff}}(s) \left(V_{\text{eff}}(s)-E_0\right)} \end{equation} with $V_{\text{eff}}(s)$ and $\mathcal{M}_{\text{eff}}(s)$ being the effective potential energy and inertia along the fission path $L(s)$, respectively. $V_{\text{eff}}$ can be obtained by subtracting the vibrational zero-point energy from the total HFB energy. (In the examples considered in this paper we assume the zero-point energy to be zero.) The integration limits $s_{\rm in}$ and $s_{\rm out}$ correspond to the classical inner and outer turning points, respectively, defined by $V_{\text{eff}}(s)=E_0$ on the two extremes of the fission path, see Fig.~\ref{fig:PES}. The collective ground state (g.s.) energy is $E_0$, and $ds$ is the element of length along $L(s)$. A one-dimensional path $L(s)$ can be defined in the multidimensional collective space by specifying the collective variables $\vec{q}(s)$ as functions of path's length $s$. The expression for $\mathcal{M}_{\text{eff}}$ is~\cite{Baran2005}: \begin{equation} \label{eff-mass} \mathcal{M}_{\text{eff}}(s)=\sum_{ij}\mathcal{M}_{ij}(\vec{q})\frac{dq_{i}}{ds}\frac{dq_{j}}{ds}. \end{equation} The least-action path (LAP) $L_{\rm min}$ is obtained by minimizing the action integral (\ref{action-integral}) with respect to all possible trajectories $L$ that connect the lines/surfaces of inner turning points $s_{\rm in}$ and outer turning points $s_{\rm out}$ \cite{Sadhukhan2013}. However, as discussed in Refs.~\cite{Schmid1986,Eckern1992} and this paper, only the pathways related to the exit points are stationary. The MEP can instead be described as the union of steepest descent paths from the saddle point(s) to the minima. The corresponding trajectory $\vec{q}(s)$ satisfies \begin{align} \frac{d \vec{q}}{ds} \propto \vec{\nabla} V\big(\vec{q}(s)\big) \label{MEP-def} \end{align} which characterizes a path of steepest descent on a surface $V(\vec{q})$ \cite{quapp1984analysis}. For the NEB, one finds the MEP by allowing the elements of the path to follow the gradient of the PES in their immediate vicinity. We shall assume that the PES in the tunneling region is free from discontinuities associated with rapid configuration changes \cite{Dubray2012,Zdeb2021,Lau2021}. This assumption is usually valid because of non-vanishing pairing correlations inside the potential barrier. It is also to be noted that, as in any optimization/minimization approach, the stationary path determined numerically corresponds to a local action minimum, which is not guaranteed to be the global minimum. Moreover, there could be many stationary pathways representing different fission modes, see Fig.~\ref{Fig: Bifurcation}. To simplify notation, we assume in the following discussions that the stationary action path found by our algorithms is indeed the LAP. Since ${\cal S}(s)=0$ on the outer turning surface $V(\vec q) = E_{0}$, it follows that paths moving on the surface $V(\vec{q}) = 0$ do not contribute to the action. This is illustrated in Fig. \ref{Fig: Bifurcation} by the path connecting the g.s. and, for example, the purple star labeled (3). Such a path consists of the cyan curve -- the exit trajectory -- and the green dashed line, connecting (1) with (3) through the OTL, which results in the same action integral as the exit trajectory. \begin{figure}[htb!] \label{Fig: Bifurcation} \includegraphics[width=\linewidth]{Figures/bimodal1.pdf} \caption{Illustration of two stationary action paths (representing competing fission modes) from the g.s. to the OTL (marked white) on the PES given by Eq.~(\ref{epsbimodal}). The cyan line shows the primary path (1). The secondary path (2) is indicated by the black line. The corresponding exit points are marked by stars. The green dashed line connects the exit point (1) with the point (3) on the OTL; the action along the dashed path is zero. The inset shows the spring force and the action force acting on the image $i$ on the NEB for the intermediate (not fully converged) grey path. For the video illustrating the NEB determination of both LAPs, see the supplemental material (SM) \cite{SM}.} \end{figure} \section{Methods/algorithms} \label{Sec: Methods algorithms} All path-optimization methods described in the following subsections, bar the EL method, have a reference implementation included in the python package, PyNEB \cite{PyNEB}. \subsection{Nudged Elastic Band}\label{Sec: NEB} The NEB method was originally formulated to provide a smooth transition of a molecular system on a potential energy surface from the reactant to the product state~\cite{Mills1994,Mills1995,Jonsson1998}. Upon application of this variant of the NEB method, one obtains the MEP as well as a series of ``images'' of the molecular system as it transitions along the path. The NEB technique has been subsequently refined, with improved numerical stability~\cite{Henkelman2000a} and a more accurate determination of a saddle point~\cite{Henkelman2000} being two key advances towards a more widely applicable numerical approach for MEP determination. To obtain the LAP, the procedure must be modified such that the images move towards the minimum of the action~\cite{Asgeirsson2018} which amounts to replacing the standard gradient of the PES with the gradient of the action \begin{equation} \vec{g}_i= -\vec{\nabla}_i S \end{equation} with respect to the image $\vec q_i$. With this prescription, the images will settle to the LAP in the collective space. While the NEB method will, by design, drive the line of images towards either the MEP or LAP, the iterative scheme chosen greatly impacts the total number of iterations required before the solution converges. In the early implementations, a simple velocity Verlet algorithm \cite{Verlet1967} was used to adjust the position of the images step to step~\cite{Mills1994,Mills1995,Jonsson1998,Henkelman2000a,Henkelman2000}. This approach is robust and relatively stable, though the convergence can be slow for flatter surfaces where the images are not pulled strongly to their optimal positions. To aid this process, the Fast Inertial Relaxation Engine (FIRE) was proposed~\cite{Bitzek2006} to accelerate convergence without sacrificing stability. The method was subsequently updated~\cite{Guenole2020} to further improve performance. Indeed, in our tests, the inertial algorithm regularly outperforms the velocity Verlet algorithm by an-order-of-magnitude reduction in iterations at the same convergence criteria. With this, our implementation of the NEB approach is defined. The algorithm itself is outlined in Algorithm~1 in SM \cite{SM}. The force used in the optimization step for each image, $\vec{F}_i^{\rm opt}$, is constructed by adding the perpendicular component of the action gradient to the spring force $\vec{F}_i^{k}$ between the images, \begin{equation} \vec{F}_i^{k} = k (|\vec{q}_{i+1} - \vec{q}_i| - |\vec{q}_i - \vec{q}_{i-1}|)\vec{\tau_i}, \end{equation} where $k$ is a tunable parameter that controls the strength of the spring force and $\vec{\tau_i}$ is the unit vector tangent to the line of images from image $i-1$ to image $i+1$. The spring force on the endpoints is defined differently: \begin{equation} \vec{F}_1^k=k|\vec{q}_2-\vec{q}_1|,\quad \vec{F}_N^k=k|\vec{q}_{N}-\vec{q}_{N-1}|. \end{equation} The total force acting on the interior images is then \begin{align}\label{Eq:ImagesForces} \vec{F}_{i}^{\rm opt} = \vec{F}^{k}_{i} + \vec{g}_{i}^{\perp}. \end{align} The NEB approach is illustrated in Fig.~\ref{Fig: Bifurcation} for the case of bimodal tunneling from the g.s. minimum to the OTL on an analytic PES defined by: \begin{align}\label{epsbimodal} V(\vec{q}) &= 3.17 + 2e^{-5\big((x-1)^{2} + (y-\frac{1}{2})^{2}\big)} -3 e^{-(x^{2} +y^{2})} \nonumber \\&-\frac{1}{2}(3x +y), \end{align} where $\vec{q} = (x,y)$. The inset shows the forces on the images of the NEB grey path, which has not converged yet to the black path. The spring force $\boldsymbol{F}^k_i$ keeps the images from drifting too much from each other, while the perpendicular part of the action gradient $\boldsymbol{g}_i^\perp$ pushes them towards the nearest stationary action path. This example shows that the NEB algorithm, depending on the initial locations of the images, will converge to a local stationary path, not necessarily the least action path. For the endpoint, $i=N$, one can choose to either fix the position of the image or to allow the image to move towards the outer turning surface. In the second case, a harmonic restraint term is added to the spring force to construct $\vec{F}_{N}^{\rm opt}$, \begin{equation} \label{eqn:endpoint} \Vec{F}_{N}^{\rm opt} = \Vec{F}_{N}^{k} - \left[\Vec{F}_{N}^{k} \cdot \vec{f}(\vec{q}_{N}) - \eta (V(\vec{q}) - E)\right]\vec{f}(\vec{q}_{N}), \end{equation} where $\vec{f} = -\vec\nabla V / |\vec\nabla V|$ and $\eta$ determines the strength of the harmonic restraint term~\cite{Asgeirsson2018}. This force pulls the endpoint $i=N$ very quickly to the outer turning surface and helps find the optimal outer turning point. The default iteration scheme used in our implementation is the inertial algorithm mentioned above, though a standard Verlet minimizer is also included in the PyNEB python package \cite{PyNEB}. The structure of the NEB solver is modular and allows for the simple replacement of components like the minimizer, allowing for easy checks on the convergence and parameters that describe the iterative scheme. \subsection{Grid-Based Methods}\label{Section: grid-methods} Some traditional methods to compute the LAP begin by computing the PES and the collective inertia on a grid of collective coordinates. The calculation of the LAP is then reduced to finding the path through the grid points that minimizes a discrete approximation of the action. Two methods that we have benchmarked are the dynamic programming method (DPM) \cite{Baran1981}, and Dijkstra's algorithm (DA) \cite{Dijkstra}. Here, both will be described for two-dimensional (2D) grid, with points labelled by $\vec{q}_{ij}=(x_i,y_j)$ ($i=1,\ldots,N$, $j=1,\ldots,M$). Both methods can be straightforwardly extended to a higher-dimensional grid. Dynamic programming is a general mathematical technique for solving multi-decision problems by breaking the problem down into simpler overlapping sub-problems. It was first adapted to the action integral minimization in Ref.~\cite{Baran1981} and used in \cite{Sadhukhan2013} to determine the LAP. This adaptation is what we refer to as the DPM. The DPM approximates the LAP between an initial point, $\vec{q}_\textrm{in}$, and a final point, $\vec{q}_\textrm{fin}$. This method finds paths that traverse diagonally from a given cell: from cell $\vec{q}_{ij}$, only cells $\vec{q}_{i+1,j}$ can be reached, for $j=1,\ldots,M$. The allowed cells are highlighted in red in Fig.~\ref{grid_method}. The LAP from $\vec{q}_\textrm{in}$ to $\vec{q}_\textrm{fin}$ is constructed iteratively as follows: for a cell $\vec{q}_{ij}$, there are $M$ possible paths, each passing through a cell at $x_{i-1}$. The LAP from $\vec{q}_\textrm{in}$ to $\vec{q}_{ij}$ is selected and stored in memory. This is repeated for every cell with $x=x_i$, for a total of $M$ possible paths. Once $\vec{q}_\textrm{fin}$ is reached, there are only $MN$ paths (out of a total of $M^N$ paths), and the LAP is selected from these. The DPM algorithm is detailed in Algorithm 2 in SM \cite{SM}. \begin{figure}[htb!] \label{grid_method} \includegraphics[width=0.48\textwidth]{Figures/Dijkstra_vs_dpm.pdf} \caption{Different types of paths that can be found in the different grid-based methods. The single node $\vec{q}_{ij}$ can reach the red (blue) regions in DPM (Dijkstra's algorithm). The initial and final points are marked.}\label{Fig: grid_methods} \end{figure} Dijkstra's method \cite{Dijkstra} is similar to DPM, in that it breaks down the large optimization problem into a set of smaller problems. Given a cell $\vec{q}_{ij}$, the action to every neighbor $\vec{q}_{i'j'}$ is calculated as if the path to $\vec{q}_{i'j'}$ passes through $\vec{q}_{ij}$. If this action integral is smaller than that along the current path to $\vec{q}_{i'j'}$, $\vec{q}_{i'j'}$ is said to come from $\vec{q}_{ij}$. This is repeated, starting from $\vec{q}_\textrm{in}$, until $\vec{q}_\textrm{fin}$ is reached. Figure \ref{grid_method} shows the nearest-neighbors of $\vec{q}_{ij}$ (the cell marked in green) in a blue square. Dijkstra's algorithm is described in Algorithm~3 in the SM \cite{SM}. Dijkstra's algorithm can find paths that pass through multiple cells with the same $x_i$ value, or even paths that backtrack. DPM cannot find such paths. However, DPM can find paths that jump from $\vec{q}_{ij}$ to $\vec{q}_{i+1,j'}$, for any $j'$, while Dijkstra's algorithm is limited to $j'=j-1,j,j+1$ (see Fig.~\ref{grid_method}). For fission calculations, one frequently takes the $x$ coordinate as the quadrupole moment $Q_{20}$, and fission can be viewed as collective motion in which $Q_{20}$ continuously increases towards scission. So, the paths that Dijkstra's algorithm can find, that DPM cannot, are rather unlikely. In general DPM tends to find paths with a smaller action than Dijkstra's algorithm, see Sec.~\ref{sec:results}. \subsection{Euler Lagrange Equations} In order to find the LAP for the functional~\eqref{action-integral} using the EL equations \cite{weinstock1974calculus}, we first parametrize the trajectory $\vec q$ by a time variable $t$, i.e., $\vec{q}=\vec q(t)$ with $t \in [0,t_f]$. This is done in order to explicitly account for the arclength $ds = (\sum dq_i^2)^{1/2}$. In terms of $t$, the action integral~\eqref{action-integral} reads: \begin{equation}\label{Eq: ELE Action} \begin{split} &S(L) = \\ &\int_{0}^{t_f} \sqrt{2\big(V_\text{eff}[\vec q(t)]-E_0 \big)}\Big(\sum_{ij}^n\mathcal{M}_{ij}[\vec q(t)]\dot{q}_i\dot{q}_j\Big)^{1/2}dt \\ &= \int_{0}^{t_f} \mathcal{L}(\vec q,\dot{\vec q})dt, \end{split} \end{equation} where $\dot{q}_i\equiv dq_i/dt$, and $\mathcal{L}$ is the corresponding Lagrangian. The associated EL equation can be written as: \begin{equation}\label{Eq: ELE} \frac{\partial \mathcal{L}}{\partial q_i} = \frac{d}{dt} \Big( \frac{\partial \mathcal{L}}{\partial \dot{q}_i}\Big), \end{equation} with the boundary conditions: $\vec q(t=0)=\vec q_\text{in}$ (the initial location) and $\vec q(t=t_f)=\vec q_\text{\textrm{fin}}$ (the final location). In order to numerically solve Eq.~\eqref{Eq: ELE} we use the shooting method \cite{Shooting}. That is, we start at the initial position $\vec q_\text{in}$ and vary the direction and orientation of the initial ``velocity'' $\dot{\vec q}(t=0)$. We use a numerical differential equation solver to propagate the solution until we find an initial condition that satisfies $\vec q(t_f)=\vec q_{\textrm{fin}}$. Finding such initial conditions can present some challenges, which we discuss in the SM \cite{SM}. The EL approach is equivalent to what is done in Ref.~\cite{Schmid1986} where the eikonal equation is solved by the method of characteristics. Each different trajectory obtained by varying $\dot{\vec q}(t=0)$ corresponds to one of the characteristics of the leading order (cf. Eqs.(2.8) and (4.3) of \cite{Schmid1986}). It is worth noting that if the imaginary part of the phase of the wave function $W(\boldsymbol{q})$ is negligible, as is the case of the motion in the deep subbarrier region, then the eikonal equation for $W$ is a valid approximation \cite{Kapur1937}. The trajectories corresponding to the stationary functional~\eqref{action-integral} are equivalent to the solutions of the eikonal equation for $W$ (see Eqs.~(11) and (13) of \cite{Kapur1937}). A connection between the eikonal equation, the dynamic programming approach, and a variational principle in the context of geometrical optics is discussed in Ref.~\cite{lakshminarayanan1997dynamic}. \section{Results}\label{sec:results} \subsection{Analytic surfaces: Illustrative examples}\label{Sec: Analytic Surfaces} We benchmark the performance of the NEB method by comparing the LAP found using NEB (denoted as NEB-LAP) to the paths found using the DPM, DA, and EL approaches for analytic surfaces defined in terms of the position vector $\vec{q}=(x,y)$. Throughout this section, we assume a constant inertia ${\cal M}_{ij}=\delta_{ij}$. Within the NEB framework, the action functional \eqref{action-integral} can develop some noise as the NEB algorithm approaches the final action. This noise is a function of the NEB hyperparameters and the optimization method used. All surfaces discussed in this section are released as example cases with PyNEB \cite{PyNEB}. In the analytic cases, the NEB is initialized by fixing an initial and final points $\vec{q}_{\textrm{in}}$ and $\vec{q}_{\textrm{fin}}$, respectively, and defining a linear trajectory connecting them. The NEB algorithm is then iterated until convergence is reached. Grid-based methods use a grid spacing of $\Delta x = 0.1$ along the x-axis and $\Delta y = 0.005$ along the y-axis for all analytic surfaces. Details of the numerical methods used for solving the EL equations for all surfaces are discussed in SM \cite{SM}. The action values for each surface considered are included in Table \ref{analytic-table}. Action integrals in Table~\ref{analytic-table} are evaluated using linearly interpolated trajectories over 500 uniformly-distributed points. We compute both LAP and MEP in the NEB framework. Since the MEP is a solution of Eq.~(\ref{MEP-def}), images along the path converge to critical points on the surface depending on the position of the boundary images at $\vec{q}_{\textrm{in}}$ and $\vec{q}_{\textrm{fin}}$. Critical points on the surface $V(\vec{q})$ contained in the MEP can be extracted by calculating $\vec{\nabla}V$ along the path and are classified by computing the eigenvalues of the Hessian at those points. \begin{table}[htb] \caption{Action integrals for the 6-Camel-Back (CB-S and CB-A) and M{\"u}ller-Brown (MB) surfaces. The integrals have been calculated using a linear spline interpolation evaluated at 500 points along each trajectory.} \label{analytic-table} \begin{ruledtabular} \begin{tabular}{cccccc} & \textrm{NEB-MEP}& \textrm{NEB-LAP}& \textrm{DPM}& \textrm{EL}& \textrm{DA}\\ \colrule CB-S & 5.522 & 5.518 & 5.524 & 5.536 & 5.563\\ CB-A & 6.793 & 6.404 & 6.405 & 6.407 & 6.886\\ MB & 28.491 & 22.875 & 22.909 & 22.871 & 23.427\\ \end{tabular} \end{ruledtabular} \end{table} First, we consider the symmetric 6-Camel Back potential (CB-S) \cite{More2004} defined as \begin{equation} V_{\rm CB-S}(\vec{q}) = \big(4 - 2.1 x^{2} + \frac{1}{3} x^{4} \big) x^{2} + x y + 4(y^{2} - 1) y^{2} \label{eq:cbs} \end{equation} In this example, we seek the LAP connecting the local minimum located at $\vec{q}_{\textrm{in}} = (1.70, -0.79)$ to the local minimum located at $\vec{q}_{\textrm{fin}} = (-1.70, 0.79)$. Figure \ref{fig-camel-symmetric} shows the CB-S PES normalized to zero at its global minimum together with the calculated NEB-MEP, NEB-LAP, DPM, EL, and DA trajectories. The action integrals along these trajectories are listed in Table \ref{analytic-table}. \begin{figure}[t!] \includegraphics[width=0.5\textwidth]{Figures/camel_symm1.pdf} \caption{The symmetric camel-back PES $V_{CB-S}(\vec{q})$ normalized to its global minimum together with the calculated NEB-MEP (red), NEB-LAP (magenta), DPM (black), EL (cyan), and DA (lime) trajectories. Black stars indicate saddle points and yellow crosses mark local minima.} \label{fig-camel-symmetric} \end{figure} \begin{figure}[t!] \includegraphics[width=0.5\textwidth]{Figures/camel_asymm1.pdf} \caption{Similar as in Fig.~\ref{fig-camel-symmetric} but for the asymmetric camel-back surface $V_{\textrm{CB-A}}(\vec{q})$. For the video illustrating the NEB determination of both LAP and MEP, see the SM \cite{SM}.} \label{fig-camel-asymmetric} \end{figure} The MEP and the LAPs computed by using the NEB, EL, DPM, and DA methods are very similar. However, the DA trajectory slightly deviates from the other ones. This is because DA is more constrained by the grid spacing than DPM: regardless of the grid spacing, DA can only consider its immediate neighbors, while DPM does not have this constraint (see Fig.~\ref{Fig: grid_methods} and \ref{Section: grid-methods}). As indicated by Fig. \ref{fig-camel-symmetric}, the final action values for the LAP obtained by the NEB, DPM, and EL methods agree well with the MEP. However, the MEP and LAP are not necessarily equivalent in general; the MEP can be viewed as an approximation of the LAP. A detailed discussion on the conditions for the MEP to be an LAP is contained in the SM \cite{SM}. To see the MEP limitations, we consider an asymmetric variant of the Camel-Back potential (CB-A) \begin{align} V_{\rm CB-A}(\vec{q}) = V_{\rm CB-S}(\vec{q}) + \frac{1}{2}y \end{align} where the end points of the local minima are $\vec{q}_{\textrm{in}} = (1.70, -0.8)$ and $\vec{q}_{\textrm{fin}} = (-1.70, 0.76)$. Figure~\ref{fig-camel-asymmetric} shows the MEP trajectory which is markedly different from the LAP solutions and corresponds to an appreciably larger action integral. Still, the MEP can be used for finding critical points (minima and saddles) on the surface. The M{\"u}ller-Brown potential is a canonical example of a PES used in theoretical chemistry \cite{muller1979location,Koistinen2017,Asgeirsson2018}. The M{\"u}ller-Brown surface shown in Fig.~\ref{fig-MB} is defined as \begin{equation} V_{\rm MB}(\vec{q}) = \sum_{i=1}^{4} A_{i}e^{a_{i}(x-x_{0_{i}})^{2} + b_{i}(x-x_{0_{i}})(y-y_{0_{i}}) + c_{i}(y-y_{0_{i}})^{2}}, \end{equation} where we use the same set of parameters as in Ref.~\cite{muller1979location}, namely: $\vec{A} = (-200, -100, -170, 15)$, $\vec{a}=(-1,-1,-6.5,0.7)$, $\vec{b}=(0,0,11,0.6)$, $\vec{c}=(-10,-10,-6.5,0.7)$, $\vec{x}_{0}=(1,0,-0.5,-1)$, and $\vec{y}_{0}=(0,0.5,1.5,1)$. \begin{figure}[t!] \includegraphics[width=\linewidth]{Figures/MB_plot.pdf} \caption{Similar as in Fig.~\ref{fig-camel-symmetric} but for the shifted M{\"u}ller-Brown surface. The inset shows the LAP pathways close to the initial point $\vec{q}_\textrm{in}$. The yellow dashed line shows the vertical. As can be seen, all paths except for the DPM curve start by moving to the left of the vertical.} \label{fig-MB} \end{figure} The MEP follows the bent trajectory that goes through the critical points: two saddle points and one local minimum. This trajectory markedly differs from the LAPs, which are in a rough agreement. The MB surface highlights a problem with the DPM. As mentioned in Sec.~\ref{Sec: Methods algorithms}, the DPM can only search a single direction of each coordinate axis of the domain. In the case of the Muller-Brown surface, the DPM cannot search for trajectories bending back in the negative-$x$ direction. As seen in the inset of Fig.~\ref{fig-camel-symmetric}, the NEB, EL, and DA methods start their trajectories moving backwards in $x$ from the initial point $\vec{q}_\textrm{in}$. The DPM path, on the other hand, always moves in the positive-$x$ direction. Consequently, the action integral along the DPM path is slightly larger than in the other methods. \subsection{Realistic calculations}\label{Sec: Realistic} To illustrate the performance of the NEB method and other approaches to the LAP in realistic cases, we carried out nuclear EDF calculations for $^{232}$U in two collective coordinates and $^{240}$Pu in three collective coordinates. In the particle-hole channel we used the Skyrme functional SkM$^{*}$ \cite{BARTEL198279}, which is often employed in fission studies. The particle-particle interaction was approximated by the mixed density-dependent pairing force~\cite{DobaczewskiPairing}. In the case of $^{232}$U, we considered two collective coordinates $\vec{q}\equiv(Q_{20},Q_{30})$ and for $^{240}$Pu we took three collective coordinates $\vec{q}\equiv(Q_{20},Q_{30},\lambda_2)$. The axial quadrupole and octupole moment operators are defined as in Ref.~\cite{DobaczewskiHFODD}: \begin{align} \hat{Q}_{\lambda 0}(r,\theta) = \mathcal{N}_{\lambda} \sqrt{\frac{2\lambda+1}{4\pi}} r^{\lambda} P_{\lambda} (\cos\theta) \end{align} where $P_{\lambda}$ is the Legendre polynomial, $\mathcal{N}_{2}=\sqrt{\frac{16\pi}{5}}$, and $\mathcal{N}_{3}=1$. The collective coordinate $\lambda_2=\lambda_{2n}+ \lambda_{2p}$ defined in Sec.~\ref{sec:EDF} represents the dynamic pairing fluctuations. The value of $\lambda_{2\tau}$= 0 corresponds to static HFB pairing. As in Ref.\,~\cite{Sadhukhan2014}, to render collective coordinates dimensionless, we use dimensionless coordinates ${x_i}$ defined as \begin{eqnarray}\label{eqn:scale_param} x_{i}&=&\frac{q_{i}}{\delta{q_{i}}}, \end{eqnarray} where $\delta{q_{i}}$ are the scale parameters used in determining numerical derivatives of density matrices in Eq.~(\ref{equation-F}). Here we took $\delta{Q_{20}} = 1$\,b, $\delta{Q_{30}} = 1$\,b$^{3/2}$ and $\delta{\lambda_2}= 0.01$\,MeV. \subsubsection{Two dimensional case: SF of $^{232}$U} The PES was computed by solving the HFB equations using the parallel axial solver HFBTHO(v3.00)\cite{PEREZ2017363}. The large stretched harmonic oscillator basis of $N= 25$ major shells was used to guarantee good convergence. We adopted a $458\times501$ grid with $0\le Q_{20} \le{457}$\,b and $0\le Q_{30}\le{50}$\,b$^{3/2}$. To apply the NEB method, which involves local gradient calculations at arbitrary values $\vec q$, we interpolate the PES and the inertia tensor on the mesh. Because the grid is two dimensional, a cubic spline interpolator suffices. Close to the $Q_{30}=0$ axis, we take into account the mirror symmetry of the PES by setting $V(-Q_{30})=V(Q_{30})$. Finally, since NEB updates occasionally push an image outside of the computed PES mesh, we extended the PES to grow exponentially with the distance outside the mesh, to smoothly push images back into the evaluated region. \begin{figure}[htb!] \includegraphics[width=1.0\columnwidth]{Figures/232U.pdf} \caption{The PES of $^{232}$U in the $(Q_{20},Q_{30})$ plane calculated with SkM$^*$. Solid lines mark the LAPs and MEP obtained with the constant inertia tensor; dotted lines correspond to the non-perturbative inertia tensor. The OTL is shown in white. The blue, orange, purple and black curves represent the LAPs calculated using the NEB, DPM, EL, and DA methods, respectively. The green curve is the MEP, which was also calculated using NEB.}\label{Fig: U232 Paths} \end{figure} Figure~\ref{Fig: U232 Paths} shows the two-dimensional PES of $^{232}$U. The least action fission pathway which goes from the g.s. at $\vec{q}_\textrm{in}$=(24\,b, 0) to the exit point $\vec{q}_\textrm{fin}$=(281\,b, 37\,b$^{3/2}$) is calculated using the methods explained in Sec.~\ref{Sec: Methods algorithms}. To select the endpoint $\vec{q}_\textrm{fin}$, we compute the LAP using DPM for all points on the OTL, and select the point with the lowest action integral. This point is then used as the exit point for the other methods. While NEB does not require a fixed endpoint in general, we fix the endpoint here in order to facilitate inter-method comparison. The MEP path is calculated using the NEB method. The action integral computed with different methods is shown in Table \ref{table:nuclei_acts}. When computing the action, we interpolate the paths using a linear spline interpolator, and the action integral is computed using 500 evaluations along the path. This reduces the differences in the action that may arise from using a different number of points along the path (for instance, NEB gives a similar path to DPM using as few as 30 images). For all paths, we compute the action using the inertia tensor evaluated along the path. As can be seen, the action values computed for $^{232}$U using different methods agree well, with DA being the worst performer. As seen in Fig.~\ref{Fig: U232 Paths} and Table\,\ref{table:nuclei_acts} the MEP is very close to the LAP. This is because the static fission pathway (i.e., MEP) is fairly straight and the fission valley is well delineated. Note that perfect agreement is not expected, and in fact was not observed for the analytic surfaces, either. This is due in part to the different approximations used in each method --- for DPM and DA, this is the grid spacing; for NEB, this is the number of images and approximate treatment of derivatives; and for EL, this is a variety of simplifications described in Sec. 3 in the SM \cite{SM}. Additional variation in the quality of the interpolator further hampers agreement beyond what is listed. \begin{table}[htb] \caption{Action integrals for $^{232}$U computed with different methods. The paths computed using the constant and non-perturbative inertia tensor are labelled as ``con.'' and ``n.-p.'', respectively.} \label{table:nuclei_acts} \begin{ruledtabular} \begin{tabular}{llccccc} & & \textrm{NEB-MEP}& \textrm{NEB-LAP}& \textrm{DPM}& \textrm{EL}&\textrm{DA}\\ \colrule \multirow{2}{*}{${}^{232}$U} & con. & 174.5 & 174.2 & 174.2 & 174.9 & 175.8 \\ & n.-p. & - & 173.6 & 173.3 & 175.0 & 178.5 \\\\[-5pt] \multirow{2}{*}{${}^{240}$Pu} & con. & 19.09 & 18.98 & 19.21 & 19.01 & 22.85 \\ & n.-p. & - & 16.54 & 16.47 & 18.18 & 30.50 \\ \end{tabular} \end{ruledtabular} \end{table} \subsubsection{Three dimensional case: SF of $^{240}$Pu} The SF of $^{240}$Pu in several collective coordinates was studied in Ref.~\cite{Sadhukhan2014} where the details pertaining to the computation, grid size, etc., can be found. Between the g.s. minimum and the fission isomer (FI), the fission pathway is affected by triaxial degrees of freedom. Between the FI and the outer turning surface (OTS), however, the predicted fission trajectory is axial. In this paper, we consider the fission of the FI of $^{240}$Pu so the OTS corresponds to the FI energy. For three-dimensional tunneling, the system of equations that must be solved to construct a global spline interpolator is too large for practical applications. Instead, we use piecewise linear interpolation. The PES at $\lambda_2=0$ for $^{240}$Pu shown in Fig.~\ref{Fig: Pu240 2d Paths} varies very smoothly in the barrier region where the potential energy is larger than the energy of the FI, and so this interpolation scheme is reasonable. \begin{figure}[htb!] \includegraphics[width=1.0\columnwidth]{Figures/240Pu_2d.pdf} \caption{The PES for $^{240}$Pu in the space of collective coordinates $Q_{20}$, $Q_{30}$ with $\lambda_2 = 0$. Only the region beyond the fission isomer is shown. The energy is normalized to the energy of the fission isomer. The OTL is shown in white. The MEP (green) practically coincides with the LAPs calculated with the constant inertia using the NEB (blue), DPM (orange), and EL (purple) methods. }\label{Fig: Pu240 2d Paths} \end{figure} Figure~\ref{Fig: Pu240 Paths} shows the LAPs for $^{240}$Pu computed with the NEB, DPM, and EL methods in three dimensions (3D). The pathways begin at the FI minimum at $\vec{q}_\textrm{in}=(Q_{20}^{0}=87\,\text{b}, Q_{30}^{0}=0\,\textrm{b}^{3/2}, \lambda_{2}=0.0)$ and the exit point was chosen for DP in the same way as the $^{232}$U results before. The NEB endpoint in this case was allowed to vary according to Eq.~\ref{eqn:endpoint}, better representing standard procedure for production runs. The exit points $\vec{q}_\textrm{fin}$ predicted by NEB ($185.1\,\text{b}, 18.4\,\text{b}^{3/2}, 3.3\,\text{MeV}$), DPM ($184.0\,\text{b}, 18.6\,\text{b}^{3/2}, 4.8\,\text{MeV}$) and EL ($179.8\,\text{b}, 17.7\,\text{b}^{3/2}, 0.0$) then differ. When the collective mass is held constant, all methods find very similar paths in the $\lambda_{2}=0.0$ plane, which are also shown in Fig.~\ref{Fig: Pu240 2d Paths}. The paths vary more when the non-perturbative inertia tensor is used, with the main difference between the NEB and DPM paths appearing in the region close to the FI minimum; beyond the saddle point, both paths are similar. \begin{figure}[htb!] \includegraphics[width=1.0\columnwidth]{Figures/3dplot.pdf} \caption{The PES for $^{240}$Pu in the collective coordinates $Q_{20}$, $Q_{30}$ and $\lambda_2$. The 2D cross section at $\lambda_2=0$ shown in Fig.~\ref{Fig: Pu240 2d Paths} is indicated. The blue, orange, and purple curves are the LAP, calculated using the NEB, DPM, and EL methods, respectively. The non-perturbative inertia tensor was used for the dashed curves. The OTS is indicated by the dark blue contour surface. } \label{Fig: Pu240 Paths} \end{figure} As seen in Table \ref{table:nuclei_acts}, the NEB and DPM are in a good agreement. In general, one would expect a better performance from NEB as this method is not constrained to a grid (this is true in the case of the analytic surfaces discussed in Sec. \ref{Sec: Analytic Surfaces}). However, in rare cases, the DPM produces a slightly lower action than the NEB. In such cases the NEB converges to an even lower action if is initialized with the DPM result. This suggests that for tunneling in more than 2D, a combination of NEB and DPM might be beneficial. \section{Conclusions}\label{sec:summary} Finding the path that minimizes the action integral can be extremely challenging since it involves searching over the space of all continuous paths that fulfill the boundary conditions. Each method explored in this paper simplifies such task in different ways. DPM and DA project the PES onto a finite grid and explore decisions in making the path between the boundary conditions. In the EL approach the surface is modified in several ways to smooth the relation between initial conditions and the end point of the trajectory. The NEB method reduces the original search over continuous path into considering only piece-wise linear paths, the number of pieces given by the number of images. It is this simplification that makes the NEB robust and accurate, since the total action now becomes a smooth function of the position of the images, a function that can straightforwardly be numerically minimized by gradient descent methods. A significant advantage of the NEB is that it can accommodate any initial positions of the images, which speeds the convergence appreciably if a good prior guess of the LAP is provided. Other methods lack for such incorporation of prior knowledge. Finally, the resolution of the NEB for a rapidly varying surface can be adjusted locally by increasing the amount of images or spring constants, while for DPM and DA the entire grid resolution would have to be increased, giving an appreciable toll on the computational cost. For both analytic and realistic potential energy surfaces the NEB robustly produces a LAP. In the cases studied, NEB outperforms the EL and DA methods, and produces close results to those of the DPM with usually lower action integral. For many-dimensional tunneling, initiating the NEB method from the DPM path might be a winning strategy. A huge advantage of the NEB over other methods is that it can efficiently and accurately estimate exit points. By exploring different initial conditions for the positions of the images which lead to distinct exit points, one can use the NEB method to study the phenomenon of multimodal fission. An example of such an application is shown in video~1 in the SM \cite{SM}. Whilst other methods can find a least-action trajectory for an arbitrary final point placed on the OTL, as done, e.g., in Refs.~\cite{Sadhukhan2013,Sadhukhan2014,Sadhukhan2016,Sadhukhan2017,Zhao2015,Zhao2016,Mercier2021}, they cannot guarantee that this trajectory is stationary. All such trajectories can be gradually transformed into a stationary pathway by moving the final point along the OTL towards the exit point, see Fig.~\ref{Fig: Bifurcation}. In this paper we also explored the minimum-energy (or static) path. We adjusted our NEB algorithm to generate MEPs, including the determination of local minima and saddle points. The necessary conditions for an MEP to also be an LAP are discussed in the SM \cite{SM}. Video~2 in the SM \cite{SM} illustrates the way the NEB method generates LAP and MEP. An important contribution of this work is providing a beta release of the PyNEB package, a python suite of codes that implement the NEB algorithm described in this paper. The package can be found in \cite{PyNEB} together with the respective documentation and code samples serving as a tutorial for its use. A comprehensive investigation into the intricacies of the numerical implementations and performance of the package itself will accompany the version $1.0$ release. The NEB approach can be readily paired with accelerated DFT calculations, such as the recent applications of Gaussian process regression to PES emulation~\cite{Koistinen2017,Torres2019}. In these works, a Gaussian process is used to emulate the PES and DFT calculations are only run if the Gaussian process is uncertain as to the actual PES value. As NEB is not a grid-based method, it can sensibly be paired with a Gaussian process emulator that is updated as necessary while NEB runs. In this way, the LAP can be determined using far fewer DFT evaluations than is necessary in DPM. The ability to determine the exit points is essential for determining fission fragment yields \cite{Sadhukhan2020,Sadhukhan2022}. The minimum action provides information on SF half-lives. In this context, the NEB method described in this paper is expected to speed up the global calculations of nuclear fission for r-process simulations and studies of superheavy nuclei stability. \noindent\makebox[\linewidth]{\resizebox{0.3333\linewidth}{1pt}{$\bullet$}}\bigskip \begin{acknowledgments} We are grateful to Edgard Bonilla and Stefan M. Wild for useful comments. This work was supported by the U.S. Department of Energy under Award Numbers DOE-DE-NA0002847 (NNSA, the Stewardship Science Academic Alliances program), DE-SC0013365 (Office of Science), and DE-SC0018083 (Office of Science, NUCLEI SciDAC-4 collaboration) and by the National Science Foundation CSSI program under award number 2004601 (BAND collaboration). \end{acknowledgments} E.F. and D.L. contributed equally to this work. \section{Algorithms} \label{algorithms} \begin{algorithm} \caption{Nudged Elastic Band}\label{alg:neb} \begin{algorithmic} \State Define start and end point \State Initialize images along line \For{Steps} \For{Images} \State Update effective force \State Compute velocity of image \State Translate image \EndFor \State Compute action \State Check convergence \EndFor \end{algorithmic} \end{algorithm} \begin{algorithm} \caption{Dynamic Programming Method}\label{alg:dpm} \begin{algorithmic} \State Define start and end points $\vec{q}_\textrm{in}$ and $\vec{q}_\textrm{fin}$ \For{$i=1,\ldots,n$} \For{$j=1,\ldots,m$} \State Select $\vec{q}_{i-1,j}$ to minimize the action from $\vec{q}_\textrm{in}$ to $\vec{q}_{ij}$, through $\vec{q}_{i-1,j}$ \State Copy the path to $\vec{q}_{i-1,j}$ and append $\vec{q}_{ij}$ to it \EndFor \EndFor \State Append $\vec{q}_\textrm{fin}$ to all paths \State Select the path with the least total action \end{algorithmic} \end{algorithm} \begin{algorithm} \caption{Dijkstra's Algorithm}\label{alg:djk} \begin{algorithmic} \State Define start point $\vec{q}_\textrm{in}$ \State Initialize distance matrix $d(\vec{q}_{ij})\to\infty$, $d(P_\textrm{in})\to0$ \State Mark all points as unvisited \While{Any point is unvisited} \State Select $\arg\min d(\vec{q}_{ij})$ with $\vec{q}_{ij}$ unvisited \For{$\vec{q}_{i'j'}$ unvisited, $\vec{q}_{i'j'}$ a neighbor of $\vec{q}_{ij}$} \State Compute tentative distance \begin{equation} d'=d(\vec{q}_{ij})+\textrm{dist}(\vec{q}_{ij},\vec{q}_{i'j'})\nonumber \end{equation} \If{$d'<d(\vec{q}_{i'j'})$} \State Set $d(\vec{q}_{i'j'})=d'$ \State Mark $\vec{q}_{i'j'}$ as coming from $\vec{q}_{ij}$ \EndIf \EndFor \State Mark $\vec{q}_{ij}$ as visited \EndWhile \end{algorithmic} \end{algorithm} \begin{algorithm} \caption{Euler-Lagrange Method}\label{alg:ELE} \begin{algorithmic} \State Define a tolerance $\epsilon$, initial $\vec{q}_{\rm in}$ and final $\vec{q}_{\rm fin}$ points \State Guess initial velocity $\dot{\vec{q}}_{\rm in}$ and solve Euler-Lagrange equations \While{Final point on path is not within tolerance $|\vec{q}(t_f)-\vec{q}_{\rm fin}|>\epsilon$} \State Vary $\dot{\vec{q}}_{\rm in}$ \EndWhile \end{algorithmic} \end{algorithm} \section{Supplemental videos} The supplemental videos (in .mp4 and .mov formats) show the evolution of the LAP as a function of the number of iteration steps in the NEB method. \begin{itemize} \item Supplementary video 1 (\href{https://pyneb.dev/assets/img/multi-modal.mov}{.mov}, \href{https://pyneb.dev/assets/img/multi-modal.mp4}{.mp4}) - This video shows the convergence behaviour of the LAP on the surface in Eq.~(12), also shown in Fig. 2, as a function of iteration step. Circles indicate the positions of the NEB images. Multiple initial pathways are shown with one boundary image fixed to the global minimum while the remaining images are subject to gradient forces. The harmonic force is applied to the outer boundary image; it pushes the end point to the outer turning line, see Eq.~(13). As the NEB algorithm iterates, the bands converge to two unique stationary paths with unique exit points. The computed action integral for each path is given in the legend. This example shows that the NEB is capable of identifying multiple stationary action paths. \item Supplementary video 2 (\href{https://pyneb.dev/assets/img/asymm-camelback.mov}{.mov}, \href{https://pyneb.dev/assets/img/asymm-camelback.mp4}{.mp4}) - This video shows the convergence behaviour of the minimum energy path (MEP), shown in red, and the least action path (LAP), shown in purple on the asymmetric camel-back surface of Fig.~5. Circles indicate the positions of the NEB images. A linear trajectory connecting two local minima with fixed endpoints is given as an initial guess. The MEP converges to a gradient curve that passes through critical points of the surface whilst the LAP bypasses two saddle points. \end{itemize} \section{MEP and LAP equivalence conditions} \label{LAP-MEP} As discussed in Sec.~I\, the MEP is obtained by computing the steepest descent line on the PES, and the LAP is determined by minimizing the collective action in the many-dimensional collective space. We want to address the necessary conditions for an MEP to be an LAP. Similar to \cite{quapp1984analysis} and Eq.~(7), we define the MEP as a gradient curve $\vec{\gamma}(\tau)$ satisfying the differential equation \begin{align} \dot{\vec{\gamma}} = \sigma(\tau) \vec{\nabla}V_\text{eff}(\vec{\gamma}(\tau)), \label{B1} \end{align} with boundary conditions $\boldsymbol\gamma(0) = \vec{q}_{\textrm{in}}$ and $\vec{\gamma}(1) = \vec{q}_{\textrm{fin}}$. $\tau$ is an arbitrary monotonic parametrization of the curve, while $\dot{\vec{\gamma}}$ represents the local velocity of the trajectory. $\sigma(\tau)$ is a factor to account for the transformation of arc length $s$ in Eq.~(7)\ to the parameter $\tau$ to be used in the EL Eq.~(15). Solutions to Eq.~\eqref{B1} define a trajectory made by a collection of steepest descents and ascents between $\vec{q}_{\textrm{in}}$ and $\vec{q}_{\textrm{fin}}$ in terms of the parameter $\tau$. \par On the other hand, the LAP trajectory $\boldsymbol{q}(\tau)$ is derived by finding a stationary point of the action integral \begin{align} S = \int_{0}^{1} \sqrt{2\Big(V_\text{eff}(\vec{q}) - E \Big)M_{\mu \nu}(\vec{q}) \dot{\vec{q}}^{\mu} \dot{\vec{q}}^{\nu}} d\tau, \label{B2} \end{align} where $M_{\mu \nu}$ is the inertia tensor, and the boundary conditions are $\boldsymbol{q}(0) = \vec{q}_{\textrm{in}}$ and $\boldsymbol{q}(1) = \vec{q}_{\textrm{fin}}$. Here and in the following, Einstein summation convention is assumed. By Beltrami's theorem, one can show that finding stationary solutions to the functional \begin{align} \tilde{S} &= \int_{0}^{1} \Big(V_\text{eff}(\vec{q}) - E \Big)M_{\mu \nu}(\vec{q})\dot{\vec{q}}^{\mu} \dot{\vec{q}}^{\nu} d\tau \\ &= \int_{0}^{1} g_{\mu \nu} \dot{\vec{q}}^{\mu} \dot{\vec{q}}^{\nu} d\tau \label{B3} \end{align} yields solutions that also fulfill the Euler Lagrangian equations for the functional \eqref{B2}. The converse of this statement is not true. Eq.~\eqref{B3} is recognized as being in the same form as the Lagrangian for a free particle in a curved space with the metric tensor $g_{\mu \nu}$. By the stationary action principle, one can derive the geodesic equation: \begin{align} \ddot{q}^{\lambda} + \Gamma^{\lambda}_{\mu \nu} \dot{q}^{\mu} \dot{q}^{\nu} = 0, \end{align} where $\Gamma^{\lambda}_{\mu \nu}$ are the Christoffel symbols: \begin{align} \Gamma^{\lambda}_{\mu \nu} = \frac{1}{2} g^{\lambda \alpha} \Big(g_{\nu \alpha,\mu} + g_{ \mu \alpha, \nu} - g_{\mu \nu, \alpha}\Big). \end{align} The solutions represent a stationary point of both action functionals $\tilde{S}$ and $S$. For the case with constant inertia, the metric tensor has the form: \begin{align} g_{\mu \nu} = \begin{pmatrix} V_\text{eff}(\vec{q}) - E & 0 \\ 0 & V_\text{eff}(\vec{q}) - E \end{pmatrix}, \end{align} where we have assumed that $M_{\mu \nu}\propto \delta_{\mu\nu}$ and is $\vec{q}$-independent. The coordinates $\vec q$ are re-scaled to be dimensionless. The geodesic equation in vector notation becomes \begin{align} \label{Eq: Equations of Motion} (V_\text{eff} - E) \ddot{\vec{q}} = \frac{1}{2} |\dot{\vec{q}}|^{2} \vec{\nabla} V_\text{eff} - \big( \dot{\vec{q}} \cdot \vec{\nabla} V_\text{eff}\big) \dot{\vec{q}}, \end{align} with the boundary condition $\vec{q}(0) = \vec{q}_{\textrm{in}}$ and $\vec{q}(1) = \vec{q}_{\textrm{fin}}$. Suppose now that there exists a trajectory $\vec{q}(\tau)$ satisfying $ \vec{\nabla} V_\text{eff}\big[\vec{q}(\tau)\big] \propto \dot{\vec{q}}$, i.e, the tangent vectors of the solution $\vec{q}(\tau)$ are parallel to the surface gradient $\vec{\nabla}V_\text{eff}$ evaluated along the curve. Since $\vec\nabla V_\text{eff}$ and $\dot{\vec{ q}}$ are parallel, Eq.~\eqref{Eq: Equations of Motion} would imply that $\ddot{\vec{q}} \propto \dot{\vec{q}}$ along the trajectory. This further implies that the tangent vectors of the solution $\vec{q}(\tau)$ are always in the same direction as its acceleration. From this we conclude that in order for the MEP to also be a solution of Eq.~$\eqref{Eq: Equations of Motion}$, such trajectory $\vec{q}(\tau)$ can only be composed of straight lines. Critical points of the PES where $V_\text{eff}-E$ and $\vec{\nabla} V_\text{eff}$ both vanish are the only locations on the surface where the trajectory can bend non-smoothly and still obey Eq.~\eqref{Eq: Equations of Motion}. \par These results significantly restrict the types of PESs for which an MEP can be an LAP. Such conditions on the surface can be obtained by the use of Eq.~\eqref{B1} as an ansatz for the equation of motion \ref{Eq: Equations of Motion}. We first note that, if the trajectory solves Eq.~\eqref{B1}, then: \begin{align} &\ddot{\vec{\gamma}}(\tau) = \dot{\sigma} \vec{\nabla} V_\text{eff} + \sigma \frac{\partial \vec{\nabla} V_\text{eff}}{\partial \vec{\gamma}} \frac{\partial \vec{\gamma}}{\partial \tau}, \\ &\frac{\partial \vec{\nabla} V_\text{eff}}{\partial \vec{\gamma}} \frac{\partial \vec{\gamma}}{\partial \tau} = H[\vec{\gamma}(\tau)] \dot{\vec{\gamma}}, \end{align} where $H$ is the Hessian matrix of second derivatives of the PES evaluated along the gradient curve $\vec{\gamma}$. For such a trajectory, the acceleration along the curve is \begin{align} \ddot{\vec{\gamma}}(\tau) = \dot{\sigma} \vec{\nabla} V_\text{eff} + \sigma^{2} H \vec{\nabla} V_{\rm eff}. \label{curve-acceleration} \end{align} Substituting this into the the geodesic equation we obtain \begin{align} H \vec{\nabla} V_\text{eff} &= -\Big(\frac{|\vec{\nabla} V_\text{eff}|^{2}}{2(V_\text{eff} - E)} + \frac{\dot{\sigma}}{\sigma^{2}} \Big) \vec{\nabla} V_\text{eff}. \label{MEP_condition} \end{align} \begin{figure}[t!] \centering \includegraphics[width=0.5\textwidth]{Figures/AnglesSc.png} \caption{Absolute value of the minimum angle (in degrees) between $\vec{\nabla} V_\text{eff}$ and the eigenvectors of $H$ for the CB-S surface of Fig.~{4} as a function of the coordinates $(x,y)$. The red curve shows the NEB-MEP, the green curve shows the NEB-LAP and the yellow stars indicate critical points of the surface. The inset shows a zoom-in on one region where there is a misalignment in the MEP path; in this region the LAP curves slightly. } \label{award_winning_plot} \end{figure} We see that the gradient $\vec{\nabla}V_\text{eff}\big[\vec{\gamma}(\tau)\big]$ must be an eigenvector of the Hessian $H\big[\vec{\gamma}(\tau)\big]$ for all $\tau$. This shows that the eigenvectors of $H$, or principal directions of the surface $V_\text{eff}$, must also be parallel to the gradient curve tangents along the trajectory. Since we know that if $\vec{\gamma}(\tau)$ is to be a solution of \eqref{Eq: Equations of Motion}, then $\dot{\vec{\gamma}}(\tau) \propto \ddot{\vec{\gamma}}(\tau)$; again showing that the curve $\vec{\gamma}(\tau)$ must be a straight line. In regions where the gradient curve is not parallel to the Hessian eigenvectors, there will be an acceleration along the gradient curve perpendicular to its tangents. In this situation, the function $\vec{\gamma}(\tau)$ cannot be a solution to \eqref{Eq: Equations of Motion} since $\ddot{\vec{q}}(\tau) \not\propto \dot{\vec{q}}(\tau)$ for all $\tau$. Equation~\eqref{MEP_condition} allows us to identify whether an MEP is an exact stationary path by checking if the gradient descent trajectory on the surface $V_{\rm eff}$ misaligns at any point with the eigenvectors of the surface's Hessian $H$. Figure \ref{award_winning_plot} shows the smallest angle between the gradient and the Hessian's eigenvectors for the CB-S surface. In the regions where the angle is small (stationary path can be approximated by straight lines) the MEP are LAP are close. From this analysis, we conclude that in many situations the MEP will not coincide with the LAP. However, in cases where a stationary path can be approximated by straight lines, such as in the CB-S surface for example shown in Fig.~4, the MEP might well approximate the LAP. Since the LAP is a stationary path, small deviations from the LAP could translate into very small errors in the action value. \section{Details on the Euler-Lagrange method} \label{appendix c} Numerically solving Eq.~(15)\ (or Eq.~\eqref{Eq: Equations of Motion} for a constant mass case) for fixed initial conditions, $\vec q(0)=\vec q_\text{in}$ and $\dot{\vec q}(0)$, is straightforward. The main challenge lies in finding the correct initial speed and direction, $|\dot{\vec q}(0)|$ and $\hat{\dot{\vec q}}(0)$, that provides a trajectory that ends on the fixed final point $\vec q(t_f)=\vec q_\text{fin}$. In practice, it is only necessary to obtain an upper value of the initial speed $|\dot{\vec q}(0)|$ since the geometry of the trajectory would be the same with the ``particle'' arriving at an earlier time than $t_f$ to $\vec q_\text{fin}$. In the following, we discuss three important elements of the EL method implementation: \begin{enumerate}[label=(\roman*)] \item How to explore the initial velocity space; \item How to deal with regions where the Lagrangian in Eq.~(14)\ is close to $0$; \item How to deal with boundaries (non-holonomic constraints) of the form $q_i\geq 0$. \end{enumerate} With respect to (i), the main issue is that the map between $\dot{\vec q}(0)$ and $\vec q(t_f)$ is highly non-linear, and it is not trivial to create a cost function that can tell us how to improve an initial condition $\dot{\vec q}(0)$ such that $\vec q(t_f)$ gets closer to $\vec q_\text{fin}$. An automatic algorithm, such as Mathematica's shooting method \cite{Shooting}, is able to find the correct initial conditions for the analytic surfaces discussed in Sec.~IV.1, but fails with the realistic cases of Sec.~IV.2. For these realistic cases we resorted to a binary search for the initial angles of $\hat{\dot{\vec q}}(0)$. This procedure worked well for the $^{232}$U case, but yielded non-optimal results (compared to the NEB and DP paths) for the 3D $^{240}$Pu case. The nonlinearity of $\vec q(t_f)$ as a function of $\dot{\vec q}(0)$ can get stronger in the regions where the effective potential and the inertia tensor in Eq.~(5)\ undergo abrupt {local} variations. Such variations can arise due to numerical noise in the underlying DFT calculation, interpolation errors, or configuration changes due to level crossings. The local fluctuations can easily deflect trajectories being computed with the EL, even though their presence should not in principle impact much the actual path of minimum action. To remedy this situation for the EL approach, we ``smoothed'' (coarse-grained) $V_\text{eff}$ and $\mathcal{M}_{\text{eff}}$ by replacing their values at each location $\vec q$ by their respective averages over a small area around $\vec q$, see Fig.~\ref{Fig: averaged surface}. This was done at the level of the grid points, before building the interpolator. The hope is that, by removing fine-grained details of the surface, the final optimal path would not change much, but the correct $\dot{\vec q}(0)$ will be easier to find. This was not necessary for the analytic surfaces of Sec.~IV.1\. \begin{figure}[t!] \includegraphics[width=0.90\linewidth]{Figures/Averaged.png} \caption{$^{232}$U PES before (a) and after (b) the smoothing procedure. The values near the boundary at $Q_{30}=0$ b have also been altered to guarantee that the vertical component of $\vec{\nabla} V_\text{eff}$ is zero. A similar effect is achieved in the outer turning line by replacing $(V_\text{eff}-E_0)\to \sqrt{(V_\text{eff}-E_0)^2+\epsilon_2^2}$. } \label{Fig: averaged surface} \end{figure} As concerns (ii), the Lagrangian $\mathcal{L}$ in Eq.~(14)\ becoming close to zero is a problem since the numerical integrator for the differential equations becomes unstable; increasingly big steps in $dt$ leave the action almost unchanged. This can happen if either $\big(V_\text{eff}-E_0 \big)$ or $\mathcal{M}_\text{eff}$ approach zero. In the case of a constant inertia tensor, such as in the analytical surfaces in Sec.~IV.1, the equations of motion Eq.~\eqref{Eq: Equations of Motion} resemble a particle of ``mass'' $\big(V_\text{eff}(\vec q)-E_0 \big)$ moving under non-conservative external forces proportional to $\nabla V_\text{eff}$ (this equation produces the same trajectories as the equations of motion of a particle moving on a inverted potential as those shown in \cite{Schmid1986}). If $V_\text{eff}\sim E_{0}$, the trajectory of the particle becomes very sensitive to any force, which is the reason why trajectories diverge near the outer turning line (see Figs. 3 and 4 in \cite{Schmid1986}), or near global minima such a those shown in Fig.~4. The fact that trajectories are unstable around the OTL is because the assumption of the eikonal approximation that the phase of the wave function is real and changes slowly breaks down. We refer the reader to \cite{Schmid1986} for more details on higher order corrections to the eikonal approximation. To avoid the instabilities in regions where $V_\text{eff}\sim E_0$, including a neighborhood around the OTL, we shifted the effective potential by a small constant $\epsilon_{0}$, $V_\text{eff}-E_0 + \epsilon_0$, and added a small diagonal term to the inertia tensor: $\mathcal{M}_\text{eff} + I\epsilon_1$, where $I$ is the identity matrix with the appropriate dimensions. For the analytic surfaces in Sec.~IV.1, we solved the EL equations for several decreasing values of $\epsilon_0$ and then extrapolated the collection of paths to obtain $\vec q(t)$ when $ \epsilon_0 \to 0$. Fig.~\ref{Fig: fig-camel-symmetric paths} shows trajectories obtained in such way for the CB-S surface. \begin{figure}[htb!] \includegraphics[width=0.90\linewidth]{Figures/Paths.png} \caption{PES from Fig.~4. The dashed paths are calculated with values of the potential shift in the range: $\epsilon_0 \in [0.05,3]$, while the solid cyan line is obtained by extrapolating the paths with $\epsilon_0 \in [0.05,0.42]$. The first red dashed path, $\epsilon_0=3$, has an action value around $10\%$ bigger than the last orange path, $\epsilon_0=0.05$, but it took nearly 400 times longer for the shooting method \cite{Shooting} to converge for $\epsilon_0=0.05$. This highlights how the regions where $V_\text{eff}\sim E$ lead to less stable trajectories.} \label{Fig: fig-camel-symmetric paths} \end{figure} In the case of the realistic calculations we only worked with fixed values of $\epsilon_0$ and $\epsilon_1$. For the 2D cases we used $\epsilon_0=1$ MeV and $\epsilon_1=0.01$ ($\epsilon_1$ being dimensionless since we rescaled $Q_{20}$ and $Q_{30}$ to be in the $[0,1]$ range from their original ranges). For the 3D case we used $\epsilon_0=0.005$ MeV and $\epsilon_1=10^{-5}$. With regard to (iii), the numerical integrator can be stopped when the trajectory arrives at such boundary. The main issue is on how to deal with trajectories which for an appreciable interval can lay extremely close to a boundary without crossing it, such as those in Fig.~7\ for $Q_{20}\in \sim[25,70]$ b. For a constant effective mass, it can be seen from Eq.~\eqref{Eq: Equations of Motion} that unless $\vec\nabla V_{\rm eff}$ has zero vertical component, the particle will either be pulled down close to the boundary and touch it, or be repelled from it in the upward direction and be driven away. To tackle this issue, we altered the effective potential and effective mass in a small vicinity around the line $Q_{30}=0$ in such a way that the PES and effective mass values are not drastically changed, but we ensure that the vertical components of their gradients are zero along such line. This creates a ``canal effect'' with the particle traveling close to the boundary for a finite time. A similar ``canal effect'' was used to explore the behavior of trajectories near the OTL and improve stability of the solution in that region. This was done by replacing the effective potential by $(V_\text{eff}-E_0)\to \sqrt{(V_\text{eff}-E_0)^2+\epsilon_2^2}$, with $\epsilon_2$ being a small constant. The gradient of this new function is zero at the OTL. In the realistic examples, we chose the value $\epsilon_2=0.5$ MeV. We conclude this section by emphasising that, for the EL approach to work, many simplifications and approximations had to be made. On the other hand, the NEB approach is free from the difficulties (i)-(iii). By using local surface derivatives and global springs forces, the NEB is effectively less sensitive to the noise in the PES. The NEB also has the advantage that the end point does not have to be predefined. For the EL approach, an undefined ending point translates into a huge computational cost. Finally, constraints of the form $q_i\geq0$ are easily implemented on the NEB, without causing any instability of the algorithm or requiring the surface to be altered in any way. \section{Introduction} \label{sec:introduction} Fission is a fundamental nuclear decay that is important in many areas of science, ranging from structure and stability of heavy and superheavy nuclei \cite{krappe2012,Schmidt2018,GiulianiRMP} to studies devoted to physics beyond the standard model of particle physics~\cite{vogel2015} and the synthesis of heavy elements~\cite{horowitz2018,vassh2019,Giuliani2020}. Theoretically, the nuclear fission process is an example of the nuclear large-amplitude collective motion originating from the single-particle motion of individual nucleons. Due to the complexity of this process, our understanding of nuclear fission is still incomplete. For the state of affairs in this field, we refer to the recent review \cite{schunck2016,Bender2020}. When it comes to realistic predictions, the self-consistent nuclear energy density functional (EDF) method \cite{BenderRMP,SchunckBook} has proven to be very successful in terms of quantitative reproduction of fission lifetimes and fragment yields. Unfortunately, realistic self-consistent fission calculations in a multidimensional collective space, based on the microscopic input, are computationally expensive when it comes to large-scale theoretical fission surveys. Given the computational cost of microscopic methods and the large number of fissioning nuclei that are, e.g., expected to contribute to the astrophysical r-process nucleosynthesis, calculations have mostly relied on simple parametrizations or highly phenomenological models. The new perspective is offered by state-of-the-art theoretical frameworks and modern computational techniques that promise to speed up the calculations to be able to carry out quantified global fission surveys for multiple inputs \cite{Bender2020}. This study is concerned with finding the optimal pathway during the tunneling motion phase of spontaneous fission (SF). Such a trajectory, dubbed the least-action path (LAP), is obtained by minimizing the collective action in a many-dimensional collective space \cite{Kapur1937,Brack1972}. A number of techniques have been proposed to deal with this challenging task. In the early application \cite{Ledergerber1973}, the trial pathways were assumed in a parametrized form and the LAP was obtained by minimizing the penetration integral with respect to the variational parameters. Grid-based techniques such as the dynamic-programming~\cite{Baran1981} and Ritz~\cite{Baran1978} methods have been used in numerous EDF calculations of LAPs \cite{Sadhukhan2013,Sadhukhan2014,Sadhukhan2016,Sadhukhan2017,Zhao2015,Zhao2016,Mercier2021}. In Refs.~\cite{Schmid1986,Eckern1992,Kindo1989,Iwamoto1992,Iwamoto1994,Scamps2015} LAPs were obtained by solving the eikonal equation by the method of characteristics. Effectively, this method can be related to a quantum mechanical propagation in imaginary time that amounts to solving the classical equations of motion in an inverted potential. Within this approach, only one trajectory, called the escape path, arrives at the outer turning surface with zero velocity. Other trajectories, corresponding to different initial conditions, cannot reach the outer turning surface. In this paper, we compare grid-based approaches to the LAPs with the nudged elastic band (NEB) method that was originally formulated in the context of molecular systems ~\cite{Garrett1983,Mills1994,Mills1995,Jonsson1998}. In NEB, the minimum action path can be obtained iteratively by continuously shifting the pathway to the nearest minimum action path \cite{Henkelman2000,Henkelman2000a,ASE}. A similar approach is a growing string method \cite{Peters2004}. To provide more insights, we also employ the Euler-Lagrange (EL) method to compute the stationary action path. In addition to the LAP, another characteristic trajectory in the collective space is the minimum-energy path (MEP), sometimes referred to as the static path. The MEP can serve as a first, rough approximation to the LAP. It is obtained by computing the steepest descent line on the potential energy surface, which passes through the local minima and saddle points. To find the MEP, a flooding, or watershed, algorithm has been applied \cite{Mamdouh1998,Moller2001,Iwamoto2002,Moller2004,Wang2019a}. The NEB approach can also be adopted to find the MEP and saddle points \cite{Asgeirsson2021}. (For a review of modern optimization methods for finding MEPs, see \cite{Sheppard2008,More2004}.) This paper is organized as follows. In Sec.~\ref{sec:EDF} we define the basics concepts of the nuclear EDF approach as applied to nuclear fission. Section~\ref{Sec: Methods algorithms} describes the path-optimization methods used. The results of our calculations and an analysis of trends are presented in Sec.~\ref{sec:results}. Finally, Sec.~\ref{sec:summary} contains the conclusions of this work. \section{Nuclear EDF approach to spontaneous fission}\label{sec:EDF} The main ingredients for a theoretical determination of SF lifetimes are the collective potential energy surface (PES) and the inertia tensor. To compute the PES, one solves the constrained Hartree-Fock-Bogoliubov (HFB) equations with the realistic energy density functional in the space of collective coordinates $\vec{q} \equiv \{q_i\}$. These are usually represented by the expectation values of the quadrupole moment operator $\hat{Q}_{20}$ (elongation), quadrupole moment operator $\hat{Q}_{22}$ (triaxiality), octupole moment operator $\hat{Q}_{30}$ (mass-asymmetry), and the particle-number dispersion term $\lambda_{2\tau}(\hat{ N}^2_\tau - \langle \hat{ N}_\tau\rangle^2)$ ($\tau=n,p)$ that controls dynamic pairing correlations \cite{Vaquero2011,Vaquero2013,Sadhukhan2014}. In some cases one also considers the hexadecapole moment $Q_{40}$ (necking coordinate) \cite{Warda2012}. That is, in practical applications, we consider 2-5 collective coordinates which describe the collective motion of the system. Figure~\ref{fig:PES} shows a representative PES of $^{256}$Fm in the space of $Q_{20}\equiv \langle \hat{Q}_{20}\rangle $ and $Q_{30}\equiv \langle \hat{Q}_{30}\rangle$. \begin{figure*}[!htb] \includegraphics[width=0.8\linewidth]{Figures/256Fm_PES.png} \caption{Potential energy surface of $^{256}$Fm calculated with nuclear EDF method using the D1S parametrization of the Gogny interaction \cite{berger1984microscopic} in the space of two collective coordinates: $Q_{20}$ (elongation) and $Q_{30}$ (mass asymmetry). The static fission pathways are marked by solid lines: red (symmetric pathway) and green (asymmetric pathway). The outer turning line (OTL) is indicated, together with the outer turning points associated with the static pathways. For simplicity, we assume that the inner turning point corresponds to the ground-state configuration (i.e., $E_0=0$). The high-energy region that is practically not accessible during collective motion is indicated in black. The intersections of fission pathways with outer turning points are indicated by dots; these are important for determining fission fragment yields \cite{Sadhukhan2020,Sadhukhan2022}. } \label{fig:PES} \end{figure*} The collective inertia (or mass) tensor $\mathcal{M}_{ij}(\vec{q})$ is obtained from the self-consistent densities by employing the the adiabatic time-dependent HFB approximation (ATDHFB) \cite{Baran2011,Giuliani2018b,Washiyama2021}. In this study, we use the non-perturbative cranking approximation~\cite{Baran2011}: \begin{equation} \label{cranking-mass} \mathcal{M}_{ij}(\vec{q})=\frac{\hbar^2}{2\dot{q}_i\dot{q}_j}\sum_{\alpha\beta}\frac{\left(F^{i*}_{\alpha\beta}F^{j}_{\alpha\beta}+ F^{i}_{\alpha\beta}F^{j*}_{\alpha\beta}\right)}{E_{\alpha}+E_{\beta}}, \end{equation} where $q_i$ is the collective coordinate, $\dot{q}_i$ represents the time derivative of $q_i$, and $E_{\alpha}$ are one-quasiparticle energies of HFB eigenstates $|\alpha\rangle$. The matrices $F^i$ are given by \begin{equation} \label{equation-F} \frac{F^{i*}}{\dot{q}_i}= A^T\frac{\partial\kappa^*}{\partial q_i}A +A^T\frac{\partial\rho^*}{\partial q_i}B -B^T\frac{\partial\rho}{\partial q_i}A - B^T\frac{\partial\kappa}{\partial q_i}B, \end{equation} where $A$ and $B$ are the matrices of the Bogoliubov transformation, and $\rho$ and $\kappa$ are particle and pairing density matrices, respectively, determined in terms of $A$ and $B$. Derivatives of the density matrices with respect to collective coordinates are calculated by employing the three-point Lagrange formula. It is important to remark that rapid variations in $\mathcal{M}_{ij}$ are expected in the regions of configuration changes (level crossings) due to strong variations of density derivatives in (\ref{equation-F}) associated with structural rearrangements \cite{Ledergerber1973,Sadhukhan2013}. Since SF is a quantum-mechanical tunneling process and the fission barriers are usually both high and wide, the SF lifetime is obtained semi-classically \cite{Brack1972} as $T_{1/2}=\ln2/(nP)$, where $n$ is the number of assaults on the fission barrier per unit time and $P$ is the penetration probability given by \begin{equation}\label{penertration} P=\left(1+\exp{[2S(L_{\rm min})]}\right)^{-1}, \end{equation} where $L_{\rm min}$ is the path that minimizes the fission action integral calculated along the one-dimensional trajectory $L(s)$ in the multidimensional collective space: \begin{equation} \label{action-integral} S(L)=\frac{1}{\hbar} \int_{s_{\rm in}}^{s_{\rm out}} {\cal S}(s)\,ds, \end{equation} where \begin{equation} \label{action-integrand} {\cal S}(s) =\sqrt{2\mathcal{M}_{\text{eff}}(s) \left(V_{\text{eff}}(s)-E_0\right)} \end{equation} with $V_{\text{eff}}(s)$ and $\mathcal{M}_{\text{eff}}(s)$ being the effective potential energy and inertia along the fission path $L(s)$, respectively. $V_{\text{eff}}$ can be obtained by subtracting the vibrational zero-point energy from the total HFB energy. (In the examples considered in this paper we assume the zero-point energy to be zero.) The integration limits $s_{\rm in}$ and $s_{\rm out}$ correspond to the classical inner and outer turning points, respectively, defined by $V_{\text{eff}}(s)=E_0$ on the two extremes of the fission path, see Fig.~\ref{fig:PES}. The collective ground state (g.s.) energy is $E_0$, and $ds$ is the element of length along $L(s)$. A one-dimensional path $L(s)$ can be defined in the multidimensional collective space by specifying the collective variables $\vec{q}(s)$ as functions of path's length $s$. The expression for $\mathcal{M}_{\text{eff}}$ is~\cite{Baran2005}: \begin{equation} \label{eff-mass} \mathcal{M}_{\text{eff}}(s)=\sum_{ij}\mathcal{M}_{ij}(\vec{q})\frac{dq_{i}}{ds}\frac{dq_{j}}{ds}. \end{equation} The least-action path (LAP) $L_{\rm min}$ is obtained by minimizing the action integral (\ref{action-integral}) with respect to all possible trajectories $L$ that connect the lines/surfaces of inner turning points $s_{\rm in}$ and outer turning points $s_{\rm out}$ \cite{Sadhukhan2013}. However, as discussed in Refs.~\cite{Schmid1986,Eckern1992} and this paper, only the pathways related to the exit points are stationary. The MEP can instead be described as the union of steepest descent paths from the saddle point(s) to the minima. The corresponding trajectory $\vec{q}(s)$ satisfies \begin{align} \frac{d \vec{q}}{ds} \propto \vec{\nabla} V\big(\vec{q}(s)\big) \label{MEP-def} \end{align} which characterizes a path of steepest descent on a surface $V(\vec{q})$ \cite{quapp1984analysis}. For the NEB, one finds the MEP by allowing the elements of the path to follow the gradient of the PES in their immediate vicinity. We shall assume that the PES in the tunneling region is free from discontinuities associated with rapid configuration changes \cite{Dubray2012,Zdeb2021,Lau2021}. This assumption is usually valid because of non-vanishing pairing correlations inside the potential barrier. It is also to be noted that, as in any optimization/minimization approach, the stationary path determined numerically corresponds to a local action minimum, which is not guaranteed to be the global minimum. Moreover, there could be many stationary pathways representing different fission modes, see Fig.~\ref{Fig: Bifurcation}. To simplify notation, we assume in the following discussions that the stationary action path found by our algorithms is indeed the LAP. Since ${\cal S}(s)=0$ on the outer turning surface $V(\vec q) = E_{0}$, it follows that paths moving on the surface $V(\vec{q}) = 0$ do not contribute to the action. This is illustrated in Fig. \ref{Fig: Bifurcation} by the path connecting the g.s. and, for example, the purple star labeled (3). Such a path consists of the cyan curve -- the exit trajectory -- and the green dashed line, connecting (1) with (3) through the OTL, which results in the same action integral as the exit trajectory. \begin{figure}[htb!] \label{Fig: Bifurcation} \includegraphics[width=\linewidth]{Figures/bimodal1.pdf} \caption{Illustration of two stationary action paths (representing competing fission modes) from the g.s. to the OTL (marked white) on the PES given by Eq.~(\ref{epsbimodal}). The cyan line shows the primary path (1). The secondary path (2) is indicated by the black line. The corresponding exit points are marked by stars. The green dashed line connects the exit point (1) with the point (3) on the OTL; the action along the dashed path is zero. The inset shows the spring force and the action force acting on the image $i$ on the NEB for the intermediate (not fully converged) grey path. For the video illustrating the NEB determination of both LAPs, see the supplemental material (SM) \cite{SM}.} \end{figure} \section{Methods/algorithms} \label{Sec: Methods algorithms} All path-optimization methods described in the following subsections, bar the EL method, have a reference implementation included in the python package, PyNEB \cite{PyNEB}. \subsection{Nudged Elastic Band}\label{Sec: NEB} The NEB method was originally formulated to provide a smooth transition of a molecular system on a potential energy surface from the reactant to the product state~\cite{Mills1994,Mills1995,Jonsson1998}. Upon application of this variant of the NEB method, one obtains the MEP as well as a series of ``images'' of the molecular system as it transitions along the path. The NEB technique has been subsequently refined, with improved numerical stability~\cite{Henkelman2000a} and a more accurate determination of a saddle point~\cite{Henkelman2000} being two key advances towards a more widely applicable numerical approach for MEP determination. To obtain the LAP, the procedure must be modified such that the images move towards the minimum of the action~\cite{Asgeirsson2018} which amounts to replacing the standard gradient of the PES with the gradient of the action \begin{equation} \vec{g}_i= -\vec{\nabla}_i S \end{equation} with respect to the image $\vec q_i$. With this prescription, the images will settle to the LAP in the collective space. While the NEB method will, by design, drive the line of images towards either the MEP or LAP, the iterative scheme chosen greatly impacts the total number of iterations required before the solution converges. In the early implementations, a simple velocity Verlet algorithm \cite{Verlet1967} was used to adjust the position of the images step to step~\cite{Mills1994,Mills1995,Jonsson1998,Henkelman2000a,Henkelman2000}. This approach is robust and relatively stable, though the convergence can be slow for flatter surfaces where the images are not pulled strongly to their optimal positions. To aid this process, the Fast Inertial Relaxation Engine (FIRE) was proposed~\cite{Bitzek2006} to accelerate convergence without sacrificing stability. The method was subsequently updated~\cite{Guenole2020} to further improve performance. Indeed, in our tests, the inertial algorithm regularly outperforms the velocity Verlet algorithm by an-order-of-magnitude reduction in iterations at the same convergence criteria. With this, our implementation of the NEB approach is defined. The algorithm itself is outlined in Algorithm~1 in SM \cite{SM}. The force used in the optimization step for each image, $\vec{F}_i^{\rm opt}$, is constructed by adding the perpendicular component of the action gradient to the spring force $\vec{F}_i^{k}$ between the images, \begin{equation} \vec{F}_i^{k} = k (|\vec{q}_{i+1} - \vec{q}_i| - |\vec{q}_i - \vec{q}_{i-1}|)\vec{\tau_i}, \end{equation} where $k$ is a tunable parameter that controls the strength of the spring force and $\vec{\tau_i}$ is the unit vector tangent to the line of images from image $i-1$ to image $i+1$. The spring force on the endpoints is defined differently: \begin{equation} \vec{F}_1^k=k|\vec{q}_2-\vec{q}_1|,\quad \vec{F}_N^k=k|\vec{q}_{N}-\vec{q}_{N-1}|. \end{equation} The total force acting on the interior images is then \begin{align}\label{Eq:ImagesForces} \vec{F}_{i}^{\rm opt} = \vec{F}^{k}_{i} + \vec{g}_{i}^{\perp}. \end{align} The NEB approach is illustrated in Fig.~\ref{Fig: Bifurcation} for the case of bimodal tunneling from the g.s. minimum to the OTL on an analytic PES defined by: \begin{align}\label{epsbimodal} V(\vec{q}) &= 3.17 + 2e^{-5\big((x-1)^{2} + (y-\frac{1}{2})^{2}\big)} -3 e^{-(x^{2} +y^{2})} \nonumber \\&-\frac{1}{2}(3x +y), \end{align} where $\vec{q} = (x,y)$. The inset shows the forces on the images of the NEB grey path, which has not converged yet to the black path. The spring force $\boldsymbol{F}^k_i$ keeps the images from drifting too much from each other, while the perpendicular part of the action gradient $\boldsymbol{g}_i^\perp$ pushes them towards the nearest stationary action path. This example shows that the NEB algorithm, depending on the initial locations of the images, will converge to a local stationary path, not necessarily the least action path. For the endpoint, $i=N$, one can choose to either fix the position of the image or to allow the image to move towards the outer turning surface. In the second case, a harmonic restraint term is added to the spring force to construct $\vec{F}_{N}^{\rm opt}$, \begin{equation} \label{eqn:endpoint} \Vec{F}_{N}^{\rm opt} = \Vec{F}_{N}^{k} - \left[\Vec{F}_{N}^{k} \cdot \vec{f}(\vec{q}_{N}) - \eta (V(\vec{q}) - E)\right]\vec{f}(\vec{q}_{N}), \end{equation} where $\vec{f} = -\vec\nabla V / |\vec\nabla V|$ and $\eta$ determines the strength of the harmonic restraint term~\cite{Asgeirsson2018}. This force pulls the endpoint $i=N$ very quickly to the outer turning surface and helps find the optimal outer turning point. The default iteration scheme used in our implementation is the inertial algorithm mentioned above, though a standard Verlet minimizer is also included in the PyNEB python package \cite{PyNEB}. The structure of the NEB solver is modular and allows for the simple replacement of components like the minimizer, allowing for easy checks on the convergence and parameters that describe the iterative scheme. \subsection{Grid-Based Methods}\label{Section: grid-methods} Some traditional methods to compute the LAP begin by computing the PES and the collective inertia on a grid of collective coordinates. The calculation of the LAP is then reduced to finding the path through the grid points that minimizes a discrete approximation of the action. Two methods that we have benchmarked are the dynamic programming method (DPM) \cite{Baran1981}, and Dijkstra's algorithm (DA) \cite{Dijkstra}. Here, both will be described for two-dimensional (2D) grid, with points labelled by $\vec{q}_{ij}=(x_i,y_j)$ ($i=1,\ldots,N$, $j=1,\ldots,M$). Both methods can be straightforwardly extended to a higher-dimensional grid. Dynamic programming is a general mathematical technique for solving multi-decision problems by breaking the problem down into simpler overlapping sub-problems. It was first adapted to the action integral minimization in Ref.~\cite{Baran1981} and used in \cite{Sadhukhan2013} to determine the LAP. This adaptation is what we refer to as the DPM. The DPM approximates the LAP between an initial point, $\vec{q}_\textrm{in}$, and a final point, $\vec{q}_\textrm{fin}$. This method finds paths that traverse diagonally from a given cell: from cell $\vec{q}_{ij}$, only cells $\vec{q}_{i+1,j}$ can be reached, for $j=1,\ldots,M$. The allowed cells are highlighted in red in Fig.~\ref{grid_method}. The LAP from $\vec{q}_\textrm{in}$ to $\vec{q}_\textrm{fin}$ is constructed iteratively as follows: for a cell $\vec{q}_{ij}$, there are $M$ possible paths, each passing through a cell at $x_{i-1}$. The LAP from $\vec{q}_\textrm{in}$ to $\vec{q}_{ij}$ is selected and stored in memory. This is repeated for every cell with $x=x_i$, for a total of $M$ possible paths. Once $\vec{q}_\textrm{fin}$ is reached, there are only $MN$ paths (out of a total of $M^N$ paths), and the LAP is selected from these. The DPM algorithm is detailed in Algorithm 2 in SM \cite{SM}. \begin{figure}[htb!] \label{grid_method} \includegraphics[width=0.48\textwidth]{Figures/Dijkstra_vs_dpm.pdf} \caption{Different types of paths that can be found in the different grid-based methods. The single node $\vec{q}_{ij}$ can reach the red (blue) regions in DPM (Dijkstra's algorithm). The initial and final points are marked.}\label{Fig: grid_methods} \end{figure} Dijkstra's method \cite{Dijkstra} is similar to DPM, in that it breaks down the large optimization problem into a set of smaller problems. Given a cell $\vec{q}_{ij}$, the action to every neighbor $\vec{q}_{i'j'}$ is calculated as if the path to $\vec{q}_{i'j'}$ passes through $\vec{q}_{ij}$. If this action integral is smaller than that along the current path to $\vec{q}_{i'j'}$, $\vec{q}_{i'j'}$ is said to come from $\vec{q}_{ij}$. This is repeated, starting from $\vec{q}_\textrm{in}$, until $\vec{q}_\textrm{fin}$ is reached. Figure \ref{grid_method} shows the nearest-neighbors of $\vec{q}_{ij}$ (the cell marked in green) in a blue square. Dijkstra's algorithm is described in Algorithm~3 in the SM \cite{SM}. Dijkstra's algorithm can find paths that pass through multiple cells with the same $x_i$ value, or even paths that backtrack. DPM cannot find such paths. However, DPM can find paths that jump from $\vec{q}_{ij}$ to $\vec{q}_{i+1,j'}$, for any $j'$, while Dijkstra's algorithm is limited to $j'=j-1,j,j+1$ (see Fig.~\ref{grid_method}). For fission calculations, one frequently takes the $x$ coordinate as the quadrupole moment $Q_{20}$, and fission can be viewed as collective motion in which $Q_{20}$ continuously increases towards scission. So, the paths that Dijkstra's algorithm can find, that DPM cannot, are rather unlikely. In general DPM tends to find paths with a smaller action than Dijkstra's algorithm, see Sec.~\ref{sec:results}. \subsection{Euler Lagrange Equations} In order to find the LAP for the functional~\eqref{action-integral} using the EL equations \cite{weinstock1974calculus}, we first parametrize the trajectory $\vec q$ by a time variable $t$, i.e., $\vec{q}=\vec q(t)$ with $t \in [0,t_f]$. This is done in order to explicitly account for the arclength $ds = (\sum dq_i^2)^{1/2}$. In terms of $t$, the action integral~\eqref{action-integral} reads: \begin{equation}\label{Eq: ELE Action} \begin{split} &S(L) = \\ &\int_{0}^{t_f} \sqrt{2\big(V_\text{eff}[\vec q(t)]-E_0 \big)}\Big(\sum_{ij}^n\mathcal{M}_{ij}[\vec q(t)]\dot{q}_i\dot{q}_j\Big)^{1/2}dt \\ &= \int_{0}^{t_f} \mathcal{L}(\vec q,\dot{\vec q})dt, \end{split} \end{equation} where $\dot{q}_i\equiv dq_i/dt$, and $\mathcal{L}$ is the corresponding Lagrangian. The associated EL equation can be written as: \begin{equation}\label{Eq: ELE} \frac{\partial \mathcal{L}}{\partial q_i} = \frac{d}{dt} \Big( \frac{\partial \mathcal{L}}{\partial \dot{q}_i}\Big), \end{equation} with the boundary conditions: $\vec q(t=0)=\vec q_\text{in}$ (the initial location) and $\vec q(t=t_f)=\vec q_\text{\textrm{fin}}$ (the final location). In order to numerically solve Eq.~\eqref{Eq: ELE} we use the shooting method \cite{Shooting}. That is, we start at the initial position $\vec q_\text{in}$ and vary the direction and orientation of the initial ``velocity'' $\dot{\vec q}(t=0)$. We use a numerical differential equation solver to propagate the solution until we find an initial condition that satisfies $\vec q(t_f)=\vec q_{\textrm{fin}}$. Finding such initial conditions can present some challenges, which we discuss in the SM \cite{SM}. The EL approach is equivalent to what is done in Ref.~\cite{Schmid1986} where the eikonal equation is solved by the method of characteristics. Each different trajectory obtained by varying $\dot{\vec q}(t=0)$ corresponds to one of the characteristics of the leading order (cf. Eqs.(2.8) and (4.3) of \cite{Schmid1986}). It is worth noting that if the imaginary part of the phase of the wave function $W(\boldsymbol{q})$ is negligible, as is the case of the motion in the deep subbarrier region, then the eikonal equation for $W$ is a valid approximation \cite{Kapur1937}. The trajectories corresponding to the stationary functional~\eqref{action-integral} are equivalent to the solutions of the eikonal equation for $W$ (see Eqs.~(11) and (13) of \cite{Kapur1937}). A connection between the eikonal equation, the dynamic programming approach, and a variational principle in the context of geometrical optics is discussed in Ref.~\cite{lakshminarayanan1997dynamic}. \section{Results}\label{sec:results} \subsection{Analytic surfaces: Illustrative examples}\label{Sec: Analytic Surfaces} We benchmark the performance of the NEB method by comparing the LAP found using NEB (denoted as NEB-LAP) to the paths found using the DPM, DA, and EL approaches for analytic surfaces defined in terms of the position vector $\vec{q}=(x,y)$. Throughout this section, we assume a constant inertia ${\cal M}_{ij}=\delta_{ij}$. Within the NEB framework, the action functional \eqref{action-integral} can develop some noise as the NEB algorithm approaches the final action. This noise is a function of the NEB hyperparameters and the optimization method used. All surfaces discussed in this section are released as example cases with PyNEB \cite{PyNEB}. In the analytic cases, the NEB is initialized by fixing an initial and final points $\vec{q}_{\textrm{in}}$ and $\vec{q}_{\textrm{fin}}$, respectively, and defining a linear trajectory connecting them. The NEB algorithm is then iterated until convergence is reached. Grid-based methods use a grid spacing of $\Delta x = 0.1$ along the x-axis and $\Delta y = 0.005$ along the y-axis for all analytic surfaces. Details of the numerical methods used for solving the EL equations for all surfaces are discussed in SM \cite{SM}. The action values for each surface considered are included in Table \ref{analytic-table}. Action integrals in Table~\ref{analytic-table} are evaluated using linearly interpolated trajectories over 500 uniformly-distributed points. We compute both LAP and MEP in the NEB framework. Since the MEP is a solution of Eq.~(\ref{MEP-def}), images along the path converge to critical points on the surface depending on the position of the boundary images at $\vec{q}_{\textrm{in}}$ and $\vec{q}_{\textrm{fin}}$. Critical points on the surface $V(\vec{q})$ contained in the MEP can be extracted by calculating $\vec{\nabla}V$ along the path and are classified by computing the eigenvalues of the Hessian at those points. \begin{table}[htb] \caption{Action integrals for the 6-Camel-Back (CB-S and CB-A) and M{\"u}ller-Brown (MB) surfaces. The integrals have been calculated using a linear spline interpolation evaluated at 500 points along each trajectory.} \label{analytic-table} \begin{ruledtabular} \begin{tabular}{cccccc} & \textrm{NEB-MEP}& \textrm{NEB-LAP}& \textrm{DPM}& \textrm{EL}& \textrm{DA}\\ \colrule CB-S & 5.522 & 5.518 & 5.524 & 5.536 & 5.563\\ CB-A & 6.793 & 6.404 & 6.405 & 6.407 & 6.886\\ MB & 28.491 & 22.875 & 22.909 & 22.871 & 23.427\\ \end{tabular} \end{ruledtabular} \end{table} First, we consider the symmetric 6-Camel Back potential (CB-S) \cite{More2004} defined as \begin{equation} V_{\rm CB-S}(\vec{q}) = \big(4 - 2.1 x^{2} + \frac{1}{3} x^{4} \big) x^{2} + x y + 4(y^{2} - 1) y^{2} \label{eq:cbs} \end{equation} In this example, we seek the LAP connecting the local minimum located at $\vec{q}_{\textrm{in}} = (1.70, -0.79)$ to the local minimum located at $\vec{q}_{\textrm{fin}} = (-1.70, 0.79)$. Figure \ref{fig-camel-symmetric} shows the CB-S PES normalized to zero at its global minimum together with the calculated NEB-MEP, NEB-LAP, DPM, EL, and DA trajectories. The action integrals along these trajectories are listed in Table \ref{analytic-table}. \begin{figure}[t!] \includegraphics[width=0.5\textwidth]{Figures/camel_symm1.pdf} \caption{The symmetric camel-back PES $V_{CB-S}(\vec{q})$ normalized to its global minimum together with the calculated NEB-MEP (red), NEB-LAP (magenta), DPM (black), EL (cyan), and DA (lime) trajectories. Black stars indicate saddle points and yellow crosses mark local minima.} \label{fig-camel-symmetric} \end{figure} \begin{figure}[t!] \includegraphics[width=0.5\textwidth]{Figures/camel_asymm1.pdf} \caption{Similar as in Fig.~\ref{fig-camel-symmetric} but for the asymmetric camel-back surface $V_{\textrm{CB-A}}(\vec{q})$. For the video illustrating the NEB determination of both LAP and MEP, see the SM \cite{SM}.} \label{fig-camel-asymmetric} \end{figure} The MEP and the LAPs computed by using the NEB, EL, DPM, and DA methods are very similar. However, the DA trajectory slightly deviates from the other ones. This is because DA is more constrained by the grid spacing than DPM: regardless of the grid spacing, DA can only consider its immediate neighbors, while DPM does not have this constraint (see Fig.~\ref{Fig: grid_methods} and \ref{Section: grid-methods}). As indicated by Fig. \ref{fig-camel-symmetric}, the final action values for the LAP obtained by the NEB, DPM, and EL methods agree well with the MEP. However, the MEP and LAP are not necessarily equivalent in general; the MEP can be viewed as an approximation of the LAP. A detailed discussion on the conditions for the MEP to be an LAP is contained in the SM \cite{SM}. To see the MEP limitations, we consider an asymmetric variant of the Camel-Back potential (CB-A) \begin{align} V_{\rm CB-A}(\vec{q}) = V_{\rm CB-S}(\vec{q}) + \frac{1}{2}y \end{align} where the end points of the local minima are $\vec{q}_{\textrm{in}} = (1.70, -0.8)$ and $\vec{q}_{\textrm{fin}} = (-1.70, 0.76)$. Figure~\ref{fig-camel-asymmetric} shows the MEP trajectory which is markedly different from the LAP solutions and corresponds to an appreciably larger action integral. Still, the MEP can be used for finding critical points (minima and saddles) on the surface. The M{\"u}ller-Brown potential is a canonical example of a PES used in theoretical chemistry \cite{muller1979location,Koistinen2017,Asgeirsson2018}. The M{\"u}ller-Brown surface shown in Fig.~\ref{fig-MB} is defined as \begin{equation} V_{\rm MB}(\vec{q}) = \sum_{i=1}^{4} A_{i}e^{a_{i}(x-x_{0_{i}})^{2} + b_{i}(x-x_{0_{i}})(y-y_{0_{i}}) + c_{i}(y-y_{0_{i}})^{2}}, \end{equation} where we use the same set of parameters as in Ref.~\cite{muller1979location}, namely: $\vec{A} = (-200, -100, -170, 15)$, $\vec{a}=(-1,-1,-6.5,0.7)$, $\vec{b}=(0,0,11,0.6)$, $\vec{c}=(-10,-10,-6.5,0.7)$, $\vec{x}_{0}=(1,0,-0.5,-1)$, and $\vec{y}_{0}=(0,0.5,1.5,1)$. \begin{figure}[t!] \includegraphics[width=\linewidth]{Figures/MB_plot.pdf} \caption{Similar as in Fig.~\ref{fig-camel-symmetric} but for the shifted M{\"u}ller-Brown surface. The inset shows the LAP pathways close to the initial point $\vec{q}_\textrm{in}$. The yellow dashed line shows the vertical. As can be seen, all paths except for the DPM curve start by moving to the left of the vertical.} \label{fig-MB} \end{figure} The MEP follows the bent trajectory that goes through the critical points: two saddle points and one local minimum. This trajectory markedly differs from the LAPs, which are in a rough agreement. The MB surface highlights a problem with the DPM. As mentioned in Sec.~\ref{Sec: Methods algorithms}, the DPM can only search a single direction of each coordinate axis of the domain. In the case of the Muller-Brown surface, the DPM cannot search for trajectories bending back in the negative-$x$ direction. As seen in the inset of Fig.~\ref{fig-camel-symmetric}, the NEB, EL, and DA methods start their trajectories moving backwards in $x$ from the initial point $\vec{q}_\textrm{in}$. The DPM path, on the other hand, always moves in the positive-$x$ direction. Consequently, the action integral along the DPM path is slightly larger than in the other methods. \subsection{Realistic calculations}\label{Sec: Realistic} To illustrate the performance of the NEB method and other approaches to the LAP in realistic cases, we carried out nuclear EDF calculations for $^{232}$U in two collective coordinates and $^{240}$Pu in three collective coordinates. In the particle-hole channel we used the Skyrme functional SkM$^{*}$ \cite{BARTEL198279}, which is often employed in fission studies. The particle-particle interaction was approximated by the mixed density-dependent pairing force~\cite{DobaczewskiPairing}. In the case of $^{232}$U, we considered two collective coordinates $\vec{q}\equiv(Q_{20},Q_{30})$ and for $^{240}$Pu we took three collective coordinates $\vec{q}\equiv(Q_{20},Q_{30},\lambda_2)$. The axial quadrupole and octupole moment operators are defined as in Ref.~\cite{DobaczewskiHFODD}: \begin{align} \hat{Q}_{\lambda 0}(r,\theta) = \mathcal{N}_{\lambda} \sqrt{\frac{2\lambda+1}{4\pi}} r^{\lambda} P_{\lambda} (\cos\theta) \end{align} where $P_{\lambda}$ is the Legendre polynomial, $\mathcal{N}_{2}=\sqrt{\frac{16\pi}{5}}$, and $\mathcal{N}_{3}=1$. The collective coordinate $\lambda_2=\lambda_{2n}+ \lambda_{2p}$ defined in Sec.~\ref{sec:EDF} represents the dynamic pairing fluctuations. The value of $\lambda_{2\tau}$= 0 corresponds to static HFB pairing. As in Ref.\,~\cite{Sadhukhan2014}, to render collective coordinates dimensionless, we use dimensionless coordinates ${x_i}$ defined as \begin{eqnarray}\label{eqn:scale_param} x_{i}&=&\frac{q_{i}}{\delta{q_{i}}}, \end{eqnarray} where $\delta{q_{i}}$ are the scale parameters used in determining numerical derivatives of density matrices in Eq.~(\ref{equation-F}). Here we took $\delta{Q_{20}} = 1$\,b, $\delta{Q_{30}} = 1$\,b$^{3/2}$ and $\delta{\lambda_2}= 0.01$\,MeV. \subsubsection{Two dimensional case: SF of $^{232}$U} The PES was computed by solving the HFB equations using the parallel axial solver HFBTHO(v3.00)\cite{PEREZ2017363}. The large stretched harmonic oscillator basis of $N= 25$ major shells was used to guarantee good convergence. We adopted a $458\times501$ grid with $0\le Q_{20} \le{457}$\,b and $0\le Q_{30}\le{50}$\,b$^{3/2}$. To apply the NEB method, which involves local gradient calculations at arbitrary values $\vec q$, we interpolate the PES and the inertia tensor on the mesh. Because the grid is two dimensional, a cubic spline interpolator suffices. Close to the $Q_{30}=0$ axis, we take into account the mirror symmetry of the PES by setting $V(-Q_{30})=V(Q_{30})$. Finally, since NEB updates occasionally push an image outside of the computed PES mesh, we extended the PES to grow exponentially with the distance outside the mesh, to smoothly push images back into the evaluated region. \begin{figure}[htb!] \includegraphics[width=1.0\columnwidth]{Figures/232U.pdf} \caption{The PES of $^{232}$U in the $(Q_{20},Q_{30})$ plane calculated with SkM$^*$. Solid lines mark the LAPs and MEP obtained with the constant inertia tensor; dotted lines correspond to the non-perturbative inertia tensor. The OTL is shown in white. The blue, orange, purple and black curves represent the LAPs calculated using the NEB, DPM, EL, and DA methods, respectively. The green curve is the MEP, which was also calculated using NEB.}\label{Fig: U232 Paths} \end{figure} Figure~\ref{Fig: U232 Paths} shows the two-dimensional PES of $^{232}$U. The least action fission pathway which goes from the g.s. at $\vec{q}_\textrm{in}$=(24\,b, 0) to the exit point $\vec{q}_\textrm{fin}$=(281\,b, 37\,b$^{3/2}$) is calculated using the methods explained in Sec.~\ref{Sec: Methods algorithms}. To select the endpoint $\vec{q}_\textrm{fin}$, we compute the LAP using DPM for all points on the OTL, and select the point with the lowest action integral. This point is then used as the exit point for the other methods. While NEB does not require a fixed endpoint in general, we fix the endpoint here in order to facilitate inter-method comparison. The MEP path is calculated using the NEB method. The action integral computed with different methods is shown in Table \ref{table:nuclei_acts}. When computing the action, we interpolate the paths using a linear spline interpolator, and the action integral is computed using 500 evaluations along the path. This reduces the differences in the action that may arise from using a different number of points along the path (for instance, NEB gives a similar path to DPM using as few as 30 images). For all paths, we compute the action using the inertia tensor evaluated along the path. As can be seen, the action values computed for $^{232}$U using different methods agree well, with DA being the worst performer. As seen in Fig.~\ref{Fig: U232 Paths} and Table\,\ref{table:nuclei_acts} the MEP is very close to the LAP. This is because the static fission pathway (i.e., MEP) is fairly straight and the fission valley is well delineated. Note that perfect agreement is not expected, and in fact was not observed for the analytic surfaces, either. This is due in part to the different approximations used in each method --- for DPM and DA, this is the grid spacing; for NEB, this is the number of images and approximate treatment of derivatives; and for EL, this is a variety of simplifications described in Sec. 3 in the SM \cite{SM}. Additional variation in the quality of the interpolator further hampers agreement beyond what is listed. \begin{table}[htb] \caption{Action integrals for $^{232}$U computed with different methods. The paths computed using the constant and non-perturbative inertia tensor are labelled as ``con.'' and ``n.-p.'', respectively.} \label{table:nuclei_acts} \begin{ruledtabular} \begin{tabular}{llccccc} & & \textrm{NEB-MEP}& \textrm{NEB-LAP}& \textrm{DPM}& \textrm{EL}&\textrm{DA}\\ \colrule \multirow{2}{*}{${}^{232}$U} & con. & 174.5 & 174.2 & 174.2 & 174.9 & 175.8 \\ & n.-p. & - & 173.6 & 173.3 & 175.0 & 178.5 \\\\[-5pt] \multirow{2}{*}{${}^{240}$Pu} & con. & 19.09 & 18.98 & 19.21 & 19.01 & 22.85 \\ & n.-p. & - & 16.54 & 16.47 & 18.18 & 30.50 \\ \end{tabular} \end{ruledtabular} \end{table} \subsubsection{Three dimensional case: SF of $^{240}$Pu} The SF of $^{240}$Pu in several collective coordinates was studied in Ref.~\cite{Sadhukhan2014} where the details pertaining to the computation, grid size, etc., can be found. Between the g.s. minimum and the fission isomer (FI), the fission pathway is affected by triaxial degrees of freedom. Between the FI and the outer turning surface (OTS), however, the predicted fission trajectory is axial. In this paper, we consider the fission of the FI of $^{240}$Pu so the OTS corresponds to the FI energy. For three-dimensional tunneling, the system of equations that must be solved to construct a global spline interpolator is too large for practical applications. Instead, we use piecewise linear interpolation. The PES at $\lambda_2=0$ for $^{240}$Pu shown in Fig.~\ref{Fig: Pu240 2d Paths} varies very smoothly in the barrier region where the potential energy is larger than the energy of the FI, and so this interpolation scheme is reasonable. \begin{figure}[htb!] \includegraphics[width=1.0\columnwidth]{Figures/240Pu_2d.pdf} \caption{The PES for $^{240}$Pu in the space of collective coordinates $Q_{20}$, $Q_{30}$ with $\lambda_2 = 0$. Only the region beyond the fission isomer is shown. The energy is normalized to the energy of the fission isomer. The OTL is shown in white. The MEP (green) practically coincides with the LAPs calculated with the constant inertia using the NEB (blue), DPM (orange), and EL (purple) methods. }\label{Fig: Pu240 2d Paths} \end{figure} Figure~\ref{Fig: Pu240 Paths} shows the LAPs for $^{240}$Pu computed with the NEB, DPM, and EL methods in three dimensions (3D). The pathways begin at the FI minimum at $\vec{q}_\textrm{in}=(Q_{20}^{0}=87\,\text{b}, Q_{30}^{0}=0\,\textrm{b}^{3/2}, \lambda_{2}=0.0)$ and the exit point was chosen for DP in the same way as the $^{232}$U results before. The NEB endpoint in this case was allowed to vary according to Eq.~\ref{eqn:endpoint}, better representing standard procedure for production runs. The exit points $\vec{q}_\textrm{fin}$ predicted by NEB ($185.1\,\text{b}, 18.4\,\text{b}^{3/2}, 3.3\,\text{MeV}$), DPM ($184.0\,\text{b}, 18.6\,\text{b}^{3/2}, 4.8\,\text{MeV}$) and EL ($179.8\,\text{b}, 17.7\,\text{b}^{3/2}, 0.0$) then differ. When the collective mass is held constant, all methods find very similar paths in the $\lambda_{2}=0.0$ plane, which are also shown in Fig.~\ref{Fig: Pu240 2d Paths}. The paths vary more when the non-perturbative inertia tensor is used, with the main difference between the NEB and DPM paths appearing in the region close to the FI minimum; beyond the saddle point, both paths are similar. \begin{figure}[htb!] \includegraphics[width=1.0\columnwidth]{Figures/3dplot.pdf} \caption{The PES for $^{240}$Pu in the collective coordinates $Q_{20}$, $Q_{30}$ and $\lambda_2$. The 2D cross section at $\lambda_2=0$ shown in Fig.~\ref{Fig: Pu240 2d Paths} is indicated. The blue, orange, and purple curves are the LAP, calculated using the NEB, DPM, and EL methods, respectively. The non-perturbative inertia tensor was used for the dashed curves. The OTS is indicated by the dark blue contour surface. } \label{Fig: Pu240 Paths} \end{figure} As seen in Table \ref{table:nuclei_acts}, the NEB and DPM are in a good agreement. In general, one would expect a better performance from NEB as this method is not constrained to a grid (this is true in the case of the analytic surfaces discussed in Sec. \ref{Sec: Analytic Surfaces}). However, in rare cases, the DPM produces a slightly lower action than the NEB. In such cases the NEB converges to an even lower action if is initialized with the DPM result. This suggests that for tunneling in more than 2D, a combination of NEB and DPM might be beneficial. \section{Conclusions}\label{sec:summary} Finding the path that minimizes the action integral can be extremely challenging since it involves searching over the space of all continuous paths that fulfill the boundary conditions. Each method explored in this paper simplifies such task in different ways. DPM and DA project the PES onto a finite grid and explore decisions in making the path between the boundary conditions. In the EL approach the surface is modified in several ways to smooth the relation between initial conditions and the end point of the trajectory. The NEB method reduces the original search over continuous path into considering only piece-wise linear paths, the number of pieces given by the number of images. It is this simplification that makes the NEB robust and accurate, since the total action now becomes a smooth function of the position of the images, a function that can straightforwardly be numerically minimized by gradient descent methods. A significant advantage of the NEB is that it can accommodate any initial positions of the images, which speeds the convergence appreciably if a good prior guess of the LAP is provided. Other methods lack for such incorporation of prior knowledge. Finally, the resolution of the NEB for a rapidly varying surface can be adjusted locally by increasing the amount of images or spring constants, while for DPM and DA the entire grid resolution would have to be increased, giving an appreciable toll on the computational cost. For both analytic and realistic potential energy surfaces the NEB robustly produces a LAP. In the cases studied, NEB outperforms the EL and DA methods, and produces close results to those of the DPM with usually lower action integral. For many-dimensional tunneling, initiating the NEB method from the DPM path might be a winning strategy. A huge advantage of the NEB over other methods is that it can efficiently and accurately estimate exit points. By exploring different initial conditions for the positions of the images which lead to distinct exit points, one can use the NEB method to study the phenomenon of multimodal fission. An example of such an application is shown in video~1 in the SM \cite{SM}. Whilst other methods can find a least-action trajectory for an arbitrary final point placed on the OTL, as done, e.g., in Refs.~\cite{Sadhukhan2013,Sadhukhan2014,Sadhukhan2016,Sadhukhan2017,Zhao2015,Zhao2016,Mercier2021}, they cannot guarantee that this trajectory is stationary. All such trajectories can be gradually transformed into a stationary pathway by moving the final point along the OTL towards the exit point, see Fig.~\ref{Fig: Bifurcation}. In this paper we also explored the minimum-energy (or static) path. We adjusted our NEB algorithm to generate MEPs, including the determination of local minima and saddle points. The necessary conditions for an MEP to also be an LAP are discussed in the SM \cite{SM}. Video~2 in the SM \cite{SM} illustrates the way the NEB method generates LAP and MEP. An important contribution of this work is providing a beta release of the PyNEB package, a python suite of codes that implement the NEB algorithm described in this paper. The package can be found in \cite{PyNEB} together with the respective documentation and code samples serving as a tutorial for its use. A comprehensive investigation into the intricacies of the numerical implementations and performance of the package itself will accompany the version $1.0$ release. The NEB approach can be readily paired with accelerated DFT calculations, such as the recent applications of Gaussian process regression to PES emulation~\cite{Koistinen2017,Torres2019}. In these works, a Gaussian process is used to emulate the PES and DFT calculations are only run if the Gaussian process is uncertain as to the actual PES value. As NEB is not a grid-based method, it can sensibly be paired with a Gaussian process emulator that is updated as necessary while NEB runs. In this way, the LAP can be determined using far fewer DFT evaluations than is necessary in DPM. The ability to determine the exit points is essential for determining fission fragment yields \cite{Sadhukhan2020,Sadhukhan2022}. The minimum action provides information on SF half-lives. In this context, the NEB method described in this paper is expected to speed up the global calculations of nuclear fission for r-process simulations and studies of superheavy nuclei stability. \noindent\makebox[\linewidth]{\resizebox{0.3333\linewidth}{1pt}{$\bullet$}}\bigskip \begin{acknowledgments} We are grateful to Edgard Bonilla and Stefan M. Wild for useful comments. This work was supported by the U.S. Department of Energy under Award Numbers DOE-DE-NA0002847 (NNSA, the Stewardship Science Academic Alliances program), DE-SC0013365 (Office of Science), and DE-SC0018083 (Office of Science, NUCLEI SciDAC-4 collaboration) and by the National Science Foundation CSSI program under award number 2004601 (BAND collaboration). \end{acknowledgments} E.F. and D.L. contributed equally to this work. \section{Algorithms} \label{algorithms} \begin{algorithm} \caption{Nudged Elastic Band}\label{alg:neb} \begin{algorithmic} \State Define start and end point \State Initialize images along line \For{Steps} \For{Images} \State Update effective force \State Compute velocity of image \State Translate image \EndFor \State Compute action \State Check convergence \EndFor \end{algorithmic} \end{algorithm} \begin{algorithm} \caption{Dynamic Programming Method}\label{alg:dpm} \begin{algorithmic} \State Define start and end points $\vec{q}_\textrm{in}$ and $\vec{q}_\textrm{fin}$ \For{$i=1,\ldots,n$} \For{$j=1,\ldots,m$} \State Select $\vec{q}_{i-1,j}$ to minimize the action from $\vec{q}_\textrm{in}$ to $\vec{q}_{ij}$, through $\vec{q}_{i-1,j}$ \State Copy the path to $\vec{q}_{i-1,j}$ and append $\vec{q}_{ij}$ to it \EndFor \EndFor \State Append $\vec{q}_\textrm{fin}$ to all paths \State Select the path with the least total action \end{algorithmic} \end{algorithm} \begin{algorithm} \caption{Dijkstra's Algorithm}\label{alg:djk} \begin{algorithmic} \State Define start point $\vec{q}_\textrm{in}$ \State Initialize distance matrix $d(\vec{q}_{ij})\to\infty$, $d(P_\textrm{in})\to0$ \State Mark all points as unvisited \While{Any point is unvisited} \State Select $\arg\min d(\vec{q}_{ij})$ with $\vec{q}_{ij}$ unvisited \For{$\vec{q}_{i'j'}$ unvisited, $\vec{q}_{i'j'}$ a neighbor of $\vec{q}_{ij}$} \State Compute tentative distance \begin{equation} d'=d(\vec{q}_{ij})+\textrm{dist}(\vec{q}_{ij},\vec{q}_{i'j'})\nonumber \end{equation} \If{$d'<d(\vec{q}_{i'j'})$} \State Set $d(\vec{q}_{i'j'})=d'$ \State Mark $\vec{q}_{i'j'}$ as coming from $\vec{q}_{ij}$ \EndIf \EndFor \State Mark $\vec{q}_{ij}$ as visited \EndWhile \end{algorithmic} \end{algorithm} \begin{algorithm} \caption{Euler-Lagrange Method}\label{alg:ELE} \begin{algorithmic} \State Define a tolerance $\epsilon$, initial $\vec{q}_{\rm in}$ and final $\vec{q}_{\rm fin}$ points \State Guess initial velocity $\dot{\vec{q}}_{\rm in}$ and solve Euler-Lagrange equations \While{Final point on path is not within tolerance $|\vec{q}(t_f)-\vec{q}_{\rm fin}|>\epsilon$} \State Vary $\dot{\vec{q}}_{\rm in}$ \EndWhile \end{algorithmic} \end{algorithm} \section{Supplemental videos} The supplemental videos (in .mp4 and .mov formats) show the evolution of the LAP as a function of the number of iteration steps in the NEB method. \begin{itemize} \item Supplementary video 1 (\href{https://pyneb.dev/assets/img/multi-modal.mov}{.mov}, \href{https://pyneb.dev/assets/img/multi-modal.mp4}{.mp4}) - This video shows the convergence behaviour of the LAP on the surface in Eq.~(12), also shown in Fig. 2, as a function of iteration step. Circles indicate the positions of the NEB images. Multiple initial pathways are shown with one boundary image fixed to the global minimum while the remaining images are subject to gradient forces. The harmonic force is applied to the outer boundary image; it pushes the end point to the outer turning line, see Eq.~(13). As the NEB algorithm iterates, the bands converge to two unique stationary paths with unique exit points. The computed action integral for each path is given in the legend. This example shows that the NEB is capable of identifying multiple stationary action paths. \item Supplementary video 2 (\href{https://pyneb.dev/assets/img/asymm-camelback.mov}{.mov}, \href{https://pyneb.dev/assets/img/asymm-camelback.mp4}{.mp4}) - This video shows the convergence behaviour of the minimum energy path (MEP), shown in red, and the least action path (LAP), shown in purple on the asymmetric camel-back surface of Fig.~5. Circles indicate the positions of the NEB images. A linear trajectory connecting two local minima with fixed endpoints is given as an initial guess. The MEP converges to a gradient curve that passes through critical points of the surface whilst the LAP bypasses two saddle points. \end{itemize} \section{MEP and LAP equivalence conditions} \label{LAP-MEP} As discussed in Sec.~I\, the MEP is obtained by computing the steepest descent line on the PES, and the LAP is determined by minimizing the collective action in the many-dimensional collective space. We want to address the necessary conditions for an MEP to be an LAP. Similar to \cite{quapp1984analysis} and Eq.~(7), we define the MEP as a gradient curve $\vec{\gamma}(\tau)$ satisfying the differential equation \begin{align} \dot{\vec{\gamma}} = \sigma(\tau) \vec{\nabla}V_\text{eff}(\vec{\gamma}(\tau)), \label{B1} \end{align} with boundary conditions $\boldsymbol\gamma(0) = \vec{q}_{\textrm{in}}$ and $\vec{\gamma}(1) = \vec{q}_{\textrm{fin}}$. $\tau$ is an arbitrary monotonic parametrization of the curve, while $\dot{\vec{\gamma}}$ represents the local velocity of the trajectory. $\sigma(\tau)$ is a factor to account for the transformation of arc length $s$ in Eq.~(7)\ to the parameter $\tau$ to be used in the EL Eq.~(15). Solutions to Eq.~\eqref{B1} define a trajectory made by a collection of steepest descents and ascents between $\vec{q}_{\textrm{in}}$ and $\vec{q}_{\textrm{fin}}$ in terms of the parameter $\tau$. \par On the other hand, the LAP trajectory $\boldsymbol{q}(\tau)$ is derived by finding a stationary point of the action integral \begin{align} S = \int_{0}^{1} \sqrt{2\Big(V_\text{eff}(\vec{q}) - E \Big)M_{\mu \nu}(\vec{q}) \dot{\vec{q}}^{\mu} \dot{\vec{q}}^{\nu}} d\tau, \label{B2} \end{align} where $M_{\mu \nu}$ is the inertia tensor, and the boundary conditions are $\boldsymbol{q}(0) = \vec{q}_{\textrm{in}}$ and $\boldsymbol{q}(1) = \vec{q}_{\textrm{fin}}$. Here and in the following, Einstein summation convention is assumed. By Beltrami's theorem, one can show that finding stationary solutions to the functional \begin{align} \tilde{S} &= \int_{0}^{1} \Big(V_\text{eff}(\vec{q}) - E \Big)M_{\mu \nu}(\vec{q})\dot{\vec{q}}^{\mu} \dot{\vec{q}}^{\nu} d\tau \\ &= \int_{0}^{1} g_{\mu \nu} \dot{\vec{q}}^{\mu} \dot{\vec{q}}^{\nu} d\tau \label{B3} \end{align} yields solutions that also fulfill the Euler Lagrangian equations for the functional \eqref{B2}. The converse of this statement is not true. Eq.~\eqref{B3} is recognized as being in the same form as the Lagrangian for a free particle in a curved space with the metric tensor $g_{\mu \nu}$. By the stationary action principle, one can derive the geodesic equation: \begin{align} \ddot{q}^{\lambda} + \Gamma^{\lambda}_{\mu \nu} \dot{q}^{\mu} \dot{q}^{\nu} = 0, \end{align} where $\Gamma^{\lambda}_{\mu \nu}$ are the Christoffel symbols: \begin{align} \Gamma^{\lambda}_{\mu \nu} = \frac{1}{2} g^{\lambda \alpha} \Big(g_{\nu \alpha,\mu} + g_{ \mu \alpha, \nu} - g_{\mu \nu, \alpha}\Big). \end{align} The solutions represent a stationary point of both action functionals $\tilde{S}$ and $S$. For the case with constant inertia, the metric tensor has the form: \begin{align} g_{\mu \nu} = \begin{pmatrix} V_\text{eff}(\vec{q}) - E & 0 \\ 0 & V_\text{eff}(\vec{q}) - E \end{pmatrix}, \end{align} where we have assumed that $M_{\mu \nu}\propto \delta_{\mu\nu}$ and is $\vec{q}$-independent. The coordinates $\vec q$ are re-scaled to be dimensionless. The geodesic equation in vector notation becomes \begin{align} \label{Eq: Equations of Motion} (V_\text{eff} - E) \ddot{\vec{q}} = \frac{1}{2} |\dot{\vec{q}}|^{2} \vec{\nabla} V_\text{eff} - \big( \dot{\vec{q}} \cdot \vec{\nabla} V_\text{eff}\big) \dot{\vec{q}}, \end{align} with the boundary condition $\vec{q}(0) = \vec{q}_{\textrm{in}}$ and $\vec{q}(1) = \vec{q}_{\textrm{fin}}$. Suppose now that there exists a trajectory $\vec{q}(\tau)$ satisfying $ \vec{\nabla} V_\text{eff}\big[\vec{q}(\tau)\big] \propto \dot{\vec{q}}$, i.e, the tangent vectors of the solution $\vec{q}(\tau)$ are parallel to the surface gradient $\vec{\nabla}V_\text{eff}$ evaluated along the curve. Since $\vec\nabla V_\text{eff}$ and $\dot{\vec{ q}}$ are parallel, Eq.~\eqref{Eq: Equations of Motion} would imply that $\ddot{\vec{q}} \propto \dot{\vec{q}}$ along the trajectory. This further implies that the tangent vectors of the solution $\vec{q}(\tau)$ are always in the same direction as its acceleration. From this we conclude that in order for the MEP to also be a solution of Eq.~$\eqref{Eq: Equations of Motion}$, such trajectory $\vec{q}(\tau)$ can only be composed of straight lines. Critical points of the PES where $V_\text{eff}-E$ and $\vec{\nabla} V_\text{eff}$ both vanish are the only locations on the surface where the trajectory can bend non-smoothly and still obey Eq.~\eqref{Eq: Equations of Motion}. \par These results significantly restrict the types of PESs for which an MEP can be an LAP. Such conditions on the surface can be obtained by the use of Eq.~\eqref{B1} as an ansatz for the equation of motion \ref{Eq: Equations of Motion}. We first note that, if the trajectory solves Eq.~\eqref{B1}, then: \begin{align} &\ddot{\vec{\gamma}}(\tau) = \dot{\sigma} \vec{\nabla} V_\text{eff} + \sigma \frac{\partial \vec{\nabla} V_\text{eff}}{\partial \vec{\gamma}} \frac{\partial \vec{\gamma}}{\partial \tau}, \\ &\frac{\partial \vec{\nabla} V_\text{eff}}{\partial \vec{\gamma}} \frac{\partial \vec{\gamma}}{\partial \tau} = H[\vec{\gamma}(\tau)] \dot{\vec{\gamma}}, \end{align} where $H$ is the Hessian matrix of second derivatives of the PES evaluated along the gradient curve $\vec{\gamma}$. For such a trajectory, the acceleration along the curve is \begin{align} \ddot{\vec{\gamma}}(\tau) = \dot{\sigma} \vec{\nabla} V_\text{eff} + \sigma^{2} H \vec{\nabla} V_{\rm eff}. \label{curve-acceleration} \end{align} Substituting this into the the geodesic equation we obtain \begin{align} H \vec{\nabla} V_\text{eff} &= -\Big(\frac{|\vec{\nabla} V_\text{eff}|^{2}}{2(V_\text{eff} - E)} + \frac{\dot{\sigma}}{\sigma^{2}} \Big) \vec{\nabla} V_\text{eff}. \label{MEP_condition} \end{align} \begin{figure}[t!] \centering \includegraphics[width=0.5\textwidth]{Figures/AnglesSc.png} \caption{Absolute value of the minimum angle (in degrees) between $\vec{\nabla} V_\text{eff}$ and the eigenvectors of $H$ for the CB-S surface of Fig.~{4} as a function of the coordinates $(x,y)$. The red curve shows the NEB-MEP, the green curve shows the NEB-LAP and the yellow stars indicate critical points of the surface. The inset shows a zoom-in on one region where there is a misalignment in the MEP path; in this region the LAP curves slightly. } \label{award_winning_plot} \end{figure} We see that the gradient $\vec{\nabla}V_\text{eff}\big[\vec{\gamma}(\tau)\big]$ must be an eigenvector of the Hessian $H\big[\vec{\gamma}(\tau)\big]$ for all $\tau$. This shows that the eigenvectors of $H$, or principal directions of the surface $V_\text{eff}$, must also be parallel to the gradient curve tangents along the trajectory. Since we know that if $\vec{\gamma}(\tau)$ is to be a solution of \eqref{Eq: Equations of Motion}, then $\dot{\vec{\gamma}}(\tau) \propto \ddot{\vec{\gamma}}(\tau)$; again showing that the curve $\vec{\gamma}(\tau)$ must be a straight line. In regions where the gradient curve is not parallel to the Hessian eigenvectors, there will be an acceleration along the gradient curve perpendicular to its tangents. In this situation, the function $\vec{\gamma}(\tau)$ cannot be a solution to \eqref{Eq: Equations of Motion} since $\ddot{\vec{q}}(\tau) \not\propto \dot{\vec{q}}(\tau)$ for all $\tau$. Equation~\eqref{MEP_condition} allows us to identify whether an MEP is an exact stationary path by checking if the gradient descent trajectory on the surface $V_{\rm eff}$ misaligns at any point with the eigenvectors of the surface's Hessian $H$. Figure \ref{award_winning_plot} shows the smallest angle between the gradient and the Hessian's eigenvectors for the CB-S surface. In the regions where the angle is small (stationary path can be approximated by straight lines) the MEP are LAP are close. From this analysis, we conclude that in many situations the MEP will not coincide with the LAP. However, in cases where a stationary path can be approximated by straight lines, such as in the CB-S surface for example shown in Fig.~4, the MEP might well approximate the LAP. Since the LAP is a stationary path, small deviations from the LAP could translate into very small errors in the action value. \section{Details on the Euler-Lagrange method} \label{appendix c} Numerically solving Eq.~(15)\ (or Eq.~\eqref{Eq: Equations of Motion} for a constant mass case) for fixed initial conditions, $\vec q(0)=\vec q_\text{in}$ and $\dot{\vec q}(0)$, is straightforward. The main challenge lies in finding the correct initial speed and direction, $|\dot{\vec q}(0)|$ and $\hat{\dot{\vec q}}(0)$, that provides a trajectory that ends on the fixed final point $\vec q(t_f)=\vec q_\text{fin}$. In practice, it is only necessary to obtain an upper value of the initial speed $|\dot{\vec q}(0)|$ since the geometry of the trajectory would be the same with the ``particle'' arriving at an earlier time than $t_f$ to $\vec q_\text{fin}$. In the following, we discuss three important elements of the EL method implementation: \begin{enumerate}[label=(\roman*)] \item How to explore the initial velocity space; \item How to deal with regions where the Lagrangian in Eq.~(14)\ is close to $0$; \item How to deal with boundaries (non-holonomic constraints) of the form $q_i\geq 0$. \end{enumerate} With respect to (i), the main issue is that the map between $\dot{\vec q}(0)$ and $\vec q(t_f)$ is highly non-linear, and it is not trivial to create a cost function that can tell us how to improve an initial condition $\dot{\vec q}(0)$ such that $\vec q(t_f)$ gets closer to $\vec q_\text{fin}$. An automatic algorithm, such as Mathematica's shooting method \cite{Shooting}, is able to find the correct initial conditions for the analytic surfaces discussed in Sec.~IV.1, but fails with the realistic cases of Sec.~IV.2. For these realistic cases we resorted to a binary search for the initial angles of $\hat{\dot{\vec q}}(0)$. This procedure worked well for the $^{232}$U case, but yielded non-optimal results (compared to the NEB and DP paths) for the 3D $^{240}$Pu case. The nonlinearity of $\vec q(t_f)$ as a function of $\dot{\vec q}(0)$ can get stronger in the regions where the effective potential and the inertia tensor in Eq.~(5)\ undergo abrupt {local} variations. Such variations can arise due to numerical noise in the underlying DFT calculation, interpolation errors, or configuration changes due to level crossings. The local fluctuations can easily deflect trajectories being computed with the EL, even though their presence should not in principle impact much the actual path of minimum action. To remedy this situation for the EL approach, we ``smoothed'' (coarse-grained) $V_\text{eff}$ and $\mathcal{M}_{\text{eff}}$ by replacing their values at each location $\vec q$ by their respective averages over a small area around $\vec q$, see Fig.~\ref{Fig: averaged surface}. This was done at the level of the grid points, before building the interpolator. The hope is that, by removing fine-grained details of the surface, the final optimal path would not change much, but the correct $\dot{\vec q}(0)$ will be easier to find. This was not necessary for the analytic surfaces of Sec.~IV.1\. \begin{figure}[t!] \includegraphics[width=0.90\linewidth]{Figures/Averaged.png} \caption{$^{232}$U PES before (a) and after (b) the smoothing procedure. The values near the boundary at $Q_{30}=0$ b have also been altered to guarantee that the vertical component of $\vec{\nabla} V_\text{eff}$ is zero. A similar effect is achieved in the outer turning line by replacing $(V_\text{eff}-E_0)\to \sqrt{(V_\text{eff}-E_0)^2+\epsilon_2^2}$. } \label{Fig: averaged surface} \end{figure} As concerns (ii), the Lagrangian $\mathcal{L}$ in Eq.~(14)\ becoming close to zero is a problem since the numerical integrator for the differential equations becomes unstable; increasingly big steps in $dt$ leave the action almost unchanged. This can happen if either $\big(V_\text{eff}-E_0 \big)$ or $\mathcal{M}_\text{eff}$ approach zero. In the case of a constant inertia tensor, such as in the analytical surfaces in Sec.~IV.1, the equations of motion Eq.~\eqref{Eq: Equations of Motion} resemble a particle of ``mass'' $\big(V_\text{eff}(\vec q)-E_0 \big)$ moving under non-conservative external forces proportional to $\nabla V_\text{eff}$ (this equation produces the same trajectories as the equations of motion of a particle moving on a inverted potential as those shown in \cite{Schmid1986}). If $V_\text{eff}\sim E_{0}$, the trajectory of the particle becomes very sensitive to any force, which is the reason why trajectories diverge near the outer turning line (see Figs. 3 and 4 in \cite{Schmid1986}), or near global minima such a those shown in Fig.~4. The fact that trajectories are unstable around the OTL is because the assumption of the eikonal approximation that the phase of the wave function is real and changes slowly breaks down. We refer the reader to \cite{Schmid1986} for more details on higher order corrections to the eikonal approximation. To avoid the instabilities in regions where $V_\text{eff}\sim E_0$, including a neighborhood around the OTL, we shifted the effective potential by a small constant $\epsilon_{0}$, $V_\text{eff}-E_0 + \epsilon_0$, and added a small diagonal term to the inertia tensor: $\mathcal{M}_\text{eff} + I\epsilon_1$, where $I$ is the identity matrix with the appropriate dimensions. For the analytic surfaces in Sec.~IV.1, we solved the EL equations for several decreasing values of $\epsilon_0$ and then extrapolated the collection of paths to obtain $\vec q(t)$ when $ \epsilon_0 \to 0$. Fig.~\ref{Fig: fig-camel-symmetric paths} shows trajectories obtained in such way for the CB-S surface. \begin{figure}[htb!] \includegraphics[width=0.90\linewidth]{Figures/Paths.png} \caption{PES from Fig.~4. The dashed paths are calculated with values of the potential shift in the range: $\epsilon_0 \in [0.05,3]$, while the solid cyan line is obtained by extrapolating the paths with $\epsilon_0 \in [0.05,0.42]$. The first red dashed path, $\epsilon_0=3$, has an action value around $10\%$ bigger than the last orange path, $\epsilon_0=0.05$, but it took nearly 400 times longer for the shooting method \cite{Shooting} to converge for $\epsilon_0=0.05$. This highlights how the regions where $V_\text{eff}\sim E$ lead to less stable trajectories.} \label{Fig: fig-camel-symmetric paths} \end{figure} In the case of the realistic calculations we only worked with fixed values of $\epsilon_0$ and $\epsilon_1$. For the 2D cases we used $\epsilon_0=1$ MeV and $\epsilon_1=0.01$ ($\epsilon_1$ being dimensionless since we rescaled $Q_{20}$ and $Q_{30}$ to be in the $[0,1]$ range from their original ranges). For the 3D case we used $\epsilon_0=0.005$ MeV and $\epsilon_1=10^{-5}$. With regard to (iii), the numerical integrator can be stopped when the trajectory arrives at such boundary. The main issue is on how to deal with trajectories which for an appreciable interval can lay extremely close to a boundary without crossing it, such as those in Fig.~7\ for $Q_{20}\in \sim[25,70]$ b. For a constant effective mass, it can be seen from Eq.~\eqref{Eq: Equations of Motion} that unless $\vec\nabla V_{\rm eff}$ has zero vertical component, the particle will either be pulled down close to the boundary and touch it, or be repelled from it in the upward direction and be driven away. To tackle this issue, we altered the effective potential and effective mass in a small vicinity around the line $Q_{30}=0$ in such a way that the PES and effective mass values are not drastically changed, but we ensure that the vertical components of their gradients are zero along such line. This creates a ``canal effect'' with the particle traveling close to the boundary for a finite time. A similar ``canal effect'' was used to explore the behavior of trajectories near the OTL and improve stability of the solution in that region. This was done by replacing the effective potential by $(V_\text{eff}-E_0)\to \sqrt{(V_\text{eff}-E_0)^2+\epsilon_2^2}$, with $\epsilon_2$ being a small constant. The gradient of this new function is zero at the OTL. In the realistic examples, we chose the value $\epsilon_2=0.5$ MeV. We conclude this section by emphasising that, for the EL approach to work, many simplifications and approximations had to be made. On the other hand, the NEB approach is free from the difficulties (i)-(iii). By using local surface derivatives and global springs forces, the NEB is effectively less sensitive to the noise in the PES. The NEB also has the advantage that the end point does not have to be predefined. For the EL approach, an undefined ending point translates into a huge computational cost. Finally, constraints of the form $q_i\geq0$ are easily implemented on the NEB, without causing any instability of the algorithm or requiring the surface to be altered in any way.

4d6d7e42069f4b5a7c59fed933f954093c088a7b

\section{Introduction} The temporal resolution of the propagation of the smallest capillary waves discretely resolved in space on a fluid interface presents the dominant time-step restriction for the majority of simulations of interfacial flows with surface tension \citep{Popinet2018}. It has long been speculated that an implicit formulation of surface tension must be able to eliminate, or at least mitigate, this time-step constraint \citep{Brackbill1992, Kothe1998, Hysing2006,Popinet2009,Sussman2009,Popinet2018}. To date, however, no numerical algorithm based on an interface capturing method, such as a Volume-of-Fluid (VOF), level-set or phase-field method, has been reported in the literature that can breach the capillary time-step constraint, while still representing the governing physics faithfully and without adding unphysical terms to the governing equations. The capillary time-step constraint arises from the phase velocity $c$ of capillary waves, given as \citep{Lamb1932} \begin{equation} c = \sqrt{\frac{2 \pi \sigma}{(\rho_\text{a}+\rho_\text{b}) \lambda}}, \label{eq:c} \end{equation} where $\sigma$ is the surface tension coefficient, $\lambda$ is the wavelength, and $\rho_\text{a}$ and $\rho_\text{b}$ are the densities of the interacting fluids. Because the phase velocity is inversely proportional to $\sqrt{\lambda}$, shorter capillary waves with an increasing phase velocity are spatially resolved as the computational mesh is refined. \citet{Brackbill1992} were the first to recognise this time-step restriction. \revB{Considering that the shortest unambiguously resolved capillary waves have a wavenumber of $k_\sigma = \pi/\Delta x$ and, consequently, a wavelength of $\lambda_\sigma=2\pi/k_\sigma=2\Delta x$, where $\Delta x$ is the mesh spacing, \citet{Brackbill1992} proposed the capillary time-step constraint as} \begin{equation} \Delta t_\sigma = \frac{\Delta x}{2 c} = \sqrt{\frac{\rho_\text{a}+\rho_\text{b}}{4 \pi \sigma} \Delta x^3}. \label{eq:tsigma_Brackbill} \end{equation} The capillary time-step constraint is, therefore, a Courant–Friedrichs–Lewy (CFL) condition \citep{Courant1928} associated with the phase velocity of capillary waves, $\Delta t_\sigma \propto \Delta x/c$, with the factor 2 in Eq.~\eqref{eq:tsigma_Brackbill} accounting for the case in which two oppositely propagating waves enter the same mesh cell simultaneously. As part of our previous work \citep{Denner2015}, we revisited the origin of the capillary time-step constraint using both numerical and signal-processing arguments, and arrived at a similar but slightly different formulation of the capillary time-step constraint than \citet{Brackbill1992}, given as\footnote{The effective phase velocity in the case where two oppositely propagating waves enter the same mesh cell simultaneously increases by factor $\sqrt{2}$, a result directly following from Eq.~\eqref{eq:c}, rather than factor 2 \citep{Denner2015}.} \begin{equation} \Delta t_\sigma = \frac{\Delta x}{\sqrt{2} \, c} =\sqrt{\frac{\rho_\text{a}+\rho_\text{b}}{2 \pi \sigma} \Delta x^3}. \label{eq:tsigma_Denner} \end{equation} We also demonstrated that, if the velocity of the flow at the interface has a magnitude similar to the phase velocity $c$, the Doppler shift associated with capillary waves propagating along a moving interface ought to be taken into account, and provided the first numerical results that clearly delineate the capillary time-step constraint. Contrary to the traditional CFL condition, $\Delta t \leq \Delta x / |\vecu|$ \citep{Courant1928}, which arises from the flow velocity $\vecu$ and is proportional to the mesh spacing $\Delta x$, the capillary time-step constraint is proportional to $\Delta x^{3/2}$, since $c \propto \Delta x^{-1/2}$. As a consequence, the capillary time-step constraint dominates the maximum applicable time-step for interfacial flows at small lengthscales ({\it e.g.}~microfluidics), applications with quasi-static heat and mass transfer ({\it e.g.}~the evaporation of a sessile drop) and, in general, simulations with high spatial resolution, as it is now routinely afforded by adaptive mesh refinement algorithms in conjunction with modern high-performance computing resources. Even at very small scales, such as in microfluidic applications, viscous contributions with a physical origin are typically not able to mitigate the capillary time-step constraint \citep{Popinet2018, Denner2015}. Already in their original proposition of the capillary time-step constraint, \citet{Brackbill1992} hypothesised that an implicit treatment of surface tension should allow to breach or even eliminate the capillary time-step constraint. \citet{Hou1994a} presented a boundary integral method with an implicit surface tension treatment for irrotational incompressible flows in two dimensions that is evidently free of the capillary time-step constraint, demonstrating the necessity of an implicit surface tension treatment in eliminating the capillary time-step constraint. Following the work of \citet{Bansch2001}, \citet{Hysing2006} proposed a semi-implicit surface tension treatment for a two-dimensional finite-element method, which, in essence, incorporates the interface position at the new time instance implicitly. \citet{Raessi2009} translated this surface tension treatment subsequently to finite-volume methods. \citet{Hysing2006} and \citet{Raessi2009} reported stable results for time-steps exceeding the capillary time-step constraint, although the solution was not stable for arbitrarily large time-steps (assuming other relevant time-step restrictions, {\it e.g.}~the CFL condition, are satisfied). We previously proposed an algorithm in which the continuity, momentum and VOF advection equations are implicitly coupled and solved as a single system of equations \citep{Denner2015}, treating the surface tension term semi-implicit with respect to the VOF colour function. However, this method does not allow to breach the capillary time-step constraint, which led us to conclude \revB{(incorrectly)} that an implicit formulation of the source term representing surface tension cannot eliminate the capillary time-step constraint using interface capturing methods \citep{Denner2015}. A fully-Lagrangian method for incompressible free-surface flows was proposed by \citet{Zheng2015}, based on a Marker-and-Cell (MAC) method in conjunction with a Lagrangian fluid mesh in the vicinity of the fluid interface, which are coupled implicitly and ensure an exact balance between surface tension and pressure. To date, this is the only method for three-dimensional interfacial flows that has been demonstrated to be practically free of the capillary time-step constraint, whereas a numerical framework based on an interface capturing method with comparable capabilities has yet to be published. In this article, we propose a numerical framework for interfacial flows based on a VOF method that is able to breach the capillary time-step constraint. We achieve this by extending a fully-coupled pressure-based algorithm \citep{Denner2020} with an algebraic VOF method that is implicitly coupled to the governing flow equations, enabling an implicit treatment of surface tension based on the Continuum Surface Force (CSF) model \citep{Brackbill1992}. The presented results demonstrate that a time-step larger than the capillary time-step constraint can be applied with this new numerical framework, as long as other relevant time-step restrictions ({\em e.g.}~the CFL condition) are satisfied. However, a new time-step constraint arises that depends on surface tension, density and viscosity, but which is less restrictive than the classical capillary time-step constraint. \section{Mathematical model} In this study, we consider interfacial flows of two immiscible and incompressible Newtonian fluids. Such an interfacial flow is governed by the continuity equation \begin{equation} \frac{\partial u_i}{\partial x_i} = 0 , \label{eq:continuity} \end{equation} with $\mathbf{x}$ the spatial coordinate and $\vecu$ the velocity vector, and the momentum equations \begin{equation} \rho \left( \frac{\partial u_j}{\partial t} + \frac{\partial u_i u_j}{\partial x_i} \right) = - \frac{\partial p}{\partial x_j} + \frac{\partial \tau_{ji}}{\partial x_i} + S_j , \label{eq:momentum} \end{equation} where $\rho$ is the density, $t$ denotes time, $p$ is the pressure and $\mathbf{S}$ is the source term representing surface tension. The stress tensor $\boldsymbol{\tau}$ is given as \begin{equation} \tau_{ji} = \mu \left(\frac{\partial u_j}{\partial x_i} + \frac{\partial u_i}{\partial x_j} \right) , \label{eq:stresstensor} \end{equation} where $\mu$ is the dynamic viscosity. A VOF method \citep{Hirt1981} is adopted to model the transport and interaction of two immiscible fluids. The two fluids are distinguished by the indicator function $\zeta$, which is defined as \begin{equation} \zeta(\mathbf{x}) = \begin{cases} 0, &\text{if} \ \mathbf{x} \in \Omega_\text{a}\\ 1, &\text{if} \ \mathbf{x} \in \Omega_\text{b} \end{cases} \end{equation} where $\Omega = \Omega_\text{a} \cup \Omega_\text{b}$ is the computational domain, with $\Omega_\text{a}$ and $\Omega_\text{b}$ the subdomains occupied by fluids ``a'' and ``b'', respectively. The indicator function $\zeta$ is advected by the underlying flow \begin{equation} \frac{\text{D}\zeta}{\text{D}t} = \frac{\partial \zeta}{\partial t} + \frac{\partial u_i \zeta}{\partial x_i} - \zeta \frac{\partial u_i}{\partial x_i} = 0. \label{eq:vofadvection_prod} \end{equation} The fluid properties are defined over the entire computational domain using the indicator function, {\em e.g.}~for density $\rho = (1-\zeta) \rho_a + \zeta \rho_b$. However, to focus the discussion on the discretisation and influence of surface tension, both interacting fluids are assumed to have the same density, $\rho=\rho_\text{a} = \rho_\text{b}$, and viscosity, $\mu=\mu_\text{a} = \mu_\text{b}$, throughout this study. \section{Numerical framework} \label{sec:numerics} The proposed numerical framework builds upon a class of fully-coupled pressure-based algorithms for single-phase \citep{Denner2018c, Denner2020} and interfacial flows \citep{Denner2014a, Denner2015}, with the aim of treating all discretised and linearised governing equations implicitly in a single linear system of governing equations, $\boldsymbol{\mathcal{A}} \cdot \boldsymbol{\phi} = \mathbf{b}$. This linear system of governing equations is solved simultaneously for the pressure $p$, the velocity $\mathbf{u} \equiv (u~v~w)^\text{T}$ and the discrete colour function $\psi$, using the Block-Jacobi pre-conditioner and the BiCGSTAB solver of the software library PETSc \citep{petsc-user-ref,petsc-web-page}. The nonlinear nature of the governing equations is accounted for by means of an inexact Newton method \citep{Dembo1982}, whereby the deferred terms resulting from the applied linearisation procedure are updated iteratively, until the nonlinear system of governing equations satisfies predefined conservation criteria, as illustrated in Figure \ref{fig:flowchart}. The discretisation of the governing equations is based on an established second-order finite-volume method and utilises a collocated variable arrangement \citep{Denner2020}, with the fluxes through the mesh faces computed by a momentum-weighted interpolation \citep{Bartholomew2018}. The proposed numerical framework is not inherently limited to VOF methods, but may similarly be used in conjunction with level-set or phase-field methods. \begin{figure}[t] \begin{center} \begin{small} \begin{tikzpicture} \node (pro0) [process] {Update \mbox{previous} time-levels: $\chi^{(t-2\Delta t)}\leftarrow \chi^{(t-\Delta t)}$ $\chi^{(t-\Delta t)}\leftarrow \chi^{(n)\phantom{...)}}$ $\vartheta_f^{(t-\Delta t)} \leftarrow \vartheta_f^{(n)\phantom{....}}$}; \node (pro1a) [process, below of=pro0, yshift=-1cm] {\mbox{Gather coefficients} \mbox{and assemble} \mbox{$\boldsymbol{\mathcal{A}}$ and $\mathbf{b}$}}; \node (pro1b) [process, below of=pro1a, yshift=-0.5cm] {Solve $\boldsymbol{\mathcal{A}} \cdot \boldsymbol{\phi} = \mathbf{b}$}; \node (pro2) [process, below of=pro1b, yshift=-0.5cm] {Update \mbox{deferred} quantities: $\chi^{(n)} \leftarrow \chi^{(n+1)}$}; \node (pro3) [process, below of=pro2, yshift=-0.5cm] {Compute $\kappa^{(n)}$ and $\vartheta_f^{(n)}$}; \node (dec1) [decision, below of=pro3, yshift=-0.7cm] {Conservation satisfied?}; \draw [arrow] (pro0) -- (pro1a); \draw [arrow] (pro1a) -- (pro1b); \draw [arrow] (pro1b) -- (pro2); \draw [arrow] (pro2) -- (pro3); \draw [arrow] (pro3) -- (dec1); \draw [arrow] (dec1) --+(-2.7cm,0) |- (pro1a); \draw [arrow] (dec1) --+(+2.7cm,0) |- (pro0); \node at (-2.1,-8) {no}; \node at (2.1,-8) {yes}; \node [rotate=90] at (-2.95,-5) {$n=n+1$}; \node [rotate=90] at (2.95,-5) {$t=t+\Delta t$}; \end{tikzpicture} \end{small} \caption{Flow chart of the solution procedure of the discretised and linearised system of governing equations, where $n$ is the nonlinear iteration counter, $\chi \in \{p,u,v,w,\psi \}$ are the solution variables, $\kappa$ is the interface curvature (see Section \ref{sec:surfacetension}) and $\vartheta_f$ is the advecting velocity through mesh face $f$ (see Section \ref{sec:mwi}). The coefficient matrix $\boldsymbol{\mathcal{A}}$ holds all coefficients for the implicitly sought solution variables $\chi^{(n+1)}$ of the discretised governing equations (see Section \ref{sec:solution}) and $\boldsymbol{\phi}$ is the solution vector. The right-hand side vector $\mathbf{b}$ holds the deferred contributions of the previous iteration ($\chi^{(n)}$, $\kappa^{(n)}$, $\vartheta_f^{(n)}$) and the contributions of the previous time-levels ($\chi^{(t-\Delta t)}$, $\chi^{(t-2\Delta t)}$, $\vartheta_f^{(t-\Delta t)}$).} \label{fig:flowchart} \end{center} \end{figure} \subsection{Discretised governing equations} In the following we assume an equidistant Cartesian mesh with mesh spacing $\Delta x$. The continuity equation \eqref{eq:continuity} for each computational cell $P$ is readily discretised as \begin{equation} \sum_f F_f\imp = 0, \label{eq:continuity_disc} \end{equation} where subscript $f$ denotes faces adjacent to mesh cell $P$, $F_f = \vartheta_f A_f$ is the flux through face $f$, $A_f$ is the area of face $f$ and $\vartheta_f$ is the advecting velocity (outward pointing with respect to cell $P$) obtained from a momentum-weighted interpolation, as further detailed in Section \ref{sec:mwi}. The superscript $(n+1)$ denotes implicitly solved quantities and the superscript $(n)$ denotes deferred quantities, where $n$ is the nonlinear iteration counter. Applying the second-order backward Euler scheme to discretise the transient term \citep{Ferziger2020}, the discretised momentum equations \eqref{eq:momentum} follow as \begin{equation} \begin{split} & \rho \left[\frac{3 u_{j,P}\imp - 4 u_{j,P}\pts + u_{j,P}^{(t-2\Delta t)}}{2\Delta t} V_P + \sum_f \left( \tilde{u}_{j,f}\imp F_f\rhs + \tilde{u}_{j,f}\rhs F_f\imp - \tilde{u}_{j,f}\rhs F_f\rhs \right)\right] \\ &= - \sum_f \overline{p}_f\imp n_{j,f} A_f + \mu \sum_f \left( \frac{u_{j,Q}\imp - u_{j,P}\imp}{\Delta x} \, + \left. \overline{\frac{\partial {u}_{i}}{\partial x_j}}\right|_f\imp n_{i,f}\right) A_f + {S}_{j,P}\imp V_P , \end{split} \label{eq:momentum_disc} \end{equation} where $\tilde{\square}$ denotes flux-limited values at face $f$ (see below), $\overline{\square}_f=(\square_P+\square_Q)/2$ denotes a linear interpolation of the cell-centred values to face $f$, with $Q$ the neighbour cell of $P$ adjacent to face $f$, $\mathbf{n}_f$ is the normal vector of face $f$ (outward pointing with respect to cell $P$) and, as further described in Section \ref{sec:surfacetension}, $\mathbf{S}$ is the source term representing surface tension. A Newton linearisation is applied to the advection term of the momentum equations \citep{Denner2018c}, to enable an implicit treatment of both the velocity $\tilde{\mathbf{u}}_f$ and the flux $F_f$. The two interacting fluids are represented discretely by the colour function $\psi$, \begin{equation} \psi_P = \frac{1}{V_P} \iiint_{V_P} \zeta \, \text{d}V, \label{eq:discretecolour} \end{equation} which is advected based on Eq.~\eqref{eq:vofadvection_prod} by \begin{equation} \begin{split} \frac{3 \psi_P\imp - 4 \psi_P\pts + \psi_P^{(t-2\Delta t)}}{2\Delta t} V_P &+ \sum_f \left( \tilde{\psi}_f\imp F_f\rhs + \tilde{\psi}_f\rhs F_f\imp - \tilde{\psi}_f\rhs F_f\rhs \right) \\ &- \left(\psi_P\imp \sum_f F_f\rhs + \psi_P\rhs \sum_f F_f\imp - \psi_P\rhs \sum_f F_f\rhs \right) = 0. \end{split} \label{eq:vof_disc} \end{equation} In analogy to the discretisation of the momentum equations \eqref{eq:momentum_disc}, the transient term is discretised with the second-order backward Euler scheme and both spatial terms are linearised with a Newton linearisation. The flux-limited face values are given as \begin{equation} \tilde{\psi}_f\imp = \begin{cases} \psi_U\imp + \xi^{(\psi)}_f \left(\psi_D\imp-\psi_U\imp\right), ~~&\text{if}~\left|\psi_U\rhs-\psi_D\rhs\right| > \varepsilon_\text{tol}\\ \psi_U\imp,~~&\text{else}\\ \end{cases} \label{eq:differencing_colour} \end{equation} and \begin{equation} \tilde{\vecu}_f\imp = \begin{cases} \vecu_U\imp + \xi^{(\psi)}_f \left(\vecu_D\imp-\vecu_U\imp\right), ~~&\text{if}~\left|\psi_U\rhs -\psi_D\rhs\right| > \varepsilon_\text{tol}\\ \dfrac{\vecu_P\imp + \vecu_Q\imp}{2},~~&\text{else}\\ \end{cases} \end{equation} where $\xi^{(\psi)}_f$ is the flux limiter of the colour function, $\varepsilon_\text{tol} = 10^{-6}$ is a predefined tolerance, and subscripts $U$ and $D$ denote the upwind and downwind cells of face $f$, respectively. For the purpose of this study, the CICSAM scheme \citep{Ubbink1999} is used to compute the flux limiter $\xi^{(\psi)}_f$, but any other advection scheme may also be applied. Note that the advection scheme applied to the colour function in Eq.~\eqref{eq:differencing_colour} away from the interface, where $\psi = \text{const.}$, is irrelevant and choosing upwind differencing is here a matter of convenience. \subsection{Surface tension} \label{sec:surfacetension} Surface tension is modelled using the CSF model \citep{Brackbill1992}, with which the source term $\mathbf{S}$ representing surface tension is given as \begin{equation} \mathbf{S} = \sigma \, \kappa \, \nabla \psi, \label{eq:csf} \end{equation} where $\sigma$ is the surface tension coefficient and $\kappa$ is the interface curvature. \revA{The gradient of the colour function is discretised using the Gauss theorem, analogous to the discretisation of the pressure gradient in the momentum equations \eqref{eq:momentum_disc}, thus it is given as \begin{equation} \nabla \psi_P \approx \frac{1}{V_P} \sum_f \overline{\psi}_f \mathbf{n}_f A_f, \label{eq:colourgrad} \end{equation} which facilitates a force-balanced discretisation \citep{Denner2014a}.} The source term $\mathbf{S}$ is treated implicitly by applying a Newton linearisation of the interface curvature and the gradient of the colour function to yield \revA{ \begin{equation} \mathbf{S}_P\imp \approx \sigma \kappa_P\rhs \nabla \psi_P\imp + \sigma \kappa_P\imp \nabla \psi_P\rhs - \sigma \kappa_P\rhs \nabla \psi_P\rhs, \label{eq:csf_imp0} \end{equation} which, by inserting Eq.~\eqref{eq:colourgrad}, becomes} \begin{equation} \mathbf{S}_P\imp \approx \frac{\sigma}{V_P} \left(\kappa_P\rhs \sum_f \overline{\psi}_f\imp \mathbf{n}_{f} A_f + \kappa_P\imp \sum_f \overline{\psi}_f\rhs \mathbf{n}_{f} A_f - \kappa_P\rhs \sum_f \overline{\psi}_f\rhs \mathbf{n}_{f} A_f \right). \label{eq:csf_imp} \end{equation} Various approaches have been proposed for estimating the interface curvature from the discrete colour function, {\it e.g.}~\citep{Brackbill1992, Cummins2005, Raessi2007, Popinet2009, Denner2014a, Owkes2014a, Evrard2017, Evrard2020}. In this work, we employ the height-function (HF) method \cite{Evrard2020} both for its superior accuracy and ease of linearisation. In the HF method, when the $z$-component of the interface normal is dominant, curvature reads as \begin{equation} \kappa = \frac{-H_{xx}\left(1+H_y^2\right) - H_{yy}\left(1+H_x^2\right)+2H_xH_yH_{xy}}{\left(H_x^2+H_y^2+1\right)^{3/2}}, \end{equation} where $H_{\{x,y,xx,yy,xy\}}$ are the first and second partial derivatives of the ``heights of fluid'' computed along the $z$-direction\footnote{In case the $x$- or $y$-component of the normal is dominant, the indices of the partial derivatives are simply permuted.}. On a Cartesian mesh, a height of fluid can be trivially obtained by summing the colour function in a column of computational cells. The partial derivatives of the heights are calculated using central differences, therefore they can be expressed as linear combinations of the discrete colour function values in the set of cells from which the heights are computed. Consider, for instance, an interfacial computational cell $P$, $\psi_P \in \left]0,1\right[$, within which the $z$-component of the interface normal is dominant. One can construct a stencil $\mathcal{S}(P)$ centred around $P$, containing at most $3\times3\times N_H$ cells, with $N_H$ an odd number typically chosen between $5$ and $9$. For the portion of interface in cell $P$, the first partial derivative of the heights along the $x$-direction is approximated as \begin{equation} H_x = \sum_{N\in\mathcal{S}(P)} \beta_{x,N} \, \psi_N , \end{equation} where $\beta_{x,N}$ are coefficients arising from the application of central differences to the heights, themselves calculated as the sum of the discrete colour function values in the columns of $N_H$ computational cells of $\mathcal{S}(P)$. The other partial derivatives of the heights are expressed similarly. In the cell $P$, applying a Newton linearisation on curvature yields \begin{equation} \kappa\imp_P = \kappa\rhs_P + \sum_{N\in\mathcal{S}(P)} \left[\left(\psi\imp_N-\psi\rhs_N\right) \left(\frac{\partial \kappa_P}{\partial \psi_N}\right)\rhs \right]. \end{equation} Introducing the quantities \begin{align} \N_P & = -H_{xx}\left(1+H_y^2\right) - H_{yy}\left(1+H_x^2\right)+2H_xH_yH_{xy} \\ \D_P & = H_x^2+H_y^2+1 \end{align} the linearisation of curvature can be reformulated as \begin{equation} \kappa\imp_P = \kappa\rhs_P + \frac{1}{\D_P\rhs} \sum_{N\in\mathcal{S}(P)} \left[\left(\psi\imp_N-\psi\rhs_N\right) \left(\left(\frac{\partial \N_P}{\partial \psi_N}\right)\rhs - \frac{3}{2} \kappa\rhs_P \sqrt{\D_P\rhs} \left(\frac{\partial \D_P}{\partial \psi_N}\right)\rhs\right) \right]. \label{eq:Klinear} \end{equation} The partial derivatives on the right-hand side of Eq.~\eqref{eq:Klinear} read as \begin{align} \frac{\partial \N_P}{\partial \psi_N} & = -\left(1+H_{y}^2\right)\beta_{xx,N} -\left(1+H_{x}^2\right)\beta_{yy,N} \nonumber \\ & \quad \quad \quad \quad \quad +2\left(H_yH_{xy}-H_{x}H_{yy}\right)\beta_{x,N} +2\left(H_xH_{xy}-H_yH_{xx}\right)\beta_{y,N} +2H_xH_y\beta_{xy,N} , \\ \frac{\partial \D_P}{\partial \psi_N} & = 2 H_{x} \beta_{x,N} + 2 H_{y} \beta_{y,N}. \end{align} Note that for two-dimensional flow in the plane $(x,z)$, the previous expressions reduce to \begin{align} \N_P & = -H_{xx} , & \D_P & = H_x^2+1, \\ \frac{\partial \N_P}{\partial \psi_N} & = -\beta_{xx,N} , & \frac{\partial \D_P}{\partial \psi_N} & = 2 H_{x} \beta_{x,N}. \end{align} \subsection{Momentum-weighted interpolation} \label{sec:mwi} The advecting velocity $\vartheta_f=\vecu_f \cdot \mathbf{n}_f$ of the flux $F_f = \vartheta_f A_f$ through face $f$ is discretised using a momentum-weighted interpolation (MWI) \citep{Rhie1983}, which provides a direct coupling of pressure and velocity that eliminates pressure-velocity decoupling as a result of the applied collocated variable arrangement. Including the source term representing surface tension, the advecting velocity is defined as \citep{Denner2014a, Bartholomew2018} \begin{equation} \vartheta_f = \overline{\vecu}_{f} \cdot \mathbf{n}_{f} - \hat{d}_f \left( \nabla p_f - \overline{\nabla p}_f \right) \cdot \mathbf{n}_{f} + \hat{d}_f \left( \mathbf{S}_f - \overline{\mathbf{S}}_f \right) \cdot \mathbf{n}_{f} + \hat{d}_f \frac{\rho}{\Delta t} \left( \vartheta_f\pts - \overline{\vecu}_{f}\pts \cdot \mathbf{n}_{f} \right). \label{eq:mwi} \end{equation} The coefficient $\hat{d}_f$, derived in detail in \citep{Bartholomew2018}, represents the weighting factor of the MWI correction terms and, in the context of a fully-coupled algorithm, the strength of the implicit coupling provided by the MWI. The pressure gradient at the face is discretised as \begin{equation} \nabla p_f \cdot \mathbf{n}_f \approx \frac{p_Q-p_P}{\Delta x} \end{equation} and, analogous to the pressure term in the momentum equations \eqref{eq:momentum_disc}, the cell-centred pressure gradient is discretised using the Gauss theorem as \begin{equation} \nabla p_P \approx \frac{1}{V_P} \sum_f \overline{p}_f \mathbf{n}_f A_f. \end{equation} Together, the pressure terms in Eq.~\eqref{eq:mwi}$, (\nabla p_f - \overline{\nabla p}_f)$, then act as a low-pass filter with respect to pressure \citep{Bartholomew2018}. Similarly, the surface tension terms are discretised as \citep{Denner2014a} \begin{equation} \mathbf{S}_f \cdot \mathbf{n}_f \approx \sigma \overline{\kappa}_f \frac{\psi_Q-\psi_P}{\Delta x} \end{equation} and \begin{equation} \mathbf{S}_P \approx \sigma \kappa_P \nabla \psi_P = \frac{\sigma \kappa_P}{V_P} \sum_f \overline{\psi}_f \mathbf{n}_f A_f. \end{equation} Applying the discretisation described above, along with a Newton linearisation of the surface tension terms analogous to the source term representing surface tension in the momentum equations, the advecting velocity is proposed as \begin{equation} \begin{split} \vartheta_f\imp = \overline{\vecu}_{f}\imp \cdot \mathbf{n}_{f} &-~\hat{d}_f \left[\frac{p_Q\imp-p_P\imp}{\Delta x} - \frac{1}{2} \left( \nabla p_P\imp + \nabla p_Q\imp \right) \cdot \mathbf{n}_{f} \right] \\ &+~\hat{d}_f \sigma \left[\overline{\kappa}_f\rhs \frac{\psi_Q\imp-\psi_P\imp}{\Delta x} - \frac{1}{2} \left(\kappa_P\rhs \nabla \psi_P\imp + \kappa_Q\rhs \nabla \psi_Q\imp \right) \cdot \mathbf{n}_{f} \right]\\ &+~\hat{d}_f \sigma \left[\overline{\kappa}_f\imp \frac{\psi_Q\rhs-\psi_P\rhs}{\Delta x} - \frac{1}{2} \left(\kappa_P\imp \nabla \psi_P\rhs + \kappa_Q\imp \nabla \psi_Q\rhs \right) \cdot \mathbf{n}_{f} \right]\\ &-~\hat{d}_f \sigma \left[\overline{\kappa}_f\rhs \frac{\psi_Q\rhs-\psi_P\rhs}{\Delta x} - \frac{1}{2} \left(\kappa_P\rhs \nabla \psi_P\rhs + \kappa_Q\rhs \nabla \psi_Q\rhs \right) \cdot \mathbf{n}_{f} \right] \\ &+~\hat{d}_f \frac{\rho}{\Delta t} \left( \vartheta_f\pts - \overline{\vecu}_{f}\pts \cdot \mathbf{n}_{f} \right). \end{split} \label{eq:mwi_implicit} \end{equation} With this discrete formulation, each term of the current time-level in the advecting velocity, Eq.~\eqref{eq:mwi}, makes an implicit contribution to the solution variables $\chi \in \{p,u,v,w,\psi\}$. \subsection{Solution procedure} \label{sec:solution} The discretised governing equations \eqref{eq:continuity_disc}, \eqref{eq:momentum_disc} and \eqref{eq:vof_disc} are solved simultaneously in a single linear system of discretised equations, given for a three-dimensional mesh with $N$ cells as \begin{equation} \begin{pmatrix} \boldsymbol{\mathcal{A}}^{p}_\text{cont.} & \boldsymbol{\mathcal{A}}^u_\text{cont.} & \boldsymbol{\mathcal{A}}^v_\text{cont.} & \boldsymbol{\mathcal{A}}^w_\text{cont.} & \boldsymbol{\mathcal{A}}^\psi_\text{cont.} \\ \boldsymbol{\mathcal{A}}^{p}_\text{$x$-mom.} & \boldsymbol{\mathcal{A}}^u_\text{$x$-mom.} & \boldsymbol{\mathcal{A}}^v_\text{$x$-mom.} & \boldsymbol{\mathcal{A}}^w_\text{$x$-mom.} & \boldsymbol{\mathcal{A}}^\psi_\text{$x$-mom.} \\ \boldsymbol{\mathcal{A}}^{p}_\text{$y$-mom.} & \boldsymbol{\mathcal{A}}^u_\text{$y$-mom.} & \boldsymbol{\mathcal{A}}^v_\text{$y$-mom.} & \boldsymbol{\mathcal{A}}^w_\text{$y$-mom.} & \boldsymbol{\mathcal{A}}^\psi_\text{$y$-mom.} \\ \boldsymbol{\mathcal{A}}^{p}_\text{$z$-mom.} & \boldsymbol{\mathcal{A}}^u_\text{$z$-mom.} & \boldsymbol{\mathcal{A}}^v_\text{$z$-mom.} & \boldsymbol{\mathcal{A}}^w_\text{$z$-mom.} & \boldsymbol{\mathcal{A}}^\psi_\text{$z$-mom.} \\ \boldsymbol{\mathcal{A}}^{p}_\text{\sc vof} & \boldsymbol{\mathcal{A}}^u_\text{\sc vof} & \boldsymbol{\mathcal{A}}^v_\text{\sc vof} & \boldsymbol{\mathcal{A}}^w_\text{\sc vof} & \boldsymbol{\mathcal{A}}^\psi_\text{\sc vof} \\ \end{pmatrix} \cdot \begin{pmatrix} \boldsymbol{\phi}^p \\ \boldsymbol{\phi}^u \\ \boldsymbol{\phi}^v \\ \boldsymbol{\phi}^w \\ \boldsymbol{\phi}^\psi \end{pmatrix} = \begin{pmatrix} \mathbf{b}_\text{cont.} \\ \mathbf{b}_\text{$x$-mom.} \\ \mathbf{b}_\text{$y$-mom.} \\ \mathbf{b}_\text{$z$-mom.} \\ \mathbf{b}_\text{\sc vof} \end{pmatrix}, \label{eq:eqsysfull} \end{equation} for the solution variables pressure $p$, velocity $\vecu \equiv (u~v~w)^\text{T}$ and colour function $\psi$. In Eq.~\eqref{eq:eqsysfull}, $\boldsymbol{\mathcal{A}}^\chi_\text{eq.}$ denotes the coefficient submatrix of size $N \times N$ of the continuity equation (eq.~= cont.), the momentum equations associated with the three Cartesian coordinate axes (eq.~= $x$-mom., eq.~= $y$-mom., eq.~= $z$-mom.) and the VOF advection equation (eq.~= {\sc vof}) for the respective solution variable $\chi \in \{p,u,v,w,\psi\}$. The solution subvectors of length $N$ for solution variable $\chi$ are denoted as $\boldsymbol{\phi}^\chi$ and the right-hand side subvectors of length $N$ of the five discretised governing equations, which contain all deferred contributions and contributions from previous time-levels, are denoted as $\mathbf{b}_\text{eq.}$. This fully-coupled system of equations~\eqref{eq:eqsysfull} is solved in an iterative fashion using the Block-Jacobi pre-conditioner and the BiCGSTAB solver of the software library PETSc \citep{petsc-user-ref,petsc-web-page}, as illustrated in Figure \ref{fig:flowchart}. The Newton linearisation of the advection terms in both the momentum equations and the VOF advection equation yields an implicit contribution of the fluxes, which in turn introduces an implicit pressure, velocity and colour function dependency in all governing equations. The implicit formulation of the fluxes by the MWI presented in Eq.~\eqref{eq:mwi_implicit} is, thus, the primary coupling term of the discretised governing equations. Notably, the implicit pressure and velocity dependency in the VOF advection is a novel building block for solving interfacial flows. Both the colour function gradient and the curvature are treated implicitly with respect to the colour function, yielding an implicit CSF model for surface tension. \section{Differences to previously proposed methods} \label{sec:comparisonprevious} Previous work aimed at breaching the capillary time-step constraint in the context of interface capturing methods, as already mentioned in the introduction of this article, has focused on incorporating the interface position at the new time instance implicitly in the momentum equations \citep{Hysing2006, Raessi2009} and on a coupled solution algorithm in which all governing equations are solved simultaneously \citep{Denner2015}. \citet{Hysing2006} proposed a semi-implicit surface tension treatment in the context of a finite-element method, whereby the interface position at the new time instance is incorporate implicitly in the source term representing surface tension. Using the CSF method to model surface tension in a finite-volume discretisation, the source term representing surface tension, accounting for the interface position at the new time instance implicitly, is given as \citep{Raessi2009} \begin{equation} \mathbf{S}_P^{(n+1)} \approx \sigma \, \kappa_P\rhs \, \nabla \psi_P\rhs + \sigma \, \Delta t \, |\nabla \psi_P\rhs| \, \Delta_\text{s} \vecu_P\imp, \label{eq:surfacetensionRaessi} \end{equation} where $\Delta_\text{s}$ is the Laplace-Beltrami operator with respect to the interface. This formulation is convenient since the implementation in existing implicit numerical frameworks, by adding an implicit contribution of velocity to the momentum equations, is straightforward. However, a simple dimensional analysis reveals that $\mu_\Sigma = \sigma \, \Delta t \, |\nabla \psi|$ represents a viscosity acting in the interface region \citep{Denner2017} and, in conjunction with a Laplacian of velocity $\Delta_\text{s} \vecu$, the second term on the right-hand side of Eq.~\eqref{eq:surfacetensionRaessi} acts as an additional viscous stress term at the interface. As discussed by \citet{Popinet2018}, the resulting increase in dissipation in the vicinity of the interface is responsible for the success of this semi-implicit surface tension formulation. To this end, we demonstrated that an explicit implementation of the additional surface dissipation term in Eq.~\eqref{eq:surfacetensionRaessi} also allows to breach the capillary time-step constraint by up to one order of magnitude \citep{Denner2017}, at the cost of artificially increasing the effective viscous stresses acting in the vicinity of the fluid interface. The implicit treatment of surface tension as part of the proposed algorithm, Eq.~\eqref{eq:csf_imp}, does not introduce an additional viscous stress term at the interface. We previously presented a coupled implicit algorithm \citep{Denner2015}, similar to the algorithm proposed in Section \ref{sec:numerics}, which also solves the VOF advection equation together with the continuity and momentum equations in a single system of linear equations, and features a semi-implicit treatment of the source term representing surface tension. Contrary to the proposed algorithm, in the previous algorithm \citep{Denner2015} the advection terms of the momentum equations \eqref{eq:momentum_disc} and the VOF advection equation \eqref{eq:vof_disc} were linearised with a Picard linearisation, \begin{equation} \iiint_{V} \frac{\partial u_i \phi}{\partial x_i} \, \text{d}V \approx \sum_f \tilde{\phi}_{f}\imp F_f\rhs, \label{eq:picard} \end{equation} where $\phi \in \{\vecu, \psi\}$, and the source term representing surface tension only considered the gradient of the colour function in an implicit manner, \begin{equation} \mathbf{S}_P\imp \approx \frac{\sigma \kappa_P\rhs}{V_P} \sum_f \overline{\psi}_f\imp \mathbf{n}_f A_f. \label{eq:surfacetension2015} \end{equation} In contrast, in the algorithm proposed in Section \ref{sec:numerics}, a Newton linearisation is applied to linearise the advection terms and the interface curvature is also treated implicitly with respect to the colour function $\psi$. In the previous algorithm \citep{Denner2015}, the advecting velocity was also formulated implicit in the solution variables $\chi \in \{p,u,v,w,\psi\}$, given as \begin{equation} \begin{split} \vartheta_f\imp = \overline{\vecu}_{f}\imp \cdot \mathbf{n}_{f} &-~\hat{d}_f \left[\frac{p_Q\imp-p_P\imp}{\Delta x} - \frac{1}{2} \left(\nabla p_P\rhs + \nabla p_Q\rhs \right) \cdot \mathbf{n}_{f} \right] \\ &+~\hat{d}_f \sigma \left[\overline{\kappa}_f\rhs \frac{\psi_Q\imp-\psi_P\imp}{\Delta x} - \frac{1}{2} \left(\kappa_P\rhs \nabla \psi_P\rhs + \kappa_Q\rhs \nabla \psi_Q\rhs \right) \cdot \mathbf{n}_{f} \right]\\ &+~\hat{d}_f \frac{\rho}{\Delta t} \left( \vartheta_f\pts - \overline{\vecu}_{f}\pts \cdot \mathbf{n}_{f} \right), \end{split} \label{eq:mwi_implicit2015} \end{equation} yet, contrary to the formulation proposed in Eq.~\eqref{eq:mwi_implicit}, only the face-based gradients of pressure and colour function were treated implicitly, while the interface curvature and the cell-centred (Gaussian) gradients were deferred. As a result of Eqs.~\eqref{eq:surfacetension2015} and \eqref{eq:mwi_implicit2015}, each governing equation has an implicit contribution of the colour function and the continuity equation has implicit contributions with respect to all solution variables $\chi \in \{p,u,v,w,\psi\}$. Nevertheless, the algorithm previously proposed in \citep{Denner2015} does not allow to breach the capillary time-step constraint. \section{Results} \label{sec:results} The ability of the numerical framework proposed in Section \ref{sec:numerics} to breach the capillary time-step constraint is tested and validated using two well-defined test-cases: the Laplace equilibrium of an interface with constant curvature ({\em i.e.}~a circular and a spherical interface) and a standing capillary wave. Both test-cases have frequently been used to scrutinise numerical methods, are governed by surface tension, and analytical solutions are available for comparison. Furthermore, as both considered test-cases yield relatively small interface deformations and the interface curvature is spatially well resolved at all times, these test-cases are not subject to issues associated with the fidelity of the interface transport or the well-posedness of the heights used for the evaluation of the interface curvature, issues that are outside the scope of this study. \subsection{Laplace equilibrium} \label{sec:laplace} A circular or spherical interface subject to surface tension in a quiescent flow is in mechanical equilibrium, with zero velocity in the entire domain and a pressure difference between the outside and the inside of the interface given by the Young-Laplace equation, \begin{equation} \Delta p = \sigma \kappa. \end{equation} However, due to errors associated with the numerical framework and the discrete representation of the interface topology, unphysical parasitic currents arise. For a force-balanced numerical framework, such as the one applied in this study \citep{Denner2014a}, these parasitic currents should decay exponentially due to viscous dissipation and converge to a value commensurate with machine precision or the chosen solver tolerance, with the interface attaining a discrete equilibrium shape \citep{Popinet2009}. Following the work of \citet{Popinet2009}, a circular interface with a diameter of $D=0.8 \, \text{m}$ is simulated, with $\rho = 1 \, \text{kg/m}^3$ and $\sigma = 1 \, \text{N/m}$. The viscosity $\mu$ is defined by the considered Laplace number, \begin{equation} \text{La} = \frac{\rho \sigma D}{\mu^2}. \end{equation} Exploiting the symmetry of the problem, only one quarter of the circular interface is simulated, situated at the origin of a square domain with edge length $1 \, \text{m}$. The domain is represented by an equidistant Cartesian mesh with $32\times 32$ cells. Figure \ref{fig:laplaceeq_2d} shows the evolution of the root-mean-square (RMS) of the velocity in the domain for $\text{La} \in \{120, 1200, 12000\}$, simulated with different time-steps $\Delta t$. A stable solution is obtained even if the time-step exceeds the capillary time-step constraint $\Delta t_\sigma$, Eq.~\eqref{eq:tsigma_Denner}, by factor $50$, and the equilibrium velocity, which is of negligible magnitude, is hardly affected. Considering the same case in three-dimensions, {\em i.e.}~a spherical interface, with $\text{La} = 120$, a similar evolution of the velocity RMS is observed in Figure \ref{fig:laplaceeq_3d}. \begin{figure}[t] \begin{center} \includegraphics[width=\linewidth]{JCOMP-D-21-01568R1_Fig2} \caption{Evolution of the root-mean-square (RMS) of the flow velocity $\mathbf{u}$ of the two-dimensional Laplace equilibrium with Laplace number $\text{La} \in \{120, 1200, 12000\}$, for different time-steps $\Delta t$, normalised by the capillary velocity $U_\sigma = \sqrt{\sigma / (\rho D)}$ and the viscous timescale $\tau_\mu = \rho D^2/\mu$. $\Delta t_\sigma$ refers to the capillary time-step constraint given in Eq.~\eqref{eq:tsigma_Denner}.} \label{fig:laplaceeq_2d} \end{center} \end{figure} \begin{figure}[t] \begin{center} \includegraphics[width=0.381\textwidth]{JCOMP-D-21-01568R1_Fig3} \caption{Evolution of the root-mean-square (RMS) of the flow velocity $\mathbf{u}$ of the three-dimensional Laplace equilibrium with Laplace number $\text{La} =120$, for different time-steps $\Delta t$, normalised by the capillary velocity $U_\sigma = \sqrt{\sigma / (\rho D)}$ and the viscous timescale $\tau_\mu = \rho D^2/\mu$. $\Delta t_\sigma$ refers to the capillary time-step constraint given in Eq.~\eqref{eq:tsigma_Denner}.} \label{fig:laplaceeq_3d} \end{center} \end{figure} Hence, the capillary time-step constraint can be breached for both two- and three-dimensional interfaces, without affecting the discrete balance between surface tension and the flow significantly. \subsection{Capillary wave} \label{sec:capillarywave} A single two-dimensional capillary wave is considered to test the fidelity and robustness with which the proposed algorithm can predict surface-tension-driven motion when the capillary time-step constraint is breached. The motion of the fluid interface and the flow are driven solely by surface tension and, considering a small initial wave amplitude and equal properties of the bulk fluids, an analytical solution to the corresponding initial-value problem is available for comparison \citep{Prosperetti1981}. The considered capillary wave with wavelength $\lambda=10^{-4} \, \text{m}$ and initial amplitude $a_0 = 0.01 \lambda$ is situated in a domain with dimensions $\lambda \times 3 \lambda$, illustrated in Figure \ref{fig:capillarywave_schematic}, resolved with a mesh spacing of $\Delta x = \lambda/100$. Periodic boundary conditions are applied on the side walls. Both fluids have the same density, $\rho=1 \, \text{kg/m}^3$, and the same viscosity, ranging from $\mu=5 \times 10^{-6} \, \text{Pa s}$ to $\mu=5 \times 10^{-2} \, \text{Pa s}$. The surface tension coefficient is $\sigma=0.01 \, \text{N/m}$. The capillary wave is fully characterised by its critical wavenumber $k_\text{c} \simeq 2 ^{2/3} \sigma (\rho_\text{a}+\rho_\text{b})/(\mu_\text{a}+\mu_\text{b})^2$, above which the oscillation of the capillary wave ceases \citep{Denner2017a}, and its undamped frequency $\omega_0 = \sqrt{\sigma k^3/(\rho_\text{a}+\rho_\text{b})}$, where $k=2 \pi/\lambda$ is the wavenumber. \begin{figure} \begin{center} \includegraphics[scale=0.7]{JCOMP-D-21-01568R1_Fig4} \caption{Schematic of the two-dimensional capillary wave with wavelength $\lambda$ and initial amplitude $a_0$.} \label{fig:capillarywave_schematic} \end{center} \end{figure} Figure \ref{fig:capillarywave_400lvc} shows the evolution of the amplitude of an oscillating capillary wave with a wavelength of $\lambda = 202 \, \lambda_\text{c}$, obtained with different time-steps $\Delta t$. The evolution of the wave amplitude is predicted accurately compared to the analytical solution, even with $\Delta t = 5 \, \Delta t_\sigma$. Although the numerical algorithm is stable for $\Delta t = 50 \, \Delta t_\sigma$, the result exhibits a visible discrepancy in comparison to the analytical solution. Nevertheless, this discrepancy is to be expected given the rather coarse temporal resolution of the oscillation, whereby each time-step is illustrated by a \revB{dot} in Figure \ref{fig:capillarywave_400lvc} for $\Delta t = 50 \, \Delta t_\sigma$. Similar observations can be made for the evolution of the amplitude of a relatively shorter capillary wave with a wavelength of $\lambda = 12.6 \, \lambda_\text{c}$ shown in Figure \ref{fig:capillarywave_100lvc}, where a time-step of $\Delta t = 100 \, \Delta t_\sigma$ yields a stable and reasonably accurate result. Changing the mesh resolution with which this capillary wave is resolved, but keeping the applied time-step unchanged, yields virtually identical results, as observed in Figure \ref{fig:capillarywave_dxcomp}. \begin{figure}[t] \includegraphics[width=\linewidth]{JCOMP-D-21-01568R1_Fig5} \caption{Evolution of the amplitude of a capillary wave with wavelength $\lambda = 202 \, \lambda_\text{c}$ and initial amplitude $a_0=0.01 \lambda$ obtained with different time-steps $\Delta t \in \{0.5,5,50\} \Delta t_\sigma$, where $\Delta t_\sigma$ is the capillary time-step constraint, Eq.~\eqref{eq:tsigma_Denner}. The evolution is presented relative to the undamped frequency $\omega_0$ of the capillary wave and the analytical solution of \citet{Prosperetti1981} is shown as a reference. For the numerical results, each \revB{dot} shows every $100^\text{th}$ time-step for $\Delta t = 0.5 \, \Delta t_\sigma$, every $10^\text{th}$ time-step for $\Delta t = 5 \, \Delta t_\sigma$ and every time-step for $\Delta t = 50 \, \Delta t_\sigma$.} \label{fig:capillarywave_400lvc} \end{figure} \begin{figure}[t] \includegraphics[width=\linewidth]{JCOMP-D-21-01568R1_Fig6} \caption{Evolution of the amplitude of a capillary wave with wavelength $\lambda = 12.6 \, \lambda_\text{c}$ and initial amplitude $a_0=0.01 \lambda$ obtained with different time-steps $\Delta t \in \{1,10,100\} \Delta t_\sigma$, where $\Delta t_\sigma$ is the capillary time-step constraint, Eq.~\eqref{eq:tsigma_Denner}. The evolution is presented relative to the undamped frequency $\omega_0$ of the capillary wave and the analytical solution of \citet{Prosperetti1981} is shown as a reference. For the numerical results, each \revB{dot} shows every $100^\text{th}$ time-step for $\Delta t = 1 \, \Delta t_\sigma$, every $10^\text{th}$ time-step for $\Delta t = 10 \, \Delta t_\sigma$ and every time-step for $\Delta t = 100 \, \Delta t_\sigma$.} \label{fig:capillarywave_100lvc} \end{figure} \begin{figure}[t] \begin{center} \includegraphics[width=0.38\linewidth]{JCOMP-D-21-01568R1_Fig7} \caption{Evolution of the amplitude of a capillary wave with wavelength $\lambda = 12.6 \, \lambda_\text{c}$ and initial amplitude $a_0=0.01 \lambda$. The results are obtained with different mesh resolutions $\Delta x \in \{\lambda/100, \lambda/200, \lambda/300\}$. The same time-step, corresponding to $\Delta t=10 \Delta t_\sigma$ for the mesh with $\Delta x =\lambda/100$, where $\Delta t_\sigma$ is the capillary time-step constraint, Eq.~\eqref{eq:tsigma_Denner}, is applied regardless of the mesh resolution. The evolution is presented relative to the undamped frequency $\omega_0$ of the capillary wave and every $10^\text{th}$ time-step is shown by a mark.} \label{fig:capillarywave_dxcomp} \end{center} \end{figure} Figure \ref{fig:capillarywave_contourplot} shows contour plots of the colour function $\psi$ after a single time-step with $\Delta t=10 \Delta t_\sigma$, using different modelling assumptions. Neglecting any of the three main implicit extensions proposed in comparison to our previously presented coupled algorithm \citep{Denner2015}, as detailed in Section \ref{sec:comparisonprevious}, yields an unphysical interface topology even after a single time-step, if the capillary time-step is breached. Only the proposed algorithm in which all contributions of the solution variables $\chi \in \{p,u,v,w,\psi\}$ are treated implicitly yields a stable result for time-steps exceeding the capillary time-step constraint. \begin{figure}[t] \begin{center} \subfloat[All implicit] {\includegraphics[width=0.485\linewidth, trim=550pt 650pt 370pt 650pt , clip=true]{JCOMP-D-21-01568R1_Fig8a}} \quad \subfloat[Picard linearisation of the advection terms] {\includegraphics[width=0.485\linewidth, trim=550pt 650pt 370pt 650pt , clip=true]{JCOMP-D-21-01568R1_Fig8b}} \\ \subfloat[Explicit interface curvature] {\includegraphics[width=0.485\linewidth, trim=550pt 650pt 370pt 650pt , clip=true]{JCOMP-D-21-01568R1_Fig8c}} \quad \subfloat[Semi-implicit MWI formulation] {\includegraphics[width=0.485\linewidth, trim=550pt 650pt 370pt 650pt , clip=true]{JCOMP-D-21-01568R1_Fig8d}} \caption{Contour plots of the VOF colour function $\psi$ after one time-step of the wave with wavelength $\lambda = 202 \, \lambda_\text{c}$ and initial amplitude $a_0=0.01 \lambda$ obtained with a time-step of $\Delta t = 10 \Delta t_\sigma$ using different modelling assumptions. (a) Shows the result obtained using the proposed {\em all implicit} algorithm; (b) a Picard linearisation, see Eq.~\eqref{eq:picard}, is applied for the linearisation of the advection terms of the momentum and VOF advection equations; (c) the interface curvature is treated explicitly, with the source term representing surface tension given by Eq.~\eqref{eq:surfacetension2015}; (d) the semi-implicit MWI formulation given in Eq.~\eqref{eq:mwi_implicit2015} is applied, but with the proposed implicit treatment of the interface curvature.} \label{fig:capillarywave_contourplot} \end{center} \end{figure} \section{Revised time-step constraint} Although the proposed algorithm exhibits a favourable behaviour for time-steps exceeding the capillary time-step constraint, the time-step that yields a stable solution is still limited. We presume that this time-step limitation is associated with the particular discretisation and linearisation chosen to cast the governing equations in a form amenable to numerical analysis using linear algebra. Because the governing equations are linearised, we consider linear stability analysis to analyse the stability of the linearised system of governing equations. Based on a linear stability analysis under the assumption of an interface perturbation with small amplitude and sufficiently small Reynolds number, \citet{Galusinski2008} proposed a maximum time-step $\Delta t_\star$ for a stable solution of surface-tension-driven flows of \begin{equation} \Delta t_\star^2 - c_2 \, \frac{{\mu} \, \Delta x}{\sigma} \Delta t_\star - c_1 \, \frac{{\rho} \, \Delta x^3}{\sigma}=0, \label{eq:dtPolynom_Galusinski} \end{equation} where $c_1$ and $c_2$ are constants. With the wavelength of the shortest spatially resolved capillary waves being $\lambda_\sigma = 2 \Delta x$, and assuming\footnote{\citet{Galusinski2008} did neither specify nor discuss how the density and viscosity of the two-phase system are defined in their linear stability analysis.} $\hat{\mu} = \mu_\text{a} + \mu_\text{b}$ and $\hat{\rho} = \rho_\text{a} + \rho_\text{b}$, we reformulate Eq.~\eqref{eq:dtPolynom_Galusinski} as \begin{equation} \Delta t_\star^2 - a_2 \, \frac{\hat{\mu} \, \lambda_\sigma}{\sigma} \Delta t_\star - a_1 \, \frac{\hat{\rho} \, \lambda_\sigma^3}{\sigma}=0, \label{eq:dtPolynom_GalusinskiNew} \end{equation} or, by inserting the capillary timescale $\tau_\sigma = \sqrt{\hat{\rho} \lambda_\sigma^3/\sigma}$ and the viscocapillary timescale $\tau_\text{vc} = \hat \mu \lambda_\sigma /\sigma$ \citep{Castrejon-Pita2015}, \begin{equation} \Delta t_\star^2 - a_2 \, \tau_\text{vc} \, \Delta t_\star - a_1 \, \tau_\sigma^2 =0. \label{eq:dtPolynom_GalusinskiNew_short} \end{equation} The maximum time-step $\Delta t_\star$ follows as the positive root of Eq.~\eqref{eq:dtPolynom_GalusinskiNew_short}, \begin{equation} \Delta t_\star = \frac{a_2 \, \tau_\text{vc} + \sqrt{a_2^2 \, \tau_\text{vc}^2 + 4 \, a_1 \, \tau_\sigma^2 }}{2}. \label{eq:dt_Galusinski} \end{equation} This suggests that the maximum applicable time-step is proportional to the capillary timescale $\tau_\sigma$ for small Ohnesorge numbers with respect to $\lambda_\sigma$, \begin{equation} \text{Oh} = \frac{\tau_\text{vc}}{\tau_\sigma} = \frac{\hat{\mu}}{\sqrt{\hat{\rho} \sigma \lambda_\sigma}}, \label{eq:Oh} \end{equation} where surface tension dominates. In contrast, a maximum time-step proportional to the viscocapillary timescale $\tau_\text{vc}$ is relevant for large $\text{Oh}$, where both viscosity and surface tension govern the interface motion. Note that the capillary time-step constraint $\Delta t_\sigma$, as presented in Eq.~\eqref{eq:tsigma_Denner}, associated with an explicit treatment of surface tension is recovered for $a_1 = (16\pi)^{-1}$ and $a_2=0$. Figure \ref{fig:newdt} shows the maximum time-step that yields a stable result for the Laplace equilibrium (Section \ref{sec:laplace}) and the capillary wave (Section \ref{sec:capillarywave}), normalised by the capillary time-step constraint $\Delta t_\sigma$ given by Eq.~\eqref{eq:tsigma_Denner}, for different Ohnesorge numbers as defined in Eq.~\eqref{eq:Oh}. For both cases, the maximum applicable time-step exceeds the capillary time-step constraint $\Delta t_\sigma$ and the revised time-step constraint presented in Eq.~\eqref{eq:dt_Galusinski} is in remarkable agreement with the maximum time-step over the considered eight orders of magnitude of the Ohnesorge number. \begin{figure}[t] \begin{center} \subfloat[Laplace equilibrium] {\includegraphics[width=0.43\linewidth]{JCOMP-D-21-01568R1_Fig9a} \label{fig:laplace_newdt}}\quad \subfloat[Capillary wave] {\includegraphics[width=0.43\linewidth]{JCOMP-D-21-01568R1_Fig9b} \label{fig:capillarywave_newdt_sigma}} \caption{Maximum applicable time-step $\Delta t$, normalised by the capillary time-step constraint $\Delta t_\sigma$ defined in Eq.~\eqref{eq:tsigma_Denner}, as a function of the Ohnesorge number $\text{Oh}$ defined in Eq.~\eqref{eq:Oh}, for (a) the Laplace equilibrium of Section \ref{sec:laplace} and (b) the capillary wave (with three different viscosities) of Section \ref{sec:capillarywave}. The approximate maximum time-step $\Delta t_\star$ using Eq.~\eqref{eq:dt_Galusinski} is shown with $a_1=6$ and $a_2=98$ in (a) and with $a_1=1$ and $a_2=20$ in (b), \revA{where suitable values for $a_1$ and $a_2$ are approximated}.} \label{fig:newdt} \end{center} \end{figure} The results presented in Figure \ref{fig:newdt} further indicate that the coefficients $a_1$ and $a_2$ are case dependent; $a_1 \approx 6$ and $a_2 \approx 98$ for the Laplace equilibrium, whereas $a_1 \approx 1$ and $a_2 \approx 20$ for the capillary wave. More generally, the results suggest a maximum applicable time-step of $\mathcal{O} (\tau_\sigma)$ in the surface-tension-dominated regime ($\text{Oh} \ll 0.01$) and of $\mathcal{O} (10 \, \tau_\text{vc}) - \mathcal{O} (100 \, \tau_\text{vc})$ in the viscocapillary regime ($\text{Oh} \gg 0.01$). \revA{We are currently not aware of a method to estimate precise values for $a_1$ and $a_2$ from first principles or based on the discretisation of the governing equations. However, given that the revised time-step constraint $\Delta t_\star$ is given by a second-order polynomial, the coefficients can be approximated for a given case for all practically relevant Ohnesorge numbers with only two results, one for $\mathrm{Oh} \ll 0.01$ and one for $\mathrm{Oh} \gg 0.01$.} \section{Conclusions} \label{sec:conclusions} The capillary time-step constraint, first formulated by \citet{Brackbill1992}, presents a severe impediment to the performance of most interfacial flow simulations with surface tension. Breaching or even eliminating the capillary time-step constraint is generally thought to be possible with an implicit treatment of surface tension. However, previous work in this direction using interface capturing methods has either been unsuccessful \citep{Denner2015} or has introduced an additional viscous dissipation term \citep{Hysing2006, Raessi2009, Denner2017}. This led us to conclude that it is not possible to breach the capillary time-step constraint with interface capturing methods \citep{Denner2015}, a conclusion that was \revB{(correctly)} met with scepticism \citep{Popinet2018}. In this study, we have presented a fully-coupled pressure-based algorithm, based on a second-order finite-volume discretisation, featuring an implicit VOF method and an implicit linearised treatment of surface tension. Three implementation principles are at the heart of this algorithm: (i) making all governing equations implicitly dependent on pressure, velocity and the colour function (through the momentum-weighted interpolation), (ii) linearising all nonlinear terms with a Newton linearisation, and (iii) treating every term involving pressure, velocity or the colour function implicitly. The ensuing system of discretised linear equations, which includes the continuity, momentum and VOF advection equations, is then solved simultaneously for pressure, velocity and the colour function. We have shown that this algorithm is able to breach the capillary time-step constraint; hence, interface capturing methods are indeed able to breach the capillary time-step constraint, which proves our early conclusion in \citep{Denner2015} to be incorrect. The presented results further indicate that the proposed algorithm features the minimum level of implicitness required for breaching the capillary time-step constraint. However, this study also highlights the limitation of the proposed approach; by how much the capillary time-step constraint can be breached depends on the fluid properties as well as the considered case. To this end, the maximum time-step that yields a stable solution is described accurately by a revised time-step constraint based on a linear stability analysis previously proposed by \citet{Galusinski2008}. As a result, the maximum time-step depends on the surface tension coefficient, density and viscosity, as well as two coefficients that are case dependent. Nevertheless, stable results with time-steps larger than the capillary time-step constraint have been obtained for all simulations considered. With this study, we provide a proof-of-concept for a fully-coupled algorithm and an implicit treatment of surface tension, based on an interface capturing method, that allows to breach the capillary time-step constraint. While the primary aim of breaching the capillary time-step constraint has been achieved, additional work is required to make such an algorithm applicable to solve problems relevant in practice and further exploit its benefits. For instance, \revA{in this proof-of-concept we have only considered density and viscosity ratios equal to unity. Making the proposed algorithm applicable to flows with density and viscosity ratios as they occur in practice, while simultaneously allowing to breach the capillary time-step, may additionally require to treat the density and viscosity in a semi-implicit fashion as a function of the colour function. Furthermore,} in the employed algebraic VOF method, the advection of the colour function is based on the CICSAM scheme, which is known for requiring very small time-steps to retain a sharp interface \citep{Gopala2008} and, thus, stands in opposition to maximising the applied time-step. The colour function in the vicinity of the interface is, hence, smeared quickly if the interface moves significantly and, as a consequence, the height-function method becomes ill-posed. Contemporary interface advection schemes, such as THINC \citep{Xiao2011}, would perhaps be better suited to maximise the time-step effectively and robustly. Also, the proposed numerical framework is not limited to implicit algebraic VOF methods, but may be based upon an implicit phase-field or level-set method, and it remains to be determined which interface capturing method is best suited to exploit the benefits of the proposed fully-coupled approach. \revB{Even a geometric VOF method \citep[\textit{e.g.}][]{Youngs1982} could, in principle, be used in conjunction with the proposed numerical framework, on the condition that it can be implemented implicitly and solved in a coupled system simultaneous with the governing equations.} \section*{Acknowledgements} This research was funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation), grant numbers 452916560 and 458610925.

36aca413f3f6eafa51b1ff72ff46a36d27360ed0

\section{Introduction} \input{tables/overview} As machine learning models are increasingly deployed in real-world scenarios, it has motivated the development of interpretable machine learning (ML) as a research field with the goal of understanding ML models, performing model debugging, and using these insights to better inform the interaction between AI and humans in joint decision making~\cite{gilpin2018explaining,bhatt2020explainable,chen2022interpretable}. Recently, the promise of multimodal models for real-world representation learning in numerous applications such as multimedia~\cite{liang2021multibench,liang2018multimodal,1667983}, affective computing~\cite{liang2018ranking,PORIA201798}, robotics~\cite{kirchner2019embedded,lee2019making}, finance~\cite{doi:10.1177/0170840618765019}, dialogue~\cite{Pittermann2010}, human-computer interaction~\cite{dumas2009multimodal,obrenovic2004modeling}, and healthcare~\cite{xu2019multimodal} has invigorated research into multimodal machine learning, which brings unique challenges for both computational and theoretical research given the heterogeneity of various data sources and difficulty of capturing correspondences between modalities~\cite{baltruvsaitis2018multimodal}. Among one of these core challenges is \textit{interpretable multimodal learning} with the end goal of empowering various stakeholders by providing insights into multimodal learning, improving model design, or debugging models and datasets. Recent work in interpretable multimodal learning has therefore focused on constructing interpretable multimodal models via careful model design~\cite{tsai2020multimodal,zadeh2018multimodal,park2018multimodal} or performing post-hoc explanations of black-box multimodal models~\cite{goyal2016towards,chandrasekaran2018explanations}. However, existing works typically focus on building interpretable models using suitable inductive biases, such as designing multimodal routing networks~\cite{tsai2020multimodal}, graph-based fusion~\cite{zadeh2018multimodal}, or multimodal explanation networks to highlight visual importance~\cite{park2018multimodal}. Some of these approaches also require the collection of specialized datasets annotated for visual explanations as intermediate steps in training interpretable models~\cite{park2018multimodal}. On the other hand, with the trend towards large-scale modeling or pre-training as an alternative over individual modality-specific or task-specific models~\cite{visualbert,liang2021multibench}, it is increasingly important to design general-purpose approaches that (1) are able to generate post-hoc explanations for arbitrary black-box models, and (2) does not assume anything about the modality or classification task itself. As a step towards more fine-grained interpretations of general-purpose multimodal models across arbitrary tasks, we propose \textsc{DIME}, an interpretation method for black-box multimodal models. While existing work has been able to generate useful explanations to help humans understand model decision-making processes~\cite{chandrasekaran2018explanations}, they are often only performed at one step of the entire multimodal decision-making process. These singular steps typically include attributing feature importance~\cite{park2018multimodal,chandrasekaran2018explanations} or representation importance~\cite{tsai2020multimodal,zadeh2018multimodal}. The core idea in \textsc{DIME}\ is to provide more fine-grained interpretations by disentangling a multimodal model into unimodal contributions (\textbf{UC}) and multimodal interactions (\textbf{MI}). We show that this key insight enables more accurate and fine-grained analysis of multimodal models while maintaining generality across arbitrary modalities, model architectures~\cite{kamath2021mdetr,lxmert}, and tasks~\cite{balanced_vqa_v2,johnson2017clevr}. Through a comprehensive suite of experiments on both synthetic and real-world multimodal tasks, we show that \textsc{DIME}\ is able to accurately perform disentanglement and generate reliable explanations for both UC and MI. Using \textsc{DIME}, we are able to gain a deeper understanding of model behavior on challenging multimodal tasks. For example, on VQA 2.0~\cite{goyal2017making}, we successfully use \textsc{DIME}\ to determine whether the model uses correct multimodal interactions to answer the questions, as shown in Figure~\ref{intro}. By providing these model explanations to a human annotator, they are able to gain additional insights on model behavior and better determine whether UC, MI, or both are the dominant factor behind the model's predictions on individual datapoints. Furthermore, \textsc{DIME}\ presents a step towards debugging and improving these models by systematically revealing certain undesirable behaviors. \section{Related Work} Interpretable machine learning as a research field aims to further our understanding of AI models, empower various stakeholders to build trust in AI models, perform model debugging, and use these insights to better inform the interaction between AI and humans in joint decision making~\cite{gilpin2018explaining,bhatt2020explainable,chen2022interpretable}. We cover related concepts in interpreting unimodal models and multimodal models. \subsection{Interpreting Unimodal Models} Related work has studied approaches for better understanding unimodal models used for vision, language, and audio modalities. These approaches can be roughly categorized into interpretable ML as designing models which are understandable by design, and explainable ML which focuses on producing post-hoc explanations for black-box models~\cite{rudin2019stop}. In the former, methods such as Concept Bottleneck Models~\cite{koh2020concept} and fitting sparse linear layers~\cite{wong2021leveraging} or decision trees on top of deep feature representations~\cite{wan2020nbdt} have emerged as promising choices marrying the expressive power of deep features with the interpretable decision-making processes of linear models or decision trees. In the latter, approaches such as saliency maps~\cite{simonyan2013deep,smilkov2017smoothgrad}, using surrogate models to interpret local decision boundaries~\cite{lime}, feature visualizations~\cite{yosinski2015understanding,erhan2009visualizing}, and assigning semantic concepts~\cite{bau2017network} all aim to provide insight into model predictions for specific input instances. We refer the reader to~\citet{chen2022interpretable} for a survey and taxonomy of interpretable ML approaches, as well as~\citet{bhatt2020explainable} for an analysis of how interpretable and explainable ML tools can be used in the real world. \subsection{Interpreting Multimodal Models} Similar to the interpretation of unimodal models, recent work in interpretable multimodal learning can be categorized into two sections: (1) constructing interpretable multimodal models via careful model design~\cite{tsai2020multimodal,zadeh2018multimodal,park2018multimodal} or (2) performing post-hoc explanations of black-box multimodal models~\cite{goyal2016towards,chandrasekaran2018explanations}. In the former, multimodal routing networks~\cite{tsai2020multimodal}, graph-based fusion techniques~\cite{zadeh2018multimodal,liang2018computational}, multimodal explanation networks to highlight visual importance~\cite{park2018multimodal}, hard-attention~\cite{chen2017multimodal}, and neuro-symbolic reasoning methods~\cite{vedantam2019probabilistic,andreas2016neural} have emerged as strong design choices as a step towards more interpretable multimodal learning. These approaches individually focus on building interpretable components for either modality importance~\cite{park2018multimodal}, cross-modal interactions~\cite{tsai2020multimodal,zadeh2018multimodal,liang2018computational}, or the reasoning process on top of cross-modal interactions~\cite{vedantam2019probabilistic,andreas2016neural}. While these approaches provide reliable interpretations by virtue of model design, they are typically restricted to a certain set of modalities or tasks. On the other hand, we propose a more general approach that is able to generate post-hoc explanations for arbitrary black-box multimodal models, and does not assume anything about the modality or classification task itself. In the latter section on post-hoc explainability of black-box multimodal models, related work has similarly gravitated towards aiming to understand either modality importance~\cite{goyal2016towards,chandrasekaran2018explanations,kanehira2019multimodal} or cross-modal interactions in pretrained language and vision transformer models~\cite{frank2021vision,cao2020behind,parcalabescu2021seeing,li2020does}. Perhaps most related to our work is~\citet{wang2021m2lens} proposing M2Lens, an interactive visual analytics system to visualize and explain black-box multimodal models for sentiment analysis through both unimodal and multimodal contributions. Our approach further disentangles the two types of contributions, which allows us to generate visualizations on each and gain insight into which input features are involved in multimodal interactions. Our approach is also not restricted to sentiment analysis. \subsection{Representation Disentanglement} Related to our work is the idea of learning disentangled data representations - mutually independent latent variables that each explain a particular variation of the data~\cite{Bengio:2013:RLR:2498740.2498889,locatello2018challenging}. Disentangled representation learning has been shown to improve both generative and discriminative performance in multimodal tasks~\cite{tsai2018learning}. If the factors of variation are known, many methods learn latent attributes that individually control each variation of data by supervised training~\cite{karaletsos2015bayesian,reed2014learning,cheung2014discovering}. If the factors are partially known or unknown, deep generative models can be used to impose an isotropic Gaussian prior on the latent variables~\cite{vae2013,rubenstein2018latent,Higgins2016VAELB}, maximize the mutual information between a subset of latent variables and the data~\cite{chen2016infogan}, or to encourage the distribution of representations to be factorial and hence independent~\cite{pmlr-v80-kim18b}. Particularly related to our work is empirical multimodally-additive function projection (EMAP)~\cite{hessel2020emap}, an approach for disentangling the effects of unimodal (additive) contributions from cross-modal interactions in multimodal tasks. \subsection{Dataset and Model Biases} One core motivation for interpretable ML is to enable a better understanding of the model's decision-making process so as to check whether model behavior is as intended. Using these tools, researchers have uncovered several biases existing in machine learning models and datasets. These biases include undesirable associations captured either in the data or the model, which do not reflect decision-making as one would expect. For example, a line of work in visualizing and understanding multimodal models has uncovered unimodal biases in the language modality of VQA tasks~\cite{jabri2016revisiting,agrawal2016analyzing,anand2018blindfold,cadene2019rubi}, which then inspired follow-up datasets to elevate the importance of visual understanding through VQA 2.0~\cite{goyal2017making}. Similar visualizations also led to improved performance on image captioning tasks by relying less on gender biases and spurious correlations~\cite{hendricks2018women}. Our approach towards better visualizing and understanding multimodal models is also inspired by these insights, and we believe that our fine-grained and general approach will motivate future work towards removing biases from a wider range of datasets and models beyond the prototypical language and vision tasks. \section{Method: \textsc{DIME}} Our approach, \textsc{DIME}\ (short for \textsc{DIsentangled Multimodal Explanations}), is primarily based on disentangling a multimodal model into unimodal contributions (\textbf{UC}) and multimodal interactions (\textbf{MI}), before performing fine-grained visualizations on each disentangled factor. In this section, we introduce precise definitions of unimodal contributions and multimodal interactions, before explaining how disentanglement and interpretations are performed. \subsection{Unimodal Contributions and Multimodal Interactions} Unimodal contributions $(\textsc{UC})$ represent information gained by only looking at one of the modalities without interacting with any other modalities, while multimodal interactions $(\textsc{MI})$ are information gained from cross-referencing inputs from multiple modalities~\cite{hessel2020emap}. Multimodal models make decisions using a combination of information from both unimodal contributions and multimodal interactions. For example, in Figure~\ref{intro}, the model assigns a high likelihood to ``glass'' because (1) just by looking at the image, there are many glass objects (unimodal contributions) and (2) by cross-referencing with text, the model focuses on the glass table and assigns a high likelihood to ``glass'' (multimodal interaction). Therefore, to performed fine-grained interpretation in a multimodal model $M$, we first propose a new method to disentangle the model into two submodels: \begin{equation} M = \textsc{UC}(M) + \textsc{MI}(M), \end{equation} where $\textsc{UC}(M)$ represents the unimodal contributions within $M$ and $\textsc{MI}(M)$ represents the multimodal interactions within $M$. We can then run visualizations on each sub-model in order to generate human-interpretable visualizations of unimodal contributions and multimodal interactions (see Figure~\ref{fig:nonexistent} for an overview of \textsc{DIME}). To generate visual explanations, we choose LIME~\cite{lime}, a widely used interpretation method for black-box models. \begin{figure*} \vspace{0mm} \includegraphics[width=1.0\textwidth]{figs/mainpic.pdf} \caption{High level illustration of \textsc{DIME}: we disentangle the model $M$ into two: unimodal contributions (UC) and multimodal interactions (MI), before running visualizations on each sub-model (e.g., using LIME~\cite{lime}) in order to generate fine-grained human-interpretable visualizations of each.} \label{fig:nonexistent} \vspace{-3mm} \end{figure*} \subsection{Model Disentanglement} \label{sec:proof} Let $M$ be the multimodal model that we wish to disentangle into unimodal contributions and multimodal interactions. For simplicity, suppose $M$ takes in two modalities as input and produces pre-softmax logits on $C$ classes as output. Therefore, we can view $M$ as a function that maps two inputs $x_1,x_2$ from two modalities to a output logit vector $V$, i.e., $V = M(x_1,x_2)$. Our goal will be to disentangle the function $M$ into a sum of two functions, one representing unimodal contributions and one representing multimodal interactions. Formally, we would like to write $M$ as $M(x_1,x_2)=g_1(x_1)+g_2(x_2)+g_{12}(x_1,x_2)$, where $g_1$ and $g_2$ are unimodal contributions from the two input modalities, respectively, and $g_{12}$ represents multimodal interactions. By definition of multimodal interactions, we require that $\mathbb{E}_{x_1}g_{12}(x_1,x_2)=0$ for all $x_2$ and $\mathbb{E}_{x_2}g_{12}(x_1,x_2)=0$ for all $x_1$ so that $g_{12}$ contains no unimodal contribution. We will show that under this definition, for each $M$ there will be a unique $g_{12}$ that satisfies these rules. We will compute $g_1(x_1)+g_2(x_2)$ using a similar method to EMAP~\cite{hessel2020emap}. We define $\textsc{UC}(M)$ as \begin{equation} \textsc{UC}(M(x_1,x_2)) = \mathbb{E}_{x_1}(M(x_1,x_2)) + \mathbb{E}_{x_2}(M(x_1,x_2)) - \mathbb{E}_{x_1,x_2}(M(x_1,x_2)). \end{equation} \textbf{Theorem 1} below (equations 3-5, proof in Appendix) states that $\textsc{UC}(M)$ indeed represents $g_1+g_2$. \begin{eqnarray} && \textsc{UC}(M(x_1,x_2)) \\ &=& \mathbb{E}_{x_1}(M(x_1,x_2))+ \mathbb{E}_{x_2}(M(x_1,x_2)) - \mathbb{E}_{x_1,x_2}(M(x_1,x_2)) \\ &=& g_1(x_1)+g_2(x_2). \end{eqnarray} Thus, we can compute $g_{12}(x_1,x_2)$ by subtracting $\textsc{UC}(M(x_1,x_2))$ from $M(x_1,x_2)$, which we name $\textsc{MI}(M)$. Formally, \begin{eqnarray} && \textsc{MI}(M(x_1,x_2)) \\ &=& M(x_1,x_2) - \textsc{UC}(M(x_1,x_2)) \\ &=& g_{12}(x_1,x_2). \end{eqnarray} This also shows that $g_{12}$ can be uniquely determined. In practice, to compute $\textsc{UC}(M(x_1,x_2))$ and $\textsc{MI}(M(x_1,x_2))$, we use a sampling method similar to~\cite{hessel2020emap}, where we sample $N$ datapoints $x^{(i)}=(x^{(i)}_1,x^{(i)}_2)$ including the point we want to explain $x=(x_1,x_2)$ as one of them, and computing each expectation in $\textsc{UC}(M(x_1,x_2))$ by approximating \begin{align} \mathbb{E}_{x_1}(M(x_1,x_2)) &= \sum_{i \in [N]} M(x^{(i)}_1,x_2), \\ \mathbb{E}_{x_2}(M(x_1,x_2)) &= \sum_{i \in [N]} M(x_1,x^{(i)}_2), \\ \mathbb{E}_{x_1,x_2}(M(x_1,x_2)) &= \sum_{i \in [N]}\sum_{j \in [N]} M(x^{(i)}_1,x^{(j)}_2). \end{align} Figure~\ref{fig:illusdisent} illustrates this disentanglement process. \begin{figure} \centering \includegraphics[width=0.45\textwidth]{figs/disent.pdf} \caption{An illustration of the disentangling process of \textsc{DIME}. We disentangle a model into two: $\textsc{UC}(M) = g_1 + g_2$ and $\textsc{MI}(M) = g_{12}$, corresponding to unimodal contributions and multimodal interactions respectively.} \label{fig:illusdisent} \end{figure} However, to compute $\textsc{UC}(M(x_1,x_2))$ and $\textsc{MI}(M(x_1,x_2))$, we will need to run forward passes through the model a total of $N^2$ times. In section~\ref{sec:fast} we will show an algorithm that computes this more efficiently by amortizing across multiple datapoints. \subsection{Interpreting Disentanglement} Now that we have disentangled the model into two, we will generate human-interpretable explanations on each modality using LIME~\cite{lime}. LIME works by subdividing the input into distinct features, and then randomly perturbing the features $S$ times to see how the perturbations on the features affect the model output logits of a specific class $c$. LIME then fits a linear model mapping the perturbations on each feature to the logits of $c$. The linear model weights on each feature gives the explanation of that feature: if the weight is positive, it means that this feature supports the decision of class $c$; if the weight is negative, it means that this feature is against the decision of class $c$; the larger the weight's absolute value, the stronger the contribution is. Visually, the weights can also be used to generate a human-interpretable visualization: for images, each feature is typically a part of the image, so the parts with the highest absolute weights can be highlighted in green for positive and red for negative contributions. For text, each feature is typically a word, so the explanation can be summarized as a histogram of weights of each word (see Figure~\ref{fig:nonexistent} for an example). When running LIME on multimodal inputs, we run LIME on one modality at a time, treating the inputs to all other modalities as constant and only perturbing the inputs to that one modality. We denote the generated explanation on model $M$, datapoint $(x_1,x_2)$, and modality $i$ as $\textsc{LIME}_i(M(x_1,x_2))$. After disentanglement into unimodal contributions $\textsc{UC}(M(x_1,x_2))$ and multimodal interactions $\textsc{MI}(M(x_1,x_2))$, our approaches enables the generation of four fine-grained explanations: \begin{itemize} \item $\textsc{UC}_1 = \textsc{LIME}_1(\textsc{UC}(M(x_1,x_2)))$, the explanation of modality 1's unimodal contributions. \item $\textsc{UC}_2 = \textsc{LIME}_2(\textsc{UC}(M(x_1,x_2)))$, the explanation of modality 2's unimodal contributions. \item $\textsc{MI}_1 = \textsc{LIME}_1(\textsc{MI}(M(x_1,x_2)))$, the explanation of modality 1's contribution to multimodal interactions. \item $\textsc{MI}_2 = \textsc{LIME}_2(\textsc{MI}(M(x_1,x_2)))$, the explanation of modality 2's contribution to multimodal interactions. \end{itemize} \subsection{Improving Efficiency} \label{sec:fast} Since running LIME on a black box model usually requires running the model many times (equal to the LIME sample size $S$), it can be costly to treat $\textsc{UC}(M)$ or $\textsc{MI}(M)$ as black-box models and run LIME on then directly - running $\textsc{UC}(M)$ involves computing $\mathbb{E}_{x_1,x_2}(M(x_1,x_2))$ which requires running $N^2$ forward passes where $N$ is the number of samples used for EMAP, so the total procedure of running \textsc{DIME}\ on one datapoint can take $O(SN^2)$ runs of $M$. In order to make the process faster, we use the following algorithmic trick: we fix $N$ datapoints from the dataset, and then run $M$ on all $N^2$ combinations of the two modalities amongst the $N$ points, and store the resulting logits in a $N\times N\times C$ array $L$ (where $C$ is the number of classes in this task). When we want to run \textsc{DIME}\ on any one of those $N$ points (let's say the $i$th point), for each perturbed LIME sample (WLOG let's say we're running LIME on modality 1, so modality 1 is perturbed in the LIME sample), we make a deep copy of $L$ called $L'$, re-run $M$ on the combination of the perturbed modality 1 input and all $N$ modality 2 inputs, replace the values in the ith row of $L'$ with the results, and compute $\textsc{UC}(M)$ on this LIME sample with the updated table $L'$. Using this trick, after amortizing the one-time initial $O(N^2)$ runs of $M$, each followup \textsc{DIME}\ run on any of the $N$ points only takes $O(SN)$ runs of $M$. See details in Algorithm 1 in the Appendix. \section{Experiments} In this section, we will perform a set of experiments to fully evaluate the reliability and usefulness of \textsc{DIME}\ in interpreting multimodal models. We will be using 3 datasets: a synthetic dataset, CLEVR~\cite{johnson2017clevr}, and VQA 2.0~\cite{balanced_vqa_v2}, and with one corresponding state-of-the-art model for each: MLP, MDETR~\cite{kamath2021mdetr} and LXMERT~\cite{lxmert}. When dealing with datasets involving image and text modalities, we will refer to the two modalities as $(V,T)$ respectively (e.g., $\textsc{UC}_V$ would refer to the \textsc{DIME}\ explanation on image unimodal contribution). Our experiments are designed to illustrate the following takeaway messages of using \textsc{DIME}\ to analyze multimodal models: \begin{enumerate} \item Our method can reliably disentangle the model and generate accurate explanations for both UC and MI, correlating highly with their respective ground truths (section~\ref{rq1}). \item In more difficult tasks such as CLEVR and VQA, and with more complex models, \textsc{DIME}\ can still disentangle the model reliably. We show that changing the text input affects $\textsc{UC}_V$ (explanation on image unimodal contribution) little but affects $\textsc{MI}_V$ (explanation on multimodal interactions from the image side) significantly (section~\ref{rq1}). \item \textsc{DIME}\ gives additional insight into understanding multimodal model behavior by answering whether the model relies mostly on UC, MI, or both in making the prediction (section~\ref{rq2}). \item \textsc{DIME}\ also enables human users to debug and improve models by identifying which input features are used in MI and revealing undesirable behavior in models (section~\ref{rq3}). \end{enumerate} Following these results, we will discuss limitations and future works (section~\ref{limit}). \subsection{Setup} \subsubsection{Datasets} We will use three datasets: a synthetic dataset to enable controlled variations between unimodal and multimodal interactions, as well as two large-scale multimodal datasets: CLEVR, and VQA 2.0. The \textbf{synthetic dataset $D$} is designed to model a task that requires both unimodal (additive) contributions and multimodal interactions to solve correctly. According to prior work~\cite{hessel2020emap}, the dot product of two modalities requires non-additive cross-modal interaction, while the sum of two vectors is additive. Therefore, we design a synthetic dataset $D$ by randomly generating two $10$-dimensional vectors following $N(0,1)$ independently for each element, and then computing the sum of all elements in both vectors plus the dot product of the two vectors. If the result's absolute value is below $0.01$, we discard this point; otherwise, we assign a $0/1$ label based on the sign of the result. We generate $100,000$ points to form $D$ and divide it into train/valid/test splits by $8/1/1$ ratio. \textbf{CLEVR}~\cite{johnson2017clevr} is a diagnostic dataset designed for language and visual reasoning. The dataset consists of synthesized images of 3D shapes of various colors, sizes, and materials on a gray background, For each image, there are several questions about the shapes' attributes, positions, and numbers. This dataset has been widely used for diagnostic purposes to find model weaknesses. \textbf{VQA 2.0}~\cite{balanced_vqa_v2} is a dataset containing various questions on real-world images. It is designed to force multimodal interactions, especially incorporating the visual aspect, by sometimes having the same question with two different answers on two different images. This dataset is interesting because models have been shown to occasionally ``guess'' correct answers purely from unimodal contributions or with the wrong visual grounding~\cite{cadene2019rubi,anand2018blindfold}. \textsc{DIME}\ will enable us to study how often models rely on undesirable unimodal biases and further understand the model's decision-making process. \subsubsection{Models} For synthetic dataset $D$, we train a \textbf{4-layer MLP} (with input size $20$ and hidden layer sizes $100,200,10,2$ respectively) on $D$ that reaches $97.3\%$ accuracy on the test split. For CLEVR dataset, we will be using a pretrained \textbf{MDETR}~\cite{kamath2021mdetr} that achieves $99.7\%$ test accuracy. For VQA 2.0, we will be using pretrained \textbf{LXMERT}~\cite{lxmert}, one of the best models on the dataset, with a $72.5\%$ test accuracy. \subsection{Research Questions and Results} \input{tables/synth} \input{tables/relia} \subsubsection{\textbf{RQ1:} Can \textsc{DIME}\ reliably disentangle a model into unimodal contributions and multimodal interactions and generate accurate explanations for both UC and MI in practice?} \label{rq1} \ In section~\ref{sec:proof}, we have theoretically shown that \textsc{DIME}\ can disentangle a model into unimodal contributions and multimodal interactions. To show that this also holds in practice (when expectation computations are replaced by sampling), we will run \textsc{DIME}\ on our trained model $M$ using $1,000$ randomly selected datapoints in the test split of our synthetic dataset $D$, on label $1$ (i.e., that the sum of all elements of both vectors plus the dot-product of the two vectors are positive). For each point $(d_1,d_2)$ in $D$, since we are classifying whether the sum of all elements in $d_1$ and $d_2$ as well as the dot product of $d_1$ and $d_2$, the ground truth UC explanation on each modality will be $d_1$ and $d_2$ respectively, and the ground truth MI explanation will be element-wise product $d_1*d_2$. Therefore, for each generated explanation on input data $(d_1,d_2)$, we will compute the Pearson Correlation between the explanation weights of the $10$ features with the values of the $10$ features of $d_1$, the values of the $10$ features of $d_2$, and the $10$ features in the element-wise product of $d_1$ and $d_2$. In addition to \textsc{DIME}, we also run LIME under the same settings as an ablation and compute average correlations. The results are shown in Table~\ref{tab:synth}. We found that within each datapoint ($d_1$,$d_2$), there is a strong correlation between each \textsc{DIME}-generated unimodal explanation ($\textsc{UC}_1, \textsc{UC}_2$) and the corresponding ground truth UC explanation, but there is neither correlation between $\textsc{UC}_1$/$\textsc{UC}_2$ and ground truth UC explanation of a different modality, nor correlation between $\textsc{UC}_1$/$\textsc{UC}_2$ and ground truth multimodal interaction explanations. This shows that \textsc{DIME}-generated UC explanations indeed capture unimodal contributions only. Moreover, we found that both \textsc{DIME}-generated multimodal interaction explanations ($\textsc{MI}_1, \textsc{MI}_2$) indeed correlate with the ground truth MI explanation, but not with either ground truth UC explanation. This shows that \textsc{DIME}-generated multimodal interaction explanation indeed captures explanations on just the multimodal interactions (i.e., the dot-product), and not any of the unimodal contributions. Meanwhile, running the original LIME on either modality just gives an explanation that weakly correlates with ground truth unimodal contributions and multimodal interactions, so the original LIME without disentangling is unable to give an accurate explanation of either unimodal contributions or multimodal interactions. In addition to using a synthetic dataset, we show that \textsc{DIME}\ can also disentangle more complex models on multimodal tasks, such as MDETR on CLEVR and LXMERT on VQA (the latter model is far from perfect in performance). As a measure of disentanglement, we check how \textsc{DIME}-generated explanations would be different given the same image but different questions. From each dataset, we randomly select $100$ points and generate their \textsc{DIME}\ explanations on the correct label. Then, for each point, we swap out the question with another different question on the same image and generate their \textsc{DIME}\ explanations on the same label (i.e., correct label before the swap). We compute cosine distance between the explanation weights from $\textsc{UC}_V$ before/after the swap, as well as cosine distance between the weights from $\textsc{MI}_V$ before/after the swap, and report average cosine distances on each dataset in Table~\ref{tab:relia}. We can see that swapping text has almost no effect on $\textsc{UC}_V$ but affects $\textsc{MI}_V$ significantly. Therefore, \textsc{DIME}\ is able to correctly disentangle a model into unimodal contributions and multimodal interaction for more complex models and tasks. \subsubsection{\textbf{RQ2:} Can \textsc{DIME}\ help researchers gain additional insight in whether unimodal contributions or multimodal interactions are the dominant factors behind a model's prediction?} \label{rq2} \ Disentangling the model into UC and MI and generating visualizations for each should provide additional insights into whether UC or MI is the main factor in the model's prediction. In the following experiments, we show that \textsc{DIME}\ can uncover which factor is dominant in a model's prediction process both across all points in the dataset (``global'') and on each individual datapoint (``local''). \textbf{Global interpretation:} CLEVR dataset is designed to force multimodal interactions, and MDETR has a $99.7\%$ accuracy on CLEVR, so we expect that MDETR will be heavily reliant on multimodal interactions. To verify this, we run \textsc{DIME}\ on MDETR for $100$ randomly sampled datapoints from the validation split of CLEVR, and compute the average absolute weight of the top-5 features in \textsc{DIME}\ explanations. As shown in Table~\ref{tab:weight}, the $\textsc{MI}_V$ and $MV_T$ weights are indeed significantly larger than $\textsc{UC}_V$ and $\textsc{UC}_T$ weights. Note that unimodal text does still give some useful information in CLEVR, such as the answer type (yes/no, attribute, or number), so that explains why $\textsc{UC}_T$ still has a weight of about $60\%$ that of $\textsc{MI}_T$. The average weight for $\textsc{MI}_V$, however, is over $4$ times higher than $\textsc{UC}_V$. Therefore, using \textsc{DIME}, we confirmed that MDETR indeed relies mostly on multimodal interactions to solve the task. \input{tables/weight} \textbf{Local interpretation:} In most datasets and models, models will not be near-perfect, and they will have different dominating factors from datapoint to datapoint. In this case, a global analysis will not suffice, and it will be necessary to look into which factor contributes more to the model's prediction on individual datapoints. We perform the following experiment to show that \textsc{DIME}\ can help users determine whether a model makes a prediction on a datapoint where (1) unimodal text is dominant, (2) unimodal image is dominant, (3) multimodal interactions are dominant, and (4) both UC and MI have significant contributions to the answer. We will use LXMERT on VQA since LXMERT is not close to perfect and often relies on different factors when predicting different datapoints. We gave five human annotators (who have some background knowledge in machine learning but do not have any knowledge about \textsc{DIME}) the same set of $52$ datapoints from VQA, as well as the prediction from LXMERT. For each datapoint, each human annotator is first given the LIME explanations without disentanglement as a baseline, and they are asked to categorize this point into one of the four categories above, while also rating how confident they are on their decision on a scale from one (least confident) to five (most confident). The human annotators are then presented with \textsc{DIME}\ explanations, and again they are asked to categorize each point as well as rate their confidence. The results are shown in Table~\ref{tab:newanno}. We can see that human annotators have significantly higher average confidence scores when presented with \textsc{DIME}\ explanations as compared to the baseline. Moreover, \textsc{DIME}\ result shows significantly higher Krippendorff's alpha score~\cite{krippendorff2011computing}, which measures inter-annotator agreements, so annotators also tend to agree a lot more on their categorizations. Therefore, \textsc{DIME}\ is able to help researchers more confidently determine whether UC or MI (or both) is the dominant factor behind the model's prediction, and thus help researchers gain additional insight into model behavior. \input{tables/anno} \subsubsection{\textbf{RQ3}: Can \textsc{DIME}\ help us better assess the qualities of the model and gain insights on how to debug or improve model performance?} \label{rq3} \ When trying to debug or improve a model on a task involving challenging reasoning, such as VQA, one important question researchers often ask is: do we know if our model actually learns to do the task ``the intended way'' (i.e., go through the same logical reasoning process as a human would to perform the task)? How often does our model perform as intended? Therefore, we conduct the following experiment to show that \textsc{DIME}\ may help answer this question. We use \textsc{DIME}\ explanations to categorize the model's behavior on each datapoint into one of the following categories: When the model answers correctly, \begin{itemize} \item (1) The model fully identifies the necessary parts of the image to answer the question logically through MI. \item (2) The model only partially identifies the parts of the image that are necessary to answer the question logically through MI and got it right with help of unimodal contributions. \item (3) The model did not correctly identify any of the parts of the image that are necessary to answer the question logically through MI. It got it right purely by unimodal contributions or by chance. \end{itemize} And when the model answers incorrectly, \begin{itemize} \item (4) The model fully identifies the necessary parts of the image to answer the question logically through MI, but still gets the answer wrong because the model does not fully understand a concept or because the question is too difficult (even for a human being). \item (5) The model only partially identifies the parts of the image that are necessary to answer the question logically through MI, thus missing some of the key parts of the image resulting in an incorrect answer. \item (6) The model did not correctly identify any of the parts of the image that are necessary to answer the question logically through MI, and thus the model fails to answer the question correctly. \end{itemize} In Figure~\ref{fig:veryinterestingexamples}, we show examples of datapoints, model predictions, and explanations that were annotated into each of the above categories. As shown in the examples, in most cases, there will be enough evidence to categorize a datapoint just by looking at the multimodal interaction explanations from the image side ($\textsc{MI}_V$), but sometimes other \textsc{DIME}\ explanations (e.g., explanations of text interactions) will be needed to gain additional understanding of the model's decision-making process. \input{tables/vqa} The results of this human study are shown in Table~\ref{tab:vqa}. With \textsc{DIME}, we were able to categorize $118$ points with evidence, out of a total of $140$ points $(84\%)$. This shows that \textsc{DIME}\ is able to highlight which input features are aligned or recognized by MI. We observe that, even though the models can fully identify the correct parts of the image that are relevant to the questions half of the time $(69/118)$, there is still a significant portion of datapoints where the model correctly aligns text and image but relies on unimodal contributions instead. This highlights several shortcomings of the model's decision-making process despite answering the question correctly. Therefore, the information gained from performing \textsc{DIME}\ can help researchers identify weaknesses in their models and debug or improve these models accordingly. \begin{figure*} \vspace{2mm} \includegraphics[width=1.0\textwidth]{figs/newfig6.pdf} \caption{Here we present examples of using \textsc{DIME}\ to categorize and explain why LXMERT makes certain predictions on datapoints in VQA 2.0. We present one example from each category. In most cases, only looking at the multimodal interaction explanations from the image side ($\textsc{MI}_V$) is sufficient to explain and categorize the model, but in certain cases, additional information from $\textsc{UC}_V$, $\textsc{UC}_T$, or $\textsc{MI}_T$ is needed as well. \textsc{DIME}\ enables researchers to gain understanding of the model's decision-making process which presents a step towards debugging and improving these models.} \label{fig:veryinterestingexamples} \end{figure*} We also observe that the model is more likely to not be able to fully identify the correct regions of the image when the model makes the wrong prediction, which is expected. In addition, we also found the following interesting observations when looking at the \textsc{DIME}\ explanations of the $118$ points: \begin{itemize} \item LXMERT often relies too heavily on unimodal text contributions: for example, in a question involving ``car'', unimodal contributions in text will prompt the model to answer ``street'' even if the model is unable to find ``street'' in the image. Sometimes, even when the model is able to interactively identify the correct regions of the image, unimodal text contributions can still dominate over the multimodal interaction (such as the fourth example in Figure~\ref{fig:veryinterestingexamples}, where the model answered ``glove'' due to unimodal text contributions even though the model was able to interactively identify the bat). \item The model sometimes interactively identifies the wrong object that happens to share the same properties in question as the correct object (such as the third example in Figure~\ref{fig:veryinterestingexamples}, where instead of the dog's paws, the model identified the nearby cat which also happens to be white). This coincidence happens more often than we expected, as there are $8$ such cases amongst the $118$ examples $(7\%)$. \item When asked about the color of an object that has two colors, LXMERT will only pick out one of the colors. \textsc{DIME}\ analysis shows that this is often due to LXMERT only identifying subregions of the object in one color while ignoring other parts of the object that are in a different color. For example, in Figure~\ref{fig:hydrant}, the model thinks that the hydrant is not ``silver and red'' because it did not classify the red tip as part of the hydrant. \end{itemize} These additional observations may guide future research in improving LXMERT (and other similar models) or designing inductive biases to avoid these undesirable behaviors. \begin{figure} \centering \vspace{2mm} \includegraphics[width=0.44\textwidth]{figs/hydrant.pdf} \caption{In this example, the model was unable to answer correctly because it did not recognize the red part in the image as part of the hydrant. As shown by the $\textsc{MI}_V$ explanation, the model actually thought that the red part is ``against'' the answer ``silver and red'', which means the model thought the red region isn't a part of the hydrant.} \label{fig:hydrant} \end{figure} \subsection{Limitations and Future Directions} \label{limit} Despite the ability of \textsc{DIME}\ in interpreting and debugging multimodal models, there remain several directions for future work: \textbf{1. Models with discrete outputs:} Even though \textsc{DIME}\ is designed to work for any black-box classification models, it requires the model to produce a continuous logit for each answer choice. \textsc{DIME}\ does not work well on the Neural-Symbolic VQA model~\cite{Mao2019NeuroSymbolic} since it only produces one discrete output instead of a continuous logit. Even when we tried to convert its outputs to logits by assigning its answer a logit of 1 and all other answer choices a logit of $-1$, \textsc{DIME}\ often fails to produce any meaningful explanation since the perturbations are unable to change the discrete answer of the model, thus having no effect on the assigned logits. \textbf{2. Number of modalities:} In all experiments, \textsc{DIME}\ was applied to tasks with 2 modalities. Disentangling a model across even 3 modalities can be very costly, as we will need to run the model $N^3$ times to compute unimodal contributions. Another challenge lies in interpreting the multimodal interaction, which would consist of bi-modal interactions between each pair of modalities as well as tri-modal interactions across all 3 modalities. Future work should tackle these challenges and try to expand \textsc{DIME}\ for high-modality scenarios. \textbf{3. Diverse modalities:} Even though the disentangling method in \textsc{DIME}\ theoretically works on any modality, our experiments have focused on image+text datasets (except the synthetic dataset experiment). This is because LIME-generated visualized explanations are relatively intuitive on image and text; it can be much harder for a human annotator to look at the results of explanations on other modalities (such as time-series of vectors) and try to make sense of them. In the future, we would like to design additional experiments to show that \textsc{DIME}\ can also be used to gain additional insight on model behavior in tasks involving modalities other than image and text as well. \textbf{4. Using these insights to improve models:} Since \textsc{DIME}\ is able to reveal several hidden undesirable behaviors in multimodal models, future work should aim to propose targeted solutions to these highlighted biases as a step towards improving multimodal models. For example, according to insights gained on VQA in RQ3, LXMERT can be improved by encouraging less reliance on unimodal text contribution, where insights from~\citet{cadene2019rubi} (which studies this research question for non-pretrained models) could be useful. Furthermore, future work could also design new training objectives which penalize models that associate wrong objects with words in MI, despite getting the correct answer. \section{Conclusion} In conclusion, \textsc{DIME}\ presents a new way to help users understand multimodal models by disentanglement into unimodal contributions and multimodal interactions before generating visual explanations for each. \textsc{DIME}\ can generate accurate disentangled explanations, help researchers and developers gain a deeper understanding of model behavior, and presents a step towards debugging and improving these models. We hope that \textsc{DIME}\ inspires the design of multimodal models that are more trustworthy, reliable, and robust for real-world applications. \section*{Acknowledgements} This material is based upon work partially supported by the National Science Foundation (Awards \#1722822 and \#1750439) and National Institutes of Health (Awards \#R01MH125740, \#R01MH096951, and \#U01MH116925). PPL is partially supported by a Facebook PhD Fellowship and a Carnegie Mellon University's Center for Machine Learning and Health Fellowship. RS is partially supported by NSF IIS1763562 and ONR Grant N000141812861. Any opinions, findings, conclusions, or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation, National Institutes of Health, Facebook, Carnegie Mellon University's Center for Machine Learning and Health, or Office of Naval Research, and no official endorsement should be inferred. We are extremely grateful to Gunjan Chhablani, Martin Ma, Chaitanya Ahuja, Volkan Cirik, Peter Wu, Amir Zadeh, Alex Wilf, Victoria Lin, Dong Won Lee, and Torsten W\"{o}rtwein for helpful discussions and feedback on initial versions of this paper. Finally, we would also like to acknowledge NVIDIA's GPU support.

ac3c520d9bd074005d320d07853fc7e9505ffd18

\section{Introduction} Conservation laws for differential equations play an important role both in the comprehension of the problem and in the analysis of the mathematical model. A conservation law relates the variation of a certain quantity within an arbitrarily small section of the space domain, to the amount of quantity that flows in and out. This quantity often has a physical meaning, such as mass, energy, momentum, electric charge. If the system is isolated, the total amount of the conserved quantity does not vary in the evolution of the process. Conservation laws further represent a fundamental tool to study the existence, uniqueness and stability of analytical solutions. Several mathematical techniques have been developed to construct conservation laws of classical partial differential equations (PDEs) involving integer order derivatives only, including methods based on Noether’s theorem, the direct method, the homotopy operator method, Ibragimov’s method \cite{AncoI,AncoII,Ibra,Olver}. In addition, a great effort has been made for the numerical preservation of conservation laws \cite{cdp16,frasca19,frasca21,Mc14,Wan16,Frasca-Caccia2021,CONTE2022} and to find structure preserving methods \cite{bras21,Conte2,Dahlby,Abdi17,pat14,pat18,pat21,hairer06,mc06,san94,Limbook,vandaele11}. In the last decade, conservation laws for fractional differential problems have been derived by suitably extending some of these known methods for PDEs. In particular, techniques that rely on generalizations of Noether’s theorem and variational Lie point symmetries have been applied to find conservation laws of fractional differential equations (FDE) with a fractional Lagrangian \cite{cw21,lhrh19}. For nonlinearly self-adjoint FDEs that do not have a Lagrangian in the classical sense, a \emph{formal} Lagrangian can be introduced and conservation laws are obtained by using modern techniques based on Lie group analysis of FDEs. This approach, proposed for the first time by Lukashchuk in 2015 \cite{Luk}, has been applied to time fractional PDEs \cite{Jafari20,lhd18,ahz21,cw21,habibi19,lhrh19,Luk} and more recently to time and space fractional PDEs (see \cite{Gupta17} and references therein). However, to the best of our knowledge, the specialized literature still misses a study on the numerical preservation of conservation laws of fractional differential problems. In the present paper, we consider a diffusion equation of fractional order in time, $\alpha$, and give sufficient conditions for identifying its conservation laws. Conservation laws of this equation with $0<\alpha<1$ and $1<\alpha<2$ have been obtained in \cite{Luk}. By exploiting the new result in this paper, we obtain a set of conservation laws for any value of $\alpha$. The main original result in this paper regards the numerical preservation of these conservation laws. In particular, we show that if a numerical method satisfies a discrete analogue of the sufficient conditions introduced in the continuous setting, then it has discrete conservation laws that approximate the continuous ones. In the integer case, $\alpha=1$, finite difference methods that preserve conservation laws have been introduced in \cite{IMA}. We propose a mixed method that combines a finite difference scheme along space with a spectral time integrator. With respect to other methods for FDEs known in literature (see e.g. \cite{ccp18,DeLuca20,DeLuca20b,milici2018,mohammadi19,moradi19,moradi20,moradi21,Pod99,vandaele19}), spectral methods present some advantages. In fact, most of these are step-by-step methods and thus they require at each time step a discretization of the long tail of the solution, arising from the hereditary nature of the fractional differential model. Thus, they are computationally expensive. Instead, spectral methods reflect the nonlocal nature of the fractional model and do not involve the discretization of the past history of the solution. Moreover, for a suitable choice of the function basis, spectral methods are exponentially convergent \cite{zk14}. Here we consider the spectral time integrator proposed in \cite{bcdp17} for time fractional PDEs of order $\alpha\in(0,1)$. This method is here suitably extended to be applied to FDEs of arbitrary fractional order, $\alpha\in(p-1,p)$, with $p\in\mathbb{N}$. Discrete conservation laws satisfied by the solutions of this method are derived, and some test examples are presented to highlight the conservation and convergence property of the new scheme. The rest of the paper is organized as follows. In Section~\ref{sec:materials}, we give some basic material and definitions of fractional differential calculus that are used in the rest of the paper. In Section~\ref{sec:contdisc}, we introduce sufficient conditions to have conservation laws in the continuous and discrete settings. Section~\ref{methsec} generalises the spectral time integrator in \cite{bcdp17} to approximate Riemann-Liouville or Caputo derivatives of arbitrary order. Considering a Riemann-Liouville derivative, in Section \ref{secmodel}, we apply this method to a finite difference discretization in space and derive its conservation laws. In Section \ref{sec:tests}, we verify on some numerical test examples the accuracy of the numerical method and its conservation laws. Finally, some conclusive remarks are drawn in Section \ref{sec:concl}. \section{Problem setting}\label{sec:materials} Let us consider a time fractional diffusion PDE of the form \begin{align}\label{FPDE} D_t^\alpha u-D_x^q K([u]_x)&=0, \end{align} where the symbol $[u]_x$ denotes the function $u$ and its integer derivatives in space and $$p-1<\alpha<p,\qquad p,q\in\mathbb{N},\qquad u=u(x,t), \qquad (x,t)\in(a,b)\times(t_0,T).$$ We assume that equation (\ref{FPDE}) is complemented by suitable Dirichlet boundary conditions, \begin{equation}\label{Dir} u(a,t)=\chi_a(t),\qquad u(b,t)=\chi_b(t), \end{equation} and $p$ initial conditions. Depending on the context, the symbol $D_t^\alpha$ denotes either the Riemann-Liouville fractional derivative, \begin{equation}\label{RLdef} ^{\text{RL}}D_t^\alpha f=D_t^p(I_t^{p-\alpha}f), \end{equation} or the Caputo fractional derivative, \begin{equation}\label{Cdef} ^{\text{C}}D_t^\alpha f= \left.I_t^{p-\alpha}(D_t^p(f)),\right. \end{equation} where $$I_t^{p-\alpha}f=\frac{1}{\Gamma (p-\alpha)}\int_{t_0}^t\frac{f(\tau,x)}{(t-\tau)^{1-p+\alpha}}\,\mathrm{d}\tau$$ is the Riemann-Liouville integral and $\Gamma(z)$ is the Gamma function. These two definitions of fractional derivative are related by \begin{equation}\label{relRLC} ^{\text{C}}D_t^\alpha f= \phantom{.}^{\text{RL}}D_t^\alpha f - \phantom{.}^{\text{RL}}D_t^\alpha \left(\sum_{k=0}^{p-1}\frac{(t-t_0)^k}{k!}f^{(k)}(t_0)\right). \end{equation} Thus, if the initial configuration is of total rest the two definitions are equivalent. For more details on the theory of fractional derivatives, we refer the reader to \cite{Pod99}. If the fractional derivative in (\ref{FPDE}) satisfies the Riemann-Liouville definition (\ref{RLdef}), the initial conditions assigned to (\ref{FPDE}) are \cite{Podlubnypaper} \begin{equation}\label{IC} D_t^{\alpha-k} u(x,t_0)=\gamma_k(x),\qquad k=0,\ldots,p-1. \end{equation} Instead, if the fractional derivative is of Caputo type (\ref{Cdef}), the initial conditions specify the initial values of the integer derivatives \begin{equation}\label{ICC} D_t^{k} u(x,t_0)=\gamma_k(x),\qquad k=0,\ldots,p-1. \end{equation} Conservation laws for time fractional diffusion problems have been object of several papers, e.g. \cite{cw21,habibi19,Luk,lhd18,lhrh19}. However, only papers \cite{Luk,lhd18} treat equation of type \eqref{FPDE} defined on a 3D and a 1D space, respectively. \section{Continuous and discrete con\-ser\-va\-tion laws}\label{sec:contdisc} \subsection*{Continuous setting} A conservation law of \eqref{FPDE} is a total divergence, \begin{equation}\label{CLaw} D_x(F(x,t,[u]_\alpha))+D_t(G(x,t,[u]_\alpha)) \end{equation} that vanishes on solutions of \eqref{FPDE}. Functions $F$ and $G$ are called the flux and the density of the conservation law (\ref{CLaw}), respectively. The symbol $[u]_\alpha$ denotes the function $u$, its fractional and integer derivatives and its fractional integrals. Differently from \cite{Luk}, in this paper we assume that $G$ depends on fractional integrals of order $p-\alpha$ only. In fact, integrals of higher order should be treated as new integral variables. Moreover, this is consistent with the limit case of $\alpha$ integer, where $F$ and $G$ are assumed to depend on $u$ and its partial derivatives but not on its integrals \cite{Olver}. When the boundary conditions are conservative (e.g., periodic) integration in space of (\ref{CLaw}) yields, $$D_t\int_a^b G(x,t,[u]_\alpha)\,\mathrm{d}x=0,$$ therefore, $$\int_a^b G(x,t,[u]_\alpha)\,\mathrm{d}x$$ is a global invariant of equation \eqref{FPDE}. However, the local conservation law (\ref{CLaw}) holds true regardless of the specific boundary conditions. The following theorem gives sufficient conditions to identify conservation laws of equation (\ref{FPDE}). \begin{theorem}\label{Theocont} If $\rho(t)$ and $\bar{G}=\bar G(x,t,[u]_\alpha)$ are two functions such that \begin{equation}\label{tclaw} \rho(t)D_t^\alpha u=D_t(\bar G(x,t,[u]_\alpha)), \end{equation} then the quantities \begin{equation}\label{xtclaw} x^k\rho(t)(D_t^\alpha u-D_x^q K([u])), \qquad k=0,\ldots q-1, \end{equation} are conservation laws of (\ref{FPDE}). \end{theorem} \begin{proof} The quantities in (\ref{xtclaw}) all vanish when $u$ is a solution of (\ref{FPDE}). Therefore, we only need to prove that these quantities can be written in the form (\ref{CLaw}). Taking into account that $\rho$ and $\bar{G}$ satisfy (\ref{tclaw}), we obtain, \begin{align*} x^k\rho(t)(D_t^\alpha u-D_x^q K([u]))=&\, D_t(x^k\bar G(x,t,[u]_\alpha))-x^kD_x^q(\rho(t)K([u])) \end{align*} that, integrating by parts, can be written as a total divergence (\ref{CLaw}) with \begin{equation}\label{exFG} F=\sum_{\ell=0}^k (-1)^{\ell+1}\frac{k!}{(k-\ell)!}x^{k-\ell}D_x^{q-1-\ell}(\rho(t)K([u])),\quad G=x^k\bar G, \end{equation} and $k=0,\ldots,q-1$. \end{proof} \begin{corollary}\label{coroll} Equation (\ref{FPDE}) with $D_t^\alpha=\,^{\text{RL}}D_t^\alpha$ has at least $p\cdot q$ conservation laws given by (\ref{CLaw}) where $F$ and $G$ are defined by (\ref{exFG}) with $$\rho(t)=t^j,\qquad \bar G=\sum_{i=0}^{j}(-1)^i\frac{j!}{(j-i)!}t^{j-i}D_t^{p-i-1}(I_t^{p-\alpha}u),\qquad j=0,\ldots p-1. $$ \end{corollary} \begin{proof} It follows from the definition of Riemann-Liouville fractional derivative \eqref{RLdef} that equation (\ref{FPDE}) is itself a conservation law. In fact, it can be written in the form (\ref{CLaw}) with \begin{equation}\label{eqclaw} F=-D_x^{q-1}K(u),\qquad G=D_t^{p-1}(I_t^{p-\alpha}u)=\left\{\begin{array}{cc}I_t^{1-\alpha}u,&\text{if}\quad p=1,\\ ^{RL}D_t^{\alpha-1}u,&\text{if}\quad p>1.\end{array}\right. \end{equation} As a consequence, Theorem~\ref{Theocont} holds true with $\rho(t)=1$ and $\bar G=G$. Hence, equation (\ref{FPDE}) has at least $q$ conservation laws given in (\ref{xtclaw}). If $p=1$, the statement is proved. If $p>1$, \begin{equation} \rho(t)D_t^\alpha u:= t^jD_t^\alpha u=t^jD_t^{p}(I_t^{p-\alpha}u)=D_t(\bar G),\qquad j=1,\ldots, p-1, \end{equation} where the last equality is obtained after integrating by parts $j$ times with \begin{equation}\label{Gbar} \bar G=\sum_{i=0}^{j}(-1)^i\frac{j!}{(j-i)!}t^{j-i}D_t^{p-i-1}(I_t^{p-\alpha}u), \qquad j=1,\ldots,p-1. \end{equation} Therefore, it follows from Theorem~\ref{Theocont} that for each $j=0,\ldots,p-1,$ there are $q$ conservation laws with flux and densities defined by \begin{equation}\label{rhot} F=\sum_{\ell=0}^k (-1)^{\ell+1}\frac{k!}{(k-\ell)!}x^{k-\ell}D_x^{q-1-\ell}(t^j K([u])),\quad G=x^k\bar G,\quad k=0,1,\ldots,q-1, \end{equation} with $\bar {G}$ given in (\ref{Gbar}), and so a total of at least $p\cdot q$ conservation laws. \end{proof} \begin{remark} The function $\bar G$ in (\ref{Gbar}) can be equivalently written as \begin{align*} &\bar G=\sum_{i=0}^{j}(-1)^i\frac{j!}{(j-i)!}t^{j-i}D_t^{\alpha-i-1}u,& \text{\,\,if\,\,}j<p-1,\\ &\bar G=\sum_{i=0}^{p-2}(-1)^i\frac{(p-1)!}{(p-i-1)!}t^{p-i-1}D_t^{\alpha-i-1}u+(-1)^{p-1}{(p-1)!}I_t^{p-\alpha}u,& \text{\,\,if\,\,}j=p-1. \end{align*} \end{remark} \begin{remark} We observe that for $\alpha=p=1$ and $q=2$, (\ref{tclaw}) and (\ref{xtclaw}) with $\rho(t)=1$ are the two conservation laws given in \cite{Ibra}. \end{remark} \subsection*{Discrete setting} In order to define a numerical approximation of (\ref{FPDE}) we define a uniform spatial grid with nodes, \begin{equation}\label{spgrid} x_i=a+i\Delta x,\qquad i=0,\ldots,M+1,\qquad \Delta x=\frac{b-a}{M+1}. \end{equation} Considering that at the endpoints the solution is known from the boundary conditions (\ref{Dir}), we define the vector of the approximations \begin{equation}\label{Uvec} \mathbf{u}=\mathbf{u}(t)\in\mathbb{R}^{M},\qquad \mathbf{u}_i(t)\simeq u(x_i,t),\quad i=1,\ldots,M. \end{equation} We denote with $D_{\Delta x}$ the forward difference operators in space and with $D^{(q)}_{\Delta x}$ the second-order centred difference operator for the $q$-th derivative. We consider here semidiscretizations of the form \begin{equation}\label{SDeq} D_t^\alpha \mathbf{u}-D_{\Delta x}^{(q)}\widetilde{K}(\mathbf{u})=0, \end{equation} where $\widetilde{K}\approx K$ is here arbitrary, and it can be defined in such a way to obtain accuracy in space of arbitrary order. In fact, high-order finite difference approximations of the $q$-th derivative are defined on larger stencils and are obtained combining $D_{\Delta x}^{(q)}$ with suitable averaging operators (see e.g. formulae in \cite{Pier}). Let be \begin{equation}\label{tjnodes} t_0<t_1<\ldots<t_{N-1}<t_N=T, \quad t_{j+1}=t_j+\Delta t_j, \quad j=0,\ldots,N-1, \end{equation} the nodes in time and $D_{\Delta t_j}$ the forward difference operator with step $\Delta t_j$. For simplicity of notation, henceforth we omit the subscript $j$ in the time difference operator. We denote with $u_{i,j}$ and $D_{\Delta t}^\alpha u_{i,j}$ the approximations of $u(x_i,t_j)$ and of $D_{t}^\alpha u(x_i,t_j)$, respectively, obtained after applying a suitable time integrator to (\ref{SDeq}). Hence, the fully discrete scheme for (\ref{FPDE}) is \begin{equation}\label{spect} D_{\Delta t}^\alpha u_{i,j}-D_{\Delta x}^{(q)} \widetilde{K}(u_{i,j})=0. \end{equation} Theorem~\ref{theodisc} will give sufficient conditions that method (\ref{spect}) has to satisfy to have discrete conservation laws in the form $$D_{\Delta x}\widetilde{F}(x_i,t_j,u_{i,j})+D_{\Delta t}\widetilde{G}(x_i,t_j,u_{i,j})=0,$$ where $\widetilde{F}$ and $\widetilde{G}$ are suitable discretizations of the flux and density of a selected continuous conservation law, respectively. This result is a discrete version of Theorem~\ref{Theocont}, and the main contribution in this section. For its proof, we recur to the following lemma that can be proved by straightforward calculations. \begin{Lemma}\label{leibniz} For any two discrete functions $f$ and $g$, the following discrete versions of Leibniz rule hold true: \begin{align}\label{Leibdisc1} -f_{i,j}D_{\Delta x}^{(1)}g_{i,j}=&\,g_{i,j}D_{\Delta x}^{(1)}f_{i,j}+D_{\Delta x}(-\tfrac{1}2(f_{i-1,j}g_{i,j}+f_{i,j}g_{i-1,j})),\\\label{Leibdisc2} -f_{i,j}D_{\Delta x}^{(2)}g_{i,j}=&\,-g_{i,j}D_{\Delta x}^{(2)}f_{i,j}+D_{\Delta x}(\tfrac{1}{\Delta x}(f_{i,j}g_{i-1,j}-f_{i-1,j}g_{i,j})). \end{align} Moreover, for $q>2$, \begin{equation*} -f_{i,j}D_{\Delta x}^{(q)}g_{i,j}=-(-1)^q g_{i,j}D_{\Delta x}^{(q)}f_{i,j}+D_{\Delta x}(\widetilde{F}),\ \end{equation*} where the function $\widetilde F$ is obtained by iterating (\ref{Leibdisc2}) $\lambda$ times, if $q=2\lambda$, or (\ref{Leibdisc2}) $\lambda$ times and (\ref{Leibdisc1}) once, if $q=2\lambda+1$. \end{Lemma} \begin{theorem}\label{theodisc} For all $\rho(t_j)$ and $\bar{G}=\bar{G}(x_i,t_j,u_{i,j})$ such that \begin{equation}\label{disctclaw} \rho(t_j)D_{\Delta t}^\alpha u_{i,j}=D_{\Delta t}(\bar{G}(x_i,t_j,u_{i,j})), \end{equation} the quantities \begin{equation}\label{discxtclaw} x_i^k\rho(t_j)(D_{\Delta t}^\alpha u_{i,j}-D_{\Delta x}^{(q)} \widetilde K(u_{i,j})), \qquad k=0,\ldots q-1, \end{equation} are conservation laws of (\ref{spect}) at the point $(x_i,t_j)$ that approximate their continuous counterparts with the same accuracy of the method. \end{theorem} \begin{proof} The proof follows along similar lines as that of Theorem~\ref{Theocont}. Multiplying method (\ref{spect}) by $x_i^k\rho(t_j)$, with $k=0,1,\ldots,q-1$, yields \begin{equation}\label{discClaws} x_i^k\rho(t_j)(D_{\Delta t}^\alpha u_{i,j}-D_{\Delta x}^{(q)}\widetilde{K}(u_{i,j})). \end{equation} These quantities clearly vanish on solutions of (\ref{spect}). Moreover, considering (\ref{disctclaw}) and Lemma~\ref{leibniz}, \begin{align*} x_i^k\rho(t_j)&(D_{\Delta t}^\alpha u_{i,j}\,-D_{\Delta x}^{(q)} \widetilde{K}(u_{i,j}))= D_{\Delta t}(x_i^k\bar{G}(x_i,t_j,u_{i,j}))-x_i^kD_{\Delta x}^{(q)}(\rho(t_j)\widetilde{K}(u_{i,j}))\\ =&\,D_{\Delta t}(x_i^k\bar{G}(x_i,t_j,u_{i,j}))-((-1)^q D_{\Delta x}^{(q)}x_i^k)\rho(t_j)\widetilde K(u_{i,j})+D_{\Delta x}(\widetilde F(x_i,t_j,u_{i,j})). \end{align*} Then, as the spatial grid is uniform and $k<q$, $$x_i^k\rho(t_j)(D_{\Delta t}^\alpha u_{i,j}-D_{\Delta x}^{(q)} \widetilde K(u_{i,j}))=D_{\Delta t}(\widetilde G)+D_{\Delta x}(\widetilde F),$$ where $\widetilde G=x_i^k\bar{G}$ and $\widetilde F \approx F$ in (\ref{exFG}). As the function $x_i^k\rho(t_j)$ is exactly evaluated at the nodes, a conservation law in the form (\ref{discxtclaw}) approximates its continuous limit (\ref{xtclaw}) with the same accuracy of the scheme. \end{proof} \section{The time integrator}\label{methsec} Given the space discretization \eqref{SDeq}, we perform the time integration by using the spectral method introduced in \cite{bcdp17} for fractional problems of order $\alpha$ with $0<\alpha<1$. This method is here generalized to deal with equations of arbitrary fractional order. We separate the treatment of equations with fractional derivative satisfying Riemann-Liouville or Caputo definition with a focus on the particular case of zero initial conditions. \subsection{Riemann-Liouville fractional derivative}\label{sec:RL} When the fractional derivative $D_t^\alpha$ in (\ref{FPDE}) satisfies the Riemann-Liouville definition (\ref{RLdef}), we look for time approximations of the solution \eqref{Uvec} and of the initial conditions (\ref{IC}) at the node $x_i$, in the form \begin{align}\label{modalexp} u_N^i(t)=&\,\sum_{j=0}^{N+p} \hat u_j^i\mathcal{P}_j(t),\\ \label{discIC} D_t^{\alpha-k} u_N^i(t_0)=&\,\sum_{j=0}^{N+p} \hat u_j^iD_t^{\alpha-k}\mathcal{P}_j(t_0)=\gamma_k(x_i),\qquad k=0,\ldots,p-1, \end{align} respectively, where $\{\mathcal P_j(t)\}_{j=0}^{N+p}$ is a suitable functional basis and $\hat u_j^i$ are unknown coefficients. By considering the set of collocation points (\ref{tjnodes}), we can equivalently write equation (\ref{modalexp}) as \begin{equation}\label{nodalexp} u_N^i(t)=\sum_{k=0}^{N} \varphi_k(t)u_N^i(t_k)+\sum_{k=0}^{p-1}\varphi_{N+k+1}(t)\gamma_{k}(x_i), \end{equation} where the functions $\varphi_k(t)$ are unknown. By defining $$\psi_k(t)=D_t^\alpha\varphi_k(t),\qquad k=0,\ldots,N+p,$$ expansion (\ref{nodalexp}) gives the following approximation of the time fractional derivative of $\mathbf{u}_i$ defined in \eqref{Uvec} at the collocation points: \begin{equation}\label{fullapp} D_t^\alpha \mathbf{u}_i(t_j)\approx D_t^\alpha u_N^i(t_j)=\sum_{k=0}^{N}\psi_k(t_j)u_N^i(t_k)+\sum_{k=0}^{p-1}\psi_{N+k+1}(t_j)\gamma_k(x_i)=: D_{\Delta t}^\alpha u_{i,j}. \end{equation} In order to determine the functions $\psi_k(t)$, and so to practically compute approximation (\ref{fullapp}), we define the following two vectors in $\mathbb{R}^{N+p+1}$, \begin{align*} \mathbf{u}_N^i&\,=\left(u_N^i(t_0),\ldots,u_N^i(t_N),\gamma_0(x_i),\ldots,\gamma_{p-1}(x_i)\right)^T,\\ \hat\mathbf{u}^i&\,=\left(\hat u^i_0,\ldots,\hat u_{N+p}^i\right)^T, \end{align*} the matrices of dimension $N+p+1$, \begin{equation}\label{matA} A=\left(\begin{array}{c} A_1\\A_2\end{array}\right),\qquad B=A^{-1}, \end{equation} with \begin{equation*} {A}_1=\left(\begin{array}{ccc} \mathcal{P}_0(t_0) & \cdots & \mathcal{P}_{N+p}(t_0)\\ \vdots & & \vdots\\ \mathcal{P}_0(t_N) & \cdots & \mathcal{P}_{N+p}(t_N) \end{array}\right), \ {A}_2=\left(\begin{array}{ccc} D_t^\alpha \mathcal{P}_0(t_0) & \cdots & D_t^{\alpha} \mathcal{P}_{N+p}(t_0)\\ \vdots & & \vdots\\ D_t^{\alpha-p+1} \mathcal{P}_0(t_0) & \cdots & D_t^{\alpha-p+1} \mathcal{P}_{N+p}(t_0)\\ \end{array}\right). \end{equation*} and the matrices \begin{align}\label{matC} C=\left(\begin{array}{ccc} \psi_0(t_0) & \cdots & \psi_{N+p}(t_0)\\ \vdots & & \vdots\\ \psi_0(t_N) & \cdots & \psi_{N+p}(t_N) \end{array}\right), \ \mathcal{P}=\left(\begin{array}{ccc} D_t^\alpha\mathcal{P}_0(t_0) & \cdots & D_t^\alpha\mathcal{P}_{N+p}(t_0)\\ \vdots & & \vdots\\ D_t^\alpha\mathcal{P}_0(t_N) & \cdots & D_t^\alpha\mathcal{P}_{N+p}(t_N) \end{array}\right). \end{align} \begin{prop} Matrix C can be computed as \begin{equation}\label{matsys} C=\mathcal PB. \end{equation} \end{prop} \begin{proof} By evaluating (\ref{modalexp}) at the nodes $t_j$, with $j=0,\ldots,N$, we obtain a set of $N+1$ equations that together with (\ref{discIC}) forms an algebraic system that can be equivalently written as $$\mathbf{u}_N^i=A\hat\mathbf{u}^i,$$ therefore, $$\hat\mathbf{u}^i=B\mathbf{u}_N^i,$$ or, entry-wise, $$\hat u^i_j=\sum_{k=0}^N B_{j+1,k+1}u_N^i(t_k)+\sum_{k=0}^{p-1}B_{j+1,N+k+2}\gamma_{k}(x_i),\qquad j=0,\ldots,N+p.$$ Substituting in (\ref{modalexp}), \begin{align}\nonumber u_N^i(t)=&\,\sum_{j=0}^{N+p} \mathcal{P}_j(t)\left(\sum_{k=0}^N B_{j+1,k+1}u_N^i(t_k)+\sum_{k=0}^{p-1}B_{j+1,N+k+2}\gamma_k(x_i)\right)\\\label{tophi} =&\,\sum_{k=0}^N\left(\sum_{j=0}^{N+p} B_{j+1,k+1}\mathcal{P}_j(t) \right)u_N^i(t_k)+\sum_{k=0}^{p-1}\left(\sum_{j=0}^{N+p}B_{j+1,N+k+2}\mathcal{P}_j(t) \right)\gamma_{k}(x_i). \end{align} Comparing (\ref{nodalexp}) and (\ref{tophi}), we find that $$\varphi_k(t)=\sum_{j=0}^{N+p} B_{j+1,k+1}\mathcal{P}_j(t), \qquad k=0,\ldots, N+p,$$ and, differentiating, \begin{equation}\label{psik} \psi_k(t)=D_t^\alpha \varphi_k(t)=\sum_{j=0}^{N+p} B_{j+1,k+1}D_t^\alpha \mathcal{P}_j(t), \qquad k=0,\ldots, N+p. \end{equation} The values $\psi_k(t_j)$ are then obtained as \begin{equation}\label{psiktj} \psi_k(t_j)=\sum_{\ell=0}^{N+p} B_{\ell+1,k+1}D_t^\alpha \mathcal{P}_\ell(t_j), \quad k=0,\ldots, N+p, \ j=0,\ldots,N. \end{equation} Equation \eqref{matsys} follows immediately. \end{proof} Note that in order to compute $\psi_k(t_j)$ in (\ref{psiktj}) we only need the values of the chosen basis $\{\mathcal P_\ell(t)\}_{\ell=0}^{N+p}$ and its fractional derivatives at the collocation points. Therefore, the basis should be chosen such that it is easy to compute these fractional derivatives. The fully discrete scheme for (\ref{FPDE}) is in the form (\ref{spect}) with the approximation of the fractional derivative (\ref{fullapp}) and $$u_{i,j}:=u_N^i(t_j).$$ Let us consider matrix $C$ defined in (\ref{matC}) and let be $C=[C_1,C_2]$, with $C_1$ a square matrix of dimension $N+1$ and $C_2$ of dimension $ (N+1)\times p$. The numerical method can be written in matrix form as \begin{equation}\label{matrixform} C_1U=\frac{1}{\Delta x^2}(\widetilde{K}(U)\mathcal{M}+\mathcal{F})-C_2\gamma, \end{equation} where the entries of $U\in\mathbb{R}^{(N+1)\times M}$ and $\gamma \in\mathbb{R}^{p\times M}$ are \begin{align*} &U_{k,i}=u_{i,k-1}, \quad \gamma_{\ell,i}=\gamma_{\ell-1}(x_i),\quad k=1,\ldots,N+1, \ i=1,\ldots,M, \ \ell=1,\ldots p, \end{align*} respectively, matrix $\widetilde{K}(U)$ is obtained by applying $\widetilde{K}$ to the entries of matrix $U$, and \begin{equation*} \mathcal{M}=\left(\begin{array}{ccccc} -2 & 1 & && \\ 1 & -2 & 1 &&\\ &\ddots & \ddots& \ddots &\\ &&1 & -2 & 1 \\ && & 1 &-2\end{array}\right),\quad \mathcal{F}=\left(\begin{array}{ccccc} \widetilde{K}(\chi_a(t_0)) & 0 &\ldots & 0 & \widetilde{K}(\chi_b(t_0))\\ \vdots & \vdots & &\vdots & \vdots\\ \widetilde{K}(\chi_a(t_N)) & 0 &\ldots & 0 & \widetilde{K}(\chi_b(t_N))\end{array}\right), \end{equation*} are matrices in $\mathbb{R}^{M\times M}$ and $\mathbb{R}^{(N+1)\times M},$ respectively. \begin{remark}\label{remr} In particular cases, the basis $\{\mathcal P_j(t)\}_j$ can be chosen such that $u_N^i$ satisfies $r\leq p$ initial conditions (\ref{IC}) by definition and do not need to be enforced. In these cases, approximation (\ref{modalexp}) is taken in a projection space of dimension $N+p+1-r$. Similarly, approximation (\ref{nodalexp}) is replaced with $$u_N^i(t)=\sum_{k=0}^{N} \varphi_k(t)u_N^i(t_k)+\sum_{k=0}^{p-1-r}\varphi_{N+k+1}(t)\gamma_{k}(x_i),$$ where (after a suitable reordering of the indexes) the second sum includes only the values $\gamma_k$ of the initial conditions to be imposed. Matrices $A$ and $B$ in \eqref{matA} have then reduced dimension $N+p+1-r$, and $\mathcal{P}$ in \eqref{matC} has dimension $(N+1)\times (N+p+1-r)$. System \eqref{matrixform} has still dimension $(N+1)\times M$ but matrices $C_2$ and $\gamma$ have dimension $(N+1)\times (p-r) $ and $(p-r)\times M$, respectively. \end{remark} \subsection{Caputo fractional derivative}\label{sec:Cap} We now adapt the method introduced in Section~\ref{sec:RL} to problems in the form \eqref{FPDE} with fractional derivative satisfying Caputo's definition \eqref{Cdef}. In this case the initial conditions are given by \eqref{ICC}. In particular, the initial value of $u$ at $t=t_0$ is known, and in order to obtain a compatible system the approximated solution must be taken in a space of reduced dimension $N+p$. Therefore, equations \eqref{modalexp} and \eqref{discIC} are replaced with \begin{align}\label{modalexpC} u_N^i(t)=&\,\sum_{j=0}^{N+p-1} \hat u_j^i\mathcal{P}_j(t),\\ \label{discICC} D_t^{k} u_N^i(t_0)=&\,\sum_{j=0}^{N+p-1} \hat u_j^iD_t^{k}\mathcal{P}_j(t_0)=\gamma_k(x_i),\qquad k=0,\ldots,p-1, \end{align} respectively. With collocation points, $$t_1<\ldots<t_N=T,\qquad t_{j+1}=t_j+\Delta t_j, \qquad j=1,\ldots,N-1,$$ equation \eqref{modalexpC} is equivalently written as \begin{equation}\label{nodalexpC} u_N^i(t)=\sum_{k=1}^{N} \varphi_k(t)u_N^i(t_k)+\sum_{k=0}^{p-1}\varphi_{N+k+1}(t)\gamma_{k}(x_i), \end{equation} where functions $\varphi_k$ are unknown. Defining $$\psi_k(t)=D_t^\alpha\varphi_k(t),\qquad k=1,\ldots,N+p,$$ the approximation of the fractional derivative is then given by \begin{equation}\label{fullappC} D_t^\alpha \mathbf{u}_i(t_j)\approx D_t^\alpha u_N^i(t_j)=\sum_{k=1}^{N}\psi_k(t_j)u_N^i(t_k)+\sum_{k=0}^{p-1}\psi_{N+k+1}(t_j)\gamma_k(x_i)=: D_{\Delta t}^\alpha u_{i,j}. \end{equation} Following similar steps as in Section \ref{sec:RL}, the values $$\psi_k(t_j),\qquad k=1,\ldots, N+p, \qquad j=1,\ldots,N,$$ can be obtained by solving \eqref{matsys} where matrices $C$ and $\mathcal{P}$ have now dimension $N\times (N+p)$ and are defined as their analogues in \eqref{matC} by deleting the first row and the first column and the first row and the last column, respectively. Matrix $B$ has dimension $(N+p)\times (N+p)$ and is defined as in \eqref{matA} after removing from $A$ the first row and the last column. By splitting matrix $C$ as $C=[C_1,C_2]$ where $C_1$ and $C_2$ have dimension $N\times N$ and $N\times p$, respectively, and defining $U\in\mathbb{R}^{N\times M}$ with entries $U_{k,i}=u_N^i(t_k),$ $k=1,\ldots,N$, $i=1,\ldots,M$, the numerical method can be written in matrix form as \begin{equation}\label{eq:Cmatform} C_1 U=\frac{1}{\Delta x^2}(\widetilde{K}(U)\mathcal{M}+\mathcal{F})-C_2\gamma, \end{equation} where $\mathcal{M}$, $\gamma$ and $\widetilde{K}(U)$ are defined as in \eqref{matrixform} and \begin{equation*} \mathcal{F}=\left(\begin{array}{ccccc} \widetilde{K}(\chi_a(t_1)) & 0 &\ldots & 0 & \widetilde{K}(\chi_b(t_1))\\ \vdots & \vdots & &\vdots & \vdots\\ \widetilde{K}(\chi_a(t_N)) & 0 &\ldots & 0 & \widetilde{K}(\chi_b(t_N))\end{array}\right)\in\mathbb{R}^{N\times M}. \end{equation*} We observe that the dimension of system \eqref{eq:Cmatform} is lower than that of system \eqref{matrixform}. \begin{remark}\label{rem:cap} If the basis $\{\mathcal P_j(t)\}_j$ is chosen such that $r\leq p$ initial conditions (\ref{IC}) are satisfied by $u_N^i$ an analogue discussion as in Remark~\ref{remr} holds considering a projection space of dimension $N+p-r$. \end{remark} \subsubsection{The case of zero initial conditions} When the initial condition is of total rest, i.e., \begin{equation}\label{IC2} D_t^k u(x,t_0)=0,\qquad k=0,\ldots p-1, \end{equation} the definitions of Caputo and Riemann-Liouville fractional derivative are equivalent. Since in this special case $\gamma_k(x_i)=0$, approximation (\ref{fullappC}) reduces to \begin{equation*} D_t^\alpha \mathbf{u}_i(t_j)\approx D_t^\alpha u_N^i(t_j)=\sum_{k=1}^{N}\psi_k(t_j)u_N^i(t_k)=: D_{\Delta t}^\alpha u_{i,j}, \end{equation*} and so only the values of $\psi_k(t_j)$ for $k=1\ldots,N,$ and $j=1\ldots,N$ need to be calculated. The matrix form of the numerical method \eqref{eq:Cmatform} reduces to \begin{equation}\label{eq:matform0} C_1 U=\frac{1}{\Delta x^2}(\widetilde{K}(U)\mathcal{M}+\mathcal{F}), \end{equation} and so it is not necessary to calculate matrix $C_2$. Matrix $C_1$ is obtained from $$C_1={\mathcal P} \bar B,$$ where $\bar B$ is obtained removing from matrix $B$ the last $p$ columns. Note that although the computation of the inverse of matrix $A$, having dimension $N+p$, is still required, system \eqref{eq:matform0} has reduced dimension, $N$. \begin{remark} The basis \begin{equation}\label{eq:powerb} \{\mathcal{P}_j(t)\}=\{(t-t_0)^{j\alpha}\}, \end{equation} identically satisfies all the initial conditions (\ref{IC2}) for $k=1,\ldots,p-1,$ and so Remark \ref{rem:cap} applies. In this case, then, matrix $A$ has dimension $N+1$. \end{remark} \section{Conservation laws of the fractional diffusion equation}\label{secmodel} Here and henceforth we consider equation (\ref{FPDE}) with $q=2$ and Riemann-Liouville fractional derivative. Although, as seen in Section~\ref{methsec}, the general discussion holds for methods of arbitrary order in space, in order to give some specific results and explicit formulae of the preserved conservation laws, we focus here on second-order accurate schemes. Therefore, henceforth we set $\widetilde{K}=K$ and method (\ref{spect}) reduces to \begin{equation}\label{spectK} D_{\Delta t}^\alpha u_{i,j}-D_{\Delta x}^{(2)} {K}(u_{i,j})=0. \end{equation} As stated in Corollary~\ref{coroll}, when the fractional derivative is defined according to the definition of Riemann-Liouville there are always at least two conservation laws. The first is equivalent to equation (\ref{FPDE}) and it is defined by (see \eqref{eqclaw}) \begin{equation}\label{SDclaw1} F_1(x,t,u)=-D_x K(u),\qquad G_1(x,t,u)=D_t^{p-1}I_t^{p-\alpha}u. \end{equation} The second is defined by (\ref{rhot}) with $k=1$ and $\bar{G}=G_1$, yielding, \begin{equation}\label{SDclaw2} F_2(x,t,u)= K(u)-xD_xK(u),\qquad G_2(x,t,u)=x G_1. \end{equation} We prove that method (\ref{spectK}) preserves these conservation laws for any value of $\alpha$ and $p$, and give their conserved approximations. According to Theorem~\ref{theodisc}, it suffices to prove (\ref{disctclaw}), i.e. to find $\widetilde{G}_1\approx G_1$ such that \begin{equation}\label{cond} D_{\Delta t}^\alpha u_{i,j}=D_{\Delta t}(\widetilde{G}_1(x_i,t_j,u_{i,j})). \end{equation} By integrating both sides in (\ref{RLdef}), we have $$G_1(x,t,u)=D_t^{p-1}I_{t}^{p-\alpha}u(x,t)=\int_{t_0}^{t}D_\tau^\alpha u(x,\tau)\,\mathrm{d}\tau.$$ We define then \begin{equation}\label{appG1} \widetilde G_1(x_i,t_{j},u_{i,j})=\Delta t \sum_{\ell=0}^{j-1} D_{\Delta t}^{\alpha}u_{i,\ell},\end{equation} so that (\ref{cond}) is satisfied. Following the steps in the proof of Theorem~\ref{theodisc}, the remaining functions that define the two discrete conservation laws are \begin{equation}\label{discF1} \widetilde{F}_1(x_i,t_{j},u_{i,j})=-D_{\Delta x}{K}(u_{i-1,j}), \end{equation} and (see (\ref{Leibdisc2})), \begin{align}\nonumber \widetilde{F}_2(x_i,t_{j},u_{i,j})=& \displaystyle \,\tfrac{1}{\Delta x}(x_iK(u_{i-1,j})-x_{i-1}K(u_{i,j}))\\ \nonumber =& A_{\Delta x}{K}(u_{i-1,j})\-A_{\Delta x}(x_{i-1})D_{\Delta x}K(u_{i-1,j}),\\ \label{discF2G2}\widetilde{G}_2(x_i,t_{j},u_{i,j})=&\,x_i\widetilde{G}_1(x_i,t_{j},u_{i,j}), \end{align} respectively, where $A_{\Delta x}$ denotes the forward average operator, $$A_{\Delta x}(f(u_{i,j}))=\frac{f(u_{i+1,j})+f(u_{i,j})}2=f(u(x_i+\tfrac{\Delta x}2,t_j))+\mathcal{O}(\Delta x^2).$$ In the rest of this section, we focus on the two cases of subdiffusion and superdiffusion equations obtained setting $p=1$ and $p=2$, respectively. The continuous and discrete conservation laws obtained are listed in Table~\ref{allcl}. \subsubsection*{Subdiffusion-wave equation} When $0<\alpha<1=p$ equation \eqref{FPDE} defines a subdiffusion problem. Using Theorem~\ref{Theocont} and Corollary~\ref{coroll} we can obtain only the two conservation laws defined by (\ref{SDclaw1}) and (\ref{SDclaw2}). Indeed, these two are the only independent conservation laws of the subdiffusion-wave equation in the generic form (\ref{FPDE}) with Riemann-Liouville fractional derivative \cite{Luk}. As shown, method (\ref{spectK}) has discrete analogues of this conservation laws defined by \eqref{appG1}--\eqref{discF2G2}, for any $p$. Some extra conservation laws are given in \cite{Luk} for special choices of the function $K$ in (\ref{FPDE}), but these depend on integrals whose order is larger than $p-\alpha$ that is a case that we do not consider in this paper. \subsubsection*{Superdiffusion equation} Let us consider now the superdiffusion equation defined by (\ref{FPDE}) with $1<\alpha<2=p$. In this case, according to Corollary \ref{coroll}, conservation laws are obtained from (\ref{xtclaw}) with $$\rho(t)=1\qquad \rho(t)=t,$$ yielding four independent conservation laws. The first two are again (\ref{SDclaw1}) and (\ref{SDclaw2}) and these are preserved by method (\ref{spectK}). The other two conservation laws are defined with (see \eqref{Gbar}--(\ref{rhot})), \begin{equation}\label{FG3} F_3(x,t,u)=-tD_xK(u),\qquad G_3(x,t,u)=tD_t^{\alpha-1}u-I_t^{2-\alpha}u, \end{equation} and \begin{equation}\label{FG4} F_4(x,t,u)=tK(u)-xtD_xK(u),\qquad G_4(x,t,u)=xtD_t^{\alpha-1}u-xI_t^{2-\alpha}u. \end{equation} The density function in (\ref{FG3}) can be equivalently written as, $$G_3(x,t,u)=tG_1(x,t,u)-\int_{t_0}^tG_1(x,z,u)\,\mathrm{d}z.$$ In fact, for $p=2$, $$D_t^{\alpha-1}u=D_tI_t^{2-\alpha}u=G_1(x,t,u),$$ and integrating twice the Riemann-Liouville fractional derivative (\ref{RLdef}) yields, $$I_t^{2-\alpha}u=\int_{t_0}^t\int_{t_0}^z D_\tau^\alpha u(x,\tau)\,\mathrm{d}\tau\,\mathrm{d}z=\int_{t_0}^tG_1(x,z,u)\,\mathrm{d}z.$$ We show that (\ref{disctclaw}) is satisfied by the solutions of (\ref{spectK}) with \begin{equation}\label{discG3} \rho(t_j)=t_j,\quad \bar{G}=\widetilde{G}_3(x_i,t_j,u_{i,j})=t_j\widetilde{G}_1(x_i,t_j,u_{i,j})-\Delta t\sum_{r=0}^{j} \widetilde G_1(x_i,t_r,u_{i,r}), \end{equation} where $\widetilde{G}_1(x_i,t_j,u_{i,j})$ is given in (\ref{appG1}). In fact, equation (\ref{cond}) yields \begin{align*} t_jD_{\Delta t}^\alpha u_{i,j}&\,=t_jD_{\Delta t}\widetilde G_1(x_i,t_{j},u_{i,j})=D_{\Delta t}(t_j \widetilde G_1(x_i,t_{j},u_{i,j}))-\widetilde G_1(x_i,t_{j+1},u_{i,j+1})\\ &\,=D_{\Delta t}(\widetilde G_3(x_i,t_{j},u_{i,j})). \end{align*} Therefore, it follows from Theorem~\ref{theodisc} that method (\ref{spectK}) has other two conservation laws that approximate (\ref{FG3}) and (\ref{FG4}), and that are defined by $\widetilde{G}_3$ in (\ref{discG3}), and \begin{align*} &\widetilde{F}_3(x_i,t_j,u_{i,j})=-t_jD_{\Delta x} ({K}(u_{i,j})),\qquad \widetilde{F}_4(x_i,t_{j},u_{i,j})=t_j\widetilde{F}_2(x_i,t_{j},u_{i,j}),\\& \widetilde{G}_4(x_i,t_{j},u_{i,j})=x_i\widetilde{G}_3(x_i,t_{j},u_{i,j}). \end{align*} \begin{table}[t] \caption{Continuous and discrete conservation laws. Subdiffusion: $\ell=1,2$. Superdiffusion: $\ell=1,\ldots,4$.}\label{allcl} \small \begingroup \setlength{\tabcolsep}{5pt} \renewcommand{\arraystretch}{1.35} \centerline{\begin{tabular}{|c||c|c||c|c|} \hline $ \ell$& $F_\ell$ & $G_\ell$ & $\widetilde F_\ell$ & $\widetilde G_\ell$ \\ \hline {$1$} & $-D_xK(u)$ & $\int D_t^\alpha u\,\mathrm{d}t$& $-D_{\Delta x}K(u_{i-1,j})$ & $\Delta t \sum_{\ell=0}^{j-1} D_{\Delta t}^\alpha u_{i,\ell}$\\ \hline $2$ & $K(u)\!-\!xD_xK(u)$ & $xG_1$& $A_{\Delta x}K(u_{i-1,j})\!-\!A_{\Delta x}(x_{i-1})D_{\Delta x}K(u_{i-1,j})$ & $x_i\widetilde{G}_1$\\ \hline {$3$} & $-tF_1$ & $tG_1-\int G_1\,\mathrm{d}t$& $-t_j\widetilde{F}_1$ & $t_j\widetilde{G}_1-\Delta t \sum_{r=0}^{j} \widetilde{G}_1$\\ \hline $4$ & $tF_2$ & $xG_3$& $t_j\widetilde{F}_2$ & $x_i\widetilde{G}_3$\\ \hline \end{tabular}} \endgroup \end{table} \section{Numerical tests}\label{sec:tests} In this section we solve problem \eqref{FPDE} with two different choices of function $K$ that define a linear and a nonlinear problem, respectively. In both cases we set $q=2$ and $(x,t)\in (0,1)\times (0,2)$, and we study the two cases of subdiffusion ($p=1$) and superdiffusion ($p=2$). We consider the boundary conditions \begin{equation}\label{eq:bctest} u(0,t)=u(1,t)=\frac{t^p}p\qquad t\in[0,2], \end{equation} and an initial configuration of total rest \eqref{IC2}, i.e. if $0<\alpha<1$ (subdiffusion case), \begin{equation}\label{eq:test_sub} u(x,0)=0, \end{equation} while, if $1<\alpha<2$ (superdiffusion case), \begin{equation}\label{eq:test_sub2} u(x,0)=\frac{\partial }{\partial t} u(x,0)=0. \end{equation} Therefore, the fractional derivative in \eqref{FPDE} is equivalently defined by either equation \eqref{RLdef} or equation \eqref{Cdef}. The space grid is defined as in (\ref{spgrid}) with $\Delta x=0.005$, that is small enough to study the rate of convergence in time of the method. Inspired by \cite{bcdp17} we choose the time nodes $t_j$ equal to the Chebyshev nodes in $[0,2]$ and the basis $\mathcal{P}_j$ in \eqref{modalexp}--\eqref{discIC} defined by the Jacobi polynomials in $[0,2]$ (see \cite{BHRAWY2015876}), \begin{equation*} \mathcal{P}_j(t)=\sum_{k=0}^j\frac{(-1)^{j+k}(j+k+p)!}{(k+p)!(j-k)!k!2^k}t^k, \qquad j=0,\ldots, N+p. \end{equation*} Different choices, such as uniform nodes and the power basis \eqref{eq:powerb}, can also be considered. However, in the experiments below, choosing Chebyshev nodes and the Jacobi basis yields a matrix $C_1$ in \eqref{eq:matform0} with lower condition number and a faster rate of convergence, respectively. \subsection*{Linear problem} We consider here the linear fractional PDE defined by \eqref{FPDE} with $K(u)=u$, i.e. \begin{equation}\label{eq:lineq} D_t^\alpha u-D_x^2 u=0. \end{equation} The exact solution satisfying the boundary conditions \eqref{eq:bctest} and the initial conditions \eqref{eq:test_sub} or \eqref{eq:test_sub2} are given in \cite{materials} and amounts to \begin{align}\label{uex} u_{\text{exact}}(x,t)=&\,\sum_{n=1}^\infty a_N(x,t) \\\nonumber =&\,\sum_{n=1}^\infty \frac{-4t^p}{(2n-1)\pi}\sin{\left((2n-1)\pi x\right)}E_{\alpha,p+1}(-(2n-1)^2\pi^2t^\alpha)+\frac{t^p}p, \end{align} where $$E_{\alpha,\beta}(z)=\sum_{k=0}^\infty \frac{z^k}{\Gamma(\alpha k+\beta)},$$ is the Mittag-Leffler function with two arguments. A reference solution, $\bar{u}$, is calculated by truncating the infinite sum in (\ref{uex}) after $R$ terms, where $R$ is the smallest integer such that $\|a_R\|<\text{tol} = 10^{-12}$, and by computing the Mittag-Leffler function using the \textsc{Matlab} routine $\texttt{ml}$ \cite{gar15,Garrappa2015}. In order to show the convergence of the proposed numerical scheme, its spectral accuracy and its conservative properties, the error in the numerical solution and in the discrete conservation laws are calculated as $$\text{Sol err}=\max_i\max_j |u_{i,j}-\bar u(x_i,t_j)|,$$ and \begin{equation}\label{CLerr} \text{Err}_\ell=\max_i\max_j |D_{\Delta t}\widetilde{G}_\ell(x_i,t_j,u_{i,j})+D_{\Delta x}\widetilde{F}_\ell(x_i,t_j,u_{i,j})|, \ \ell=1,\ldots,p\cdot q, \end{equation} with functions $\widetilde F_\ell$ and $\widetilde G_\ell$ defined in Section~\ref{secmodel}. \begin{table}[t] \caption{Linear problem. Errors in solution and conservation laws. \label{tab:CLaws}} \small \begingroup \setlength{\tabcolsep}{6pt} \renewcommand{\arraystretch}{1.12} \centerline{\begin{tabular}{|c||c|c|c|c|c|} \hline $\alpha$ & Sol err & Err$_1$ & Err$_2$ & Err$_3$ & Err$_4$\\ \hline 0.1& 2.37e-06& 1.54e-11& 1.21e-11& N.A.& N.A.\\ 0.5& 4.29e-04& 1.59e-11& 1.35e-11& N.A.& N.A.\\ 0.9& 1.04e-03& 1.57e-11& 1.41e-11& N.A.& N.A.\\ 1.1& 2.99e-05& 1.60e-11& 1.30e-11& 3.02e-11& 2.43e-11\\ 1.5& 1.11e-04& 1.67e-11& 1.58e-11& 3.23e-11& 2.20e-11\\ 1.9& 2.34e-03& 1.69e-11& 1.52e-11& 3.30e-11& 2.97e-11\\ \hline \end{tabular}} \endgroup \end{table} In Table~\ref{tab:CLaws} we show the error in the solution and conservation laws given by method \eqref{spectK} with $N=10$ applied to \eqref{eq:lineq}. We consider three different values of $\alpha$ corresponding to subdiffusion problems. Two of these values are close to the integer cases ($\alpha=0.1$ and $\alpha=0.9$). The third is an intermediate value ($\alpha=0.5$). Similarly, we consider three different superdiffusion problems ($\alpha=1.1,1.5,1.9$). The error in the solution is larger for values of $\alpha$ that are closer to $p$. In all cases the error in the conservation laws is only due to the roundoffs. In particular, we verified that the errors in the conservation laws are comparable in magnitude to the residual of equation \eqref{eq:matform0}. \begin{figure}[p] \begin{center} \includegraphics[scale=0.7]{figures/05-3dC.eps} \includegraphics[scale=0.7]{figures/15-3dC.eps} \caption{Linear problem. Numerical solution with $N=10$, $\Delta x=0.005$, $\alpha=0.5$ (left) and $\alpha=1.5$ (right).}\label{diff3d} \end{center} \end{figure} \begin{figure}[p] \begin{center} \input{figures/01JacobiC.tex} \input{figures/05JacobiC.tex} \input{figures/09JacobiC.tex}\\ \input{figures/11JacobiC.tex} \input{figures/15JacobiC.tex} \input{figures/19JacobiC.tex} \end{center} \caption{Linear problem. Rate of convergence in time. $\Delta x=0.005$ (logarithmic scale on $y$-axis).} \label{a2} \end{figure} In Fig.~\ref{diff3d} we show the solutions of method \eqref{spectK} with $\alpha=0.5$ and $\alpha=1.5$. These two graphs very well reproduce the behaviour of the exact solution shown in \cite{materials}. In Figure~\ref{a2} we study the convergence of the method by plotting the logarithm of the error in the solution against $N$ for $N=2,\ldots,18$. These graphs show that the convergence of the method is exponential in time. \subsection*{Nonlinear problem} We consider now equation \eqref{FPDE} with $K(u)=\sqrt{u}$, therefore we solve equation \begin{equation}\label{eq:sqrteq} D_t^\alpha u-D_x^2 (\sqrt{u})=0, \end{equation} with boundary conditions \eqref{eq:bctest} and initial conditions \eqref{eq:test_sub} if $0< \alpha <1$ or \eqref{eq:test_sub2} if $1<\alpha <2$. The exact solution of this problem is not known, and so we compute a reference solution, $\bar{u}$, setting $\bar N=13$. For $N<\bar N$, the error in the solution at the final time is estimated by $$\text{Sol err}=\sqrt{\sum_i |u_{i,N}-\bar u_{i,\bar N}|}.$$ The error in the conservation laws is evaluated as in (\ref{CLerr}). \begin{table}[t] \caption{Nonlinear problem. Errors in solution and conservation laws.\label{tab:CLawsNL}} \small \begingroup \setlength{\tabcolsep}{6pt} \renewcommand{\arraystretch}{1.12} \centerline{\begin{tabular}{|c||c|c|c|c|c|} \hline $\alpha$ & Sol err & Err$_1$ & Err$_2$ & Err$_3$ & Err$_4$\\ \hline 0.1& 6.78e-07 & 2.39e-11 & 1.99e-11 & N.A. & N.A.\\ 0.5& 4.08e-15 & 2.46e-11 & 1.88e-11 & N.A. & N.A.\\ 0.9& 4.37e-05 & 2.76e-11 & 2.25e-11 & N.A.& N.A.\\ 1.1& 1.10e-05 & 2.33e-11 & 2.00e-11 & 4.35e-11 & 3.78e-11 \\ 1.5& 2.00e-04 & 2.94e-11 & 2.83e-11 & 4.83e-11 & 4.64e-11 \\ 1.9& 4.30e-03 & 2.89e-11 & 2.82e-11 & 5.33e-11 & 4.64e-11\\ \hline \end{tabular}} \endgroup \end{table} Table~\ref{tab:CLawsNL} shows the solution error and the error in the conservation laws given by method \eqref{spectK} with $N=10$ applied to \eqref{eq:sqrteq}. The error in the conservation laws is only due to the accuracy in the Newton method to solve the nonlinear system \ref{eq:matform0}. As in the linear case, the error in the solution is larger for $\alpha$ close to $p$, except that in this case the method exactly solves the subdiffusion problem with $\alpha=0.5$. \begin{figure}[p] \begin{center} \includegraphics[scale=0.7]{figures/NL05-3d.eps} \includegraphics[scale=0.7]{figures/NL15-3d.eps} \caption{Nonlinear problem. Numerical solution with $N=10$, $\Delta x=0.005$, $\alpha=0.5$ (left) and $\alpha=1.5$ (right).}\label{NLdiff3d} \end{center} \end{figure} \begin{figure}[p] \begin{center} \input{figures/NL01.tex} \hspace{-0.3cm} \input{figures/NL05.tex} \hspace{-0.3cm} \input{figures/NL09.tex}\\ \input{figures/NL11.tex} \input{figures/NL15.tex} \input{figures/NL19.tex} \end{center} \caption{Nonlinear problem. Rate of convergence in time. $\Delta x=0.005$ (logarithmic scale on $y$-axis).} \label{NL1} \end{figure} The numerical solutions obtained for $\alpha=0.5$ and $\alpha=1.5$ are shown in Figure~\ref{NLdiff3d}. The convergence of the method is analysed in Figure~\ref{NL1} where we plot the logarithm of the error in the solution against $N$ for $N=2,\ldots,12$. Also in this case the rate of convergence in time is exponential, except for $\alpha=0.5$, where the method is exact for all values of $N$. \section{Conclusions}\label{sec:concl} In the present paper we have investigated a class of time fractional diffusion PDEs of arbitrary fractional order. The purpose is twofold. On one hand, we have derived sufficient conditions to find conservation laws of a FDE of this kind, and derived an analogue result for a numerical method to have discrete conservation laws. On the other hand, we have generalised the spectral method in \cite{bcdp17} to approximate Caputo and Riemann-Liouville derivatives of arbitrary order. This method has been coupled with a finite difference approximation in space, and proved to have conservation laws. In the cases of subdiffusion and superdiffusion, we have derived the expressions of the conservation laws and performed numerical tests to confirm the theoretical findings and show the convergence of the method. \subsection*{Acknowledgements} The authors are members of the GNCS group. This work is supported by GNCS-INDAM project and by the Italian Ministry of University and Research, through the PRIN 2017 project (No. 2017JYCLSF) “Structure preserving approximation of evolutionary problems”. \bibliographystyle{plain}

91cf9a53025e6d35313b7f9580e239db190ff8bf

\section{Introduction} In 2019 experts of Solar Cycle 25 Prediction Panel predicted that the maximum of the solar cycle 25 (SC25) will lie in the sunspot number range between 95 and 130, that is similar to SC24 with the maximal SN=116 (\url{https://www.weather.gov/news/190504-sun-activity-in-solar-cycle}). The opinion that SC25 will be equal or lower than SC24 is shared by many authors (\opencite{du22}; \opencite{burud21}; \opencite{nandy21}; \opencite{wu21}; \opencite{courtillot21} etc). If the odd SC25 is lower than the even SC24, it will violate the Gnevyshev-Ohl correlation rule \cite{gnevyshev48}. Many regard this as a sign that the Sun is entering to a Dalton-like global minimum. However, there is an alternative point of view that the forecast of the experts has been underestimated and the forecoming cycle will be higher (\opencite{macintosh20}; \opencite{koutchmy21}; \opencite{prasad22}; \opencite{lu22} etc). None of the two positions have prevailed so far. In the following we propose a simple method of solar maximums predictions based on linear relations between amplitudes of cycles and lengths of its phases. This method provides one more argument that SC25 will be higher that SC24 and comparable with SC23. \section{Data and notation} Hereafter we will use for analysys and prediction the 13-month smoothed monthly averages of the recalibrated sunspot number SN for 1749--2021 \cite{clette14} (\url{https://wwwbis.sidc.be/silso/DATA/SN_ms_tot_V2.0.txt}). We will introduce the followig notation: ${\rm SN}_{\rm max}(i)$ is the sunspot index in the maximum of the $i$th cycle (i.e. its amplitude), $T_{\rm min}(i)$ and $T_{\rm Max}(i)$ are moments of its minimum and maximum, $T_{\rm mm}(i) = T_{\rm min}(i+1) - T_{\rm min}(i)$ is its length (from minimum to minimum), $T_{\rm mM}(i) = T_{\rm Max}(i) - T_{\rm min}(i)$ is the length of the ascending branch of the cycle (from minimum to maximum). \section{The Waldmeiers rule} The well-known empirical Waldmeier rule (WR) states that the length of the ascending branch of the cycle anticorrelates with its amplitude \cite{waldmeier35} (Fig.~\ref{figwr}). One can find the linear regression equation \begin{eqnarray} {\rm SN}_{\rm max}(i) &=& a_1 + b_1\,T_{\rm mM}(i)\,,\nonumber\\ a_1 &=&340.2 \pm 31.3\,,\nonumber\\ b_1 &=& (-36.76 \pm 6.90)\;{\rm yr}^{-1}\,, \label{eqwr} \end{eqnarray} and the correlation coefficient between the regression parameters is \[ r(a_1,b_1) =-0.976\,. \] \begin{figure} \begin{center} \fb{\includegraphics[width=0.8\textwidth]{figwr.eps}} \end{center} \caption{The relationship between the length of the ascending phase $T_{\rm mM}(i)$ and the amplitude of the cycle ${\rm SN}_{\rm max}(i)$ (the Waldmeier rule). The correlation coefficient is $r=-0.75$. The solid curve is the linear regression \eq{eqwr}. The dashed curves mark the standard error ranges for predictions.} \label{figwr} \end{figure} \section{The length-to-next-amplitude rule (LNAR)} Another empirical rule, similar to WR but less known (and still nameless, to the best of our knowledge), states that the length of a given cycle anticorrelates with the amplitude of the next one \cite{hathaway94, solanki02, hathaway15, ivanov21} (Fig.~\ref{figlnar}). The corresponding regression is \begin{eqnarray} {\rm SN}_{\rm max}(i+1) &=& a_2 + b_2\,T_{\rm mm}(i)\,,\nonumber\\ a_2 &=& 544.5 \pm 86.6\,,\nonumber\\ b_2 &=& (-33.01 \pm 7.81)\;{\rm yr}^{-1}\,,\nonumber\\ r(a_2,b_2) &=& -0.994\,. \label{eqlnar} \end{eqnarray} \begin{figure} \begin{center} \fb{\includegraphics[width=0.8\textwidth]{figlnar.eps}} \end{center} \caption{The relationship between the length of the cycle $T_{\rm mm}(i)$ and the amplitude of the next one ${\rm SN}_{\rm max}(i+1)$ (the LNA rule). The correlation coefficient is $r=-0.68$. The solid curve is the linear regression \eq{eqlnar}. The red circle is the prediction for SC25. The dashed curves mark the standard error ranges for predictions.} \label{figlnar} \end{figure} \section{The length-to-ascending-length rule (LALR)} A direct consequence of the WR and LNAR is the third rule, that binds the length of the current cycle with the length of the ascending phase of the next one (Fig.~\ref{figlalr}). These parameters correlate, and the regression is \begin{eqnarray} {\rm T}_{\rm mM}(i+1) &=& a_3 + b_3\,T_{\rm mm}(i) \,,\nonumber\\ a_3 &=& (-1.98 \pm 1.81)\;{\rm yr}\,,\nonumber\\ b_3 &=& 0.570 \pm 0.163\,,\nonumber\\ r(a_3,b_3) &=& -0.994\,. \label{eqlalr} \end{eqnarray} \begin{figure} \begin{center} \fb{\includegraphics[width=0.8\textwidth]{figlalr.eps}} \end{center} \caption{The relationship between the length of the cycle $T_{\rm mm}(i)$ and the length of the next ascending phase $T_{\rm mM}(i+1)$ (the LAL rule). The correlation coefficient is $r=+0.61$. The red circle is the prediction for SC25. The solid curve is the linear regression \eq{eqlalr}. The dashed curves mark the standard error ranges for predictions.} \label{figlalr} \end{figure} \section{Prediction of SC25} The LNA and LAL rules stated above allow to estimate the time and magnitude of the maximum of the next cycle provided we know the length of the current one. Since the minimum between SC24 and SC25 in the 13-month smoothed index occurs in December 2019, the length of the SC24 is 11.0~yr, and it follows from \eq{eqwr} and \eq{eqlalr} that \[ {\rm SN}_{\rm max}(25)=181\pm46\,, \] \[ {\rm T}_{\rm mM}(25)=(4.23\pm0.96)\;{\rm yr}\,, \] and \[ T_{\rm max}(25)=(2024.24\pm0.96)\;{\rm yr} \] (see Figs.~\ref{figlnar} and~\ref{figlalr}). The errors of the predictions are large enough. Nevertheless, according to the prediction, with probability 0.92 SC25 will be higher than SC24 (${\rm SN}_{\rm max}(24) = 116$). Probably, it will be similar to SC23 (${\rm SN}_{\rm max}(23)=176$) or SC17 (${\rm SN}_{\rm max}(17)=189$) (Fig.~\ref{figscxx}). It means that the odd SC25 will be higher than the even SC24, and the Gnevyshev-Ohl rule will keep valid. \begin{figure} \begin{center} \fb{\includegraphics[width=0.8\textwidth]{figscxx.eps}} \end{center} \caption{ The observed SN index for SC24 and SC25 (black line), SC17 (short dashes) and SC23 (long dashes) shifted to match the minimums to that of SC25, our prediction of SC25 maximum (red point with errorbars), the prediction of SC25 (green line) and its maximum (green point with errorbars) by the Solar Cycle Prediction Panel (\protect\url{https://services.swpc.noaa.gov/json/solar-cycle/predicted-solar-cycle.json}).} \label{figscxx} \end{figure} Earlier \cite{ivanov21}, using the same method, we have obtained for SC25 weaker estimates: the amplitude $136\pm36$ and the moment of maximum $2025.7\pm0.7$ (for the Gaussian smoothing of SN with $\sigma=8$~months). This underestimate can be explained by the fact that in January 2021, when the mentioned article was being prepared, it was difficult to determine the moment of SC25 minimum in the smoothed index accurately, and we assumed it to occur too late, in October 2020, when the last local minimum of the index before its fast growth took place. \section{Control of stability} To control the stability of the prediction method we will do the following procedure. Let's take the subseries of parameters from 1st to $(i-1)$th cycle, construct the regressions for the LNAR and LALR using the truncated series and obtain the predictions for the moment and amplitude of the $i$th cycle. The results of the predictions compared with the observed values are plotted in Figs.~\ref{figcamp} and~\ref{figctime}. All predicted values are in $1.5\sigma$ bands relative to the observed values. The number of predictions in the one sigma band is 8 out of 13 (62\%) for amplitudes and 11 out of 13 (85\%) for lengths, which does not contradict to the probability to be in the $1\sigma$ band for the normal distribution (68\%). \section{Conclusions} The proposed method of prediction, despite its simplicity and rather large errors, provides an accuracy of about 15--30\% for the moment of maximum and 25--50\% in its magnitude, which is enough to distinguish between cycles of low, medium and high power. In particular, the method predicts SC25 of medium magnitude ${\rm SN}_{\rm max}(25) = 181\pm46$, which will be higher than the previous one (${\rm SN}_{\rm max}(24) = 116$) with probability 0.92, and with the same probability the Gnevyshev-Ohl correlation rule will be valid for SC25. \begin{figure} \begin{center} \fb{\includegraphics[width=0.8\textwidth]{figcamp.eps}} \end{center} \caption{The control predictions for ${\rm SN}_{\rm max}$ on truncated series. The red circles with errorbars are predictions of the $i$th SC. The blue circles are observed values.} \label{figcamp} \end{figure} \begin{figure} \begin{center} \fb{\includegraphics[width=0.8\textwidth]{figctime.eps}} \end{center} \caption{The control predictions for $T_{\rm mM}$ on truncated series. The red circles with errorbars are predictions of the $i$th SC. The blue circles are observed values.} \label{figctime} \end{figure}

e1ed4cd0321b168314f288e33fe6c61be83c9a5b

\section{Introduction}\label{s.i} \subsection{Main results and structure of the paper}\label{s:sg} Let $d\in {\mathbb N}:=\{1,2,\ldots\}$, $\alpha\in (0,2)$ and $\alpha<d$. We consider the semigroup $\tilde P_t$, $t>0$, of the following Hardy operator on ${{\mathbb R}^d}$, \begin{equation} \label{eq:SchrOp} \Delta^{\alpha/2}+\kappa|x|^{-\alpha}. \end{equation} We call $\kappa$, and $\Delta^{\alpha/2}+\kappa|x|^{-\alpha}$, subcritical if $\kappa<\kappa^*$, critical if $\kappa=\kappa^*$ and supercritical if $\kappa>\kappa^*$. Here $\Delta^{\alpha/2}:=-(-\Delta)^{\alpha/2}$ is the fractional Laplacian, \begin{equation*} \kappa^*:=\frac{2^{\alpha} \Gamma((d+\alpha)/4)^2 }{\Gamma((d-\alpha)/4)^{2}}, \end{equation*} and $\Gamma(t)=\int_0^\infty y^{t-1}e^{-y}\,{\rm d} y$ is the Gamma function. It is well known that $\kappa^*$ is the best constant in the Hardy inequality for the quadratic form of $\Delta^{\alpha/2}$, see Herbst \cite[Theorem~2.5]{MR0436854}, Beckner \cite[Theorem~2]{MR1254832} or Yafaev \cite[(1.1)]{MR1717839}; see also Frank and Seiringer \cite[Theorem~1.1]{MR2469027} and Bogdan, Dyda and Kim \cite[Proposition~5]{MR3460023}. Following \cite[Section~4]{MR3460023}, for $\beta\in [0,d-\alpha]$ we let \begin{equation}\label{e.dkd} \kappa_\beta=\frac{2^\alpha\Gamma((\beta+\alpha)/2) \Gamma((d-\beta)/2)}{\Gamma(\beta/2)\Gamma((d-\beta-\alpha)/2)}, \end{equation} where $\kappa_0 =\kappa_{d-\alpha}= 0$, according to the convention $1/\Gamma(0)=0$. The function $\beta \mapsto \kappa_\beta$ is increasing on $[0,(d-\alpha)/2]$, decreasing on $[(d-\alpha)/2,d-\alpha]$, and $\kappa_\beta = \kappa_{d-\alpha-\beta}$, see \cite[Proof of Proposition~5]{MR3460023}. The maximal or {\it critical} value of $\kappa_\beta$ is, therefore, $\kappa_{(d-\alpha)/2}=\kappa^*$, and for each $\kappa\in [0,\kappa^*]$ there is a unique number $\delta$ such that \begin{equation}\label{e.cdk} 0\le\delta\leq (d-\alpha)/2 \quad \text{ and } \quad \kappa=\kappa_\delta=\frac{2^\alpha\Gamma((\delta+\alpha)/2) \Gamma((d-\delta)/2)}{\Gamma(\delta/2)\Gamma((d-\delta-\alpha)/2)}. \end{equation} In what follows, $\delta$ and $\kappa$ shall satisfy \eqref{e.cdk}. We let $h(x)=h_\delta(x):=|x|^{-\delta}$, $x\in {{\mathbb R}^d}$. By Bogdan, Grzywny, Jakubowski and Pilarczyk \cite{MR3933622}, the Schr\"odinger operator \eqref{eq:SchrOp} has heat kernel $\tilde{p}$ with singularity at the origin in ${{\mathbb R}^d}$ and sharp explicit estimates (given by \eqref{eq:mainThmEstz} below). The first main result of the present paper is a description of the limiting behavior of $\tilde{p}$, as follows. \begin{theorem}\label{t:eta2} The limit $\Psi_t(x):= \lim\limits_{y \to 0} \frac{\tilde{p}(t,x,y)}{h(y)}$ exists whenever $0\le\delta\leq \frac{d-\alpha}{2}$, $t>0$, $x\in {{\mathbb R}^d}$. \end{theorem} The proof of Theorem~\ref{t:eta2} is given in Section~\ref{s:csss}. The function $\Psi_t(x)$ is a self-similar semigroup solution of the heat equation for the Hardy operator, as we assert in \rf{e.s} and \rf{ss form eta1} below. It has an important application to large-time asymptotics of the semigroup $\tilde P_t$, which we now present. To this end we consider Doob--conditioned and weighted $L^q$ spaces. Let $$H=\max\{1,h\}.$$ As usual, $L^1=L^1({{\mathbb R}^d},\,{\rm d} x)$, $L^1(H)=L^1({{\mathbb R}^d},H(x)\,{\rm d} x)$, etc. We have $L^1(H)=\{f/H: f\in L^1\}=L^1(h) \cap L^1$. We then define, for $1\le q <\infty$, \begin{equation}\label{normqh} \|f\|_{q,h }:=\|f/h\|_{L^q(h^2)} =\bigg( \int_{{{\mathbb R}^d}} |f(x)|^q h^{2-q}(x)\,{\rm d} x \bigg) ^{\frac{1}{q}} =\|f\|_{L^q(h^{2-q})} , \end{equation} and, for $q=\infty$, \begin{equation*} \|f\|_{\infty, h}:={\rm ess} \sup_{x\in {\mathbb R}^d} |f(x)|/h(x). \end{equation*} Of course, $\|f\|_{2,h}=\|f\|_2$ and $\|f\|_{1,h}=\|f\|_{L^1(h)}$. For $f\in L^1(H)$ we let \begin{equation}\label{e.dtp} \tilde P_t f(x):=\int_{{\mathbb R}^d} \tilde p(t,x,y)f(y) \,{\rm d} y,\quad t>0,\; x\in {{\mathbb R}^d}\setminus \{0\}. \end{equation} Our second main result is the following large-time asymptotics for $\tilde P_t$. \begin{theorem}\label{lim norm th} If $f\in L^1(H)$, $A=\int_{{{\mathbb R}^d}}f (x) h(x) \,{\rm d} x$, $u(t,x)=\tilde P_t f(x)$, and $q\in[1,\infty)$, then \begin{equation}\label{e.as} \lim_{t\rightarrow \infty }t^{\frac{d-2\delta}{\alpha}(1-\frac{1}{q})}\| u(t, \cdot)-A\Psi_t\|_{q, h}=0. \end{equation} \end{theorem} The structure of the paper is as follows. The proof of Theorem~\ref{lim norm th} is given at the end of Section~\ref{s.ssa}, where we also show that the result is optimal. In Section~\ref{s:csss} we state and prove Theorem \ref{t:eta}, of which Theorem \ref{t:eta2} is a direct consequence. In Section~\ref{s.fatp} we discuss the Feynman-Kac semigroup $\tilde{P}_t$ from the point of view of functional analysis, in particular we prove hypercontractivity of the semigroup in Theorem~\ref{hyperc}, which is then used in Section~\ref{s.ssa}. The last main result of the paper, Theorem~\ref{thm:intu} in Section~\ref{s.psss}, gives an explicit formula for the potential $\int_0^\infty \Psi_t(x) \,{\rm d} t$ of the self-similar solution. Notably, Theorem~\ref{thm:intu} and Corollary~\ref{cor:intPsi} further the integral analysis which is the foundation of \cite{MR3933622}. They were inspired by one of our earlier attempts to prove Theorem~\ref{t:eta2} and are particularly interesting for $\kappa=\kappa^*$, see \eqref{e.pcc}. \subsection{Motivation and methods} \label{s:Mm} The classical result of Baras and Goldstein \cite{MR742415} asserts the existence of nontrivial nonnegative solutions of the heat equation $ \partial_t = \Delta + \kappa|x|^{-2}$ in ${{\mathbb R}^d}$ for (subcritical) $\kappa \in [0, (d - 2)^2/4]$, and non-existence of such solutions for (supercritical) $\kappa>(d-2)^2/4$. Later on, the upper and lower bounds for the heat kernel of the subcritical Hardy operator $\Delta + \kappa|x|^{-2}$ were obtained by Liskevich and Sobol \cite{MR1953267}, Milman and Semenov \cite{MR2064932,MR2180080}, Moschini and Tesei \cite{MR2328115}, Filippas, Moschini and Tertikas \cite{MR2308757}. The classical Hardy operator $\Delta+\kappa|x|^{-2}$ plays a distinctive role in limiting and self-similar phenomena in probability \cite{2017-LM-EP} and partial differential equations \cite{MR3020137}. This is related to the scaling of the corresponding heat kernel, which is the same as for the Gauss-Weierstrass kernel, and to the asymptotics at the origin in ${{\mathbb R}^d}$, which is very different. Such applications motivate our work on the Hardy perturbation of the fractional Laplacian. In fact the paper \cite{MR3933622} was a preparation for the present work, which now comes to fruition. The strategy of the proof of Theorem~\ref{t:eta2} is to prove and use the existence of a stationary density of a corresponding Ornstein-Uhlenbeck semigroup. Then the large-time asymptotics of the Ornstein-Uhlenbeck semigroup yields the asymptotics of $\tilde p$ at the origin. To the best of our knowledge the approach is new and should apply to other heat kernels with scaling. Let us also comment on Theorem~\ref{lim norm th}. We start by recalling the initial value problem for the classical heat equation, \begin{equation}\label{c.h.eq} \begin{cases} \partial_t u(x,t)=\Delta u(x,t) , \quad x\in {\mathbb R}^d, \quad t>0 ,\\ u(x,0)=f(x) . \end{cases} \end{equation} For $f \in L^1 ({\mathbb R}^d )$ the following is asymptotics is well-known: \begin{equation}\label{heat lim} \lim_{t\rightarrow \infty }t^{\frac{d}{2}(1-\frac{1}{p})}\| u(t,\cdot)-Mg_t\|_{L^p({\mathbb R}^d)} =0, \end{equation} see, e.g., Giga, Giga and Saal \cite[Theorem in Sect. 1.1.4]{Giga} or Duoandikoetxea and Zuazua \cite{DZ}). Here $p\in [1, \infty ]$, $M:=\int_{{\mathbb R}^d}f(x) \,{\rm d} x $, $u(t,x) = g_t \ast f(x)$ is the semigroup solution of \rf{c.h.eq}, and $g_t(x)=\big(4\pi t\big)^{-d/2}\exp( -|x|^2/(4 t))$ is the Gauss-Weierstrass kernel. Of course, we have \textit{scaling}: $g_t(x)=t^{-d/2}g_1(t^{-1/2}x)$, that is, the function is \textit{self-similar}. The function also satisfies the first equality in \eqref{c.h.eq}. We can consider \eqref{heat lim} as a statement about the universality of the self-similar solution $g_t(x)$ for the large-time behavior of all solutions to \eqref{c.h.eq}. Theorem~\ref{lim norm th} gives an analogous result for $u(t,x)=\tilde P_t f(x)$ and the initial value problem \begin{equation}\label{equation} \begin{cases} \partial_t u(x,t)=\left(\Delta^{\alpha/2} +\kappa |x|^{-\alpha}\right) u(x,t) , \quad x\in {\mathbb R}^d, \quad t>0 ,\\ u(x,0)=f(x) . \end{cases} \end{equation} for $\kappa$ satisfying \eqref{e.cdk} and $f\in L^1(H)$. The use of $h$ is novel in this setting -- this and connections to the classical literature will be discussed in more detail in Section~\ref{s.ssa}. \subsection{General conventions} \label{s:n} We tend to use ``:='' to indicate definitions, e.g., $a\land b := \min\{a, b\}$ and $a\vee b := \max\{a, b\}$, $a_+:=a \vee 0$ and $a_-:=(-a)\vee 0$. Throughout, we only consider Borel measurable functions and Borel measures. As usual, integrals are considered well-defined when the integrands are nonnegative or absolutely integrable with respect to a given measure. In the case of integral kernels, the corresponding integrals should at least be well-defined pointwise almost everywhere ($a.e.$). For $x\in {{\mathbb R}^d}$ and $r>0$ we define $B(x,r)=\{y\in {{\mathbb R}^d}: |y-x|<r\}$, the ball with center at $x$ and radius $r$. We write $f\approx g$, which we call \textit{approximation} or \textit{comparison}, and say $f$ and $g$ are \textit{comparable}, if $f,g$ are nonnegative functions, $c^{-1}g\le f\le cg$ with some constant $c$, that is a number in $(0,\infty)$. The values of constants may change without notice from line to line in a chain of estimates. Of course, we shall also use constants in inequalities (one-sided comparisons of functions), e.g., $f\le cg$. We occasionally write $c=c(a,\ldots,z)$ to assert that the constant $c$ may be so selected as to depend only on $a,\ldots,z$. As usual, for $1\le p\le \infty$, $L^p:=L^p({{\mathbb R}^d}, \,{\rm d} x)$, with norm $\|\cdot \|_p$, and $L^p(g):=L^p({{\mathbb R}^d}, g \,{\rm d} x)$ with norm $\| f \|_{L^p(g)}$ and nonnegative (weight) function $g$. \subsection*{Acknowledgements} We thank Tomasz Grzywny, Piotr Knosalla, Tomasz Komorowski, Alex Kulik, Markus Kunze, \L{}ukasz Le\.zaj, Grzegorz Serafin, \L{}ukasz Stettner, and Tomasz Szarek for helpful discussions, comments and references. We particularly thank Tomasz Komorowski for much advice on the proof of Theorem~\ref{c.gsall} and Alex Kulik for discussions on \eqref{e.convL}. \section{Preliminaries} \subsection{Fractional Laplacian} Let $$ \nu(y) =\frac{ \alpha 2^{\alpha-1}\Gamma\big((d+\alpha)/2\big)}{\pi^{d/2}\Gamma(1-\alpha/2)}|y|^{-d-\alpha}\,,\quad y\in {{\mathbb R}^d}\, . $$ The coefficient is so chosen that \begin{equation} \label{eq:trf} \int_{{{\mathbb R}^d}} \left[1-\cos(\xi\cdot y)\right]\nu(y)\,{\rm d} y=|\xi|^\alpha\,,\quad \xi\in {{\mathbb R}^d}\,, \end{equation} see, e.g., Bogdan, Byczkowski, Kulczycki, Ryznar, Song, and Vondra{\v{c}}ek \cite[(1.28)]{MR2569321}. The fractional Laplacian for (smooth compactly supported) \textit{test functions} $\varphi\in C^\infty_c({{\mathbb R}^d})$ is \begin{equation*} \Delta^{\alpha/2}\varphi(x) = \lim_{\varepsilon \downarrow 0}\int_{|y|>\varepsilon} \left[\varphi(x+y)-\varphi(x)\right]\nu(y)\,{\rm d} y\,, \quad x\in {{\mathbb R}^d}\,. \end{equation*} Many authors use the notation $-(-\Delta)^{\alpha/2}$ for the operator. In terms of the Fourier transform, $\widehat{\Delta^{\alpha/2}\varphi}(\xi)=-|\xi|^{\alpha} \widehat{\varphi}(\xi)$, see, e.g., \cite[Section~1.1.2]{MR2569321} or \cite{MR3613319}. \subsection{The semigroup of $\Delta^{\alpha/2}$} We consider the convolution semigroup of functions \begin{equation} \label{eq:dpt} p_t(x):=(2\pi)^{-d}\int_ {{\mathbb R}^d} e^{-t|\xi|^\alpha}e^{-ix\cdot\xi}\,{\rm d} \xi\,,\quad\ t>0,\ x\in {{\mathbb R}^d}\,. \end{equation} According to (\ref{eq:trf}) and the L\'evy-Khinchine formula, each $p_t$ is a radial probability density function and $\nu(y)\,{\rm d} y$ is the L\'evy measure of the semigroup, see, e.g., \cite{MR2569321}. From (\ref{eq:dpt}) we have \begin{equation} \label{eq:sca} p_t(x)=t^{-d/\alpha}p_1(t^{-{1/\alpha}}x)\,. \end{equation} It is well-known that $p_1(x) \approx1\land |x|^{-d-\alpha}$ (see \cite{MR2569321}, Bogdan, Grzywny and Ryznar \cite[remarks after Theorem~21]{MR3165234} or \cite{MR3613319}), so \begin{equation}\label{eq:oppt} p_t(x) \approx t^{-d/\alpha}\land \frac{t}{|x|^{d+\alpha}} \,,\quad t>0,\ \ x\in {{\mathbb R}^d}\,. \end{equation} Since to $\alpha<d$, we get (the Riesz kernel) \begin{align}\label{eq:pot} \int_0^\infty p_t(x)\,{\rm d}{t} = {\mathcal A}_{d,\alpha} |x|^{\alpha-d}, \quad x\in {{\mathbb R}^d}, \end{align} where \begin{equation}\label{e.dAda} {\mathcal A}_{d,\alpha} = \frac{\Gamma(\frac{d-\alpha}{2})}{\Gamma(\frac{\alpha}{2})2^\alpha \pi^{d/2}}, \end{equation} see, e.g., \cite[Section 1.1.2]{MR2569321}. We denote $$ p(t,x,y)=p_t(y-x), \quad t>0,\ x,y\in {{\mathbb R}^d}. $$ Clearly, $p$ is symmetric: \begin{equation*}\label{pt} p(t,x, y) = p(t,y, x), \quad t>0,\ x,y\in {{\mathbb R}^d}, \end{equation*} and satisfies the Chapman-Kolmogorov equations: \begin{equation}\label{eq:ChKforp} \int_{{{\mathbb R}^d}} p(s,x, y)p(t,y, z)\,{\rm d} y = p(t+s,x, z), \quad x, z \in {{\mathbb R}^d},\, s, t > 0. \end{equation} We denote, as usual, $P_t g(x) = \int_{{{\mathbb R}^d}} p(t,x, y) g(y)\,{\rm d} y$. The fractional Laplacian extends to the generator of the semigroup $\{P_t\}_{t>0}$ on many Banach spaces, see, e.g., \cite{MR3613319}. \subsection{Schr\"odinger perturbation by Hardy potential} We recall elements of the integral analysis of \cite{MR3460023} and \cite{MR3933622}, which was used to handle the heat kernel $\tilde p$ of $\Delta^{\alpha/2}+\kappa |x|^{-\alpha}$. Thus, $$f_\beta(t) := c_\beta t_+^{(d-\alpha -\beta)/\alpha },\quad t\in {\mathbb R},$$ for $\beta \in (0, d)$. The constant $c_\beta$ is so chosen that \begin{equation}\label{h_beta} h_\beta(x) := \int_0^\infty f_\beta(t) p_t(x) {\rm d}t =|x|^{-\beta}, \quad x\in {{\mathbb R}^d}. \end{equation} The existence of such $c_\beta\in (0, \infty)$ follows from \eqref{eq:sca} and the estimate $p_1(x) \approx1\land |x|^{-d-\alpha}$. Of course, $f_\beta'(t)=c_\beta\frac{d-\alpha-\beta}{{\alpha}}t^{(d-2\alpha -\beta)/\alpha }$. Accordingly, for $\beta \in (0, d-\alpha)$ we may define \begin{equation*}\label{def_q} q_\beta(x) = \frac{1}{h_\beta(x)}\int_0^{\infty }f_\beta'(t) p_t(x) {\rm d}t,\quad x\in {\mathbb R}_0^d, \end{equation*} where ${\mathbb R}_0^d :={{\mathbb R}^d}\setminus \{0\}$. By \cite[(26)]{MR3460023}\footnote{The exponent $(d-\alpha -\beta)/\alpha$ in the definition of $f$ is denoted $\beta$ in \cite[Corollary~6]{MR3460023}.}, \begin{equation*}\label{q_beta} q_\beta(x) = \kappa_\beta |x|^{-\alpha }, \end{equation*} where $\kappa_\beta$ is defined by \eqref{e.dkd}. In what follows we keep our notation from \eqref{e.cdk}, that is, we let \begin{equation}\label{eq:dk0} \delta \in [0,(d-\alpha)/2], \quad \kappa = \kappa_\delta,\quad h(x)=h_\delta(x)=|x|^{-\delta},\quad q(x) = q_\delta(x)=\kappa |x|^{-\alpha}. \end{equation} To wit, the case of $\delta \in (0,(d-\alpha)/2]$ is covered by the discussion of $\beta$ above, and $\delta=0$ yields the trivial $\kappa=0$, $q=0$ and $h=1$. We then define the Schr\"odinger perturbation of $p$ by $q$: \begin{equation}\label{def_p_tilde} \tilde p =\tilde p_\delta= \sum_{n=0}^{\infty }p_n. \end{equation} Here for $t>0$ and $x,y\in {{\mathbb R}^d}$ we let $p_0(t,x, y) = p(t,x, y)$ and then proceed by induction: \begin{align}\label{pn} p_n(t,x, y) &= \int_0^t \int_{{{\mathbb R}^d}} p(s,x, z) q(z)p_{n-1}(t-s, z, y) \,{\rm d} z \,{\rm d} s \\ &=\int_0^t \int_{{{\mathbb R}^d}} p_{n-1}(s,x, z) q(z)p(t-s, z, y) \,{\rm d} z \,{\rm d} s , \quad n \geq 1.\notag \end{align} Of course, $\tilde{p}_0=p$. By \eqref{eq:oppt}, for $t>0$ and $y\in {{\mathbb R}^d}$ we have \begin{align*} p_1(t,0,y)&=\int_0^t \int_{{{\mathbb R}^d}} p(s,0, z) q(z)p(t-s, z, y) \,{\rm d} z \,{\rm d} s \\ &\ge c_{t,y}\int_0^{t/2} \int_{|z|<s^{1/{\alpha}}} s^{-d/{\alpha}} |z|^{-{\alpha}} \,{\rm d} z \,{\rm d} s =\infty. \end{align*} By symmetry, $p_1(t,x,0)=\infty$, too, for all $x\in {{\mathbb R}^d}$ and $t>0$, therefore $\tilde{p}(t,x,y)=\infty$ if $x=0$ or $y=0$. By \cite[Theorem 1.1]{MR3933622}, the above discussion of $x=0$ and $y=0$ and the usual notational conventions we have for all $x,y\in {{\mathbb R}^d}$, $t>0$, \begin{equation} \label{eq:mainThmEstz} \tilde{p}(t,x,y)\approx \left( t^{-d/\alpha}\wedge \frac{t}{|x-y|^{d+\alpha}} \right)\left(1+t^{\delta/\alpha}|x|^{-\delta} \right)\left(1+t^{\delta/\alpha}|y|^{-\delta} \right). \end{equation} Clearly, if $0<t_1<t_2<\infty$, then \begin{equation}\label{e.crt1t2p} \tilde p(s,x,y)\approx \tilde p(t, x,y), \quad x,y\in {{\mathbb R}^d}, \quad t_1\le s,t\le t_2 \end{equation} (the comparability constant does depend on $t_1$, $t_2$ and $d$, $\alpha$, $\delta$). Recall that \begin{equation}\label{e.dH} H(x)=1\vee h(x)=1\vee |x|^{-\delta}\approx 1+|x|^{-\delta},\quad x\in {{\mathbb R}^d}. \end{equation} Thus, we can reformulate \eqref{eq:mainThmEstz} as follows, \begin{equation} \label{eq:mainThmEstz2} \tilde{p}(t,x,y)\approx p(t,x,y)\ H(t^{-1/\alpha}x)\ H(t^{-1/\alpha}y), \qquad t>0,\quad x,y\in {{\mathbb R}^d}. \end{equation} By \cite{MR3933622} and the above conventions, $\tilde p$ is a symmetric time-homogeneous transition density on ${{\mathbb R}^d}$, in particular the Chapman-Kolmogorov equations hold: \begin{align}\label{eq:ChK} \int_{{\mathbb R}^d} \tilde{p}(s,x,z) \tilde{p}(t,z,y) \,{\rm d} z = \tilde{p}(t+s,x,y)\,,\quad {x,y\in {{\mathbb R}^d}, \ s,t>0.} \end{align} The following Duhamel formulae hold for $p$ and $\tilde p$, \begin{align}\label{eq:Df1} \tilde{p}(t,x,y) &= p(t,x,y) + \int_0^t \int_{{\mathbb R}^d} p(s,x,z) q(z) \tilde{p}(t-s,z,y) \,{\rm d} z \,{\rm d} s \\ &= p(t,x,y) + \int_0^t \int_{{\mathbb R}^d} \tilde{p}(s,x,z) q(z) p(t-s,z,y) \,{\rm d} z \,{\rm d} s ,\quad t>0,\ x,y\in {\mathbb R}^d. \label{eq:Df2} \end{align} In passing we refer to Bogdan, Hansen and Jakubowski \cite{MR2457489} and Bogdan, Jakubowski and Sydor \cite{MR3000465} for a general setting of Schr\"odinger perturbations of transition semigroups and other families of integral kernels. The function $\tilde p$ is self-similar, i.e., has the following scaling \cite[Lemma~2.2]{MR3933622}: \begin{equation}\label{ss form} \tilde p(t,x, y)=t^{-d/\alpha } \tilde p\big(1, t^{-1/\alpha }x , t^{-1/\alpha }y \big), \qquad t>0, \ x,y \in {\mathbb R}^d. \end{equation} This is the same scaling as for $p$. Furthermore, if $T$ is a (linear) isometry of ${{\mathbb R}^d}$, then for all $t>0$, $x,y \in {{\mathbb R}^d}$ we have $p(t,Tx,Ty)=p_t(T(y-x))=p_t(y-x)=p(t,x,y)$, because $p_t$ is radial. Of course, $q$ is radial, too. By the change of variables $z=Tv$ and induction, \begin{align*} &p_n(t,Tx, Ty) = \int_0^t \int_{{{\mathbb R}^d}} p(s,Tx, z) q(z)p_{n-1}(t-s, z, Ty) \,{\rm d} z \,{\rm d} s \\ &= \int_0^t \int_{{{\mathbb R}^d}} p(s,Tx, Tv) q(Tv)p_{n-1}(t-s, Tv, Ty) \,{\rm d} v \,{\rm d} s \\ &=\int_0^t \int_{{{\mathbb R}^d}} p(s,x, v) q(v)p_{n-1}(t-s, v, y) \,{\rm d} v \,{\rm d} s = p_n(t,x, y), \quad n \geq 1. \end{align*} Therefore, \begin{equation}\label{isom} \tilde p(t,Tx, Ty)=\tilde{p} (t,x,y), \qquad t>0, \ x,y \in {\mathbb R}^d. \end{equation} \subsection{Doob's conditioning} Recall that $\delta \in [0,(d-\alpha)/2]$ and $\kappa = \kappa_\delta$, $h(x)=h_\delta (x)=|x|^{-\delta}$, $\tilde p$ depends on $\delta$. By \cite[Theorem~3.1]{MR3933622} and the preceding discussions, the function $h$ is \textit{invariant} in the following sense: \begin{align}\label{eq:2} \int_{{\mathbb R}^d} \tilde{p}(t,x,y) h(y) \,{\rm d} y = h(x), \qquad t>0,\; x \in {{\mathbb R}^d}\,. \end{align} We define the following Doob-conditioned (renormalized) kernel \begin{equation}\label{e:drhot} \rho_t(x,y)=\frac{\tilde p(t,x,y)}{h(x)h(y)},\quad t>0,\quad x,y\in {{\mathbb R}^d_0} \end{equation} (later on we shall extend $\rho$ to $(0,\infty)\times{{\mathbb R}^d}\times{{\mathbb R}^d}$). We consider the integral weight $h^2(x), x\in {{\mathbb R}^d}$. {Doob-type conditioning is also called Davies' method, see Murugan and Saloff-Coste \cite{MR3601569}, and $h$ is sometimes called desingularizing weight, see Milman and Norsemen \cite{MR2064932}.} By \eqref{eq:2}, \begin{equation}\label{e:rhois1} \int_{{\mathbb R}^d} \rho_t(x,y)h^2(y)\,{\rm d} y =1, \quad x\in {{\mathbb R}^d_0}, \quad t>0. \end{equation} By \rf{eq:ChKforp}, \begin{align} &\int_{{{\mathbb R}^d}} \rho_s(x, y)\rho_t(y, z)h^2(y)\,{\rm d} y = \int_{{{\mathbb R}^d}} \frac{\tilde p(s,x, y)}{h(x)h(y)}\frac{\tilde p(t,y, z)}{h(y)h(z)}h^2(y)\,{\rm d} y\nonumber\\ &=\frac{1}{h(x)h(z)} \tilde p(t+s,x, z) =\rho_{t+s}(x,z), \quad x, z \in {{\mathbb R}^d_0},\, s, t > 0.\label{e.CKfrho} \end{align} We see that $\rho$ is a symmetric time-homogeneous transition probability density on ${{\mathbb R}^d_0}$ with the reference measure $h^2(y)\,{\rm d} y$. For nonnegative $f\in L^1(h^2)$, by Fubini-Tonelli and \eqref{e:rhois1}, \begin{align*} \int_{{\mathbb R}^d} \int_{{\mathbb R}^d} f(y)\rho_t(x,y)h^2(y) \,{\rm d} y \,h^2(x) \,{\rm d} x &=\int_{{\mathbb R}^d} f(y) h^2(y) \,{\rm d} y , \end{align*} so the operators $$ \mathcal{R}_t f(x):=\int_{{\mathbb R}^d} f(y)\rho_t(x,y)\,h^2(y) \,{\rm d} y , \quad t>0, $$ are contractions on $L^1(h^2)$. By \eqref{ss form}, $\rho$ is self-similar: for $t>0$ and $x,y \in {\mathbb R}^d$ we have \begin{equation}\label{ss form rho} \rho_t(x, y)=\frac{t^{-d/\alpha } \tilde p\big(1, t^{-1/\alpha }x , t^{-1/\alpha }y \big)} {t^{-\delta/\alpha}h(t^{-1/\alpha}x)t^{-\delta/\alpha}h(t^{-1/\alpha}y)} =t^{\frac{2\delta-d}{\alpha} } \rho_1\big(t^{-1/\alpha }x , t^{-1/\alpha }y \big), \end{equation} hence \begin{equation}\label{ss form rhos} \rho_{st}(t^{1/\alpha }x, t^{1/\alpha }y) =t^{\frac{2\delta-d}{\alpha} } \rho_s(x , y), \quad s>0. \end{equation} For each (linear) isometry $T$ of ${{\mathbb R}^d}$, $h\circ T=h$, thus by \eqref{isom} we get \begin{equation}\label{risom} \rho_t(Tx, Ty)=\rho_t (t,x,y), \qquad t>0, \ x,y \in {{\mathbb R}^d_0}. \end{equation} By \eqref{eq:mainThmEstz}, \begin{equation} \label{eq:mainThmEstrho} \rho_1(x,y)\approx \left(1 \wedge {|x-y|^{-d-\alpha}} \right)\left(1+|x|^{\delta} \right)\left(1+|y|^{\delta} \right) ,\quad \ x,y \in {{\mathbb R}^d_0}. \end{equation} Notably, $\rho_1$ is not bounded (consider large $x=y$). By \eqref{ss form rho}, \begin{equation} \label{eq:mainThmEstrhot} \rho_t(x,y)\approx \left( t^{-d/\alpha}\wedge \frac{t}{|x-y|^{d+\alpha}} \right)\left(t^{\delta/\alpha}+|x|^{\delta} \right)\left(t^{\delta/\alpha}+|y|^{\delta} \right),\quad t>0, \ x,y \in {{\mathbb R}^d_0}. \end{equation} We also note that if $0<t_1<t_2<\infty$, then \begin{equation}\label{e.crt1t2} \rho_{s}(x,y)\approx \rho_{t}(x,y), \quad x,y\in {{\mathbb R}^d_0}, \quad t_1\le s,t\le t_2 \end{equation} (the comparability constant depends on $t_1$, $t_2$ and $d$, $\alpha$, $\delta$). By \eqref{eq:mainThmEstrho}, for every $M\in (0,\infty)$, \begin{equation} \label{eq:mainThmEstrho2} \rho_1(x,y)\approx (1+|y|)^{-d-\alpha+\delta}, \quad 0<|x|\le M,\quad y\in {{\mathbb R}^d_0}. \end{equation} Clearly, the right-hand side is integrable with respect to $h^2(y) \,{\rm d} y $ and bounded. Our aim is to prove the continuity of $\rho_1$, notably at $x=0$. \section{Limiting behaviour at the origin }\label{s:csss} Here is a full-fledged variant of Theorem~\ref{t:eta2}. \begin{theorem}\label{t:eta} The function $\rho$ has a continuous extension to $(0,\infty)\times {{\mathbb R}^d}\times {{\mathbb R}^d}$ and \begin{equation}\label{e.det} \rho_t(0,y):=\lim_{x\to 0}\rho_t(x,y), \quad t>0,\quad y\in {{\mathbb R}^d_0} , \end{equation} satisfies: \begin{equation}\label{ss form eta} \rho_t(0,y) = t^{\frac{2\delta-d}{\alpha} } \rho_1(0,t^{-1/\alpha }y), \quad t>0,\quad y\in {{\mathbb R}^d_0} , \end{equation} \begin{equation}\label{e:ssf} \int_{{\mathbb R}^d} \rho_t(0,y)\rho_s(y,z)h^2(y)\,{\rm d} y=\rho_{t+s}(0,z),\quad s,t>0,\quad z\in {{\mathbb R}^d_0} . \end{equation} \end{theorem} The proof of Theorem~\ref{t:eta} is given below in this section. Let us explain the line of attack. Given \eqref{e.det} and taking the limit in \eqref{e.CKfrho} as $x\to 0$, we should get \eqref{e:ssf}. By \eqref{ss form rho} we should then obtain \eqref{ss form eta} and \begin{equation}\label{e:ssf2} \int_{{\mathbb R}^d} t^{\frac{2\delta-d}{\alpha}}\rho_1(0,t^{-1/\alpha}y)\rho_s(y,z)h^2(y)\,{\rm d} y=(t+s)^{\frac{2\delta-d}{\alpha}} \rho_1(0,(t+s)^{-1/\alpha}z),\quad s,t>0, z \in {{\mathbb R}^d_0}. \end{equation} Changing variables $u=(t+s)^{-1/\alpha}z$ and $x=t^{-1/\alpha}y$, we see that \eqref{e:ssf2} is equivalent to \begin{equation}\label{e:ssf3} \int_{{\mathbb R}^d} \rho_1(0,x)(t+s)^{\frac{d-2\delta}{\alpha}}\rho_s(t^{1/\alpha}x,(t+s)^{1/\alpha}u)h^2(x)\,{\rm d} x= \rho_1(0,u). \end{equation} By \eqref{ss form rhos}, this is the same as \begin{equation}\label{e:ssf3b} \int\limits_{{\mathbb R}^d} \rho_{\frac{s}{s+t}}\left(\left(t/(s+t)\right)^{1/\alpha}x,u\right)\rho_1(0,x)h^2(x)\,{\rm d} x= \rho_1(0,u). \end{equation} In what follows we shall \textit{define} $\rho_1(0,\cdot)$ as a solution to the integral equation \rf{e:ssf3b} and then essentially reverse the above reasoning. Additionally, to simplify the notation and arguments we introduce an auxiliary Ornstein-Uhlenbeck-type semigroup. \subsection{Ornstein-Uhlenbeck semigroup} For $f\ge 0$ we let \begin{equation} L_tf(y)=\int_{{\mathbb R}^d} l_t(x,y) f(x)h^2(x) \,{\rm d} x , \label{e:ssf4e} \end{equation} where \begin{equation}\label{e.dlt} l_t(x,y)=\rho_{1-e^{-t}}(e^{-t/\alpha}x,y), \qquad t>0,\quad x,y\in {{\mathbb R}^d_0}. \end{equation} By \eqref{e.CKfrho} and \eqref{ss form rhos}, \begin{align*} &\int_{{\mathbb R}^d} l_s(x,y) l_t(y,z)h^2(y)\,{\rm d} y= \int_{{\mathbb R}^d} \rho_{1-e^{-s}}(e^{-s/\alpha}x,y) \rho_{1-e^{-t}}(e^{-t/\alpha}y,z)h^2(y) \,{\rm d} y \\ &=\int_{{\mathbb R}^d} \rho_{1-e^{-s}}(e^{-s/\alpha}x,y) (e^{-t})^{\frac{2\delta-d}{\alpha}} \rho_{e^{t}-1}(y,e^{t/\alpha}z)h^2(y) \,{\rm d} y \\ &=(e^{t})^{\frac{2\delta-d}{\alpha}} \rho_{e^t-e^{-s}}(e^{-s/\alpha}x, e^{t/\alpha} z) =\rho_{1-e^{-s-t}}(e^{-(s+t)/\alpha}x,z) =l_{t+s}(x,z). \end{align*} Thus $l_t(x,y)$ is a transition density on ${{\mathbb R}^d_0}$ with respect to measure $h^2(y) \,{\rm d} y $. By Fubini's theorem, $\{L_t\}_{ t>0}$ is a semigroup of operators on $L^1(h^2)$, an Ornstein-Uhlenbeck-type semigroup, see \cite[solution to E 18.17 on p. 462--463]{MR3185174}. It shall be a major technical tool in our development. \subsection{Stationary density} If $\varphi\ge 0$ and $\int \varphi(x)h^2(x)\,{\rm d} x=1$, then we say that $\varphi$ is a \textit{density}. By Fubini-Tonelli Theorem, for $t>0$ and $f\ge0$ \eqref{e:rhois1}, \begin{align}\label{e.iso-1} \int_{{\mathbb R}^d} L_t f(y)h^2(y) \,{\rm d} y &=\int_{{\mathbb R}^d} f(x)h^2(x) \,{\rm d} x . \end{align} Thus, the operators $L_t$ preserve densities. So they are Markov, see Komorowski \cite{MR1162571}, Lasota and Mackey \cite{MR1244104}, Lasota and York \cite{MR1265226} for this setting. We say that density $\varphi$ is \textit{stationary} for $L_t$ if $ L_t\varphi=\varphi. $ \begin{theorem}\label{c.gsall} There is a unique stationary density $\varphi$ for the operators $L_t$, $t>0$. \end{theorem} \begin{proof} Fix $t>0$ and let $P=L_t$ so that $ P^k=L_{kt} $, $k=1,2,\ldots$. By \eqref{e.dlt} and \eqref{ss form rho}, for all $f\ge 0$ and $k\in {\mathbb N}$ we have \begin{eqnarray}\label{e.pts} P^{k}f(u) &=& \int_{{\mathbb R}^d} f(y) \rho_{1-e^{-kt}}(e^{-kt/\alpha}y,u) h^2(y) \,{\rm d} y \\ &=& \int_{{\mathbb R}^d} f(y) (e^{-kt})^{\frac{2\delta-d}{\alpha}} \rho_{e^{kt}-1}(y,e^{kt/\alpha}u) h^2(y) \,{\rm d} y , \qquad u \in {{\mathbb R}^d_0}.\nonumber \end{eqnarray} Let $B=\{x\in {{\mathbb R}^d}: 0<|x|\le 1\}$. We write $f\in F$ if \begin{equation}\label{e.comp} f(y)=\int_B \rho_1(x,y)\mu(\!\,{\rm d} x) \end{equation} for some subprobability measure $\mu$ concentrated on $B$. Then, by \eqref{e.CKfrho} and \eqref{ss form rho}, \begin{align}\nonumber P^{k}f(u) &= \int_B (e^{-kt})^{\frac{2\delta-d}{\alpha}} \int_{{\mathbb R}^d} \rho_1(x,y) \rho_{e^{kt}-1}(y,e^{kt/\alpha}u) h^2(y) \,{\rm d} y \mu(\!\,{\rm d} x)\\ \nonumber &= \int_B (e^{-kt})^{\frac{2\delta-d}{\alpha}} \rho_{e^{kt}}(x,e^{kt/\alpha}u)\mu(\!\,{\rm d} x)\\ \nonumber &= \int_B \rho_1\left(e^{-kt/\alpha}x, u\right)\mu(\!\,{\rm d} x)\\ \label{e.pts3} &= \int_B \rho_1(x, u)\tilde \mu(\!\,{\rm d} x)\approx (1+|u|)^{-d-\alpha+\delta}, \qquad u \in {{\mathbb R}^d_0}, \end{align} where $\tilde \mu$ is a subprobability measure concentrated on $e^{-kt/\alpha}B\subset B$. Thus, $P^k F\subset F$. We note that the comparison \eqref{e.pts3} is independent of $k$, see \eqref{eq:mainThmEstrho2}. We next argue that the hypotheses of \cite[Theorem~3.1]{MR1162571} hold true for $P$ and $F$: (C1) -- the lower bound, and (C2) -- the uniform absolute continuity. Indeed, by \eqref{e.pts3}, $$\liminf_{n\to \infty}\frac1n\sum_{k=1}^n \int_B P^k f(u) h^2(u)\,{\rm d} u \approx \int_B (1+|u|)^{-d-\alpha+ \delta} h^2(u)\,{\rm d} u >0,$$ which yields (C1) in \cite{MR1162571}. Also, (C2) therein is satisfied because $$\int_A P^kf(y)h^2(y) \,{\rm d} y \le C\int_A (1+|y|)^{-d-\alpha+ \delta}h^2(y) \,{\rm d} y \to 0,$$ uniformly in $k$ as $\int_A h^2(y) \,{\rm d} y\to 0$, due to the integrability of $(1+|y|)^{-d-\alpha+2\delta}$. By \cite[Theorem~3.1]{MR1162571}, a density $\varphi$ exists satisfying $L_t\varphi=P\varphi=\varphi$. Moreover, the density $\varphi$ is unique. Indeed, if $\psi$ is a probability density with respect to $h^2(x)\,{\rm d} x$ and $P\psi=\psi$, then $r:=\varphi-\psi$ satisfies $Pr=r$, too. If $r=0$ $a.e.$, then we are done, otherwise $\int_{{\mathbb R}^d} r_+(x) h^2(x) \,{\rm d} x =\int_{{\mathbb R}^d} r_-(x) h^2(x) \,{\rm d} x >0$ and we observe what follows. On the one hand $|Pr|=|r|$, on the other hand for $a.e.$ $x\in {{\mathbb R}^d}$ we have \begin{align*} Pr(x)&=P r_+(x)-Pr_-(x)\\ &= \int_{{\mathbb R}^d} r_+(y) l_t(y,x) h^2(y) \,{\rm d} y -\int_{{\mathbb R}^d} r_-(y) l_t(y,x) h^2(y) \,{\rm d} y \end{align*} and both terms are nonzero, because $l_t$ is positive. Then, $$|Pr(x)|<P r_+(x)\vee Pr_-(x)\le P|r|(x).$$ By this and \eqref{e.iso-1}, \begin{align*} \int_{{\mathbb R}^d} |r(x)|h^2(x) \,{\rm d} x &=\int_{{\mathbb R}^d} |Pr(x)|h^2(x) \,{\rm d} x <\int_{{\mathbb R}^d} P|r|(x)h^2(x) \,{\rm d} x =\int_{{\mathbb R}^d} |r(x)| h^2(x) \,{\rm d} x . \end{align*} We obtain a contradiction, therefore $r=0$, and $\psi=\varphi$ in $L^1(h^2)$. In passing we note that the above argument is a part of the proof of Doob's theorem \cite[Theorem~4.2.1]{MR1417491}. Since the operators $L_t$, $t>0$, commute, they have the same stationary density, that is $\varphi$. Indeed, if $s>0$ and $Q=L_s$, then $P(Q\varphi)=QP\varphi=(Q\varphi)$, and $Q\varphi$ is a density, so by the uniqueness, $Q\varphi=\varphi$. \end{proof} In passing we like to refer the interested reader to additional literature on the existence and uniqueness of stationary densities and measures. The book of Foguel \cite{MR0261686} gives a concise introduction to ergodic theory of Markov processes, in particular to the $L^1$ setting. The stationary measures and densities are discussed in Da Prato and Zabczyk \cite[Remark~3.1.3]{MR1417491}, \cite[Theorem~3.1]{MR1265226}, Lasota \cite[Theorem~6.1]{MR1452617}, and Komorowski, Peszat, Szarek \cite{MR2663632}. A nice presentation of asymptotic stability (and periodicity) of Markov operators is given by Komornik \cite{MR826761}, see also Lasota and Mackey \cite[p. 373, 11.9.4]{MR1244104}, Komorowski and Tyrcha \cite{MR1101473}, and Stettner \cite{MR1810695}. We note that the existence of a stationary density in our setting would also follow from the -- perhaps more constructive -- approach using the weak compactness of the set of functions given by \eqref{e.comp} and Schauder-Tychonoff fixed point theorem, see, e.g., Rudin \cite[5.28 Theorem]{MR1157815}, see also \cite[Section~3.8]{MR1157815}, the Dunford-Pettis theorem in Voigt \cite[Chapter~15]{MR4182424} and \eqref{e.pts3} above. We could also use the Krylov-Bogolioubov theorem, see, e.g., \cite{MR1244104} or the minicourse of Hairer \cite{mc2021MH}; and it is always worthy to check with Doob \cite{MR25097}, especially that his example (0.9) touches upon the classical Ornstein-Uhlenbeck semigroup. In what follows, $\varphi$ denotes the stationary density for the operators $L_t$, $t>0$ (in Lemma \ref{l.ext} below we verify that $\varphi$ can be defined pointwise so as to be continuous). \subsection{Asymptotic stability} The state space ${{\mathbb R}^d_0} $ is a Polish space and $l_t$ is positive. By Theorem~\ref{c.gsall} and the results of Kulik and Scheutzov \cite[Theorem~1 and Remark~2]{MR3345324}, for every $x\in {{\mathbb R}^d_0} $ we get \begin{equation}\label{e.convL} \int_{{\mathbb R}^d} \left|l_t(x,y) -\varphi(y)\right|h^2(y) \,{\rm d} y \to 0 \mbox{ as } t\to \infty. \end{equation} This is the (large-time) asymptotic stability of the semigroup aforementioned in the title of this subsection. Let $A$ be a bounded subset of ${{\mathbb R}^d_0}$ and $x,x_0\in A$. By Theorem~\ref{c.gsall} and \rf{eq:mainThmEstrho}, \begin{align} \int_{{\mathbb R}^d} \left|l_{1+t}(x,y) -\varphi(y)\right|h^2(y) \,{\rm d} y &=\int_{{\mathbb R}^d} \left| \int_{{\mathbb R}^d} l_1(x,z) \left(l_{t}(z,y) -\varphi(y)\right) h^2(z) \,{\rm d} z \right|h^2(y) \,{\rm d} y \nonumber \\ & \le c\int_{{\mathbb R}^d} l_1(x_0,z) \int_{{\mathbb R}^d} \left|l_{t}(z,y) -\varphi(y)\right| h^2(y) \,{\rm d} y\ h^2(z)\,{\rm d} z.\label{e.li} \end{align} We have $I_t(z):=\int_{{\mathbb R}^d} \left|l_{t}(z,y) -\varphi(y)\right| h^2(y) \,{\rm d} y \to 0$ as $t\to \infty$. Furthermore, for every $z\in {{\mathbb R}^d_0}$, $I_t(z)\le \int_{{\mathbb R}^d} \left(l_{t}(z,y) +\varphi(y)\right) h^2(y) \,{\rm d} y=2$. Of course, $\int_{{\mathbb R}^d} 2 l_1(x_0,z) \ h^2(z)\,{\rm d} z=2<\infty$. By the dominated convergence theorem, the iterated integral in \eqref{e.li} converges to $0$, therefore the convergence in \eqref{e.convL} is uniform for $x\in A$. In terms of $\rho$, \eqref{e.convL} reads as follows: uniformly in $x\in A$, \begin{equation}\label{e.conv} \int_{{\mathbb R}^d} \left|\rho_{1-e^{-t}}(e^{-t/\alpha}x,y) -\varphi(y)\right|h^2(y) \,{\rm d} y \to 0 \mbox{ as } t\to \infty. \end{equation} As a consequence we obtain the following spatial convergence in $L^1(h^2)$. \begin{lemma}\label{l.crho} We have $\int_{{\mathbb R}^d} \left|\rho_{1}(x,y) -\varphi(y)\right|h^2(y) \,{\rm d} y \to 0$ as $x\to 0$. \end{lemma} \begin{proof} It suffices to make cosmetic changes to \eqref{e.conv}. Of course, $e^{-t/\alpha}x\to 0$ and $1-e^{-t}\to 1$ as $t\to \infty$. By scaling, $$\rho_{1-e^{-t}}(e^{-t/\alpha}x,y)=(1-e^{-t})^{(2\delta-d)/\alpha}\rho_1\big((e^t-1)^{-1/\alpha}x,(1-e^{-t})^{-1/\alpha}y\big),$$ so \begin{align}\label{rho_et} \int_{{\mathbb R}^d} \left|\rho_1\big((e^t-1)^{-1/\alpha}x,(1-e^{-t})^{-1/\alpha}y\big) -\varphi(y)\right|h^2(y) \,{\rm d} y \to 0 \mbox{ as } t\to \infty. \end{align} By the continuity of dilations on $L^1(\!\,{\rm d} x)$, \begin{align*} \int_{{\mathbb R}^d} \left|\varphi\big((1-e^{-t})^{-1/\alpha}y\big) h^2(y)-\varphi(y)h^2(y)\right| \,{\rm d} y \to 0 \mbox{ as } t\to \infty. \end{align*} Using the triangle inequality and changing variables in \rf{rho_et} we get \begin{align*} \int_{{\mathbb R}^d} \left|\rho_1\big((e^{t}-1)^{-1/\alpha}z,y\big) -\varphi(y)\right|h^2(y) \,{\rm d} y \to 0 \mbox{ as } t\to \infty \end{align*} uniformly in $z\in A$ for bounded $A\subset {{\mathbb R}^d_0}$. We take $A$ as the unit sphere, for $x\in {{\mathbb R}^d_0}$ we write $x=(e^t-1)^{-1/\alpha}z$, where $t=\ln (1+|x|^{-\alpha})$ and $z=x/|x|\in A$, and we obtain the result. \end{proof} By Lemma~\ref{l.crho} and \eqref{eq:mainThmEstrho2} we obtain the following estimate. \begin{corollary}\label{c.bphi} $\varphi(y)\approx (1+|y|)^{-d-\alpha+\delta}$ for almost all $y\in {{\mathbb R}^d}$. \end{corollary} By the next result we may actually consider ${\varphi}$ as defined pointwise. \begin{lemma}\label{l.ext} After modification on a set of Lebesgue measure zero, $\varphi$ is a continuous radial function on ${{\mathbb R}^d}$ and $\varphi(y)\approx (1+|y|)^{-d-\alpha+\delta}$ for all $y\in {{\mathbb R}^d}$. \end{lemma} \begin{proof}By Theorem~\ref{c.gsall}, ${\varphi}=L_1{\varphi}$ ($a.e.$), so it suffices to prove that $L_1 {\varphi}$ is $a.e.$ equal to a continuous function on ${{\mathbb R}^d}$. By \eqref{e.crt1t2} and \eqref{eq:mainThmEstrho2}, $c(1+|y|)^{-d-\alpha+\delta}$ is an integrable majorant of $l_1(x,y)$ for bounded $x$. Furthermore, ${\varphi}$ is essentially bounded ($a.e.$), by Corollary~\ref{c.bphi}. The function $l_1(x,y)$ is continuous in $y\in {{\mathbb R}^d_0}$ so, by the dominated convergence theorem, $L_1{\varphi}(y)$ is continuous for $y\in {{\mathbb R}^d_0}$. It remains to prove the convergence of $L_1{\varphi}(y)$ to a finite limit as $y\to 0$. We let $t>0$ and $y \to 0$. By scaling, changing variables, the symmetry of $\rho_1$ and Lemma~\ref{l.crho}, \begin{align} \nonumber L_t {\varphi}(y)&=\int_{{\mathbb R}^d_0} \rho_{1-e^{-t}}\big(e^{-t/\alpha}x,y\big){\varphi}(x)h^2(x) \,{\rm d} x \\\nonumber &=\int_{{\mathbb R}^d_0} e^{t(d-2\delta)/\alpha}\rho_{1}\big(z, (1-e^{-t/\alpha})^{1/\alpha}y\big){\varphi}\big((e^t-1)^{1/\alpha}z\big)h^2(z) \,{\rm d} z \\\nonumber &\to \int_{{\mathbb R}^d_0} e^{t(d-2\delta)/\alpha}{\varphi}(z){\varphi}\big((e^t-1)^{1/\alpha}z\big)h^2(z) \,{\rm d} z \\ &=\int_{{\mathbb R}^d_0} {\varphi}((e^{-t})^{1/\alpha}x){\varphi}\big((1-e^{-t})^{1/\alpha}x\big)h^2(x) \,{\rm d} x <\infty.\label{e.dvf0} \end{align} Since $-2d-2\alpha+2\delta<-d-3\alpha<-d$, the finiteness in \eqref{e.dvf0} follows from Corollary \ref{c.bphi}. This proves the continuity of the extension of $L_1{\varphi}$ on the whole of ${{\mathbb R}^d}$. The rest of the lemma follows from the continuity, Lemma~\ref{l.crho} and \eqref{risom}, and from Corollary~\ref{c.bphi}. \end{proof} Needless to say, the continuous modification of ${\varphi}$ is unique and pointwise defined for every $x\in {{\mathbb R}^d}$. From now on the extension shall be denoted by ${\varphi}$. By the equality in \eqref{e.dvf0}, \begin{equation}\label{e.wvf0} {\varphi}(0)=\int_{{\mathbb R}^d_0} {\varphi}\big(\lambda^{1/\alpha}x\big){\varphi}\big((1-\lambda)^{1/\alpha}x\big)h^2(x) \,{\rm d} x \end{equation} for every $\lambda \in [0,1]$, including the endpoint cases, since $\varphi$ is a density. In particular, \begin{equation}\label{e.wvf02} {\varphi}(0)=\int_{{\mathbb R}^d} {\varphi}(2^{-1/\alpha}x)^2 h^2(x) \,{\rm d} x . \end{equation} \subsection{Regularization of $\rho$} We are now in a position to prove convergence of $\rho_1(x,y)$ as $x\to 0$ to a finite limit. By scaling, Chapman-Kolmogorov, Lemma~\ref{l.crho} and the boundedness of ${\varphi}$, for $y\in {{\mathbb R}^d_0}$ and ${{\mathbb R}^d_0}\ni x\to 0$ we get \begin{eqnarray*} \rho_1(x,y)&=&2^\frac{d-2\delta}{\alpha}\rho_2(2^{1/\alpha}x,2^{1/\alpha}y)\\ &=&2^\frac{d-2\delta}{\alpha}\int \rho_1(2^{1/\alpha}x,z)\rho_1(z,2^{1/\alpha}y)h^2(z) \,{\rm d} z \\ &\to& 2^\frac{d-2\delta}{\alpha}\int {\varphi}(z)\rho_1(z,2^{1/\alpha}y)h^2(z) \,{\rm d} z \\ &=&\int {\varphi}(z)\rho_{1/2}(2^{-1/\alpha}z,y)h^2(z) \,{\rm d} z =L_{\ln 2}{\varphi}(y)={\varphi}(y). \end{eqnarray*} Thus, for every $y\neq 0$, \begin{equation}\label{rho_lim} \rho_1(0,y):=\lim_{x\to 0} \rho_1 (x,y) ={\varphi}(y). \end{equation} By scaling, for all $t>0$ we get \begin{equation}\label{e.srty} \rho_t(0,y) := \lim_{x\to 0} \rho_t (x,y)= t^{\frac{2\delta -d}{\alpha}}\rho_1\left(0, t^{-1/{\alpha}}y\right)=t^{\frac{2\delta-d}{\alpha}} {\varphi} \left(t^{-1/\alpha}y \right). \end{equation} By \eqref{e:rhois1}, for $x\not=0$ we have \begin{align*} \int_{{{\mathbb R}^d}} \rho_1(x,y)|y|^{-2\delta} \,{\rm d} y = 1. \end{align*} By \rf{eq:mainThmEstrho2} and \rf{rho_lim}, \begin{equation}\label{rhosim1} \rho_{1} (x,y)|y|^{-2\delta} \approx (1+|y|)^{-d-\alpha+\delta}|y|^{-2\delta}, \quad |x|<1, \ y\neq 0. \end{equation} Thus, applying \rf{rho_lim} and the dominated convergence theorem, we get \begin{equation}\label{est_t} \int_{{{\mathbb R}^d}} \rho_t(0,y)h^2(y) \,{\rm d} y =1,\quad t>0. \end{equation} By the symmetry of $\rho_t$ and \eqref{e.srty}, for all $t>0$, $x\in {{\mathbb R}^d_0}$ we also have \begin{equation}\label{e.srtx} \rho_t(x,0) := \lim_{y\to 0} \rho_t (x,y)=\rho_t(0,x)= t^{\frac{2\delta -d}{\alpha}}\rho_1\left(t^{-1/{\alpha}}x,0\right)=t^{\frac{2\delta-d}{\alpha}}{\varphi} \left( t^{-1/\alpha}x\right). \end{equation} \begin{lemma}\label{Lemma_rho_lim_0} The function $\rho$ has a unique continuous positive extension to $(0,\infty)\times {{\mathbb R}^d}\times {{\mathbb R}^d}$. \end{lemma} \begin{proof} By Chapman-Kolmogorov and the symmetry of $\rho_1$, for $x,y\neq 0$, we have \begin{align}\label{rho1_2} \rho_1(x,y) = \int_{{{\mathbb R}^d}} \rho_{\nicefrac{1}{2}} (z,x) \rho_{\nicefrac{1}{2}}(z,y) \,{\rm d} z . \end{align} Recall that $\rho_1$ is continuous on ${{\mathbb R}^d_0}\times {{\mathbb R}^d_0}$. This and \rf{e.srty} yield for all $z\in {{\mathbb R}^d_0}$ that $\rho_{\nicefrac{1}{2}} (z,x) \to \rho_{\nicefrac{1}{2}} (z,x_0)$ and $\rho_{\nicefrac{1}{2}} (z,y)\to \rho_{\nicefrac{1}{2}} (z,y_0)$ if $x\to x_0\in {{\mathbb R}^d}$ and $y\to y_0\in {{\mathbb R}^d}$, hence \begin{equation}\label{d.r100} \rho_1 (x,y) \to \int_{{{\mathbb R}^d}}\rho_{\nicefrac{1}{2}} (z,x_0)\rho_{\nicefrac{1}{2}} (z,y_0) \,{\rm d} z , \end{equation} by the dominated convergence theorem, since for bounded $x,y\neq 0$, by and \rf{e.crt1t2} and \rf{eq:mainThmEstrho2}, \begin{equation}\label{rhosim} \rho_{\nicefrac{1}{2}} (z,x)\rho_{\nicefrac{1}{2}} (z,y) \approx (1+|z|)^{-2d-2\alpha+2\delta}. \end{equation} The latter function is integrable because $-2d-2\alpha+2\delta<-d-3\alpha<-d$. Furthermore, \eqref{rhosim} implies that the limit in \eqref{d.r100} is positive and, by \rf{rho1_2}, it is an extension of $\rho_1$ to ${{\mathbb R}^d}\times {{\mathbb R}^d}$, which we shall denote by $\rho_1$ again. In view of \rf{ss form rho}, \begin{equation}\label{e.et} t^{\frac{2\delta-d}{\alpha} } \rho_1\big(t^{-1/\alpha }x , t^{-1/\alpha }y \big), \quad x,y\in {{\mathbb R}^d},\quad t>0, \end{equation} is finite and defines a continuous extension of $\rho_t$ for each $t>0$. The extension is clearly positive and jointly continuous in $t,x,y$. It is also unique, since $(0,\infty)\times {{\mathbb R}^d_0}\times {{\mathbb R}^d_0}$ is dense in $(0,\infty)\times {{\mathbb R}^d}\times {{\mathbb R}^d}$. \end{proof} We recall the reader that the boundedness of $\rho_1$ near the origin readily follows from \eqref{eq:mainThmEstrho}. On the other hand the continuity was turned out difficult for us to capture directly or indirectly. For instance we considered \cite{MR4164845} in this connection. The proof presented above seems to be underpinned by self-regularization of functions satisfying (implicit) integral equations. In what follows, $\rho$ will denote the continuous extension of $\rho$ to $(0,\infty)\times {{\mathbb R}^d}\times {{\mathbb R}^d}$. \begin{corollary} $\rho_1(0,0)= \lim_{x,y\to 0} \rho_1(x,y)\in (0,\infty)$. \end{corollary} The following Chapman-Kolmogorov equations hold, so $\rho$ is a transition density on ${{\mathbb R}^d}$. \begin{corollary}\label{c.CH} For all $s,t>0$, $x,y\in {{\mathbb R}^d}$, we have $ \int_{{\mathbb R}^d} \rho_s(x,z)\rho_t(z,y) h^2(z) \,{\rm d} z =\rho_{s+t}(x,y). $ \end{corollary} \begin{proof} Since $(1+|z|)^{-2d-2\alpha+2\delta}|z|^{-2\delta}$ is integrable, we can use \rf{rho_lim} and \rf{e.CKfrho}. \end{proof} \begin{proof}[Proof of Theorem~\ref{t:eta}] The first statement of the theorem is proved in Lemma~\ref{Lemma_rho_lim_0}, \eqref{ss form eta} is proved more generally as \eqref{e.et}, and \eqref{e:ssf} is proved more generally in Corollary~\ref{c.CH}. See also \eqref{rho_lim} and \eqref{e.srty} for a more explicit expression of \eqref{e.det}. \end{proof} \begin{proof}[Proof of Theorem~\ref{t:eta2}] The result is immediate from \eqref{e:drhot} and \eqref{e.det}. \end{proof} We can now extend \eqref{eq:mainThmEstrhot} as follows. \begin{corollary}\label{cor:mainThmEstrhot} \begin{align*} \rho_t(x,y)\approx \left( t^{-d/\alpha}\wedge \frac{t}{|x-y|^{d+\alpha}} \right)\left(t^{\delta/\alpha}+|x|^{\delta} \right)\left(t^{\delta/\alpha}+|y|^{\delta} \right) ,\quad t>0, \ x,y \in {{\mathbb R}^d}. \end{align*} \end{corollary} \begin{example} For $\kappa=0$ we have $H=h=1$, $\rho=\tilde p=p$, $l_t(x,y)=p_{1-e^{-t}}(e^{-t/\alpha}x,y)$, $\varphi(x)=p_1(0,x)$, and $\Psi_t(x)=p_t(0,x)$. \end{example} \section{Functional analysis of $\tilde P_t$}\label{s.fatp} We recall that $\tilde P_t{\varphi} (x) := \int_{{\mathbb R}^d} \tilde p(t,x,y) {\varphi} (y) \,{\rm d} y$. By \eqref{eq:2}, $\tilde P_t h = h$ for all $t>0$. \begin{lemma}\label{contraction} $\{\tilde P_t\}_{t>0}$ is a contraction semigroup on $L^1(h)$ and for every $f\in L^1(h)$, we have \begin{equation}\label{e.cl} \int_{{\mathbb R}^d} \tilde P_t f(x) h(x)\,{\rm d} x=\int_{{\mathbb R}^d} f(x)h(x)\,{\rm d} x,\quad t>0. \end{equation} \end{lemma} \begin{proof} The semigroup property of $\tilde P_t$ follows from \eqref{eq:ChK}. Let $ f \ge 0$ be a measurable function. By Fubini-Tonelli, the symmetry of $\tilde p$ and \eqref{eq:2}, \begin{align*} \int_{{\mathbb R}^d} \tilde P_t f (x) h(x) \,{\rm d} x & = \int_{{\mathbb R}^d} \int_{{\mathbb R}^d} \tilde p(t, x,y) f (y) h(x) \,{\rm d} y \,{\rm d} x = \int_{{\mathbb R}^d} h(y) f (y) \,{\rm d} y. \end{align*} For arbitrary $ f \in L^1(h)$ we write $ f = f _+- f _-$ and use the nonnegative case. \end{proof} \begin{lemma}\label{contr.L1} $\{P_t\}_{t>0}$ is a strongly continuous contraction semigroup on $L^1(h)$. \end{lemma} \begin{proof} Since $p\le\tilde{p}$, for nonnegative function $ f $ and $t>0$ we get $\|P_t f \|_{L^1(h)}\le \|\tilde{P}_t f \|_{L^1(h)}\le \| f \|_{L^1(h)}$, so the contractivity follows as in the proof of Lemma~\ref{contraction}. By the contractivity and the semigroup property of $P_t$, it suffices to prove strong continuity as $t\to 0$. Let $ f \in L^1(h)$. Then $g:= f h\in L^1$. There are functions $g_n \in C_c^{\infty } \left( {{\mathbb R}^d_0} \right)$ such that $\|g-g_n\|_{L^1} \rightarrow 0$ as $n\rightarrow \infty$. To this end let $ f _n=g_n/h$. Of course, $ f _n\in C_c^\infty({{\mathbb R}^d_0})$ and $\| f - f _n\|_{L^1(h)}=\|g-g_n\|_{L^1}$ for every $n$, and, by the contractivity, we have \begin{align*} \|P_t f - f \|_{L^1(h)} & \le \| P_t f -P_t f _n\|_{L^1(h)} + \| P_t f _n - f _n\|_{L^1(h)}+\| f _n - f \|_{L^1(h)} \\ &\le 2\| f - f _n\|_{L^1(h)} + \| P_t f _n - f _n\|_{L^1(h)}. \end{align*} Therefore, it suffices to prove that $ \|P_t f - f \|_{L^1(h)}\to 0$ as $t\to 0$ for every $ f \in C_c^{\infty } \left( {{\mathbb R}^d_0} \right)$. To this end take $K\in (0,\infty)$ such that $\text{supp } f \subset B(0, K/2)$. Then, \begin{align*} \|P_t f &- f \|_{L^1(h)} \le \int_{|x| \le K} | P_t f (x) - f (x) | h(x) \,{\rm d} x + \int_{|x| >K} P_t| f | (x)h(x) \,{\rm d} x. \end{align*} When $t\to 0$, the above converges to $0$ because $P_t f \to f $ uniformly, $h$ is locally integrable and $P_t| f |(x)\le ct(1+|x|)^{-d-{\alpha}}$ on $B(0,K)^c$, see \eqref{eq:oppt}. The proof is complete. \end{proof} \begin{lemma}\label{strong1} $\{\tilde P_t\}_{t>0}$ is a strongly continuous contraction semigroup on $L^1(h)$. \end{lemma} \begin{proof} The contractivity is resolved in Lemma~\ref{contraction}. Since $P_t$ is strongly continuous as $t\to 0$, by \eqref{eq:Df2} it suffices to consider nonnegative $ f \in L^1(h)$ and verify that $$ I_t := \int_{{\mathbb R}^d} \int_{{\mathbb R}^d} \int_0^t \int_{{\mathbb R}^d} \tilde p(s, x,z) q(z)p({t-s},z,y) f (y)\,{\rm d} z \,{\rm d} s \,{\rm d} y h(x) \,{\rm d} x \to 0 $$ as $t\to 0$. By \eqref{eq:2} and \cite[Lemma 3.3]{MR3933622}, \begin{align*} I_t&=\int_{{\mathbb R}^d} \int_0^t \int_{{\mathbb R}^d} h(z)q(z)p({t-s}, z,y) f (y) \,{\rm d} z \,{\rm d} s \,{\rm d} y \\ &=\int_{{\mathbb R}^d} \big(h(y)- P_t h(y)\big) f (y) \,{\rm d} y =\|f\|_{L^1(h)}-\|P_t f \|_{L^1(h)} \to 0, \end{align*} indeed, because of Lemma~\ref{contr.L1}. \end{proof} We consider the resolvent operators $R_\lambda = \int_0^{\infty } e^{-\lambda t}P_t \,{\rm d} t$ and $\tilde R_\lambda = \int_0^{\infty } e^{-\lambda t}\tilde{P}_t \,{\rm d} t$ for $\lambda>0$. The contractivities of $P_t$ and $\tilde{P}_t$ yield the following result. \begin{corollary} $\lambda \tilde R_\lambda $ and $\lambda R_\lambda $ are contractions on $L^1(h)$. \end{corollary} Let ${\mathcal L}$ be the $L^1(h)$-generator of the semigroup $P_t$, with domain ${\mathcal D}$, and let $\tilde{\mathcal L}$ be the $L^1(h)$-generator of the semigroup $\tilde P_t$, with domain $\tilde {\mathcal D}$. The following result is consistent with \eqref{eq:SchrOp}. \begin{proposition} We have ${\mathcal D}\subset \tilde {\mathcal D}$ and $\tilde {\mathcal L} f = {\mathcal L} f + q f $ for $ f \in {\mathcal D}$. \end{proposition} \begin{proof} Let $\lambda>0$. Recall that $ f \in {\mathcal D}$ if and only if $ f = R_\lambda g$ for some (unique) $g\in L^1(h)$ and then $ {\mathcal L}\varphi = \lambda R_\lambda g-g$. The same holds for $\tilde R_\lambda $. By integrating \eqref{eq:Df2} with respect to $e^{-\lambda t} \,{\rm d} t$ and using Fubini-Tonelli and the following self-explanatory notation we get \begin{align*} \tilde R_\lambda (x,y) &= R_\lambda (x, y) + \int_0^{\infty } e^{-\lambda t } \int_0^t \int_{{\mathbb R}^d} \tilde p(s,x,z) q(z) p(t-s,z,y) \,{\rm d} z \,{\rm d} s\,{\rm d} t \\ & = R_\lambda (x, y) + \int_0^{\infty } \int_s^{\infty }\int_{{\mathbb R}^d} e^{-\lambda s }\tilde p(s,x,z) q(z) e^{-\lambda (t -s)}p(t-s,z,y) \,{\rm d} z \,{\rm d} t\,{\rm d} s \\ & = R_\lambda (x, y) + \int_{{\mathbb R}^d}\tilde R_{\lambda } (x, z) q(z) R_{\lambda } (z, y) \,{\rm d} z,\quad x,y \in {{\mathbb R}^d}. \end{align*} Accordingly, for $ f \ge 0$, we get $ \tilde R_{\lambda } f = R_{\lambda } f + \tilde R_{\lambda } q R_{\lambda } f $. In particular, $\tilde R_{\lambda } q R_{\lambda } f \in L^1(h)$ if $ f \in L^1(h)$. Moreover, $q R_{\lambda }$ is a bounded operator on $L^1(h)$. Indeed, if $0\le f \in L^1(h)$, then using Fubini-Tonelli, the identity $e^{-\lambda t}=\int_t^\infty \lambda e^{-\lambda s}ds$ and \cite[Lemma 3.3]{MR3933622}, we get \begin{align*} 0&\le \int_{{\mathbb R}^d} q(x)R_\lambda f (x)h(x) \,{\rm d} x = \int_{{\mathbb R}^d} \int_{{{\mathbb R}^d}}\int_0^\infty e^{-\lambda t} p(t,x,y) q(x)h(x) f (y) \,{\rm d} t\,{\rm d} y \,{\rm d} x \\ &= \int_0^\infty \int_{{\mathbb R}^d} \lambda e^{-\lambda s} f (y) \left(\int_0^s\int_{{{\mathbb R}^d}} p(t,x,y) q(x)h(x) \,{\rm d} x \,{\rm d} t \right) \,{\rm d} y \,{\rm d} s \\ &\le \int_0^\infty \lambda e^{-\lambda s}\,{\rm d} s \int_{{\mathbb R}^d} h(y) f (y) \,{\rm d} y = \int_{{\mathbb R}^d} h(y) f (y) \,{\rm d} y. \end{align*} It follows that $R_{\lambda } = \tilde R_{\lambda } (I - qR_{\lambda })$ on $L^1(h)$, and so \begin{equation*} {\mathcal D} = R_{\lambda } \big( L^1(h) \big) \subset \tilde R_{\lambda } \big( L^1(h)\big) = \tilde{\mathcal D}. \end{equation*} To prove that $\tilde {\mathcal L} \varphi ={\mathcal L}\varphi +q$ for $\varphi \in \mathcal{D}$, we let $\psi ={\mathcal L}\varphi$, $\varphi, \psi \in L^1(h)$, that is $\varphi =\lambda R_\lambda f -R_\lambda \psi$. It is enough to prove that $\tilde {\mathcal L} f = \psi +q f $, that is \begin{equation}\label{vf_psi} f = \lambda \tilde R_\lambda f -\tilde R_\lambda q f . \end{equation} But the right-hand-side of \rf{vf_psi} is \begin{align*} \lambda R_\lambda f +\tilde R_\lambda \lambda q R_\lambda f -R_\lambda \psi -R_\lambda qR_\lambda \psi -\tilde R_\lambda q f = f +\tilde R_\lambda q f -\tilde R_\lambda q f = f . \end{align*} The proof is complete. \end{proof} Recall the notation of \rf{normqh}. Our proof of the large-time asymptotics of $\tilde P_t$ hinges on the following hypercontractivity result (for $q=1$ see Lemma \ref{contraction} and for $q=\infty$ see Lemma~\ref{l.qinf}). \begin{theorem}\label{hyperc} If $1<q< \infty $ then for all $t>0$ and nonnegative functions $f$ on ${{\mathbb R}^d}$, \begin{equation}\label{w norm} \|\tilde P_t f\|_{q,h} \le Ct^{-\frac{d-2\delta}{\alpha}(1-\frac{1}{q})} \|f\|_{L^1(h)}+Ct^{-\frac{d-2\delta}{\alpha}(1-\frac{1}{q})-\frac{\delta}{q\alpha}}\|f\|_{L^1({\mathbb R}^d)}. \end{equation} \end{theorem} \begin{proof} We first prove the inequality \rf{w norm} for $t=1$. Let $q'=\frac{q}{q-1}$. It is enough to verify that for every function $g\ge 0$ on ${\mathbb R}^d$, \begin{align}\label{observation} \int_{{\mathbb R}^d} \tilde P_1 f(x)g(x) h^{2-q}(x) \,{\rm d} x \le C \|f\|_{L^1(H)} \|g\|_{L^{q'}(h^{2-q})}, \end{align} since then, by {the duality of $L^p(h^{2-q})$}, we get \begin{equation}\label{e.op1} \|\tilde P_1 f\|_{q,h} =\|\tilde P_1 f\|_{L^q(h^{2-q})}=\sup_{ \|b\|_{L^{q'}(h^{2-q})}=1}\left|\int_{{\mathbb R}^d} \tilde P_1 f(x)b(x)h^{2-q}(x)\,{\rm d} x\right| \le C\|f\|_{L^1(H)}, \end{equation} as needed. To prove \rf{observation}, by \eqref{eq:mainThmEstz} it is enough to estimate \begin{align} I(f,g) &:= \int_{{\mathbb R}^d}\int_{{\mathbb R}^d} H(x)H(y)p(1,x,y)f(y)g(x)h^{2-q}(x) \,{\rm d} y \,{\rm d} x \nonumber\\ &=I(f\mathbf{1}_B,g\mathbf{1}_{B})+I(f\mathbf{1}_B,g\mathbf{1}_{B^c})+I(f\mathbf{1}_{B^c},g\mathbf{1}_{B})+I(f\mathbf{1}_{B^c},g\mathbf{1}_{B^c}), \label{e.4t} \end{align} where $B=B(0,1)\subset {\mathbb R}^d$. By H\"older inequality, \begin{align*} I(f\mathbf{1}_B,g\mathbf{1}_{B}) & \le \int_B \int_B h(x)h(y)p(1,x,y)f(y)g(x) \,{\rm d} y h(x)^{2-q} \,{\rm d} x \\ & \le c \left( \int_B h(y)f(y) \,{\rm d} y \right) \left( \int_B h(x)g(x) h^{2-q}(x)\,{\rm d} x \right)\\ & \le c \|f\|_{L^1(H)} \left(\int_B g^{q'}(x) h^{2-q}(x) \,{\rm d} x \right)^{\frac{1}{q'}}\left( \int_Bh^{2}(x) \,{\rm d} x \right)^{\frac{1}{q}}\\ & \le C\|f\|_{L^1(H)} \|g\|_{L^{q'}(h^{2-q})}, \end{align*} where $\int_Bh^{2}(x) \,{\rm d} x<\infty$, because $2\delta \le d-\alpha<d$. We next deal with the second term in \eqref{e.4t}, \begin{align} \label{last} \nonumber I(f\mathbf{1}_B,g\mathbf{1}_{B^c}) & \le \int_B \int_{B^c} f(y)h(y)p(1,x,y)g(x) h^{2-q}(x)\,{\rm d} x \,{\rm d} y \\ \nonumber& \le c\int_B \int_{B^c} f(y)h(y)|x|^{-d-\alpha}g(x) h^{2-q}(x)\,{\rm d} x \,{\rm d} y \\ \nonumber & \le c \|f\|_{L^1(H)}\int_{B^c} |x|^{-d-\alpha} g(x) h^{2-q}(x)\,{\rm d} x \\ &\le c\|f\|_{L^1(H)}\left( \int_{B^c} g^{q'}(x)h^{2-q}(x) \,{\rm d} x \right)^{\frac{1}{q'}}\left(\int_{B^c}|x|^{-(d+\alpha)q}h^{2-q}(x) \,{\rm d} x \right)^{\frac{1}{q}}\\ \nonumber& \le C\|f\|_{L^1(H)} \|g\|_{L^{q'}(h^{2-q})}. \end{align} The last integral in \rf{last} is finite because $0\le \delta \le (d-\alpha)/2$, so $|x|^{-(d+\alpha)q-\delta(2-q)} \le |x|^{-d-\alpha} \in L^1(B^c, \,{\rm d} x )$ for every $q\in (1,+\infty)$. We then estimate the third term, \begin{align*} I(f\mathbf{1}_{B^c},g\mathbf{1}_{B})& \le \int_{B^c}\int_B f(y)h(y)p(1,x,y)g(x) h^{2-q}(x) \,{\rm d} x \,{\rm d} y \\ & \le c \int_{B^c} \int_B f(y)h(y)|y|^{-d-\alpha} g(x) h^{2-q}(x)\,{\rm d} x \,{\rm d} y \\ & \le c \|f\|_{L^1(H)}\int_B g(x)h^{2-q}(x) \,{\rm d} x \\ & \le c\|f\|_{L^1(H)}\left(\int_B g^{q'}(x) h^{2-q}(x) \,{\rm d} x \right)^{\frac{1}{q'}}\left( \int_B h^{2-q}(x) \,{\rm d} x \right)^{\frac{1}{q}}, \end{align*} and the last integral is finite, as before. For the fourth term in \eqref{e.4t} we note that \begin{align*} \int_{{\mathbb R}^d} p(1,x,y)g(x)h^{2-q}(x)\,{\rm d} x & \le \left( \int_{{\mathbb R}^d} g^{q'}(x)h^{2-q}(x) \,{\rm d} x \right)^{\frac{1}{q'}} \left( \int_{{\mathbb R}^d} p^q(1,x,y) h^{2-q}(x) \,{\rm d} x \right)^{\frac{1}{q}}\\ & \le \|g\|_{L^{q'}(h^{2-q})}\left( \int_{{\mathbb R}^d} p_1^q (x) h^{2-q}(x) \,{\rm d} x \right)^{\frac{1}{q}}, \end{align*} by rearrangement inequalities \cite[Sec.3]{MR1817225}. The last integral is finite, because \begin{align*} p_1^q(x)h^{2-q}(x) & \le C|x|^{-\delta},\quad x\in B,\\ p_1^q(x)h^{2-q}(x) & \le C|x|^{-(d+\alpha)q -\delta(2-q)} \le |x|^{-d-\alpha}, \quad x\in B^c. \end{align*} Thus, \begin{align*} I(f\mathbf{1}_{B^c},g\mathbf{1}_{B^c})& \le \int_{B^c}\int_{B^c} p(1,x,y)f(y)g(x)h^{2-q}(x) \,{\rm d} x \,{\rm d} y \\ & \le C \|g\|_{L^{q'}(h^{2-q})}\int_{B^c} f(y)\,{\rm d} y \le C\|f\|_{L^1(H)} \|g\|_{L^{q'}(h^{2-q})}. \end{align*} Therefore \eqref{e.4t} is bounded above by $C\|f\|_{L^1(H)} \|g\|_{L^{q'}(h^{2-q})}$, which yields \eqref{w norm} for $t=1$. By this and a change of variables, for all $t>0$ we get \begin{align} \|\tilde P_t f\|_{q,h}&= t^{\frac{d-\delta(2-q)}{\alpha q}}\| \tilde P_1 f(t^{\nicefrac{1}{\alpha}}\cdot )\|_{q,h} \label{e.spt} \\ \nonumber & \le Ct^{\frac{d-\delta(2-q)}{\alpha q}}\left( \| f (t^{\nicefrac{1}{\alpha}}\cdot) \|_{L^1(h)}+\| f(t^{\nicefrac{1}{\alpha}}\cdot )\|_{L^1}\right)\\ \nonumber &=Ct^{\frac{d-\delta(2-q)}{\alpha q}}\left( t^{-\frac{d-\delta}{\alpha}}\|f\|_{L^1(h)}+t^{-\frac{d}{\alpha}}\|f\|_{L^1}\right)\\&=Ct^{-\frac{d-\delta}{\alpha }(1-\frac{1}{q})}\|f\|_{L^1(h)}+Ct^{-\frac{d-\delta}{\alpha }(1-\frac{1}{q})-\frac{\delta}{\alpha q}}\|f\|_{L^1}.\nonumber \end{align} The proof of \eqref{w norm} is complete. \end{proof} We note in passing that \eqref{w norm} for $t=1$ is equivalent to \rf{e.op1} but $H$ is incompatible with the scaling \eqref{e.spt}, so the long form of \eqref{w norm} seems inevitable. Here is a more trivial bound. \begin{lemma}\label{l.qinf} $\|\tilde P_t f\|_{\infty,h}\le \|f\|_{\infty,h}$. \end{lemma} \begin{proof} For $x\in {{\mathbb R}^d_0}$ we have $|\tilde P_t f (x)\,{\rm d} x|\le \|f\|_{\infty,h}\int_{{\mathbb R}^d} \tilde P_t(x,y)h(y)\,{\rm d} y \le h(x)\|f\|_{\infty,h} $. \end{proof} \section{Asymptotic behavior for large time}\label{s.ssa} This section is devoted to the proof of Theorem \ref{lim norm th}. For the sake of comparison let us discuss the following classical analogue of \rf{equation}, \begin{equation}\label{lam p} \begin{cases} \partial_t u(x,t)=\left(\Delta +\kappa |x|^{-2}\right) u(x,t) , \quad x\in {\mathbb R}^d, \quad t>0 ,\\ u(x,0)=f(x) . \end{cases} \end{equation} where $d \ge 3$ and $\kappa \in {\mathbb R}$. As we already mentioned in Section \ref{s:Mm}, the Cauchy problem \rf{lam p} was popularized by Baras and Goldstein \cite{MR742415}, who discovered that \rf{lam p} has no positive local-in-time solutions if $\kappa >(d-2)^2/4$, which is called the instantaneous blow up. See also Goldstein and Kombe \cite{GK} for a simple proof via Harnack inequality. V\'azquez and Zuazua \cite{MR1760280} then studied the large time behavior of solutions to \rf{lam p} with $0<\kappa \le (d-2)^2/4$. Using a weighted version of the Hardy-Poincar\'e inequality, they proved in \cite[Theorem 10.3]{MR1760280} the stabilization of some solutions toward the following self-similar solution of \rf{lam p} \begin{equation*} V(x,t)=t^{\sigma -\frac{d}{2}}|x|^{-\sigma }{\text{\rm{e}}}^{-\frac{|x|^2}{4t}}, \end{equation*} where $\sigma=\frac{d-2}{2}-\sqrt{(d-2)^2/4- \kappa}$. Then, Pilarczyk \cite{MR3020137} proved that if $u=u(t,x)$ is the solution of \rf{lam p} with $\kappa \in \big(-\infty , (d-2)^2/4\big)$, then for every $1 \le q \le \infty$, \begin{equation*} \lim_{t\rightarrow \infty }t^{\frac{d}{2}(1-\frac{1}{q})-\frac{\sigma }{2}}\| u(\cdot ,t)-AV(\cdot , t)\|_{q, {\varphi}_\sigma (t)}=0. \end{equation*} Here $f \in L^1({\mathbb R}^d) \cap L^1(|x|^{-\sigma}\,{\rm d} x)$, \begin{equation*} A=\frac{\int_{{\mathbb R}^d} |x|^{-\sigma } f(x) \,{\rm d} x }{\int_{{\mathbb R}^d} |x|^{-2\sigma }{\text{\rm{e}}}^{-\frac{|x|^2}{4}} \,{\rm d} x }, \end{equation*} $ {\varphi}_\sigma (x,t)= 1 \vee ({\sqrt{t}}/{|x|}) ^\sigma$, and \begin{align*} \| g\|_{q,{\varphi}_\sigma (t)}= \begin{cases} \Big( \int_{{\mathbb R}^d}| g(x)|^q {\varphi}_{\sigma }^{2-q}(x,t) \,{\rm d} x \Big)^{1/q} & \textrm{for}\quad 1 \le q<\infty,\\ {\rm ess}\sup_{x \in {\mathbb R}^d}| g(x)|/{\varphi}_{\sigma }(x,t) &\text{for}\quad q=\infty. \end{cases} \end{align*} In Theorem \ref{lim norm th} we extend the results of V\'azquez and Zuazua \cite{MR1760280} and Pilarczyk \cite{MR3020137} to the Cauchy problem \eqref{equation}. In fact we also propose a novel ramification of the asymptotics, using simpler, time-independent, norms $\|f\|_{q,h }$ defined in \eqref{normqh}. We now return to the setting of \eqref{equation} and \rf{eq:dk0}. The solutions to \rf{equation} will be defined as $u(t,x)= \tilde P_t f(x)$, $t>0$, $x\in {{\mathbb R}^d}$ for $f\in L^1(H)$. By Theorem~\ref{t:eta2}, \begin{align*} \Psi_t(x):=\lim_{y \to 0} \frac{\tilde{p}(t,x,y)}{h(y)}=\rho_t(0,x)h(x) = t^{\frac{2\delta-d}{\alpha}}{\varphi}\big( t^{-1/\alpha}x\big)h(x),\quad t>0, \quad x\in {{\mathbb R}^d}. \nonumber \end{align*} In particular, $\Psi_t(0)=\infty$. Since, for $s>0$, \begin{align*} \rho_{t+s}(0,x)&=\int_{{\mathbb R}^d}\rho_t(0,y) \rho_s (y,x) h^2(y)\,{\rm d} y= \int_{{\mathbb R}^d}\rho_t(0,y) h(y) \tilde p (s,y,x)/h(x) \,{\rm d} y, \end{align*} we get the following evolution property \begin{align}\label{e.s} \int_{{\mathbb R}^d} \tilde p (s,y,x)\Psi_t(y) \,{\rm d} y =\Psi_{t+s}(x). \end{align} By \eqref{e.et} we also have \begin{equation}\label{ss form eta1} \Psi_t(x)= t^{\frac{\delta-d}{\alpha} } \Psi_1(t^{-1/\alpha }x), \quad t>0,\quad x\in {{\mathbb R}^d}. \end{equation} We summarize \eqref{e.s} and \eqref{ss form eta1} by saying that $\Psi_t(x)$ is a self-similar semigroup solution of \rf{equation}. By \rf{est_t}, \begin{equation}\label{psi_int} \int_{{{\mathbb R}^d}} \Psi _t (x) h(x)\,{\rm d} x =\int_{{{\mathbb R}^d}} \Psi_1 (x) h(x) \,{\rm d} x =1. \end{equation} Furthermore, $\Psi_t(x)$ is a mild solution of \eqref{equation} since it satisfies the following Duhamel formula. \begin{lemma}\label{lem:PsiDuh} \begin{align}\label{e.DfPsi} \Psi_t(x) &= \int_0^t\int_{{{\mathbb R}^d}} \Psi_r(z) q(z) p(t-r,x,z) \,{\rm d}{z}\,{\rm d}{r}, \quad t>0,\; x\in {{\mathbb R}^d}. \end{align} \end{lemma} \begin{proof} By \eqref{eq:mainThmEstz}, we have \begin{align*} \frac{\tilde{p}(t,x,y)}{h(y)} \approx (|y|^{\delta}+t^{\delta/\alpha})(1+t^{\delta/\alpha}|x|^{-\delta})p(t,x,y), \end{align*} hence \begin{align}\label{eq:u_est} \Psi_t(x) \approx t^{\delta/\alpha}(1+t^{\delta/\alpha}|x|^{-\delta}) \big(t^{-d/\alpha} \land \tfrac{t}{|x|^{d+\alpha}}\big), \qquad t>0,\; x \in {{\mathbb R}^d}. \end{align} Now, by \eqref{e.s} and \eqref{eq:Df1} for $0<s<t$ we obtain \begin{align} \Psi_t(x) &= \int_{\mathbb{R}^d} \tilde{p}(t-s,x,y) \Psi_s(y)\,{\rm d} y \nonumber \\ &+ \int_0^{t-s} \int_{\mathbb{R}^d} \int_{\mathbb{R}^d} p(r,x,y) q(y) \tilde{p}(t-s-r,y,z) \Psi_s(z)\,{\rm d} z \,{\rm d} y\,{\rm d} r \nonumber\\ &= \int_{\mathbb{R}^d} p(t-s,x,y) \Psi_s(y)\,{\rm d} y + \int_0^{t-s} \int_{\mathbb{R}^d} p(r,x,y) q(y) \Psi_{t-r}(,y)\,{\rm d} y\,{\rm d} r \nonumber\\ &= \int_{\mathbb{R}^d} p(t-s,x,y) \Psi_s(y)\,{\rm d} y + \int_s^t \int_{\mathbb{R}^d} p(t-r,x,y) q(y) \Psi_r(y)\,{\rm d} y\,{\rm d} r. \label{eq:PsiDuh2} \end{align} Using \eqref{eq:u_est} and \eqref{ss form eta1}, we get \begin{align*} 0 \le \limsup_{s\to0} \int_{\mathbb{R}^d} p(t-s,x,z) \Psi_s(z) \,{\rm d}{z}&\le c \limsup_{s\to0} (t-s)^{-d/\alpha} \int_{\mathbb{R}^d} \Psi_s(z) \,{\rm d}{z} \\ &= c \limsup_{s\to0} (t-s)^{-d/\alpha} \int_{\mathbb{R}^d} s^{(\delta-d)/\alpha} \Psi_1(s^{-1/\alpha}z) \,{\rm d}{z} \\ & = c t^{-d/\alpha} \int_{\mathbb{R}^d} \Psi_1(z) \,{\rm d}{z} \lim_{s\to0} s^{\delta/\alpha}= 0. \end{align*} Therefore, taking $s \to 0$ in \eqref{eq:PsiDuh2}, we obtain \rf{e.DfPsi}. \end{proof} \begin{remark} By the self-similarity \rf{ss form eta1}, the definition of $\| \cdot \|_{q, h }$, the change of variables $y=t^{-1/\alpha}x$, and \rf{eq:u_est}, we have for all $t>0$ \begin{equation}\label{e.sq} 0<\|\Psi_t\|_{q, h }=t^{-\frac{d-2\delta}{\alpha}(1-\frac{1}{q})}\|\Psi_1\|_{q,h }<\infty, \end{equation} see \rf{e.spt}. It follows that \begin{equation}\label{e.sqst} \|\Psi_{st}\|_{q, h }=t^{-\frac{d-2\delta}{\alpha}(1-\frac{1}{q})}\|\Psi_s\|_{q,h } \end{equation} and \begin{align}\label{Psi_norm} \| \Psi_t\|_{1,h} = \| \Psi_1\|_{1,h}=1, \end{align} which is the same as \rf{psi_int}. Furthermore, \begin{equation*} t^{-\frac{d-2\delta}{\alpha}(1-\frac{1}{q})}\| u(t, \cdot)-A\Psi_t\|_{q,h }=\| t^{\frac{d-\delta }{\alpha}}u\left(t,t^{\frac{1}{\alpha}} \cdot \right) -A\Psi_1\|_{q,h }, \end{equation*} in analogy with the results of Pilarczyk \cite{MR3020137} and V\'azquez and Zuazua \cite{MR1760280}. \end{remark} \begin{remark} Similar arguments yield the following result, to be proved in a forthcoming paper, \begin{equation}\label{H_asymp} \lim_{t\rightarrow \infty }t^{\frac{d}{\alpha}\left(1-\frac{1}{q}\right)-\frac{\delta}{\alpha}}\| u(t,\cdot )-A\Psi_t(\cdot )\|_{q, H_t}=0, \end{equation} where $1 \le q \le \infty$, $H_t(z) =H(t^{-1/\alpha}z)$ and the norms are defined by \begin{equation*} \|f\|_{q,H_t}= \begin{cases} \bigg( \int_{{\mathbb R}^d} |f(x)|^q H_t ^{2-q} (x) \,{\rm d} x \bigg) ^{\frac{1}{q}} & \textrm{for}\quad 1 \le q<\infty,\\ \sup_{x\in {\mathbb R}^d} |f(x)|/H_t(x) &\text{for}\quad q=\infty. \end{cases} \end{equation*} The asymptotics \rf{H_asymp} is an exact analogue of the result in \cite[Theorem 2.1]{MR3020137}. \end{remark} \begin{remark}\label{R:new1} By \eqref{e.cl} and \eqref{psi_int}, $A$ in Theorem~\ref{lim norm th} satisfies $\int_{{\mathbb R}^d} \big(\tilde P_t f(x)-A \Psi_s(x)\big)h(x) \,{\rm d} x =0$ for all $t\ge0, s >0$. \end{remark} \begin{remark}\label{r.cit} By Lemma~\ref{contraction} we have $\int_{{\mathbb R}^d} u(t,x)h(x) \,{\rm d} x=\int_{{\mathbb R}^d} f(x)h(x)\,{\rm d} x$ for all $t\ge 0$. Here is a rather heuristic alternative argument for the equality: If we multiply \rf{equation} by the function $h$ and integrate, then indeed we get \begin{equation*} \frac{\ d }{\ d t} \int_{{\mathbb R}^d} |x|^{-\delta }u(t,x) \,{\rm d} x = -\int_{{\mathbb R}^d}(-\Delta )^{\nicefrac{\alpha }{2}} u(t,x) |x|^{-\delta } \,{\rm d} x + \kappa_\delta \int_{{\mathbb R}^d} |x|^{-\delta -\alpha }u(t,x) \,{\rm d} x . \end{equation*} Using the Fourier symbol of $\Delta^{\alpha/2}$, the Parseval's relation and \rf{eq:dk0} we get \begin{align*} \frac{\ d }{\ d t} \int_{{\mathbb R}^d} |x|^{-\delta }u(t,x) \,{\rm d} x &= -\frac{2^{\alpha }\pi^{\alpha+\delta - \nicefrac{d}{2}} \Gamma \left( \frac{d-\delta}{2} \right)}{ \Gamma \left(\frac{\delta}{2}\right)}\int_{{\mathbb R}^d} |\xi |^{\alpha + \delta -d}\hat u(t,\xi ) \,{\rm d} \xi + \kappa_\delta \int_{{\mathbb R}^d} |x|^{-\delta -\alpha }u(t,x) \,{\rm d} x \\ & =\Big( \kappa_\delta - \kappa_\delta\Big) \int_{{\mathbb R}^d} |x|^{ -\delta -\alpha } u(t,x) \,{\rm d} x = 0, \end{align*} see \cite[Theorem 2.2.14 p.102]{MR2445437} and \cite[Lemma 2 p.117]{MR0290095}. \end{remark} Recall that $\tilde P_t$ is defined in \eqref{e.dtp}. The next lemma is crucial for the proof of Theorem~\ref{lim norm th}. \begin{lemma}\label{lim 0} If $f\in L^1(H)$, $\int_{{\mathbb R}^d } f(x)h(x) \,{\rm d} x =0$ and $1 \le q < \infty$, then \begin{equation}\label{lim_0_cond} \lim_{t \rightarrow \infty }t^{\frac{d-2\delta}{\alpha}(1-\frac{1}{q})}\| \tilde P_t f\|_{q,h }=0. \end{equation} If, additionally, $f$ has compact support, then \eqref{lim_0_cond} is true for $q=\infty$, too. \end{lemma} \begin{proof} First, we consider a compactly supported function $\psi $ such that $\psi \in L^1(H)$ and \begin{equation}\label{0 con 1} \int_{{{\mathbb R}^d}} \psi (x) h(x) \,{\rm d} x =0 , \end{equation} and we intend to prove \rf{lim_0_cond} with $f$ replaced by $\psi$. \noindent {\it Step 1. Case $q=\infty $.}\\ By the definition of the norm $\| \cdot \|_{\infty , h}$, \begin{align*} I(t)&:= t^{\frac{d-2\delta}{\alpha}}\| \tilde P_t \psi \|_{\infty ,h } =t^{\frac{d-2\delta}{\alpha}}\sup_{x\in {{\mathbb R}^d}} \left| \int_{{{\mathbb R}^d}} \rho _t(x,y)h(y)\psi(y) \,{\rm d} y \right|. \end{align*} Using the condition \rf{0 con 1}, we obtain \begin{align*} I(t) = t^{\frac{d-2\delta}{\alpha}}\sup_{x\in {{\mathbb R}^d}}\left| \int_{{{\mathbb R}^d}} \left( \rho_t(x,y)-\rho_t(x,0)\right)h(y)\psi (y) \,{\rm d} y \right|. \end{align*} We fix $\omega>0$. Since $\psi$ has a compact support, for sufficiently large $t>0$ we get \begin{align*} I(t) &=t^{\frac{d-2\delta}{\alpha}}\sup_{x\in {{\mathbb R}^d}}\left| \int_{|y| \le t^{\frac{1}{\alpha}}\omega} \left( \rho_t(x,y)-\rho_t(x,0)\right)h(y)\psi (y) \,{\rm d} y \right|\\ & \le t^{\frac{d-2\delta}{\alpha}}\sup_{\substack {x\in {\mathbb R}^d \\ |y| \le \omega t^{\frac{1}{\alpha}}}}\big|\rho_t(x,y) -\rho_ t(x,0)\big| \int_{|y| \le \omega t^{\frac{1}{\alpha}}}h(y)|\psi (y)| \,{\rm d} y , \end{align*} so by scaling of $\rho$, \begin{align*} I(t)& \le \sup_{\substack {x\in {{\mathbb R}^d} \\ |y| \le \omega t^{\frac{1}{\alpha}}}}\left| \rho_1\left(t^{-\frac{1}{\alpha}}x,t^{-\frac{1}{\alpha}}y \right) - \rho_1\left( t^{-\frac{1}{\alpha}}x, 0\right) \right| \int_{|y| \le t^{\frac{1}{\alpha}}\omega} h(y)| \psi(y)| \,{\rm d} y \\ & \le \sup_{\substack {x\in {{\mathbb R}^d} \\ |y| \le \omega}}\left| \rho_1\left(x,y \right) - \rho_1\left( x, 0\right) \right| \int_{ {{\mathbb R}^d}}h(y)| \psi(y)| \,{\rm d} y. \end{align*} By Lemma \ref{Lemma_rho_lim_0} and \rf{eq:mainThmEstrho2} we choose $\omega $ small enough to have \begin{equation*} \sup_{\substack {x\in {{\mathbb R}^d} \\ |y| \le \omega}}\left| \rho_1\left(x,y \right) - \rho_1\left( x, 0\right) \right|<{\varepsilon }. \end{equation*} This proves \rf{lim_0_cond} in the considered setting.\\ {\it Step 2. Case $q=1$. }\\ By the definition of the norm $\| \cdot \|_{1, h}$, \begin{equation*} J(t):=\| \tilde P_t \psi \|_{1,h}= \int_{{{\mathbb R}^d}}\left| \int_{{{\mathbb R}^d}}\tilde p(t,x,y)h(x)\psi(y) \,{\rm d} y \right| \,{\rm d} x =\int_{{\mathbb R}^d }\Big| \int_{{\mathbb R}^d }\rho_t (x,y)h^2(x)\psi (y)h(y) \,{\rm d} y \Big| \,{\rm d} x . \end{equation*} Applying \rf{0 con 1}, we get \begin{equation*} J(t) \le \int_{{\mathbb R}^d } \int_{{\mathbb R}^d } \big| \rho_t(x,y)-\rho_t(x,0)\big|h^2(x) |\psi (y)|h(y) \,{\rm d} y \,{\rm d} x . \end{equation*} We fix $\omega >0$ and notice that \begin{align*} J(t)& \le \int_{{\mathbb R}^d } \int_{|y| \le \omega t^{\frac{1}{\alpha}}} \big| \rho_t (x,y)-\rho_t(x,0)\big|h^2(x) |\psi (y)| h(y) \,{\rm d} y \,{\rm d} x , \end{align*} for sufficiently large $t$, since the function $\psi $ has a compact support. Then \begin{align*} J(t) & \le \sup_{|y| \le \omega t^{\frac{1}{\alpha}}}\| \rho_t(\cdot , y)-\rho_t(\cdot ,0)\|_{L^1(h^2)} \int_{{\mathbb R}^d} h(y)|\psi (y)| \,{\rm d} y . \end{align*} By scaling of $\rho$, substituting $x=t^{\frac{1}{\alpha}}z$ we obtain \begin{align*} &\sup_{|y| \le \omega t^{\frac{1}{\alpha}}}\| \rho_t(\cdot , y)-\rho_t(\cdot ,0)\|_{L^1(h^2)} \\ &= t^{\frac{2\delta-d}{\alpha}} \sup_{|y| \le \omega t^{\frac{1}{\alpha}}}\int \left | \rho_1 \left(t^{-{1}/{\alpha}}x,t^{-{1}/{\alpha}}y \right) -\rho_1\left( t^{-{1}/{\alpha}}x,0\right)\right|h^2(x) \,{\rm d} x \\ &=\sup_{|y|<\omega} \| \rho_1(\cdot ,y) -\rho_1(\cdot, 0)\|_{L^1(h^2)}. \end{align*} Applying Lemma~\ref{l.crho}, for $\omega >0$ sufficiently small we get \begin{equation*} J(t) \le {\varepsilon } \| \psi \|_{L^1(h)} . \end{equation*} {\it Step 3. Case $q\in (1,\infty)$. }\\ By the definition of the norm $\| \cdot \|_{q, h }$ and H\"older inequality, \begin{align*} t^{\frac{d-2\delta}{\alpha}(1-\frac{1}{q})}\|\tilde P_t \psi \|_{q, h }&=t^{\frac{d-2\delta}{\alpha}(1-\frac{1}{q})}\bigg(\int_{{\mathbb R}^d}\big| \tilde P_t\psi (x)/h(x)\big|^{q-1}\big|\tilde P_t\psi (x)h(x)\big| \,{\rm d} x \bigg)^{\frac{1}{q}} \\ & \le \left(t^{\frac{d-2\delta}{\alpha}}\|\tilde P_t\psi \|_{\infty ,h}\right)^{1-\frac{1}{q}} \left(|\tilde P_t\psi \|_{1, h} \right)^{\frac{1}{q}}. \end{align*} Both factors converge to zero as $t\rightarrow \infty $ by {\it Step 1.} and {\it Step 2.} \\ {\it Step 4. General initial datum.}\\ Let $R>0$, $c_R=\int_{|x|\le R} f(x)h(x)\,{\rm d} x/ \int_{|x|\le R} h(x)\,{\rm d} x$ and $\psi_R (x)= (f (x)-c_R)\mathbf{1}_{|x| \le R}$. Of course, $\psi_R$ is compactly supported and $$\int_{{\mathbb R}^d} h(x) \psi_R (x) \,{\rm d} x =0.$$ Furthermore, \begin{align*} \|f- \psi_R \|_{L^1(h )}& = |c_R| \int_{|x| \le R}h(x) \,{\rm d} x+\int_{|x|> R}h(x)|f (x)| \,{\rm d} x \\ & = \Big| \int_{|x| \le R} h(x)f (x) \,{\rm d} x \Big| + \int_{|x|> R}h(x)|f (x)| \,{\rm d} x \to 0 \end{align*} as $R\rightarrow \infty $, due to the assumption \rf{0 con 1} and the condition $f \in L^1(h)$. Let ${\varepsilon } >0$ and choose $R>0$ so large that \begin{equation*} \| f -\psi_R \|_{1,h}<{\varepsilon }. \end{equation*} For $q=1$, by using the triangle inequality and Lemma \ref{contraction}, we have \begin{align*} \|\tilde P_t f\|_{1,h }& \le \|\tilde P_t\psi_R \|_{1,h }+ \|\tilde P_t(f-\psi_R) \|_{1,h }\\ & \le \|\tilde P_t\psi_R \|_{1,h }+ \|f -\psi_R \|_{1, h } , \end{align*} and {\it Step 2.} of this proof we get \begin{equation}\label{lim epsilon1} \limsup_{t\rightarrow \infty } \|\tilde P_t f\|_{1,h } \le {\varepsilon }. \end{equation} The proof of \rf{lim_0_cond} is complete. If $1<q<\infty$, then using the triangle inequality and Theorem \ref{hyperc}, we obtain \begin{align*} t^{\frac{d-2\delta}{\alpha}(1-\frac{1}{q})}\|\tilde P_t f\|_{q,h }& \le t^{\frac{d-2\delta}{\alpha}(1-\frac{1}{q})}\|\tilde P_t\psi_R \|_{q,h }+ t^{\frac{d-2\delta}{\alpha}(1-\frac{1}{q})}\|\tilde P_t(f-\psi_R) \|_{q,h }\\ & \le t^{\frac{d-2\delta}{\alpha}(1-\frac{1}{q})}\|\tilde P_t\psi_R \|_{q,h }+ C\|f -\psi_R \|_{1, h } +Ct^{-\frac{\delta}{q\alpha}}\|f-\psi_R\|_{L^1({\mathbb R}^d)}. \end{align*} By {\it Step 3.} of this proof, \begin{equation}\label{lim epsilon} \limsup_{t\rightarrow \infty } t^{\frac{d-2\delta}{\alpha}(1-\frac{1}{q})}\|\tilde P_t f\|_{q,h } \le 2C{\varepsilon }, \end{equation} which is valid both for $\delta>0$ and $\delta=0$. This completes the proof of \rf{lim_0_cond} for $q\in (1,\infty)$. \end{proof} We prove Theorem \ref{lim norm th} immediately after the following discussion of \textit{solutions} to \eqref{equation}. \begin{remark}\label{r.ww} {\rm By \eqref{eq:PsiDuh2} we conclude that $\Psi_t(x)$ is a mild solution of \eqref{equation}. In passing we note that, by \eqref{eq:u_est}, $\Psi_t(x)\to 0$ when $t\to 0^+$ and $x\neq 0$. Further, if $f\in L^1(H)$, $t>0$, $x\in {{\mathbb R}^d_0}$ and $u(t,x)=\tilde P_t f(x)$, then \begin{align} u(t,x) &= \int_{\mathbb{R}^d} p(t,x,y) f(y)\,{\rm d} y + \int_0^t \int_{\mathbb{R}^d} p(t-r,x,y) q(y) u(r,y)\,{\rm d} y\,{\rm d} r, \label{eq:PsiDuh23} \end{align} as follows from \eqref{eq:Df1}. This justifies calling $u$ in Theorem \ref{lim norm th} solution to \eqref{equation}. By definition, $u$ can also be called the semigroup solution and \eqref{e.s} yields an analogue for $\Psi_t(x)$. } \end{remark} \begin{proof}[Proof of Theorem \ref{lim norm th}] By \rf{eq:ChK}, \rf{e.s}, Remark \ref{R:new1} and Lemma \ref{lim 0}, \begin{align*} \lim_{t\rightarrow \infty }t^{\frac{d-2\delta}{\alpha}(1-\frac{1}{q})}\| u(t, \cdot)-A\Psi_t\|_{q,h}& = \lim_{t\to \infty}t^{\frac{d-2\delta}{\alpha}(1-\frac{1}{q})}\|u(t+1,\cdot) -A\Psi_{t+1}\|_{q,h} \\ &= \lim_{t\to \infty}t^{\frac{d-2\delta}{\alpha}(1-\frac{1}{q})}\|\tilde P_t\left( \tilde P_1f-A\Psi_1\right)\|_{q,h}=0. \end{align*} \end{proof} Theorem~\ref{lim norm th} is optimal, as asserted by the following two observations. \begin{proposition}\label{p.es} Let $q\in[1,\infty)$ and $\tau:[0,\infty)\to [0,\infty)$ be increasing with $\lim_{t\to \infty}\tau(t)=\infty$. Then there is $f\in L^1(H)$ such that $\int_{{{\mathbb R}^d}}f (x) h(x) \,{\rm d} x=1$ and $u:=\tilde P_t f$ satisfies \begin{equation}\label{e.ass} \lim_{t\rightarrow \infty }\tau(t)t^{\frac{d-2\delta}{\alpha}(1-\frac{1}{q})}\| u(t, \cdot)-\Psi_t\|_{q, h}=\infty. \end{equation} \begin{proof} The proof builds on the fact that $\|\Psi_{1}-\Psi_s\|_{q,h}>0$ for $s>1$, which, as we shall see below, is a consequence of the scaling of $\Psi_t(x)$. Without loss of generality we may assume that $\tau$ is continuous, strictly increasing and $\tau(0)=0$, by replacing $\tau(t)$ by $t/(t+1)\left[t/(t+1)+\frac1t \int_0^t \tau(s)ds\right]\le \tau(t)+1$, see also \eqref{e.as}. In particular, $\tau^{-1}$ is well defined $: [0,\infty)\to [0,\infty)$. For $n\in {\mathbb N}$ we let $t_n=\tau^{-1}(2^{2n})$. Let $f=\sum_{n=1}^\infty 2^{-n}\Psi_{t_n}$. According to \eqref{Psi_norm}, $\int_{{{\mathbb R}^d}}f (x) h(x) \,{\rm d} x=\|f\|_{1,h}=1$. By \eqref{e.s}, $$u(t,x):=\tilde P_t f(x)=\sum_{n=1}^\infty 2^{-n} \Psi_{t+t_n}(x)\ge 0,\quad t>0,\ x\in {{\mathbb R}^d}.$$ We remark that if $\tau(t)\ge 1$, then $t_n\ge t$ is equivalent to $n\ge \left \lceil{\frac12 \log_2 \tau(t)}\right \rceil$. In this case, by \eqref{e.sqst}, and triangle inequality, we get \begin{align}\label{e.sln} &t^{\frac{d-2\delta}{\alpha}(1-\frac1q)}\| u(t, \cdot)-\Psi_t\|_{q, h}= \|\sum_{n=1}^\infty 2^{-n} \Psi_{1+t_n/t}-\Psi_1\|_{q,h}\\ \nonumber &\ge \|\Psi_1\|_{q,h} -\sum_{n=1}^\infty 2^{-n} \|\Psi_{1+t_n/t}\|_{q,h} = \|\Psi_1\|_{q,h}\sum_{n=1}^\infty 2^{-n} \left(1-(1+t_n/t)^{-\frac{d-2\delta}{\alpha}(1-\frac{1}{q})}\right)\\ \nonumber &\ge \|\Psi_1\|_{q,h}\left(1-2^{-\frac{d-2\delta}{\alpha}(1-\frac{1}{q})}\right)\sum_{t_n\ge t}^\infty 2^{-n}= \|\Psi_1\|_{q,h}\left(1-2^{-\frac{d-2\delta}{\alpha}(1-\frac{1}{q})}\right) 2^{1-\left \lceil{\frac12 \log_2 \tau(t)}\right \rceil} \\ &\ge \|\Psi_1\|_{q,h}\left(1-2^{-\frac{d-2\delta}{\alpha}(1-\frac{1}{q})}\right)\tau(t)^{-1/2}.\nonumber \end{align} The case of $q\in (1,\infty)$ in \eqref{e.ass} is resolved, because $1-2^{-\frac{d-2\delta}{\alpha}(1-\frac{1}{q})}>0$ in this case. We next consider $q=1$. Then, starting from \eqref{e.sln}, we get \begin{align} \| u(t, \cdot)-\Psi_t\|_{1, h}&= \|\sum_{n=1}^\infty 2^{-n} \Psi_{1+t_n/t}-\Psi_1\|_{1,h} \nonumber \\\nonumber &\ge \int_{B(0,1)} |\Psi_1(x)-\sum_{n=1}^\infty 2^{-n} \Psi_{1+t_n/t}(x)|h(x)\,{\rm d} x\\\nonumber &\ge \int_{B(0,1)}\left(\Psi_1(x)-\sum_{n=1}^\infty 2^{-n} \Psi_{1+t_n/t}(x)\right)h(x)\,{\rm d} x\\ &= \sum_{n=1}^\infty 2^{-n}\left(\int_{B(0,1)}\Psi_1(x)h(x)\,{\rm d} x - \int_{B(0,1)}\Psi_{1+t_n/t}(x)h(x)\,{\rm d} x\right). \label{e.sln1aa} \end{align} By the same change of variables as in \eqref{e.sq}, for all $t>0$ and $U\subset {{\mathbb R}^d}$ we get \begin{align}\int_{U}\Psi_t(x) h(x)\,{\rm d} x= &\int_{t^{-1/\alpha}U}\Psi_1(x) h(x)\,{\rm d} x. \label{e.sq2a} \end{align} Therefore by \eqref{e.sln1aa}, \begin{align*} &\| u(t, \cdot)-\Psi_t\|_{1, h} \ge \sum_{n=1}^\infty 2^{-n}\int_{B(0,1)\setminus B(0,(1+t_n/t)^{-1/\alpha})}\Psi_1(x)h(x)\,{\rm d} x\\ &\ge \int_{B(0,1)\setminus B(0,2^{-1/\alpha})}\Psi_1(x)h(x)\,{\rm d} x\sum_{t_n\ge t}^\infty 2^{-n}. \end{align*} By \eqref{eq:u_est}, $\int_{B(0,1)\setminus B(0,2^{-1/\alpha})}\Psi_1(x)h(x)\,{\rm d} x>0$, and we conclude as before. \end{proof} \end{proposition} \begin{remark} We note that \eqref{e.as} does not hold for $q = \infty$. Indeed, let $f_n(x) = {\bf 1}_{B(x_n,1)}(x)$, where $\{x_n\}$ is a sequence in ${{\mathbb R}^d}$ such that $|x_n| = 2^{n}$. Then, by \eqref{eq:oppt}, for $t>1$ we have \begin{align*} \tilde{P}_t f_n(x_n) \ge \int_{B(x_n,1)} p(t,x_n,y) \,{\rm d}{y} \ge c t^{-d/\alpha}. \end{align*} Let $f(x) = \sum_{n=2}^{\infty} f_n(x)$. Clearly $f \in L^1(h)$ and $\tilde{P}_t f(x_n) \ge c t^{-d/\alpha}$. Moreover, by \eqref{eq:u_est} and \eqref{eq:oppt}, $\Psi_t(x_n) \le c t^{\delta/\alpha} (1 + t^{-\delta/\alpha}|x_n|^{-\delta}) t|x_n|^{-d-\alpha}$. Hence for every $t>1$ we actually have, \begin{align*} \| u(t, \cdot)-A\Psi_t\|_{\infty, h} = \sup_{x \in {{\mathbb R}^d}} |\tilde{P}_t f(x) - A \Psi_t (x)|/h(x) \ge c_t \sup_{n \in \mathbb{N}}(|x_n|^{\delta} - |x_n|^{\delta-d-\alpha}) = \infty. \end{align*} \end{remark} \section{The potential of the self-similar solution}\label{s.psss} In this section we assume that $0\le \kappa \le \kappa^*$, except that on several occasions we explicitly exclude the critical case $\kappa=\kappa^*$. As usual, $\kappa$ and $\delta$ are related by \rf{eq:dk0}. \textit{If} $\kappa < \kappa^*$ then, by \eqref{ss form eta1} and \eqref{eq:u_est}, there is $c\in (0,\infty)$, such that \begin{align*} c\int_0^\infty \Psi_s(x) \,{\rm d} s = |x|^{\delta+\alpha-d}, \qquad x \in {{\mathbb R}^d}. \end{align*} Furthermore, \begin{align*} \int_{{\mathbb R}^d} \tilde{p}_t(x,y) |y|^{\delta+\alpha-d} \,{\rm d} y &= c \int_0^\infty \int_{{\mathbb R}^d} \Psi_s(y) \tilde{p}_t(x,y) \,{\rm d} y \,{\rm d} s = c\int_0^\infty \Psi_{t+s}(x) \,{\rm d} s\\ &= c\int_0^\infty \Psi_s(x) \,{\rm d} s - c\int_0^t \Psi_s(x) \,{\rm d} s = |x|^{\delta+\alpha-d} - c\int_0^t \Psi_s(x) \,{\rm d} s. \end{align*} Hence, for $t>0$ and $y\in \mathbb{R}^d \setminus \{0\}$, \begin{align*} c\int_0^t \Psi_s(x) \,{\rm d} s = |x|^{\delta+\alpha-d} - \int_{{\mathbb R}^d} \tilde{p}_t(x,y) |y|^{\delta+\alpha-d} \,{\rm d} y. \end{align*} The main goal of this section is to calculate the constant $c$, and derive a similar formula for $\kappa=\kappa^*$ (see Corollary \ref{cor:intPsi}). \begin{lemma}\label{lem:PsiAI} For $t>0$ and $x \in {{\mathbb R}^d}$ we have \begin{align*} \Psi_t(x) = \lim_{\beta \to 0^+} \frac{\Gamma(d/2)}{2\pi^{d/2}}\int_{{{\mathbb R}^d}} \beta |z|^{\beta+\delta-d} \tilde{p}(t,x,z) \,{\rm d}{z}. \end{align*} \end{lemma} \begin{proof} Let $x \in {{\mathbb R}^d_0}$. For $z\in{{\mathbb R}^d_0}$ and $\beta>0$, put $f_\beta(z) = \frac{\Gamma(d/2)}{2\pi^{d/2}} \beta |z|^{\beta-d}$. Let $\varepsilon >0$ and $\beta \to 0$. We have \begin{align*} \beta \int_{B(0,\varepsilon)} |z|^{\beta-d} \,{\rm d} z &= d|B(0,1)| \beta \int_0^\varepsilon r^{\beta-d} r^{d-1} \,{\rm d}{r} = d|B(0,1)|\varepsilon^\beta \longrightarrow \frac{2 \pi^{d/2}}{\Gamma(d/2)}. \end{align*} Furthermore, by \eqref{eq:mainThmEstz} and the dominated convergence theorem for every $x \in {{\mathbb R}^d_0}$ we get \begin{align*} \lim_{\beta \to 0^+} \beta \int_{|z|>\varepsilon} \frac{\tilde{p}(t,x,z)}{h(z)} |z|^{\beta-d} \,{\rm d}{z} = 0. \end{align*} Then, since $z \mapsto \tilde{p}(t,x,z)/h(z)$ has a continuous extension to ${{\mathbb R}^d}$ with the value $\Psi_t(x)$ at the origin, we get \begin{align*} \Psi_t(x) = \lim_{\beta \to 0^+} \int_{{{\mathbb R}^d}} \frac{\tilde{p}(t,x,z)}{h(z)} f_\beta(z) \,{\rm d}{z}. \end{align*} The result follows (in the case of $x=0$, the statement is trivial). \end{proof} \begin{lemma}\label{lem:intlimPsi} For $t>0$ and $x \in {{\mathbb R}^d_0}$, \begin{align*} \int_0^t \Psi_s(x) \,{\rm d}{s} = \lim_{\beta \to 0^+} \frac{\Gamma(d/2)}{2\pi^{d/2}} \frac{\beta}{\kappa_{\delta+\beta}-\kappa_\delta} \left( |x|^{-(d-\delta -\alpha-\beta)} - \int_{{{\mathbb R}^d}} \tilde{p}(t,x,z)|z|^{-(d-\delta -\alpha-\beta)} \,{\rm d}{z}\right). \end{align*} \end{lemma} \begin{proof} Fix $x \in {{\mathbb R}^d_0}$ and $t>0$. By \eqref{eq:oppt}, there is a constant $c=c(d,\alpha)$ such that \begin{align*} &\int_{{{\mathbb R}^d}} p(s,x,z) |z|^{\gamma-d} \,{\rm d}{z} \le \int_{|z|>|x|/2} p(s,x,z) (|x|/2)^{\gamma-d} \,{\rm d}{z} + c\int_{|z|\le|x|/2} \frac{s}{|x|^{d+\alpha}} |z|^{\gamma-d} \,{\rm d}{z} \\ &\le (|x|/2)^{\gamma-d} + \frac{c}{\gamma} t|x|^{\gamma-d-\alpha} \le c \frac{t + 1}{\gamma} \left(1+ |x|^{-d-\alpha}\right), \quad s \in (0,t),\; \gamma \in (0,d). \end{align*} Then, by \eqref{eq:mainThmEstz2}, there is $c'=c'(d,\alpha,t)$ such that \begin{align*} \int_{{{\mathbb R}^d}} \beta \tilde{p}(s,x,z) |z|^{\beta+\delta-d} \,{\rm d}{z} &\le c' H(x) \beta \int_{{{\mathbb R}^d}} p(s,x,z) (|z|^{\beta+\delta-d} + |z|^{\beta-d}) \,{\rm d}{z} \\ &\le c'H(x) \left(1 + |x|^{-d-\alpha} \right), \quad \beta \in (0,d-\delta). \end{align*} Therefore, by Lemma \ref{lem:PsiAI}, the dominated convergence theorem and \cite[Theorem~3.1]{MR3933622}, \begin{align} &\int_0^t \Psi_s(x) \,{\rm d}{s} = \lim_{\beta \to 0^+}\frac{\Gamma(d/2)}{2\pi^{d/2}}\int_0^t \int_{{{\mathbb R}^d}} \beta \tilde{p}(s,x,z) |z|^{\beta+\delta-d} \,{\rm d}{z} \,{\rm d}{s} \notag\\ &= \lim_{\beta \to 0^+} \frac{\Gamma(d/2)}{2\pi^{d/2}} \frac{\beta}{\kappa_{d-\delta -\alpha-\beta}-\kappa_\delta} \left( |x|^{-(d-\delta -\alpha-\beta)} - \int_{{{\mathbb R}^d}} \tilde{p}(t,x,z)|z|^{-(d-\delta -\alpha-\beta)} \,{\rm d}{z}\right). \label{eq:intu} \end{align} The result follows from the symmetry of $\kappa$. \end{proof} \begin{corollary}\label{cor:intPsi} For $0 \le \delta< \frac{d-\alpha}{2}$, we have \begin{align}\label{eq:intu2} \int_0^t \Psi_s(x) \,{\rm d}{s} = \frac{\Gamma(d/2)}{2\pi^{d/2}\kappa'_\delta} \left( |x|^{-(d-\delta -\alpha)} - \int_{{{\mathbb R}^d}} \tilde{p}(t,x,z)|z|^{-(d-\delta -\alpha)} \,{\rm d}{z}\right), \quad t>0,\, x\in{{\mathbb R}^d_0}. \end{align} For $\delta= \frac{d-\alpha}{2}$, \begin{align}\label{e.pcc} \int_0^t \Psi_s(x)\,{\rm d}{s} = \frac{\Gamma(d/2)}{ \pi^{d/2}\kappa_\delta''} \left(|x|^{-\delta}\ln|x| - \int_{{\mathbb R}^d} \tilde{p}(t,z,x) |z|^{-\delta}\ln|z| \,{\rm d}{z} \right), \quad t>0,\, x\in{{\mathbb R}^d_0}. \end{align} Here, $\kappa_\delta'$ and $\kappa_\delta''$ are the first and the second derivatives of $\kappa_\delta$, respectively. \end{corollary} \begin{proof} For $0 \le \delta< \frac{d-\alpha}{2}$ the statement follows directly from Lemma \ref{lem:intlimPsi}, \eqref{eq:mainThmEstz} and the dominated convergence theorem. For $\delta= \frac{d-\alpha}{2}$, we have $\kappa_\delta' =0$ and \begin{align}\label{eq:invfun} \int_{{\mathbb R}^d} \tilde{p}(t,x,z) |z|^{\delta+\alpha -d} \,{\rm d}{z} = \int_{{\mathbb R}^d} \tilde{p}(t,x,z) |z|^{-\delta} \,{\rm d}{z} = |x|^{-\delta} = |x|^{\delta+\alpha-d}, \end{align} see \rf{eq:2}, so the difference in parentheses on the right hand side of \rf{eq:intu} tends to $0$. So, as $\beta \to 0$, we use \eqref{eq:invfun}, \eqref{eq:mainThmEstz} and the dominated convergence theorem, obtaining \begin{align*} \int_0^t \Psi_s(x)\,{\rm d}{s} &= \lim_{\beta \to 0^+} \frac{\Gamma(d/2)}{2\pi^{d/2}} \frac{\beta^2}{\kappa_{\delta+\beta}-\kappa^*} \left(\frac{|x|^{\beta-\delta} - |x|^{-\delta} }{\beta}- \int_{{\mathbb R}^d} \tilde{p}(t,x,z) \frac{|z|^{\beta-\delta} - |z|^{-\delta}}{\beta} \,{\rm d}{z} \right) \\ &= \frac{\Gamma(d/2)}{ \pi^{d/2}\kappa_\delta''} \left(|x|^{-\delta}\ln|x| - \int_{{\mathbb R}^d} \tilde{p}(t,x,z) |z|^{-\delta}\ln|z| \,{\rm d}{z} \right). \end{align*} \end{proof} \begin{theorem}\label{thm:intu} For $\delta \in [0,\frac{d-\alpha}{2})$ and $x\in {{\mathbb R}^d}$ we have \begin{align}\label{Eq:intu2} \int_0^\infty \Psi_s(x)\,{\rm d}{s} = \frac{\Gamma(d/2)}{2 \pi^{d/2}\kappa_\delta'} |x|^{\delta+\alpha-d}, \end{align} and \begin{align}\label{Eq:intu} \int_t^\infty \Psi_s(x)\,{\rm d}{s} = \frac{\Gamma(d/2)}{2 \pi^{d/2}\kappa_\delta'} \int_{{\mathbb R}^d} \tilde{p}(t,x,z)|z|^{\delta+\alpha -d} \,{\rm d} z, \quad t>0. \end{align} For $\delta=(d-\alpha)/2$, \begin{align}\label{Eq:intuk} \int_t^\infty \Psi_s(x)\,{\rm d}{s} = \infty, \quad x\in {{\mathbb R}^d}, \quad t\ge 0. \end{align} \end{theorem} \begin{proof} We shall prove \eqref{Eq:intu2} by letting $t \to \infty$ in \eqref{eq:intu2}. To this end let $x \in {{\mathbb R}^d_0}$ and $T>1$. By \eqref{eq:intu2}, $\int_{{\mathbb R}^d} \tilde{p}(t,x,z)|z|^{\delta+\alpha -d} \,{\rm d} z $ is finite and decreases as $t$ increases to $\infty$. Hence, for every $\varepsilon>0$ there is $R>0$ such that for every $t>T$, \begin{align}\label{eq:ptIntBr} \int_{B(0,R)^c} \tilde{p}(t,x,z) |z|^{\delta+\alpha -d} \,{\rm d} z \le \int_{B(0,R)^c} \tilde{p}(T,x,z) |z|^{\delta+\alpha -d} \,{\rm d} z < \varepsilon. \end{align} By \eqref{eq:mainThmEstz}, for $t>T$ we get \begin{align*} |z|^{\delta+\alpha -d} \tilde{p}(t,x,z) &\le c |z|^{\delta+ \alpha -d} (1+t^{\delta/\alpha}|z|^{-\delta})(1+t^{\delta/\alpha}|x|^{-\delta}) t^{-d/\alpha} \\ &\le c |z|^{\delta+ \alpha -d} (1+|z|^{-\delta})(1+|x|^{-\delta}) T^{-(d-2\delta)/\alpha}. \end{align*} By the dominated convergence theorem, \begin{align*} \lim_{t\to\infty} \int_{B(0,R)} \tilde{p}(t,x,z) |z|^{\delta+\alpha -d} \,{\rm d} z =0. \end{align*} This and \eqref{eq:ptIntBr} yields \eqref{Eq:intu2}. Then \eqref{Eq:intu} follows by \eqref{eq:intu2} and \eqref{Eq:intu2}. If $x=0$, then we trivially have infinity on both sides of \eqref{Eq:intu2} and \eqref{Eq:intu}. Finally, \eqref{Eq:intuk} follows from \eqref{eq:u_est}. \end{proof} \begin{remark}\label{r.asss} We note that the function \begin{align} \mu_t(x) &:= \int_0^t \Psi_s(x)\,{\rm d} s\label{e.dmutx} \end{align} is self-similar, too. Namely, by \eqref{ss form eta1} and changing variables $s=tu$ in \rf{e.dmutx}, we get \begin{align*} \mu_t(x)=t^{(\alpha+\delta-d)/\alpha}\mu_1(t^{-1/\alpha}x),\quad t>0, x\in {{\mathbb R}^d}. \end{align*} Furthermore, $\mu$ satisfies the Duhamel formula, \begin{align*} \mu_t(x) = \int_0^t\int_{{{\mathbb R}^d}} \mu_s(z) q(z) p(t-s,x,z) \,{\rm d}{z}\,{\rm d}{s}. \end{align*} Indeed, by \eqref{e.DfPsi} and Fubini-Tonelli, \begin{align*} \int_0^t \int_{\mathbb{R}^d} p(t-s,x,z)q(z)\mu_s(z) \,{\rm d}{z} \,{\rm d}{s} &= \int_0^t \int_{\mathbb{R}^d} p(t-s,x,z)q(z)\int_0^s\Psi_{s-r}(z) \,{\rm d}{r} \,{\rm d}{z} \,{\rm d}{s} \\ &= \int_0^t \int_r^t\int_{\mathbb{R}^d} p(t-s,x,z)q(z)\Psi_{s-r}(z) \,{\rm d}{s} \,{\rm d}{z} \,{\rm d}{r} \\ &= \int_0^t \int_0^{t-r}\int_{\mathbb{R}^d} p(t-r-s,x,z)q(z)\Psi_s(z) \,{\rm d}{s} \,{\rm d}{z} \,{\rm d}{r} \\ &= \int_0^t \Psi_{t-r}(z) \,{\rm d}{r} = \mu_t(x). \\ \end{align*} Of course, it is $\Psi_t$, not $\mu_t$, that captures the large time asymptotics for the solutions of the equation \eqref{equation} in Theorem~\ref{lim norm th}. Interestingly, it seems feasible, if not easy, to construct $\mu_t$ directly as $\lim_{y\to0} \int_0^t \tilde{p}(s,x,y)/h(y)\,{\rm d}{s}$ and then attempt \textit{to define} $\Psi_t(x)=\partial \mu_t(x)/\partial t$. \end{remark}

b3d2119dccacce2fa6f213f191a330c8c4cb8470

\section{Introduction} \label{introduction} Neural network pruning has become an essential tool to reduce the size of modern-day neural networks. Small-sized networks are important for faster inference and many real-world tasks, for example, deployment on edge devices. With scaling model parameters becoming a popular way to scale performance, model sizes are becoming huge and an increasing cost to the environment \cite{strubell2019energy}. Neural network compression and neural network pruning in particular, has therefore, seen a lot of new work in the last few years. This has also coincided with the insight by Han et al. in 2015 \cite{han2015learning}, that neural networks can be pruned to a significantly large extent without a drop in accuracy. New methods have utilized a myriad of pruning techniques consisting of gradient-based methods, sensitivity to or feedback from an objective function, distance or similarity measures, regularization-based techniques, amongst others. The state-of-the-art (SOTA) pruning techniques use complex rules like iterative pruning and re-growth of weight parameters using heuristics rules every few hundred iterations for DSR \cite{ICML-2019-MostafaW}. SM \cite{sparse_momentum} uses sparse momentum that uses exponentially smoothed gradients (momentum) to find layers and weights that reduce error and then redistribute the pruned weights across layers using the mean momentum magnitude of each layer. For each layer, sparse momentum grows the weights using the momentum magnitude of zero-valued weights. Another popular SOTA technique, RigL \cite{pmlr-v119-evci20a}, also works by iteratively pruning and re-growing weights every few iterations. They use either uniform or Erdos-Renyi-Kernel (ERK) for pruning connections and re-grow connections based on the highest magnitude gradients. Among the most recent techniques, DPF \cite{Lin2020Dynamic} uses dynamic allocation of the sparsity pattern and incorporates a feedback signal to re-activate prematurely pruned weights, while STR \cite{pmlr-v119-kusupati20a} utilises Soft Threshold Reparameterization and uses back-propagation to find sparsity ratios for each layer. Despite the high number of new pruning algorithms proposed, the tangible benefits of many of them are still questionable. For instance, recently it has been shown that many pruning at initialization (PAI) schemes do not perform as well as expected \cite{frankle2021pruning}. In that paper, it is shown through a number of experiments that these PAI schemes are actually no better than random pruning, which is one of the most naive pruning baselines with no complexity involved. Similarly, in this paper, we bring attention to the trend of proposing increasingly complex pruning algorithms and question whether such complexity is really required to achieve superior results. We benchmark popular state-of-the-art (SOTA) pruning techniques against a naive pruning baseline, namely, Global Magnitude Pruning (Global MP). Global MP ranks all the weights in a neural network by their magnitudes and then prunes off the smallest ones (Fig.~\ref{fig:workflow}). Thus, in its vanilla form, it is a very simple pruning technique and contrasts sharply with the rest of the algorithms in the literature in terms of complexity. Despite its simplicity, Global MP has not been comprehensively analyzed and evaluated in the literature. Although, some prior works have used Global MP as a baseline \cite{frankle2018lottery, NEURIPS2019_a4613e8d, blalock2020state, NEURIPS2020_46a4378f, Renda2020Comparing, lee2021layeradaptive}, they missed out on conducting rigorous experiments with it; for example, in settings of both gradual and one-shot pruning or comparing it with SOTA. Similarly, many SOTA papers do not use Global MP for benchmarking and miss out on capturing its remarkable performance \cite{pmlr-v119-evci20a, pmlr-v119-kusupati20a, Zhu2018ToPO, gale2019state, DNW}. We bridge this gap in evaluating the efficacy of Global MP under multiple experimental conditions and demonstrate its superior performance. In this paper, we show that naive Global MP surpasses the other pruning techniques and sets a new SOTA result for ImageNet experiments. This performance is also valid across different datasets, neural network models, and target sparsity levels. While achieving such performance, Global MP does not require any additional algorithm-specific hyper-parameters to be tuned. Unlike many pruning techniques in the literature, it is very straightforward to implement. We conduct experiments with Global MP in both one-shot and gradual settings, and find that Global MP in a gradual fashion helps to increase the FLOPs sparsity even further, without compromising accuracy. Aside to its benefits, we also shed light into a potential problem with Global MP, known as layer-collapse, whereby an entire layer is pruned away, leading to a drastic loss in accuracy. In fact, this is a long-standing issue for many pruning algorithms in the literature, but the fix for it in Global MP is rather simple through introducing a minimum threshold to retain a minimum number of weights in every layer, while it is likely to be more complicated in other algorithms. We conduct experiments on WRN-28-8, ResNet-32, ResNet-50, MobileNet-V1, and FastGRNN models, and on CIFAR-10, ImageNet, and HAR-2 datasets. We test Global MP for both unstructured and structured as well as one-shot and gradual settings, and share our findings. \begin{figure*}[!h] \centering \includegraphics[trim=0cm 5.5cm 0cm 8cm,clip,width=\textwidth]{Figures/GlobalMPv4.pdf} \caption{Illustration of how Global MP works. Global MP ranks all the weights in a network by their magnitudes and prunes off the smallest weights until the target sparsity is met. Light green weights refer to the smaller-magnitude weights which are pruned off. A pruned network consisting of larger-magnitude weights (dark green weights) is obtained after the process.} \label{fig:workflow} \end{figure*} \section{Related Work} \label{related_work} Compression of neural networks has become an important research area due to the rapid increase in size of neural networks \cite{brown2020language}, the need for fast inference \cite{camci2020deep}, application to real-world tasks \cite{9516010, 8818358, 8756206, Liu2021ARA, 8693518} and concerns about the carbon footprint of training large neural networks \cite{strubell2019energy}. Over the years, several compression techniques have emerged in the literature \cite{cheng2017survey, 9478787}, such as quantisation, factorisation, attention, knowledge distillation, architecture search and pruning \cite{almahairi2016dynamic,ashok2017n2n,i2016squeezenet,pham2018efficient}. Quantisation techniques which restrict the bitwidth of parameters \cite{Rastegari_2016,courbariaux2016binarized} and tensor factorisation and decomposition which aim to break large kernels into smaller components \cite{mathieu2013fast,gong2014compressing,lebedev2014speedingup,Masana_2017} are popular methods. However, they need to be optimised for specific architectures. Attention networks \cite{almahairi2016dynamic} have two separate networks to focus on only a small patch of the input image. Training smaller student networks in a process called knowledge distillation \cite{ashok2017n2n, 9461003} has also proved effective, although it can potentially require a large training budget. Architecture search techniques, such as new kernel design \cite{i2016squeezenet} or whole architecture design \cite{pmlr-v80-pham18a,Tan_2019} have also become popular. Nevertheless, the large search space size requires ample computational resources to do the architecture search. Different from all these approaches, we focus on pruning deep neural networks in this work. As compared to other categories, pruning is more general in nature and has shown strong performance \cite{gale2019state}. Many pruning techniques have been developed over the years, which use first or second order derivatives \cite{NIPS1989_250,NIPS1992_647}, gradient based methods \cite{lee2018snip, Wang2020Picking}, sensitivity to or feedback from some objective function \cite{Lin2020Dynamic, molchanov2016pruning, LIU2020Dynamic, jorge2021progressive, 9097925}, distance or similarity measures \cite{Srinivas_2015}, regularization-based techniques \cite{pmlr-v119-kusupati20a, ContinuousSparsification2020, wang2021neural, 9398648}, and magnitude-based criterion \cite{pmlr-v119-evci20a, lee2021layeradaptive, Zhu2018ToPO, Strom97sparseconnection, Park2020LookaheadAF}. A key trick has been discovered in \cite{han2015learning} to iteratively prune and retrain a network, thereby preserving high accuracy. Runtime Neural Pruning \cite{NIPS2017_6813} attempts to use reinforcement learning (RL) for compression by training an RL agent to select smaller sub-networks during inference. \cite{he2018amc} design the first approach using RL for pruning. However, RL training approaches typically require additional RL training budgets and careful RL action and state space design \cite{gupta2020learning, qlp}. Global Magnitude Pruning (Global MP) on the other hand works by ranking all the parameters in a network by their absolute magnitudes and then pruning the smallest ones. It is therefore, quite intuitive, logical and straightforward to implement. It is also not to be confused with methods utilizing Global Pruning but not conducting magnitude pruning, for example, SNIP \cite{lee2018snip}. Many methods can do Global Pruning but they cannot be called Global MP because they do not conduct magnitude-based pruning. Some prior works have utilised Global MP but have missed out on rigorously benchmarking it, for example in settings of both gradual and one-shot pruning, and have also not compared it to SOTA algorithms \cite{frankle2018lottery, NEURIPS2019_a4613e8d, blalock2020state, NEURIPS2020_46a4378f, Renda2020Comparing, lee2021layeradaptive}. Also, many SOTA algorithms miss out on benchmarking their algorithms against Global MP and hence are unable to capture its efficacy \cite{pmlr-v119-evci20a, pmlr-v119-kusupati20a, Zhu2018ToPO, gale2019state, DNW}. We conduct systematic experiments to bridge this gap and demonstrate its superior performance by evaluating its efficacy under multiple experimental conditions. \section{Method} \label{approach} In this section, we explain how Global MP works by describing its key components. We shed light into its practical details and implementation. We present a pseudocode to explain the algorithmic flow of Global MP (Algorithm~\ref{algo1} and Table~\ref{table:func_list}). We also introduce a simple thresholding mechanism, called \textit{Minimum Threshold (MT)}, to avoid the issue of layer-collapse at high sparsity levels. \subsection{Global Magnitude Pruning (Global MP)} \label{algo:gp} Global MP is a magnitude-based pruning approach, whereby weights larger than a certain threshold are kept, and weights smaller than the threshold are pruned across a neural network. The threshold is calculated based on the target sparsity rate and is not a hyper-parameter that needs to be tuned or learnt. Given a target sparsity rate $\kappa_{target}$, the threshold $t$ is simply calculated as the weight magnitude that serves as a separation point between the smallest $\kappa_{target}$ percent of weights and the rest, once all weights are sorted into an array based on their magnitude. Formally, for a calculated threshold $t$ and each individual weight $w$ in any layer, the new weight $w_{new}$ is defined as follows: \begin{equation} w_{new} = \begin{cases} 0 & |w| < t, \\ w & otherwise. \\ \end{cases} \end{equation} In Global MP, a single threshold is set for the entire network based on the target sparsity for the network. This is in contrast to layer-wise pruning, in which different threshold values have to be searched for each layer individually. In the case of uniform pruning on the other hand, a threshold for each layer needs to be calculated based on the sparsity target assigned to the layers uniformly across the network. In this aspect, Global MP is more efficient than layer-wise or uniform pruning because the threshold does not need to be searched or calculated for every layer individually. \subsection{Minimum Threshold (MT)} \label{algo:MT} The Minimum Threshold (MT) refers to the fixed number of weights that are preserved in every layer of the neural network post pruning. The MT is a scalar value that is fixed before the start of the pruning cycle. The weights in a layer are sorted by their magnitude and the largest MT number of weights are preserved. For instance, an MT of 500 implies that 500 of the largest weights in every layer need to be preserved post pruning. If a layer originally has a smaller number of weights than the MT number, then all the weights of that layer will be preserved. Therefore, MT is simple to apply and also computationally inexpensive. This corresponds to: \begin{equation} {\|W_l\|_0} \geq \begin{cases} \sigma & \text{if } m \geq \sigma_l, \\ m & \text{otherwise.} \\ \end{cases} \label{eq:MT} \end{equation} The term $W_l \in \mathbb{R}^{m}$ denotes the weight vector for layer $l$, $\sigma$ is the MT value in terms of the number of weights and ${\|W_l\|_0}$ indicates the number of non-zero elements in $W_l$. We explain in the below section how the actual pruning using MT is implemented. \begin{algorithm}[t!] \caption{Global MP} \label{algo1} \begin{algorithmic} \STATE{\textbf{Input:} $DNN_{init}$, pre-trained or untrained DNN} \STATE{\hspace{1.1cm}$\kappa_{target}$, target sparsity} \STATE{\hspace{1.1cm}$\sigma$, minimum threshold (MT)} \STATE{\hspace{1.1cm}$e_{total}$, total epochs} \STATE{\hspace{1.1cm}$isGradual$, gradual or one-shot} \STATE{\hspace{1.1cm}$isMT$, MT is applied or not} \STATE{\textbf{Output:} $DNN_{final}$, pruned and trained DNN} \vspace{0.35cm} \STATE{$e = 0$} \STATE{$DNN(w_e)$ = $DNN_{init}$} \WHILE{$e < e_{total}$} \IF{$\kappa_{DNN(w_e)} < \kappa_{target}$} \STATE{$t_e$ $\leftarrow$ $CalcThreshold(e, \kappa_{target}, isGradual)$} \IF{$isMT$} \STATE{$DNN(w_{e'}) \hspace{0.1cm} \leftarrow Mask(DNN(w_e), t_e)$} \STATE{$W \hspace{1.43cm} \leftarrow MTcheck(DNN(w_{e'}), \sigma)$} \STATE{$DNN(w_{e^+}) \leftarrow MTprune(DNN(w_{e'}), W)$} \ELSE \STATE{$DNN(w_{e^+}) \leftarrow Prune(DNN(w_e), t_e)$} \ENDIF \ELSE \STATE{$DNN(w_{e^+}) = DNN(w_e)$} \ENDIF \STATE{$DNN(w_{e^{++}}) \leftarrow BackProp(DNN(w_{e^+}))$} \STATE{$DNN(w_e) = DNN(w_{e^{++}})$} \STATE{$e = e+1$} \ENDWHILE \STATE{$DNN_{final}$ = $DNN(w_e)$} \end{algorithmic} \end{algorithm} \begin{table}[t!] \caption{Function explanations for Algorithm 1.} \label{table:func_list} \centering \begin{tabular}{rl} \hline\noalign{\smallskip} \textbf{Function} & \textbf{Explanation} \\ \noalign{\smallskip}\hline\noalign{\smallskip} \multirow{4}{*}{$CalcThreshold()$:} & Calculates the magnitude threshold below which \\ & the weights are to be pruned. Assigns all pruning \\ & budget in one epoch or distributes it to epochs \\ & based on $isGradual$. \\ \noalign{\smallskip} \multirow{2}{*}{$Mask()$:} & Identifies the weights to be pruned without \\ & actually pruning them. \\ \noalign{\smallskip} \multirow{3}{*}{$MTcheck()$:} & Checks the layers that violate the MT condition \\ & and returns a new mask by distributing \\ & the pruning budget among other layers. \\ \noalign{\smallskip} \multirow{2}{*}{$MTprune()$:} & Prunes based on the mask returned by \\ & $MTcheck()$. \\ \noalign{\smallskip} $Prune()$: & Prunes based on a threshold. \\ \noalign{\smallskip} $BackProp()$: & Conducts a single back-propagation for training. \\ \hline \vspace{-1cm} \end{tabular} \end{table} \subsection{The Pruning Workflow} The pruning pipeline for Global MP is specified in Algorithm \ref{algo1}. It consists of taking a starting model, pruning it until the desired sparsity target is met and training or fine-tuning it for the specified number of epochs. It supports both one-shot and gradual pruning settings as well as with or without MT. The users may choose any pruning setting as per their use-case. The procedure starts by first taking a pre-trained model for the case of one-shot pruning or untrained model for the case of gradual pruning. Next, the sparsity of the model is checked and if the sparsity is lower than the target sparsity, then the model is pruned using either vanilla Global MP or Global MP with MT, as per the choice of the user. Once, the model is pruned then it is trained for the case of gradual pruning or fine-tuned for the case of one-shot pruning. The above procedure repeats until the final epoch is reached. For the case of one-shot pruning, the later epochs are just used for doing fine-tuning as the pruning happens in one-go in the first epoch itself. This finishes the procedure and the final result is a pruned and trained (or fine-tuned) model. \section{Experiments} \label{Results} Below we describe experiments related to Global Magnitude Pruning (Global MP) compared to state-of-the-art (SOTA) pruning algorithms. We conduct experiments on well-known image classification datasets, such as CIFAR-10 and ImageNet. We also include a human activity recognition dataset (HAR-2) to demonstrate generalization to other domains. We report hyper-parameters and training-related information for all the experiments in supplementary materials (Section A). \subsection{Comparison with SOTA} We compare Global MP with various popular SOTA algorithms that are well known for pruning including SNIP \cite{lee2018snip}, SM \cite{sparse_momentum}, DSR \cite{ICML-2019-MostafaW}, DPF \cite{Lin2020Dynamic}, GMP \cite{Zhu2018ToPO}, DNW \cite{DNW}, RigL \cite{pmlr-v119-evci20a}, and STR \cite{pmlr-v119-kusupati20a}. These include a broad spectrum of methods involving iteratively pruning and re-growing weights every few iterations, pruning at initialization, using gradients and feedback signals for pruning and pruning using regularization. We report results from these algorithms whenever they report results for the specific dataset that is being experimented upon. We report performance on weight sparsity (i.e., the number of parameters pruned) vs. accuracy, the default metric reported by all pruning papers, for all our experiments. \subsubsection{CIFAR-10} \label{sota_cifar10} We conduct experiments to compare Global MP to SOTA pruning algorithms on the CIFAR-10 dataset, which features 60,000 tiny, 32$\times$32-sized RGB images with 10 classes. It is a commonly used dataset for benchmarking DNN pruning algorithms. We compare Global MP with various algorithms including SNIP \cite{lee2018snip}, SM \cite{sparse_momentum}, DSR \cite{ICML-2019-MostafaW}, and DPF \cite{Lin2020Dynamic}. We report results on two popular and widely pruned network architectures, namely, WideResNet-28-8 (WRN-28-8) and ResNet-32 \cite{DBLP:journals/corr/HeZRS15}. For both architectures, we start off with the original model having the same initial accuracy as the other algorithms to have a fair comparison. Table~\ref{table:wrn-28-8_cifar_10} includes the results for WRN-28-8 experiments. As can be seen, Global MP performs better than the rest of the competitors at 90\% and 95\% sparsity levels. At 97.5\% sparsity level, Global MP is the second-best algorithm with a very small margin after a strong competitor DPF, which takes the second-best place in other two target sparsity levels. As for ResNet-32, since it is a smaller network with less redundancy, we conduct experiments only up to 95\% sparsity. Table \ref{table:resnet32_cifar10} depicts these results. Similar to the results in WRN-28-8, Global MP and DPF take the first two places at both 90\% and 95\% sparsity levels, while margins are being very small in between. This is an indication of the capabilities of Global MP as compared to the other algorithms, while featuring no added complexity. \begin{table}[t!] \small \centering \begin{tabular}{p{1.5cm}p{2.1cm}p{1.1cm}p{1.1cm}} \toprule \multirow{1}{*}{Method} & Top-1 Acc & Params & Sparsity\\ \midrule WRN-28-8 & 96.06\% & 23.3M\ & 0.0\%\\ \midrule SNIP & $95.49 \pm 0.21\%$ & 2.33M\ & 90\%\\ SM & $95.67 \pm 0.14\%$ & 2.33M\ & 90\%\\ DSR & $95.81 \pm 0.10\%$ & 2.33M\ & 90\%\\ \underline{DPF} & $\underline{96.08 \pm 0.15\%}$ & 2.33M\ & 90\%\\ \textbf{Global MP} & $\textbf{96.30} \pm \textbf{0.03\%}$ & 2.33M & 90\%\\ \midrule SNIP & $94.93 \pm 0.13\%$ & 1.17M\ & 95\%\\ SM & $95.64 \pm 0.07\%$ & 1.17M\ & 95\%\\ DSR & $95.55 \pm 0.12\%$ & 1.17M\ & 95\%\\ \underline{DPF} & $\underline{95.98 \pm 0.10\%}$ & 1.17M\ & 95\%\\ \textbf{Global MP} & $\textbf{96.16} \pm \textbf{0.02\%}$ & 1.17M\ & 95\%\\ \midrule SNIP & $94.11 \pm 0.19\%$ & 0.58M\ & 97.5\%\\ SM & $95.31 \pm 0.20\%$ & 0.58M\ & 97.5\%\\ DSR & $95.11 \pm 0.07\%$ & 0.58M\ & 97.5\%\\ \textbf{DPF} & $\textbf{95.84} \pm \textbf{0.04\%}$ & 0.58M\ & 97.5\%\\ \underline{Global MP} & $\underline{95.68 \pm 0.08\%}$ & 0.58M\ & 97.5\%\\ \bottomrule \end{tabular} \vspace{3pt} \captionof{table}{Results of SOTA pruning algorithms on WideResNet-28-8 on CIFAR-10. Global MP outperforms or yields comparable performance to other algorithms.} \label{table:wrn-28-8_cifar_10} \end{table} \begin{table}[t!] \vspace{0pt} \small \centering \begin{tabular}{p{1.5cm}p{2.3cm}p{1.1cm}p{1.1cm}} \toprule \multirow{1}{*}{Method} & Top-1 Acc & Params. & Sparsity\\ \midrule ResNet-32 & 93.83 $\pm$ 0.12 \% & 0.46M\ & 0.00\%\\ \midrule SNIP & 90.40 $\pm$ 0.26\% & 0.046M\ & 90\%\\ SM & 91.54 $\pm$ 0.18\% & 0.046M\ & 90\%\\ DSR & 91.41 $\pm$ 0.23\% & 0.046M\ & 90\%\\ \underline{DPF} & \underline{92.42 $\pm$ 0.18\%} & 0.046M\ & 90\%\\ \textbf{Global MP} & $\textbf{92.67} \pm \textbf{0.03\%}$ & 0.046M\ & 90\%\\ \midrule SNIP & 87.23 $\pm$ 0.29\% & 0.023M\ & 95\%\\ SM & 88.68 $\pm$ 0.22\% & 0.023M\ & 95\%\\ DSR & 84.12 $\pm$ 0.32\% & 0.023M\ & 95\%\\ \textbf{DPF} & \textbf{90.94 $\pm$ 0.35\%} & 0.023M\ & 95\%\\ \underline{Global MP} & \underline{90.65 $\pm$ 0.13\%} & 0.023M\ & 95\%\\ \bottomrule \end{tabular} \vspace{3pt} \captionof{table}{Results of pruning algorithms on ResNet-32 on CIFAR-10. Global MP outperforms or yields comparable performance to other algorithms.} \label{table:resnet32_cifar10} \end{table} \subsubsection{ImageNet} \label{sota_imagenet} Following the favorable performance on CIFAR-10 dataset, we benchmark Global MP against other competitors in the literature over ImageNet dataset, also known as ILSVRC 2012 dataset. This is a highly compelling dataset as compared to CIFAR-10, featuring around 1.3 million RGB images with 1,000 classes. In the pruning context, it typically serves as an ultimate benchmark and has been utilized by many other papers during comparison. Using this dataset, we compare Global MP with SOTA algorithms like GMP \cite{Zhu2018ToPO}, DSR \cite{ICML-2019-MostafaW}, DNW \cite{DNW}, SM \cite{sparse_momentum}, RigL \cite{pmlr-v119-evci20a}, DPF \cite{Lin2020Dynamic} and STR \cite{pmlr-v119-kusupati20a}. The two network architectures that we use for this comparison are ResNet-50 and MobileNet-V1 \cite{howard2017mobilenets}, the two most popular architectures for benchmarking on ImageNet \cite{blalock2020state}. We again start from the same initial accuracy for the non-pruned models for all algorithms, either by matching the results in their original papers or reproducing their results whenever their code is available. The remarkable performance of Global MP becomes clearly visible in ResNet-50 over ImageNet experiments. As can be seen from Table~\ref{table:resnet50_imagenet}, Global MP outperforms all the other competitors in every weight sparsity level from 80\% to 98\%. The first two places at each of these target sparsity levels belong to either the gradual version or the one-shot version of Global MP. Different from CIFAR-10 results, the performance margins that Global MP surpasses the others are also fairly high in ImageNet experiments. For instance, Global MP (Gradual) yields about 2\% higher accuracy at 95\% target sparsity level, while this number goes up to about 5\% at 98\% target sparsity level. This is an important finding that such a simple algorithm like Global MP can highly outperform other competitors that incorporates very complex design choices or computationally demanding procedures. \begin{table}[t!] \small \centering \begin{tabular}[t]{p{3cm}p{0.9cm}p{0.8cm}p{1cm}p{0.8cm}} \toprule \multirow{1}{*}{Method} & Top-1 Acc & Params & Sparsity & FLOPs pruned\\ \midrule ResNet-50 & 77.0\% & 25.6M\ & 0.00\% & 0.0\%\\ \midrule GMP & 75.60\% & 5.12M\ & 80.00\% & 80.0\%\\ DSR*\#\ & 71.60\% & 5.12M\ & 80.00\% & 69.9\%\\ DNW & 76.00\% & 5.12M\ & 80.00\% & 80.0\%\\ SM & 74.90\% & 5.12M\ & 80.00\% & -\\ SM + ERK & 75.20\% & 5.12M\ & 80.00\% & 58.9\%\\ RigL* & 74.60\% & 5.12M\ & 80.00\% & 77.5\%\\ RigL + ERK & 75.10\% & 5.12M\ & 80.00\% & 58.9\%\\ DPF & 75.13\% & 5.12M\ & 80.00\% & 80.0\%\\ STR & 76.19\% & 5.22M\ & 79.55\% & 81.3\%\\ \textbf{Global MP (One-shot)} & \textbf{76.84\%} & 5.12M & 80.00\% & 72.4\%\\ \underline{Global MP (Gradual)} & \underline{76.12\%} & 5.12M & 80.00\% & 76.7\%\\ \midrule GMP & 73.91\% & 2.56M\ & 90.00\% & 90.0\%\\ DNW & 74.00\% & 2.56M\ & 90.00\% & 90.0\%\\ SM & 72.90\% & 2.56M\ & 90.00\% & 60.1\%\\ SM + ERK & 72.90\% & 2.56M\ & 90.00\% & 76.5\%\\ RigL* & 72.00\% & 2.56M\ & 90.00\% & 87.4\%\\ RigL + ERK & 73.00\% & 2.56M\ & 90.00\% & 76.5\%\\ DPF\# & 74.55\% & 4.45M\ & 82.60\% & 90.0\%\\ STR & 74.73\% & 3.14M\ & 87.70\% & 90.2\%\\ \textbf{Global MP (One-shot)} & \textbf{75.28\%} & 2.56M & 90.00\% & 82.8\%\\ \underline{Global MP (Gradual)} & \underline{74.83\%} & 2.56M & 90.00\% & 87.8\%\\ \midrule GMP & 70.59\% & 1.28M\ & 95.00\% & 95.0\%\\ DNW & 68.30\% & 1.28M\ & 95.00\% & 95.0\%\\ RigL* & 67.50\% & 1.28M\ & 95.00\% & 92.2\%\\ RigL + ERK & 70.00\% & 1.28M\ & 95.00\% & 85.3\%\\ STR & 70.97\% & 1.33M\ & 94.80\% & 95.6\%\\ STR & 70.40\% & 1.27M\ & 95.03\% & 96.1\%\\ STR & 70.23\% & 1.24M\ & 95.15\% & 96.0\%\\ \underline{Global MP (One-shot)} & \underline{71.56\%} & 1.20M & 95.30\% & 89.3\%\\ \textbf{Global MP (Gradual)} & \textbf{72.14\%} & 1.20M & 95.30\% & 93.1\%\\ \midrule GMP & 57.90\% & 0.51M\ & 98.00\% & 98.0\%\\ DNW & 58.20\% & 0.51M\ & 98.00\% & 98.0\%\\ STR & 61.46\% & 0.50M & 98.05\% & 98.2\%\\ \underline{Global MP (One-shot)} & \underline{61.80}\% & 0.50M & 98.05\% & 93.7\%\\ \textbf{Global MP (Gradual)} & \textbf{66.57}\% & 0.50M & 98.05\% & 96.2\%\\ \bottomrule \end{tabular} \vspace{3pt} \captionof{table}{Results on ResNet-50 on ImageNet. Global MP outperforms SOTA pruning algorithms at all sparsity levels. * and \# imply the first and the last layer are dense, respectively.} \label{table:resnet50_imagenet} \end{table} \begin{figure}[t!] \begin{minipage}[t]{0.49\textwidth} \centering \includegraphics[trim={0 0 0 1.3cm}, clip, width=\textwidth]{Figures/ResNet50_results.pdf} \caption{Global MP surpasses all the unstructured pruning baselines at all sparsity ratios on ResNet-50 over ImageNet, showcasing that it is state-of-the-art for weight sparsity.} \label{fig:resnet50_sparsity} \end{minipage} \end{figure} We also test another architecture on ImageNet, MobileNet-V1, which is a much smaller and more efficient architecture than ResNet-50. In this case, strong competitors are limited in the literature; only two of the aforementioned algorithms are able to present competitive results due to the fact that this architecture has less redundancy. We benchmark Global MP with two other competitors at two target sparsity levels: 75\% and 90\%. As can be seen in Table~\ref{table:mobilenetv1_imagenet}, Global MP outperforms SOTA algorithms by a margin of more than 2\% at 75\% sparsity, which is a significant result given how compact the MobileNet-V1 is. At 90\% sparsity on the other hand, the same compactness causes Global MP to over-prune certain layers in the network, which result in a significant accuracy drop. This is the above-mentioned problem of layer-collapse, and it is easily rectified when MT is introduced to Global MP. We use an MT value of 0.2\% which is determined using the same search procedure as any other hyper-parameter. The accuracy of Global MP at 90\% sparsity goes beyond SOTA again with such a simple fix, and the accuracy margin to the next competitor gets higher than 2\%. MT comes at the cost of a less FLOPs reduction, but it is useful especially for accuracy-critical applications where decreasing the size of the network is still important. All these findings clearly indicates that Global MP is a simple yet competitive pruning algorithm. \begin{table}[t!] \centering \small \begin{tabular}{p{3cm}p{0.9cm}p{0.8cm}p{1cm}p{0.8cm}} \toprule \multirow{1}{*}{Method} & Top-1 Acc & Params. & Sparsity & FLOPs pruned\\ \midrule MobileNet-V1 & 71.95\% & 4.21M\ & 0.00\% & 0.0\%\\ \midrule GMP & 67.70\% & 1.09M\ & 74.11\% & 71.4\%\\ \underline{STR} & \underline{68.35\%} & 1.04M\ & 75.28\% & 82.2\%\\ \textbf{Global MP} & \textbf{70.74\%} & 1.04M & 75.28\% & 68.9\%\\ \midrule GMP & 61.80\% & 0.46M\ & 89.03\% & 85.6\%\\ STR & 61.51\% & 0.44M\ & 89.62\% & 93.0\%\\ \underline{Global MP} & \underline{59.49\%} &0.42M\ & 90.00\% & 83.7\%\\ \textbf{Global MP with MT} & \textbf{63.94\%} & 0.42M & 90.00\% & 72.9\%\\ \bottomrule \end{tabular} \captionof{table}{Results of pruning algorithms on MobileNet-V1 on ImageNet. Global MP with MT surpasses SOTA algorithms on weight sparsity.} \label{table:mobilenetv1_imagenet} \end{table} \subsection{Generalizing to other domains and RNN architectures} \label{rnn} We experiment with Global MP on other domains and non-convolutional networks as well to measure the generalizability of the algorithm on different domains and network types. We experiment on a FastGRNN model \cite{Kusupati2018FastGRNNAF} on the HAR-2 Human Activity Recognition dataset \cite{HAR}. HAR-2 dataset is a binarized version of the 6-class Human Activity Recognition dataset. From the full-rank model with $r_W = 9$ and $r_U = 80$ as suggested on the STR paper \cite{pmlr-v119-kusupati20a}, we apply Global MP on the matrices $W_1$ and $W_2$. To do this, we find the weight mask by ranking the columns of $W_1$ and $W_2$ based on their absolute sum, then we prune the $9 - r_W^{new}$ lowest columns and $80 - r_U^{new}$ lowest columns from $W_1$ and $W_2$ respectively. In the end, we fine-tune this pruned model by retraining it with FastGRNN's trainer and applying the weight mask at every epoch. We test Global MP under different network configurations. We find that Global MP surpasses the other baselines on all the configurations (Table \ref{table:fastgrnn_har2}) and successfully prunes the model on a very different architecture and domain. \subsection{Mitigating Layer-Collapse} \label{high_sparsity} Layer-collapse is an issue that many pruning algorithms run into \cite{NEURIPS2020_46a4378f, Lee2020A, hayou2021robust} and occurs when an entire layer is pruned by the pruning algorithm, rendering the network untrainable. We investigate this phenomena and find that the performance of a pruning algorithm can be substantially affected by the architecture of the neural network being pruned, especially in the high sparsity domain. We conduct experiments on MobileNet-V2 and WRN-22-8 models over the CIFAR-10 dataset. We report results averaged over multiple runs where each run uses a different pre-trained model to provide more robustness. We first prune a WRN-22-8 model to 99.9\% sparsity. We find that at 99.9\% sparsity, the WRN is still able to get decent accuracy (Table \ref{table:highsparsity_wrn}). We then prune a MobileNet-V2 model to 98\% sparsity. For the MobileNet, however, accuracy drops to 10\% using only Global MP, and the model is not able to learn (Table \ref{table:highsparsity_mnet}). \begin{figure*}[h!] \centering \includegraphics[trim=3cm 2.7cm 12cm 2cm,clip,width=0.7\textwidth]{Figures/architectures.pdf} \caption{Difference in architectures between WRN and MobileNet. WRN does not have any prunable residual connections in the last layers (dotted lines) while MobileNet does. This leads to different pruning behaviors on the two architectures.} \vspace*{-0.3cm} \label{fig:highsparsity_archs} \end{figure*} \begin{figure*}[h!] \centering \includegraphics[trim=1.5cm 11cm 1.5cm 11cm,clip,width=\textwidth]{Figures/MobileNetv2-remaining_weights_93.89.pdf} \vspace*{-0.9cm} \caption{For MobileNet-V2 at 98\% sparsity, MT helps retain some weights in the heavily pruned layers (Layers 55, 56, and 57) and allows the model to learn successfully.} \vspace*{-0.3cm} \label{fig:highsparsity_mnet} \end{figure*} \begin{figure*}[h!] \centering \includegraphics[trim=1.5cm 11cm 1.5cm 11cm,clip,width=\textwidth]{Figures/MobileNetv2-remaining_weights_3runs_GP.pdf} \vspace*{-0.9cm} \caption{Layer-wise pruning results produced by Global MP on MobileNet-V2 model on CIFAR-10. Pruning is conducted on three different pre-trained models and the pruning results across the three runs are very stable.} \vspace*{-0.3cm} \label{fig:highsparsity_mnet_gp} \end{figure*} The reason for this wide discrepancy in learning behavior lies in the shortcut connections \cite{DBLP:journals/corr/HeZRS15}. Both WRN-22-8 and MobileNet-V2 use shortcut connections, however, their placement is different. Referring to Fig. \ref{fig:highsparsity_archs}, WRN uses identity shortcut connections from Layer 20 to Layer 23. This type of shortcut connections are simple identity mappings and do not require any extra parameters, and hence, they do not count towards the weights. However, MobileNet-V2 uses a convolutional shortcut mapping from Layer 52 to Layer 57 and hence, it adds to the model's weights, and thus, it is prunable layer for the pruning algorithm. Global MP completely prunes the two preceding layers before the last layer. However, because WRN uses identity mappings, it is still able to relay information to the last layer, and the model is still able to learn, whereas MobileNet-V2 faces catastrophic accuracy drop due to layer-collapse. Pruning algorithms can be susceptible to such catastrophic layer-collapse issues especially in the high sparsity domain. The MT rule can help overcome this issue. Retaining a small MT of 0.02\% was sufficient for the MobileNet-V2 model to avoid layer-collapse and learn successfully. We provide layer-wise weight snapshot for the model, before and after applying MT, to illustrate what MT does (Fig. \ref{fig:highsparsity_mnet}). Hence, retaining a small amount of weights can help in the learning dynamics of models in high sparsity settings. \begin{table}[t!] \centering \small \begin{tabular}{p{2.6cm}p{1.45cm}p{0.5cm}p{0.5cm}} \toprule \multirow{1}{*}{Method} & Top-1 Acc & $r_W$ & $r_U$\\ \midrule FastGRNN & 96.10\% & 9\ & 80\\ \midrule Vanilla Training & 94.06\% & 9\ & 8\\ \underline{STR} & \underline{95.76\%} & 9\ & 8\\ \textbf{Global MP} & \textbf{95.89}\% & 9 & 8\\ \midrule Vanilla Training & 93.15\% & 9\ & 7\\ \underline{STR} & \underline{95.62\%} & 9\ & 7\\ \textbf{Global MP} & \textbf{95.72}\% & 9 & 7\\ \midrule Vanilla Training & 94.88\% & 8\ & 7\\ \underline{STR} & \underline{95.59\%} & 8\ & 7\\ \textbf{Global MP} & \textbf{95.62}\% & 8 & 7\\ \bottomrule \end{tabular} \captionof{table}{Results on FastGRNN on HAR-2 dataset. Global MP outperforms other pruning algorithms.} \label{table:fastgrnn_har2} \end{table} \begin{table}[t!] \small \begin{tabular}{p{1.5cm}p{0.8cm}p{2.2cm}p{2.2cm}} \toprule \multirow{2}{*}{Method} & \multicolumn{2}{c}{WRN-22-8 on CIFAR-10} \\ \cmidrule(lr){2-4} & Sparsity & Starting Acc. & Pruned Acc.\\ \midrule \textbf{Global MP} & 99.9\% & $94.07\% \pm 0.05\%$ & $\textbf{67.68\%} \pm \textbf{0.78\%}$ \\ \bottomrule \end{tabular} \captionof{table}{Performance of Global MP on WideResNet-22-8 in the high sparsity regime at 99.9\% sparsity.} \label{table:highsparsity_wrn} \end{table} \begin{table}[t!] \small \centering \begin{tabular}{m{1.5cm}m{0.7cm}m{2.2cm}m{2.2cm}} \toprule \multirow{2}{*}{Method} & \multicolumn{2}{c}{MobileNet-V2 on CIFAR-10} \\ \cmidrule(lr){2-4} & Sparsity & Starting Acc. & Pruned Acc.\\ \midrule Global MP & 98.0\% & $94.15\%\pm0.23\%$ & $10\%$ \textit{(Unable to learn)} \\ \textbf{Global MP with MT} & 98.0\% & $94.15\% \pm 0.23\%$ & $\textbf{82.97\%} \pm \textbf{0.57\%}$\\ \bottomrule \end{tabular} \captionof{table}{Adding MT enables the MobileNet-V2 model to learn in the high sparsity regime.} \label{table:highsparsity_mnet} \end{table} \section{Discussion, Limitations and Future Work} We have seen that Global MP works very well and achieves superior performance on all the datasets and architectures tested. It can work as a one-shot pruning algorithm or as a gradual pruning algorithm. It is very stable and produces similar pruning results across multiple runs and pre-trained models (Fig. \ref{fig:highsparsity_mnet_gp}). It also surpasses SOTA algorithms on ResNet-50 over ImageNet and sets the new SOTA results across many sparsity levels. At the same time, Global MP has very low algorithmic complexity and arguably is one of the simplest pruning algorithms. It is simpler than many other pruning algorithms like custom loss based regularization, RL-based procedures, heuristics-based layerwise pruning ratios, etc. It just ranks weights based on their magnitude and removes the smallest ones. This raises a key question on whether complexity is really required for pruning and if complex pruning algorithms have tangible advantages over naive baselines. According to our results, the advantages seem to be narrow. The only advantages maybe that while Global MP gets competitive FLOPs performance, some algorithms may have higher FLOPs sparsity, though at the cost of accuracy. Therefore, practitioners may opt for another algorithm to get SOTA FLOPs performance if the accuracy loss incurred is reasonable for their application. A limitation of Global MP is that the theoretical foundations for it have not been well-established yet. While empirically it gets superior performance, we still do not understand mathematically why this is the case. A richer understanding of the dynamics of Global MP can enable researchers to build upon it and further improve its performance, and it is an area for future work. Another area for future work is jointly optimizing both weights and FLOPs during the pruning process. Currently, Global MP is used to reach a certain parameter sparsity, and FLOPs reduction comes as a by-product. In the future, FLOPs can also be added to the optimization function to jointly sparsify both parameters and FLOPs. This can lead to further gains in the FLOPs performance of Global MP. \section{Conclusions} \label{conclusion} In this work, we raised the question of whether utilizing complex and computationally demanding algorithms are really required to achieve superior DNN pruning results. This stemmed from the hike in the number of new pruning algorithms proposed in the recent years, each with a marginal performance increment, though with complicated procedures, which makes it hard for a practitioner to select the correct algorithm and the best set of algorithm-specific hyper-parameters for their application. We benchmarked these algorithms against a naive baseline, namely, Global MP, which does not incorporate any complex procedure or any hard-to-tune hyper-parameter. Despite its simplicity, we found that Global MP outperforms many SOTA pruning algorithms over multiple datasets such as CIFAR-10, ImageNet and HAR-2; with different network architectures such as ResNet-50 and MobileNet-V1; and at various sparsity levels from 50\% up to 99.9\%. We also presented a few variants of Global MP, i.e., one-shot and gradual, together with a new, complementary technique, MT. We demonstrated that through the selection of an appropriate variant of Global MP, the performance at different metrics, i.e., accuracy vs. weight sparsity or accuracy vs. FLOPs sparsity, can be maximized. While our results serves as an empirical proof that a naive pruning algorithm like Global MP can achieve SOTA results, it remains as a promising future research direction to shed light into theoretical aspects of how such performance is possible with Global MP. Another future direction includes extending the capabilities of Global MP, such as jointly optimizing both FLOPs and the number of weights. \section{Acknowledgement} This research is supported by the Agency for Science, Technology and Research (A*STAR) under its AME Programmatic Funds (Project No: A1892b0026 and A19E3b0099). Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not reflect the views of the A*STAR. \subsection{Hyper-parameters and Experimental Setup} \label{hyperparams} No data augmentation is done apart from standard data pre-processing. Difference in batch size for training and testing in some experiments is due to GPU RAM availability. Averaged results are reported over three runs. \begin{table*}[h] \vspace{-1cm} \centering \small \begin{tabular}{p{2cm}p{2cm}p{2cm}} \toprule \multicolumn{3}{c}{MT Values} \\ \cmidrule(lr){1-3} Experiment & Sparsity & MT Value\\ \midrule Table 4 & 90\% & 0.2\% \\ \midrule Table 7 & 98\% & 0.02\% \\ \bottomrule \end{tabular} \label{table:MT_values} \end{table*} \begin{table*}[hbt!] \vspace{-0.2cm} \centering \small \begin{tabular}{lccccccc} \toprule \multicolumn{8}{c}{\centering{Setup for Table 2}} \\ \cmidrule(lr){1-8} \multirow{2}{*}{Stage} & \multirow{2}{*}{Epochs} & Batch & \multirow{2}{*}{Momentum} & Weight & Initial & LR & \multirow{2}{*}{Nesterov}\\ & & Size & & Decay & LR & Scheduler & \\ \midrule Training & 200 & 128 & 0.875 & 5e-4 & 0.1 & Cosine & Yes\\ Finetuning (GP 90\%) & 200 & 128 & 0.9 & 0 & 0.0512 & Cosine & Yes\\ Finetuning (GP 95\%) & 200 & 128 & 0.9 & 2e-5 & 0.0256 & Cosine & Yes\\ Finetuning (GP 97.5\%) & 200 & 128 & 0.9 & 0 & 0.0128 & Cosine & Yes\\ \bottomrule \end{tabular} \end{table*} \begin{table*}[hbt!] \vspace{-0.1cm} \centering \small \begin{tabular}{lccccccc} \toprule \multicolumn{8}{c}{\centering{Setup for Table 3}} \\ \cmidrule(lr){1-8} \multirow{2}{*}{Stage} & \multirow{2}{*}{Epochs} & Batch & \multirow{2}{*}{Momentum} & Weight & Initial & LR & \multirow{2}{*}{Nesterov}\\ & & Size & & Decay & LR & Scheduler & \\ \midrule Training & 300 & 128 & 0.9 & 0.001 & 0.05 & Cosine & No\\ Finetuning (GP 90\%) & 300 & 128 & 0.9 & 0.001 & 0.01 & Cosine & No\\ Finetuning (GP 95\%) & 300 & 128 & 0.9 & 1e-5 & 0.01 & Cosine & No\\ \bottomrule \end{tabular} \end{table*} \begin{table*}[hbt!] \vspace{-0.1cm} \centering \small \begin{tabular}{lccccccc} \toprule \multicolumn{8}{c}{\centering{Setup for Table 4}} \\ \cmidrule(lr){1-8} \multirow{2}{*}{Stage} & \multirow{2}{*}{Epochs} & Batch & \multirow{2}{*}{Momentum} & Weight & Initial & LR & Label\\ & & Size & & Decay & LR & Scheduler & Smoothing \\ \midrule Training & 100 & 256 & 0.875 & 0.000031 & 0.256 & Cosine (warmup = 5) & 0.1\\ One-shot GP 80\% & 100 & 256 & 0.875 & 0.000023 & 0.0256 & Cosine (warmup = 5) & 0.1\\ Gradual GP 80\% & 100 & 256 & 0.875 & 0.000031 & 0.256 & Cosine (warmup = 5) & 0.1\\ One-shot GP 90\% & 100 & 256 & 0.875 & 0.000007 & 0.1024 & Cosine (no warmup) & 0.1\\ Gradual GP 90\% & 100 & 256 & 0.875 & 0.000031 & 0.256 & Cosine (warmup = 5) & 0.1\\ One-shot GP 95.3\% & 100 & 256 & 0.95 & 0.0 & 0.0512 & Cosine (no warmup) & 0.05\\ Gradual GP 95.3\% & 100 & 256 & 0.875 & 0.000031 & 0.256 & Cosine (warmup = 5) & 0.1\\ One-shot GP 98.05\% & 100 & 256 & 0.95 & 0.0 & 0.0512 & Cosine (no warmup) & 0.05\\ Gradual GP 98.05\% & 100 & 256 & 0.875 & 0.000031 & 0.256 & Cosine (warmup = 5) & 0.1\\ \bottomrule \end{tabular} \end{table*} \begin{table*}[hbt!] \vspace{-0.1cm} \centering \small \begin{tabular}{lccccccc} \toprule \multicolumn{8}{c}{\centering{Setup for Table 5}} \\ \cmidrule(lr){1-8} \multirow{2}{*}{Stage} & \multirow{2}{*}{Epochs} & Batch & \multirow{2}{*}{Momentum} & Weight & Initial & LR & Label\\ & & Size & & Decay & LR & Scheduler & Smoothing \\ \midrule Training & 100 & 256 & 0.875 & 3.1e-5 & 0.256 & Cosine (warmup=5) & 0.1\\ Finetuning (GP 75\%) & 120 & 256 & 0.875 & 1e-5 & 0.0512 & Cosine (no warmup) & 0.1\\ Finetuning (GP 90\%) & 120 & 256 & 0.875 & 1e-5 & 0.0256 & Cosine (no warmup) & 0.1\\ Finetuning (GP + MT 90\%) & 120 & 256 & 0.875 & 0 & 0.0512 & Cosine (no warmup) & 0.1\\ \bottomrule \end{tabular} \end{table*} \begin{table*}[hbt!] \vspace{-0.1cm} \centering \small \begin{tabular}{lccccc} \toprule \multicolumn{6}{c}{\centering{Setup for Table 6}} \\ \cmidrule(lr){1-6} Stage & Epochs & Batch Size & Initial LR & hd & Optimizer \\ \midrule Training & 300 & 100 & 0.0064 & 80 & Adam\\ Finetuning (GP 9,8) & 300 & 64 & 0.5 & 80 & Adam\\ Finetuning (GP 9,7) & 300 & 100 & 0.5 & 80 & Adam\\ Finetuning (GP 8,7) & 300 & 100 & 0.55 & 80 & Adam\\ \bottomrule \end{tabular} \end{table*} \begin{table*}[hbt!] \vspace{-0.1cm} \centering \small \begin{tabular}{lccccccc} \toprule \multicolumn{8}{c}{\centering{Setup for Table 7}} \\ \cmidrule(lr){1-8} \multirow{2}{*}{Stage} & \multirow{2}{*}{Epochs} & Batch & \multirow{2}{*}{Momentum} & Weight & Initial & LR & \multirow{2}{*}{Nesterov}\\ & & Size & & Decay & LR & Scheduler & \\ \midrule Training & 30 & 256 & 0.9 & 5e-4 & 0.1 & Step decay (Step size 25, gamma 0.1) & Yes\\ Finetuning (GP 99.9\%) & 80 & 64 & 0.9 & 5e-4 & 0.1 & Step decay (Step size 40, gamma 0.1) & Yes\\ \bottomrule \end{tabular} \end{table*} \begin{table*}[t!] \centering \small \begin{tabular}{lccccccc} \toprule \multicolumn{8}{c}{\centering{Setup for Table 8}} \\ \cmidrule(lr){1-8} \multirow{2}{*}{Stage} & \multirow{2}{*}{Epochs} & Batch & \multirow{2}{*}{Momentum} & Weight & Initial & LR & \multirow{2}{*}{Nesterov}\\ & & Size & & Decay & LR & Scheduler & \\ \midrule Training & 200 & 450 & 0.9 & 5e-4 & 0.1 & Step decay (Step size 25, gamma 0.56) & Yes\\ Finetuning (GP 98\%) & 200 & 64 & 0.9 & 5e-4 & 0.1 & Step decay (Step size 25, gamma 0.56) & Yes\\ Finetuning (GP + MT 98\%) & 200 & 64 & 0.9 & 5e-4 & 0.1 & Step decay (Step size 25, gamma 0.56) & Yes\\ \bottomrule \end{tabular} \end{table*}

5b6e01a006ff4fc66907aeb3d053a7f11956201b

\section{Introduction} Random sampling is a fundamental area of study in statistics. Countless processes can be simulated as Monte Carlo experiments, with far-reaching applications in a range of fields from biology and economics to business. A rather simple but ubiquitous and surprisingly non-trivial experiment is that of taking $D$ samples from a distribution and picking out one of these samples according to some omniscient "strategy." This problem is dual to K-means clustering, a well-studied process where one groups a large set of samples into a fixed $K$ number of buckets in order to minimize an error function called the quadratic distortion (see \cite{liu2020convergence} for state-of-the-art). We frame our model in the realm of percolation theory and ask what is the limiting behaviour of empirical measures along the event-dependent path of choices. Answering this question provides a gold mine of information about the model - normalized passage time along a path, after all, is just the identity function integrated against the path's empirical measure. Recent efforts have gone into better understanding the weak convergence of empirical measures. Some, such as \cite{riekert2021wasserstein} and \cite{lei2020convergence}, establish upper bounds on the \mbox{rate} of Wasserstein convergence of empirical measures to the Markov invariant measure/i.i.d. measure respectively they are sampled from. Others focus on describing the limits of empirical measures. In \cite{bates}, Bates shows that the set $\mathcal{R}$ of limit points of normalized empirical measures along paths in a generic lattice model is deterministic and derives an explicit variational formula for the limit shape of first passage time as the minimum value of a linear functional over this deterministic set. In \cite{gatea}, this author builds on Bates' paper by developing the notion of grid entropy, a deterministic quantity capturing not just whether a certain target measure is a limit point of empirical measures. Grid entropy has previously appeared in \cite{rassoul2014quenched}. \cite{gatea} proposes a novel approach, realizing grid entropy both as a subadditive limit of entropies of paths with empirical measure within an $\epsilon$-Levy-Prokhorov distance of the target measure, and as the critical exponent of canonical order statistics associated with the Levy-Prokhorov metric, something a priori only known for i.i.d. Bernoulli$p)$ edge labels (see \cite{carmona2010directed}). Though using different approaches, both this author and Rassoul-Agha et al establish the same convex duality between grid entropy and Gibbs Free Energy. Furthermore, this author observes that Bates' set $\mathcal{R}$ almost surely coincides with the set of probability measures with finite grid entropy: \[ \mathcal{R} = \{\nu \ \mbox{prob meas}: ||\nu|| > -\infty\} \ \mbox{a.s.} \] One limitation of \cite{gatea} is that only empirical measures along deterministic or polymer paths are considered, rather than allowing for arbitrary mixtures of paths. Even in \cite{bates}, the focus is mainly on empirical measures along geodesics. In this paper, we seek to rectify this by studying empirical measures along \emph{random} paths (picked according to some probabilistic "strategy," which may or may not yield geodesics or polymer paths). We will show that $\mathcal{R}$ is still the set of limit points, whether or not we assume the strategy is omniscient. Another deficiency is that grid entropy and the set $\mathcal{R}$, like limit shape and other quantities in this area, are not known to be explicitly computable in most cases. One noteworthy exception is the paper \cite{martinAllan} by Martin, in which an explicit formula for the weak limits of empirical measures along geodesics is derived in the solvable Exponential LPP on $ {\mathbb{Z}} ^2$ model. Also, in \cite{bates} it is established that replacing the lattice $ {\mathbb{Z}} ^D$ by the infinite complete $D$-ary tree $\mathcal{T}_D$, the set $\mathcal{R}$ can be precisely described as a specific sublevel of relative entropy (in terms of $D$). The $D$-ary tree is in a way the dual model to the one we explore in this paper, and as one might expect, we will be able to give a description of $\mathcal{R}$ in our setting too. We will precisely characterize the extreme points of $\mathcal{R}$ as those measures whose density is $D \cdot Beta(1,D)$ distributed; $\mathcal{R}$ is then the closed convex hull of these measures. In addition, we compute the grid entropy of the extreme points of $\mathcal{R}$ to be 0, and we give a simplification of a general formula for grid entropy from \cite{gatea} for this model. \section{Definitions and Results} Let us be more precise about our setup and goals. We consider vertices on $ {\mathbb{Z}} _{\geq 0}$, $D$ parallel edges $(e_i^1,\ldots, e_i^D)$ for every $i \geq 0$, and an i.i.d. array of edge labels $U_i^j \sim $ Unif[0,1]. We denote by $\Lambda$ the Lebesgue measure on $[0,1]$. It is convenient to work on the compact space $\mathcal{M}_1$ of probability measures on [0,1], and as we will explain in Section \ref{coupling} we do not lose any generality by restricting to this setting. \begin{center} \begin{tikzpicture} \node (p1) at (0pt,0pt) {}; \node (p2) at (50pt,0pt) {}; \node (p3) at (100pt,0pt) {}; \filldraw (50pt,0pt)circle(2pt) (0pt,0pt)circle(2pt)(100pt,0pt)circle(2pt) (110pt,0pt)circle(2pt)(120pt,0pt)circle(2pt)(130pt,0pt)circle(2pt); \draw (25pt,0pt) node {$e_0^2$}; \draw (25pt,16pt) node [ above ] {$e_0^1$}; \draw (25pt,-16pt) node [ below ] {$e_0^3$}; \draw (75pt,0pt) node {$e_1^2$}; \draw (75pt,16pt) node [ above ] {$e_1^1$}; \draw (75pt,-16pt) node [ below ] {$e_1^3$}; \draw (0pt,0pt) node [ below right] {$0$}; \draw (50pt,0pt) node [ below right] {$1$}; \draw (100pt,0pt) node [ below right] {$2$}; \draw [->] (p1) to [out=90,in=90] (p2); \draw [->] (p1) to [out=-90,in=-90] (p2); \draw [->] (p2) to [out=90,in=90] (p3); \draw [->] (p2) to [out=-90,in=-90] (p3); \begin{scope}[every path/.style={->}] \draw (p1) -- (p2); \draw (p2) -- (p3); \end{scope} \end{tikzpicture} \end{center} Next, we wish to formalize the notion of strategies. These will be probability measures on the product space coupling the environment $(U_i^j)_{i \geq 0, 1 \leq j \leq D}$ with the infinite target sequence of indices of the $U_i$ corresponding to the choices $(X_i)_{i \geq 0}$. \begin{definition} A strategy is a probability measure $\chi$ on the product space\\ $([0,1]^D)^{ {\mathbb{Z}} _{\geq 0}} \times \{1,\ldots,D\}^{ {\mathbb{Z}} _{\geq 0}}$ s.t. the marginal distribution of the first coordinate (the environment $(U_i^j) \in ([0,1]^D)^{ {\mathbb{Z}} _{\geq 0}}$) is a sequence of i.i.d. Unif[0,1]. We denote by $(J_i)_{i \geq 0}$ the second coordinate (a random sequence of indices) and define the random vector of choices \[ (X_0(\chi), X_1(\chi),\ldots) := (U_0^{J_0}, U_1^{J_1},\ldots) \] Denote by $\frac1n \mu_{0 \rightarrow n}(\chi)$ the empirical measures of this vector: \[ \frac1n \mu_{0 \rightarrow n}(\chi) = \frac1n \sum \limits_{i=0}^{n-1} \delta_{X_i(\chi)} \in \mathcal{M}_1\] Also, let $\sigma_{\chi}^i$ be the law of $X_i(\chi)$. \end{definition} \begin{remark} If $\chi$ conditioned on the environment is a delta mass for a.a. environments $(U_i^j)$ with respect to the product measure $(\Lambda^{\times D})_{\infty}$, it means that the strategy picks exactly one sequence $(x_i)_{i \geq 0}$ for each set of observed labels. In other words, the strategy is deterministic. \end{remark} \begin{definition} We also define (strategy-free) empirical measures along a fixed path $\pi: 0 \rightarrow n$ consisting of edges $(e_0^{j_0},\ldots, e_{n-1}^{j_{n-1}})$ by \[ \frac1n \mu_{\pi} = \frac1n \sum_{i=0}^{n-1} \delta_{U_i^{j_i}} \] \end{definition} An important type of strategy is one which chooses each $X_k$ independent of $(U_i^j)_{i \neq k, 1 \leq j \leq D}$, the observed values from all but the $k$th trial. These are $\chi$ which are infinite products of measures arising from "single-edge strategies," i.e. micro-strategies operating at the individual trial level. We call such $\chi$ "independent strategies," as they give rise to independent $X_i$. \begin{definition} A single-edge strategy is a probability measure $\psi$ on the product space\\ $([0,1]^D) \times \{1,\ldots,D\}$ s.t. the marginal distribution of the first coordinate $(U^j)_{1 \leq j \leq D}$ is a sequence of $D$ i.i.d. Unif[0,1]. We denote by $J$ the second coordinate (a random index) and define the random choice \[ X(\psi):= U^J \] Also, let $\sigma_{\psi}$ be the law of $X(\psi)$. \end{definition} \begin{remark} If $\psi$ conditioned on $(U^j)_{1 \leq j \leq D}$ is a delta mass for a.a. $(u^j)_{1 \leq j \leq D}$ with respect to the product measure $\Lambda^{\times D}$, it means that the single-edge strategy picks exactly one sequence $x$ for each set of $D$ observed labels, so $\psi$ is deterministic. \end{remark} A single-edge strategy $\psi$ is completely determined by the $\psi$-conditional probabilities \\ $p_k: [0,1]^D \rightarrow [0,1]$, $1 \leq k \leq D$ defined as \[ p_k(u_1,\ldots, u_D) := P_{\psi}[J=k \mid (U^j)_{1 \leq j \leq D} = (u^j)_{1 \leq j \leq D} ] \] We expand on this in more detail in Section \ref{vectorFcns}. The key takeaway is that we may interchangeably refer to both $\psi$ and $\vec{p}$ as a single-edge strategy, and therefore we define $X(\vec{p}):=X(\psi), \sigma_{\vec{p}} :=\sigma_{\psi}$. \bigskip In this paper we are interested in the weak limit points of $\frac1n \mu_{0 \rightarrow n}(\chi)$. As it turns out, this set of limit points almost surely coincides with the set of limit points of $\frac1n \mu_{0 \rightarrow n}(\chi)$ over independent strategies $\chi$ only, with the distributions $\sigma_{\vec{p}}$ over single-edge strategies $\vec{p}$, as well as with the set $\mathcal{R}$ of probability measures with finite grid entropy. The following theorem is the main objective of Section \ref{indepReduction} . \begin{manualtheorem}{A}\label{part2_A} A.s. we have \begin{align*} \mathcal{R} &= \bigg\{\mbox{limit pts of} \ \frac1n \mu_{0\rightarrow n}(\chi): \mbox{strategies} \ \chi \bigg\} \\ &=\bigg\{\mbox{limit pts of} \ \frac1n \mu_{0\rightarrow n}(\chi): \mbox{independent strategies} \ \chi \bigg\}\\ &= \bigg\{\sigma_{\vec{p}}: \mbox{single-edge strategies} \ \vec{p} \bigg\} \end{align*} \end{manualtheorem} This reduces the problem to working with single-edge strategies $\vec{p}$, which are simpler to handle than generic strategies. Moreover, every $\sigma_{\vec{p}}$ can be achieved by a "consistent" single-edge strategy $\vec{p}$ with the properties that $p_1(\vec{u})$ is invariant under 1-fixing permutations in $S_D$ and $p_k(\vec{u}) = p_1(u_{D-k+2}, \ldots, u_1,\ldots, u_{D-k+1}) \ \forall k$. Now, the sets from Theorem \ref{part2_A} are convex and weakly compact. In Section \ref{extreme} we fully characterize their extreme points. \begin{manualtheorem}{B}\label{B} Let $\vec{p}$ be a single-edge strategy. The following are equivalent: \begin{enumerate}[label=(\roman*)] \item $\sigma_{\vec{p}}$ is an extreme point \item Any consistent single-edge strategy achieving $\sigma_{\vec{p}}$ must be deterministic. \item $\sigma_{\vec{p}}$ has a density $f_{\vec{p}}$ which is not constant on sets of positive $\Lambda$-measure and $\sigma_{\vec{p}}$ is given by the following single-edge strategy $\vec{q}$: \[ q_k(u_1,\ldots, u_D) = \textbf{1}_{\{f_{\vec{p}}(u_k) \geq f_{\vec{p}}(u_i) \ \forall i\}} \ \mbox{for $\Lambda^{\times D}$-a.a.} \ (u_1,\ldots, u_D) \] In other words, $\sigma_{\vec{p}}$ is achieved by the deterministic greedy single-edge strategy "choose whichever label yields a higher value when evaluating the density $f_{\vec{p}}$." \item $\sigma_{\vec{p}}$ has a density $f_{\vec{p}}$ s.t. $\frac1D f_{\vec{p}}(Unif[0,1]) \sim \mbox{Beta}(1,D)$. \end{enumerate} \end{manualtheorem} As an immediate corollary, $\mathcal{R}$ is the closed convex hull of these measures. \begin{remark} It is important to stress that Theorem \ref{B} holds even in a more general setting where the i.i.d. labels $U_i^j$ follow some finite mean distribution on $\mathcal{R}$ that is not necessarily Unif[0,1]. The only caveat is that the value distribution of the densities $f_{\vec{p}}$ of extreme points might not have as explicit a form as $D \cdot Beta(1,D$); however, the value distribution of the densities $f_{\vec{p}}$ is still identical to the value distribution of $f_{MAX}$, whatever that may be. \end{remark} \begin{remark} In fact, Theorem \ref{B} holds in a discrete setting as well. If the i.i.d. labels $U^j$ are Unif$\{1,\ldots, K\}$ then the convex set of consistent single-edge strategies forms a permutohedron of order $K$. \end{remark} \begin{center} \captionsetup{type=figure} \includegraphics[scale=0.4]{permutahedron.png} \captionof{figure}{Permutohedron of order 4, \cite{holroyd}} \end{center} The following is the discrete analogue of Theorem \ref{B}. \begin{manualtheorem}{B'}\label{B'} Let $\vec{p}$ be a single-edge strategy. The following are equivalent: \begin{enumerate}[label=(\roman*)] \item $\sigma_{\vec{p}}$ is an extreme point \item Any consistent single-edge strategy achieving $\sigma_{\vec{p}}$ must be deterministic. \item $\sigma_{\vec{p}}$ is given by the single-edge strategy "choose whichever label is maximal with respect to the ordering $\alpha(1) < \cdots < \alpha(K)$" for some permutation $\alpha \in S_K$. \item $\sigma_{\vec{p}}$ has a probability mass function whose value distribution is \[\bigg\{\frac1{K^D}, \frac{2^D-1}{K^D}, \frac{3^D-2^D}{K^D}, \ldots, \frac{K^D-(K-1)^D}{K^D} \bigg\}\] \end{enumerate} \end{manualtheorem} The ordering mentioned in (iii) gives the natural bijection between the extreme consistent single-edge strategies and the extreme points of the permutohedron. For example, the extreme point $1342$ corresponds to the single-edge strategy of choosing the maximal $u^i$ according to the ordering $1<3<4<2$, hence the bijection maps \[ 1342 \mapsto \sigma_{1342} = \frac1{4^D} \delta_1 + \frac{2^D-1}{4^D} \delta_3 + \frac{3^D-2^D}{4^D} \delta_4 + \frac{4^D-3^D}{4^D} \delta_2 \] This bijection extends to a bijection between the entire permutohedron and $\mathcal{R}$ by taking convex combinations. We walk through an explicit example of this phenomenon in Section \ref{discrete}. \bigskip Generally speaking there is no known way of computing grid entropy, however, it amazingly can be computed to be 0 for these extreme points. This effectively means that a.s. the number of paths $0 \rightarrow n$ with empirical measures weakly converging to any one of these extreme points is $e^{o(n)}$. \begin{manualtheorem}{C} Let $\sigma_{\vec{p}}$ be an extreme point. Then \[ || \sigma_{\vec{p}} || = 0 \] \end{manualtheorem} Whether \emph{every} measure in $\mathcal{R}$ with grid entropy 0 is an extreme point remains an open question. Section \ref{grid} focuses on this result, as well as simplified formulas for the Gibbs Free Energy and grid entropy in this model. \begin{manualtheorem}{D} Suppose $\tau: [0,1] \rightarrow {\mathbb{R}} $ is a bounded measurable function. Then Gibbs Free Energy with respect to $\tau$ is given by \[ G(\tau) = E \bigg[\log \sum_{j=1}^D e^{\tau(U^j)} \bigg] \ \mbox{a.s.}\] where $U^j$ are i.i.d. Unif[0,1]. For all probability measures $\nu$, grid entropy is given by \[ -||\nu|| = \sup_{\tau} \bigg[ \langle \tau, \nu \rangle - G(\tau) \bigg] = \sup_{\tau} \bigg[ \langle \tau, \nu \rangle - E \bigg[\log \sum_{j=1}^D e^{\tau(U^j)} \bigg] \bigg] \] where the supremum is over bounded measurable functions $\tau:[0,1] \rightarrow {\mathbb{R}} $ and where $\langle \tau, \nu \rangle$ denotes the integral $ \int_0^1 \tau(u) d\nu$. \end{manualtheorem} \begin{remark} This theorem actually holds for measurable $\tau$ s.t. $E[e^{\beta \tau(U)}] < \infty$ for $U \sim Unif[0,1]$ and can easily be extended to all measurable $\tau$ by truncating $\tau$ in $\langle \tau, \nu \rangle$ at some $C>0$ and taking a supremum over $C$ of the variational formula. We leave the details to the reader and focus on the bounded $\tau$ case. \end{remark} But first we delve deeper into the setup and known relevant results. \section{Preliminaries} \subsection{More on Strategies} We list some miscellaneous observations about strategies. It is trivial to see that the sets of strategies $\chi$, of single-edge strategies $\psi$, of distributions $\sigma_{\chi}^i$ for strategies $\chi$ and $i \geq 0$, and of $\sigma_{\psi}$ for single-edge strategies $\psi$ are each closed under convex combinations. The extreme points of the sets of strategies/single-edge strategies are clearly the sets of deterministic strategies/single-edge strategies respectively. Moreover, if $\chi$ is an independent strategy corresponding to a sequence $(\psi_i)$ of single-edge strategies then $\sigma_{\chi}^i = \sigma_{\psi_i} \ \forall i \geq 0$. In particular, this implies \begin{equation} \label{EQ2} \{\sigma_{\chi}^i: \mbox{independent strategies} \ \chi\} = \{\sigma_{\psi}: \mbox{single-edge strategies} \ \psi\} \ \forall i \geq 0 \end{equation} In fact, observe that for any strategy $\chi$ and $i \geq 0$ if we define the strategy $\chi'$ to be the measure determined on product sets by \[ \chi'(A \times B) := \chi( (0 \ i) A \times (0 \ i) B) \ \forall \ \mbox{measurable} \ A \subseteq ([0,1]^D)^{ {\mathbb{Z}} _{\geq 0}}, B \subseteq (\{1,\ldots,D\})^{ {\mathbb{Z}} _{\geq 0}}\] where $(0 \ i) C$ swaps the 0th and $i$th coordinates of sequences in $C$, then this is easily checked to be a strategy with \[ \sigma_{\chi}^i = \sigma_{\chi'}^0 \] Therefore \begin{equation} \label{EQswap} \{\sigma_{\chi}^i: \mbox{strategies} \ \chi\} = \{\sigma_{\chi}^0: \mbox{strategies} \ \chi\} \ \forall i \geq 0 \end{equation} \subsection{Coupling the Edge Weights with I.I.D. Uniforms}\label{coupling} We briefly explain why we can work in the setting of edge labels $U_i^j \sim Unif[0,1]$ without losing generality. Suppose instead we start with i.i.d. edge weights $(\tau_i^j)_{i \geq 0, 1 \leq j \leq D} \sim \theta$ for some distribution $\theta$ on $ {\mathbb{R}} $ with finite mean. We can then couple the environment to uniform random variables as is done in \cite[Sect.~2.1]{bates}. More specifically, we let our the $\tau_i^j$ be given by \[ \tau_i^j = \tau(U_i^j) \] for a measurable function $\tau: [0,1] \rightarrow {\mathbb{R}} $ and i.i.d. $U_i^j \sim Unif[0,1]$. If we let $F_{\theta}$ be the cumulative distribution function of $\theta$ then the quantile function \[ F_{\theta}^{-}(t) := \inf \{t \in {\mathbb{R}} : F_{\theta}(t) \geq x\} \] is an example of such a $\tau$; however, our results hold for any such $\tau$ chosen so we let $\tau$ be arbitrary. We can move from the simplified $(\Lambda, [0,1])$ setting we wish to work in to the initial $(\theta, {\mathbb{R}} )$ setting by applying the $\tau$-pushforward. Now, as mentioned in Theorem \ref{part2_A}, we show that, in the $(\Lambda, [0,1])$ setting, the set of limit points of the empirical measures $\frac1n \mu_{0\rightarrow n}(\chi)$ a.s. coincides with $\mathcal{R}$, which is shown in \ref{gridEntropyPart1} to a.s. coincide with the set of limit points of (strategy-free) empirical measures $\frac1n \mu_{\pi_n}$ along paths $\pi_n:0 \rightarrow n$. We do not lose generality by performing this coupling because a.s., $\frac1{n_k} \mu_{\pi_{n_k}} \Rightarrow \nu$ implies $\tau_{\ast}(\frac1{n_k} \mu_{\pi_{n_k}}) \Rightarrow \tau_{\ast}(\nu)$. This is proved in \cite[Lemma 6.15]{bates}. In fact, if we take $\tau $ to be the quantile function and we assume $\theta$ has continuous cdf $F_{\theta}$ then this implication becomes an if and only if, as seen in \cite[Lemma 5]{gatea}. \subsection{Grid Entropy} \begin{definition}The Levy-Prokhorov metric on the space $\mathcal{M}_+$ of finite non-negative Borel measures on $[0,1]$ is defined by \[\rho(\mu, \nu) = \inf \{\epsilon > 0: \mu(A) \leq \nu(A^{\epsilon}) + \epsilon \ \mbox{and} \ \nu(A) \leq \mu(A^{\epsilon}) + \epsilon \ \forall A \in \mathcal{B}([0,1])\}\] \end{definition} It is standard that the Levy-Prokhorov metric $\rho$ metrizes weak convergence. Some elementary properties include that the Levy-Prokhorov metric is upper bounded by total variation and it satisfies a certain subadditivity: \[ \rho(\mu_1 + \mu_2, \nu_1 + \nu_2) \leq \rho(\mu_1, \nu_1) + \rho(\mu_2, \nu_2) \] For $t \geq 0$ let $\mathcal{M}_t$ denote the space of non-negative Borel measures on $[0,1]$ with total mass $t$. In a concurrent paper \cite{gatea}, this author studies not just the convergence of empirical measures in a lattice model such as ours but the \emph{entropy} of empirical measures converging to a certain weak limit. Three quite different but equivalent definitions of grid entropy are given in said paper, and they are linked to the original description of this entropy provided in a 2014 paper \cite{rassoul2014quenched} by Rassoul-Agha and Sepp{\"a}l{\"a}inen. Here we work with the two most relevant to our needs, coming from \cite{gatea}. Note that in the setting of \cite{gatea}, both direction-fixed and direction-free grid entropy are considered, but in the model we focus on in this paper there is but one unit direction so we will use the direction-free versions of the results from \cite{gatea}. We go into a brief description of this work, as fits our needs in this paper. Fix any $t \geq 0$ and a target measure $\nu$. We consider the order statistics of the Levy-Prokhorov distance between $\nu$ and the empirical measures $\frac1n \mu_{\pi}$ varying over all $D^{\lfloor nt\rfloor}$ possible origin-anchored, length $\lfloor nt \rfloor$ paths $\pi$. That is, for every $n \in {\mathbb{N} } $ we let \[ \min_{\pi: 0\rightarrow \lfloor nt \rfloor}^1 \rho\bigg(\frac1n \mu_{\pi}, \nu \bigg) \leq \min_{\pi: 0\rightarrow \lfloor nt \rfloor}^2 \rho\bigg(\frac1n \mu_{\pi}, \nu \bigg) \leq \ldots \leq \min_{\pi: 0\rightarrow \lfloor nt \rfloor}^{D^{\lfloor nt \rfloor}} \rho\bigg(\frac1n \mu_{\pi}, \nu \bigg) \] denote the order statistics value of $ \rho(\frac1n \mu_{\pi}, \nu)$. It is convenient to define \[ \min_{\pi: 0\rightarrow \lfloor nt \rfloor}^j \rho\bigg(\frac1n \mu_{\pi}, \nu \bigg) := +\infty \ \mbox{for} \ j > D^{\lfloor nt \rfloor}\] These order statistics and the paths corresponding to them are of course event-dependent. However, the following theorem from \cite{gatea} shows there is a deterministic critical exponent where the empirical measures along the paths corresponding to these order statistics change from converging a.s. to $\nu$ to a.s. diverging away from $\nu$. This critical exponent turns out to be precisely the negative convex conjugate of Gibbs Free Energy. We state the full theorem as presented in \cite{gatea}, noting that part (i) has been previously shown in \cite{carmona2010directed} to hold in the case of Bernoulli edge labels, and that parts (ii) and (iv) follow immediately from the earlier work of Rassoul-Agha and Sepp{\"a}l{\"a}inen \cite{rassoul2014quenched}. \begin{theorem}\label{gridEntropyPart1} \begin{enumerate}[label=(\roman*)] \item[] \item For any target measure $\nu \in \mathcal{M}$, its grid entropy is defined to be the deterministic quantity \begin{align*} || \nu|| &:= \sup \bigg\{\alpha \geq 0: \lim_{n \rightarrow \infty} \min_{\pi: 0 \rightarrow \lfloor nt \rfloor}^{\lfloor e^{\alpha n} \rfloor} \rho\bigg(\frac1{n} \mu_{\pi}, \nu \bigg) = 0 \ \mbox{a.s.} \bigg\} \\ &\in \{-\infty\} \cup [0, t\log D] \end{align*} where the value is $-\infty$ if the set of $\alpha$'s is empty. Then grid entropy is the critical exponent where the $\min \limits_{\pi: 0\rightarrow \lfloor nt \rfloor}^{\lfloor e^{\alpha n} \rfloor} \rho(\frac1n \mu_{\pi}, \nu )$ change from converging to 0 to a.s. having a $\liminf \limits_{n \rightarrow \infty} >0$.\\ For example, the grid entropy of the original distribution, $\Lambda$, is maximal:\\ $||\Lambda|| = t \log D$. \item Grid entropy is positive-homogeneous, satisfies the reverse-triangle inequality \[ ||\nu|| + ||\xi|| \leq || \nu + \xi||\] and is concave and upper-semicontinuous in $\nu$. \item Consider the deterministic, weakly closed set \[ \mathcal{R}^t:= \{\nu \in \mathcal{M}: ||\nu|| \geq 0 \}\] Then $\mathcal{R}^t \subseteq \mathcal{M}_t$, $\nu \ll \Lambda \ \forall \nu \in \mathcal{R}^t$ and \[ \mathcal{R}^t = \bigg\{\mbox{limit pts of} \ \frac1n \mu_{\pi} \ \mbox{for} \ \pi: 0\rightarrow \lfloor nt \rfloor \bigg\} \ \mbox{a.s.} \] \item Grid entropy is the negative convex conjugate of $\beta$-Gibbs Free Energy. More concretely, for $\beta > 0$, \begin{equation}\label{dualFormula} -||\nu|| = \sup_{\tau} \bigg[ \beta \langle \tau, \nu \rangle - G_t^{\beta}(\tau) \bigg] \ \forall \nu \in \mathcal{M}_t \end{equation} where the supremum is taken over bounded measurable $\tau: [0,1] \rightarrow {\mathbb{R}} $, where \\ $\langle \tau, \nu \rangle = \int_0^1 \tau(u) d\nu$, where $G_t^{\beta}(\tau)$ is the length $t$ $\beta$-Gibbs Free Energy \[ G_t^{\beta}(\tau) := \lim_{n \rightarrow \infty} \frac1n \log \sum_{\pi: 0 \rightarrow \lfloor nt \rfloor} e^{\beta T(\pi)} \] and where $T(\pi) = \sum \limits_{e\in \pi} \tau(U_e) = \langle \tau, \mu_{\pi} \rangle$ is the passage time along $\pi$. \item Any $\nu \in \mathcal{R}^t$ satisfies the following upper bound on the sum of the grid entropy and the relative entropy with respect to Lebesgue measure on [0,1]: $$t D_{KL} (\nu||\Lambda) + ||\nu|| \leq t \log D$$ where $D_{KL}$ denotes relative entropy (or Kullback-Leibler divergence). \end{enumerate} \end{theorem} \begin{remark} An immediate but highly non-trivial consequence of (i) is the existence of random length $\lfloor nt \rfloor$ paths whose empirical measures converge weakly to a given target $\nu \in \mathcal{R}^t$ a.s.. Indeed, the event-dependent paths corresponding to $\min \limits_{\pi: 0 \rightarrow \lfloor nt \rfloor}^1 \rho(\frac1n \mu_{\pi}, \nu)$ do the job. \end{remark} \begin{remark} (i), seen as the statement that the negative convex conjugate of Gibbs Free Energy is the critical exponent of the order statistics $\min_{\pi: 0\rightarrow \lfloor nt \rfloor}^j \rho(\frac1n \mu_{\pi}, \nu)$, has been previously proved in the case of Bernoulli($p$) edge labels in \cite{carmona2010directed} [Corollary 2]. \end{remark} By positive-homogeneity it suffices to work with the case $t=1$. For the sake of notation we drop the 1 in $\mathcal{R}^1, G_1^{\beta}(\tau)$ for the rest of the paper. Once we determine the extreme points of $\mathcal{R}$, we compute their grid entropies to be 0. We also give a limit-free formula for the $\beta$-Gibbs Free Energy $G^{\beta}(\tau)$ in our model, which renders the convex duality formula \eqref{dualFormula} more practical. \subsection{Measure-Preserving Bijections in \texorpdfstring{$ {\mathbb{R}} ^n$}{Rn}} We will encounter measure-preserving bijections in $ {\mathbb{R}} $, so we briefly outline the required notions in this section. \begin{definition} Let $\mathcal{B}( {\mathbb{R}} ^n)$ denote the Borel $\sigma$-algebra of $ {\mathbb{R}} ^n$ and let $\mu$ be a Borel probability measure on $ {\mathbb{R}} ^n$. Let $A, A' \in \mathcal{B}( {\mathbb{R}} ^n)$ with $\mu(A') = \mu(A) > 0$. A bijection $\phi: A \rightarrow A'$ is said to be $\mu$-measure-preserving if $\mu(B) = \mu(\phi(B)) $ for all $B \in \mathcal{B}( {\mathbb{R}} ^n), B \subseteq A$. \end{definition} Nishiura proves in \cite[Theorem 4]{nishiura} the existence of a $\mu$-measure-preserving bijection between Borel sets of equal, positive but not full $\mu$-measure, given that $\mu$ is nonatomic. We will only need the following simplified version of this theorem. \begin{theorem}\label{thmNishiura} Suppose $\theta$ is a Borel probability measure on $[0,1]$ that is absolutely continuous with respect to Lebesgue measure $\Lambda$ on $[0,1]$. Let $A, A' \in \mathcal{B}([0,1])$ s.t. $\theta(A) = \theta(A') \in (0,1)$. Then there exists a $\theta$-measure-preserving bijection $\phi: A \rightarrow A'$. \end{theorem} Preliminaries finished, our next goal is to show that we may restrict ourselves to working with single-edge strategies without losing generality. \section{Reducing the Problem}\label{indepReduction} Over the course of this section, we prove the following theorem, which allows us to reduce the problem of describing the set of limit points of the empirical measures $\frac1n \mu_{0\rightarrow n}(\chi)$ to characterizing the set of achievable distributions $\sigma_{\psi}$ for single-edge strategies $\psi$. \begin{theorem}\label{limitPts} A.s., the following sets are equal: \begin{enumerate}[label=(\roman*)] \item $\{\mbox{limit pts of} \ \frac1n \mu_{0\rightarrow n}(\chi): \mbox{ strategies} \ \chi\}$ \item $\mathcal{R} := \{\nu \in \mathcal{M}_1: ||\nu|| \geq 0\}$ \item $\{\sigma_{\chi}^0: \mbox{ strategies} \ \chi\}$ \item $\{\sigma_{\psi}: \mbox{single-edge strategies} \ \psi\}$ \item $\{\mbox{limit pts of} \ \frac1n \mu_{0\rightarrow n}(\chi): \mbox{independent strategies} \ \chi\}$ \item $\{\mbox{limit pts of} \ \frac1n \mu_{0\rightarrow n}(\chi): \mbox{i.i.d. strategies} \ \chi\}$ \end{enumerate} \end{theorem} \begin{remark} A priori it not clear that (iii)-(iv) are weakly closed. We show this as part of our proof. \end{remark} We argue via a chain of inclusions. We begin with the most trivial of these. \begin{lemma}\label{1st} A.s. we have \[ \bigg\{\mbox{limit pts of} \ \frac1n \mu_{0\rightarrow n}(\chi): \mbox{ strategies} \ \chi \bigg\} \subseteq \mathcal{R} \] \end{lemma} \begin{proof} Fix a strategy $\chi$. Suppose $\frac1{n_k} \mu_{0 \rightarrow n_k}(\chi) \Rightarrow \nu$ for some observed edge labels \\ $(u_i^j)_{i \geq 0, 1 \leq j \leq D}$. Then letting $\pi_{n}: 0 \rightarrow n$ be the paths corresponding to the indices \\ $(J_0(\chi), \ldots, J_{n-1}(\chi))$ conditioned on these same observed edge labels, we get\\ $ \frac1{n_k} \mu_{\pi_{n_k}}=\frac1{n_k} \mu_{0 \rightarrow n_k}(\chi) \Rightarrow \nu$. Thus $\nu$ is a limit point of the $\frac1n \mu_{\pi}$. The desired inclusion follows from \[ \mathcal{R} = \bigg\{\mbox{limit pts of} \ \frac1n \mu_{\pi}: \mbox{ paths} \ \pi:0 \rightarrow n \bigg\} \ \mbox{a.s.} \ \mbox{by Theorem \ref{gridEntropyPart1}(iii)}\] \end{proof} \subsection{Expected Value of Empirical Measures}\label{sect31} For any strategy $\chi$, \[E[\delta_{X_i(\chi)}] = \int_{[0,1]} \delta_x d\sigma_{\chi}^i(x) = \sigma_{\chi}^i \ \mbox{and} \ E \bigg[\frac1n \mu_{0\rightarrow n}(\chi)\bigg] = \frac1n \sum_{i=0}^{n-1} \sigma^i_{\chi} \] Recalling \eqref{EQswap} and the fact that $\{\sigma^0_{\chi}: \mbox{strategies} \ \chi\}$ is closed under convex combinations, it follows that every $E [\frac1n \mu_{0\rightarrow n}(\chi)]$ is contained in $\{\sigma^0_{\chi}: \mbox{strategies} \ \chi\}$. Therefore \begin{align*} \bigg\{E \bigg[\frac1n \mu_{0\rightarrow n}(\chi)\bigg]: \mbox{ strategies} \ \chi \bigg\} &= \{E[\delta_{X_0(\chi)}]: \mbox{ strategies} \ \chi\} \\ &= \{\sigma_{\chi}^0: \mbox{ strategies} \ \chi \} \end{align*} Also it is clear that \[ \{E[\delta_{X(\psi)}]: \mbox{single-edge strategies} \ \psi\} = \{\sigma_{\psi}: \mbox{ single-edge strategies} \ \psi \}\] A simple Tonelli argument establishes that the closures of the two sets above coincide. \begin{lemma} \label{equivalenceIndep} \[ \mbox{cl}\{\sigma_{\psi}: \mbox{ single-edge strategies} \ \psi \} = \mbox{cl}\{\sigma^0_{\chi}: \mbox{ strategies} \ \chi \} \] \end{lemma} \begin{proof} By \eqref{EQ2} $\{\sigma_{\psi}\} = \{\sigma^0_{\chi}: \mbox{indep strategies} \ \chi\}$ so it suffices to show \[ \mbox{cl}\{\sigma_{\psi}: \mbox{ single-edge strategies} \ \psi \} \supseteq \{\sigma^0_{\chi}: \mbox{ strategies} \ \chi \} \] Given a strategy $\chi$ and observed "future" edge labels $(u_i^j)_{i \geq 1, 1 \leq j \leq D} \in [0,1]^{ {\mathbb{Z}} _{\geq 1}}$, define $\psi_{(u_i^j)_{i \geq 1, 1 \leq j \leq D}}$ to be $\chi$ conditioned on the rest of the environment $(u_i^j)_{i \geq 1, 1 \leq j \leq D}$. The $\psi_{(u_i^j)_{i \geq 1, 1 \leq j \leq D}}$ are easily seen to be single-edge strategies. Since integrating over the entire environment $(U_i^j)_{i \geq 0, 1 \leq j \leq D}$ is equivalent to integrating over $(U_0^j)_{1 \leq j \leq D}$ first and then over $(U_i^j)_{i \geq 1, 1 \leq j \leq D}$, we have \[ \sigma_{\chi}^0 = \int_{([0,1]^D)^{ {\mathbb{Z}} _{\geq 1}}} \sigma_{\psi_{(u_i^j)_{i \geq 1, 1 \leq j \leq D}}} d\Lambda^{\infty}((u_i^j)_{i \geq 1, 1 \leq j \leq D}) \] By convexity and weak closure it follows that \[\sigma_{\chi}^0 \in \mbox{cl}\{\sigma_{\psi}: \mbox{ single-edge strategies} \ \psi \} \] \end{proof} Next, we show that this closure of $\{\sigma_{\chi}^0: \mbox{strategies} \ \chi\}$ contains all probability measures with finite grid entropy. \begin{lemma} \label{3rd} We have \[\mathcal{R}:= \{\nu \in \mathcal{M}_1: ||\nu|| \geq 0\} \subseteq \mbox{cl}\{E[\delta_{X_0(\chi)}]: \mbox{ strategies} \ \chi \} \] as deterministic sets. \end{lemma} \begin{proof} Suppose this is not the case, say $\exists \nu \in \mathcal{R} \cap (\mbox{cl}\{E[\delta_{X_0(\chi)}]: \mbox{ strategies} \ \chi \})^C$. Thus there exists $\epsilon > 0$ s.t. $B_{\epsilon}(\nu) \cap cl\{E[\delta_{X_0(\chi)}]: \chi\} = \emptyset$. Let $\pi_n: 0 \rightarrow n$ be the environment-dependent path corresponding to $\min \limits_{\pi: 0 \rightarrow n}^1 \rho(\frac1n \mu_{\pi}, \nu)$. It is crucial to note that $\pi_n$ depends on the observed edge labels $((U_i^j)_{0 \leq i < n, 1\leq j \leq D})$ of the first $n$ trials only. By definition of grid entropy, $\frac1n \mu_{\pi_n} \Rightarrow \nu$ a.s.. In particular, we get convergence in probability: \[ \lim_{n \rightarrow \infty} P \bigg( \rho \bigg(\frac1n \mu_{\pi_n} , \nu \bigg) \geq \frac{\epsilon}2 \bigg) = 0\] Thus $\exists n \in {\mathbb{N} } $ s.t. \[ P \bigg( \rho \bigg(\frac1n \mu_{\pi_n} , \nu \bigg) \geq \frac{\epsilon}2 \bigg) < \frac{\epsilon}4 \] We claim that $\rho(E[\frac1n \mu_{\pi_n}], \nu)<\epsilon$. Let $\mathcal{E}$ be the event $\{\rho (\frac1n \mu_{\pi_n} , \nu ) \geq \frac{\epsilon}2\}$ so $P(\mathcal{E})< \frac{\epsilon}4$. We split the expectation: \[E \bigg[ \frac1n \mu_{\pi_n} \bigg] = \int_{\mathcal{E}} \frac1n \mu_{\pi_n} dP + \int_{\mathcal{E}^C} \frac1n \mu_{\pi_n} dP \] Since the Levy-Prokhorov metric is upper bounded by the total variation, then \[\rho \bigg(\int_{\mathcal{E}} \frac1n \mu_{\pi_n} dP, \nu \cdot P(\mathcal{E}) \bigg) \leq \bigg( \bigg| \bigg |\frac1n\mu_{\pi_n} \bigg|\bigg|_{TV} + ||\nu||_{TV} \bigg)P(\mathcal{E}) = 2P(\mathcal{E}) < \frac{\epsilon}2 \] On $\mathcal{E}^C$, we have $\rho (\frac1n \mu_{\pi_n} , \nu ) < \frac{\epsilon}2$. By definition of $\rho$, for all measurable $A \in \mathcal{B}([0,1])$, \begin{align*} \bigg(\int_{\mathcal{E}^C} \frac1n \mu_{\pi_n} dP\bigg) (A) &= \int_{\mathcal{E}^C} \frac1n \mu_{\pi_n} (A) dP \\ & \leq \int_{\mathcal{E}^C} \nu(A^{\frac{\epsilon}2} )+ \frac{\epsilon}2 dP \leq \nu(A^{\frac{\epsilon}2} ) P(\mathcal{E}^C) +\frac{\epsilon}2 \end{align*} and similarly \begin{align*} \nu(A) P(\mathcal{E}^C) &= \int_{\mathcal{E}^C} \nu(A) dP \\ &\leq \int_{G^C} \frac1n \mu_{\pi_n} (A^{\frac{\epsilon}2}) + \frac{\epsilon}2 dP \leq \bigg( \int_{\mathcal{E}^C} \frac1n \mu_{\pi_n} dP\bigg) (A^{\frac{\epsilon}2}) + \frac{\epsilon}2 \end{align*} hence \[ \rho \bigg(\int_{\mathcal{E}^C} \frac1n \mu_{\pi_n} dP, \nu P(\mathcal{E}^C) \bigg) \leq \frac{\epsilon}2 \] By subadditivity of $\rho$, \begin{align*} \rho \bigg( E \bigg[\frac1n \mu_{\pi_n} \bigg], \nu \bigg) &= \rho \bigg( \int_{\mathcal{E}} \frac1n \mu_{\pi_n} dP + \int_{\mathcal{E}^C} \frac1n \mu_{\pi_n} dP, \nu P(\mathcal{E}) + \nu P(\mathcal{E}^C) \bigg) \\ &\leq \rho \bigg( \int_{\mathcal{E}} \frac1n \mu_{\pi_n} dP, \nu P(\mathcal{E}) \bigg) + \rho\bigg(\int_{\mathcal{E}^C} \frac1n \mu_{\pi_n} dP, \nu P(\mathcal{E}^C) \bigg) \\ & < \epsilon \end{align*} It remains to construct a strategy $\chi$ for which $\frac1n \mu_{\pi_n} = \frac1n \mu_{0 \rightarrow n} (\chi)$ for this fixed $n$. But this is trivial since conditioned on the observed edge labels $(U_i^j)_{0 \leq i \leq n-1, 1 \leq j \leq D}$ from the first $n$ trials, $\pi_n$ is a deterministic path $0 \rightarrow n$ (namely the one which minimizes $\rho(\frac1n \mu_{\pi}, \nu)$). To be concrete, consider the product measure $\chi'$ on $([0,1]^D)^{ {\mathbb{Z}} _{\geq n}} \times \{1,\ldots,D\}^{ {\mathbb{Z}} _{\geq n}}$ given by \[ \chi' = \Lambda^{\infty} \times \delta_{(1,1,\ldots)} \] Heuristically, $\chi'$ is a partial strategy always picking the "top" edge label $U_i^1$ for $i \geq n$. Also let $\chi''$ be the measure on $([0,1]^D)^{ {\mathbb{Z}} _{0 \leq i < n}} \times \{1,\ldots,D\}^{ {\mathbb{Z}} _{0 \leq i < n}}$ determined by \begin{align*} &\chi''(A \times \{(j_0,\ldots,j_{n-1}) \}) \\ &= \Lambda^{\infty}(\{(u_i^j)_{0 \leq i < n, 1 \leq j \leq D} \in A: \pi_n((u_i^j)_{0 \leq i < n, 1 \leq j \leq D}) = (j_0,\ldots, j_{n-1})\}) \end{align*} $\forall A \in \mathcal{B}(([0,1]^D)^{ {\mathbb{Z}} _{0 \leq i < n}}), (j_i)_{0 \leq i < n} \in \{1,\ldots,D\}^{ {\mathbb{Z}} _{0 \leq i < n}}$ where $\pi_n((u_i^j)_{0 \leq i < n, 1 \leq j \leq D})$ denotes the sequence of $n$ indices $(j_0,\ldots, j_{n-1})$ corresponding to the path $\pi_n$ when the first $n$ trials yield observed labels $(u_i^j)_{0 \leq i < n, 1 \leq j \leq D}$. That is, conditioned on the observed values from the first $n$ trials, $\chi''$ always picks the path corresponding to $\pi_n$. \bigskip Now consider the strategy $\chi = \chi'' \times \chi'$. $\mu_{0 \rightarrow n}(\chi) = \mu_{\pi_n}$ hence $E[\frac1n \mu_{\pi_n}] = E[\frac1n \mu_{0\rightarrow n}(\chi)]$. This contradicts \[B_{\epsilon}(\nu) \cap cl \bigg\{E \bigg[\frac1n \mu_{0 \rightarrow n}(\chi) \bigg]: \chi \bigg\} = B_{\epsilon}(\nu) \cap cl\{E[\delta_{X_0(\chi)}]: \chi\} = \emptyset\] \end{proof} Combining Lemmas \ref{1st}-\ref{3rd}, we get that a.s., (vi) $\subseteq$ (v) $\subseteq$ (i) $\subseteq$ (ii) $\subseteq $ cl(iii) = cl(iv) in Theorem \ref{limitPts}. To complete the proof we only need to show the last inclusion cl (iv) $\subseteq$ (vi) and the fact that (iv) is weakly closed. \subsection{What Happens with Independent Strategies} Let us now focus on the case of independent strategies. Recall that these look like $\chi = \bigtimes \limits_{i=0}^{\infty} \psi_i$ where $(\psi_i)_{i \geq 0}$ are single-edge strategies. First, consider the even simpler case of i.i.d. strategies $\chi$, where all the $\psi_i$ are identical, resulting in i.i.d. $X_i$. Then the Glivenko-Cantelli Theorem \cite[Thm.~2.4.7]{durrett} gives that a.s. the empirical measures $\frac1n \mu_{0 \rightarrow n}(\chi)$ converge weakly to the common law of the $X_i$'s, $\sigma_{\psi_0}$. \begin{theorem}[Glivenko-Cantelli Theorem]\label{thm1} Let $F_{\gamma}$ be the cumulative distribution function of $\gamma$, let $Y_i \sim \gamma$ be i.i.d. random variables and let \[F_{n}(y) = \frac1{n} \sum_{i=1}^n \mathbf{1}_{\{Y_i \leq y\}} \] be the cumulative distribution functions of the empirical measures. Then \[\sup_y |F_n(y) - F_{\gamma}(y)| \rightarrow 0 \ \mbox{a.s. as} \ n \rightarrow \infty \] \end{theorem} As an immediate corollary we get the outstanding inclusion (vi) $\supseteq $ cl (iv) mentioned at the end of Section \ref{sect31}. \begin{corollary} \end{corollary} A.s. we have \[ \bigg\{\mbox{limit pts of} \ \frac1n \mu_{0 \rightarrow n}(\chi): \ \mbox{i.i.d. strategies $\chi$}\bigg\} \supseteq cl\{\sigma_{\psi}: \mbox{single-edge strategies} \ \psi\} \] \begin{proof} Fix a dense subset $\mathcal{O}$ of the deterministic, weakly compact set cl (iv) s.t. $\mathcal{O} \subseteq \mbox{(iv).}$ For every $\sigma_{\psi} \in \mathcal{O}$, apply Theorem \ref{thm1} to the i.i.d. strategy $\chi = \bigtimes \limits_{i=0}^{\infty} \psi$ to get $\frac1n \mu_{0\rightarrow n}(\chi) \Rightarrow \sigma_{\psi}$ a.s.. Therefore $\mathcal{O} \subseteq$ (vi) a.s.. The inclusion follows since (vi) is weakly closed.\\ \end{proof} Thus we have shown that the sets (i), (ii), cl(iii), cl(iv), (v), (vi) in Theorem \ref{limitPts} are equal a.s.. Before proceeding to look further into single-edge strategies, it is worth mentioning a version of the Glivenko-Cantelli Theorem for independent but not necessarily i.d. sequences from \cite{wellner}. It gives further insight into the limit points of the empirical measures $\frac1n \mu_{0\rightarrow n}(\chi)$ for independent strategies $\chi$. \begin{theorem}\label{generalGlivenko} Let $Y_i$ be a sequence of independent random variables with distributions $\sigma_i$. Define $\overline{\sigma}_n = \frac1n (\sigma_0+\ldots + \sigma_{n-1})$ to be the averages of these distributions and let \\ $\frac1n \mu_n = \frac1n (\delta_{Y_0}+\ldots+\delta_{Y_{n-1}})$ be the empirical measures. If $\{\overline{\sigma}_n\}$ is tight then $\rho(\overline{\sigma}_n, \frac1n \mu_n) \rightarrow 0$ a.s.. \end{theorem} In our case, $\mathcal{M}_1$ is weakly compact so any sequence of probability measures in $\mathcal{M}_1$ is tight. Thus, for independent strategies $\chi = \bigtimes \limits_{i=0}^{\infty} \psi_i$ with $\psi_i$ single-edge strategies, we have \\ $\rho(\overline{\sigma_n}, \frac1n \mu_{0 \rightarrow n}(\chi)) \rightarrow 0 \ \mbox{a.s.}$ where $\overline{\sigma_n}$ denote the averages of the distributions of $X_i(\chi)$: \[\overline{\sigma}_n = \frac1n (\sigma^0_{\chi} + \ldots +\sigma^{n-1}_{\chi}) = \frac1n (\sigma_{\psi_0} + \ldots +\sigma_{\psi_{n-1}})= \sigma_{\frac1n(\psi_0+\ldots + \psi_{n-1})} \] This not only confirms that cl(iv) and (v) are equal, but it tells us that a.s. the empirical measures $\frac1n \mu_{0 \rightarrow n}(\chi)$ for independent strategies $\chi$ have the exact same limit points as the distributions $\sigma_{\frac1n(\psi_0+\ldots + \psi_{n-1})}$ corresponding to the law of $X$ chosen according to the average of the single-edge strategies $\psi_i$. \section{Single-Edge Strategies Revisited} \subsection{Single-Edge Strategies in Terms of Conditional Probabilities}\label{vectorFcns} From what we have shown thus far, we only need to focus on single-edge strategies, as the set of limit points of empirical measures $\frac1n \mu_{0 \rightarrow n}(\chi)$ coincides with \[ cl \{\sigma_{\psi}: \mbox{single-edge strategies} \ \psi\} \] Rather than using the measure definition of single-edge strategies, it is more practical to work with the vector $\vec{p}$ of conditional probabilities defined as \[ p_k(u_1,\ldots, u_D) := P_{\psi}[J = k \mid (U^j)_{1 \leq j \leq D} = (u^j)_{1 \leq j \leq D}] \] Then $\psi$ evaluated on product sets is given by \begin{equation}\label{PsiFormula} \psi(A \times B) = \int_A \sum_{j=1}^D p_j (u_1,\ldots, u_D) \mathbf{1}_{\{j \in B\}} \ d\Lambda^D(u_1,\ldots, u_D) \ \forall A \in \mathcal{B}([0,1]^D), B \subseteq \{1,\ldots, D\} \end{equation} and $\sum \limits_{j=1}^D p_j \equiv 1$ everywhere. Intuitively, each $p_k(u_1,\ldots, u_D)$ is the probability of choosing $u_k$ when the observed samples are $(u_1,\ldots,u_D)$. This justifies the term "single-edge strategy," because $\vec{p}$ is prescribing the strategy by which we make our choice once we have the $D$ observed samples. Of course, the vector $\vec{p}=(p_1,\ldots, p_D)$ determines $\psi$ by \eqref{PsiFormula}. And if all the $p_j$ are 0-1 valued $\Lambda^{\times D}$-a.e. then the single-edge strategy $\psi$ is deterministic. Even though different $\vec{p}$ may give rise to the same $\psi$, we conflate the two notions and call both $\psi$ and $\vec{p}$ the single-edge strategy. We use $X(\vec{p}), X(\chi)$ and $\sigma_{\vec{p}}, \sigma_{\psi}$ and $F_{\vec{p}}, F_{\psi}$ interchangeably. At the end of the day, all that matters is whether two vectors $\vec{p}, \vec{q}$ yield the same law $\sigma_{\vec{p}}$ of $X(\vec{p})$. We may now write the cumulative distribution function of $\sigma_{\vec{p}}$ in terms of $\vec{p}$: \begin{equation} \label{cdf} \begin{split} F_{\vec{p}}(y) &=\int_{[0,1]^D} \sum_{j=1}^D p_j (u_1,\ldots, u_D) \mathbf{1}_{[0,y]}(u_j) d\Lambda^D(u_1,\ldots, u_D) \\ &= \int_{[0,y] \times [0,1]^{D-1}} p_1 (u_1,\ldots, u_D) d\Lambda^D(u_1,\ldots, u_D) \\ &+ \ldots + \int_{[0,1]^{D-1} \times [0,y] } p_D (u_1,\ldots, u_D) d\Lambda^D(u_1,\ldots, u_D) \end{split} \end{equation} Therefore $\sigma_{\vec{p}} \ll \Lambda$ with density \begin{equation}\label{density} \begin{split} f_{\vec{p}}(y) &=\int_{[0,1]^{D-1}} p_1(y,u_2,\ldots, u_D) d\Lambda^{D-1}(u_2,\ldots, u_D) \\ &+ \cdots + \int_{[0,1]^{D-1}} p_D(u_1,\ldots,u_{D-1},y) d\Lambda^{D-1}(u_1,\ldots, u_{D-1}) \end{split} \end{equation} It is clear that $f_{\vec{p}} \in [0,D]$.\\ Let us give an example that will make everything clear. Consider the deterministic single-edge strategy $\vec{p}^{MAX}$ that always chooses the maximum of the observed edge labels. In our notation, \[ p_j^{MAX}(u_1, \ldots, u_D) = \mathbf{1}_{\{u_j \geq u_k \ \forall 1 \leq k \leq D\}}\] The resulting distribution $\sigma_{MAX}$ has cdf \[F_{MAX}(y) = P(U^j \leq y \ \forall 1 \leq j \leq D) = y^D \] and density \[f_{MAX}(y) = D y^{D-1}\] This single-edge strategy plays a crucial role later in our description of the extreme points of $\{\sigma_{\vec{p}}\}$. We will also be interested in single-edge strategies "scrambled" by a $\Lambda$-measure-preserving bijection. \begin{definition} Let $\vec{p}$ be a single-edge strategy and let $\phi: [0,1] \rightarrow [0,1]$ be a $\Lambda$-measure-preserving bijection. We define $\vec{p}^\phi$ to be the single-edge strategy with coordinate functions \[p_j^{\phi}(u_1, \ldots, u_D) := p_j (\phi(u_1), \ldots, \phi(u_D))\] It is clear that $\sum p_j^{\phi} \equiv 1$ still so it is a valid single-edge strategy. Furthermore, since $\phi$ is measure-preserving with respect to $\Lambda$, then by a change of variables in the integral formula for $f_{\vec{p}}$ we get \[ f_{\vec{p}^{\phi}} = f_{\vec{p}} \circ \phi \] \end{definition} The following lemma collects some basic facts about single-edge strategies. \begin{lemma}\label{lemma3} Let $\vec{p}$ be a single-edge strategy. \begin{enumerate}[label=(\roman*)] \item For any Borel set $A \in \mathcal{B}([0,1])$, \[\Lambda(A)^D \leq \sigma_{\vec{p}}(A) \leq 1 - (1-\Lambda(A))^D \] In particular, $X(\vec{p}^{MAX})$ stochastically dominates $X(\vec{p})$ and $\sigma_{\vec{p}} \ll \Lambda, \Lambda \ll \sigma_{\vec{p}}$. \item We have \[ \sup_{\phi} E[\sigma_{\vec{p}^{\phi}}] \leq E[\sigma_{MAX}]\] where the supremum is taken over $\Lambda$-measure-preserving bijections $\phi: [0,1] \rightarrow [0,1]$. \item Convex combinations of single-edge strategies $\vec{p}$ translate to convex combinations of $\sigma_{\vec{p}}, f_{\vec{p}}, F_{\vec{p}}$. \end{enumerate} \end{lemma} \begin{remark} It is not clear whether the supremum in (ii) is achieved, but this is beyond the scope of this paper. \end{remark} \begin{proof} (i) If all $D$ edge labels $u_1, \ldots, u_D$ are in the set $A$, then so must be the edge label chosen from among them. Therefore $\Lambda(A)^D \leq \sigma_{\vec{p}}(A)$. The other inequality follows by replacing $A$ with $A^C$. Recalling that $F_{MAX}(y) = y^D$, we get \[F_{MAX}(y) \leq F_{\vec{p}}(y) \ \forall y\] (ii) For any $\Lambda$-measure-preserving bijection $\phi$, we apply the tail integral formula for expectation and use (i) to get \[E[\sigma_{\vec{p}^{\phi}}] = \int_{0}^{1} 1-F_{\vec{p}^{\phi}}(t) dt \leq \int_{0}^{1} 1-F_{MAX}(t) dt = E[\sigma_{MAX}]\] (iii) For single-edge strategies $\vec{p}, \vec{q}$ and $t \in [0,1]$, $t\vec{p} + (1-t)\vec{q}$ is itself a single-edge strategy with \[ F_{t\vec{p} + (1-t)\vec{q}} = t F_{\vec{p}} + (1-t) F_{\vec{q}}, \ f_{t\vec{p} + (1-t)\vec{q}} = t f_{\vec{p}} + (1-t) f_{\vec{q}}, \ \mu_{t\vec{p} + (1-t)\vec{q}} = t \sigma_{\vec{p}} + (1-t) \sigma_{\vec{q}}\] \end{proof} \subsection{"Consistent" Single-Edge Strategies} Before proceeding further, we explain why we can restrict ourselves to single-edge strategies $\vec{p}$ with some convenient symmetries that render $\vec{p}$ consistent. We say a permutation $\iota \in S_D$ acts on a $D$-tuple $\vec{u}$ by applying $\iota$ to the indices: \[ \iota(\vec{u}) = (u_{\iota(1)}, \ldots, u_{\iota(D)})\] \begin{lemma}\label{lemma4} Let $\vec{p}$ be any single-edge strategy. Then there exists another single-edge strategy $\vec{q}$ that gives rise to the same distribution $\sigma_{\vec{p}} = \sigma_{\vec{q}}$ s.t. $\forall x \in [0,1], 1 \leq i \leq D$ \begin{equation}\label{eq17} f_{\vec{p}}(x) = f_{\vec{q}}(x) = D \int_{[0,1]^{D-1}} q_i(u_1, u_2, \ldots, x, \ldots, u_D) d\Lambda^{D-1}(u_1,\ldots, \widehat{u_i}, \ldots, u_D) \end{equation} where the $x$ occurs at position $i$, s.t. \begin{equation}\label{eq18} q_1(\iota_i(\vec{u})) = q_2(\iota_{i-1}(\vec{u})) = \ldots = q_D(\iota_{i+1}(\vec{u})) \ \forall 1 \leq i \leq D \ \forall \vec{u} \end{equation} where each $\iota_j$ is the cyclic shift \[\iota_j(\vec{u}) = (u_j, u_{j+1}, \ldots, u_D, u_1,\ldots, u_{j-1})\] and s.t. \begin{equation}\label{eq19} q_i(\vec{u}) = q_i(\iota(\vec{u})) \ \forall 1 \leq i \leq D, \forall \vec{u}, \mbox{and} \ \forall \iota \in S_D \ \mbox{with} \ \iota(i) = i \end{equation} \end{lemma} \begin{remark}\label{RMK11} The new single-edge strategy $\vec{q}$ is consistent across all permutations of a tuple $\vec{u}$. That is, given an unordered tuple $(u_1, \ldots, u_D)$ we can say that $\vec{q}$ chooses each $u_i$ with probability $t_i$. Then every $q_j(\iota(\vec{u}))$ for $\iota \in S_D$ picks out the probability $t_{\iota(j)}$ i.e. $q_j$ outputs the probability of choosing the $j$th entry in its input tuple. Furthermore, both the density $f_{\vec{q}}$ and the entire single-edge strategy $\vec{q}$ are uniquely determined by any one of the $q_i$. That is, given a measurable function $q_i: [0,1]^D \rightarrow [0,1]$ whose integral over $[0,1]^D$ with respect to the product measure $\Lambda^{\times D}$ is $\frac1D$, which is invariant under permutations in $S_D$ fixing $i$, and which satisfies \[ \sum_j q_i(\iota_j(\vec{u})) = 1 \ \forall \vec{u} \] we can use cyclic shifts to define a valid corresponding single-edge strategy $\vec{q}$ (that satisfies $\sum q_j \equiv 1$ and $\int_{[0,1]^D} q_j \ d\Lambda^{\times D} = \frac1D$) and we can compute $f_{\vec{q}}$ directly from $q_i$. \end{remark} \begin{remark} It is easy to check that $\vec{p}^{MAX}$ satisfies \eqref{eq17}-\eqref{eq19}. Furthermore, if $\vec{p}$ is a consistent single-edge strategy and $\phi:[0,1] \rightarrow [0,1]$ is a $\Lambda$-measure-preserving bijection then so is $\vec{p}^{\phi}$. \end{remark} \begin{proof} The intuition is that we take the average of the original $p_i$ over the desired symmetries. We do this in two steps. For each $1 \leq i \leq D$ define \[p_i'(\vec{u}) := \frac{p_1(\iota_i(\vec{u})) + p_2(\iota_{i-1}(\vec{u})) + \ldots + p_i(\iota_1(\vec{u}))+ \ldots p_D(\iota_{i+1}(\vec{u}))}{D}\] \[q_i(\vec{u}) = \frac1{(D-1)!} \sum_{\iota \in S_{D}, \iota(i) = i} p_i'(u_{\iota(1)}, u_{\iota(2)} \ldots, u_{\iota(D)})\] A straightforward computation shows that $\vec{q}$ satisfies the required properties \eqref{eq17}-\eqref{eq19}. \end{proof} For the rest of the paper we restrict ourselves to these consistent single-edge strategies, which will simplify our computations. \subsection{Closure of \texorpdfstring{$\{\sigma_{\vec{p}}\}$}{mP}} The last remaining part of Theorem \ref{limitPts} is to show that the set of distributions $\sigma_{\vec{p}}$ is weakly closed. \begin{theorem}\label{closure} $\{\sigma_{\vec{p}}: \mbox{ single-edge strategies} \ \vec{p}\}$ is weakly closed \end{theorem} \begin{proof} Suppose $\sigma_{\vec{p}^{n}} \Rightarrow \xi$ for some consistent single-edge strategies $\vec{p}^n$. We seek a single-edge strategy yielding the distribution $\xi$. Consider any $n$. Let $\nu_{\vec{p}^n}$ be the distribution on $[0,1]^D$ given by integration against \\ $D p_1^n d\Lambda^{\times D}$. Observe that \eqref{eq17} implies \[ f_{\vec{p}^n}(y) = \int_{[0,1]^{D-1}} D p_1^n(y, u_2, \ldots, u_D) d\Lambda^{D-1}(u_2, \ldots, u_D) \] so $\sigma_{\vec{p}^n}$ is just the first coordinate marginal of $\nu_{\vec{p}^n}$. Compactness yields a weakly convergent subsequence $\nu_{\vec{p}^{n_j}} \Rightarrow \nu$. By a standard argument, since $Dp_1^n$ are uniformly bounded by $D$, then $\nu \ll \Lambda^{\times D}$ and has a density of the form $Dp_1$ for some measurable function $p_1: [0,1]^D \rightarrow [0,1]$. This along with the fact that $\int_{[0,1]^D} p_1 d\Lambda^{\times D} = \frac1D$ are enough for $p_1$ to give rise to a single-edge strategy $\vec{p}$ by the first Remark following Lemma \ref{lemma4}. But then the corresponding first coordinate marginals must also converge weakly, so we get $\sigma_{\vec{p}^{n_j}} \Rightarrow \xi'$ where $\xi'$ is the corresponding marginal of $\nu$: \[ \xi'(A) =\int_{A\times [0,1]^{D-1}} Dp_1 d\Lambda^D = \sigma_{\vec{p}}(A) \ \forall A \in \mathcal{B}([0,1])\] Thus $\xi' = \sigma_{\vec{p}}$. On the other hand, $\sigma_{\vec{p}^n} \Rightarrow \xi$ so by uniqueness of weak limits, $\xi = \xi' = \sigma_{\vec{p}}$. \end{proof} Therefore the set of possible distributions $\{\sigma_{\vec{p}}\}$ of $X(\vec{p})$ over single-edge strategies $\vec{p}$ almost surely coincides with the set of limit points of the empirical measures $\frac1n \mu_{0 \rightarrow n}(\chi)$ over all strategies $\chi$. In particular these have the same extreme points. \section{Extreme Points of \texorpdfstring{$\{\sigma_{\vec{p}}\}$}{mp}}\label{extreme} Our goal in this section is to characterize the extreme points of the possible distributions of $X$ we can observe as we vary the underlying single-edge strategy. It turns out that the extreme points of $\{\sigma_{\vec{p}}: \mbox{ single-edge strategies} \ \vec{p}\}$ are precisely those $\sigma_{\vec{p}}$ which have scramblings with mean converging to the mean of $\sigma_{MAX}$, or, equivalently, those $\sigma_{\vec{p}}$ with deterministic $\vec{p}$. \begin{theorem}\label{thm4} Let $\vec{p}$ be a single-edge strategy. The following are equivalent: \begin{enumerate}[label=(\roman*)] \item $\sigma_{\vec{p}}$ is an extreme point \item Any consistent single-edge strategy achieving $\sigma_{\vec{p}}$ must be deterministic \item $f_{\vec{p}}$ is not constant on sets of positive $\Lambda$-measure and $\sigma_{\vec{p}}$ is given by the following single-edge strategy $\vec{q}$: \[ q_k(u_1,\ldots, u_D) = \textbf{1}_{\{f_{\vec{p}}(u_k) \geq f_{\vec{p}}(u_i) \ \forall i\}} \ \mbox{for $\Lambda^{\times D}$-a.a.} \ (u_1,\ldots, u_D) \] In other words, $\sigma_{\vec{p}}$ is achieved by the deterministic single-edge strategy "choose \\ whichever label yields a higher value when evaluating the density $f_{\vec{p}}$" \item $\sup_{\phi} E[\sigma_{\vec{p}^{\phi}}] = E[\sigma_{MAX}]$ \item For $U \sim [0,1]$ if $f_{\vec{p}}(U), f_{MAX}(U)$ as $[0,D]$-valued random variables on the probability space $([0,1], \mathcal{B}([0,1]), \Lambda)$, then they have the same distribution. That is, \[P(f_{\vec{p}} \leq x) = P(f_{MAX} \leq x) \ \forall x\] \end{enumerate} \end{theorem} \begin{remark} Since $f_{MAX}(y) = Dy^{D-1}$ on $[0,1]$ then $\frac1D f_{MAX} \sim \mbox{Beta}(D,1)$. Thus (v) implies that for $U\sim \mbox{Unif}[0,1]$, $f_{\vec{p}}(U)$ has a continuous probability distribution. \end{remark} We prove this theorem over the next couple of sections. \subsection{Deterministic Single-Edge Strategies } We begin by characterizing the extreme points of the set of distributions $\{\sigma_{\vec{p}}\}$ in terms of deterministic single-edge strategies. \begin{lemma}\label{lemma6} (i) $\Leftrightarrow$ (ii) in Theorem \ref{thm4} \end{lemma} \begin{proof} We prove both contrapositives. First, consider a distribution $\sigma_{\vec{p}}$ achieved by a non-deterministic consistent single-edge strategy $\vec{p}$, i.e. a single-edge strategy satisfying \eqref{eq17}-\eqref{eq19}. We will construct a set of positive $\Lambda^{\times D}$ measure on which we perturb $p_1$ in a way that allows us to write $\vec{p}$ as a non-trivial average of two single-edge strategies $\vec{q}, \vec{r}$. By our assumption, there exists $0 < \epsilon < \frac12$ s.t. the set \[A = \{\vec{u}: p_1(\vec{u}) \in (\epsilon, 1-\epsilon), p_1(\vec{u}) \geq p_1(\iota_i(\vec{u})) \ \forall 1 \leq i \leq D\} \ \mbox{has positive $\Lambda^{\times D}$ measure}\] where $\iota_i$ are the cyclic shifts as before. Then there are $a_1 < b_1, a_2 < b_2$ s.t. $[a_1, b_1], [a_2, b_2]$ are disjoint and \[A' = A \cap [a_1, b_1] \times [a_2, b_2]^{D-1} \ \mbox{has positive $\Lambda^{\times D}$ measure}\] The purpose of this is to ensure that the first coordinate in a tuple in $A'$ cannot appear in any other position in another tuple in $A'$. By \eqref{eq18}, the fact that $p_1(\vec{u}) < 1-\epsilon$ on $A'$, and the definition of single-edge strategies, \[\sum_{k=2}^D p_1(\iota_k(\vec{u})) = \sum_{k=2}^D p_k(\vec{u}) = 1-p_1(\vec{u}) > \epsilon \ \forall \vec{u} \in A'\] Hence for any $\vec{u} \in A'$, there exists $2 \leq k \leq D$ s.t. $p_1(\iota_k(\vec{u})) > \frac{\epsilon}D$. It follows that there is $2 \leq k \leq D$ s.t. \[A'' = \bigg\{\vec{u} \in A': p_1(\iota_k(\vec{u})) \in \bigg(\frac{\epsilon}D, 1-\epsilon\bigg)\bigg\} \ \mbox{has positive $\Lambda^{\times D}$ measure}\] Define \[q_1 := p_1 + \frac{\epsilon}{D!} \sum_{\iota \in S_D, \iota(1)= 1} ( \mathbf{1}_{\iota(A'')} - \mathbf{1}_{\iota(\iota_k^{-1}(A''))}), r_1 := p_1 + \frac{\epsilon}{D!} \sum_{\iota \in S_D, \iota(1)= 1} ( - \mathbf{1}_{\iota(A'')} + \mathbf{1}_{\iota(\iota_k^{-1}(A''))})\] It is clear that $q_1, r_1$ average to $p_1$, and are both invariant under permutations in $S_D$ fixing 1. Since $p_1(\vec{u}) \in (\frac{\epsilon}D, 1-\epsilon) \ \forall \vec{u} \in \iota(A'') \cup \iota(\iota_k^{-1}(A'')) \ \forall \iota \in S_D, \iota(1) = 1$ then both $q_1, r_1$ have ranges in $[0,1]$. Also, note that \[\int_{[0,1]^D} \mathbf{1}_{\iota(B)} d\Lambda^{\times D} = \Lambda^{\times D}(B) \ \forall B \in \mathcal{B}([0,1]^D), \iota \in S_D \] so it follows that \[\int_{[0,1]^D} q_1 d\Lambda^{\times D} = \int_{[0,1]^D} r_1 d\Lambda^{\times D} = \int_{[0,1]^D} p_1 d\Lambda^{\times D} = \frac1D \] Finally, note that by a simple counting argument, for any $\vec{u}$, \[\sum_i \sum_{\iota \in S_D, \iota(1)= 1} \mathbf{1}_{\iota(A'')}(\iota_i(\vec{u})) = \sum_{\iota \in S_D} \mathbf{1}_{\iota(A'')}(\vec{u}) = \sum_i \sum_{\iota \in S_D, \iota(1)= 1} \mathbf{1}_{\iota(\iota_k^{-1}(A''))}(\iota_i(\vec{u})) \] so \[\sum_i q_1(\iota_i(\vec{u})) = \sum_i r_1(\iota_i(\vec{u})) = \sum_i p_1(\iota_i(\vec{u})) = 1 \ \forall \vec{u}\] Thus $q_1, r_1$ are valid probability functions which average to $p_1$. By the remark after Lemma \ref{lemma4}, these uniquely determine single-edge strategies $\vec{q}, \vec{r}$ whose average is $\vec{p}$. Thus $f_{\vec{p}}$ is the average of $f_{\vec{q}}, f_{\vec{r}}$. It remains to show this convex combination is non-trivial. Consider any $\vec{u} = (u_1, u_2,\ldots, u_D)$ with $u_1 \in [a_1, b_1]$. For any $\iota \in S_D, \iota(1) = 1$, \[\iota^{-1}(k) \neq 1 \Rightarrow u_{\iota^{-1}(k)} \in [a_2, b_2] \Rightarrow u_{\iota^{-1}(k)} \notin [a_1, b_1] \Rightarrow \vec{u} \notin \iota(\iota_k^{-1}(A''))\] It follows that \[ q_1(\vec{u}) = p_1(\vec{u}) + \frac{\epsilon}{D!} \sum_{\iota \in S_D, \iota(1)= 1} \mathbf{1}_{\iota(A'')}(\vec{u}) \geq p_1(\vec{u}) + \frac{\epsilon}{D!} \mathbf{1}_{A''}(\vec{u})\] so by \eqref{eq17}, \[ f_{\vec{q}}(u_1) \geq f_{\vec{p}}(u_1) + \frac{\epsilon}{(D-1)!}\int_{[0,1]^{D-1}} \mathbf{1}_{A''} (u_1, u_2, \ldots, u_D) \ d\Lambda^{\times (D-1)}(u_2,\ldots, u_D) \] Since $\Lambda^{\times D} (A'') > 0$ then there is $\delta > 0$ and a Borel set $B \subset [a_1, b_1]$ with $\Lambda(B)> 0$ s.t. the integral above is $\geq \delta$ for $u_1 \in B$. It follows that $f_{\vec{q}} \geq f_{\vec{p}} + \frac{\epsilon}{(D-1)!}\delta$ on $B$ so the convex combination is non-trivial. Therefore $\sigma_{\vec{p}}$ is not an extreme point. Now suppose $\sigma_{\vec{p}}$ is not an extreme point, say $\sigma_{\vec{p}} = t\sigma_{\vec{q}} + (1-t) \sigma_{\vec{r}}$ for some consistent single-edge strategies $\vec{q}, \vec{r}$ with $\sigma_{\vec{q}} \neq \sigma_{\vec{r}}$. Then $\sigma_{\vec{p}}$ is achieved by the single-edge strategy $t \vec{q} + (1-t) \vec{r}$, which trivially satisfies \eqref{eq17}-\eqref{eq19} so is consistent. Since $\vec{q}, \vec{r}$ have range in $[0,1]$ and differ on some set of positive $\Lambda^{\times D}$-measure then on this same set $t \vec{q} + (1-t) \vec{r}$ is not $\{0,1\}$-valued. Thus $\sigma_{\vec{p}}$ is achieved by a non-deterministic single-edge strategy satisfying \eqref{eq17}-\eqref{eq19}. \end{proof} \subsection{"Weight Tuples" of Single-Edge Strategies} $\sigma_{\vec{p}}$ being an extreme point will give us further properties relating to $(D+1)$-tuples \\ $(u_1, \ldots, u_{D+1})$, but these properties hold a.s. and we must be extra careful about which tuples living in measure 0 sets we are excluding. We need some setup to address this issue. Let $\vec{p}$ be a consistent single-edge strategy s.t. $\sigma_{\vec{p}}$ is an extreme point. By Lemma \ref{lemma6}, $p_1$ is 0-1 valued $\Lambda^{\times D}$-a.e.. Also, the set of tuples in $[0,1]^D$ with two or more duplicates has $\Lambda^{\times D}$-measure 0. Define \[ S := \{\mbox{ordered} \ (u_1,\ldots,u_D) \in {\mathbb{R}} ^D \ \mbox{ duplicate-free s.t.} \ p_1(u_1,\ldots,u_D) \in \{0,1\}\} \] to be the set of "good" ordered $D$-tuples we consider. Note that $\Lambda^D(S)= 1$. Once we prove more a.s. properties of the elements of $S$, we will amend this definition of $S$ accordingly. Combining the fact that $p_1$ is 0-1 valued on $S$ and the remark after Lemma \ref{lemma4}, we see that $p_1$ is equivalent to a choice function taking in \emph{unordered} tuples $(u_1, \ldots, u_D)$ and outputting one of the coordinates $u_j$; then for any $\iota \in S_D$, $p_1(\iota(\vec{u})) = \mathbf{1}_{\{j = \iota(1)\}}$ i.e. $p_1$ returns whether or not the first coordinate of the input ordered tuple is the choice associated with the corresponding unordered tuple. Let \[ T = \left\{ \begin{array}{l} \mbox{ordered} \ (u_1,\ldots,u_{D+1}) \in [0,1]^{D+1} \ \mbox{ s.t.} \\ \mbox{all $(D+1)!$ ordered subtuples of size $D$ are in $S$} \end{array}\right\} \] It is clear that $\Lambda^{\times (D+1)}(T) = 1$ and the tuples in $T$ contain no duplicate entries. $T$ is the set of "good" ordered $(D+1)$-tuples we restrict ourselves to. For ordered $(D+1)$-tuples in $T$, we wish to study the possible sets of choices we can make for the $(D+1)!$ ordered subtuples of size $D$. Since the choice for some $\vec{u} \in [0,1]^D$ is also the choice for $\iota(\vec{u})$ for all $\iota \in S_D$, then rather than keeping a factor of $D!$ everywhere, we instead consider the possible sets of choices we can make for the $D+1$ \emph{unordered} subtuples of size $D$. We can encode these choices as an ordered "weight tuple" $(w_1, \ldots, w_{D+1})$ where $w_j$ is the number of times $u_j$ is the choice made. In terms of the consistent single-edge strategy $\vec{p}$, \begin{equation}\label{equation4} w_j = w_j(u_1,\ldots, u_{D+1}) := \sum_{1 \leq k \leq D+1, k \neq j} p_1(u_j,u_1,\ldots, \widehat{u_j},\ldots,\widehat{u_k}, \ldots, u_{D+1}) \end{equation} For example, the weight tuple $(D, 1, 0, \ldots, 0)$ corresponds to the choices \[ (u_1,\ldots, \widehat{u_k}, \ldots, u_{D+1}) \mapsto u_1 \ \forall 2 \leq k \leq D+1 \ \mbox{and} \ (u_2,\ldots, u_{D+1})\mapsto u_2\] Note that each $u_j$ appears in exactly $D$ of the $D+1$ $D$-subtuples and there is exactly one choice for each subtuple so \begin{equation}\label{eq20} w_i \in \{0, 1,\ldots, D\} \ \forall 1 \leq i \leq D+1, \sum w_i \equiv D+1 \end{equation} Even though a weight tuple may correspond to more than one set of choices for the $D+1$ $D$-subtuples, it is easily checked that every valid weight tuple (satisfying \eqref{eq20}) corresponds to at least one set of choices. The following flow chart summarizes this process: \begin{center} \begin{tikzpicture} \node[draw] at (0,100pt) {Start with ordered $(D+1)$-tuple $(u_1,\ldots,u_{D+1}) \in T$}; \node[draw, align=left] at (0,50pt) {Consider the choice for each unordered $D$-subtuple $(u_1,\ldots, \widehat{u_k}, \ldots, u_{D+1})$. \\ Count occurrences of each $u_j$ among these choices.}; \node[draw] at (0,0pt) {Get an ordered $(D+1$)-weight tuple $(w_1, \ldots, w_{D+1})$ satisfying \eqref{eq20}}; \draw [->] (0,90pt) to (0,68pt); \draw [->] (0,33pt) to (0,10pt); \end{tikzpicture} \end{center} One last observation we make about the weight tuples is that they can be used to compute the density $f_{\vec{p}}$ directly. Fix any $1 \leq j \leq D+1$. Observe that \eqref{eq17} with a change of variables gives \begin{equation*} \begin{split} f_{\vec{p}}(t) &= D \int_{[0,1]^{D-1}} p_1(t,u_2, \ldots, u_D) d\Lambda^{D-1}(u_2, \ldots, u_D) \\ &= \sum_{1 \leq k \leq D+1, k \neq j} \int_{[0,1]^{D-1}} p_1(t,u_1,\ldots, \widehat{u_j},\ldots,\widehat{u_k}, \ldots, u_{D+1}) \\ &\qquad \qquad \qquad d\Lambda^{D-1}(u_1, \ldots, \widehat{u_j}, \ldots,\widehat{u_k}, \ldots, u_{D+1}) \end{split} \end{equation*} But $\Lambda$ is a probability measure so we can integrate everything against $du_k$ without changing the value: \begin{equation} \label{eqn6} \begin{split} f_{\vec{p}}(t) &= \sum_{1 \leq k \leq D+1, k \neq j} \int_{[0,1]^{D}} p_1(t,u_1,\ldots, \widehat{u_j},\ldots,\widehat{u_k}, \ldots, u_{D+1}) d\Lambda^{D}(u_1, \ldots, \widehat{u_j}, \ldots, u_{D+1}) \\ &=\int_{[0,1]^{D}} w_j(u_1,\ldots, t, \ldots, u_{D+1} ) d\Lambda^{D}(u_1, \ldots, \widehat{u_j}, \ldots, u_{D+1}) \ \mbox{by \eqref{equation4}} \end{split} \end{equation} where $t$ occurs in position $j$. Note that the above formula holds for any $t$ for which \[\{(u_1,\ldots, u_{j-1}, u_{j+1},\ldots, u_{D+1}) | (u_1,\ldots, u_{j-1},t,u_{j+1},\ldots, u_{D+1}) \in T\} \ \mbox{has} \ \Lambda^{\times D}-\mbox{measure 1}\] i.e. it holds for $t$ in a set of $\Lambda$-measure 1. \subsection{"Weight Tuples" of the Extreme Points} We claim that $\sigma_{\vec{p}}$ being an extreme point implies that $\Lambda^{\times (D+1)}$-almost all weight tuples must be a permutation of $(D,1,0,\ldots, 0)$. That means that almost all $(D+1)$-tuples in $T$ have one coordinate that "dominates" the others in terms of the choice function, and this behaviour will imply precisely that $f_{\vec{p}}$ can be approximated by "scrambles" of $f_{MAX}$. \begin{lemma} Suppose $\vec{p}$ is a consistent, single-edge strategy s.t. $\sigma_{\vec{p}}$ is an extreme point. For $\Lambda^{\times (D+1)}$-a.a. ordered tuples in $T$, the corresponding weight tuple is a permutation of $(D,1,0,\ldots,0)$. \end{lemma} \begin{proof} We sketch the details of the proof in the case $D \geq 3$. The case $D = 2$ is similar. Suppose the claim is false, say the set $U$ of ordered $(D+1)$-tuples $(u_1,\ldots, u_{D+1}) \in T$ whose weight tuple $(w_1,\ldots, w_{D+1})$ is not a permutation of $(D,1,0,\ldots,0)$ has positive $\Lambda^{\times (D+1)}$-measure. Note that $U$ is clearly invariant under permutations in $S_{D+1}$. We follow a similar approach as in the proof of Lemma \ref{lemma6}, in that we seek write $\vec{p}$ as a convex combination of two new single-edge strategies $\vec{q}, \vec{r}$ obtained by perturbing $\vec{p}$. First, we further restrict $U$. There exist $a_i < b_i$ s.t. $[a_i, b_i]$ are pairwise disjoint and s.t. \[U' := U \cap [a_1, b_1] \times \cdots \times [a_{D+1}, b_{D+1}] \ \mbox{has positive $\Lambda^{\times (D+1)}$ measure} \] In this way, each coordinate in a tuple in $U'$ cannot appear in any other position in another tuple in $U'$. Now consider any $(u_1, \ldots, u_{D+1}) \in U'$ so it has a maximal weight $1 \leq w_K \leq D-1$ in its weight tuple. If $w_K \geq 2$ then either there exists $i \neq K$ s.t. $w_i \geq 2$ or there are at least two $i \neq K$ s.t. $w_i = 1$ (this is because $\sum w_m = D+1$ and $w_K \leq D-1$). In either case, we can pick $I \neq K$ so that the choice for $(u_1, \ldots, \widehat{u_K}, \ldots, u_{D+1})$ is not $u_I$, hence there is another $D$-subtuple containing $u_K$ for which the choice is $u_I$. Furthermore, since $w_K \geq 2$, then there also is a $D$-subtuple containing $u_I$ for which the choice is $u_K$. On the other hand, if $w_K = 1$ then $w_m = 1 \ \forall m$. If $(u_{m_1}, \ldots, u_{m_D})$ is the $D$-subtuple for which the choice is $u_K$ then of the remaining $D-1 \geq 2$ coordinates $u_{m_{n}}$ with $m_n \neq K$ at least one, call it $u_I$, is the choice for a $D$-subtuple other than $(u_1, \ldots, \widehat{u_K}, \ldots, u_{D+1})$, i.e. for a $D$-subtuple containing $u_K$. Since this holds for all $\vec{u} \in U'$ then there exist distinct $1 \leq I, J, K, L \leq D+1$ s.t. \[ U'' := \left\{ \vec{u} \in U'\ \middle\vert \begin{array}{l} 1 \leq w_I, \ \mbox{and} \ w_m \leq w_K \ \forall m, \ \mbox{and} \\ \mbox{ the choice for the $D$-subtuple $(u_1,\ldots, \widehat{u_J}, \ldots, u_{D+1})$ is $u_I$, and} \\ \mbox{ the choice for the $D$-subtuple $(u_1,\ldots, \widehat{u_L}, \ldots, u_{D+1})$ is $u_K$} \end{array}\right\} \] has positive $\Lambda^{\times (D+1)}$ measure. The idea is that we write the weight tuples for $(D+1)$-tuples in $U''$ as averages of two different weight tuples, and use these new weight tuples to obtain two new non-trivial, valid single-edge strategies whose densities average to $f_{\vec{p}}$. Consider any $\vec{u} \in U''$. We can write the corresponding weight tuple $(w_1, \ldots, w_{D+1})$ as an average of two other weight tuples: \[(\ldots, w_I, \ldots, w_K, \ldots) = \frac12 ( \ldots, w_I+1, \ldots, w_K-1, \ldots) + \frac12 (\ldots, w_I-1, \ldots, w_K+1, \ldots) \] The weight tuple $( \ldots, w_I+1, \ldots, w_K-1, \ldots)$ can be achieved simply by changing the choice of $(u_1, \ldots, \widehat{u_L}, \ldots, u_{D+1})$ from $u_K$ to $u_I$, whereas the weight tuple $( \ldots, w_I-1, \ldots, w_K+1, \ldots)$ can be achieved by changing the choice of $(u_1, \ldots, \widehat{u_J}, \ldots, u_{D+1})$ from $u_I$ to $u_K$. It is trivial to see that these two new tuples are also valid weight tuples that satisfy \eqref{eq20} so they correspond to new single-edge strategies $\vec{q}, \vec{r}$ respectively. By \eqref{eqn6}, we see that $f_{\vec{p}}(y)$ is the average of $f_{\vec{q}}(y), f_{\vec{r}}(y)$. The only difference in the $D=2$ case is that the new weight tuples are achieved by making two changes to the choice function rather than one. It remains to check that this average is non-trivial. Let \[V := \{u_K: \exists u_1, \ldots, \widehat{u_K}, \ldots, u_{D+1} \ \mbox{s.t.} \ (u_1,\ldots, u_{D+1}) \in U''\} \] and for $x \in [0,1]$ let \[W^{x} := \{(u_1, \ldots, \widehat{u_K}, \ldots, u_{D+1}): \ (u_1,\ldots, , u_{K-1}, x, u_{K+1}, \ldots, u_{D+1}) \in U''\} \] Then $\Lambda^{(D+1) \times}(U'') > 0$ implies $\Lambda(V) > 0$ and for $\Lambda$-a.a. $u_K \in V$, $\Lambda^{\times D} (W^{u_K}) > 0$. Let us consider how the function $w_K(\cdot, \ldots, \cdot, u_K, \cdot, \ldots, \cdot): [0,1]^D \rightarrow \{0,1,\ldots, D\}$ changes from $\vec{p}$ to $\vec{r}$ for any given fixed $u_K \in V$. On $W^{u_K}$, $w_K(\cdot, \ldots, \cdot, u_K, \cdot, \ldots, \cdot)$ increases by 1 by construction of $\vec{r}$. Note that $u_K$ can only appear in the $K$th coordinate of a tuple in $U''$ (by construction of $U''$) hence $w_K(\cdot, \ldots, \cdot, u_K, \cdot, \ldots, \cdot)$ either remains unchanged or increases by 1 off of $W^{u_K}$ (because the only way it decreases by 1 is if $u_K$ had appeared at index $I$ in a tuple in $U''$). Thus, for any $u_K \in V$, \[f_{\vec{r}}(u_K) \geq f_{\vec{p}}(u_K) + \Lambda^{\times D}(W^{u_K}) \] by \eqref{eqn6}. It follows that $\sigma_{\vec{r}} \neq \sigma_{\vec{p}}$ so this is indeed a non-trivial convex combination. This contradicts the assumption that $\sigma_{\vec{p}}$ was an extreme point. Therefore $\Lambda^{\times (D+1)}$-a.a. weight tuples must be a permutation of $(D, 1,0,\ldots, 0)$. \end{proof} We amend our definition of "good" $(D+1)$-tuples: \[T' := \{(u_1,\ldots, u_{D+1}) \in T: (w_1,\ldots, w_{D+1}) \ \mbox{is a perm. of } \ (D,1,0,\ldots,0)\}\] We still have $\Lambda^{\times (D+1)}(T') = 1$ (provided of course that $\sigma_{\vec{p}}$ is an extreme point). We now prove the (i) $\Rightarrow$ (ii) direction of Theorem \ref{thm4}. We split the proof into several claims. \begin{lemma}\label{lemma8} Let $\sigma_{\vec{p}}$ be an extreme point.\\ (i) Suppose $(u_1, \ldots, u_{D+1}) \in T'$ and $u_i$ has weight $D$ in $(u_1, \ldots, u_{D+1})$. Then for any $1 \leq j \leq D+1$, $j \neq i$ and any $u_j'$ s.t. $(u_1, \ldots, u_{j-1}, u_j', u_{j+1}, \ldots, u_{D+1}) \in T'$, either $u_i$ or $u_j'$ has weight $D$ in $(u_1, \ldots, u_{j-1}, u_j', u_{j+1}, \ldots, u_{D+1})$. (ii) Suppose $(u_1, u_1', u_2, \ldots, u_D) \in T'$. Then in the corresponding weight tuple, $u_1'$ has weight $D$ and $u_1$ has weight 0 if and only if \[p_1(u_1, u_2, u_3, \ldots, u_D) < p_1(u_1', u_2, u_3, \ldots, u_D) \] (iii) For $\Lambda^{\times 2}$-a.a. $(u_1, u_1')$, if \[p_1(u_1, u_2, u_3, \ldots, u_D) < p_1(u_1', u_2, u_3, \ldots, u_D) \] on a set of positive $\Lambda^{\times (D-1)}$ measure then \[p_1(u_1, u_2, u_3, \ldots, u_D) \leq p_1(u_1', u_2, u_3, \ldots, u_D) \] for $\Lambda^{\times (D-1)}$-a.a. $(u_2, \ldots, u_D)$. (iv) For $\Lambda^{\times 2}$-a.a. $(u_1, u_1')$, if $f_{\vec{p}}(u_1) \leq f_{\vec{p}}(u_1')$ then \[p_1(u_1, u_2, u_3, \ldots, u_D) \leq p_1(u_1', u_2, u_3, \ldots, u_D) \] for $\Lambda^{\times (D-1)}$-a.a. $(u_2, \ldots, u_D)$. (v) $f_{\vec{p}}$ is not constant on any set of positive $\Lambda$ measure. (vi) For $\Lambda^{\times D}$-a.a. $(u_1, \ldots, u_D)$, \[ p_1(u_1,\ldots, u_D) \geq \mathbf{1}_{\{f_{\vec{p}}(u_1) \geq f_{\vec{p}}(u_i) \ \forall i\}} \] (vii) For $\Lambda^{\times D}$-a.a. $(u_1, \ldots, u_D)$, \[ p_1(u_1,\ldots, u_D) = \mathbf{1}_{\{f_{\vec{p}}(u_1) \geq f_{\vec{p}}(u_i) \ \forall i\}} \] \end{lemma} \begin{remark} We give heuristics to aid in understanding these claims:\\ (i) If $u_i$ is the dominant choice in $(u_1,\ldots, u_{D+1})$ then $u_i$ can never be dominated by one of the $u_j$ for $i \neq j$ in tuples containing $u_i$ and $u_j$.\\ (ii) Dominance can be determined by evaluating $p_1$.\\ (iii) If $u_1'$ dominates $u_1$ once then $u_1'$ almost always dominate $u_1$.\\ (iv) The ordering imposed by domination coincides with the ordering imposed by the values of $f_{\vec{p}}$.\\ (vii) $u_1$ is the choice in $(u_1,\ldots,u_D)$ if and only of $u_1$ maximizes the value of $f_{\vec{p}}$. \end{remark} \begin{proof} (i) Let $1 \leq i \leq D$ be the index for which $u_i$ has weight $D$ in $(u_1,\ldots, u_{D+1})$. Recall that this means we have the choices \[ (u_1,\ldots, \widehat{u_{\ell}},\ldots, u_{D+1}) \mapsto u_i \ \forall \ell \neq i\] Taking $\ell = j$, this means $u_i$ has weight 1 or $D$ in $(u_1,\ldots, u_{j-1}, u_j', u_{j+1},\ldots,u_{D+1})$ (because this $(D+1)$-tuple is in $T'$). If it is $D$, we are done. If it is 1, the fact that we have the choice \[(u_1,\ldots, u_{j-1}, u_{j+1},\ldots, u_{D+1}) \mapsto u_i \] implies no $u_k, k \neq j$ can have weight $D$ so $u_j'$ must have weight $D$ in \\ $(u_1,\ldots, u_{j-1}, u_j', u_{j+1},\ldots,u_{D+1})$. (ii) Recall that $p_1$ is 0-1 valued on $D$-subtuples of tuples in $T'$. We have the following sequence of if-and-only-if statements: \begin{align*} & p_1(u_1, u_2, u_3, \ldots, u_D) < p_1(u_1', u_2, u_3, \ldots, u_D) \\ &\Leftrightarrow p_1(u_1, u_2, u_3, \ldots, u_D) = 0, p_1(u_1', u_2, u_3, \ldots, u_D) = 1 \\ & \Leftrightarrow (u_1, u_2, u_3, \ldots, u_D) \mapsto u_i \ \mbox{for some} \ 2 \leq i \leq D \ \mbox{and} \ (u_1', u_2, u_3, \ldots, u_D) \mapsto u_1'\\ \end{align*} Now if $u_1'$ has weight $D$ and $u_1$ has weight 0 in $(u_1, u_1',u_2,\ldots, u_D$), then we must have \begin{equation} \label{equation6} (u_1, u_2, u_3, \ldots, u_D) \mapsto u_i \ \mbox{for some} \ 2 \leq i \leq D \ \mbox{and} \ (u_1', u_2, u_3, \ldots, u_D) \mapsto u_1' \end{equation} On the other hand, if we know \eqref{equation6} then the weight of $u_i$ cannot be $D$ since the choice for $(u_1', u_2, \ldots, u_i, \ldots, u_D)$ is $u_1'$. Thus $u_i$ has weight 1, so $u_1'$ has weight $D$ and $u_1$ has weight 0. (iii) Consider $u_1, u_1'$ s.t. $(u_1, u_1', u_2,\ldots, u_D) \in T'$ for $\Lambda^{\times (D-1)}$-a.a. $(u_2, \ldots, u_D)$. This clearly holds for $\Lambda^{\times 2}$-a.a. $(u_1, u_1')$ with $u_1 \neq u_1'$. Let \[U = \{(u_2, \ldots, u_D): (u_1, u_1', u_2,\ldots, u_D) \in T', p_1(u_1, u_2, u_3, \ldots, u_D) < p_1(u_1', u_2, u_3, \ldots, u_D) \}\] so $\Lambda^{\times (D-1)}(U) > 0$ by assumption. Consider any $(u_2, \ldots, u_D) \in U$ and $u_2'$ s.t. $(u_1, u_1', u_2', u_3, \ldots, u_D) \in T'$. By (ii), $u_1'$ has weight $D$ and $u_1$ has weight 0 in the weight tuple for $(u_1, u_1', u_2, u_3, \ldots, u_D)$. By (i), it follows that either $u_1'$ or $u_2'$ has weight $D$ in $(u_1, u_1', u_2', u_3, \ldots, u_D)$. By the contrapositive of (ii), \[p_1(u_1, u_2', u_3, \ldots, u_D) \leq p_1(u_1', u_2', u_3, \ldots, u_D) \] Note that for $\Lambda^{\times 2}$-a.a. $(u_1, u_1')$ it is true that for $\Lambda^{\times (D-1)}(U)>0$ of $(u_2, \ldots, u_D) \in U$ and $\Lambda$-a.a. $u_2' \in {\mathbb{R}} $ we have $(u_1, u_1', u_2', u_3, \ldots, u_D) \in T'$ and the inequality above holds. We repeat this argument, "replacing" $u_2, u_3, \ldots, u_D$ with $\Lambda$-almost arbitrary $u_2', u_3', \ldots, u_D'$ and noting the invariant that one of $u_1',\ldots, u_k'$ has weight of $D$ in the weight tuple of $(u_1, u_1', u_2', \ldots, u_k', u_{k+1}, \ldots, u_D) \in T'$ (which means we may apply (ii) to replace $u_{k+1}$ with $\Lambda$-almost arbitrary $u_{k+1}'$). Thus, for $\Lambda^{\times 2}$-a.a. $(u_1, u_1')$, \[p_1(u_1, u_2', u_3', \ldots, u_D') \leq p_1(u_1', u_2', u_3', \ldots, u_D') \] for $\Lambda^{(D-1)}$-a.a. $(u_2', \ldots, u_D')$. (iv) This follows immediately from (iii) and \eqref{eq17}. (v) Suppose $f_{\vec{p}}$ is constant on a set $V$ with $\Lambda(V) > 0$. By (iv), can restrict $V$ to a set $V'$ with $\Lambda(V') = \Lambda(V) > 0$ s.t. for all distinct $u_1, u_2, \ldots, u_{D+1} \in V'$, $(u_1, \ldots, u_{D+1}) \in T'$ and \[ p_1(u_i, u_2, \ldots, \widehat{u_i}, \ldots, \widehat{u_j}, \ldots, u_{D+1}) = p_1(u_j, u_2, \ldots, \widehat{u_i}, \ldots, \widehat{u_j}, \ldots, u_{D+1}) \ \forall i, j \] We claim this is impossible. Consider distinct $u_1, \ldots, u_{D+1} \in V'$. Suppose $u_i$ has weight $D$ and $u_k$ has weight 1 in the corresponding weight tuple. Pick $j \notin \{i, k\}$ (which exists since $D+1 \geq 3$). Then $u_j$ has weight 0 in the corresponding weight tuple. But this implies the contradiction \[ 1 = p_1(u_i, u_2, \ldots, \widehat{u_i}, \ldots, \widehat{u_j}, \ldots, u_{D+1}) = p_1(u_j, u_2, \ldots, \widehat{u_i}, \ldots, \widehat{u_j}, \ldots, u_{D+1}) = 0 \] Therefore $f_{\vec{p}}$ is not constant on any set of positive $\Lambda$-measure. (vi) It suffices to show that for $\Lambda$-a.a. $u_1$, we have \[p_1(u_1,u_2,\ldots, u_D) = 1 \] for all but a $\Lambda^{\times (D-1)}$-null subset of tuples $(u_2,\ldots, u_D) \in \bigg(f_{\vec{p}}^{-1}([0, f_{\vec{p}}(u_1)]) \bigg)^{D-1}$. Suppose not, say there is a set $G$ of $u_1$ of $\Lambda$-positive measure on which \[p_1(u_1,u_2,\ldots, u_D) = 0 \] for a tuples $(u_2,\ldots, u_D)$ in a set $H_{u_1} \subseteq \bigg(f_{\vec{p}}^{-1}([0, f_{\vec{p}}(u_1)])\bigg)^{D-1}$ of $\Lambda^{\times (D-1)}$-positive measure. By (iv), for $\Lambda$-a.a. $u_1 \in G$, \[0= p_1(u_1,u_2,\ldots, u_D) \geq p_1(u_{D+1}, u_2,\ldots, u_D)\] for $\Lambda$-a.a. $u_{D+1} \in f_{\vec{p}}^{-1}([0, f_{\vec{p}}(u_1)])$ and $\Lambda^{\times (D-1)}$-a.a. $(u_2,\ldots, u_D) \in H_{u_1}$. In particular, there exists $2 \leq i \leq D$ and a subset $H_{u_1}' \subset H_{u_1}$ of positive $\Lambda^D$-measure s.t. the choice for $(u_1, u_2, \ldots, u_D)$ is $u_i$ and s.t. for $\Lambda$-a.a. $u_{D+1} \in f_{\vec{p}}^{-1}([0, f_{\vec{p}}(u_1)])$ we have \begin{equation}\label{eqn9} p_1(u_1,u_2,\ldots, u_D) = p_1(u_{D+1}, u_2,\ldots, u_D) = 0 \end{equation} For each such tuple consider the weight of $u_i$ in $(u_1,\ldots, u_{D+1})$; it must be 1 or $D$ and if it were 1 then the choice $(u_1,u_2,\ldots, u_D) \mapsto u_i$ implies none of $u_2,\ldots, u_D$ can have weight $D$ so $u_{D+1}$ must have weight $D$, contradicting $p_1(u_{D+1}, u_2,\ldots, u_D)= 0$. Thus $u_i$ has weight $D$ in $(u_1,\ldots, u_{D+1})$ so we must have the choice $(u_{D+1}, u_2,\ldots, u_D) \mapsto u_i$. Therefore \begin{equation}\label{eqn10} p_1(u_i, u_2, \ldots, u_{i-1}, u_{D+1}, u_{i+1}, \ldots, u_D) = 1 \end{equation} Since \eqref{eqn9} holds for $\Lambda$-a.a. $u_{D+1} \in f_{\vec{p}}^{-1}([0, f_{\vec{p}}(u_1)])$ and $\Lambda^{\times(D-1)}$-a.a. $(u_2,\ldots, u_D) \in H_{u_1}'$, then there is a set of tuples $(u_2,\ldots, u_{D+1})$ of positive $\Lambda^{\times D}$-measure for which \eqref{eqn9} (hence \eqref{eqn10}) holds for both $(u_2, \ldots, u_{D+1})$ and $(u_2,\ldots, u_{i-1}, u_{D+1}, u_{i+1}, \ldots, u_D, u_i)$ (i.e. it holds with the values $u_i, u_{D+1}$ swapped). Combining the two sets of \eqref{eqn9}, \eqref{eqn10} gives the contradiction \[0 = p_1(u_i, u_2, \ldots, u_{i-1}, u_{D+1}, u_{i+1}, \ldots, u_D) = 1 \] (vii) By (v) and (vi), for $\Lambda^{\times D}$-a.a. $(u_1,\ldots,u_D)$, $f_{\vec{p}}(u_i)$ are distinct and \[ p_1(u_k, u_1,\ldots, \widehat{u_k}, \ldots, u_D) \geq \mathbf{1}_{\{f_{\vec{p}}(u_k) \geq f_{\vec{p}}(u_i) \ \forall i\}} \ \forall k\] Consider any such $(u_1,\ldots, u_D)$ and let $u_k$ maximize $f_{\vec{p}}$. Then \[ p_1(u_k, u_1,\ldots, \widehat{u_k}, \ldots, u_D) = \mathbf{1}_{\{f_{\vec{p}}(u_k) \geq f_{\vec{p}}(u_i) \ \forall i\}} = 1\] so the choice for $(u_1,\ldots, u_D)$ is $u_k$. It follows that for all other $j \neq k$, \[ p_1(u_j, u_1,\ldots, \widehat{u_j}, \ldots, u_D) = 0 = \mathbf{1}_{\{f_{\vec{p}}(u_j) \geq f_{\vec{p}}(u_i) \ \forall i\}}\] We have shown that the a.s. inequality in (vi) is an a.s. equality. \end{proof} The next lemma shows that (ii) implies (iii) in Theorem \ref{thm4}. \begin{lemma}\label{Lemma10} Suppose $\vec{p}$ is a single-edge strategy s.t. \begin{enumerate}[label=(\roman*)] \item $f_{\vec{p}}$ is not constant on any set of positive $\Lambda$ measure, and \item $p_k(u_1,\ldots, u_D) = \textbf{1}_{\{f_{\vec{p}}(u_k) \geq f_{\vec{p}}(u_i) \ \forall i\}} \ \mbox{for $\Lambda^{\times D}$-a.a.} \ (u_1,\ldots, u_D) \ \forall 1 \leq k \leq D$ \end{enumerate} Then $\sup_{\phi} E[\sigma_{\vec{p}^{\phi}}] = E[\sigma_{MAX}]$ where the supremum is taken over $\Lambda$-measure-preserving bijections $\phi: [0,1] \rightarrow [0,1]$. \end{lemma} \begin{proof} We begin by constructing a sequence $(\phi_n)$ of $\Lambda$-measure-preserving bijections for which \begin{equation}\label{24} \mathbf{1}_{\{f_{\vec{p}}(\phi_n(u_1)) \geq f_{\vec{p}}(\phi_n(u_i)) \ \forall i\}} \rightarrow \mathbf{1}_{\{u_1 \geq u_i \ \forall i\}} \ \mbox{$\Lambda^{\times D}$-a.s.} \end{equation} Consider any $n \in {\mathbb{N} } $. By (i), inverse images of singletons $f_{\vec{p}}(\{c\})$ have $\Lambda$-measure 0. There exist $ 0 = a_0^n \leq a_1^n \leq \cdots \leq a_{D 2^n}^n = 1$ s.t. \begin{equation}\label{eq24} \Lambda ([a_i^n, a_{i+1}^n)) = \Lambda \bigg(f_{\vec{p}}^{-1} \bigg( \bigg[ \frac{i}{2^n}, \frac{i+1}{2^n} \bigg) \bigg)\bigg) \ \forall 0 \leq i \leq D2^n-1 \end{equation} The indices stop at $D2^n-1$ since the range of $f_{\vec{p}}$ is a subset of $[0,D]$. Clearly, we can make it so that the partition for $n+1$ refines the one for $n$, for all $n$. We construct $\phi_n$ by pasting together $\Lambda$-measure-preserving bijections we get from Theorem \ref{thmNishiura} between sets \[ [a_i^n, a_{i+1}^n) \rightarrow f_{\vec{p}}^{-1} \bigg( \bigg[ \frac{i}{2^n}, \frac{i+1}{2^n} \bigg) \bigg)\] for $0 \leq i \leq D2^n-1$ for which $f_{\vec{p}}^{-1} ( [ \frac{i}{2^n}, \frac{i+1}{2^n} ) )$ has positive $\Lambda$-measure. Then for each such $i$, $f_{\vec{p}^{\phi_n}} = f_{\vec{p}} \circ \phi_n$ is a map \[ [a_i^n, a_{i+1}^n) \rightarrow \bigg[ \frac{i}{2^n}, \frac{i+1}{2^n} \bigg) \] Let us show \eqref{24}. It suffices to show that for $\Lambda^{\times 2}$-a.a. $(u_1, u_2)$ we have \[\mathbf{1}_{\{f_{\vec{p}}(\phi_n(u_1)) \geq f_{\vec{p}}(\phi_n(u_2)) \}} \rightarrow \mathbf{1}_{\{u_1 \geq u_2 \}}\] since multiplying $D-1$ of these sequences of indicators gives $\eqref{24}$. By (i), the sets $f_{\vec{p}}^{-1}(\{c\})$ have $\Lambda$-measure 0 hence it follows that for $\Lambda^{\times 2}$-a.a. $(u_1, u_2)$ we have $f_{\vec{p}}(u_1) \neq f_{\vec{p}}(u_2)$. Consider any such $u_1, u_2$. Since $\Lambda([a_i^n, a_{i+1}^n)) \downarrow 0$ as $n \rightarrow \infty$ (because $\Lambda \ll $ Lebesgue measure), then for large enough $n$, $u_1$ and $u_2$ are in different intervals of the form $[a_i^n, a_{i+1}^n)$ and they stay in these intervals (because the partitions get progressively more refined). Thus $f_{\vec{p}}\circ \phi_n (u_1), f_{\vec{p}}\circ \phi_n (u_2)$ are in different intervals of the form $[\frac{i}{2^n}, \frac{i+1}{2^n})$ and the order is preserved: \[u_1 < u_2 \Leftrightarrow u_1 \in [a_{i_1}^n, a_{i_1+1}^n), u_2 \in [a_{i_2}^n, a_{i_2+1}^n) \ \mbox{with} \ i_1 < i_2 \Leftrightarrow f_{\vec{p}}\circ \phi_n (u_1) < f_{\vec{p}}\circ \phi_n (u_2)\] Therefore we get \eqref{24}. Recall that $\vec{p}^{\phi_n}$ are themselves consistent single-edge strategies. Using (ii) and \eqref{cdf}, it follows that for any $t \in [0,1]$, \begin{equation}\label{limit} \begin{split} 1-F_{\vec{p}^{\phi_n}}(t) &= D\int_{[t,1] \times [0,1]^{D-1}} p_1(\phi_n(\vec{u})) d\Lambda^{D}(u_1, \ldots, u_{D}) \\ & = D\int_{[t,1] \times [0,1]^{D-1}} \mathbf{1}_{\{f_{\vec{p}}(\phi_n(u_1)) \geq f_{\vec{p}}(\phi_n(u_i)) \ \forall i\}} d\Lambda^{D}(u_1, \ldots, u_{D}) \end{split} \end{equation} By \eqref{24} and the Bounded Convergence Theorem, this last expression converges to the integral of the density $f_{MAX}$: \begin{equation*} \begin{split} & D\int_{[t,1] \times [0,1]^{D-1}} \mathbf{1}_{\{f_{\vec{p}}(\phi_n(u_1)) \geq f_{\vec{p}}(\phi_n(u_i)) \ \forall i\}} d\Lambda^{D}(u_1, \ldots, u_{D}) \\ & \rightarrow D\int_{[t,1] \times [0,1]^{D-1}} \mathbf{1}_{\{u_k \geq u_i \ \forall i\}}d\Lambda^{D}(u_1, \ldots, u_{D}) \\ & = \int_{t}^{1} f_{MAX}(u_1) du_1 \\ & = 1- F_{MAX}(t) \end{split} \end{equation*} But we already established that $F_{\vec{q}} \geq F_{MAX}$ for any single-edge strategy $\vec{q}$. Therefore $F_{\vec{p}^{\phi_n}} \rightarrow F_{MAX}$ pointwise from below hence $E[\sigma_{\vec{p}^{\phi_n}}] \rightarrow E[\sigma_{MAX}]$ by the Bounded Convergence Theorem combined with the tail integral formula for expectation. \end{proof} \subsection{Theorem \ref{thm4}} We wish to prove the rest of Theorem \ref{thm4}, but first we make some useful observations about $\sigma_{MAX}$. \begin{lemma}\label{lemma9} Let $f_{MAX}^-$ be the quantile function \[ f_{MAX}^-(t) = \inf \{x: t \leq f_{MAX}(x) \} \] Then treating $f_{MAX}$ itself as a $[0,D]$-valued random variable on the probability space \\ $([0,1], \mathcal{B}([0,1]), \Lambda)$, \[ P(f_{MAX} \leq t) = f_{MAX}^-(t) \ \forall t\] In particular, \[ P(f_{MAX} \leq f_{MAX}(y)) = y \ \forall y \] \end{lemma} \begin{proof} This is trivial since the density $f_{MAX}(y) = Dy^{D-1}$ is invertible so the quantile function $f_{MAX}^-$ is just its inverse. \end{proof} To prove Theorem \ref{thm4}, we need one more short lemma exploring what \\ $\sup_{\phi} E[\sigma_{\vec{p}^{\phi}}] = E[\sigma_{MAX}]$ tells us. \begin{lemma}\label{lemma10} Let $\vec{p}$ be a single-edge strategy and $\phi_n: [0,1] \rightarrow [0,1]$ be $\Lambda$-measure-preserving bijections s.t. $E[\sigma_{\vec{p}^{\phi}}] \uparrow E[\sigma_{MAX}]$. Then there is a subsequence $\phi_{n_j}$ s.t. $F_{\vec{p}^{\phi_{n_j}}} (y) \rightarrow F_{MAX}(y)$, $f_{\vec{p}^{\phi_{n_j}}}(y) = f_{\vec{p}} \circ \phi_{n_j}(y) \rightarrow f_{MAX}(y)$ $\Lambda$-a.e. and almost uniformly on compact subsets of $[0,1]$. \end{lemma} \begin{proof} By the tail integral formula for expectation, \[ \int_{0}^{1} F_{\vec{p}^{\phi_n}}(t)- F_{MAX}(t) dt = E[\sigma_{MAX}] - E[\sigma_{\vec{p}^{\phi}}] \downarrow 0\] where the integrands $F_{\vec{p}^{\phi_n}}(t)- F_{MAX}(t)$ are non-negative from Lemma \ref{lemma3} (i), hence this is $L^1$ convergence. It follows that there exists a subsequence $n_j$ s.t. $F_{\vec{p}^{\phi_{n_j}}}(t) \rightarrow F_{MAX}(t)$ a.e. and a.u. on compact subsets of $[0,1]$. The latter convergence gives us that $f_{\vec{p}^{\phi_{n_j}}}(t) \rightarrow f_{MAX}(t)$ a.e. and a.u. on compact subsets of $[0,1]$. \end{proof} We are now ready to prove Theorem \ref{thm4} in its entirety. \begin{manualtheorem}{\ref{thm4}} Let $\vec{p}$ be a single-edge strategy. The following are equivalent: \begin{enumerate}[label=(\roman*)] \item $\sigma_{\vec{p}}$ is an extreme point \item Any consistent single-edge strategy achieving $\sigma_{\vec{p}}$ must be deterministic \item $f_{\vec{p}}$ is not constant on sets of positive $\Lambda$-measure and $\sigma_{\vec{p}}$ is given by the following single-edge strategy $\vec{q}$: \[ q_k(u_1,\ldots, u_D) = \textbf{1}_{\{f_{\vec{p}}(u_k) \geq f_{\vec{p}}(u_i) \ \forall i\}} \ \mbox{for $\Lambda^{\times D}$-a.a.} \ (u_1,\ldots, u_D) \] In other words, $\sigma_{\vec{p}}$ is achieved by the deterministic single-edge strategy "choose whichever label yields a higher value when evaluating the density $f_{\vec{p}}$" \item $\sup_{\phi} E[\sigma_{\vec{p}^{\phi}}] = E[\sigma_{MAX}]$ \item For $U \sim [0,1]$ if $f_{\vec{p}}(U), f_{MAX}(U)$ as $[0,D]$-valued random variables on the probability space $([0,1], \mathcal{B}([0,1]), \Lambda)$, then they have the same distribution. That is, \[P(f_{\vec{p}} \leq t) = P(f_{MAX} \leq t) = \bigg(\frac{t}D\bigg)^{\frac1{D-1}} \ \forall t \in [0,D]\] \end{enumerate} \end{manualtheorem} \begin{proof} (i) $\Leftrightarrow $ (ii), (i) $\Rightarrow$ (iii), (iii) $\Rightarrow $ (iv) These are given by Lemmas \ref{lemma6}, \ref{lemma8}, \ref{Lemma10}. (iv) $\Rightarrow$ (i) Suppose (iv) holds but $\sigma_{\vec{p}}$ is not an extreme point. Since Radon-Nikodym derivatives are additive, then \[ f_{\vec{p}} = \alpha f_{\vec{q}} + (1-\alpha) f_{\vec{r}}\] for $\alpha \in (0,1)$ and single-edge strategies $\vec{q}, \vec{r}$ s.t. $f_{\vec{p}} \neq f_{\vec{q}}$ on a set of positive $\Lambda$ measure. Take $\phi_n$ $\Lambda$-measure-preserving bijections s.t. $E[\sigma_{\vec{p}^{\phi_n}}] \uparrow E[\sigma_{MAX}]$. Then \[\alpha E[\sigma_{\vec{q}^{\phi_n}}] + (1-\alpha) E[\sigma_{\vec{r}^{\phi_n}}] \uparrow \alpha E[\sigma_{MAX}] + (1-\alpha) E[\sigma_{MAX}] \] where $E[\sigma_{\vec{q}^{\phi_n}}], E[\sigma_{\vec{r}^{\phi_n}}] \leq E[\sigma_{MAX}]$. It follows that \[E[\sigma_{\vec{q}^{\phi_n}}], E[\sigma_{\vec{r}^{\phi_n}}] \uparrow E[\sigma_{MAX}]\] By Lemma \ref{lemma10}, there is a subsequence $\phi_{n_j}$ s.t. \[ f_{\vec{q}} \circ \phi_{n_j}, f_{\vec{r}} \circ \phi_{n_j} \rightarrow f_{MAX} \] $\Lambda$-a.e. and $\Lambda$-a.u. on compact subsets of $[0,1]$. Since $f_{\vec{q}} \neq f_{\vec{r}}$ on a set of positive $\Lambda$-measure then $\exists \epsilon > 0$ s.t. \[Q := \{x: |f_{\vec{q}}(x) - f_{\vec{r}}(x)| \geq \epsilon \} \] has measure $\delta := \Lambda(Q) >0$. Now, by the choice of $\phi_{n_j}$, there is a subset $R \subset [0, 1]$ of $\Lambda$ measure $\leq \frac{\delta}2$ s.t. $f_{\vec{q}}\circ \phi_{n_j}, f_{\vec{r}} \circ \phi_{n_j} \rightarrow f_{MAX}$ uniformly on $[0, 1] \setminus R$. In particular, $\exists J \in {\mathbb{N} } $ s.t. for all $j \geq J$ and $y \in [0, 1] \cap R$ we have \[|f_{\vec{q}} \circ \phi_{n_j}(y) - f_{MAX}(y)|, |f_{\vec{r}} \circ \phi_{n_j}(y) - f_{MAX}(y)| < \frac{\epsilon}2 \Rightarrow |f_{\vec{q}} \circ \phi_{n_j}(y) - f_{\vec{r}} \circ \phi_{n_j}(y) | < \epsilon\] Thus \begin{equation}\label{eq25} \phi_{n_J}([0,1] \setminus R) \subseteq [0,1] \setminus Q \end{equation} But $\phi_{n_J}$ is $\Lambda$-measure-preserving and \begin{align*} \Lambda(\phi_{n_J}([0,1] \setminus R)) &= \Lambda([0,1]) - \Lambda(R)\\ &\geq \Lambda([0,1]-\frac{\delta}2 \\ &> \Lambda([0,1])-\Lambda(Q) \\ &= \Lambda([0,1] \setminus Q) \end{align*} Contradiction of \eqref{eq25}. Therefore $\sigma_{\vec{p}}$ is an extreme point. (iv) $\Rightarrow$ (v) Take $\phi_n$ $\Lambda$-measure-preserving bijections s.t. $E[\sigma_{\vec{p}^{\phi_n}}] \uparrow E[\sigma_{MAX}]$. By Lemma \ref{lemma10} there is a subsequence $\phi_{n_j}$ s.t. $f_{\vec{p}^{\phi_{n_j}}} = f_{\vec{p}} \circ \phi_{n_j} \rightarrow f_{MAX}$ $\Lambda$-a.e and a.u. on compact subsets of $[0,1]$. Now for any $z \in [0,1]$, \[P(f_{\vec{p}} \leq z) = \Lambda(f_{\vec{p}}^{-1}([0,z])) = \Lambda( (f_{\vec{p}^{\phi_{n_j}}})^{-1} ([0,z]))\] We claim this last expression equals $\Lambda(f_{MAX}^{-1}([0,z])) = P(f_{MAX} \leq z)$. Suppose $P(f_{\vec{p}} \leq z) \neq P(f_{MAX} \leq z)$ for some $z$, say $P(f_{\vec{p}} \leq z) < \Lambda(f_{MAX}^{-1}([0,z]))$; the $>$ case is very similar. By continuity from below and since $\Lambda(f_{MAX}^{-1}(\{z\})) = 0$, there exists $ \delta > 0$ s.t. \[\Lambda( f_{\vec{p}} ^{-1} ([0,z])) < \Lambda(f_{MAX}^{-1} ([0, z-\delta]))\] By a.u. convergence on $[0,1]$, there is a subset $R \subset [0, 1]$ off of which $f_{\vec{p}} \circ \phi_{n_j} \rightarrow f_{MAX}$ uniformly s.t. $\Lambda(R)$ is sufficiently small so that for all $j$, \begin{equation}\label{eq26} \Lambda( (f_{\vec{p}} \circ \phi_{n_j})^{-1} ([0,z])) = \Lambda( f_{\vec{p}} ^{-1} ([0,z])) < \Lambda(f_{MAX}^{-1} ([0, z-\delta]) \cap [0,1] \setminus R) \end{equation} Uniform convergence gives a $J$ s.t. for all $j \geq J$ and all $t \in f_{MAX}^{-1} ([0, z-\delta]) \cap [0,1] \setminus R$, \[ f_{\vec{p}} \circ \phi_{n_j}(t) \in (f_{MAX}(t)- \delta, f_{MAX}(t) + \delta) \cap [0,\infty) \subseteq [0, z] \] (here we use the fact that $f_{\vec{p}}$ is a non-negative function). Thus \[f_{MAX}^{-1} ([0, z-\delta]) \cap [0,1] \setminus R \subseteq (f_{\vec{p}} \circ \phi_{n_j}) ^{-1} ([0,z]) \] which is a contradiction of \eqref{eq26}. Therefore \[P(f_{\vec{p}} \leq z) = P(f_{MAX} \leq z) = f_{MAX}^-(z) \] where the last inequality holds by Lemma \ref{lemma9}. The expression on the right is precisely $ (\frac{z}D)^{\frac1{D-1}}$. (v) $\Rightarrow$ (iv) We construct $\Lambda$-measure-preserving bijections $\phi_n : [0,1] \rightarrow [0,1]$ s.t.\\ $E[\sigma_{\vec{p}^{\phi_n}}] \rightarrow E[\sigma_{MAX}]$. Consider any $n$. By (v) combined with Lemma \ref{lemma9}, for any $z \in \{0,1,\ldots, 2^n-1\}$ we have \begin{align*} &\Lambda \bigg( f_{\vec{p}}^{-1} \bigg( \bigg[f_{MAX} \bigg(\frac{z}{2^n} \bigg), f_{MAX} \bigg(\frac{z+1}{2^n} \bigg) \bigg) \bigg) \bigg) \\ &= P\bigg(f_{MAX} \bigg(\frac{z}{2^n} \bigg) \leq f_{\vec{p}} < f_{MAX} \bigg(\frac{z+1}{2^n} \bigg) \bigg) \\ &= P\bigg(f_{MAX} \bigg(\frac{z}{2^n} \bigg) \leq f_{MAX} < f_{MAX} \bigg(\frac{z+1}{2^n} \bigg) \bigg) \\ &= \Lambda\bigg( \bigg[\frac{z}{2^n}, \frac{z+1}{2^n} \bigg) \bigg) \end{align*} We construct $\phi_n$ by pasting together $\Lambda$-measure-preserving bijections we get from Theorem \ref{thmNishiura} between sets \[\bigg[\frac{z}{2^n}, \frac{z+1}{2^n} \bigg) \rightarrow f_{\vec{p}}^{-1} \bigg( \bigg[f_{MAX} \bigg(\frac{z}{2^n} \bigg), f_{MAX} \bigg(\frac{z+1}{2^n} \bigg) \bigg) \bigg) \] Then for each $z$, $f_{\vec{p}^{\phi_n}} = f_{\vec{p}} \circ \phi_n$ is a map \[\bigg[\frac{z}{2^n}, \frac{z+1}{2^n} \bigg) \rightarrow \bigg[f_{MAX} \bigg(\frac{z}{2^n} \bigg), f_{MAX} \bigg(\frac{z+1}{2^n} \bigg) \bigg) \] We now compute the expectation of each $X(\vec{p}^{\phi_n})$: \begin{align*} E[X(p^{MAX}) ] &\geq E[X_0(\vec{p}^{\phi_n}) ] \\ &= \sum_{z=0}^{2^n-1} \int_{\frac{z}{2^n}}^{\frac{z+1}{2^n}} f_{\vec{p}}(\phi_n(t)) t\ dt\\ &\geq \sum_{z=0}^{2^n-1} \int_{\frac{z}{2^n}}^{\frac{z+1}{2^n}} f_{MAX} \bigg(\frac{z}{2^n} \bigg) t \ dt \end{align*} But $f_{MAX}(t) t$ is nondecreasing hence is of bounded variation on $[0,1]$ hence this lower bound converges to $E[X(p^{MAX}) ]$ as $n \rightarrow \infty$. Thus \[ E[\sigma_{MAX}] = E[X(p^{MAX}) ] = \sup_n E[X(\vec{p}^{\phi_n}) ] =\sup_{\phi} E[\sigma_{\vec{p}^{\phi}} ] \] as desired. \end{proof} By the Krein-Milman Theorem, $\mathcal{R}$ is exactly the closed convex hull of its extreme points. Thus we obtain a description of the set of limit points of empirical measures. As a final remark, observe that we only used the compactness of the space $\mathcal{M}_1$ of probability measures on [0,1] to reduce from generic strategies to single-edge strategies. Furthermore, the fact that $\Lambda$ has nice formulas for its cdf and pdf was convenient but unnecessary. The proof of this theorem could be tweaked to hold with $\Lambda$ replaced by an arbitrary distribution $\theta$ on $ {\mathbb{R}} $ with finite mean. We could thus get a similar characterization for the extreme points of $\{\sigma_{\psi}: \mbox{single-edge strategies} \ \psi\}$. The only caveats would be that this set might not coincide with $\{\mbox{limit points of } \ \frac1n \mu_{0 \rightarrow n}(\chi): \ \mbox{strategies} \ \chi\}$ and that the value distribution of the densities $f_{\vec{p}}$ of extreme points might not have as nice a form as $D \cdot Beta(1,D)$. \subsection{The Discrete Case with Example}\label{discrete} The same argument with the weight tuples can be used to show a discrete version of Theorem \ref{thm4}, where the i.i.d. labels $U^j$ are $Unif\{1,\ldots,K\}$. \begin{manualtheorem}{13'}\label{13discrete} Let $\vec{p}$ be a single-edge strategy. The following are equivalent \begin{enumerate}[label=(\roman*)] \item $\sigma_{\vec{p}}$ is an extreme point \item Any consistent single-edge strategy achieving $\sigma_{\vec{p}}$ must be deterministic. \item $\sigma_{\vec{p}}$ is given by the single-edge strategy "choose whichever label is maximal with respect to the ordering $\alpha(1) < \cdots < \alpha(K)$" for some $\alpha \in S_K$. \item There exists a permutation $\beta \in S_K$ s.t. $\vec{p}(\beta(\vec{u})) = \vec{p}^{MAX}(\vec{u}) \ \forall \vec{u}$. \item $\sigma_{\vec{p}}$ has a probability mass function whose value distribution is \[\bigg\{\frac1{K^D}, \frac{2^D-1}{K^D}, \frac{3^D-2^D}{K^D}, \ldots, \frac{K^D-(K-1)^D}{K^D}\bigg\}\] \end{enumerate} \end{manualtheorem} Let us begin by highlighting the main differences between Theorems \ref{thm4} and \ref{13discrete}. In the continuous case, the extreme points are those $\sigma_{\vec{p}}$ achieved by the deterministic single-edge strategy of choosing whichever $u_i$ maximizes the value of the density $f_{\vec{p}}$, or equivalently those $\sigma_{\vec{p}}$ whose density has the same value distribution as $\sigma_{MAX} = Dx^{D-1}dx$ (namely $D \cdot \mbox{Beta}(1,D)$). Contrast this with the discrete case, where the extreme points are those $\sigma_{\vec{p}}$ achieved by the deterministic single-edge strategy of choosing whichever $u_i$ maximizes the value of the probability mass function, or equivalently those $\sigma_{\vec{p}}$ whose pmf has the same value distribution as $\sigma_{MAX} = \frac1{K^D} \delta_1 + \frac{2^D-1}{K^D}\delta_2+ \cdots + \frac{K^D-(K-1)^D}{K^D} \delta_K$. Furthermore, in the discrete case, if we write the extremal $\sigma_{\vec{p}}$ as \[\sigma_{\vec{p}} = \frac1{K^D} \delta_{\alpha(1)} +\frac{2^D-1}{K^D}\delta_{\alpha(2)} + \cdots + \frac{K^D-(K-1)^D}{K^D} \delta_{\alpha(K)} \] for some permutation $\alpha \in S_K$ then clearly the single-edge strategy of choosing whichever $u_i$ maximizes the value of the pmf can be equivalently described as the single-edge strategy of choosing whichever $u_i$ is maximal according to the ordering $\alpha(1) < \alpha(2) < \cdots < \alpha(K)$. It is easy to see that this $\alpha$ satisfies $\vec{p}(\alpha^{-1}(\vec{u})) = \vec{p}^{MAX}(\vec{u})$ (which is where (iv) above comes from). In the discrete setting, the existence of this $\alpha \in S_K$ leads to a natural bijection between the extreme points of the permutohedron and of $\mathcal{R}$, the set of achievable distributions $\{\sigma_{\vec{p}}\}$, by mapping $\sigma_{\vec{p}}$ to the ordering $\alpha(1) < \cdots < \alpha(K)$ associated with it. Since convex combinations of single-edge strategies translate to convex combinations of the distributions $\sigma_{\vec{p}}$ then this bijection extends to a bijection between the permutohedron and $\mathcal{R}$. Of course, the permutohedron is the same for different $D$ but the bijection depends on $D$. It is also worth noting that our proof of Theorem \ref{thm4} adapted to this discrete setting becomes much simpler and more intuitive. Recall we showed that a consistent and deterministic single-edge strategy is one which induces an ordering on the possible values of the edge labels in the sense that $y$ "dominates" $x$ if and only if the choice for $\vec{u}$ is never $x$ if both $x$ and $y$ appear in $\vec{u}$. In the discrete case, there are only $K$ possible values of the edge labels so immediately this gives the ordering $\alpha(1) < \cdots < \alpha(K)$ from which Theorem \ref{13discrete} follows. As an example, let us briefly work through the $D=2, K =4$ discrete case. \begin{center} \captionsetup{type=figure} \includegraphics[scale=0.4]{permutahedron.png} \captionof{figure}{Permutohedron of order 4, \cite{holroyd}} \end{center} The bijection between the extreme points of the permutohedron of order 4 and the extreme points of the set of consistent single-edge strategies goes as follows: \[ 1342 \mapsto \sigma_{1342} = \frac1{16} \delta_1 + \frac{2^2-1}{16} \delta_3 + \frac{3^2-2^2}{16} \delta_4 + \frac{4^2-3^2}{16} \delta_2\] where $\sigma_{1342}$ is the distribution obtained by choosing whichever of the samples $u^1, u^2 \in \{1,2,3,4\}$ is maximal with respect to the ordering $1 < 3 < 4 < 2$. It is plain to see that all the deterministic consistent single-edge strategies correspond to such an ordering, and the corresponding distribution has a pmf with the same value distribution $\{\frac1{16},\frac3{16},\frac5{16},\frac7{16}\}$. \section{Grid Entropy in this Model}\label{grid} In this short section, we compute the grid entropy of the extreme points described in the previous section and we describe grid entropy in general using a simplified formula for the Gibbs Free Energy. \subsection{Grid Entropy of Extreme Points} Recall that the extreme points of $\{\sigma_{\vec{p}}\}$ are given by the single-edge strategy of choosing whichever edge label maximizes the density $f_{\vec{p}}$. We show that such extreme points $\sigma_{\vec{p}}$ have grid entropy 0. However, we first need a short lemma about the partial averages of the expectations of order statistics being bounded away from 0. \begin{lemma}\label{orderStats} Let $Z_i \sim \theta, i \geq 0$ be i.i.d. random variables s.t. $\theta$ is a distribution on $ {\mathbb{R}} $ with cdf $F_{\theta}$ satisfying $F_{\theta}(0) = 0$. Consider the order statistics $Z_{0:n} \leq \cdots \leq Z_{n:n}$. Then there exists a constant $C> 0$ s.t. for large enough $n$, \[ \frac1n \sum_{k=0}^{\lfloor n \epsilon \rfloor} E[Z_{k:n}] \geq C \] \end{lemma} \begin{remark} We allow for the case where some of these expected values may be $\infty$. \end{remark} \begin{proof} Observe that half the terms in this sum are bounded below by $E[Z_{\lfloor \frac{n \epsilon}2 \rfloor:n}]$ hence \[ \frac1n \sum_{k=0}^{\lfloor n \epsilon \rfloor} E[Z_{k:n}] \geq \frac1n \bigg \lfloor \frac{n \epsilon}2 \bigg \rfloor E[Z_{\lfloor \frac{n \epsilon}2 \rfloor:n}] \] Thus it suffices to find $C>0$ s.t. for large $n$, $E[Z_{\lfloor \frac{n \epsilon}2 \rfloor:n}] \geq \frac12 C$. By right-continuity of the cdf $F_{\theta}$, we can take $C > 0$ s.t. $F_{\theta}(C) \leq \frac18 \epsilon$. Consider the i.i.d. Bernoulli random variables $ \textbf{1}_{\{Z_i \leq C\}}$ with success probability $F_{\theta}(C)$. Then Markov's Inequality yields \[P(Z_{\lfloor \frac{n \epsilon}2 \rfloor:n} \leq C) = P \bigg(\#\mbox{successes in Bin($n, F_{\theta}(C)$)} \geq \bigg\lfloor \frac{n \epsilon}2 \bigg \rfloor \bigg) \leq \frac{n F_{\theta}(C)}{\lfloor \frac{n \epsilon}2 \rfloor} \leq \frac12 \] for large $n$ by the choice of $C$. It follows that \[ E[Z_{\lfloor \frac{n \epsilon}2 \rfloor:n}] \geq \frac12 C \] for large $n$, which completes the proof. \end{proof} \begin{theorem}\label{entropy0} Fix $\tau: [0,1] \rightarrow {\mathbb{R}} $ measurable and bounded s.t. $\tau$ is not constant on sets of positive $\Lambda$ measure, and consider the single-edge strategy of picking the maximal $\tau(U^j)$ over $1 \leq j \leq D$, given by \[ \vec{p}(u_1,\ldots,u_D) = \textbf{1}_{\{\tau(u_k) \geq \tau(u_i) \ \forall i\}} \] Then $||\sigma_{\vec{p}}|| = 0$. \end{theorem} \begin{remark} The types of strategies considered in this theorem are deterministic so the resulting $\sigma_{\vec{p}}$ are all extreme points by Theorem \ref{thm4}. On the other hand, Theorem \ref{thm4} establishes that all extreme points can be realized as $\sigma_{\vec{p}}$ for a single-edge strategy $\vec{p}$ choosing whichever observed label maximizes the density $f_{\vec{p}} \in [0,D]$ where $f_{\vec{p}}(\mbox{Unif}[0,1])$ is the $D \cdot \mbox{Beta}(D,1)$ distribution. Thus Theorem \ref{entropy0} captures what happens for all extreme points. \end{remark} \begin{proof} Suppose $||\sigma_{\vec{p}}|| > \delta > 0$. Let $\alpha = \frac{\delta}{\ln 2}$. For $n \in {\mathbb{N} } $ and $ 1 \leq m_n \leq \lfloor e^{n \delta} \rfloor = \lfloor 2^{n \alpha} \rfloor$ consider the event-dependent paths $\pi_{n,m_n}$ corresponding to \[ \min_{\pi: 0\rightarrow n}^{m_n} \rho\bigg(\frac1n \mu_{\pi}, \sigma_{\vec{p}} \bigg) \] Since $\delta < ||\sigma_{\vec{p}}||$ then by definition of grid entropy, \[ \min_{\pi: 0\rightarrow n}^{\lfloor e^{n\delta} \rfloor } \rho\bigg(\frac1n \mu_{\pi}, \sigma_{\vec{p}} \bigg) \rightarrow 0 \ \mbox{a.s.}\] hence $\frac1n \mu_{\pi_{n, m_n}} \Rightarrow \sigma_{\vec{p}}$ a.s. regardless of the sequence $(m_n)$. Also for $i \geq 0$ define the random variables \[Y_i := \max \limits_{1 \leq j \leq D} (\tau(U_i^j)), Y_i' = \mbox{2nd} \ \max \limits_{1 \leq j \leq D} (\tau(U_i^j)), Z_i := Y_i - Y_i'\] Now it is a classic result that for $\epsilon > 0$, \[ \sum_{i=0}^{\lfloor n\epsilon \rfloor} \binom{n}{i} \leq 2^{nL} \ \mbox{with} \ L = L(\epsilon)= \epsilon \log \epsilon + (1-\epsilon) \log (1-\epsilon) \] Take $\epsilon > 0$ small enough so that $L +\epsilon \log(D-1) < \frac12 \alpha$ and take $N \in {\mathbb{N} } $ so that $ \forall n \geq N$, $2^{nL} (D-1)^{n\epsilon} < \lfloor 2^{n \alpha}\rfloor$. Consider any $n \geq N$. Then the number of paths $\pi: 0 \rightarrow n$ with $< \lceil n\epsilon \rceil$ of its edges not having the maximal edge label in their trial is at most \[ \sum_{i=0}^{\lfloor n\epsilon \rfloor} \binom{n}{i} (D-1)^{i} \leq 2^{nL} (D-1)^{n\epsilon} < \lfloor 2^{n \alpha} \rfloor \] By the Pigeonhole Principle, there is an event-dependent path $\pi_{n, m_n}: \vec{0} \rightarrow n$ s.t. at least $\lceil n\epsilon \rceil$ of its edges do not have the maximal edge label in their trial. Let $\pi_{n, m_n}$ have edges $e_0^{j_0}, \ldots, e_{n-1}^{j_{n-1}}$ and let $I_n \subseteq \{0,\ldots,n-1\}, |I_n| = \lceil n\epsilon \rceil$ be a set of indices $i$ for which $\tau(U_i^{j_i}) \neq Y_i$. We compute an upper bound for the passage time along $\pi_{n, m_n}$ (with respect to $\tau$) by splitting the sum over the edges with index in $I_n$ and those not in $I_n$: \begin{align*} \langle \tau, \frac1n \mu_{\pi_{n,m_n}} \rangle &= \frac1n\sum_{i=0}^{n-1} \tau(U_i^{j_i}) \\ & \leq \frac1n \sum_{i\in I_n^C} \max_{1\leq j \leq D} (\tau(U_i^j)) + \frac1n \sum_{i \in I_n} \mbox{2nd} \max_{1\leq j \leq D} (\tau(U_i^j)) \\ &= \frac1n \sum_{i\in I_n^C} Y_i + \frac1n \sum_{i \in I_n} Y_i' \end{align*} On the other hand, the passage time along the $\tau$-optimal path $\pi_{0\rightarrow n}(\vec{p})$ is \begin{equation*} \langle \tau, \frac1n \mu_{\pi_n(\vec{p})} \rangle = \frac1n\sum_{i=0}^{n-1} \max_{1\leq j \leq D} (\tau(U_i^j)) = \frac1n\sum_{i=0}^{n-1} Y_i \end{equation*} hence \begin{equation}\label{lowerB} \langle \tau, \frac1n \mu_{\pi_n(\vec{p})} \rangle - \langle \tau, \frac1n \mu_{\pi_{n,m_n}} \rangle \geq \frac1n \sum_{i \in I_n} (Y_i - Y_i') = \frac1n \sum_{i \in I_n} Z_i \geq \frac1n \sum_{k=0}^{\lfloor n\epsilon \rfloor} Z_{k:n} \end{equation} Now $Z_i$ are i.i.d., non-negative and satisfy \begin{align*} P(Z_i = 0) &= P\bigg(\max \limits_{1 \leq j \leq D} (\tau(U_i^j)) = \mbox{2nd} \ \max \limits_{1 \leq j \leq D} (\tau(U_i^j))\bigg) \\ &\leq P(\exists 1 \leq j_1 < j_2 \leq D \ \mbox{s.t.} \ \tau(U_i^{j_1}) = \tau(U_i^{j_2}) )\\ &=0 \end{align*} since $\tau$ is not constant on sets of positive $\Lambda$ measure. Thus we can take the expectation in \eqref{lowerB} and apply Lemma \ref{orderStats} to get that $\exists C > 0$ s.t. \begin{equation}\label{EQ21} E \bigg[ \langle \tau, \frac1n \mu_{\pi_n(\vec{p})} \rangle -\langle \tau, \frac1n \mu_{\pi_{n,m_n}} \rangle \bigg] \geq \frac1n \sum_{k=0}^{\lfloor n\epsilon \rfloor} E[Z_{k:n}] \geq C \end{equation} for large $n$. Now, by assumption, $\frac1n \mu_{\pi_{n,m_n}} \Rightarrow \sigma_{\vec{p}}$ a.s.. On the other hand, by nature of the model, the $\tau$-optimal length $m$ path $\pi_{m}(\vec{p})$ contains the $\tau$-optimal length $n$ path $\pi_n(\vec{p})$ for any $m \geq n$; thus we can apply the Glivenko-Cantelli Theorem (Theorem \ref{thm1}) to get that the empirical measures $\frac1n \mu_{\pi_{n}(\vec{p})} $ converge weakly to $\sigma_{\vec{p}}$ a.s.. Recall from Section \ref{coupling} that a.s. the pushforward $\tau_{\ast}$ preserves weak limits of empirical measures so \begin{align*} \langle \tau, \frac1n \mu_{\pi_n(\vec{p})} \rangle -\langle \tau, \frac1n \mu_{\pi_{n,m_n}} \rangle &= \langle 1, \tau_{\ast}(\frac1n \mu_{\pi_n(\vec{p})}) \rangle -\langle 1, \tau_{\ast}(\frac1n \mu_{\pi_{n,m_n}}) \rangle \\ &\rightarrow \langle 1, \tau_{\ast}(\sigma_{\vec{p}}) \rangle - \langle 1, \tau_{\ast}(\sigma_{\vec{p}}) \rangle \\ &= 0 \end{align*} a.s.. But $\tau$ is bounded so by the Bounded Convergence Theorem we get \[E \bigg[ \langle \tau, \frac1n \mu_{\pi_n(\vec{p})} \rangle -\langle \tau, \frac1n \mu_{\pi_{n,m_n}} \rangle \bigg] = 0\] which contradicts \eqref{EQ21}. Thus $||\sigma_{\vec{p}}||=0$. \end{proof} \subsection{Grid Entropy via Gibbs Free Energy} Fix $\beta > 0$. Suppose $\tau: [0,1] \rightarrow {\mathbb{R}} $ is a bounded measurable function. From the definition of $\beta$-Gibbs Free Energy, \[ G^{\beta}(\tau) := \lim_{n \rightarrow \infty} \frac1n \log \sum_{\pi: 0 \rightarrow n} e^{\beta T(\pi)} = \lim_{n \rightarrow \infty} \frac1n \sum_{i=0}^{n-1} \log \sum_{j=1}^D e^{\beta \tau(U_i^j)} \ \mbox{a.s.} \] But $\log \sum \limits_{j=1}^D e^{\beta \tau(U_i^j)}$ are i.i.d. in $i$ hence by the SLLN, \[ G^{\beta} (\tau) = E \bigg[\log \sum_{j=1}^D e^{\beta \tau(U^j)} \bigg] \ \mbox{a.s.}\] where $U^j$ are i.i.d. Unif[0,1]. Of course, grid entropy is simply the negative convex conjugate of $\beta$-Gibbs Free Energy by Theorem \ref{gridEntropyPart1}: \begin{equation} -||\nu|| = \sup_{\tau} \bigg[ \beta \langle \tau, \nu \rangle - G^{\beta}(\tau) \bigg] = \sup_{\tau} \bigg[\beta \langle \tau, \nu \rangle - E \bigg[\log \sum_{j=1}^D e^{\beta \tau(U^j)} \bigg] \bigg] \ \forall \nu \in \mathcal{M} \end{equation} where the supremum is over bounded measurable functions $\tau:[0,1] \rightarrow {\mathbb{R}} $ and where $\langle \tau, \nu \rangle$ denotes the integral $ \int_0^1 \tau(u) d\nu$. \section{Next Steps} We have characterized the extreme points of the set of limit points of empirical measures and have shown that these extreme points have grid entropy 0. A natural next question is whether only the extreme points have grid entropy 0. Recalling that grid entropy is concave, another key question to ask is whether it is strictly concave in the model used in this paper; this might provide insights on whether it is strictly concave in general, which would imply that the Gibbs Free Energy is strictly concave, a major open research problem. Furthermore, it would be nice to compute grid entropy of non-extremal points, and also try to describe the subset of $\mathcal{R}$ of maximizers for the variational formula for the Gibbs Free Energy presented in Section 5.2 of \cite{gatea}. These are all questions well worth exploring in the future. \section{Acknowledgments} I would like to thank my friends and family for their continual support during my graduate studies. Special thanks goes to my advisor B\'{a}lint Vir\'{a}g, for his patience, guidance, and optimism in the face of setbacks. Finally, this work would not have been possible without the funding from my NSERC Canadian Graduate Scholarship-Doctoral. \section{Introduction} Random sampling is a fundamental area of study in statistics. Countless processes can be simulated as Monte Carlo experiments, with far-reaching applications in a range of fields from biology and economics to business. A rather simple but ubiquitous and surprisingly non-trivial experiment is that of taking $D$ samples from a distribution and picking out one of these samples according to some omniscient "strategy." This problem is dual to K-means clustering, a well-studied process where one groups a large set of samples into a fixed $K$ number of buckets in order to minimize an error function called the quadratic distortion (see \cite{liu2020convergence} for state-of-the-art). We frame our model in the realm of percolation theory and ask what is the limiting behaviour of empirical measures along the event-dependent path of choices. Answering this question provides a gold mine of information about the model - normalized passage time along a path, after all, is just the identity function integrated against the path's empirical measure. Recent efforts have gone into better understanding the weak convergence of empirical measures. Some, such as \cite{riekert2021wasserstein} and \cite{lei2020convergence}, establish upper bounds on the \mbox{rate} of Wasserstein convergence of empirical measures to the Markov invariant measure/i.i.d. measure respectively they are sampled from. Others focus on describing the limits of empirical measures. In \cite{bates}, Bates shows that the set $\mathcal{R}$ of limit points of normalized empirical measures along paths in a generic lattice model is deterministic and derives an explicit variational formula for the limit shape of first passage time as the minimum value of a linear functional over this deterministic set. In \cite{gatea}, this author builds on Bates' paper by developing the notion of grid entropy, a deterministic quantity capturing not just whether a certain target measure is a limit point of empirical measures. Grid entropy has previously appeared in \cite{rassoul2014quenched}. \cite{gatea} proposes a novel approach, realizing grid entropy both as a subadditive limit of entropies of paths with empirical measure within an $\epsilon$-Levy-Prokhorov distance of the target measure, and as the critical exponent of canonical order statistics associated with the Levy-Prokhorov metric, something a priori only known for i.i.d. Bernoulli$p)$ edge labels (see \cite{carmona2010directed}). Though using different approaches, both this author and Rassoul-Agha et al establish the same convex duality between grid entropy and Gibbs Free Energy. Furthermore, this author observes that Bates' set $\mathcal{R}$ almost surely coincides with the set of probability measures with finite grid entropy: \[ \mathcal{R} = \{\nu \ \mbox{prob meas}: ||\nu|| > -\infty\} \ \mbox{a.s.} \] One limitation of \cite{gatea} is that only empirical measures along deterministic or polymer paths are considered, rather than allowing for arbitrary mixtures of paths. Even in \cite{bates}, the focus is mainly on empirical measures along geodesics. In this paper, we seek to rectify this by studying empirical measures along \emph{random} paths (picked according to some probabilistic "strategy," which may or may not yield geodesics or polymer paths). We will show that $\mathcal{R}$ is still the set of limit points, whether or not we assume the strategy is omniscient. Another deficiency is that grid entropy and the set $\mathcal{R}$, like limit shape and other quantities in this area, are not known to be explicitly computable in most cases. One noteworthy exception is the paper \cite{martinAllan} by Martin, in which an explicit formula for the weak limits of empirical measures along geodesics is derived in the solvable Exponential LPP on $ {\mathbb{Z}} ^2$ model. Also, in \cite{bates} it is established that replacing the lattice $ {\mathbb{Z}} ^D$ by the infinite complete $D$-ary tree $\mathcal{T}_D$, the set $\mathcal{R}$ can be precisely described as a specific sublevel of relative entropy (in terms of $D$). The $D$-ary tree is in a way the dual model to the one we explore in this paper, and as one might expect, we will be able to give a description of $\mathcal{R}$ in our setting too. We will precisely characterize the extreme points of $\mathcal{R}$ as those measures whose density is $D \cdot Beta(1,D)$ distributed; $\mathcal{R}$ is then the closed convex hull of these measures. In addition, we compute the grid entropy of the extreme points of $\mathcal{R}$ to be 0, and we give a simplification of a general formula for grid entropy from \cite{gatea} for this model. \section{Definitions and Results} Let us be more precise about our setup and goals. We consider vertices on $ {\mathbb{Z}} _{\geq 0}$, $D$ parallel edges $(e_i^1,\ldots, e_i^D)$ for every $i \geq 0$, and an i.i.d. array of edge labels $U_i^j \sim $ Unif[0,1]. We denote by $\Lambda$ the Lebesgue measure on $[0,1]$. It is convenient to work on the compact space $\mathcal{M}_1$ of probability measures on [0,1], and as we will explain in Section \ref{coupling} we do not lose any generality by restricting to this setting. \begin{center} \begin{tikzpicture} \node (p1) at (0pt,0pt) {}; \node (p2) at (50pt,0pt) {}; \node (p3) at (100pt,0pt) {}; \filldraw (50pt,0pt)circle(2pt) (0pt,0pt)circle(2pt)(100pt,0pt)circle(2pt) (110pt,0pt)circle(2pt)(120pt,0pt)circle(2pt)(130pt,0pt)circle(2pt); \draw (25pt,0pt) node {$e_0^2$}; \draw (25pt,16pt) node [ above ] {$e_0^1$}; \draw (25pt,-16pt) node [ below ] {$e_0^3$}; \draw (75pt,0pt) node {$e_1^2$}; \draw (75pt,16pt) node [ above ] {$e_1^1$}; \draw (75pt,-16pt) node [ below ] {$e_1^3$}; \draw (0pt,0pt) node [ below right] {$0$}; \draw (50pt,0pt) node [ below right] {$1$}; \draw (100pt,0pt) node [ below right] {$2$}; \draw [->] (p1) to [out=90,in=90] (p2); \draw [->] (p1) to [out=-90,in=-90] (p2); \draw [->] (p2) to [out=90,in=90] (p3); \draw [->] (p2) to [out=-90,in=-90] (p3); \begin{scope}[every path/.style={->}] \draw (p1) -- (p2); \draw (p2) -- (p3); \end{scope} \end{tikzpicture} \end{center} Next, we wish to formalize the notion of strategies. These will be probability measures on the product space coupling the environment $(U_i^j)_{i \geq 0, 1 \leq j \leq D}$ with the infinite target sequence of indices of the $U_i$ corresponding to the choices $(X_i)_{i \geq 0}$. \begin{definition} A strategy is a probability measure $\chi$ on the product space\\ $([0,1]^D)^{ {\mathbb{Z}} _{\geq 0}} \times \{1,\ldots,D\}^{ {\mathbb{Z}} _{\geq 0}}$ s.t. the marginal distribution of the first coordinate (the environment $(U_i^j) \in ([0,1]^D)^{ {\mathbb{Z}} _{\geq 0}}$) is a sequence of i.i.d. Unif[0,1]. We denote by $(J_i)_{i \geq 0}$ the second coordinate (a random sequence of indices) and define the random vector of choices \[ (X_0(\chi), X_1(\chi),\ldots) := (U_0^{J_0}, U_1^{J_1},\ldots) \] Denote by $\frac1n \mu_{0 \rightarrow n}(\chi)$ the empirical measures of this vector: \[ \frac1n \mu_{0 \rightarrow n}(\chi) = \frac1n \sum \limits_{i=0}^{n-1} \delta_{X_i(\chi)} \in \mathcal{M}_1\] Also, let $\sigma_{\chi}^i$ be the law of $X_i(\chi)$. \end{definition} \begin{remark} If $\chi$ conditioned on the environment is a delta mass for a.a. environments $(U_i^j)$ with respect to the product measure $(\Lambda^{\times D})_{\infty}$, it means that the strategy picks exactly one sequence $(x_i)_{i \geq 0}$ for each set of observed labels. In other words, the strategy is deterministic. \end{remark} \begin{definition} We also define (strategy-free) empirical measures along a fixed path $\pi: 0 \rightarrow n$ consisting of edges $(e_0^{j_0},\ldots, e_{n-1}^{j_{n-1}})$ by \[ \frac1n \mu_{\pi} = \frac1n \sum_{i=0}^{n-1} \delta_{U_i^{j_i}} \] \end{definition} An important type of strategy is one which chooses each $X_k$ independent of $(U_i^j)_{i \neq k, 1 \leq j \leq D}$, the observed values from all but the $k$th trial. These are $\chi$ which are infinite products of measures arising from "single-edge strategies," i.e. micro-strategies operating at the individual trial level. We call such $\chi$ "independent strategies," as they give rise to independent $X_i$. \begin{definition} A single-edge strategy is a probability measure $\psi$ on the product space\\ $([0,1]^D) \times \{1,\ldots,D\}$ s.t. the marginal distribution of the first coordinate $(U^j)_{1 \leq j \leq D}$ is a sequence of $D$ i.i.d. Unif[0,1]. We denote by $J$ the second coordinate (a random index) and define the random choice \[ X(\psi):= U^J \] Also, let $\sigma_{\psi}$ be the law of $X(\psi)$. \end{definition} \begin{remark} If $\psi$ conditioned on $(U^j)_{1 \leq j \leq D}$ is a delta mass for a.a. $(u^j)_{1 \leq j \leq D}$ with respect to the product measure $\Lambda^{\times D}$, it means that the single-edge strategy picks exactly one sequence $x$ for each set of $D$ observed labels, so $\psi$ is deterministic. \end{remark} A single-edge strategy $\psi$ is completely determined by the $\psi$-conditional probabilities \\ $p_k: [0,1]^D \rightarrow [0,1]$, $1 \leq k \leq D$ defined as \[ p_k(u_1,\ldots, u_D) := P_{\psi}[J=k \mid (U^j)_{1 \leq j \leq D} = (u^j)_{1 \leq j \leq D} ] \] We expand on this in more detail in Section \ref{vectorFcns}. The key takeaway is that we may interchangeably refer to both $\psi$ and $\vec{p}$ as a single-edge strategy, and therefore we define $X(\vec{p}):=X(\psi), \sigma_{\vec{p}} :=\sigma_{\psi}$. \bigskip In this paper we are interested in the weak limit points of $\frac1n \mu_{0 \rightarrow n}(\chi)$. As it turns out, this set of limit points almost surely coincides with the set of limit points of $\frac1n \mu_{0 \rightarrow n}(\chi)$ over independent strategies $\chi$ only, with the distributions $\sigma_{\vec{p}}$ over single-edge strategies $\vec{p}$, as well as with the set $\mathcal{R}$ of probability measures with finite grid entropy. The following theorem is the main objective of Section \ref{indepReduction} . \begin{manualtheorem}{A}\label{part2_A} A.s. we have \begin{align*} \mathcal{R} &= \bigg\{\mbox{limit pts of} \ \frac1n \mu_{0\rightarrow n}(\chi): \mbox{strategies} \ \chi \bigg\} \\ &=\bigg\{\mbox{limit pts of} \ \frac1n \mu_{0\rightarrow n}(\chi): \mbox{independent strategies} \ \chi \bigg\}\\ &= \bigg\{\sigma_{\vec{p}}: \mbox{single-edge strategies} \ \vec{p} \bigg\} \end{align*} \end{manualtheorem} This reduces the problem to working with single-edge strategies $\vec{p}$, which are simpler to handle than generic strategies. Moreover, every $\sigma_{\vec{p}}$ can be achieved by a "consistent" single-edge strategy $\vec{p}$ with the properties that $p_1(\vec{u})$ is invariant under 1-fixing permutations in $S_D$ and $p_k(\vec{u}) = p_1(u_{D-k+2}, \ldots, u_1,\ldots, u_{D-k+1}) \ \forall k$. Now, the sets from Theorem \ref{part2_A} are convex and weakly compact. In Section \ref{extreme} we fully characterize their extreme points. \begin{manualtheorem}{B}\label{B} Let $\vec{p}$ be a single-edge strategy. The following are equivalent: \begin{enumerate}[label=(\roman*)] \item $\sigma_{\vec{p}}$ is an extreme point \item Any consistent single-edge strategy achieving $\sigma_{\vec{p}}$ must be deterministic. \item $\sigma_{\vec{p}}$ has a density $f_{\vec{p}}$ which is not constant on sets of positive $\Lambda$-measure and $\sigma_{\vec{p}}$ is given by the following single-edge strategy $\vec{q}$: \[ q_k(u_1,\ldots, u_D) = \textbf{1}_{\{f_{\vec{p}}(u_k) \geq f_{\vec{p}}(u_i) \ \forall i\}} \ \mbox{for $\Lambda^{\times D}$-a.a.} \ (u_1,\ldots, u_D) \] In other words, $\sigma_{\vec{p}}$ is achieved by the deterministic greedy single-edge strategy "choose whichever label yields a higher value when evaluating the density $f_{\vec{p}}$." \item $\sigma_{\vec{p}}$ has a density $f_{\vec{p}}$ s.t. $\frac1D f_{\vec{p}}(Unif[0,1]) \sim \mbox{Beta}(1,D)$. \end{enumerate} \end{manualtheorem} As an immediate corollary, $\mathcal{R}$ is the closed convex hull of these measures. \begin{remark} It is important to stress that Theorem \ref{B} holds even in a more general setting where the i.i.d. labels $U_i^j$ follow some finite mean distribution on $\mathcal{R}$ that is not necessarily Unif[0,1]. The only caveat is that the value distribution of the densities $f_{\vec{p}}$ of extreme points might not have as explicit a form as $D \cdot Beta(1,D$); however, the value distribution of the densities $f_{\vec{p}}$ is still identical to the value distribution of $f_{MAX}$, whatever that may be. \end{remark} \begin{remark} In fact, Theorem \ref{B} holds in a discrete setting as well. If the i.i.d. labels $U^j$ are Unif$\{1,\ldots, K\}$ then the convex set of consistent single-edge strategies forms a permutohedron of order $K$. \end{remark} \begin{center} \captionsetup{type=figure} \includegraphics[scale=0.4]{permutahedron.png} \captionof{figure}{Permutohedron of order 4, \cite{holroyd}} \end{center} The following is the discrete analogue of Theorem \ref{B}. \begin{manualtheorem}{B'}\label{B'} Let $\vec{p}$ be a single-edge strategy. The following are equivalent: \begin{enumerate}[label=(\roman*)] \item $\sigma_{\vec{p}}$ is an extreme point \item Any consistent single-edge strategy achieving $\sigma_{\vec{p}}$ must be deterministic. \item $\sigma_{\vec{p}}$ is given by the single-edge strategy "choose whichever label is maximal with respect to the ordering $\alpha(1) < \cdots < \alpha(K)$" for some permutation $\alpha \in S_K$. \item $\sigma_{\vec{p}}$ has a probability mass function whose value distribution is \[\bigg\{\frac1{K^D}, \frac{2^D-1}{K^D}, \frac{3^D-2^D}{K^D}, \ldots, \frac{K^D-(K-1)^D}{K^D} \bigg\}\] \end{enumerate} \end{manualtheorem} The ordering mentioned in (iii) gives the natural bijection between the extreme consistent single-edge strategies and the extreme points of the permutohedron. For example, the extreme point $1342$ corresponds to the single-edge strategy of choosing the maximal $u^i$ according to the ordering $1<3<4<2$, hence the bijection maps \[ 1342 \mapsto \sigma_{1342} = \frac1{4^D} \delta_1 + \frac{2^D-1}{4^D} \delta_3 + \frac{3^D-2^D}{4^D} \delta_4 + \frac{4^D-3^D}{4^D} \delta_2 \] This bijection extends to a bijection between the entire permutohedron and $\mathcal{R}$ by taking convex combinations. We walk through an explicit example of this phenomenon in Section \ref{discrete}. \bigskip Generally speaking there is no known way of computing grid entropy, however, it amazingly can be computed to be 0 for these extreme points. This effectively means that a.s. the number of paths $0 \rightarrow n$ with empirical measures weakly converging to any one of these extreme points is $e^{o(n)}$. \begin{manualtheorem}{C} Let $\sigma_{\vec{p}}$ be an extreme point. Then \[ || \sigma_{\vec{p}} || = 0 \] \end{manualtheorem} Whether \emph{every} measure in $\mathcal{R}$ with grid entropy 0 is an extreme point remains an open question. Section \ref{grid} focuses on this result, as well as simplified formulas for the Gibbs Free Energy and grid entropy in this model. \begin{manualtheorem}{D} Suppose $\tau: [0,1] \rightarrow {\mathbb{R}} $ is a bounded measurable function. Then Gibbs Free Energy with respect to $\tau$ is given by \[ G(\tau) = E \bigg[\log \sum_{j=1}^D e^{\tau(U^j)} \bigg] \ \mbox{a.s.}\] where $U^j$ are i.i.d. Unif[0,1]. For all probability measures $\nu$, grid entropy is given by \[ -||\nu|| = \sup_{\tau} \bigg[ \langle \tau, \nu \rangle - G(\tau) \bigg] = \sup_{\tau} \bigg[ \langle \tau, \nu \rangle - E \bigg[\log \sum_{j=1}^D e^{\tau(U^j)} \bigg] \bigg] \] where the supremum is over bounded measurable functions $\tau:[0,1] \rightarrow {\mathbb{R}} $ and where $\langle \tau, \nu \rangle$ denotes the integral $ \int_0^1 \tau(u) d\nu$. \end{manualtheorem} \begin{remark} This theorem actually holds for measurable $\tau$ s.t. $E[e^{\beta \tau(U)}] < \infty$ for $U \sim Unif[0,1]$ and can easily be extended to all measurable $\tau$ by truncating $\tau$ in $\langle \tau, \nu \rangle$ at some $C>0$ and taking a supremum over $C$ of the variational formula. We leave the details to the reader and focus on the bounded $\tau$ case. \end{remark} But first we delve deeper into the setup and known relevant results. \section{Preliminaries} \subsection{More on Strategies} We list some miscellaneous observations about strategies. It is trivial to see that the sets of strategies $\chi$, of single-edge strategies $\psi$, of distributions $\sigma_{\chi}^i$ for strategies $\chi$ and $i \geq 0$, and of $\sigma_{\psi}$ for single-edge strategies $\psi$ are each closed under convex combinations. The extreme points of the sets of strategies/single-edge strategies are clearly the sets of deterministic strategies/single-edge strategies respectively. Moreover, if $\chi$ is an independent strategy corresponding to a sequence $(\psi_i)$ of single-edge strategies then $\sigma_{\chi}^i = \sigma_{\psi_i} \ \forall i \geq 0$. In particular, this implies \begin{equation} \label{EQ2} \{\sigma_{\chi}^i: \mbox{independent strategies} \ \chi\} = \{\sigma_{\psi}: \mbox{single-edge strategies} \ \psi\} \ \forall i \geq 0 \end{equation} In fact, observe that for any strategy $\chi$ and $i \geq 0$ if we define the strategy $\chi'$ to be the measure determined on product sets by \[ \chi'(A \times B) := \chi( (0 \ i) A \times (0 \ i) B) \ \forall \ \mbox{measurable} \ A \subseteq ([0,1]^D)^{ {\mathbb{Z}} _{\geq 0}}, B \subseteq (\{1,\ldots,D\})^{ {\mathbb{Z}} _{\geq 0}}\] where $(0 \ i) C$ swaps the 0th and $i$th coordinates of sequences in $C$, then this is easily checked to be a strategy with \[ \sigma_{\chi}^i = \sigma_{\chi'}^0 \] Therefore \begin{equation} \label{EQswap} \{\sigma_{\chi}^i: \mbox{strategies} \ \chi\} = \{\sigma_{\chi}^0: \mbox{strategies} \ \chi\} \ \forall i \geq 0 \end{equation} \subsection{Coupling the Edge Weights with I.I.D. Uniforms}\label{coupling} We briefly explain why we can work in the setting of edge labels $U_i^j \sim Unif[0,1]$ without losing generality. Suppose instead we start with i.i.d. edge weights $(\tau_i^j)_{i \geq 0, 1 \leq j \leq D} \sim \theta$ for some distribution $\theta$ on $ {\mathbb{R}} $ with finite mean. We can then couple the environment to uniform random variables as is done in \cite[Sect.~2.1]{bates}. More specifically, we let our the $\tau_i^j$ be given by \[ \tau_i^j = \tau(U_i^j) \] for a measurable function $\tau: [0,1] \rightarrow {\mathbb{R}} $ and i.i.d. $U_i^j \sim Unif[0,1]$. If we let $F_{\theta}$ be the cumulative distribution function of $\theta$ then the quantile function \[ F_{\theta}^{-}(t) := \inf \{t \in {\mathbb{R}} : F_{\theta}(t) \geq x\} \] is an example of such a $\tau$; however, our results hold for any such $\tau$ chosen so we let $\tau$ be arbitrary. We can move from the simplified $(\Lambda, [0,1])$ setting we wish to work in to the initial $(\theta, {\mathbb{R}} )$ setting by applying the $\tau$-pushforward. Now, as mentioned in Theorem \ref{part2_A}, we show that, in the $(\Lambda, [0,1])$ setting, the set of limit points of the empirical measures $\frac1n \mu_{0\rightarrow n}(\chi)$ a.s. coincides with $\mathcal{R}$, which is shown in \ref{gridEntropyPart1} to a.s. coincide with the set of limit points of (strategy-free) empirical measures $\frac1n \mu_{\pi_n}$ along paths $\pi_n:0 \rightarrow n$. We do not lose generality by performing this coupling because a.s., $\frac1{n_k} \mu_{\pi_{n_k}} \Rightarrow \nu$ implies $\tau_{\ast}(\frac1{n_k} \mu_{\pi_{n_k}}) \Rightarrow \tau_{\ast}(\nu)$. This is proved in \cite[Lemma 6.15]{bates}. In fact, if we take $\tau $ to be the quantile function and we assume $\theta$ has continuous cdf $F_{\theta}$ then this implication becomes an if and only if, as seen in \cite[Lemma 5]{gatea}. \subsection{Grid Entropy} \begin{definition}The Levy-Prokhorov metric on the space $\mathcal{M}_+$ of finite non-negative Borel measures on $[0,1]$ is defined by \[\rho(\mu, \nu) = \inf \{\epsilon > 0: \mu(A) \leq \nu(A^{\epsilon}) + \epsilon \ \mbox{and} \ \nu(A) \leq \mu(A^{\epsilon}) + \epsilon \ \forall A \in \mathcal{B}([0,1])\}\] \end{definition} It is standard that the Levy-Prokhorov metric $\rho$ metrizes weak convergence. Some elementary properties include that the Levy-Prokhorov metric is upper bounded by total variation and it satisfies a certain subadditivity: \[ \rho(\mu_1 + \mu_2, \nu_1 + \nu_2) \leq \rho(\mu_1, \nu_1) + \rho(\mu_2, \nu_2) \] For $t \geq 0$ let $\mathcal{M}_t$ denote the space of non-negative Borel measures on $[0,1]$ with total mass $t$. In a concurrent paper \cite{gatea}, this author studies not just the convergence of empirical measures in a lattice model such as ours but the \emph{entropy} of empirical measures converging to a certain weak limit. Three quite different but equivalent definitions of grid entropy are given in said paper, and they are linked to the original description of this entropy provided in a 2014 paper \cite{rassoul2014quenched} by Rassoul-Agha and Sepp{\"a}l{\"a}inen. Here we work with the two most relevant to our needs, coming from \cite{gatea}. Note that in the setting of \cite{gatea}, both direction-fixed and direction-free grid entropy are considered, but in the model we focus on in this paper there is but one unit direction so we will use the direction-free versions of the results from \cite{gatea}. We go into a brief description of this work, as fits our needs in this paper. Fix any $t \geq 0$ and a target measure $\nu$. We consider the order statistics of the Levy-Prokhorov distance between $\nu$ and the empirical measures $\frac1n \mu_{\pi}$ varying over all $D^{\lfloor nt\rfloor}$ possible origin-anchored, length $\lfloor nt \rfloor$ paths $\pi$. That is, for every $n \in {\mathbb{N} } $ we let \[ \min_{\pi: 0\rightarrow \lfloor nt \rfloor}^1 \rho\bigg(\frac1n \mu_{\pi}, \nu \bigg) \leq \min_{\pi: 0\rightarrow \lfloor nt \rfloor}^2 \rho\bigg(\frac1n \mu_{\pi}, \nu \bigg) \leq \ldots \leq \min_{\pi: 0\rightarrow \lfloor nt \rfloor}^{D^{\lfloor nt \rfloor}} \rho\bigg(\frac1n \mu_{\pi}, \nu \bigg) \] denote the order statistics value of $ \rho(\frac1n \mu_{\pi}, \nu)$. It is convenient to define \[ \min_{\pi: 0\rightarrow \lfloor nt \rfloor}^j \rho\bigg(\frac1n \mu_{\pi}, \nu \bigg) := +\infty \ \mbox{for} \ j > D^{\lfloor nt \rfloor}\] These order statistics and the paths corresponding to them are of course event-dependent. However, the following theorem from \cite{gatea} shows there is a deterministic critical exponent where the empirical measures along the paths corresponding to these order statistics change from converging a.s. to $\nu$ to a.s. diverging away from $\nu$. This critical exponent turns out to be precisely the negative convex conjugate of Gibbs Free Energy. We state the full theorem as presented in \cite{gatea}, noting that part (i) has been previously shown in \cite{carmona2010directed} to hold in the case of Bernoulli edge labels, and that parts (ii) and (iv) follow immediately from the earlier work of Rassoul-Agha and Sepp{\"a}l{\"a}inen \cite{rassoul2014quenched}. \begin{theorem}\label{gridEntropyPart1} \begin{enumerate}[label=(\roman*)] \item[] \item For any target measure $\nu \in \mathcal{M}$, its grid entropy is defined to be the deterministic quantity \begin{align*} || \nu|| &:= \sup \bigg\{\alpha \geq 0: \lim_{n \rightarrow \infty} \min_{\pi: 0 \rightarrow \lfloor nt \rfloor}^{\lfloor e^{\alpha n} \rfloor} \rho\bigg(\frac1{n} \mu_{\pi}, \nu \bigg) = 0 \ \mbox{a.s.} \bigg\} \\ &\in \{-\infty\} \cup [0, t\log D] \end{align*} where the value is $-\infty$ if the set of $\alpha$'s is empty. Then grid entropy is the critical exponent where the $\min \limits_{\pi: 0\rightarrow \lfloor nt \rfloor}^{\lfloor e^{\alpha n} \rfloor} \rho(\frac1n \mu_{\pi}, \nu )$ change from converging to 0 to a.s. having a $\liminf \limits_{n \rightarrow \infty} >0$.\\ For example, the grid entropy of the original distribution, $\Lambda$, is maximal:\\ $||\Lambda|| = t \log D$. \item Grid entropy is positive-homogeneous, satisfies the reverse-triangle inequality \[ ||\nu|| + ||\xi|| \leq || \nu + \xi||\] and is concave and upper-semicontinuous in $\nu$. \item Consider the deterministic, weakly closed set \[ \mathcal{R}^t:= \{\nu \in \mathcal{M}: ||\nu|| \geq 0 \}\] Then $\mathcal{R}^t \subseteq \mathcal{M}_t$, $\nu \ll \Lambda \ \forall \nu \in \mathcal{R}^t$ and \[ \mathcal{R}^t = \bigg\{\mbox{limit pts of} \ \frac1n \mu_{\pi} \ \mbox{for} \ \pi: 0\rightarrow \lfloor nt \rfloor \bigg\} \ \mbox{a.s.} \] \item Grid entropy is the negative convex conjugate of $\beta$-Gibbs Free Energy. More concretely, for $\beta > 0$, \begin{equation}\label{dualFormula} -||\nu|| = \sup_{\tau} \bigg[ \beta \langle \tau, \nu \rangle - G_t^{\beta}(\tau) \bigg] \ \forall \nu \in \mathcal{M}_t \end{equation} where the supremum is taken over bounded measurable $\tau: [0,1] \rightarrow {\mathbb{R}} $, where \\ $\langle \tau, \nu \rangle = \int_0^1 \tau(u) d\nu$, where $G_t^{\beta}(\tau)$ is the length $t$ $\beta$-Gibbs Free Energy \[ G_t^{\beta}(\tau) := \lim_{n \rightarrow \infty} \frac1n \log \sum_{\pi: 0 \rightarrow \lfloor nt \rfloor} e^{\beta T(\pi)} \] and where $T(\pi) = \sum \limits_{e\in \pi} \tau(U_e) = \langle \tau, \mu_{\pi} \rangle$ is the passage time along $\pi$. \item Any $\nu \in \mathcal{R}^t$ satisfies the following upper bound on the sum of the grid entropy and the relative entropy with respect to Lebesgue measure on [0,1]: $$t D_{KL} (\nu||\Lambda) + ||\nu|| \leq t \log D$$ where $D_{KL}$ denotes relative entropy (or Kullback-Leibler divergence). \end{enumerate} \end{theorem} \begin{remark} An immediate but highly non-trivial consequence of (i) is the existence of random length $\lfloor nt \rfloor$ paths whose empirical measures converge weakly to a given target $\nu \in \mathcal{R}^t$ a.s.. Indeed, the event-dependent paths corresponding to $\min \limits_{\pi: 0 \rightarrow \lfloor nt \rfloor}^1 \rho(\frac1n \mu_{\pi}, \nu)$ do the job. \end{remark} \begin{remark} (i), seen as the statement that the negative convex conjugate of Gibbs Free Energy is the critical exponent of the order statistics $\min_{\pi: 0\rightarrow \lfloor nt \rfloor}^j \rho(\frac1n \mu_{\pi}, \nu)$, has been previously proved in the case of Bernoulli($p$) edge labels in \cite{carmona2010directed} [Corollary 2]. \end{remark} By positive-homogeneity it suffices to work with the case $t=1$. For the sake of notation we drop the 1 in $\mathcal{R}^1, G_1^{\beta}(\tau)$ for the rest of the paper. Once we determine the extreme points of $\mathcal{R}$, we compute their grid entropies to be 0. We also give a limit-free formula for the $\beta$-Gibbs Free Energy $G^{\beta}(\tau)$ in our model, which renders the convex duality formula \eqref{dualFormula} more practical. \subsection{Measure-Preserving Bijections in \texorpdfstring{$ {\mathbb{R}} ^n$}{Rn}} We will encounter measure-preserving bijections in $ {\mathbb{R}} $, so we briefly outline the required notions in this section. \begin{definition} Let $\mathcal{B}( {\mathbb{R}} ^n)$ denote the Borel $\sigma$-algebra of $ {\mathbb{R}} ^n$ and let $\mu$ be a Borel probability measure on $ {\mathbb{R}} ^n$. Let $A, A' \in \mathcal{B}( {\mathbb{R}} ^n)$ with $\mu(A') = \mu(A) > 0$. A bijection $\phi: A \rightarrow A'$ is said to be $\mu$-measure-preserving if $\mu(B) = \mu(\phi(B)) $ for all $B \in \mathcal{B}( {\mathbb{R}} ^n), B \subseteq A$. \end{definition} Nishiura proves in \cite[Theorem 4]{nishiura} the existence of a $\mu$-measure-preserving bijection between Borel sets of equal, positive but not full $\mu$-measure, given that $\mu$ is nonatomic. We will only need the following simplified version of this theorem. \begin{theorem}\label{thmNishiura} Suppose $\theta$ is a Borel probability measure on $[0,1]$ that is absolutely continuous with respect to Lebesgue measure $\Lambda$ on $[0,1]$. Let $A, A' \in \mathcal{B}([0,1])$ s.t. $\theta(A) = \theta(A') \in (0,1)$. Then there exists a $\theta$-measure-preserving bijection $\phi: A \rightarrow A'$. \end{theorem} Preliminaries finished, our next goal is to show that we may restrict ourselves to working with single-edge strategies without losing generality. \section{Reducing the Problem}\label{indepReduction} Over the course of this section, we prove the following theorem, which allows us to reduce the problem of describing the set of limit points of the empirical measures $\frac1n \mu_{0\rightarrow n}(\chi)$ to characterizing the set of achievable distributions $\sigma_{\psi}$ for single-edge strategies $\psi$. \begin{theorem}\label{limitPts} A.s., the following sets are equal: \begin{enumerate}[label=(\roman*)] \item $\{\mbox{limit pts of} \ \frac1n \mu_{0\rightarrow n}(\chi): \mbox{ strategies} \ \chi\}$ \item $\mathcal{R} := \{\nu \in \mathcal{M}_1: ||\nu|| \geq 0\}$ \item $\{\sigma_{\chi}^0: \mbox{ strategies} \ \chi\}$ \item $\{\sigma_{\psi}: \mbox{single-edge strategies} \ \psi\}$ \item $\{\mbox{limit pts of} \ \frac1n \mu_{0\rightarrow n}(\chi): \mbox{independent strategies} \ \chi\}$ \item $\{\mbox{limit pts of} \ \frac1n \mu_{0\rightarrow n}(\chi): \mbox{i.i.d. strategies} \ \chi\}$ \end{enumerate} \end{theorem} \begin{remark} A priori it not clear that (iii)-(iv) are weakly closed. We show this as part of our proof. \end{remark} We argue via a chain of inclusions. We begin with the most trivial of these. \begin{lemma}\label{1st} A.s. we have \[ \bigg\{\mbox{limit pts of} \ \frac1n \mu_{0\rightarrow n}(\chi): \mbox{ strategies} \ \chi \bigg\} \subseteq \mathcal{R} \] \end{lemma} \begin{proof} Fix a strategy $\chi$. Suppose $\frac1{n_k} \mu_{0 \rightarrow n_k}(\chi) \Rightarrow \nu$ for some observed edge labels \\ $(u_i^j)_{i \geq 0, 1 \leq j \leq D}$. Then letting $\pi_{n}: 0 \rightarrow n$ be the paths corresponding to the indices \\ $(J_0(\chi), \ldots, J_{n-1}(\chi))$ conditioned on these same observed edge labels, we get\\ $ \frac1{n_k} \mu_{\pi_{n_k}}=\frac1{n_k} \mu_{0 \rightarrow n_k}(\chi) \Rightarrow \nu$. Thus $\nu$ is a limit point of the $\frac1n \mu_{\pi}$. The desired inclusion follows from \[ \mathcal{R} = \bigg\{\mbox{limit pts of} \ \frac1n \mu_{\pi}: \mbox{ paths} \ \pi:0 \rightarrow n \bigg\} \ \mbox{a.s.} \ \mbox{by Theorem \ref{gridEntropyPart1}(iii)}\] \end{proof} \subsection{Expected Value of Empirical Measures}\label{sect31} For any strategy $\chi$, \[E[\delta_{X_i(\chi)}] = \int_{[0,1]} \delta_x d\sigma_{\chi}^i(x) = \sigma_{\chi}^i \ \mbox{and} \ E \bigg[\frac1n \mu_{0\rightarrow n}(\chi)\bigg] = \frac1n \sum_{i=0}^{n-1} \sigma^i_{\chi} \] Recalling \eqref{EQswap} and the fact that $\{\sigma^0_{\chi}: \mbox{strategies} \ \chi\}$ is closed under convex combinations, it follows that every $E [\frac1n \mu_{0\rightarrow n}(\chi)]$ is contained in $\{\sigma^0_{\chi}: \mbox{strategies} \ \chi\}$. Therefore \begin{align*} \bigg\{E \bigg[\frac1n \mu_{0\rightarrow n}(\chi)\bigg]: \mbox{ strategies} \ \chi \bigg\} &= \{E[\delta_{X_0(\chi)}]: \mbox{ strategies} \ \chi\} \\ &= \{\sigma_{\chi}^0: \mbox{ strategies} \ \chi \} \end{align*} Also it is clear that \[ \{E[\delta_{X(\psi)}]: \mbox{single-edge strategies} \ \psi\} = \{\sigma_{\psi}: \mbox{ single-edge strategies} \ \psi \}\] A simple Tonelli argument establishes that the closures of the two sets above coincide. \begin{lemma} \label{equivalenceIndep} \[ \mbox{cl}\{\sigma_{\psi}: \mbox{ single-edge strategies} \ \psi \} = \mbox{cl}\{\sigma^0_{\chi}: \mbox{ strategies} \ \chi \} \] \end{lemma} \begin{proof} By \eqref{EQ2} $\{\sigma_{\psi}\} = \{\sigma^0_{\chi}: \mbox{indep strategies} \ \chi\}$ so it suffices to show \[ \mbox{cl}\{\sigma_{\psi}: \mbox{ single-edge strategies} \ \psi \} \supseteq \{\sigma^0_{\chi}: \mbox{ strategies} \ \chi \} \] Given a strategy $\chi$ and observed "future" edge labels $(u_i^j)_{i \geq 1, 1 \leq j \leq D} \in [0,1]^{ {\mathbb{Z}} _{\geq 1}}$, define $\psi_{(u_i^j)_{i \geq 1, 1 \leq j \leq D}}$ to be $\chi$ conditioned on the rest of the environment $(u_i^j)_{i \geq 1, 1 \leq j \leq D}$. The $\psi_{(u_i^j)_{i \geq 1, 1 \leq j \leq D}}$ are easily seen to be single-edge strategies. Since integrating over the entire environment $(U_i^j)_{i \geq 0, 1 \leq j \leq D}$ is equivalent to integrating over $(U_0^j)_{1 \leq j \leq D}$ first and then over $(U_i^j)_{i \geq 1, 1 \leq j \leq D}$, we have \[ \sigma_{\chi}^0 = \int_{([0,1]^D)^{ {\mathbb{Z}} _{\geq 1}}} \sigma_{\psi_{(u_i^j)_{i \geq 1, 1 \leq j \leq D}}} d\Lambda^{\infty}((u_i^j)_{i \geq 1, 1 \leq j \leq D}) \] By convexity and weak closure it follows that \[\sigma_{\chi}^0 \in \mbox{cl}\{\sigma_{\psi}: \mbox{ single-edge strategies} \ \psi \} \] \end{proof} Next, we show that this closure of $\{\sigma_{\chi}^0: \mbox{strategies} \ \chi\}$ contains all probability measures with finite grid entropy. \begin{lemma} \label{3rd} We have \[\mathcal{R}:= \{\nu \in \mathcal{M}_1: ||\nu|| \geq 0\} \subseteq \mbox{cl}\{E[\delta_{X_0(\chi)}]: \mbox{ strategies} \ \chi \} \] as deterministic sets. \end{lemma} \begin{proof} Suppose this is not the case, say $\exists \nu \in \mathcal{R} \cap (\mbox{cl}\{E[\delta_{X_0(\chi)}]: \mbox{ strategies} \ \chi \})^C$. Thus there exists $\epsilon > 0$ s.t. $B_{\epsilon}(\nu) \cap cl\{E[\delta_{X_0(\chi)}]: \chi\} = \emptyset$. Let $\pi_n: 0 \rightarrow n$ be the environment-dependent path corresponding to $\min \limits_{\pi: 0 \rightarrow n}^1 \rho(\frac1n \mu_{\pi}, \nu)$. It is crucial to note that $\pi_n$ depends on the observed edge labels $((U_i^j)_{0 \leq i < n, 1\leq j \leq D})$ of the first $n$ trials only. By definition of grid entropy, $\frac1n \mu_{\pi_n} \Rightarrow \nu$ a.s.. In particular, we get convergence in probability: \[ \lim_{n \rightarrow \infty} P \bigg( \rho \bigg(\frac1n \mu_{\pi_n} , \nu \bigg) \geq \frac{\epsilon}2 \bigg) = 0\] Thus $\exists n \in {\mathbb{N} } $ s.t. \[ P \bigg( \rho \bigg(\frac1n \mu_{\pi_n} , \nu \bigg) \geq \frac{\epsilon}2 \bigg) < \frac{\epsilon}4 \] We claim that $\rho(E[\frac1n \mu_{\pi_n}], \nu)<\epsilon$. Let $\mathcal{E}$ be the event $\{\rho (\frac1n \mu_{\pi_n} , \nu ) \geq \frac{\epsilon}2\}$ so $P(\mathcal{E})< \frac{\epsilon}4$. We split the expectation: \[E \bigg[ \frac1n \mu_{\pi_n} \bigg] = \int_{\mathcal{E}} \frac1n \mu_{\pi_n} dP + \int_{\mathcal{E}^C} \frac1n \mu_{\pi_n} dP \] Since the Levy-Prokhorov metric is upper bounded by the total variation, then \[\rho \bigg(\int_{\mathcal{E}} \frac1n \mu_{\pi_n} dP, \nu \cdot P(\mathcal{E}) \bigg) \leq \bigg( \bigg| \bigg |\frac1n\mu_{\pi_n} \bigg|\bigg|_{TV} + ||\nu||_{TV} \bigg)P(\mathcal{E}) = 2P(\mathcal{E}) < \frac{\epsilon}2 \] On $\mathcal{E}^C$, we have $\rho (\frac1n \mu_{\pi_n} , \nu ) < \frac{\epsilon}2$. By definition of $\rho$, for all measurable $A \in \mathcal{B}([0,1])$, \begin{align*} \bigg(\int_{\mathcal{E}^C} \frac1n \mu_{\pi_n} dP\bigg) (A) &= \int_{\mathcal{E}^C} \frac1n \mu_{\pi_n} (A) dP \\ & \leq \int_{\mathcal{E}^C} \nu(A^{\frac{\epsilon}2} )+ \frac{\epsilon}2 dP \leq \nu(A^{\frac{\epsilon}2} ) P(\mathcal{E}^C) +\frac{\epsilon}2 \end{align*} and similarly \begin{align*} \nu(A) P(\mathcal{E}^C) &= \int_{\mathcal{E}^C} \nu(A) dP \\ &\leq \int_{G^C} \frac1n \mu_{\pi_n} (A^{\frac{\epsilon}2}) + \frac{\epsilon}2 dP \leq \bigg( \int_{\mathcal{E}^C} \frac1n \mu_{\pi_n} dP\bigg) (A^{\frac{\epsilon}2}) + \frac{\epsilon}2 \end{align*} hence \[ \rho \bigg(\int_{\mathcal{E}^C} \frac1n \mu_{\pi_n} dP, \nu P(\mathcal{E}^C) \bigg) \leq \frac{\epsilon}2 \] By subadditivity of $\rho$, \begin{align*} \rho \bigg( E \bigg[\frac1n \mu_{\pi_n} \bigg], \nu \bigg) &= \rho \bigg( \int_{\mathcal{E}} \frac1n \mu_{\pi_n} dP + \int_{\mathcal{E}^C} \frac1n \mu_{\pi_n} dP, \nu P(\mathcal{E}) + \nu P(\mathcal{E}^C) \bigg) \\ &\leq \rho \bigg( \int_{\mathcal{E}} \frac1n \mu_{\pi_n} dP, \nu P(\mathcal{E}) \bigg) + \rho\bigg(\int_{\mathcal{E}^C} \frac1n \mu_{\pi_n} dP, \nu P(\mathcal{E}^C) \bigg) \\ & < \epsilon \end{align*} It remains to construct a strategy $\chi$ for which $\frac1n \mu_{\pi_n} = \frac1n \mu_{0 \rightarrow n} (\chi)$ for this fixed $n$. But this is trivial since conditioned on the observed edge labels $(U_i^j)_{0 \leq i \leq n-1, 1 \leq j \leq D}$ from the first $n$ trials, $\pi_n$ is a deterministic path $0 \rightarrow n$ (namely the one which minimizes $\rho(\frac1n \mu_{\pi}, \nu)$). To be concrete, consider the product measure $\chi'$ on $([0,1]^D)^{ {\mathbb{Z}} _{\geq n}} \times \{1,\ldots,D\}^{ {\mathbb{Z}} _{\geq n}}$ given by \[ \chi' = \Lambda^{\infty} \times \delta_{(1,1,\ldots)} \] Heuristically, $\chi'$ is a partial strategy always picking the "top" edge label $U_i^1$ for $i \geq n$. Also let $\chi''$ be the measure on $([0,1]^D)^{ {\mathbb{Z}} _{0 \leq i < n}} \times \{1,\ldots,D\}^{ {\mathbb{Z}} _{0 \leq i < n}}$ determined by \begin{align*} &\chi''(A \times \{(j_0,\ldots,j_{n-1}) \}) \\ &= \Lambda^{\infty}(\{(u_i^j)_{0 \leq i < n, 1 \leq j \leq D} \in A: \pi_n((u_i^j)_{0 \leq i < n, 1 \leq j \leq D}) = (j_0,\ldots, j_{n-1})\}) \end{align*} $\forall A \in \mathcal{B}(([0,1]^D)^{ {\mathbb{Z}} _{0 \leq i < n}}), (j_i)_{0 \leq i < n} \in \{1,\ldots,D\}^{ {\mathbb{Z}} _{0 \leq i < n}}$ where $\pi_n((u_i^j)_{0 \leq i < n, 1 \leq j \leq D})$ denotes the sequence of $n$ indices $(j_0,\ldots, j_{n-1})$ corresponding to the path $\pi_n$ when the first $n$ trials yield observed labels $(u_i^j)_{0 \leq i < n, 1 \leq j \leq D}$. That is, conditioned on the observed values from the first $n$ trials, $\chi''$ always picks the path corresponding to $\pi_n$. \bigskip Now consider the strategy $\chi = \chi'' \times \chi'$. $\mu_{0 \rightarrow n}(\chi) = \mu_{\pi_n}$ hence $E[\frac1n \mu_{\pi_n}] = E[\frac1n \mu_{0\rightarrow n}(\chi)]$. This contradicts \[B_{\epsilon}(\nu) \cap cl \bigg\{E \bigg[\frac1n \mu_{0 \rightarrow n}(\chi) \bigg]: \chi \bigg\} = B_{\epsilon}(\nu) \cap cl\{E[\delta_{X_0(\chi)}]: \chi\} = \emptyset\] \end{proof} Combining Lemmas \ref{1st}-\ref{3rd}, we get that a.s., (vi) $\subseteq$ (v) $\subseteq$ (i) $\subseteq$ (ii) $\subseteq $ cl(iii) = cl(iv) in Theorem \ref{limitPts}. To complete the proof we only need to show the last inclusion cl (iv) $\subseteq$ (vi) and the fact that (iv) is weakly closed. \subsection{What Happens with Independent Strategies} Let us now focus on the case of independent strategies. Recall that these look like $\chi = \bigtimes \limits_{i=0}^{\infty} \psi_i$ where $(\psi_i)_{i \geq 0}$ are single-edge strategies. First, consider the even simpler case of i.i.d. strategies $\chi$, where all the $\psi_i$ are identical, resulting in i.i.d. $X_i$. Then the Glivenko-Cantelli Theorem \cite[Thm.~2.4.7]{durrett} gives that a.s. the empirical measures $\frac1n \mu_{0 \rightarrow n}(\chi)$ converge weakly to the common law of the $X_i$'s, $\sigma_{\psi_0}$. \begin{theorem}[Glivenko-Cantelli Theorem]\label{thm1} Let $F_{\gamma}$ be the cumulative distribution function of $\gamma$, let $Y_i \sim \gamma$ be i.i.d. random variables and let \[F_{n}(y) = \frac1{n} \sum_{i=1}^n \mathbf{1}_{\{Y_i \leq y\}} \] be the cumulative distribution functions of the empirical measures. Then \[\sup_y |F_n(y) - F_{\gamma}(y)| \rightarrow 0 \ \mbox{a.s. as} \ n \rightarrow \infty \] \end{theorem} As an immediate corollary we get the outstanding inclusion (vi) $\supseteq $ cl (iv) mentioned at the end of Section \ref{sect31}. \begin{corollary} \end{corollary} A.s. we have \[ \bigg\{\mbox{limit pts of} \ \frac1n \mu_{0 \rightarrow n}(\chi): \ \mbox{i.i.d. strategies $\chi$}\bigg\} \supseteq cl\{\sigma_{\psi}: \mbox{single-edge strategies} \ \psi\} \] \begin{proof} Fix a dense subset $\mathcal{O}$ of the deterministic, weakly compact set cl (iv) s.t. $\mathcal{O} \subseteq \mbox{(iv).}$ For every $\sigma_{\psi} \in \mathcal{O}$, apply Theorem \ref{thm1} to the i.i.d. strategy $\chi = \bigtimes \limits_{i=0}^{\infty} \psi$ to get $\frac1n \mu_{0\rightarrow n}(\chi) \Rightarrow \sigma_{\psi}$ a.s.. Therefore $\mathcal{O} \subseteq$ (vi) a.s.. The inclusion follows since (vi) is weakly closed.\\ \end{proof} Thus we have shown that the sets (i), (ii), cl(iii), cl(iv), (v), (vi) in Theorem \ref{limitPts} are equal a.s.. Before proceeding to look further into single-edge strategies, it is worth mentioning a version of the Glivenko-Cantelli Theorem for independent but not necessarily i.d. sequences from \cite{wellner}. It gives further insight into the limit points of the empirical measures $\frac1n \mu_{0\rightarrow n}(\chi)$ for independent strategies $\chi$. \begin{theorem}\label{generalGlivenko} Let $Y_i$ be a sequence of independent random variables with distributions $\sigma_i$. Define $\overline{\sigma}_n = \frac1n (\sigma_0+\ldots + \sigma_{n-1})$ to be the averages of these distributions and let \\ $\frac1n \mu_n = \frac1n (\delta_{Y_0}+\ldots+\delta_{Y_{n-1}})$ be the empirical measures. If $\{\overline{\sigma}_n\}$ is tight then $\rho(\overline{\sigma}_n, \frac1n \mu_n) \rightarrow 0$ a.s.. \end{theorem} In our case, $\mathcal{M}_1$ is weakly compact so any sequence of probability measures in $\mathcal{M}_1$ is tight. Thus, for independent strategies $\chi = \bigtimes \limits_{i=0}^{\infty} \psi_i$ with $\psi_i$ single-edge strategies, we have \\ $\rho(\overline{\sigma_n}, \frac1n \mu_{0 \rightarrow n}(\chi)) \rightarrow 0 \ \mbox{a.s.}$ where $\overline{\sigma_n}$ denote the averages of the distributions of $X_i(\chi)$: \[\overline{\sigma}_n = \frac1n (\sigma^0_{\chi} + \ldots +\sigma^{n-1}_{\chi}) = \frac1n (\sigma_{\psi_0} + \ldots +\sigma_{\psi_{n-1}})= \sigma_{\frac1n(\psi_0+\ldots + \psi_{n-1})} \] This not only confirms that cl(iv) and (v) are equal, but it tells us that a.s. the empirical measures $\frac1n \mu_{0 \rightarrow n}(\chi)$ for independent strategies $\chi$ have the exact same limit points as the distributions $\sigma_{\frac1n(\psi_0+\ldots + \psi_{n-1})}$ corresponding to the law of $X$ chosen according to the average of the single-edge strategies $\psi_i$. \section{Single-Edge Strategies Revisited} \subsection{Single-Edge Strategies in Terms of Conditional Probabilities}\label{vectorFcns} From what we have shown thus far, we only need to focus on single-edge strategies, as the set of limit points of empirical measures $\frac1n \mu_{0 \rightarrow n}(\chi)$ coincides with \[ cl \{\sigma_{\psi}: \mbox{single-edge strategies} \ \psi\} \] Rather than using the measure definition of single-edge strategies, it is more practical to work with the vector $\vec{p}$ of conditional probabilities defined as \[ p_k(u_1,\ldots, u_D) := P_{\psi}[J = k \mid (U^j)_{1 \leq j \leq D} = (u^j)_{1 \leq j \leq D}] \] Then $\psi$ evaluated on product sets is given by \begin{equation}\label{PsiFormula} \psi(A \times B) = \int_A \sum_{j=1}^D p_j (u_1,\ldots, u_D) \mathbf{1}_{\{j \in B\}} \ d\Lambda^D(u_1,\ldots, u_D) \ \forall A \in \mathcal{B}([0,1]^D), B \subseteq \{1,\ldots, D\} \end{equation} and $\sum \limits_{j=1}^D p_j \equiv 1$ everywhere. Intuitively, each $p_k(u_1,\ldots, u_D)$ is the probability of choosing $u_k$ when the observed samples are $(u_1,\ldots,u_D)$. This justifies the term "single-edge strategy," because $\vec{p}$ is prescribing the strategy by which we make our choice once we have the $D$ observed samples. Of course, the vector $\vec{p}=(p_1,\ldots, p_D)$ determines $\psi$ by \eqref{PsiFormula}. And if all the $p_j$ are 0-1 valued $\Lambda^{\times D}$-a.e. then the single-edge strategy $\psi$ is deterministic. Even though different $\vec{p}$ may give rise to the same $\psi$, we conflate the two notions and call both $\psi$ and $\vec{p}$ the single-edge strategy. We use $X(\vec{p}), X(\chi)$ and $\sigma_{\vec{p}}, \sigma_{\psi}$ and $F_{\vec{p}}, F_{\psi}$ interchangeably. At the end of the day, all that matters is whether two vectors $\vec{p}, \vec{q}$ yield the same law $\sigma_{\vec{p}}$ of $X(\vec{p})$. We may now write the cumulative distribution function of $\sigma_{\vec{p}}$ in terms of $\vec{p}$: \begin{equation} \label{cdf} \begin{split} F_{\vec{p}}(y) &=\int_{[0,1]^D} \sum_{j=1}^D p_j (u_1,\ldots, u_D) \mathbf{1}_{[0,y]}(u_j) d\Lambda^D(u_1,\ldots, u_D) \\ &= \int_{[0,y] \times [0,1]^{D-1}} p_1 (u_1,\ldots, u_D) d\Lambda^D(u_1,\ldots, u_D) \\ &+ \ldots + \int_{[0,1]^{D-1} \times [0,y] } p_D (u_1,\ldots, u_D) d\Lambda^D(u_1,\ldots, u_D) \end{split} \end{equation} Therefore $\sigma_{\vec{p}} \ll \Lambda$ with density \begin{equation}\label{density} \begin{split} f_{\vec{p}}(y) &=\int_{[0,1]^{D-1}} p_1(y,u_2,\ldots, u_D) d\Lambda^{D-1}(u_2,\ldots, u_D) \\ &+ \cdots + \int_{[0,1]^{D-1}} p_D(u_1,\ldots,u_{D-1},y) d\Lambda^{D-1}(u_1,\ldots, u_{D-1}) \end{split} \end{equation} It is clear that $f_{\vec{p}} \in [0,D]$.\\ Let us give an example that will make everything clear. Consider the deterministic single-edge strategy $\vec{p}^{MAX}$ that always chooses the maximum of the observed edge labels. In our notation, \[ p_j^{MAX}(u_1, \ldots, u_D) = \mathbf{1}_{\{u_j \geq u_k \ \forall 1 \leq k \leq D\}}\] The resulting distribution $\sigma_{MAX}$ has cdf \[F_{MAX}(y) = P(U^j \leq y \ \forall 1 \leq j \leq D) = y^D \] and density \[f_{MAX}(y) = D y^{D-1}\] This single-edge strategy plays a crucial role later in our description of the extreme points of $\{\sigma_{\vec{p}}\}$. We will also be interested in single-edge strategies "scrambled" by a $\Lambda$-measure-preserving bijection. \begin{definition} Let $\vec{p}$ be a single-edge strategy and let $\phi: [0,1] \rightarrow [0,1]$ be a $\Lambda$-measure-preserving bijection. We define $\vec{p}^\phi$ to be the single-edge strategy with coordinate functions \[p_j^{\phi}(u_1, \ldots, u_D) := p_j (\phi(u_1), \ldots, \phi(u_D))\] It is clear that $\sum p_j^{\phi} \equiv 1$ still so it is a valid single-edge strategy. Furthermore, since $\phi$ is measure-preserving with respect to $\Lambda$, then by a change of variables in the integral formula for $f_{\vec{p}}$ we get \[ f_{\vec{p}^{\phi}} = f_{\vec{p}} \circ \phi \] \end{definition} The following lemma collects some basic facts about single-edge strategies. \begin{lemma}\label{lemma3} Let $\vec{p}$ be a single-edge strategy. \begin{enumerate}[label=(\roman*)] \item For any Borel set $A \in \mathcal{B}([0,1])$, \[\Lambda(A)^D \leq \sigma_{\vec{p}}(A) \leq 1 - (1-\Lambda(A))^D \] In particular, $X(\vec{p}^{MAX})$ stochastically dominates $X(\vec{p})$ and $\sigma_{\vec{p}} \ll \Lambda, \Lambda \ll \sigma_{\vec{p}}$. \item We have \[ \sup_{\phi} E[\sigma_{\vec{p}^{\phi}}] \leq E[\sigma_{MAX}]\] where the supremum is taken over $\Lambda$-measure-preserving bijections $\phi: [0,1] \rightarrow [0,1]$. \item Convex combinations of single-edge strategies $\vec{p}$ translate to convex combinations of $\sigma_{\vec{p}}, f_{\vec{p}}, F_{\vec{p}}$. \end{enumerate} \end{lemma} \begin{remark} It is not clear whether the supremum in (ii) is achieved, but this is beyond the scope of this paper. \end{remark} \begin{proof} (i) If all $D$ edge labels $u_1, \ldots, u_D$ are in the set $A$, then so must be the edge label chosen from among them. Therefore $\Lambda(A)^D \leq \sigma_{\vec{p}}(A)$. The other inequality follows by replacing $A$ with $A^C$. Recalling that $F_{MAX}(y) = y^D$, we get \[F_{MAX}(y) \leq F_{\vec{p}}(y) \ \forall y\] (ii) For any $\Lambda$-measure-preserving bijection $\phi$, we apply the tail integral formula for expectation and use (i) to get \[E[\sigma_{\vec{p}^{\phi}}] = \int_{0}^{1} 1-F_{\vec{p}^{\phi}}(t) dt \leq \int_{0}^{1} 1-F_{MAX}(t) dt = E[\sigma_{MAX}]\] (iii) For single-edge strategies $\vec{p}, \vec{q}$ and $t \in [0,1]$, $t\vec{p} + (1-t)\vec{q}$ is itself a single-edge strategy with \[ F_{t\vec{p} + (1-t)\vec{q}} = t F_{\vec{p}} + (1-t) F_{\vec{q}}, \ f_{t\vec{p} + (1-t)\vec{q}} = t f_{\vec{p}} + (1-t) f_{\vec{q}}, \ \mu_{t\vec{p} + (1-t)\vec{q}} = t \sigma_{\vec{p}} + (1-t) \sigma_{\vec{q}}\] \end{proof} \subsection{"Consistent" Single-Edge Strategies} Before proceeding further, we explain why we can restrict ourselves to single-edge strategies $\vec{p}$ with some convenient symmetries that render $\vec{p}$ consistent. We say a permutation $\iota \in S_D$ acts on a $D$-tuple $\vec{u}$ by applying $\iota$ to the indices: \[ \iota(\vec{u}) = (u_{\iota(1)}, \ldots, u_{\iota(D)})\] \begin{lemma}\label{lemma4} Let $\vec{p}$ be any single-edge strategy. Then there exists another single-edge strategy $\vec{q}$ that gives rise to the same distribution $\sigma_{\vec{p}} = \sigma_{\vec{q}}$ s.t. $\forall x \in [0,1], 1 \leq i \leq D$ \begin{equation}\label{eq17} f_{\vec{p}}(x) = f_{\vec{q}}(x) = D \int_{[0,1]^{D-1}} q_i(u_1, u_2, \ldots, x, \ldots, u_D) d\Lambda^{D-1}(u_1,\ldots, \widehat{u_i}, \ldots, u_D) \end{equation} where the $x$ occurs at position $i$, s.t. \begin{equation}\label{eq18} q_1(\iota_i(\vec{u})) = q_2(\iota_{i-1}(\vec{u})) = \ldots = q_D(\iota_{i+1}(\vec{u})) \ \forall 1 \leq i \leq D \ \forall \vec{u} \end{equation} where each $\iota_j$ is the cyclic shift \[\iota_j(\vec{u}) = (u_j, u_{j+1}, \ldots, u_D, u_1,\ldots, u_{j-1})\] and s.t. \begin{equation}\label{eq19} q_i(\vec{u}) = q_i(\iota(\vec{u})) \ \forall 1 \leq i \leq D, \forall \vec{u}, \mbox{and} \ \forall \iota \in S_D \ \mbox{with} \ \iota(i) = i \end{equation} \end{lemma} \begin{remark}\label{RMK11} The new single-edge strategy $\vec{q}$ is consistent across all permutations of a tuple $\vec{u}$. That is, given an unordered tuple $(u_1, \ldots, u_D)$ we can say that $\vec{q}$ chooses each $u_i$ with probability $t_i$. Then every $q_j(\iota(\vec{u}))$ for $\iota \in S_D$ picks out the probability $t_{\iota(j)}$ i.e. $q_j$ outputs the probability of choosing the $j$th entry in its input tuple. Furthermore, both the density $f_{\vec{q}}$ and the entire single-edge strategy $\vec{q}$ are uniquely determined by any one of the $q_i$. That is, given a measurable function $q_i: [0,1]^D \rightarrow [0,1]$ whose integral over $[0,1]^D$ with respect to the product measure $\Lambda^{\times D}$ is $\frac1D$, which is invariant under permutations in $S_D$ fixing $i$, and which satisfies \[ \sum_j q_i(\iota_j(\vec{u})) = 1 \ \forall \vec{u} \] we can use cyclic shifts to define a valid corresponding single-edge strategy $\vec{q}$ (that satisfies $\sum q_j \equiv 1$ and $\int_{[0,1]^D} q_j \ d\Lambda^{\times D} = \frac1D$) and we can compute $f_{\vec{q}}$ directly from $q_i$. \end{remark} \begin{remark} It is easy to check that $\vec{p}^{MAX}$ satisfies \eqref{eq17}-\eqref{eq19}. Furthermore, if $\vec{p}$ is a consistent single-edge strategy and $\phi:[0,1] \rightarrow [0,1]$ is a $\Lambda$-measure-preserving bijection then so is $\vec{p}^{\phi}$. \end{remark} \begin{proof} The intuition is that we take the average of the original $p_i$ over the desired symmetries. We do this in two steps. For each $1 \leq i \leq D$ define \[p_i'(\vec{u}) := \frac{p_1(\iota_i(\vec{u})) + p_2(\iota_{i-1}(\vec{u})) + \ldots + p_i(\iota_1(\vec{u}))+ \ldots p_D(\iota_{i+1}(\vec{u}))}{D}\] \[q_i(\vec{u}) = \frac1{(D-1)!} \sum_{\iota \in S_{D}, \iota(i) = i} p_i'(u_{\iota(1)}, u_{\iota(2)} \ldots, u_{\iota(D)})\] A straightforward computation shows that $\vec{q}$ satisfies the required properties \eqref{eq17}-\eqref{eq19}. \end{proof} For the rest of the paper we restrict ourselves to these consistent single-edge strategies, which will simplify our computations. \subsection{Closure of \texorpdfstring{$\{\sigma_{\vec{p}}\}$}{mP}} The last remaining part of Theorem \ref{limitPts} is to show that the set of distributions $\sigma_{\vec{p}}$ is weakly closed. \begin{theorem}\label{closure} $\{\sigma_{\vec{p}}: \mbox{ single-edge strategies} \ \vec{p}\}$ is weakly closed \end{theorem} \begin{proof} Suppose $\sigma_{\vec{p}^{n}} \Rightarrow \xi$ for some consistent single-edge strategies $\vec{p}^n$. We seek a single-edge strategy yielding the distribution $\xi$. Consider any $n$. Let $\nu_{\vec{p}^n}$ be the distribution on $[0,1]^D$ given by integration against \\ $D p_1^n d\Lambda^{\times D}$. Observe that \eqref{eq17} implies \[ f_{\vec{p}^n}(y) = \int_{[0,1]^{D-1}} D p_1^n(y, u_2, \ldots, u_D) d\Lambda^{D-1}(u_2, \ldots, u_D) \] so $\sigma_{\vec{p}^n}$ is just the first coordinate marginal of $\nu_{\vec{p}^n}$. Compactness yields a weakly convergent subsequence $\nu_{\vec{p}^{n_j}} \Rightarrow \nu$. By a standard argument, since $Dp_1^n$ are uniformly bounded by $D$, then $\nu \ll \Lambda^{\times D}$ and has a density of the form $Dp_1$ for some measurable function $p_1: [0,1]^D \rightarrow [0,1]$. This along with the fact that $\int_{[0,1]^D} p_1 d\Lambda^{\times D} = \frac1D$ are enough for $p_1$ to give rise to a single-edge strategy $\vec{p}$ by the first Remark following Lemma \ref{lemma4}. But then the corresponding first coordinate marginals must also converge weakly, so we get $\sigma_{\vec{p}^{n_j}} \Rightarrow \xi'$ where $\xi'$ is the corresponding marginal of $\nu$: \[ \xi'(A) =\int_{A\times [0,1]^{D-1}} Dp_1 d\Lambda^D = \sigma_{\vec{p}}(A) \ \forall A \in \mathcal{B}([0,1])\] Thus $\xi' = \sigma_{\vec{p}}$. On the other hand, $\sigma_{\vec{p}^n} \Rightarrow \xi$ so by uniqueness of weak limits, $\xi = \xi' = \sigma_{\vec{p}}$. \end{proof} Therefore the set of possible distributions $\{\sigma_{\vec{p}}\}$ of $X(\vec{p})$ over single-edge strategies $\vec{p}$ almost surely coincides with the set of limit points of the empirical measures $\frac1n \mu_{0 \rightarrow n}(\chi)$ over all strategies $\chi$. In particular these have the same extreme points. \section{Extreme Points of \texorpdfstring{$\{\sigma_{\vec{p}}\}$}{mp}}\label{extreme} Our goal in this section is to characterize the extreme points of the possible distributions of $X$ we can observe as we vary the underlying single-edge strategy. It turns out that the extreme points of $\{\sigma_{\vec{p}}: \mbox{ single-edge strategies} \ \vec{p}\}$ are precisely those $\sigma_{\vec{p}}$ which have scramblings with mean converging to the mean of $\sigma_{MAX}$, or, equivalently, those $\sigma_{\vec{p}}$ with deterministic $\vec{p}$. \begin{theorem}\label{thm4} Let $\vec{p}$ be a single-edge strategy. The following are equivalent: \begin{enumerate}[label=(\roman*)] \item $\sigma_{\vec{p}}$ is an extreme point \item Any consistent single-edge strategy achieving $\sigma_{\vec{p}}$ must be deterministic \item $f_{\vec{p}}$ is not constant on sets of positive $\Lambda$-measure and $\sigma_{\vec{p}}$ is given by the following single-edge strategy $\vec{q}$: \[ q_k(u_1,\ldots, u_D) = \textbf{1}_{\{f_{\vec{p}}(u_k) \geq f_{\vec{p}}(u_i) \ \forall i\}} \ \mbox{for $\Lambda^{\times D}$-a.a.} \ (u_1,\ldots, u_D) \] In other words, $\sigma_{\vec{p}}$ is achieved by the deterministic single-edge strategy "choose \\ whichever label yields a higher value when evaluating the density $f_{\vec{p}}$" \item $\sup_{\phi} E[\sigma_{\vec{p}^{\phi}}] = E[\sigma_{MAX}]$ \item For $U \sim [0,1]$ if $f_{\vec{p}}(U), f_{MAX}(U)$ as $[0,D]$-valued random variables on the probability space $([0,1], \mathcal{B}([0,1]), \Lambda)$, then they have the same distribution. That is, \[P(f_{\vec{p}} \leq x) = P(f_{MAX} \leq x) \ \forall x\] \end{enumerate} \end{theorem} \begin{remark} Since $f_{MAX}(y) = Dy^{D-1}$ on $[0,1]$ then $\frac1D f_{MAX} \sim \mbox{Beta}(D,1)$. Thus (v) implies that for $U\sim \mbox{Unif}[0,1]$, $f_{\vec{p}}(U)$ has a continuous probability distribution. \end{remark} We prove this theorem over the next couple of sections. \subsection{Deterministic Single-Edge Strategies } We begin by characterizing the extreme points of the set of distributions $\{\sigma_{\vec{p}}\}$ in terms of deterministic single-edge strategies. \begin{lemma}\label{lemma6} (i) $\Leftrightarrow$ (ii) in Theorem \ref{thm4} \end{lemma} \begin{proof} We prove both contrapositives. First, consider a distribution $\sigma_{\vec{p}}$ achieved by a non-deterministic consistent single-edge strategy $\vec{p}$, i.e. a single-edge strategy satisfying \eqref{eq17}-\eqref{eq19}. We will construct a set of positive $\Lambda^{\times D}$ measure on which we perturb $p_1$ in a way that allows us to write $\vec{p}$ as a non-trivial average of two single-edge strategies $\vec{q}, \vec{r}$. By our assumption, there exists $0 < \epsilon < \frac12$ s.t. the set \[A = \{\vec{u}: p_1(\vec{u}) \in (\epsilon, 1-\epsilon), p_1(\vec{u}) \geq p_1(\iota_i(\vec{u})) \ \forall 1 \leq i \leq D\} \ \mbox{has positive $\Lambda^{\times D}$ measure}\] where $\iota_i$ are the cyclic shifts as before. Then there are $a_1 < b_1, a_2 < b_2$ s.t. $[a_1, b_1], [a_2, b_2]$ are disjoint and \[A' = A \cap [a_1, b_1] \times [a_2, b_2]^{D-1} \ \mbox{has positive $\Lambda^{\times D}$ measure}\] The purpose of this is to ensure that the first coordinate in a tuple in $A'$ cannot appear in any other position in another tuple in $A'$. By \eqref{eq18}, the fact that $p_1(\vec{u}) < 1-\epsilon$ on $A'$, and the definition of single-edge strategies, \[\sum_{k=2}^D p_1(\iota_k(\vec{u})) = \sum_{k=2}^D p_k(\vec{u}) = 1-p_1(\vec{u}) > \epsilon \ \forall \vec{u} \in A'\] Hence for any $\vec{u} \in A'$, there exists $2 \leq k \leq D$ s.t. $p_1(\iota_k(\vec{u})) > \frac{\epsilon}D$. It follows that there is $2 \leq k \leq D$ s.t. \[A'' = \bigg\{\vec{u} \in A': p_1(\iota_k(\vec{u})) \in \bigg(\frac{\epsilon}D, 1-\epsilon\bigg)\bigg\} \ \mbox{has positive $\Lambda^{\times D}$ measure}\] Define \[q_1 := p_1 + \frac{\epsilon}{D!} \sum_{\iota \in S_D, \iota(1)= 1} ( \mathbf{1}_{\iota(A'')} - \mathbf{1}_{\iota(\iota_k^{-1}(A''))}), r_1 := p_1 + \frac{\epsilon}{D!} \sum_{\iota \in S_D, \iota(1)= 1} ( - \mathbf{1}_{\iota(A'')} + \mathbf{1}_{\iota(\iota_k^{-1}(A''))})\] It is clear that $q_1, r_1$ average to $p_1$, and are both invariant under permutations in $S_D$ fixing 1. Since $p_1(\vec{u}) \in (\frac{\epsilon}D, 1-\epsilon) \ \forall \vec{u} \in \iota(A'') \cup \iota(\iota_k^{-1}(A'')) \ \forall \iota \in S_D, \iota(1) = 1$ then both $q_1, r_1$ have ranges in $[0,1]$. Also, note that \[\int_{[0,1]^D} \mathbf{1}_{\iota(B)} d\Lambda^{\times D} = \Lambda^{\times D}(B) \ \forall B \in \mathcal{B}([0,1]^D), \iota \in S_D \] so it follows that \[\int_{[0,1]^D} q_1 d\Lambda^{\times D} = \int_{[0,1]^D} r_1 d\Lambda^{\times D} = \int_{[0,1]^D} p_1 d\Lambda^{\times D} = \frac1D \] Finally, note that by a simple counting argument, for any $\vec{u}$, \[\sum_i \sum_{\iota \in S_D, \iota(1)= 1} \mathbf{1}_{\iota(A'')}(\iota_i(\vec{u})) = \sum_{\iota \in S_D} \mathbf{1}_{\iota(A'')}(\vec{u}) = \sum_i \sum_{\iota \in S_D, \iota(1)= 1} \mathbf{1}_{\iota(\iota_k^{-1}(A''))}(\iota_i(\vec{u})) \] so \[\sum_i q_1(\iota_i(\vec{u})) = \sum_i r_1(\iota_i(\vec{u})) = \sum_i p_1(\iota_i(\vec{u})) = 1 \ \forall \vec{u}\] Thus $q_1, r_1$ are valid probability functions which average to $p_1$. By the remark after Lemma \ref{lemma4}, these uniquely determine single-edge strategies $\vec{q}, \vec{r}$ whose average is $\vec{p}$. Thus $f_{\vec{p}}$ is the average of $f_{\vec{q}}, f_{\vec{r}}$. It remains to show this convex combination is non-trivial. Consider any $\vec{u} = (u_1, u_2,\ldots, u_D)$ with $u_1 \in [a_1, b_1]$. For any $\iota \in S_D, \iota(1) = 1$, \[\iota^{-1}(k) \neq 1 \Rightarrow u_{\iota^{-1}(k)} \in [a_2, b_2] \Rightarrow u_{\iota^{-1}(k)} \notin [a_1, b_1] \Rightarrow \vec{u} \notin \iota(\iota_k^{-1}(A''))\] It follows that \[ q_1(\vec{u}) = p_1(\vec{u}) + \frac{\epsilon}{D!} \sum_{\iota \in S_D, \iota(1)= 1} \mathbf{1}_{\iota(A'')}(\vec{u}) \geq p_1(\vec{u}) + \frac{\epsilon}{D!} \mathbf{1}_{A''}(\vec{u})\] so by \eqref{eq17}, \[ f_{\vec{q}}(u_1) \geq f_{\vec{p}}(u_1) + \frac{\epsilon}{(D-1)!}\int_{[0,1]^{D-1}} \mathbf{1}_{A''} (u_1, u_2, \ldots, u_D) \ d\Lambda^{\times (D-1)}(u_2,\ldots, u_D) \] Since $\Lambda^{\times D} (A'') > 0$ then there is $\delta > 0$ and a Borel set $B \subset [a_1, b_1]$ with $\Lambda(B)> 0$ s.t. the integral above is $\geq \delta$ for $u_1 \in B$. It follows that $f_{\vec{q}} \geq f_{\vec{p}} + \frac{\epsilon}{(D-1)!}\delta$ on $B$ so the convex combination is non-trivial. Therefore $\sigma_{\vec{p}}$ is not an extreme point. Now suppose $\sigma_{\vec{p}}$ is not an extreme point, say $\sigma_{\vec{p}} = t\sigma_{\vec{q}} + (1-t) \sigma_{\vec{r}}$ for some consistent single-edge strategies $\vec{q}, \vec{r}$ with $\sigma_{\vec{q}} \neq \sigma_{\vec{r}}$. Then $\sigma_{\vec{p}}$ is achieved by the single-edge strategy $t \vec{q} + (1-t) \vec{r}$, which trivially satisfies \eqref{eq17}-\eqref{eq19} so is consistent. Since $\vec{q}, \vec{r}$ have range in $[0,1]$ and differ on some set of positive $\Lambda^{\times D}$-measure then on this same set $t \vec{q} + (1-t) \vec{r}$ is not $\{0,1\}$-valued. Thus $\sigma_{\vec{p}}$ is achieved by a non-deterministic single-edge strategy satisfying \eqref{eq17}-\eqref{eq19}. \end{proof} \subsection{"Weight Tuples" of Single-Edge Strategies} $\sigma_{\vec{p}}$ being an extreme point will give us further properties relating to $(D+1)$-tuples \\ $(u_1, \ldots, u_{D+1})$, but these properties hold a.s. and we must be extra careful about which tuples living in measure 0 sets we are excluding. We need some setup to address this issue. Let $\vec{p}$ be a consistent single-edge strategy s.t. $\sigma_{\vec{p}}$ is an extreme point. By Lemma \ref{lemma6}, $p_1$ is 0-1 valued $\Lambda^{\times D}$-a.e.. Also, the set of tuples in $[0,1]^D$ with two or more duplicates has $\Lambda^{\times D}$-measure 0. Define \[ S := \{\mbox{ordered} \ (u_1,\ldots,u_D) \in {\mathbb{R}} ^D \ \mbox{ duplicate-free s.t.} \ p_1(u_1,\ldots,u_D) \in \{0,1\}\} \] to be the set of "good" ordered $D$-tuples we consider. Note that $\Lambda^D(S)= 1$. Once we prove more a.s. properties of the elements of $S$, we will amend this definition of $S$ accordingly. Combining the fact that $p_1$ is 0-1 valued on $S$ and the remark after Lemma \ref{lemma4}, we see that $p_1$ is equivalent to a choice function taking in \emph{unordered} tuples $(u_1, \ldots, u_D)$ and outputting one of the coordinates $u_j$; then for any $\iota \in S_D$, $p_1(\iota(\vec{u})) = \mathbf{1}_{\{j = \iota(1)\}}$ i.e. $p_1$ returns whether or not the first coordinate of the input ordered tuple is the choice associated with the corresponding unordered tuple. Let \[ T = \left\{ \begin{array}{l} \mbox{ordered} \ (u_1,\ldots,u_{D+1}) \in [0,1]^{D+1} \ \mbox{ s.t.} \\ \mbox{all $(D+1)!$ ordered subtuples of size $D$ are in $S$} \end{array}\right\} \] It is clear that $\Lambda^{\times (D+1)}(T) = 1$ and the tuples in $T$ contain no duplicate entries. $T$ is the set of "good" ordered $(D+1)$-tuples we restrict ourselves to. For ordered $(D+1)$-tuples in $T$, we wish to study the possible sets of choices we can make for the $(D+1)!$ ordered subtuples of size $D$. Since the choice for some $\vec{u} \in [0,1]^D$ is also the choice for $\iota(\vec{u})$ for all $\iota \in S_D$, then rather than keeping a factor of $D!$ everywhere, we instead consider the possible sets of choices we can make for the $D+1$ \emph{unordered} subtuples of size $D$. We can encode these choices as an ordered "weight tuple" $(w_1, \ldots, w_{D+1})$ where $w_j$ is the number of times $u_j$ is the choice made. In terms of the consistent single-edge strategy $\vec{p}$, \begin{equation}\label{equation4} w_j = w_j(u_1,\ldots, u_{D+1}) := \sum_{1 \leq k \leq D+1, k \neq j} p_1(u_j,u_1,\ldots, \widehat{u_j},\ldots,\widehat{u_k}, \ldots, u_{D+1}) \end{equation} For example, the weight tuple $(D, 1, 0, \ldots, 0)$ corresponds to the choices \[ (u_1,\ldots, \widehat{u_k}, \ldots, u_{D+1}) \mapsto u_1 \ \forall 2 \leq k \leq D+1 \ \mbox{and} \ (u_2,\ldots, u_{D+1})\mapsto u_2\] Note that each $u_j$ appears in exactly $D$ of the $D+1$ $D$-subtuples and there is exactly one choice for each subtuple so \begin{equation}\label{eq20} w_i \in \{0, 1,\ldots, D\} \ \forall 1 \leq i \leq D+1, \sum w_i \equiv D+1 \end{equation} Even though a weight tuple may correspond to more than one set of choices for the $D+1$ $D$-subtuples, it is easily checked that every valid weight tuple (satisfying \eqref{eq20}) corresponds to at least one set of choices. The following flow chart summarizes this process: \begin{center} \begin{tikzpicture} \node[draw] at (0,100pt) {Start with ordered $(D+1)$-tuple $(u_1,\ldots,u_{D+1}) \in T$}; \node[draw, align=left] at (0,50pt) {Consider the choice for each unordered $D$-subtuple $(u_1,\ldots, \widehat{u_k}, \ldots, u_{D+1})$. \\ Count occurrences of each $u_j$ among these choices.}; \node[draw] at (0,0pt) {Get an ordered $(D+1$)-weight tuple $(w_1, \ldots, w_{D+1})$ satisfying \eqref{eq20}}; \draw [->] (0,90pt) to (0,68pt); \draw [->] (0,33pt) to (0,10pt); \end{tikzpicture} \end{center} One last observation we make about the weight tuples is that they can be used to compute the density $f_{\vec{p}}$ directly. Fix any $1 \leq j \leq D+1$. Observe that \eqref{eq17} with a change of variables gives \begin{equation*} \begin{split} f_{\vec{p}}(t) &= D \int_{[0,1]^{D-1}} p_1(t,u_2, \ldots, u_D) d\Lambda^{D-1}(u_2, \ldots, u_D) \\ &= \sum_{1 \leq k \leq D+1, k \neq j} \int_{[0,1]^{D-1}} p_1(t,u_1,\ldots, \widehat{u_j},\ldots,\widehat{u_k}, \ldots, u_{D+1}) \\ &\qquad \qquad \qquad d\Lambda^{D-1}(u_1, \ldots, \widehat{u_j}, \ldots,\widehat{u_k}, \ldots, u_{D+1}) \end{split} \end{equation*} But $\Lambda$ is a probability measure so we can integrate everything against $du_k$ without changing the value: \begin{equation} \label{eqn6} \begin{split} f_{\vec{p}}(t) &= \sum_{1 \leq k \leq D+1, k \neq j} \int_{[0,1]^{D}} p_1(t,u_1,\ldots, \widehat{u_j},\ldots,\widehat{u_k}, \ldots, u_{D+1}) d\Lambda^{D}(u_1, \ldots, \widehat{u_j}, \ldots, u_{D+1}) \\ &=\int_{[0,1]^{D}} w_j(u_1,\ldots, t, \ldots, u_{D+1} ) d\Lambda^{D}(u_1, \ldots, \widehat{u_j}, \ldots, u_{D+1}) \ \mbox{by \eqref{equation4}} \end{split} \end{equation} where $t$ occurs in position $j$. Note that the above formula holds for any $t$ for which \[\{(u_1,\ldots, u_{j-1}, u_{j+1},\ldots, u_{D+1}) | (u_1,\ldots, u_{j-1},t,u_{j+1},\ldots, u_{D+1}) \in T\} \ \mbox{has} \ \Lambda^{\times D}-\mbox{measure 1}\] i.e. it holds for $t$ in a set of $\Lambda$-measure 1. \subsection{"Weight Tuples" of the Extreme Points} We claim that $\sigma_{\vec{p}}$ being an extreme point implies that $\Lambda^{\times (D+1)}$-almost all weight tuples must be a permutation of $(D,1,0,\ldots, 0)$. That means that almost all $(D+1)$-tuples in $T$ have one coordinate that "dominates" the others in terms of the choice function, and this behaviour will imply precisely that $f_{\vec{p}}$ can be approximated by "scrambles" of $f_{MAX}$. \begin{lemma} Suppose $\vec{p}$ is a consistent, single-edge strategy s.t. $\sigma_{\vec{p}}$ is an extreme point. For $\Lambda^{\times (D+1)}$-a.a. ordered tuples in $T$, the corresponding weight tuple is a permutation of $(D,1,0,\ldots,0)$. \end{lemma} \begin{proof} We sketch the details of the proof in the case $D \geq 3$. The case $D = 2$ is similar. Suppose the claim is false, say the set $U$ of ordered $(D+1)$-tuples $(u_1,\ldots, u_{D+1}) \in T$ whose weight tuple $(w_1,\ldots, w_{D+1})$ is not a permutation of $(D,1,0,\ldots,0)$ has positive $\Lambda^{\times (D+1)}$-measure. Note that $U$ is clearly invariant under permutations in $S_{D+1}$. We follow a similar approach as in the proof of Lemma \ref{lemma6}, in that we seek write $\vec{p}$ as a convex combination of two new single-edge strategies $\vec{q}, \vec{r}$ obtained by perturbing $\vec{p}$. First, we further restrict $U$. There exist $a_i < b_i$ s.t. $[a_i, b_i]$ are pairwise disjoint and s.t. \[U' := U \cap [a_1, b_1] \times \cdots \times [a_{D+1}, b_{D+1}] \ \mbox{has positive $\Lambda^{\times (D+1)}$ measure} \] In this way, each coordinate in a tuple in $U'$ cannot appear in any other position in another tuple in $U'$. Now consider any $(u_1, \ldots, u_{D+1}) \in U'$ so it has a maximal weight $1 \leq w_K \leq D-1$ in its weight tuple. If $w_K \geq 2$ then either there exists $i \neq K$ s.t. $w_i \geq 2$ or there are at least two $i \neq K$ s.t. $w_i = 1$ (this is because $\sum w_m = D+1$ and $w_K \leq D-1$). In either case, we can pick $I \neq K$ so that the choice for $(u_1, \ldots, \widehat{u_K}, \ldots, u_{D+1})$ is not $u_I$, hence there is another $D$-subtuple containing $u_K$ for which the choice is $u_I$. Furthermore, since $w_K \geq 2$, then there also is a $D$-subtuple containing $u_I$ for which the choice is $u_K$. On the other hand, if $w_K = 1$ then $w_m = 1 \ \forall m$. If $(u_{m_1}, \ldots, u_{m_D})$ is the $D$-subtuple for which the choice is $u_K$ then of the remaining $D-1 \geq 2$ coordinates $u_{m_{n}}$ with $m_n \neq K$ at least one, call it $u_I$, is the choice for a $D$-subtuple other than $(u_1, \ldots, \widehat{u_K}, \ldots, u_{D+1})$, i.e. for a $D$-subtuple containing $u_K$. Since this holds for all $\vec{u} \in U'$ then there exist distinct $1 \leq I, J, K, L \leq D+1$ s.t. \[ U'' := \left\{ \vec{u} \in U'\ \middle\vert \begin{array}{l} 1 \leq w_I, \ \mbox{and} \ w_m \leq w_K \ \forall m, \ \mbox{and} \\ \mbox{ the choice for the $D$-subtuple $(u_1,\ldots, \widehat{u_J}, \ldots, u_{D+1})$ is $u_I$, and} \\ \mbox{ the choice for the $D$-subtuple $(u_1,\ldots, \widehat{u_L}, \ldots, u_{D+1})$ is $u_K$} \end{array}\right\} \] has positive $\Lambda^{\times (D+1)}$ measure. The idea is that we write the weight tuples for $(D+1)$-tuples in $U''$ as averages of two different weight tuples, and use these new weight tuples to obtain two new non-trivial, valid single-edge strategies whose densities average to $f_{\vec{p}}$. Consider any $\vec{u} \in U''$. We can write the corresponding weight tuple $(w_1, \ldots, w_{D+1})$ as an average of two other weight tuples: \[(\ldots, w_I, \ldots, w_K, \ldots) = \frac12 ( \ldots, w_I+1, \ldots, w_K-1, \ldots) + \frac12 (\ldots, w_I-1, \ldots, w_K+1, \ldots) \] The weight tuple $( \ldots, w_I+1, \ldots, w_K-1, \ldots)$ can be achieved simply by changing the choice of $(u_1, \ldots, \widehat{u_L}, \ldots, u_{D+1})$ from $u_K$ to $u_I$, whereas the weight tuple $( \ldots, w_I-1, \ldots, w_K+1, \ldots)$ can be achieved by changing the choice of $(u_1, \ldots, \widehat{u_J}, \ldots, u_{D+1})$ from $u_I$ to $u_K$. It is trivial to see that these two new tuples are also valid weight tuples that satisfy \eqref{eq20} so they correspond to new single-edge strategies $\vec{q}, \vec{r}$ respectively. By \eqref{eqn6}, we see that $f_{\vec{p}}(y)$ is the average of $f_{\vec{q}}(y), f_{\vec{r}}(y)$. The only difference in the $D=2$ case is that the new weight tuples are achieved by making two changes to the choice function rather than one. It remains to check that this average is non-trivial. Let \[V := \{u_K: \exists u_1, \ldots, \widehat{u_K}, \ldots, u_{D+1} \ \mbox{s.t.} \ (u_1,\ldots, u_{D+1}) \in U''\} \] and for $x \in [0,1]$ let \[W^{x} := \{(u_1, \ldots, \widehat{u_K}, \ldots, u_{D+1}): \ (u_1,\ldots, , u_{K-1}, x, u_{K+1}, \ldots, u_{D+1}) \in U''\} \] Then $\Lambda^{(D+1) \times}(U'') > 0$ implies $\Lambda(V) > 0$ and for $\Lambda$-a.a. $u_K \in V$, $\Lambda^{\times D} (W^{u_K}) > 0$. Let us consider how the function $w_K(\cdot, \ldots, \cdot, u_K, \cdot, \ldots, \cdot): [0,1]^D \rightarrow \{0,1,\ldots, D\}$ changes from $\vec{p}$ to $\vec{r}$ for any given fixed $u_K \in V$. On $W^{u_K}$, $w_K(\cdot, \ldots, \cdot, u_K, \cdot, \ldots, \cdot)$ increases by 1 by construction of $\vec{r}$. Note that $u_K$ can only appear in the $K$th coordinate of a tuple in $U''$ (by construction of $U''$) hence $w_K(\cdot, \ldots, \cdot, u_K, \cdot, \ldots, \cdot)$ either remains unchanged or increases by 1 off of $W^{u_K}$ (because the only way it decreases by 1 is if $u_K$ had appeared at index $I$ in a tuple in $U''$). Thus, for any $u_K \in V$, \[f_{\vec{r}}(u_K) \geq f_{\vec{p}}(u_K) + \Lambda^{\times D}(W^{u_K}) \] by \eqref{eqn6}. It follows that $\sigma_{\vec{r}} \neq \sigma_{\vec{p}}$ so this is indeed a non-trivial convex combination. This contradicts the assumption that $\sigma_{\vec{p}}$ was an extreme point. Therefore $\Lambda^{\times (D+1)}$-a.a. weight tuples must be a permutation of $(D, 1,0,\ldots, 0)$. \end{proof} We amend our definition of "good" $(D+1)$-tuples: \[T' := \{(u_1,\ldots, u_{D+1}) \in T: (w_1,\ldots, w_{D+1}) \ \mbox{is a perm. of } \ (D,1,0,\ldots,0)\}\] We still have $\Lambda^{\times (D+1)}(T') = 1$ (provided of course that $\sigma_{\vec{p}}$ is an extreme point). We now prove the (i) $\Rightarrow$ (ii) direction of Theorem \ref{thm4}. We split the proof into several claims. \begin{lemma}\label{lemma8} Let $\sigma_{\vec{p}}$ be an extreme point.\\ (i) Suppose $(u_1, \ldots, u_{D+1}) \in T'$ and $u_i$ has weight $D$ in $(u_1, \ldots, u_{D+1})$. Then for any $1 \leq j \leq D+1$, $j \neq i$ and any $u_j'$ s.t. $(u_1, \ldots, u_{j-1}, u_j', u_{j+1}, \ldots, u_{D+1}) \in T'$, either $u_i$ or $u_j'$ has weight $D$ in $(u_1, \ldots, u_{j-1}, u_j', u_{j+1}, \ldots, u_{D+1})$. (ii) Suppose $(u_1, u_1', u_2, \ldots, u_D) \in T'$. Then in the corresponding weight tuple, $u_1'$ has weight $D$ and $u_1$ has weight 0 if and only if \[p_1(u_1, u_2, u_3, \ldots, u_D) < p_1(u_1', u_2, u_3, \ldots, u_D) \] (iii) For $\Lambda^{\times 2}$-a.a. $(u_1, u_1')$, if \[p_1(u_1, u_2, u_3, \ldots, u_D) < p_1(u_1', u_2, u_3, \ldots, u_D) \] on a set of positive $\Lambda^{\times (D-1)}$ measure then \[p_1(u_1, u_2, u_3, \ldots, u_D) \leq p_1(u_1', u_2, u_3, \ldots, u_D) \] for $\Lambda^{\times (D-1)}$-a.a. $(u_2, \ldots, u_D)$. (iv) For $\Lambda^{\times 2}$-a.a. $(u_1, u_1')$, if $f_{\vec{p}}(u_1) \leq f_{\vec{p}}(u_1')$ then \[p_1(u_1, u_2, u_3, \ldots, u_D) \leq p_1(u_1', u_2, u_3, \ldots, u_D) \] for $\Lambda^{\times (D-1)}$-a.a. $(u_2, \ldots, u_D)$. (v) $f_{\vec{p}}$ is not constant on any set of positive $\Lambda$ measure. (vi) For $\Lambda^{\times D}$-a.a. $(u_1, \ldots, u_D)$, \[ p_1(u_1,\ldots, u_D) \geq \mathbf{1}_{\{f_{\vec{p}}(u_1) \geq f_{\vec{p}}(u_i) \ \forall i\}} \] (vii) For $\Lambda^{\times D}$-a.a. $(u_1, \ldots, u_D)$, \[ p_1(u_1,\ldots, u_D) = \mathbf{1}_{\{f_{\vec{p}}(u_1) \geq f_{\vec{p}}(u_i) \ \forall i\}} \] \end{lemma} \begin{remark} We give heuristics to aid in understanding these claims:\\ (i) If $u_i$ is the dominant choice in $(u_1,\ldots, u_{D+1})$ then $u_i$ can never be dominated by one of the $u_j$ for $i \neq j$ in tuples containing $u_i$ and $u_j$.\\ (ii) Dominance can be determined by evaluating $p_1$.\\ (iii) If $u_1'$ dominates $u_1$ once then $u_1'$ almost always dominate $u_1$.\\ (iv) The ordering imposed by domination coincides with the ordering imposed by the values of $f_{\vec{p}}$.\\ (vii) $u_1$ is the choice in $(u_1,\ldots,u_D)$ if and only of $u_1$ maximizes the value of $f_{\vec{p}}$. \end{remark} \begin{proof} (i) Let $1 \leq i \leq D$ be the index for which $u_i$ has weight $D$ in $(u_1,\ldots, u_{D+1})$. Recall that this means we have the choices \[ (u_1,\ldots, \widehat{u_{\ell}},\ldots, u_{D+1}) \mapsto u_i \ \forall \ell \neq i\] Taking $\ell = j$, this means $u_i$ has weight 1 or $D$ in $(u_1,\ldots, u_{j-1}, u_j', u_{j+1},\ldots,u_{D+1})$ (because this $(D+1)$-tuple is in $T'$). If it is $D$, we are done. If it is 1, the fact that we have the choice \[(u_1,\ldots, u_{j-1}, u_{j+1},\ldots, u_{D+1}) \mapsto u_i \] implies no $u_k, k \neq j$ can have weight $D$ so $u_j'$ must have weight $D$ in \\ $(u_1,\ldots, u_{j-1}, u_j', u_{j+1},\ldots,u_{D+1})$. (ii) Recall that $p_1$ is 0-1 valued on $D$-subtuples of tuples in $T'$. We have the following sequence of if-and-only-if statements: \begin{align*} & p_1(u_1, u_2, u_3, \ldots, u_D) < p_1(u_1', u_2, u_3, \ldots, u_D) \\ &\Leftrightarrow p_1(u_1, u_2, u_3, \ldots, u_D) = 0, p_1(u_1', u_2, u_3, \ldots, u_D) = 1 \\ & \Leftrightarrow (u_1, u_2, u_3, \ldots, u_D) \mapsto u_i \ \mbox{for some} \ 2 \leq i \leq D \ \mbox{and} \ (u_1', u_2, u_3, \ldots, u_D) \mapsto u_1'\\ \end{align*} Now if $u_1'$ has weight $D$ and $u_1$ has weight 0 in $(u_1, u_1',u_2,\ldots, u_D$), then we must have \begin{equation} \label{equation6} (u_1, u_2, u_3, \ldots, u_D) \mapsto u_i \ \mbox{for some} \ 2 \leq i \leq D \ \mbox{and} \ (u_1', u_2, u_3, \ldots, u_D) \mapsto u_1' \end{equation} On the other hand, if we know \eqref{equation6} then the weight of $u_i$ cannot be $D$ since the choice for $(u_1', u_2, \ldots, u_i, \ldots, u_D)$ is $u_1'$. Thus $u_i$ has weight 1, so $u_1'$ has weight $D$ and $u_1$ has weight 0. (iii) Consider $u_1, u_1'$ s.t. $(u_1, u_1', u_2,\ldots, u_D) \in T'$ for $\Lambda^{\times (D-1)}$-a.a. $(u_2, \ldots, u_D)$. This clearly holds for $\Lambda^{\times 2}$-a.a. $(u_1, u_1')$ with $u_1 \neq u_1'$. Let \[U = \{(u_2, \ldots, u_D): (u_1, u_1', u_2,\ldots, u_D) \in T', p_1(u_1, u_2, u_3, \ldots, u_D) < p_1(u_1', u_2, u_3, \ldots, u_D) \}\] so $\Lambda^{\times (D-1)}(U) > 0$ by assumption. Consider any $(u_2, \ldots, u_D) \in U$ and $u_2'$ s.t. $(u_1, u_1', u_2', u_3, \ldots, u_D) \in T'$. By (ii), $u_1'$ has weight $D$ and $u_1$ has weight 0 in the weight tuple for $(u_1, u_1', u_2, u_3, \ldots, u_D)$. By (i), it follows that either $u_1'$ or $u_2'$ has weight $D$ in $(u_1, u_1', u_2', u_3, \ldots, u_D)$. By the contrapositive of (ii), \[p_1(u_1, u_2', u_3, \ldots, u_D) \leq p_1(u_1', u_2', u_3, \ldots, u_D) \] Note that for $\Lambda^{\times 2}$-a.a. $(u_1, u_1')$ it is true that for $\Lambda^{\times (D-1)}(U)>0$ of $(u_2, \ldots, u_D) \in U$ and $\Lambda$-a.a. $u_2' \in {\mathbb{R}} $ we have $(u_1, u_1', u_2', u_3, \ldots, u_D) \in T'$ and the inequality above holds. We repeat this argument, "replacing" $u_2, u_3, \ldots, u_D$ with $\Lambda$-almost arbitrary $u_2', u_3', \ldots, u_D'$ and noting the invariant that one of $u_1',\ldots, u_k'$ has weight of $D$ in the weight tuple of $(u_1, u_1', u_2', \ldots, u_k', u_{k+1}, \ldots, u_D) \in T'$ (which means we may apply (ii) to replace $u_{k+1}$ with $\Lambda$-almost arbitrary $u_{k+1}'$). Thus, for $\Lambda^{\times 2}$-a.a. $(u_1, u_1')$, \[p_1(u_1, u_2', u_3', \ldots, u_D') \leq p_1(u_1', u_2', u_3', \ldots, u_D') \] for $\Lambda^{(D-1)}$-a.a. $(u_2', \ldots, u_D')$. (iv) This follows immediately from (iii) and \eqref{eq17}. (v) Suppose $f_{\vec{p}}$ is constant on a set $V$ with $\Lambda(V) > 0$. By (iv), can restrict $V$ to a set $V'$ with $\Lambda(V') = \Lambda(V) > 0$ s.t. for all distinct $u_1, u_2, \ldots, u_{D+1} \in V'$, $(u_1, \ldots, u_{D+1}) \in T'$ and \[ p_1(u_i, u_2, \ldots, \widehat{u_i}, \ldots, \widehat{u_j}, \ldots, u_{D+1}) = p_1(u_j, u_2, \ldots, \widehat{u_i}, \ldots, \widehat{u_j}, \ldots, u_{D+1}) \ \forall i, j \] We claim this is impossible. Consider distinct $u_1, \ldots, u_{D+1} \in V'$. Suppose $u_i$ has weight $D$ and $u_k$ has weight 1 in the corresponding weight tuple. Pick $j \notin \{i, k\}$ (which exists since $D+1 \geq 3$). Then $u_j$ has weight 0 in the corresponding weight tuple. But this implies the contradiction \[ 1 = p_1(u_i, u_2, \ldots, \widehat{u_i}, \ldots, \widehat{u_j}, \ldots, u_{D+1}) = p_1(u_j, u_2, \ldots, \widehat{u_i}, \ldots, \widehat{u_j}, \ldots, u_{D+1}) = 0 \] Therefore $f_{\vec{p}}$ is not constant on any set of positive $\Lambda$-measure. (vi) It suffices to show that for $\Lambda$-a.a. $u_1$, we have \[p_1(u_1,u_2,\ldots, u_D) = 1 \] for all but a $\Lambda^{\times (D-1)}$-null subset of tuples $(u_2,\ldots, u_D) \in \bigg(f_{\vec{p}}^{-1}([0, f_{\vec{p}}(u_1)]) \bigg)^{D-1}$. Suppose not, say there is a set $G$ of $u_1$ of $\Lambda$-positive measure on which \[p_1(u_1,u_2,\ldots, u_D) = 0 \] for a tuples $(u_2,\ldots, u_D)$ in a set $H_{u_1} \subseteq \bigg(f_{\vec{p}}^{-1}([0, f_{\vec{p}}(u_1)])\bigg)^{D-1}$ of $\Lambda^{\times (D-1)}$-positive measure. By (iv), for $\Lambda$-a.a. $u_1 \in G$, \[0= p_1(u_1,u_2,\ldots, u_D) \geq p_1(u_{D+1}, u_2,\ldots, u_D)\] for $\Lambda$-a.a. $u_{D+1} \in f_{\vec{p}}^{-1}([0, f_{\vec{p}}(u_1)])$ and $\Lambda^{\times (D-1)}$-a.a. $(u_2,\ldots, u_D) \in H_{u_1}$. In particular, there exists $2 \leq i \leq D$ and a subset $H_{u_1}' \subset H_{u_1}$ of positive $\Lambda^D$-measure s.t. the choice for $(u_1, u_2, \ldots, u_D)$ is $u_i$ and s.t. for $\Lambda$-a.a. $u_{D+1} \in f_{\vec{p}}^{-1}([0, f_{\vec{p}}(u_1)])$ we have \begin{equation}\label{eqn9} p_1(u_1,u_2,\ldots, u_D) = p_1(u_{D+1}, u_2,\ldots, u_D) = 0 \end{equation} For each such tuple consider the weight of $u_i$ in $(u_1,\ldots, u_{D+1})$; it must be 1 or $D$ and if it were 1 then the choice $(u_1,u_2,\ldots, u_D) \mapsto u_i$ implies none of $u_2,\ldots, u_D$ can have weight $D$ so $u_{D+1}$ must have weight $D$, contradicting $p_1(u_{D+1}, u_2,\ldots, u_D)= 0$. Thus $u_i$ has weight $D$ in $(u_1,\ldots, u_{D+1})$ so we must have the choice $(u_{D+1}, u_2,\ldots, u_D) \mapsto u_i$. Therefore \begin{equation}\label{eqn10} p_1(u_i, u_2, \ldots, u_{i-1}, u_{D+1}, u_{i+1}, \ldots, u_D) = 1 \end{equation} Since \eqref{eqn9} holds for $\Lambda$-a.a. $u_{D+1} \in f_{\vec{p}}^{-1}([0, f_{\vec{p}}(u_1)])$ and $\Lambda^{\times(D-1)}$-a.a. $(u_2,\ldots, u_D) \in H_{u_1}'$, then there is a set of tuples $(u_2,\ldots, u_{D+1})$ of positive $\Lambda^{\times D}$-measure for which \eqref{eqn9} (hence \eqref{eqn10}) holds for both $(u_2, \ldots, u_{D+1})$ and $(u_2,\ldots, u_{i-1}, u_{D+1}, u_{i+1}, \ldots, u_D, u_i)$ (i.e. it holds with the values $u_i, u_{D+1}$ swapped). Combining the two sets of \eqref{eqn9}, \eqref{eqn10} gives the contradiction \[0 = p_1(u_i, u_2, \ldots, u_{i-1}, u_{D+1}, u_{i+1}, \ldots, u_D) = 1 \] (vii) By (v) and (vi), for $\Lambda^{\times D}$-a.a. $(u_1,\ldots,u_D)$, $f_{\vec{p}}(u_i)$ are distinct and \[ p_1(u_k, u_1,\ldots, \widehat{u_k}, \ldots, u_D) \geq \mathbf{1}_{\{f_{\vec{p}}(u_k) \geq f_{\vec{p}}(u_i) \ \forall i\}} \ \forall k\] Consider any such $(u_1,\ldots, u_D)$ and let $u_k$ maximize $f_{\vec{p}}$. Then \[ p_1(u_k, u_1,\ldots, \widehat{u_k}, \ldots, u_D) = \mathbf{1}_{\{f_{\vec{p}}(u_k) \geq f_{\vec{p}}(u_i) \ \forall i\}} = 1\] so the choice for $(u_1,\ldots, u_D)$ is $u_k$. It follows that for all other $j \neq k$, \[ p_1(u_j, u_1,\ldots, \widehat{u_j}, \ldots, u_D) = 0 = \mathbf{1}_{\{f_{\vec{p}}(u_j) \geq f_{\vec{p}}(u_i) \ \forall i\}}\] We have shown that the a.s. inequality in (vi) is an a.s. equality. \end{proof} The next lemma shows that (ii) implies (iii) in Theorem \ref{thm4}. \begin{lemma}\label{Lemma10} Suppose $\vec{p}$ is a single-edge strategy s.t. \begin{enumerate}[label=(\roman*)] \item $f_{\vec{p}}$ is not constant on any set of positive $\Lambda$ measure, and \item $p_k(u_1,\ldots, u_D) = \textbf{1}_{\{f_{\vec{p}}(u_k) \geq f_{\vec{p}}(u_i) \ \forall i\}} \ \mbox{for $\Lambda^{\times D}$-a.a.} \ (u_1,\ldots, u_D) \ \forall 1 \leq k \leq D$ \end{enumerate} Then $\sup_{\phi} E[\sigma_{\vec{p}^{\phi}}] = E[\sigma_{MAX}]$ where the supremum is taken over $\Lambda$-measure-preserving bijections $\phi: [0,1] \rightarrow [0,1]$. \end{lemma} \begin{proof} We begin by constructing a sequence $(\phi_n)$ of $\Lambda$-measure-preserving bijections for which \begin{equation}\label{24} \mathbf{1}_{\{f_{\vec{p}}(\phi_n(u_1)) \geq f_{\vec{p}}(\phi_n(u_i)) \ \forall i\}} \rightarrow \mathbf{1}_{\{u_1 \geq u_i \ \forall i\}} \ \mbox{$\Lambda^{\times D}$-a.s.} \end{equation} Consider any $n \in {\mathbb{N} } $. By (i), inverse images of singletons $f_{\vec{p}}(\{c\})$ have $\Lambda$-measure 0. There exist $ 0 = a_0^n \leq a_1^n \leq \cdots \leq a_{D 2^n}^n = 1$ s.t. \begin{equation}\label{eq24} \Lambda ([a_i^n, a_{i+1}^n)) = \Lambda \bigg(f_{\vec{p}}^{-1} \bigg( \bigg[ \frac{i}{2^n}, \frac{i+1}{2^n} \bigg) \bigg)\bigg) \ \forall 0 \leq i \leq D2^n-1 \end{equation} The indices stop at $D2^n-1$ since the range of $f_{\vec{p}}$ is a subset of $[0,D]$. Clearly, we can make it so that the partition for $n+1$ refines the one for $n$, for all $n$. We construct $\phi_n$ by pasting together $\Lambda$-measure-preserving bijections we get from Theorem \ref{thmNishiura} between sets \[ [a_i^n, a_{i+1}^n) \rightarrow f_{\vec{p}}^{-1} \bigg( \bigg[ \frac{i}{2^n}, \frac{i+1}{2^n} \bigg) \bigg)\] for $0 \leq i \leq D2^n-1$ for which $f_{\vec{p}}^{-1} ( [ \frac{i}{2^n}, \frac{i+1}{2^n} ) )$ has positive $\Lambda$-measure. Then for each such $i$, $f_{\vec{p}^{\phi_n}} = f_{\vec{p}} \circ \phi_n$ is a map \[ [a_i^n, a_{i+1}^n) \rightarrow \bigg[ \frac{i}{2^n}, \frac{i+1}{2^n} \bigg) \] Let us show \eqref{24}. It suffices to show that for $\Lambda^{\times 2}$-a.a. $(u_1, u_2)$ we have \[\mathbf{1}_{\{f_{\vec{p}}(\phi_n(u_1)) \geq f_{\vec{p}}(\phi_n(u_2)) \}} \rightarrow \mathbf{1}_{\{u_1 \geq u_2 \}}\] since multiplying $D-1$ of these sequences of indicators gives $\eqref{24}$. By (i), the sets $f_{\vec{p}}^{-1}(\{c\})$ have $\Lambda$-measure 0 hence it follows that for $\Lambda^{\times 2}$-a.a. $(u_1, u_2)$ we have $f_{\vec{p}}(u_1) \neq f_{\vec{p}}(u_2)$. Consider any such $u_1, u_2$. Since $\Lambda([a_i^n, a_{i+1}^n)) \downarrow 0$ as $n \rightarrow \infty$ (because $\Lambda \ll $ Lebesgue measure), then for large enough $n$, $u_1$ and $u_2$ are in different intervals of the form $[a_i^n, a_{i+1}^n)$ and they stay in these intervals (because the partitions get progressively more refined). Thus $f_{\vec{p}}\circ \phi_n (u_1), f_{\vec{p}}\circ \phi_n (u_2)$ are in different intervals of the form $[\frac{i}{2^n}, \frac{i+1}{2^n})$ and the order is preserved: \[u_1 < u_2 \Leftrightarrow u_1 \in [a_{i_1}^n, a_{i_1+1}^n), u_2 \in [a_{i_2}^n, a_{i_2+1}^n) \ \mbox{with} \ i_1 < i_2 \Leftrightarrow f_{\vec{p}}\circ \phi_n (u_1) < f_{\vec{p}}\circ \phi_n (u_2)\] Therefore we get \eqref{24}. Recall that $\vec{p}^{\phi_n}$ are themselves consistent single-edge strategies. Using (ii) and \eqref{cdf}, it follows that for any $t \in [0,1]$, \begin{equation}\label{limit} \begin{split} 1-F_{\vec{p}^{\phi_n}}(t) &= D\int_{[t,1] \times [0,1]^{D-1}} p_1(\phi_n(\vec{u})) d\Lambda^{D}(u_1, \ldots, u_{D}) \\ & = D\int_{[t,1] \times [0,1]^{D-1}} \mathbf{1}_{\{f_{\vec{p}}(\phi_n(u_1)) \geq f_{\vec{p}}(\phi_n(u_i)) \ \forall i\}} d\Lambda^{D}(u_1, \ldots, u_{D}) \end{split} \end{equation} By \eqref{24} and the Bounded Convergence Theorem, this last expression converges to the integral of the density $f_{MAX}$: \begin{equation*} \begin{split} & D\int_{[t,1] \times [0,1]^{D-1}} \mathbf{1}_{\{f_{\vec{p}}(\phi_n(u_1)) \geq f_{\vec{p}}(\phi_n(u_i)) \ \forall i\}} d\Lambda^{D}(u_1, \ldots, u_{D}) \\ & \rightarrow D\int_{[t,1] \times [0,1]^{D-1}} \mathbf{1}_{\{u_k \geq u_i \ \forall i\}}d\Lambda^{D}(u_1, \ldots, u_{D}) \\ & = \int_{t}^{1} f_{MAX}(u_1) du_1 \\ & = 1- F_{MAX}(t) \end{split} \end{equation*} But we already established that $F_{\vec{q}} \geq F_{MAX}$ for any single-edge strategy $\vec{q}$. Therefore $F_{\vec{p}^{\phi_n}} \rightarrow F_{MAX}$ pointwise from below hence $E[\sigma_{\vec{p}^{\phi_n}}] \rightarrow E[\sigma_{MAX}]$ by the Bounded Convergence Theorem combined with the tail integral formula for expectation. \end{proof} \subsection{Theorem \ref{thm4}} We wish to prove the rest of Theorem \ref{thm4}, but first we make some useful observations about $\sigma_{MAX}$. \begin{lemma}\label{lemma9} Let $f_{MAX}^-$ be the quantile function \[ f_{MAX}^-(t) = \inf \{x: t \leq f_{MAX}(x) \} \] Then treating $f_{MAX}$ itself as a $[0,D]$-valued random variable on the probability space \\ $([0,1], \mathcal{B}([0,1]), \Lambda)$, \[ P(f_{MAX} \leq t) = f_{MAX}^-(t) \ \forall t\] In particular, \[ P(f_{MAX} \leq f_{MAX}(y)) = y \ \forall y \] \end{lemma} \begin{proof} This is trivial since the density $f_{MAX}(y) = Dy^{D-1}$ is invertible so the quantile function $f_{MAX}^-$ is just its inverse. \end{proof} To prove Theorem \ref{thm4}, we need one more short lemma exploring what \\ $\sup_{\phi} E[\sigma_{\vec{p}^{\phi}}] = E[\sigma_{MAX}]$ tells us. \begin{lemma}\label{lemma10} Let $\vec{p}$ be a single-edge strategy and $\phi_n: [0,1] \rightarrow [0,1]$ be $\Lambda$-measure-preserving bijections s.t. $E[\sigma_{\vec{p}^{\phi}}] \uparrow E[\sigma_{MAX}]$. Then there is a subsequence $\phi_{n_j}$ s.t. $F_{\vec{p}^{\phi_{n_j}}} (y) \rightarrow F_{MAX}(y)$, $f_{\vec{p}^{\phi_{n_j}}}(y) = f_{\vec{p}} \circ \phi_{n_j}(y) \rightarrow f_{MAX}(y)$ $\Lambda$-a.e. and almost uniformly on compact subsets of $[0,1]$. \end{lemma} \begin{proof} By the tail integral formula for expectation, \[ \int_{0}^{1} F_{\vec{p}^{\phi_n}}(t)- F_{MAX}(t) dt = E[\sigma_{MAX}] - E[\sigma_{\vec{p}^{\phi}}] \downarrow 0\] where the integrands $F_{\vec{p}^{\phi_n}}(t)- F_{MAX}(t)$ are non-negative from Lemma \ref{lemma3} (i), hence this is $L^1$ convergence. It follows that there exists a subsequence $n_j$ s.t. $F_{\vec{p}^{\phi_{n_j}}}(t) \rightarrow F_{MAX}(t)$ a.e. and a.u. on compact subsets of $[0,1]$. The latter convergence gives us that $f_{\vec{p}^{\phi_{n_j}}}(t) \rightarrow f_{MAX}(t)$ a.e. and a.u. on compact subsets of $[0,1]$. \end{proof} We are now ready to prove Theorem \ref{thm4} in its entirety. \begin{manualtheorem}{\ref{thm4}} Let $\vec{p}$ be a single-edge strategy. The following are equivalent: \begin{enumerate}[label=(\roman*)] \item $\sigma_{\vec{p}}$ is an extreme point \item Any consistent single-edge strategy achieving $\sigma_{\vec{p}}$ must be deterministic \item $f_{\vec{p}}$ is not constant on sets of positive $\Lambda$-measure and $\sigma_{\vec{p}}$ is given by the following single-edge strategy $\vec{q}$: \[ q_k(u_1,\ldots, u_D) = \textbf{1}_{\{f_{\vec{p}}(u_k) \geq f_{\vec{p}}(u_i) \ \forall i\}} \ \mbox{for $\Lambda^{\times D}$-a.a.} \ (u_1,\ldots, u_D) \] In other words, $\sigma_{\vec{p}}$ is achieved by the deterministic single-edge strategy "choose whichever label yields a higher value when evaluating the density $f_{\vec{p}}$" \item $\sup_{\phi} E[\sigma_{\vec{p}^{\phi}}] = E[\sigma_{MAX}]$ \item For $U \sim [0,1]$ if $f_{\vec{p}}(U), f_{MAX}(U)$ as $[0,D]$-valued random variables on the probability space $([0,1], \mathcal{B}([0,1]), \Lambda)$, then they have the same distribution. That is, \[P(f_{\vec{p}} \leq t) = P(f_{MAX} \leq t) = \bigg(\frac{t}D\bigg)^{\frac1{D-1}} \ \forall t \in [0,D]\] \end{enumerate} \end{manualtheorem} \begin{proof} (i) $\Leftrightarrow $ (ii), (i) $\Rightarrow$ (iii), (iii) $\Rightarrow $ (iv) These are given by Lemmas \ref{lemma6}, \ref{lemma8}, \ref{Lemma10}. (iv) $\Rightarrow$ (i) Suppose (iv) holds but $\sigma_{\vec{p}}$ is not an extreme point. Since Radon-Nikodym derivatives are additive, then \[ f_{\vec{p}} = \alpha f_{\vec{q}} + (1-\alpha) f_{\vec{r}}\] for $\alpha \in (0,1)$ and single-edge strategies $\vec{q}, \vec{r}$ s.t. $f_{\vec{p}} \neq f_{\vec{q}}$ on a set of positive $\Lambda$ measure. Take $\phi_n$ $\Lambda$-measure-preserving bijections s.t. $E[\sigma_{\vec{p}^{\phi_n}}] \uparrow E[\sigma_{MAX}]$. Then \[\alpha E[\sigma_{\vec{q}^{\phi_n}}] + (1-\alpha) E[\sigma_{\vec{r}^{\phi_n}}] \uparrow \alpha E[\sigma_{MAX}] + (1-\alpha) E[\sigma_{MAX}] \] where $E[\sigma_{\vec{q}^{\phi_n}}], E[\sigma_{\vec{r}^{\phi_n}}] \leq E[\sigma_{MAX}]$. It follows that \[E[\sigma_{\vec{q}^{\phi_n}}], E[\sigma_{\vec{r}^{\phi_n}}] \uparrow E[\sigma_{MAX}]\] By Lemma \ref{lemma10}, there is a subsequence $\phi_{n_j}$ s.t. \[ f_{\vec{q}} \circ \phi_{n_j}, f_{\vec{r}} \circ \phi_{n_j} \rightarrow f_{MAX} \] $\Lambda$-a.e. and $\Lambda$-a.u. on compact subsets of $[0,1]$. Since $f_{\vec{q}} \neq f_{\vec{r}}$ on a set of positive $\Lambda$-measure then $\exists \epsilon > 0$ s.t. \[Q := \{x: |f_{\vec{q}}(x) - f_{\vec{r}}(x)| \geq \epsilon \} \] has measure $\delta := \Lambda(Q) >0$. Now, by the choice of $\phi_{n_j}$, there is a subset $R \subset [0, 1]$ of $\Lambda$ measure $\leq \frac{\delta}2$ s.t. $f_{\vec{q}}\circ \phi_{n_j}, f_{\vec{r}} \circ \phi_{n_j} \rightarrow f_{MAX}$ uniformly on $[0, 1] \setminus R$. In particular, $\exists J \in {\mathbb{N} } $ s.t. for all $j \geq J$ and $y \in [0, 1] \cap R$ we have \[|f_{\vec{q}} \circ \phi_{n_j}(y) - f_{MAX}(y)|, |f_{\vec{r}} \circ \phi_{n_j}(y) - f_{MAX}(y)| < \frac{\epsilon}2 \Rightarrow |f_{\vec{q}} \circ \phi_{n_j}(y) - f_{\vec{r}} \circ \phi_{n_j}(y) | < \epsilon\] Thus \begin{equation}\label{eq25} \phi_{n_J}([0,1] \setminus R) \subseteq [0,1] \setminus Q \end{equation} But $\phi_{n_J}$ is $\Lambda$-measure-preserving and \begin{align*} \Lambda(\phi_{n_J}([0,1] \setminus R)) &= \Lambda([0,1]) - \Lambda(R)\\ &\geq \Lambda([0,1]-\frac{\delta}2 \\ &> \Lambda([0,1])-\Lambda(Q) \\ &= \Lambda([0,1] \setminus Q) \end{align*} Contradiction of \eqref{eq25}. Therefore $\sigma_{\vec{p}}$ is an extreme point. (iv) $\Rightarrow$ (v) Take $\phi_n$ $\Lambda$-measure-preserving bijections s.t. $E[\sigma_{\vec{p}^{\phi_n}}] \uparrow E[\sigma_{MAX}]$. By Lemma \ref{lemma10} there is a subsequence $\phi_{n_j}$ s.t. $f_{\vec{p}^{\phi_{n_j}}} = f_{\vec{p}} \circ \phi_{n_j} \rightarrow f_{MAX}$ $\Lambda$-a.e and a.u. on compact subsets of $[0,1]$. Now for any $z \in [0,1]$, \[P(f_{\vec{p}} \leq z) = \Lambda(f_{\vec{p}}^{-1}([0,z])) = \Lambda( (f_{\vec{p}^{\phi_{n_j}}})^{-1} ([0,z]))\] We claim this last expression equals $\Lambda(f_{MAX}^{-1}([0,z])) = P(f_{MAX} \leq z)$. Suppose $P(f_{\vec{p}} \leq z) \neq P(f_{MAX} \leq z)$ for some $z$, say $P(f_{\vec{p}} \leq z) < \Lambda(f_{MAX}^{-1}([0,z]))$; the $>$ case is very similar. By continuity from below and since $\Lambda(f_{MAX}^{-1}(\{z\})) = 0$, there exists $ \delta > 0$ s.t. \[\Lambda( f_{\vec{p}} ^{-1} ([0,z])) < \Lambda(f_{MAX}^{-1} ([0, z-\delta]))\] By a.u. convergence on $[0,1]$, there is a subset $R \subset [0, 1]$ off of which $f_{\vec{p}} \circ \phi_{n_j} \rightarrow f_{MAX}$ uniformly s.t. $\Lambda(R)$ is sufficiently small so that for all $j$, \begin{equation}\label{eq26} \Lambda( (f_{\vec{p}} \circ \phi_{n_j})^{-1} ([0,z])) = \Lambda( f_{\vec{p}} ^{-1} ([0,z])) < \Lambda(f_{MAX}^{-1} ([0, z-\delta]) \cap [0,1] \setminus R) \end{equation} Uniform convergence gives a $J$ s.t. for all $j \geq J$ and all $t \in f_{MAX}^{-1} ([0, z-\delta]) \cap [0,1] \setminus R$, \[ f_{\vec{p}} \circ \phi_{n_j}(t) \in (f_{MAX}(t)- \delta, f_{MAX}(t) + \delta) \cap [0,\infty) \subseteq [0, z] \] (here we use the fact that $f_{\vec{p}}$ is a non-negative function). Thus \[f_{MAX}^{-1} ([0, z-\delta]) \cap [0,1] \setminus R \subseteq (f_{\vec{p}} \circ \phi_{n_j}) ^{-1} ([0,z]) \] which is a contradiction of \eqref{eq26}. Therefore \[P(f_{\vec{p}} \leq z) = P(f_{MAX} \leq z) = f_{MAX}^-(z) \] where the last inequality holds by Lemma \ref{lemma9}. The expression on the right is precisely $ (\frac{z}D)^{\frac1{D-1}}$. (v) $\Rightarrow$ (iv) We construct $\Lambda$-measure-preserving bijections $\phi_n : [0,1] \rightarrow [0,1]$ s.t.\\ $E[\sigma_{\vec{p}^{\phi_n}}] \rightarrow E[\sigma_{MAX}]$. Consider any $n$. By (v) combined with Lemma \ref{lemma9}, for any $z \in \{0,1,\ldots, 2^n-1\}$ we have \begin{align*} &\Lambda \bigg( f_{\vec{p}}^{-1} \bigg( \bigg[f_{MAX} \bigg(\frac{z}{2^n} \bigg), f_{MAX} \bigg(\frac{z+1}{2^n} \bigg) \bigg) \bigg) \bigg) \\ &= P\bigg(f_{MAX} \bigg(\frac{z}{2^n} \bigg) \leq f_{\vec{p}} < f_{MAX} \bigg(\frac{z+1}{2^n} \bigg) \bigg) \\ &= P\bigg(f_{MAX} \bigg(\frac{z}{2^n} \bigg) \leq f_{MAX} < f_{MAX} \bigg(\frac{z+1}{2^n} \bigg) \bigg) \\ &= \Lambda\bigg( \bigg[\frac{z}{2^n}, \frac{z+1}{2^n} \bigg) \bigg) \end{align*} We construct $\phi_n$ by pasting together $\Lambda$-measure-preserving bijections we get from Theorem \ref{thmNishiura} between sets \[\bigg[\frac{z}{2^n}, \frac{z+1}{2^n} \bigg) \rightarrow f_{\vec{p}}^{-1} \bigg( \bigg[f_{MAX} \bigg(\frac{z}{2^n} \bigg), f_{MAX} \bigg(\frac{z+1}{2^n} \bigg) \bigg) \bigg) \] Then for each $z$, $f_{\vec{p}^{\phi_n}} = f_{\vec{p}} \circ \phi_n$ is a map \[\bigg[\frac{z}{2^n}, \frac{z+1}{2^n} \bigg) \rightarrow \bigg[f_{MAX} \bigg(\frac{z}{2^n} \bigg), f_{MAX} \bigg(\frac{z+1}{2^n} \bigg) \bigg) \] We now compute the expectation of each $X(\vec{p}^{\phi_n})$: \begin{align*} E[X(p^{MAX}) ] &\geq E[X_0(\vec{p}^{\phi_n}) ] \\ &= \sum_{z=0}^{2^n-1} \int_{\frac{z}{2^n}}^{\frac{z+1}{2^n}} f_{\vec{p}}(\phi_n(t)) t\ dt\\ &\geq \sum_{z=0}^{2^n-1} \int_{\frac{z}{2^n}}^{\frac{z+1}{2^n}} f_{MAX} \bigg(\frac{z}{2^n} \bigg) t \ dt \end{align*} But $f_{MAX}(t) t$ is nondecreasing hence is of bounded variation on $[0,1]$ hence this lower bound converges to $E[X(p^{MAX}) ]$ as $n \rightarrow \infty$. Thus \[ E[\sigma_{MAX}] = E[X(p^{MAX}) ] = \sup_n E[X(\vec{p}^{\phi_n}) ] =\sup_{\phi} E[\sigma_{\vec{p}^{\phi}} ] \] as desired. \end{proof} By the Krein-Milman Theorem, $\mathcal{R}$ is exactly the closed convex hull of its extreme points. Thus we obtain a description of the set of limit points of empirical measures. As a final remark, observe that we only used the compactness of the space $\mathcal{M}_1$ of probability measures on [0,1] to reduce from generic strategies to single-edge strategies. Furthermore, the fact that $\Lambda$ has nice formulas for its cdf and pdf was convenient but unnecessary. The proof of this theorem could be tweaked to hold with $\Lambda$ replaced by an arbitrary distribution $\theta$ on $ {\mathbb{R}} $ with finite mean. We could thus get a similar characterization for the extreme points of $\{\sigma_{\psi}: \mbox{single-edge strategies} \ \psi\}$. The only caveats would be that this set might not coincide with $\{\mbox{limit points of } \ \frac1n \mu_{0 \rightarrow n}(\chi): \ \mbox{strategies} \ \chi\}$ and that the value distribution of the densities $f_{\vec{p}}$ of extreme points might not have as nice a form as $D \cdot Beta(1,D)$. \subsection{The Discrete Case with Example}\label{discrete} The same argument with the weight tuples can be used to show a discrete version of Theorem \ref{thm4}, where the i.i.d. labels $U^j$ are $Unif\{1,\ldots,K\}$. \begin{manualtheorem}{13'}\label{13discrete} Let $\vec{p}$ be a single-edge strategy. The following are equivalent \begin{enumerate}[label=(\roman*)] \item $\sigma_{\vec{p}}$ is an extreme point \item Any consistent single-edge strategy achieving $\sigma_{\vec{p}}$ must be deterministic. \item $\sigma_{\vec{p}}$ is given by the single-edge strategy "choose whichever label is maximal with respect to the ordering $\alpha(1) < \cdots < \alpha(K)$" for some $\alpha \in S_K$. \item There exists a permutation $\beta \in S_K$ s.t. $\vec{p}(\beta(\vec{u})) = \vec{p}^{MAX}(\vec{u}) \ \forall \vec{u}$. \item $\sigma_{\vec{p}}$ has a probability mass function whose value distribution is \[\bigg\{\frac1{K^D}, \frac{2^D-1}{K^D}, \frac{3^D-2^D}{K^D}, \ldots, \frac{K^D-(K-1)^D}{K^D}\bigg\}\] \end{enumerate} \end{manualtheorem} Let us begin by highlighting the main differences between Theorems \ref{thm4} and \ref{13discrete}. In the continuous case, the extreme points are those $\sigma_{\vec{p}}$ achieved by the deterministic single-edge strategy of choosing whichever $u_i$ maximizes the value of the density $f_{\vec{p}}$, or equivalently those $\sigma_{\vec{p}}$ whose density has the same value distribution as $\sigma_{MAX} = Dx^{D-1}dx$ (namely $D \cdot \mbox{Beta}(1,D)$). Contrast this with the discrete case, where the extreme points are those $\sigma_{\vec{p}}$ achieved by the deterministic single-edge strategy of choosing whichever $u_i$ maximizes the value of the probability mass function, or equivalently those $\sigma_{\vec{p}}$ whose pmf has the same value distribution as $\sigma_{MAX} = \frac1{K^D} \delta_1 + \frac{2^D-1}{K^D}\delta_2+ \cdots + \frac{K^D-(K-1)^D}{K^D} \delta_K$. Furthermore, in the discrete case, if we write the extremal $\sigma_{\vec{p}}$ as \[\sigma_{\vec{p}} = \frac1{K^D} \delta_{\alpha(1)} +\frac{2^D-1}{K^D}\delta_{\alpha(2)} + \cdots + \frac{K^D-(K-1)^D}{K^D} \delta_{\alpha(K)} \] for some permutation $\alpha \in S_K$ then clearly the single-edge strategy of choosing whichever $u_i$ maximizes the value of the pmf can be equivalently described as the single-edge strategy of choosing whichever $u_i$ is maximal according to the ordering $\alpha(1) < \alpha(2) < \cdots < \alpha(K)$. It is easy to see that this $\alpha$ satisfies $\vec{p}(\alpha^{-1}(\vec{u})) = \vec{p}^{MAX}(\vec{u})$ (which is where (iv) above comes from). In the discrete setting, the existence of this $\alpha \in S_K$ leads to a natural bijection between the extreme points of the permutohedron and of $\mathcal{R}$, the set of achievable distributions $\{\sigma_{\vec{p}}\}$, by mapping $\sigma_{\vec{p}}$ to the ordering $\alpha(1) < \cdots < \alpha(K)$ associated with it. Since convex combinations of single-edge strategies translate to convex combinations of the distributions $\sigma_{\vec{p}}$ then this bijection extends to a bijection between the permutohedron and $\mathcal{R}$. Of course, the permutohedron is the same for different $D$ but the bijection depends on $D$. It is also worth noting that our proof of Theorem \ref{thm4} adapted to this discrete setting becomes much simpler and more intuitive. Recall we showed that a consistent and deterministic single-edge strategy is one which induces an ordering on the possible values of the edge labels in the sense that $y$ "dominates" $x$ if and only if the choice for $\vec{u}$ is never $x$ if both $x$ and $y$ appear in $\vec{u}$. In the discrete case, there are only $K$ possible values of the edge labels so immediately this gives the ordering $\alpha(1) < \cdots < \alpha(K)$ from which Theorem \ref{13discrete} follows. As an example, let us briefly work through the $D=2, K =4$ discrete case. \begin{center} \captionsetup{type=figure} \includegraphics[scale=0.4]{permutahedron.png} \captionof{figure}{Permutohedron of order 4, \cite{holroyd}} \end{center} The bijection between the extreme points of the permutohedron of order 4 and the extreme points of the set of consistent single-edge strategies goes as follows: \[ 1342 \mapsto \sigma_{1342} = \frac1{16} \delta_1 + \frac{2^2-1}{16} \delta_3 + \frac{3^2-2^2}{16} \delta_4 + \frac{4^2-3^2}{16} \delta_2\] where $\sigma_{1342}$ is the distribution obtained by choosing whichever of the samples $u^1, u^2 \in \{1,2,3,4\}$ is maximal with respect to the ordering $1 < 3 < 4 < 2$. It is plain to see that all the deterministic consistent single-edge strategies correspond to such an ordering, and the corresponding distribution has a pmf with the same value distribution $\{\frac1{16},\frac3{16},\frac5{16},\frac7{16}\}$. \section{Grid Entropy in this Model}\label{grid} In this short section, we compute the grid entropy of the extreme points described in the previous section and we describe grid entropy in general using a simplified formula for the Gibbs Free Energy. \subsection{Grid Entropy of Extreme Points} Recall that the extreme points of $\{\sigma_{\vec{p}}\}$ are given by the single-edge strategy of choosing whichever edge label maximizes the density $f_{\vec{p}}$. We show that such extreme points $\sigma_{\vec{p}}$ have grid entropy 0. However, we first need a short lemma about the partial averages of the expectations of order statistics being bounded away from 0. \begin{lemma}\label{orderStats} Let $Z_i \sim \theta, i \geq 0$ be i.i.d. random variables s.t. $\theta$ is a distribution on $ {\mathbb{R}} $ with cdf $F_{\theta}$ satisfying $F_{\theta}(0) = 0$. Consider the order statistics $Z_{0:n} \leq \cdots \leq Z_{n:n}$. Then there exists a constant $C> 0$ s.t. for large enough $n$, \[ \frac1n \sum_{k=0}^{\lfloor n \epsilon \rfloor} E[Z_{k:n}] \geq C \] \end{lemma} \begin{remark} We allow for the case where some of these expected values may be $\infty$. \end{remark} \begin{proof} Observe that half the terms in this sum are bounded below by $E[Z_{\lfloor \frac{n \epsilon}2 \rfloor:n}]$ hence \[ \frac1n \sum_{k=0}^{\lfloor n \epsilon \rfloor} E[Z_{k:n}] \geq \frac1n \bigg \lfloor \frac{n \epsilon}2 \bigg \rfloor E[Z_{\lfloor \frac{n \epsilon}2 \rfloor:n}] \] Thus it suffices to find $C>0$ s.t. for large $n$, $E[Z_{\lfloor \frac{n \epsilon}2 \rfloor:n}] \geq \frac12 C$. By right-continuity of the cdf $F_{\theta}$, we can take $C > 0$ s.t. $F_{\theta}(C) \leq \frac18 \epsilon$. Consider the i.i.d. Bernoulli random variables $ \textbf{1}_{\{Z_i \leq C\}}$ with success probability $F_{\theta}(C)$. Then Markov's Inequality yields \[P(Z_{\lfloor \frac{n \epsilon}2 \rfloor:n} \leq C) = P \bigg(\#\mbox{successes in Bin($n, F_{\theta}(C)$)} \geq \bigg\lfloor \frac{n \epsilon}2 \bigg \rfloor \bigg) \leq \frac{n F_{\theta}(C)}{\lfloor \frac{n \epsilon}2 \rfloor} \leq \frac12 \] for large $n$ by the choice of $C$. It follows that \[ E[Z_{\lfloor \frac{n \epsilon}2 \rfloor:n}] \geq \frac12 C \] for large $n$, which completes the proof. \end{proof} \begin{theorem}\label{entropy0} Fix $\tau: [0,1] \rightarrow {\mathbb{R}} $ measurable and bounded s.t. $\tau$ is not constant on sets of positive $\Lambda$ measure, and consider the single-edge strategy of picking the maximal $\tau(U^j)$ over $1 \leq j \leq D$, given by \[ \vec{p}(u_1,\ldots,u_D) = \textbf{1}_{\{\tau(u_k) \geq \tau(u_i) \ \forall i\}} \] Then $||\sigma_{\vec{p}}|| = 0$. \end{theorem} \begin{remark} The types of strategies considered in this theorem are deterministic so the resulting $\sigma_{\vec{p}}$ are all extreme points by Theorem \ref{thm4}. On the other hand, Theorem \ref{thm4} establishes that all extreme points can be realized as $\sigma_{\vec{p}}$ for a single-edge strategy $\vec{p}$ choosing whichever observed label maximizes the density $f_{\vec{p}} \in [0,D]$ where $f_{\vec{p}}(\mbox{Unif}[0,1])$ is the $D \cdot \mbox{Beta}(D,1)$ distribution. Thus Theorem \ref{entropy0} captures what happens for all extreme points. \end{remark} \begin{proof} Suppose $||\sigma_{\vec{p}}|| > \delta > 0$. Let $\alpha = \frac{\delta}{\ln 2}$. For $n \in {\mathbb{N} } $ and $ 1 \leq m_n \leq \lfloor e^{n \delta} \rfloor = \lfloor 2^{n \alpha} \rfloor$ consider the event-dependent paths $\pi_{n,m_n}$ corresponding to \[ \min_{\pi: 0\rightarrow n}^{m_n} \rho\bigg(\frac1n \mu_{\pi}, \sigma_{\vec{p}} \bigg) \] Since $\delta < ||\sigma_{\vec{p}}||$ then by definition of grid entropy, \[ \min_{\pi: 0\rightarrow n}^{\lfloor e^{n\delta} \rfloor } \rho\bigg(\frac1n \mu_{\pi}, \sigma_{\vec{p}} \bigg) \rightarrow 0 \ \mbox{a.s.}\] hence $\frac1n \mu_{\pi_{n, m_n}} \Rightarrow \sigma_{\vec{p}}$ a.s. regardless of the sequence $(m_n)$. Also for $i \geq 0$ define the random variables \[Y_i := \max \limits_{1 \leq j \leq D} (\tau(U_i^j)), Y_i' = \mbox{2nd} \ \max \limits_{1 \leq j \leq D} (\tau(U_i^j)), Z_i := Y_i - Y_i'\] Now it is a classic result that for $\epsilon > 0$, \[ \sum_{i=0}^{\lfloor n\epsilon \rfloor} \binom{n}{i} \leq 2^{nL} \ \mbox{with} \ L = L(\epsilon)= \epsilon \log \epsilon + (1-\epsilon) \log (1-\epsilon) \] Take $\epsilon > 0$ small enough so that $L +\epsilon \log(D-1) < \frac12 \alpha$ and take $N \in {\mathbb{N} } $ so that $ \forall n \geq N$, $2^{nL} (D-1)^{n\epsilon} < \lfloor 2^{n \alpha}\rfloor$. Consider any $n \geq N$. Then the number of paths $\pi: 0 \rightarrow n$ with $< \lceil n\epsilon \rceil$ of its edges not having the maximal edge label in their trial is at most \[ \sum_{i=0}^{\lfloor n\epsilon \rfloor} \binom{n}{i} (D-1)^{i} \leq 2^{nL} (D-1)^{n\epsilon} < \lfloor 2^{n \alpha} \rfloor \] By the Pigeonhole Principle, there is an event-dependent path $\pi_{n, m_n}: \vec{0} \rightarrow n$ s.t. at least $\lceil n\epsilon \rceil$ of its edges do not have the maximal edge label in their trial. Let $\pi_{n, m_n}$ have edges $e_0^{j_0}, \ldots, e_{n-1}^{j_{n-1}}$ and let $I_n \subseteq \{0,\ldots,n-1\}, |I_n| = \lceil n\epsilon \rceil$ be a set of indices $i$ for which $\tau(U_i^{j_i}) \neq Y_i$. We compute an upper bound for the passage time along $\pi_{n, m_n}$ (with respect to $\tau$) by splitting the sum over the edges with index in $I_n$ and those not in $I_n$: \begin{align*} \langle \tau, \frac1n \mu_{\pi_{n,m_n}} \rangle &= \frac1n\sum_{i=0}^{n-1} \tau(U_i^{j_i}) \\ & \leq \frac1n \sum_{i\in I_n^C} \max_{1\leq j \leq D} (\tau(U_i^j)) + \frac1n \sum_{i \in I_n} \mbox{2nd} \max_{1\leq j \leq D} (\tau(U_i^j)) \\ &= \frac1n \sum_{i\in I_n^C} Y_i + \frac1n \sum_{i \in I_n} Y_i' \end{align*} On the other hand, the passage time along the $\tau$-optimal path $\pi_{0\rightarrow n}(\vec{p})$ is \begin{equation*} \langle \tau, \frac1n \mu_{\pi_n(\vec{p})} \rangle = \frac1n\sum_{i=0}^{n-1} \max_{1\leq j \leq D} (\tau(U_i^j)) = \frac1n\sum_{i=0}^{n-1} Y_i \end{equation*} hence \begin{equation}\label{lowerB} \langle \tau, \frac1n \mu_{\pi_n(\vec{p})} \rangle - \langle \tau, \frac1n \mu_{\pi_{n,m_n}} \rangle \geq \frac1n \sum_{i \in I_n} (Y_i - Y_i') = \frac1n \sum_{i \in I_n} Z_i \geq \frac1n \sum_{k=0}^{\lfloor n\epsilon \rfloor} Z_{k:n} \end{equation} Now $Z_i$ are i.i.d., non-negative and satisfy \begin{align*} P(Z_i = 0) &= P\bigg(\max \limits_{1 \leq j \leq D} (\tau(U_i^j)) = \mbox{2nd} \ \max \limits_{1 \leq j \leq D} (\tau(U_i^j))\bigg) \\ &\leq P(\exists 1 \leq j_1 < j_2 \leq D \ \mbox{s.t.} \ \tau(U_i^{j_1}) = \tau(U_i^{j_2}) )\\ &=0 \end{align*} since $\tau$ is not constant on sets of positive $\Lambda$ measure. Thus we can take the expectation in \eqref{lowerB} and apply Lemma \ref{orderStats} to get that $\exists C > 0$ s.t. \begin{equation}\label{EQ21} E \bigg[ \langle \tau, \frac1n \mu_{\pi_n(\vec{p})} \rangle -\langle \tau, \frac1n \mu_{\pi_{n,m_n}} \rangle \bigg] \geq \frac1n \sum_{k=0}^{\lfloor n\epsilon \rfloor} E[Z_{k:n}] \geq C \end{equation} for large $n$. Now, by assumption, $\frac1n \mu_{\pi_{n,m_n}} \Rightarrow \sigma_{\vec{p}}$ a.s.. On the other hand, by nature of the model, the $\tau$-optimal length $m$ path $\pi_{m}(\vec{p})$ contains the $\tau$-optimal length $n$ path $\pi_n(\vec{p})$ for any $m \geq n$; thus we can apply the Glivenko-Cantelli Theorem (Theorem \ref{thm1}) to get that the empirical measures $\frac1n \mu_{\pi_{n}(\vec{p})} $ converge weakly to $\sigma_{\vec{p}}$ a.s.. Recall from Section \ref{coupling} that a.s. the pushforward $\tau_{\ast}$ preserves weak limits of empirical measures so \begin{align*} \langle \tau, \frac1n \mu_{\pi_n(\vec{p})} \rangle -\langle \tau, \frac1n \mu_{\pi_{n,m_n}} \rangle &= \langle 1, \tau_{\ast}(\frac1n \mu_{\pi_n(\vec{p})}) \rangle -\langle 1, \tau_{\ast}(\frac1n \mu_{\pi_{n,m_n}}) \rangle \\ &\rightarrow \langle 1, \tau_{\ast}(\sigma_{\vec{p}}) \rangle - \langle 1, \tau_{\ast}(\sigma_{\vec{p}}) \rangle \\ &= 0 \end{align*} a.s.. But $\tau$ is bounded so by the Bounded Convergence Theorem we get \[E \bigg[ \langle \tau, \frac1n \mu_{\pi_n(\vec{p})} \rangle -\langle \tau, \frac1n \mu_{\pi_{n,m_n}} \rangle \bigg] = 0\] which contradicts \eqref{EQ21}. Thus $||\sigma_{\vec{p}}||=0$. \end{proof} \subsection{Grid Entropy via Gibbs Free Energy} Fix $\beta > 0$. Suppose $\tau: [0,1] \rightarrow {\mathbb{R}} $ is a bounded measurable function. From the definition of $\beta$-Gibbs Free Energy, \[ G^{\beta}(\tau) := \lim_{n \rightarrow \infty} \frac1n \log \sum_{\pi: 0 \rightarrow n} e^{\beta T(\pi)} = \lim_{n \rightarrow \infty} \frac1n \sum_{i=0}^{n-1} \log \sum_{j=1}^D e^{\beta \tau(U_i^j)} \ \mbox{a.s.} \] But $\log \sum \limits_{j=1}^D e^{\beta \tau(U_i^j)}$ are i.i.d. in $i$ hence by the SLLN, \[ G^{\beta} (\tau) = E \bigg[\log \sum_{j=1}^D e^{\beta \tau(U^j)} \bigg] \ \mbox{a.s.}\] where $U^j$ are i.i.d. Unif[0,1]. Of course, grid entropy is simply the negative convex conjugate of $\beta$-Gibbs Free Energy by Theorem \ref{gridEntropyPart1}: \begin{equation} -||\nu|| = \sup_{\tau} \bigg[ \beta \langle \tau, \nu \rangle - G^{\beta}(\tau) \bigg] = \sup_{\tau} \bigg[\beta \langle \tau, \nu \rangle - E \bigg[\log \sum_{j=1}^D e^{\beta \tau(U^j)} \bigg] \bigg] \ \forall \nu \in \mathcal{M} \end{equation} where the supremum is over bounded measurable functions $\tau:[0,1] \rightarrow {\mathbb{R}} $ and where $\langle \tau, \nu \rangle$ denotes the integral $ \int_0^1 \tau(u) d\nu$. \section{Next Steps} We have characterized the extreme points of the set of limit points of empirical measures and have shown that these extreme points have grid entropy 0. A natural next question is whether only the extreme points have grid entropy 0. Recalling that grid entropy is concave, another key question to ask is whether it is strictly concave in the model used in this paper; this might provide insights on whether it is strictly concave in general, which would imply that the Gibbs Free Energy is strictly concave, a major open research problem. Furthermore, it would be nice to compute grid entropy of non-extremal points, and also try to describe the subset of $\mathcal{R}$ of maximizers for the variational formula for the Gibbs Free Energy presented in Section 5.2 of \cite{gatea}. These are all questions well worth exploring in the future. \section{Acknowledgments} I would like to thank my friends and family for their continual support during my graduate studies. Special thanks goes to my advisor B\'{a}lint Vir\'{a}g, for his patience, guidance, and optimism in the face of setbacks. Finally, this work would not have been possible without the funding from my NSERC Canadian Graduate Scholarship-Doctoral.

59d213ffa170f7434109908be864ccce4fe99c28

\section{Introduction} A polyp is a grossly visible mass of epithelial cells that protrudes from the mucosal surface into the lumen of the intestine. Most colon polyps are harmless. However, some colon polyps can develop into colon cancer over time, which is fatal when found in its late stages. Hence, it is essential to have regular screenings. Colonoscopy is one of the most common techniques used to detect colon polyps. With the recent success of deep learning, various image segmentation methods have been employed to help oncologists save significant time. However, deep learning models often require a large amount of labeled data. In real-world scenarios, this can be difficult, especially for medical data. Producing a labeled medical dataset for training semantic segmentation models requires the laborious creation of pixel-level labels and annotators' expertise and medical knowledge in medical image diagnosis. Semi-supervised learning is a method that utilizes a massive amount of unlabeled data for training deep neural networks on actual tasks. This method is significant with medical image data since the cost of labeling the data is often high and requires much effort from many experts. Many studies applied semi-supervised learning on medical image data, such as pseudo labeling \cite{wang2022semi}, cross pseudo supervision \cite{zhang2022semi}, few-shot learning \cite{li2021few}, deep adversarial learning, and so on. These methods use both labeled and unlabeled data to train deep learning models for the semantic segmentation task. Most of them aim to generate high-quality pseudo labels and have more robust representations of the domain distribution of the data. Previous works \cite{araslanov2021self} have shown the effectiveness of the momentum network - a slow copy version of the training model by taking an exponential moving average. Empirical results show that the momentum network often produces more stable results than the original model trained directly through the back-propagation process. In current semi-supervised learning methods, pseudo label generation can be done online as in \cite{araslanov2021self} or offline as in \cite{yang2020fda, zou2019confidence}. The advantage of online pseudo labeling is that it is simple and easy to implement with only one training stage. In addition, this method will be effective when the pseudo label generation model is updated online to help enhance the quality of pseudo labels after each iteration. This paper combines online pseudo labeling and momentum networks to improve the quality of pseudo labels generated after each iteration. This method improves Dice Score by 3\% compared with offline pseudo labeling and outperforms supervised models in some out-of-domain datasets. Our main contributions in this paper are: \begin{itemize} \item We propose a novel training strategy for semi-supervised learning which combines online pseudo labeling and momentum networks for the polyp segmentation task. Our results on Dice Score are about 3\% better than the offline pseudo labeling method and outperform supervised models in some out-of-domain datasets. \item Comprehensive analysis of the effects of online pseudo labeling and momentum network on our results. \item To the best of our knowledge, this is the first study to combine online pseudo labeling and momentum network for polyp segmentation task. \end{itemize} The rest of this paper is organized as follows. Related work is presented in Section 2. The proposed method is presented in Section 3. Section 4 describes the datasets and details implementation methods. The experimental results are presented in Section 5. Finally, discussions and future research directions are presented in Section 6. \section{Related Work} \subsection{Semantic Segmentation} Semantic segmentation aims to classify images at the pixel level. Since the Fully Convolutional Neural Network was applied to semantic segmentation in the study of Long et al. \cite{long2015fully}, deep learning models have achieved remarkable results in this field. Currently, most segmentation models are designed based on the encoder-decoder architecture \cite{long2015fully}. A few studies also aim to build a specific model for polyp segmentation. HarDNet-MSEG \cite{huang2021hardnet} achieved high performance on the Kvasir-SEG dataset with processing speed up to 86 FPS. AG-ResUNet++ improved UNet++ with attention gates and ResNet backbone. Another study called Transfuse combined Transformer and CNN using BiFusion module \cite{zhang2021transfuse}. ColonFormer \cite{duc2022colonformer} used MiT backbone, Upper decoder, and residual axial reverse attention to further boost the polyp segmentation accuracy. NeoUNet \cite{ngoc2021neounet} and BlazeNeo \cite{an2022blazeneo} proposed effective encoder-decoder networks for polyp segmentation and neoplasm detection. Generally, research in changing model architecture is still a potential approach. However, this approach often requires a large amount of labeled data. Collecting dense pixel-level annotations for semantic segmentation is costly and difficult, especially for medical segmentation data. Therefore, applying an efficient learning method to help deal with the lack of data is critical for practical. \subsection{Semi supervised learning for semantic segmentation} One of the disadvantages of supervised learning is the requirement of massive amounts of labeled data. This problem becomes even more difficult with medical image data such as polyp segmentation because it requires effort from people with experience and expertise in imaging diagnostic. Semi-supervised learning solves this problem by combining labeled and unlabelled data during the training model. Several studies have applied semi-supervised learning to semantic segmentation problems, such as adversarial learning, consistency regularization, or pseudo-labeling \cite{feng2020semi}. General objective of semi-supervised learning is improving base model trained on dataset of $N$ labelled samples $D_{sup}=\left\{ (x_n, y_n) | n=1...N \right\}$ by utilizing a dataset of M unlabeled samples $D_{unsup}=\left\{ x_m | m=N+1...N+M \right\}$. In this paper, we focus on improving the quality of pseudo labels by combining the online pseudo labeling strategy with momentum networks. \subsection{Momentum Network} A momentum network is a slow copy version of weights of the original model during the training process through Exponential Moving Average (EMA). The update process can be done after each epoch or after a certain number of steps in the training process. $$\theta'_t=\alpha\theta'_{t-1} + (1 - \alpha)\theta_t$$ where $\theta'_t$ and $\theta_t$ are the weights of the momentum and original models, respectively, at the $t$-th step during the training process. The studies applying momentum network have shown experimentally that momentum network gives more stable accuracy than the original model \cite{araslanov2021self}. Accordingly, the momentum model can be considered an ensemble of several versions of the original model at different time steps during training. \subsection{Online vs. offline pseudo label generation} Pseudo-label generation methods can be divided into two types: online generation and offline generation. Online generation generates pseudo labels directly during the forwarding process \cite{sohn2020fixmatch}. The offline generation will generate pseudo labels once in each semi-supervised training, and pseudo labels will only change in the subsequent iterations \cite{iscen2019label, chen2020big}. The advantage of offline learning is that the quality of the pseudo label will not change in each iteration, and the more iterations we train, the better the quality of the pseudo label will be. However, offline pseudo label generation requires more complicated installation and computational costs than online pseudo label generation. The installation is straightforward with the online pseudo label generation method and only requires one-stage training with the student model. However, the online generation method demands a stable quality of the pseudo label during the label generation process. That needs the pseudo-label generation model to be smooth. We solve this problem by using a momentum network during training semi-supervised learning. \section{Proposed method} In this section, we present an online prototyping strategy that ensures the stability of fake label quality during prototyping with a combination of online pseudo labeling and momentum networks. Details of the available pipeline, loss function configurations as well as augmentation methods for each data type are also detailed in this section. \subsection{Overall pipeline} The overall architecture of the system is described in Fig.~\ref{fig:pipeline}. We still apply a two-stage strategy for semi-supervised training, where the teacher model is first trained on the labeled dataset. During the training process, both the original model and the slow copy version of the model updated by EMA (Momentum teacher network) are saved and served for student training. We perform the online pseudo label generation during the student training process using the momentum teacher model. At the same time, the weights of the original student model are updated for the momentum teacher and its student momentum version. Finally, the momentum student version will be used to predict the outcome after the training finishes. Details of each training step will be described in the sections below. \begin{figure*}[ht \centering \includegraphics[width=0.8\linewidth]{images/pipeline.png} \caption{Overview of online pseudo labeling with momentum network pipeline} \label{fig:pipeline} \end{figure*} \subsubsection{Teacher training} The training teacher model is performed on the entire labeled dataset. We use the Tversky loss function during training. Besides, we use the FPN architecture with the DenseNet169 backbone as the primary model for our experiment. The best momentum teacher network on the validation set will be used to generate pseudo labels during student training. \subsubsection{Student training} After having the momentum teacher model trained on the dataset labeled $D_{sup}$, we proceed to train the student model with the pseudo label generated from the momentum teacher model. The difference is that the weights of the momentum teacher model will also be updated during training after a certain number of training steps. This strategy helps the fake labels to be constantly updated and stable due to the nature of the momentum network. The detailed algorithm is presented in \textbf{Algorithm} \ref{algo:main_stategy}. \begin{algorithm} \caption{Online Pseudo Labeling with Momentum Teacher Algorithm} \label{algo:main_stategy} \textbf{Input}: Labeled images $D_{sup}$, Unlabeled images $D_{unsup}$, Trained momentum teacher $MT_0$ \\ \textbf{Output}: Best momentum student $MS_{best}$ trained with combination of supervised and unsupervised data \begin{algorithmic} \Procedure{OnlineLabelingMomentum}{} \For {$t=0$ to $n\_epochs$} \State \textbf{Step 1:} Generate pseudo labels of $D_{unsup}$ with $MT_t$ model. \State \textbf{Step 2:} Train student $S_t$ with combination of $D_{sup}$ and $D_{unsup}$ with generated pseudo labels. \State \textbf{Step 3:} Update momentum teacher $MT_{t}$ with $S_t$ weights via EMA. \State \textbf{Step 3:} Update momentum student $MS_{t}$ with $S_t$ weights via EMA. \State \textbf{Step 4:} Save the best models in validation set. \EndFor \EndProcedure \end{algorithmic} \end{algorithm} \subsection{Data Augmentation} \subsubsection{Weak Augmentation on labeled dataset} We choose two different data augmentation strategies for labeled and unlabelled data. We apply the weak or basic augmentation methods to the labeled dataset. For the weak augmentation, the training images are random flips with a probability of 0.5. \subsubsection{Strong Augmentation on unlabeled dataset} When applied to unlabeled data, noise has the essential benefit of enforcing invariance in the decision function on both labeled and unlabeled data. When the student is deliberately noised, it is trained to be consistent with the teacher that is not noised when it generates pseudo labels. We only use the input noise in our experiments via a strong augmentation method. We use the combination of ShiftScaleRotate, RGBShift, RandomBrightnessContrast, and RandomFlip with the probability of 0.5. \section{Dataset and Experiments Setup} \subsection{Dataset} To compare the method results, we use the same Polyp dataset as in the study of HardDet-MSEG \cite{huang2021hardnet}. This dataset includes 1450 endoscopic images of size $384 \times 288 \times 3$ corresponding to 1450 masks of size $384 \times 288 \times 1$. The mask is a binary image, the pixels have a value of 255 corresponding to the area containing the polyp, and the pixels in background areas have a value of 0. The training dataset includes 900 images in Kvasir-SEG and 550 in the CVC-ClinicDB. The test dataset includes 798 images synthesized from different data sets, which are the CVC-300 dataset, CVC-ClinicDB, CVC-ColonDB, ETIS- LaribPolypDB and Kvasir-SEG. Our experiment uses a 10\% dataset equivalent to 145 images as the validation dataset. The remaining total of 1305 images was used to scale the labeled and unlabeled datasets in each experiment. \subsection{Evaluation Metrics} \textbf{Mean IoU} and \textbf{Mean Dice} are used to evaluate the model and compare our experiments. There are defined as follows: \begin{raggedleft} \begin{tabular}{p{4cm}p{3.5cm}} $$mIoU = \frac{TP}{TP + FP + FN}$$ & $$mDice = \frac{2*TP}{2*TP + FP + FN}$$ \end{tabular} \end{raggedleft} \subsection{System configuration} Our experiments are conducted on a computer with Intel Core i5-7500 CPU @3.4GHz, 32GB of RAM, GPU GeForce GTX 1080 Ti, and 1TB SSD hard disk. The models are implemented with the PyTorch Lightning framework. \section{Result of Experiments} This section presents teacher training results on the labeled dataset and semi-supervised learning with offline and online pseudo-labeling. We use the base model and its momentum network for the above experiments to compare the results. We also divide the dataset with different ratios of labeled data to measure the effect of the amount of labeled data on our method. \subsection{Training teacher model in a supervised manner in labeled data} First, we train the model on the labeled set. Original model weights are optimized by Adam optimizer with back-propagation algorithms. The slow copy of the original model is updated simultaneously with the momentum ratio is 0.95. We save the best checkpoint of both models in the validation set during training as the final teacher model. Table \ref{tab:compare_teacher} shows the performance comparison between the original teacher model and its momentum network. We found that the momentum model gives better results than the original model. Therefore, we choose the momentum teacher model as the basis for the semi-supervised experiments presented in the following sections. \begin{table*}[h] \centering \caption{A comparison of the original teacher with the momentum model teacher in different ratios of labeled data} \def1.1{1.1} \begin{tabular}{|c|c|cc|cc|cc|cc|cc|cc|} \hline \textbf{Ratio of} & \textbf{Momentum} & \multicolumn{2}{c|}{\textbf{CVC-ClinicDB}} & \multicolumn{2}{c|}{\textbf{ETIS}} & \multicolumn{2}{c|}{\textbf{CVC-ColonDB}} & \multicolumn{2}{c|}{\textbf{CVC-300}} & \multicolumn{2}{c|}{\textbf{Kvarsir}} & \multicolumn{2}{c|}{\textbf{Average}} \\ \cline{3-14} \textbf{labeled data} & \textbf{teacher} & \textbf{mIoU} & \textbf{mDice} & \textbf{mIoU} & \textbf{mDice} & \textbf{mIoU} & \textbf{mDice} & \textbf{mIoU} & \textbf{mDice} & \textbf{mIoU} & \textbf{mDice} & \textbf{mIoU} & \textbf{mDice} \\ \hline \multirow{2}{*}{20\%} & - & 0.794 & 0.856 & 0.618 & 0.698 & 0.625 & 0.702 & 0.803 & 0.875 & 0.819 & 0.883 & 0.732 & 0.803 \\ & \checkmark & 0.792 & 0.854 & 0.616 & 0.699 & 0.616 & 0.693 & 0.800 & 0.875 & 0.820 & 0.884 & 0.730 & 0.801 \\ \hline \multirow{2}{*}{40\%} & - & 0.787 & 0.857 & 0.587 & 0.674 & 0.624 & 0.700 & 0.824 & 0.892 & 0.817 & 0.879 & 0.728 & 0.808 \\ & \checkmark & 0.803 & 0.868 & 0.616 & 0.694 & 0.642 & 0.716 & 0.831 & 0.898 & 0.819 & 0.876 & 0.743 & 0.811 \\ \hline \multirow{2}{*}{60\%} & - & 0.835 & 0.889 & 0.610 & 0.677 & 0.645 & 0.721 & 0.820 & 0.885 & 0.842 & 0.894 & 0.751 & 0.814 \\ & \checkmark & 0.850 & 0.902 & 0.620 & 0.685 & 0.663 & 0.741 & 0.839 & 0.904 & 0.847 & 0.898 & 0.764 & 0.826 \\ \hline \end{tabular} \label{tab:compare_teacher} \end{table*} \subsection{Training student with offline pseudo labeling with semi-supervised manner} We use the momentum teacher model for our semi-supervised learning experiments. The student model has the same architecture as the teacher model. In the offline pseudo labeling strategy, the pseudo labels will be generated directly from the momentum teacher model. The momentum teacher model did not change during training. That means the pseudo-labels for the same image will be the same in all training iterations. The student model kept the original and momentum versions as the teacher training on the labeled dataset. \subsection{Training student with online pseudo labeling with semi-supervised manner} Online pseudo labeling follows the same training strategy as offline labeling. The only difference is that the Teacher model will also be updated via EMA with the weights of the student model after each epoch. Since then, the pseudo label for an image is also continuously updated during model training. The online learning method combined with the momentum student model gives better results than offline learning. The results are shown in Table \ref{tab:compare_online_offline} and illustrated in Fig. \ref{fig:visualize_result}. \begin{table*}[h] \centering \caption{A comparison of the online pseudo labeling and offline pseudo labeling strategy for semi-supervised training} \def1.1{1.1} \resizebox{\textwidth}{!}{ \begin{tabular}{|c|c|c|cc|cc|cc|cc|cc|cc|} \hline \textbf{Ratio of} & \textbf{Online} & \textbf{Momentum} & \multicolumn{2}{c|}{\textbf{CVC-ClinicDB}} & \multicolumn{2}{c|}{\textbf{ETIS}} & \multicolumn{2}{c|}{\textbf{CVC-ColonDB}} & \multicolumn{2}{c|}{\textbf{CVC-300}} & \multicolumn{2}{c|}{\textbf{Kvarsir}} & \multicolumn{2}{c|}{\textbf{Average}} \\ \cline{4-15} \textbf{labeled data} & \textbf{pseudo-labels} & \textbf{student} & \textbf{mIoU} & \textbf{mDice} & \textbf{mIoU} & \textbf{mDice} & \textbf{mIoU} & \textbf{mDice} & \textbf{mIoU} & \textbf{mDice} & \textbf{mIoU} & \textbf{mDice} & \textbf{mIoU} & \textbf{mDice} \\ \hline \multirow{4}{*}{20\%} & - & - & 0.789 & 0.847 & 0.590 & 0.670 & 0.647 & 0.727 & 0.821 & 0.891 & 0.832 & 0.891 & 0.736 & 0.805 \\ & - & \checkmark & 0.830 & 0.887 & 0.676 & 0.754 & 0.676 & 0.755 & 0.828 & 0.897 & 0.843 & 0.897 & 0.770 & 0.838 \\ & \checkmark & - & 0.801 & 0.857 & 0.673 & 0.748 & 0.641 & 0.717 & 0.830 & 0.897 & 0.835 & 0.895 & 0.756 & 0.823 \\ & \checkmark & \checkmark & 0.816 & 0.870 & 0.701 & 0.772 & 0.669 & 0.742 & 0.836 & 0.903 & 0.856 & 0.911 & 0.778 & 0.841 \\ \hline \multirow{4}{*}{40\%} & - & - & 0.792 & 0.850 & 0.628 & 0.704 & 0.650 & 0.733 & 0.833 & 0.902 & 0.813 & 0.875 & 0.743 & 0.813 \\ & - & \checkmark & 0.825 & 0.882 & 0.673 & 0.749 & 0.671 & 0.745 & 0.824 & 0.894 & 0.837 & 0.893 & 0.766 & 0.833 \\ & \checkmark & - & 0.824 & 0.883 & 0.601 & 0.671 & 0.668 & 0.750 & 0.829 & 0.896 & 0.838 & 0.899 & 0.752 & 0.820 \\ & \checkmark & \checkmark & 0.825 & 0.882 & 0.702 & 0.777 & 0.689 & 0.768 & 0.825 & 0.895 & 0.850 & 0.908 & 0.778 & 0.846 \\ \hline \multirow{4}{*}{60\%} & - & - & 0.833 & 0.888 & 0.617 & 0.686 & 0.651 & 0.725 & 0.802 & 0.871 & 0.842 & 0.899 & 0.749 & 0.814 \\ & - & \checkmark & 0.856 & 0.905 & 0.700 & 0.773 & 0.685 & 0.763 & 0.839 & 0.904 & 0.865 & 0.915 & 0.789 & 0.852 \\ & \checkmark & - & 0.832 & 0.881 & 0.652 & 0.724 & 0.674 & 0.752 & 0.819 & 0.880 & 0.856 & 0.910 & 0.767 & 0.830 \\ & \checkmark & \checkmark & 0.855 & 0.901 & 0.694 & 0.772 & 0.701 & 0.777 & 0.833 & 0.898 & 0.865 & 0.916 & 0.790 & 0.853 \\ \hline \end{tabular} } \label{tab:compare_online_offline} \end{table*} \subsection{Comparison with some difference supervised methods} We evaluated our model’s performance with a benchmark consisting of the state-of-the-art models, namely UNet \cite{ronneberger2015u}, UNet++ \cite{zhou2019unet++}, SFA \cite{fang2019selective}, PraNet \cite{pranet}, MSNet \cite{zhao2021automatic} and Shallow Attention \cite{wei2021shallow}. Table \ref{tab:compare_with_supervised} shows our model’s performance results for mIoU and mDice metrics compared to the results of the benchmark studies. Our model outperformed UNet, UNet++, PraNet, SFA, and Shallow Attention on all datasets for Dice and IoU metrics with only using a maximum of 60\% of labeled data. We are only about 2\% mDice worse than MSNET on the CVC-ClinicDB set but much better on the rest of the datasets. In particular, our method gives better generalization. Our method outperforms all above-supervised methods when tested in out-of-distribution datasets such as ETIS-LabribPolypDB, CVC-300, and CVC-ColonDB. \begin{table*}[h] \centering \caption{A comparison of our method with state-of-the-art supervised models} \def1.1{1.1} \begin{tabular}{|c|c|cc|cc|cc|cc|cc|} \hline \multirow{2}{*}{\textbf{Methods}} & \textbf{Ratio of} & \multicolumn{2}{c|}{\textbf{CVC-ClinicDB}} & \multicolumn{2}{c|}{\textbf{ETIS}} & \multicolumn{2}{c|}{\textbf{CVC-ColonDB}} & \multicolumn{2}{c|}{\textbf{CVC-300}} & \multicolumn{2}{c|}{\textbf{Kvarsir}} \\ \cline{3-12} & \textbf{labeled data} & \textit{\textbf{mIOU}} & \textit{\textbf{mDice}} & \textit{\textbf{mIOU}} & \textit{\textbf{mDice}} & \textit{\textbf{mIOU}} & \textit{\textbf{mDice}} & \textit{\textbf{mIOU}} & \textit{\textbf{mDice}} & \textit{\textbf{mIOU}} & \textit{\textbf{mDice}} \\ \hline Unet \cite{ronneberger2015u} & 100\% & 0.755 & 0.823 & 0.335 & 0.398 & 0.444 & 0.512 & 0.627 & 0.710 & 0.746 & 0.818 \\ Unet++ \cite{zhou2019unet++} & 100\% & 0.729 & 0.794 & 0.344 & 0.401 & 0.410 & 0.483 & 0.624 & 0.707 & 0.743 & 0.821 \\ SFA \cite{fang2019selective} & 100\% & 0.607 & 0.700 & 0.217 & 0.297 & 0.347 & 0.469 & 0.329 & 0.467 & 0.611 & 0.723 \\ PraNet \cite{pranet} & 100\% & 0.849 & 0.899 & 0.567 & 0.628 & 0.640 & 0.709 & 0.797 & 0.871 & 0.840 & 0.898 \\ MSNET \cite{zhao2021automatic} & 100\% & \textbf{0.879} & \textbf{0.921} & 0.664 & 0.719 & 0.678 & 0.755 & 0.807 & 0.869 & 0.862 & 0.907 \\ Shallow Attention \cite{wei2021shallow} & 100\% & 0.859 & 0.916 & 0.654 & 0.750 & 0.670 & 0.753 & 0.815 & 0.888 & 0.847 & 0.904 \\ \hline \multirow{3}{*}{Ours} & 20\% & 0.816 & 0.870 & 0.702 & 0.777 & 0.669 & 0.743 & \textbf{0.836} & \textbf{0.904} & 0.856 & 0.912 \\ & 40\% & 0.825 & 0.883 & \textbf{0.702} & \textbf{0.777} & 0.690 & 0.757 & 0.825 & 0.895 & 0.850 & 0.909 \\ & 60\% & 0.855 & 0.902 & 0.694 & 0.772 & \textbf{0.701} & \textbf{0.767} & 0.833 & 0.899 & \textbf{0.865} & \textbf{0.916} \\ \hline \end{tabular} \label{tab:compare_with_supervised} \end{table*} \begin{figure*}[ht \centering \includegraphics[width=0.6\linewidth]{images/result_online.png} \caption{Qualitative result comparison between offline pseudo labeling and online pseudo labeling, with/without using momentum network. From left to right: input image in RGB format, ground truth, the output of teacher (momentum network) trained on the labeled dataset, the output of original student trained with offline pseudo labeling, the output of momentum network of original student trained with offline pseudo labeling, the output of original student trained with online pseudo labeling and the output of momentum network of original student trained with online pseudo labeling.} \label{fig:visualize_result} \end{figure*} \section{Conclusion and Future Works} This paper proposes a semi-supervised learning method that combines online pseudo label generation and momentum network. This method has been shown to generate pseudo labels without complicated settings. Our method achieves an average mDice of 85.33\% on a test set of 5 different data sets, CVC-300, CVC-ClinicDB, CVC-ColonDB, ETIS-LaribPolypDB, Kvasir-SEG. Especially, mDice on some datasets such as CVC-300, CVC-ColonDB, and ETIS-LaribPolypDB are outperformed with state-of-the-art supervised learning methods. This research is a promising direction in many practical applications because of the enormous amount of current and future unlabelled data, especially in support systems for diagnostics on medical images. We hope this research can promote the applications of semi-supervised learning in medical image segmentation problems in the future. \section*{Acknowledgement} This work is funded by Vingroup Innovation Foundation (VINIF) under project code VINIF.2020.DA17. This work is also partially supported by Sun-Asterisk Inc. We would like to thank our colleagues at Sun-Asterisk Inc for their advice and expertise. Without their support, this experiment would not have been accomplished. \printbibliography \end{document} }

a037df5c4ed44c81599fa6532767bce02d42dc07

\section{Introduction} More than one hundred years ago, H. Weyl has advocated a new theory to unify all the interactions known at that time, namely gravitational interaction and electro-magnetic force, within the framework of a newly established geometry which is nowadays called ``Weyl geometry'' \cite{Weyl, Scholz}. In Riemann geometry both length and angle are preserved under parallel transport while in Weyl geometry, only angle, but not length, is preserved by the Weyl gauge field. Soon after the advent of the Weyl's idea, A. Einstein has criticized that regarding the spacing of atomic spectral lines, the prediction obtained from Weyl's theory and the experimental observations were in contradiction (this problem is sometimes called the second clock problem \cite{Penrose}), thus Weyl theory has been buried in oblivion for a long time.\footnote{Even during this period, there were some papers dealing with Weyl theory \cite{Dirac}-\!\cite{Cesare}.} However, in recent years a considerate interest has been developed for Weyl conformal geometry. This is because it was found that the Weyl gauge field acquires a huge mass around the Planck scale and decouples at low energies, thereby avoiding the second clock problem \cite{Ghilencea1, Ghilencea2, Oda-P, Oda-Corf}. In addition, we have noticed the importance of global scale invariance and also local scale invariance, which is also called Weyl invariance, in formulating a theory beyond the Standard Model (BSM) \cite{Bardeen} and quantum gravity. Hence, Weyl geometry provides us with a natural playground for describing Weyl symmetry. The study of Weyl conformal gravity in Weyl geometry has been mainly limited to a classical analysis thus far.\footnote{At one-loop level, the effective potential has been already calculated in \cite{Oda-P, Oda-Corf}.} One of motivations behind the present article is to present a quantum theory of Weyl conformal gravity. To this end, we construct a BRST formalism of the theory from which we can shed some light on important features of quantum aspects of Weyl conformal gravity in Weyl geometry. For instance, as been already shown in case of Weyl invariant scalar-tensor gravity \cite{Oda-Q, Oda-W, Oda-V}, there is an extended ${I\kern-.06em O\kern-.11em S\kern-.04em p}(10|10)$ choral symmetry compared with the ${I\kern-.06em O\kern-.11em S\kern-.04em p}(8|8)$ choral symmetry in Einstein's gravity \cite{Nakanishi, N-O-text}. This extended symmetry is not confined to the sector of the Nakanishi-Lautrup auxiliary fields and the Faddeev-Popov (FP) (anti-)ghosts but relevant to a classical theory. Moreover, it can be shown that we have a gravitational analog of conformal algebra as a subalgebra of the ${I\kern-.06em O\kern-.11em S\kern-.04em p}(10|10)$ choral symmetry. That algebra then gives rise to a spontaneous symmetry breakdown to the Poincar\'e symmetry, by which we can prove that the graviton \cite{NO} and the dilaton \cite{Oda-W, Oda-V} are exactly massless since they are the Nambu-Goldstone particles. The paper is organized as follows. In Section 2, we give a brief review of Weyl geometry. In Section 3, we consider a classical theory which is not only invariant under Weyl gauge transformation but also is free of ghosts, that generally exist in the higher-derivative gravities. Based on the classical theory in Section 3, we fix the gauge symmetries by the extended de Donder gauge and new scalar gauge conditions and construct a BRST invariant quantum Lagrangian in Section 4. In Section 5, we perform the canonical quantization of the quantum Lagrangian where we meet primary and secondary constraints associated with Weyl symmetry. They are the second-class constraints and hence are treated by applying the Dirac brackets. In Section 6, we prove the unitarity of the physical S-matrix on the basis of the BRST quartet mechanism. We find that physical modes are the two polarizations of the massless graviton and the three modes of the massive Weyl gauge fields. Furthermore, it is shown that the massless dilaton, which is eaten by the Weyl gauge field via the Higgs mechanism, belongs to the unphysical sector. In Section 7, we show that the quantum Lagrangian of Weyl conformal gravity possesses the huge global ${I\kern-.06em O\kern-.11em S\kern-.04em p}(10|10)$ choral symmetry. In Section 8, we point out that there exists a gravitational analog of conformal symmetry in quantum gravity and investigate the spontaneous symmetry breaking. We find that the graviton and the dilaton are massless Nambu-Goldstone bosons. The final section is devoted to discussion. \section{Review of Weyl conformal geometry} In this section, we briefly review the basic concepts and definitions of Weyl conformal geometry \cite{Oda-Corf}.\footnote{We follow the notation and conventions of MTW textbook \cite{MTW}. Lower case Greek letters $\mu, \nu, \cdots$ and Latin ones $i, j, \cdots$ are used for space-time and spatial indices, respectively; for instance, $\mu= 0, 1, 2, 3$ and $i = 1, 2, 3$. The Riemann curvature tensor and the Ricci tensor are respectively defined by $R^\rho{}_{\sigma\mu\nu} = \partial_\mu \Gamma^\rho_{\sigma\nu} - \partial_\nu \Gamma^\rho_{\sigma\mu} + \Gamma^\rho_{\lambda\mu} \Gamma^\lambda_{\sigma\nu} - \Gamma^\rho_{\lambda\nu} \Gamma^\lambda_{\sigma\mu}$ and $R_{\mu\nu} = R^\rho{}_{\mu\rho\nu}$. The Minkowski metric tensor is denoted by $\eta_{\mu\nu}$; $\eta_{00} = - \eta_{11} = - \eta_{22} = - \eta_{33} = -1$ and $\eta_{\mu\nu} = 0$ for $\mu \neq \nu$.} In Weyl geometry, the Weyl gauge transformation, which is the sum of a local scale transformation for a generic field $\Phi (x)$ and a gauge transformation for the Weyl gauge field $S_\mu(x)$, is defined as \begin{eqnarray} \Phi (x) \rightarrow \Phi^\prime (x) = e^{w \Lambda(x)} \Phi (x), \qquad S_\mu (x) \rightarrow S^\prime_\mu (x) = S_\mu (x) - \frac{1}{f} \partial_\mu \Lambda (x), \label{Weyl transf} \end{eqnarray} where $w$ is called the ``Weyl weight'', or simply ``weight'' henceforth, $f$ is the coupling constant for the non-compact Abelian gauge group, and $\Lambda(x)$ is a local parameter for the Weyl transformation. The Weyl gauge transformation for various fields is explicitly given by \begin{eqnarray} g_{\mu\nu} (x) &\rightarrow& g_{\mu\nu}^\prime (x) = e^{2 \Lambda(x)} g_{\mu\nu}(x), \qquad \phi (x) \rightarrow \phi^\prime (x) = e^{- \Lambda(x)} \phi (x), \nonumber\\ \psi (x) &\rightarrow& \psi^\prime (x) = e^{- \frac{3}{2} \Lambda(x)} \psi (x), \qquad A_\mu (x) \rightarrow A^\prime_\mu (x) = A_\mu (x), \label{Weyl transf 2} \end{eqnarray} where $g_{\mu\nu} (x)$, $\phi (x)$, $\psi (x)$ and $A_\mu (x)$ are the metric tensor, scalar, spinor, and electromagnetic gauge fields, respectively. The covariant derivative $D_\mu$ for the Weyl gauge transformation for a generic field $\Phi (x)$ of weight $w$ is defined as \begin{eqnarray} D_\mu \Phi \equiv \partial_\mu \Phi + w f S_\mu \Phi, \label{W-cov-deriv} \end{eqnarray} which transforms covariantly under the Weyl transformation: \begin{eqnarray} D_\mu \Phi \rightarrow (D_\mu \Phi)^\prime = e^{w \Lambda(x)} D_\mu \Phi. \label{S-cov-transf} \end{eqnarray} The Weyl geometry is defined as a geometry with a real symmetric metric tensor $g_{\mu\nu} (= g_{\nu\mu})$ and a symmetric connection $\tilde \Gamma^\lambda_{\mu\nu} (= \tilde \Gamma^\lambda_{\nu\mu})$ which is defined as\footnote{We often use the tilde characters to express quantities belonging to Weyl geometry.} \begin{eqnarray} \tilde \Gamma^\lambda_{\mu\nu} &=& \frac{1}{2} g^{\lambda\rho} \left( D_\mu g_{\nu\rho} + D_\nu g_{\mu\rho} - D_\rho g_{\mu\nu} \right) \nonumber\\ &=& \Gamma^\lambda_{\mu\nu} + f \left( S_\mu \delta^\lambda_\nu + S_\nu \delta^\lambda_\mu - S^\lambda g_{\mu\nu} \right), \label{W-connection} \end{eqnarray} where $\Gamma^\lambda_{\mu\nu}$ is the standard Christoffel symbol in Riemann geometry. The most important difference between Riemann geometry and Weyl geometry lies in the fact that in Riemann geometry the metric condition is satisfied: \begin{eqnarray} \nabla_\lambda g_{\mu\nu} \equiv \partial_\lambda g_{\mu\nu} - \Gamma^\rho_{\lambda\mu} g_{\rho\nu} - \Gamma^\rho_{\lambda\nu} g_{\mu\rho} = 0, \label{Metric cond} \end{eqnarray} while in Weyl geometry we have: \begin{eqnarray} \tilde \nabla_\lambda g_{\mu\nu} \equiv \partial_\lambda g_{\mu\nu} - \tilde \Gamma^\rho_{\lambda\mu} g_{\rho\nu} - \tilde \Gamma^\rho_{\lambda\nu} g_{\mu\rho} = - 2 f S_\lambda g_{\mu\nu}, \label{W-metric cond} \end{eqnarray} where $\nabla_\mu$ and $\tilde \nabla_\mu$ are covariant derivatives for diffeomorphisms in Riemann and Weyl geometries, respectively. Since the metric condition (\ref{Metric cond}) implies that both length and angle are preserved under parallel transport, Eq. (\ref{W-metric cond}) shows that only angle, but not length, is preserved by the Weyl connection. The general covariant derivative for both diffeomorphisms and Weyl gauge transformation, for instance, for a covariant vector of weight $w$, is defined as \begin{eqnarray} {\cal D}_\mu V_\nu &\equiv& D_\mu V_\nu - \tilde \Gamma^\rho_{\mu\nu} V_\rho \nonumber\\ &=& \tilde \nabla_\mu V_\nu + w f S_\mu V_\nu \nonumber\\ &=& \nabla_\mu V_\nu + w f S_\mu V_\nu - f ( S_\mu \delta^\rho _\nu + S_\nu \delta^\rho _\mu - S^\rho g_{\mu\nu} ) V_\rho \nonumber\\ &=& \partial_\mu V_\nu + w f S_\mu V_\nu - \Gamma^\rho_{\mu\nu} V_\rho - f ( S_\mu \delta^\rho _\nu + S_\nu \delta^\rho _\mu - S^\rho g_{\mu\nu} ) V_\rho. \label{Gen-cov-deriv} \end{eqnarray} One can verify that using the general covariant derivative, the following metric condition is satisfied: \begin{eqnarray} {\cal D}_\lambda g_{\mu\nu} = 0. \label{Gen-metric cond} \end{eqnarray} Moreover, under Weyl gauge transformation the general covariant derivative for a generic field $\Phi$ of weight $w$ transforms in a covariant manner as desired: \begin{eqnarray} {\cal D}_\mu \Phi \rightarrow ({\cal D}_\mu \Phi)^\prime = e^{w \Lambda(x)} {\cal D}_\mu \Phi, \label{Gen-cov-transf} \end{eqnarray} because the Weyl connection is invariant under Weyl gauge transformation, i.e., $\tilde \Gamma^{\prime \rho}_{\mu\nu} = \tilde \Gamma^\rho_{\mu\nu}$. As in Riemann geometry, in Weyl geometry one can also construct a Weyl invariant curvature tensor $\tilde R_{\mu\nu\rho} \, ^\sigma$ via a commutator of the covariant derivative $\tilde \nabla_\mu$: \begin{eqnarray} [ \tilde \nabla_\mu, \tilde \nabla_\nu ] V_\rho = \tilde R_{\mu\nu\rho} \, ^\sigma V_\sigma. \label{Commutator} \end{eqnarray} Calculating this commutator, one finds that \begin{eqnarray} \tilde R_{\mu\nu\rho} \, ^\sigma &=& \partial_\nu \tilde \Gamma^\sigma_{\mu\rho} - \partial_\mu \tilde \Gamma^\sigma_{\nu\rho} + \tilde \Gamma^\alpha_{\mu\rho} \tilde \Gamma^\sigma_{\alpha\nu} - \tilde \Gamma^\alpha_{\nu\rho} \tilde \Gamma^\sigma_{\alpha\mu} \nonumber\\ &=& R_{\mu\nu\rho} \, ^\sigma + 2 f \left( \delta^\sigma_{[\mu} \nabla_{\nu]} S_\rho - \delta^\sigma_\rho \nabla_{[\mu} S_{\nu]} - g_{\rho [\mu} \nabla_{\nu]} S^\sigma \right) \nonumber\\ &+& 2 f^2 \left( S_{[\mu} \delta^\sigma_{\nu]} S_\rho - S_{[\mu} g_{\nu]\rho} S^\sigma + \delta^\sigma_{[\mu} g_{\nu]\rho} S_\alpha S^\alpha \right), \label{W-curv-tensor} \end{eqnarray} where $R_{\mu\nu\rho} \, ^\sigma$ is the curvature tensor in Riemann geometry and we have defined the antisymmetrization by the square bracket, i.e., $A_{[\mu} B_{\nu]} \equiv \frac{1}{2} ( A_\mu B_\nu - A_\nu B_\mu )$. Then, it is straightforward to prove the following identities: \begin{eqnarray} \tilde R_{\mu\nu\rho} \, ^\sigma = - \tilde R_{\nu\mu\rho} \, ^\sigma, \qquad \tilde R_{[\mu\nu\rho]} \, ^\sigma = 0, \qquad \tilde \nabla_{[\lambda} \tilde R_{\mu\nu]\rho} \, ^\sigma = 0. \label{W-curv-identity} \end{eqnarray} From $\tilde R_{\mu\nu\rho} \, ^\sigma$ one can define a Weyl invariant Ricci tensor: \begin{eqnarray} \tilde R_{\mu\nu} &\equiv& \tilde R_{\mu\rho\nu} \, ^\rho \nonumber\\ &=& R_{\mu\nu} + f \left( - 2 \nabla_\mu S_\nu - H_{\mu\nu} - g_{\mu\nu} \nabla_{\alpha} S^\alpha \right) \nonumber\\ &+& 2 f^2 \left( S_\mu S_\nu - g_{\mu\nu} S_\alpha S^\alpha \right). \label{W-Ricci-tensor} \end{eqnarray} Let us note that \begin{eqnarray} \tilde R_{[\mu\nu]} \equiv \frac{1}{2} ( \tilde R_{\mu\nu} - \tilde R_{\nu\mu} ) = - 2 f H_{\mu\nu}. \label{W-Ricci-tensor 2} \end{eqnarray} Similarly, one can define not a Weyl invariant but a Weyl covariant scalar curvature: \begin{eqnarray} \tilde R \equiv g^{\mu\nu} \tilde R_{\mu\nu} = R - 6 f \nabla_\mu S^\mu - 6 f^2 S_\mu S^\mu. \label{W-scalar-curv} \end{eqnarray} One finds that under Weyl gauge transformation, $\tilde R \rightarrow \tilde R^\prime = e^{- 2 \Lambda(x)} \tilde R$ while $\tilde \Gamma^\lambda_{\mu\nu}, \tilde R_{\mu\nu\rho} \, ^\sigma$ and $\tilde R_{\mu\nu}$ are all invariant. We close this section by discussing a spinor field as an example of matter fields in Weyl geometry. As is well known, to describe a spinor field it is necessary to introduce the vierbein $e^a _\mu$, which is defined as \begin{eqnarray} g_{\mu\nu} = \eta_{ab} e^a _\mu e^b _\nu, \label{Vierbein} \end{eqnarray} where $a, b, \cdots$ are local Lorentz indices taking $0, 1, 2, 3$ and $\eta_{ab} = \rm{diag} ( - 1, 1, 1, 1)$. Now the metric condition (\ref{Gen-metric cond}) takes the form: \begin{eqnarray} {\cal D}_\mu e^a _\nu \equiv D_\mu e^a _\nu + \tilde \omega^a \, _{b \mu} e^b _\nu - \tilde \Gamma^\rho_{\mu\nu} e^a _\rho = 0, \label{Gen-vierbein cond} \end{eqnarray} where the general covariant derivative is extended to include the local Lorentz transformation whose gauge connection is the spin connection $\tilde \omega^a \, _{b \mu}$ of weight $0$ in Weyl geometry, and $D_\mu e^a _\nu = \partial_\mu e^a _\nu + f S_\mu e^a _\nu$ since the vierbein $e^a _\mu$ has weight $1$. Solving the metric condition (\ref{Gen-vierbein cond}) leads to the expression of the spin connection in Weyl geometry: \begin{eqnarray} \tilde \omega_{a b \mu} = \omega_{a b \mu} + f e^c _\mu ( \eta_{ac} S_b - \eta_{bc} S_a ), \label{spin connection} \end{eqnarray} where $\omega_{a b \mu}$ is the spin connection in Riemann geometry and we have defined $S_a \equiv e^\mu _a S_\mu$. Then, the general covariant derivative for a spinor field $\Psi$ of weight $- \frac{3}{2}$ reads: \begin{eqnarray} {\cal D}_\mu \Psi = D_\mu \Psi + \frac{i}{2} \tilde \omega_{a b \mu} S^{a b} \Psi, \label{spinor CD} \end{eqnarray} where $D_\mu \Psi = \partial_\mu \Psi - \frac{3}{2} f S_\mu \Psi$ and the Lorentz generator $S^{a b}$ for a spinor field is defined as $S^{a b} = \frac{i}{4} [ \gamma^a, \gamma^b ]$. Here we define the gamma matrices to satisfy the Clifford algebra $\{ \gamma^a, \gamma^b \} = - 2 \eta^{ab}$. Since the spin connection $\tilde \omega^a \, _{b \mu}$ has weight $0$, the covariant derivative ${\cal D}_\mu \Psi$ transforms covariantly under Weyl gauge transformation: \begin{eqnarray} {\cal D}_\mu \Psi \rightarrow ( {\cal D}_\mu \Psi )^\prime = e^{- \frac{3}{2} \Lambda(x)} {\cal D}_\mu \Psi. \label{spinor covariance} \end{eqnarray} Then, the Lagrangian density for a massless Dirac spinor field is of form: \begin{eqnarray} {\cal L} = \frac{i}{2} e \ e^\mu _a ( \bar \Psi \gamma^a {\cal D}_\mu \Psi - {\cal D}_\mu \bar \Psi \gamma^a \Psi ), \label{spinor Lag} \end{eqnarray} where $e \equiv \sqrt{-g}, \bar \Psi \equiv \Psi^\dagger \gamma^0$, and ${\cal D}_\mu \bar \Psi$ is given by \begin{eqnarray} {\cal D}_\mu \bar \Psi = D_\mu \bar \Psi - \bar \Psi \frac{i}{2} \tilde \omega_{a b \mu} S^{a b}. \label{spinor CD2} \end{eqnarray} Inserting Eqs. (\ref{spinor CD}) and (\ref{spinor CD2}) to the Lagrangian density (\ref{spinor Lag}), we find that \begin{eqnarray} {\cal L} &=& \frac{i}{2} e \Bigl[ e^\mu _a \left( \bar \Psi \gamma^a \partial_\mu \Psi - \partial_\mu \bar \Psi \gamma^a \Psi + \frac{i}{2} \omega_{b c \mu} \bar \Psi \{ \gamma^a, S^{bc} \} \Psi \right) \nonumber\\ &+& \frac{i}{2} f ( \eta_{ab} S_c - \eta_{ac} S_b ) \bar \Psi \{ \gamma^a, S^{bc} \} \Psi \Bigr]. \label{spinor Lag2} \end{eqnarray} The last term identically vanishes owing to the relation: \begin{eqnarray} \{ \gamma^a, S^{bc} \} = - \varepsilon^{abcd} \gamma_5 \gamma_d, \label{gamma rel} \end{eqnarray} where we have defined as $\gamma_5 = i \gamma^0 \gamma^1 \gamma^2 \gamma^3$ and $\varepsilon^{0123} = +1$. Thus, as is well known, the Weyl gauge field $S_\mu$ does not couple minimally to a spinor field $\Psi$. Technically speaking, it is the absence of imaginary unit $i$ in the covariant derivative $D_\mu \Psi = \partial_\mu \Psi - \frac{3}{2} f S_\mu \Psi$ that induced this decoupling of the Weyl gauge field from the spinor field. Without the imaginary unit, the terms including the Weyl gauge field cancel out each other in Eq. (\ref{spinor Lag}). In a similar manner, we can prove that the Weyl gauge field does not couple to a gauge field, i.e., the electromagnetic potential $A_\mu$ either. On the other hand, the Weyl gauge field can couple to a scalar field such as the Higgs field as well as a graviton. \section{Classical theory} We wish to consider a model of Weyl conformal gravity in Weyl geometry. It is of interest to recall that without matter fields we have a unique classical Lagrangian which is invariant under the Weyl gauge transformation; the Lagrangian must be of form of quadratic gravity: \begin{eqnarray} {\cal L}_{QG} = \sqrt{- g} \left ( - \frac{1}{2 \xi^2} \tilde C_{\mu\nu\rho\sigma} \tilde C^{\mu\nu\rho\sigma} + \alpha \tilde R^2 \right), \label{L-QG} \end{eqnarray} where $\xi$ and $\alpha$ are dimensionless coupling constants, and $\tilde C_{\mu\nu\rho\sigma}$ and $\tilde R$ are a generalization of conformal tensor and scalar curvature in Weyl geometry, respectively. Note that the Lagrangian of the Einstein-Hilbert type or the higher-derivative terms involving more than quadratic terms are prohibited to be present by Weyl gauge symmetry. The fatal defect of the Lagrangian (\ref{L-QG}), however, is the existence of a massless ghost which breaks unitarity in quantum regime. Another unsatisfactory feature of the Lagrangian (\ref{L-QG}) is that it does not reduce to Einstein's general relativity at low energies which is known to be a good description of the physics relevant to gravitational phenomena at such long range scales. Provided that we are allowed to use matter fields\footnote{As explained in the previous section, fermions and the conventional gauge fields do not couple to the Weyl gauge field, but only the scalar field does.}, the situation changes and we can construct a scalar-tensor gravity of the Einstein-Hilbert type which includes at most the second-order derivatives of the metric tensor \cite{Dirac}: \begin{eqnarray} {\cal L}_{ST} = \sqrt{- g} \, \frac{1}{2} \xi \phi^2 \tilde R, \label{L-ST} \end{eqnarray} where $\phi$ is a real scalar field.\footnote{The extension to a complex scalar field or multiple scalar fields is straightforward.} The most general classical Lagrangian, which is invariant under Weyl gauge transformation and is free of the massless ghost, reads: \begin{eqnarray} {\cal L}_G &=& \sqrt{- g} \biggl[ \frac{1}{2} \xi \phi^2 \tilde R - \frac{1}{4} H_{\mu\nu} H^{\mu\nu} - \frac{1}{2} \epsilon g^{\mu\nu} D_\mu \phi D_\nu \phi - \frac{\lambda}{4 !} \phi^4 \nonumber\\ &+& \eta \left( \frac{1}{12} \phi^2 R + \frac{1}{2} g^{\mu\nu} \partial_\mu \phi \partial_\nu \phi \right) \biggr], \label{L-G} \end{eqnarray} where $\xi, \lambda, \eta$ are all dimensionless constants, and $\epsilon = \pm 1$ depending on a normal field $\epsilon = 1$ or a ghost field $\epsilon = -1$. In this article, we limit ourselves to the case $6 \xi + \epsilon \neq 0$ since the specific case $6 \xi + \epsilon = 0$ leads to the same expression as the last term with the constant $\eta$, which is called ``Weyl invariant scalar-tensor gravity'', when surface terms are ignored. Finally, the scalar field $\phi$ has the weight $-1$ so the Weyl covariant derivative in (\ref{L-G}) takes the form:\footnote{In what follows, we will set $f = 1$ for the coupling constant for the non-compact Abelian gauge group.} \begin{eqnarray} D_\mu \phi = \partial_\mu \phi - S_\mu \phi. \label{W-covd-S} \end{eqnarray} Since we have already analyzed the Weyl invariant scalar-tensor gravity in Riemann geometry \cite{Oda-W} and the quartic potential term has no essential role in the BRST formalism, we will put $\lambda = \eta = 0$. Thus, the classical Lagrangian which is treated in this article reads: \begin{eqnarray} {\cal L}_c &=& \sqrt{- g} \biggl[ \frac{1}{2} \xi \phi^2 \tilde R - \frac{1}{4} H_{\mu\nu} H^{\mu\nu} - \frac{1}{2} \epsilon g^{\mu\nu} D_\mu \phi D_\nu \phi \biggr] \nonumber\\ &=& \sqrt{- g} \biggl[ \frac{1}{2} \xi \phi^2 ( R - 6 \nabla_\mu S^\mu - 6 S_\mu S^\mu ) - \frac{1}{4} H_{\mu\nu} H^{\mu\nu} \nonumber\\ &-& \frac{1}{2} \epsilon g^{\mu\nu} ( \partial_\mu \phi - S_\mu \phi ) ( \partial_\nu \phi - S_\nu \phi ) \biggr]. \label{L-c} \end{eqnarray} \section{Quantum theory} The classical Lagrangian (\ref{L-c}) is invariant under both general coordinate transformation (GCT) and Weyl gauge transformation. For a quantum theory we have to fix such gauge symmetries by introducing suitable gauge-fixing conditions. After introducing the gauge-fixing conditions the quantum Lagrangian is not longer invariant under the gauge transformations, but as residual global symmetries the quantum Lagrangian is invariant under two BRST transformations, one of which is denoted as $\delta_B$, corresponding to the GCT is defined as \begin{eqnarray} \delta_B g_{\mu\nu} &=& - ( \nabla_\mu c_\nu+ \nabla_\nu c_\mu) = - ( c^\alpha\partial_\alpha g_{\mu\nu} + \partial_\mu c^\alpha g_{\alpha\nu} + \partial_\nu c^\alpha g_{\mu\alpha} ), \nonumber\\ \delta_B \phi &=& - c^\lambda \partial_\lambda \phi, \quad \delta_B S_\mu = - c^\lambda \nabla_\lambda S_\mu - \nabla_\mu c^\lambda S_\lambda, \nonumber\\ \delta_B c^\rho &=& - c^\lambda\partial_\lambda c^\rho, \quad \delta_B \bar c_\rho = i B_\rho, \quad \delta_B B_\rho = 0, \label{GCT-BRST} \end{eqnarray} where $c^\rho$ and $\bar c_\rho$ are respectively the Faddeev-Popov (FP) ghost and anti-ghost, $B_\rho$ is the Nakanishi-Lautrup (NL) field. For convenience, in place of the NL field $B_\rho$ we will introduce a new NL field defined as \begin{eqnarray} b_\rho= B_\rho- i c^\lambda\partial_\lambda\bar c_\rho, \label{b-rho-field} \end{eqnarray} and its BRST transformation reads: \begin{eqnarray} \delta_B b_\rho= - c^\lambda\partial_\lambda b_\rho. \label{b-BRST} \end{eqnarray} The other BRST transformation, which is denoted as $\bar \delta_B$, corresponding to the Weyl transformation is defined as \begin{eqnarray} \bar \delta_B g_{\mu\nu} &=& 2 c g_{\mu\nu}, \quad \bar \delta_B \phi = - c \phi, \quad \bar \delta_B S_\mu = - \partial_\mu c, \nonumber\\ \bar \delta_B \bar c &=& i B, \quad \bar \delta_B c = \bar \delta_B B = 0, \label{Weyl-BRST} \end{eqnarray} where $c$ and $\bar c$ are respectively the FP ghost and FP anti-ghost, $B$ is the NL field. Note that the two BRST transformations are nilpotent, i.e., \begin{eqnarray} \delta_B^2 = \bar \delta_B^2 = 0. \label{Nilpotent} \end{eqnarray} To complete the two BRST transformations, we have to fix not only the GCT BRST transformation $\delta_B$ on $c, \bar c$ and $B$ but also the Weyl BRST transformation $\delta_B$ on $c^\rho, \bar c_\rho$ and $b_\rho$. The BRST transformations on these fields are fixed by requiring that the two BRST transformations anti-commute with each other, that is, \cite{Oda-W} \begin{eqnarray} \{ \delta_B, \bar \delta_B \} \equiv \delta_B \bar \delta_B + \bar \delta_B \delta_B = 0. \label{GCT-Weyl-BRST} \end{eqnarray} Then, the resultant BRST transformations take the form: \begin{eqnarray} \delta_B B &=& - c^\lambda\partial_\lambda B, \quad \delta_B c = - c^\lambda\partial_\lambda c, \quad \delta_B \bar c = - c^\lambda\partial_\lambda \bar c, \nonumber\\ \bar \delta_B b_\rho &=& \bar \delta_B c^\rho = \bar \delta_B \bar c_\rho = 0. \label{BRST2} \end{eqnarray} In this context, it is worthwhile to recall that the gauge condition for the GCT must be invariant under Weyl gauge transformation while the one for Weyl transformation must be invariant under GCT in order for the two BRST transformations to anti-commute. In that case we can consider the two BRST transformations separately. The suitable gauge condition for the GCT is almost unique and is called ``the extended de Donder gauge'' \cite{Oda-W}:\footnote{Let us note that this gauge condition breaks the general coordinate invariance, but it is invariant under the general linear transformation $GL(4)$. Thus, the quantum Lagrangian which is obtained shortly is also invaraint under the $GL(4)$.} \begin{eqnarray} \partial_\mu ( \tilde g^{\mu\nu} \phi^2 ) = 0, \label{Ext-de-Donder} \end{eqnarray} where we have defined $\tilde g^{\mu\nu} \equiv \sqrt{-g} g^{\mu\nu}$. On the other hand, we have a few candidates for the gauge-fixing condition for the Weyl transformation, which must be invariant under the GCT, i.e., a scalar quantity. The first one is the well-known ``unitary gauge'', $\phi = \rm{constant}$, which is taken to show that Weyl invariant scalar-tensor gravity is equivalent to the Einstein-Hilbert term. The other gauge condition is the Lorenz gauge, $\nabla_\mu S^\mu = 0$, which is usually adopted in quantum field theories. However, it turns out that these gauge conditions are not so interesting in the present context since they do not allow for conformal symmetry to remain. Hence, we shall choose, what we call, ``the scalar gauge condition'' \cite{Oda-W}: \begin{eqnarray} \partial_\mu ( \tilde g^{\mu\nu} \phi \partial_\nu \phi ) = 0, \label{Scalar-gauge} \end{eqnarray} which can be alternatively written as \begin{eqnarray} \Box \, \phi^2 = 0. \label{Alt-Scalar-gauge} \end{eqnarray} After taking the extended de Donder gauge condition (\ref{Ext-de-Donder}) for the GCT and the scalar gauge condition (\ref{Scalar-gauge}) for the Weyl transformation, the gauge-fixed and BRST invariant quantum Lagrangian is given by \begin{eqnarray} {\cal L}_q &=& {\cal L}_c + {\cal L}_{GF+FP} + \bar {\cal L}_{GF+FP} \nonumber\\ &=& {\cal L}_c + i \delta_B ( \tilde g^{\mu\nu} \phi^2 \partial_\mu \bar c_\nu ) + i \bar \delta_B \left[ \bar c \partial_\mu ( \tilde g^{\mu\nu} \phi \partial_\nu \phi ) \right] \nonumber\\ &=& \sqrt{- g} \biggl[ \frac{1}{2} \xi \phi^2 ( R - 6 \nabla_\mu S^\mu - 6 S_\mu S^\mu ) - \frac{1}{4} H_{\mu\nu} H^{\mu\nu} - \frac{1}{2} \epsilon g^{\mu\nu} D_\mu \phi D_\nu \phi \biggr] \nonumber\\ &-& \tilde g^{\mu\nu} \phi^2 ( \partial_\mu b_\nu + i \partial_\mu \bar c_\lambda \partial_\nu c^\lambda ) + \tilde g^{\mu\nu} \phi \partial_\mu B \partial_\nu \phi - i \tilde g^{\mu\nu} \phi^2 \partial_\mu \bar c \partial_\nu c, \label{ST-q-Lag} \end{eqnarray} where surface terms are dropped. From the Lagrangian ${\cal L}_q$, it is straightforward to derive the field equations by taking the variation with respect to $g_{\mu\nu}, S_\mu, \phi, b_\nu, B, c^\rho, \bar c_\rho, c$ and $\bar c$ in order: \begin{eqnarray} &{}& \frac{1}{2} \xi \phi^2 G_{\mu\nu} - \frac{1}{2} \xi ( \nabla_\mu \nabla_\nu - g_{\mu\nu} \Box ) \phi^2 - 3 \xi \phi^2 ( S_\mu S_\nu - \frac{1}{2} g_{\mu\nu} S_\alpha S^\alpha ) \nonumber\\ &{}& + 3 \xi \phi ( S_\mu \partial_\nu \phi + S_\nu \partial_\mu \phi - g_{\mu\nu} S^\alpha \partial_\alpha \phi ) - \frac{1}{2} H_{\mu\alpha} H_\nu \, ^\alpha + \frac{1}{8} g_{\mu\nu} H_{\alpha\beta}^2 \nonumber\\ &{}& - \frac{1}{2} \epsilon \left[ D_\mu \phi D_\nu \phi - \frac{1}{2} g_{\mu\nu} ( D_\alpha \phi )^2 \right] - \frac{1}{2} ( E_{\mu\nu} - \frac{1}{2} g_{\mu\nu} E ) = 0, \nonumber\\ &{}& ( 6 \xi + \epsilon ) g^{\mu\nu} \phi D_\nu \phi - \nabla_\nu H^{\mu\nu} = 0, \nonumber\\ &{}& \xi \phi^2 ( R - 6 \nabla_\mu S^\mu - 6 S_\mu S^\mu ) + \epsilon \frac{1}{\sqrt{-g}} \phi D_\mu ( \tilde g^{\mu\nu} \phi D_\nu \phi ) - E \nonumber\\ &{}& - 2 g^{\mu\nu} \phi \partial_\mu B \partial_\nu \phi - \phi^2 \Box B = 0, \nonumber\\ &{}& \partial_\mu ( \tilde g^{\mu\nu} \phi^2 ) = 0, \qquad \partial_\mu ( \tilde g^{\mu\nu} \phi \partial_\nu \phi ) = 0, \nonumber\\ &{}& g^{\mu\nu} \partial_\mu \partial_\nu\bar c_\rho = g^{\mu\nu} \partial_\mu \partial_\nu c^\rho = g^{\mu\nu} \partial_\mu \partial_\nu\bar c = g^{\mu\nu} \partial_\mu \partial_\nu c = 0. \label{q-field-eq} \end{eqnarray} where $G_{\mu\nu} \equiv R_{\mu\nu} - \frac{1}{2} g_{\mu\nu} R$ denotes the Einstein tensor, while $E_{\mu\nu}$ and $E$ are defined as \begin{eqnarray} E_{\mu\nu} &=& \phi^2 ( \partial_\mu b_\nu + i \partial_\mu \bar c_\lambda \partial_\nu c^\lambda ) - \phi \partial_\mu B \partial_\nu \phi + i \phi^2 \partial_\mu \bar c \partial_\nu c + ( \mu \leftrightarrow \nu ), \nonumber\\ E &=& g^{\mu\nu} E_{\mu\nu}. \label{E} \end{eqnarray} Moreover, since $\tilde g^{\mu\nu} D_\nu \phi$ has the weight $1$, the Weyl covariant derivative is defined as \begin{eqnarray} D_\mu ( \tilde g^{\mu\nu} D_\nu \phi ) = \partial_\mu ( \tilde g^{\mu\nu} D_\nu \phi ) + S_\mu \tilde g^{\mu\nu} D_\nu \phi. \label{W-cov-d} \end{eqnarray} When we introduce the dilaton $\sigma (x)$ by defining \begin{eqnarray} \phi (x) \equiv e^{\sigma (x)}, \label{Dilaton} \end{eqnarray} the two gauge-fixing conditions in (\ref{q-field-eq}), or equivalently, Eqs. (\ref{Ext-de-Donder}) and (\ref{Scalar-gauge}) lead to a very simple d'Alembert-like equation for the dilaton: \begin{eqnarray} g^{\mu\nu} \partial_\mu \partial_\nu \sigma = 0. \label{Dilaton-eq} \end{eqnarray} It is worthwhile to notice that it is not the scalar field $\phi$ but the dilaton $\sigma$ that satisfies this type of equation. In order to show that the auxiliary field $B$ also obeys the same type of equation, let us take account of the trace part of the Einstein equation, i.e., the first field equation in (\ref{q-field-eq}), which gives us the equation: \begin{eqnarray} \xi \phi^2 R - 6 \xi \phi^2 S_\alpha S^\alpha + 12 \xi \phi S^\alpha \partial_\alpha \phi - \epsilon ( D_\alpha \phi )^2 - E = 0. \label{Trace-E-eq} \end{eqnarray} Next, we can rewrite the field equation for $\phi$, the third equation in (\ref{q-field-eq}), as \begin{eqnarray} &{}& \xi \phi^2 R - 6 \xi \phi^2 S_\alpha S^\alpha + 12 \xi \phi S^\alpha \partial_\alpha \phi - \epsilon ( D_\alpha \phi )^2 - E \nonumber\\ &{}& - ( 6 \xi + \epsilon ) \phi^2 g^{\mu\nu} \partial_\mu S_\nu - 2 g^{\mu\nu} \phi \partial_\mu B \partial_\nu \phi - \phi^2 \Box B = 0. \label{Phi-eq} \end{eqnarray} Using Eqs. (\ref{Trace-E-eq}) and (\ref{Phi-eq}), we can obtain the equation: \begin{eqnarray} g^{\mu\nu} \partial_\mu \partial_\nu B + ( 6 \xi + \epsilon ) g^{\mu\nu} \partial_\mu S_\nu = 0. \label{Trace&Phi-eq} \end{eqnarray} Now we are ready to prove \begin{eqnarray} g^{\mu\nu} \partial_\mu S_\nu = 0. \label{S-eq} \end{eqnarray} To do that, let us consider the field equation for $S_\mu$ in (\ref{q-field-eq}), multiply by $\sqrt{-g}$, and then operate the covariant derivative consequently leading to: \begin{eqnarray} \sqrt{-g} \nabla_\mu \nabla_\nu H^{\mu\nu} = ( 6 \xi + \epsilon ) \nabla_\mu ( \tilde g^{\mu\nu} \phi D_\nu \phi ). \label{d-S-eq} \end{eqnarray} The LHS of Eq. (\ref{d-S-eq}) is identically zero and $6 \xi + \epsilon \neq 0$ by our assumption, we find that \begin{eqnarray} \nabla_\mu ( \tilde g^{\mu\nu} \phi D_\nu \phi ) = 0. \label{d-S-eq2} \end{eqnarray} Using the formula: \begin{eqnarray} \nabla_\mu ( \tilde g^{\mu\nu} A_\nu ) = \partial_\mu ( \tilde g^{\mu\nu} A_\nu ), \label{Math-formula} \end{eqnarray} which holds for an arbitrary covariant vector $A_\mu$, Eq. (\ref{d-S-eq2}) is reduced to the form: \begin{eqnarray} \partial_\mu ( \tilde g^{\mu\nu} \phi \partial_\nu \phi - \tilde g^{\mu\nu} \phi^2 S_\nu ) = 0. \label{d-S-eq3} \end{eqnarray} Then, using the gauge conditions (\ref{Ext-de-Donder}) and (\ref{Scalar-gauge}), we can reach the equation (\ref{S-eq}). Hence, Eq. (\ref{Trace&Phi-eq}) implies that the auxiliary field $B$ obeys the equation: \begin{eqnarray} g^{\mu\nu} \partial_\mu \partial_\nu B = 0. \label{B-eq} \end{eqnarray} Surprisingly enough, using the Weyl BRST transformation, we can show this equation (\ref{B-eq}) in the simplest way. For this aim, let us start with the field equation for $\bar c$ in (\ref{q-field-eq}): \begin{eqnarray} g^{\mu\nu} \partial_\mu \partial_\nu \bar c = 0. \label{Bar-c-eq} \end{eqnarray} Operating $\bar \delta_B$ on this equation leads to \begin{eqnarray} - 2 c g^{\mu\nu} \partial_\mu \partial_\nu \bar c + i g^{\mu\nu} \partial_\mu \partial_\nu B = 0. \label{Bar-c-eq2} \end{eqnarray} The first term on the LHS is vanishing owing to (\ref{Bar-c-eq}), so we can arrive at the equation (\ref{B-eq}). In a perfectly similar manner, we can show that the Nakanishi-Lautrup auxiliary field $b_\rho$ satisfies the d'Alembert-like equation by either an explicit calculation or using the BRST transformation for the GCT. Here we present only the latter proof since the former one was given in our previous paper \cite{Oda-Q}. Let us start with the field equation for $\bar c_\rho$ in (\ref{q-field-eq}): \begin{eqnarray} g^{\mu\nu} \partial_\mu \partial_\nu \bar c_\rho = 0. \label{Bar-c-rho-eq} \end{eqnarray} Taking the GCT BRST transformation of this equation yields: \begin{eqnarray} ( - \partial_\lambda g^{\mu\nu} c^\lambda + g^{\mu\alpha} \partial_\alpha c^\nu + g^{\nu\alpha} \partial_\alpha c^\mu ) \partial_\mu \partial_\nu \bar c_\rho + i g^{\mu\nu} \partial_\mu \partial_\nu B_\rho = 0, \label{Bar-c-rho-eq2} \end{eqnarray} where we have used the GCT BRST transformation (\ref{GCT-BRST}). Substituting the definition of $b_\rho$ in Eq. (\ref{b-rho-field}) into (\ref{Bar-c-rho-eq2}), we have the equation for $b_\rho$: \begin{eqnarray} i g^{\mu\nu} \partial_\mu \partial_\nu b_\rho = g^{\mu\nu} \partial_\mu \partial_\nu ( c^\lambda \partial_\lambda \bar c_\rho ) - ( - \partial_\lambda g^{\mu\nu} c^\lambda + 2 g^{\mu\alpha} \partial_\alpha c^\nu ) \partial_\mu \partial_\nu \bar c_\rho. \label{Bar-c-rho-eq3} \end{eqnarray} With the help of Eq. (\ref{Bar-c-rho-eq}) and the field equation for $c^\rho$ in (\ref{q-field-eq}), the RHS is found to be vanishing so we have the desired equation: \begin{eqnarray} g^{\mu\nu} \partial_\mu \partial_\nu b_\rho = 0. \label{b-rho-eq} \end{eqnarray} In other words, setting $X^M = \{ x^\mu, b_\mu, \sigma, B, c^\mu, \bar c_\mu, c, \bar c \}$, $X^M$ turns out to obey the very simple equation: \begin{eqnarray} g^{\mu\nu} \partial_\mu \partial_\nu X^M = 0. \label{X-M-eq} \end{eqnarray} This fact, together with the gauge condition $\partial_\mu ( \tilde g^{\mu\nu} \phi^2 ) = 0$ produces the two kinds of conserved currents: \begin{eqnarray} {\cal P}^{\mu M} &\equiv& \tilde g^{\mu\nu} \phi^2 \partial_\nu X^M = \tilde g^{\mu\nu} \phi^2 \bigl( 1 \overset{\leftrightarrow}{\partial}_\nu X^M \bigr) \nonumber\\ {\cal M}^{\mu M N} &\equiv& \tilde g^{\mu\nu} \phi^2 \bigl( X^M \overset{\leftrightarrow}{\partial}_\nu Y^N \bigr), \label{Cons-currents} \end{eqnarray} where we have defined $X^M \overset{\leftrightarrow}{\partial}_\mu Y^N \equiv X^M \partial_\mu Y^N - ( \partial_\mu X^M ) Y^N$. These conserved currents constitute a Poincar${\rm{\acute{e}}}$-like ${I\kern-.06em O\kern-.11em S\kern-.04em p}(10|10)$ supersymmetry as will be shown later. \section{Canonical quantization and equal-time commutation relations} In this section, after introducing the Poisson brackets, we will evaluate various equal-time commutation relations (ETCRs) among fundamental variables. To simplify various expressions, we will obey the following abbreviations adopted in the textbook of Nakanishi and Ojima \cite{N-O-text}: \begin{eqnarray} [ A, B^\prime ] &=& [ A(x), B(x^\prime) ] |_{x^0 = x^{\prime 0}}, \qquad \delta^3 = \delta(\vec{x} - \vec{x}^\prime), \nonumber\\ \tilde f &=& \frac{1}{\tilde g^{00}} = \frac{1}{\sqrt{-g} g^{00}}, \label{abbreviation} \end{eqnarray} where we assume that $\tilde g^{00}$ is invertible. Here the above brackets $[ A, B^\prime ]$ symbolically describe the Poisson brackets and the ETCRs. First of all, let us set up the Poisson brackets of canonical variables: \begin{eqnarray} &{}& \{ g_{\mu\nu}, \pi_g^{\rho\lambda\prime} \}_P = \frac{1}{2} ( \delta_\mu^\rho\delta_\nu^\lambda + \delta_\mu^\lambda\delta_\nu^\rho) \delta^3, \quad \{ \phi, \pi_\phi^\prime \}_P = \delta^3, \quad \{ S_\mu, \pi_S^{\nu\prime} \}_P = \delta_\mu^\nu \delta^3, \nonumber\\ &{}& \{ c^\sigma, \pi_{c \lambda}^\prime \}_P = \{ \bar c_\lambda, \pi_{\bar c}^{\sigma\prime} \}_P = \delta_\lambda^\sigma \delta^3, \quad \{ B, \pi_B^\prime \}_P = \{ c, \pi_c^\prime \}_P \nonumber\\ &{}& = \{ \bar c, \pi_{\bar c}^\prime \}_P = \delta^3, \label{CCRs} \end{eqnarray} where the other Poisson brackets vanish. Here the canonical variables are $g_{\mu\nu}, \phi, S_\mu, B, c^\rho, \bar c_\rho, c, \bar c$ and the corresponding canonical conjugate momenta are $\pi_g^{\mu\nu}, \pi_\phi, \pi_S^\mu, \pi_B, \pi_{c \rho}, \pi_{\bar c}^\rho, \pi_c, \pi_{\bar c}$, respectively and the $b_\mu$ field is regarded as not a canonical variable but a conjugate momentum of $\tilde g^{0 \mu}$. To remove second order derivatives of the metric involved in $R$, we perform the integration by parts once and rewrite the Lagrangian (\ref{ST-q-Lag}) as \begin{eqnarray} {\cal L}_q &=& - \frac{1}{2} \xi \tilde g^{\mu\nu} \phi^2 ( \Gamma^\sigma_{\mu\nu} \Gamma^\alpha_{\sigma\alpha} - \Gamma^\sigma_{\mu\alpha} \Gamma^\alpha_{\sigma\nu} + 6 S_\mu S_\nu ) - \xi \phi \partial_\mu \phi ( \tilde g^{\alpha\beta} \Gamma^\mu_{\alpha\beta} - \tilde g^{\mu\nu} \Gamma^\alpha_{\nu\alpha} ) \nonumber\\ &+& 6 \xi \tilde g^{\mu\nu} \phi S_\mu \partial_\nu \phi - \frac{1}{4} \sqrt{-g} H_{\mu\nu} H^{\mu\nu} - \frac{1}{2} \epsilon \tilde g^{\mu\nu} D_\mu \phi D_\nu \phi + \partial_\mu ( \tilde g^{\mu\nu} \phi^2 ) b_\nu \nonumber\\ &-& i \tilde g^{\mu\nu} \phi^2 \partial_\mu \bar c_\rho \partial_\nu c^\rho + \tilde g^{\mu\nu} \partial_\mu B \phi \partial_\nu \phi - i \tilde g^{\mu\nu} \phi^2 \partial_\mu \bar c \partial_\nu c + \partial_\mu {\cal{V}}^\mu, \label{Mod-ST-q-Lag} \end{eqnarray} where we have also integrated by parts two terms with the linear $S_\mu$ and $b_\mu$, and a surface term ${\cal{V}}^\mu$ is thus given by \begin{eqnarray} {\cal{V}}^\mu = \frac{1}{2} \xi \phi^2 ( \tilde g^{\alpha\beta} \Gamma^\mu_{\alpha\beta} - \tilde g^{\mu\nu} \Gamma^\alpha_{\nu\alpha} ) - 3 \xi \tilde g^{\mu\nu} \phi^2 S_\nu - \tilde g^{\mu\nu} \phi^2 b_\nu. \label{surface} \end{eqnarray} Using this Lagrangian, the concrete expressions for canonical conjugate momenta become: \begin{eqnarray} \pi_g^{\mu\nu} &=& \frac{\partial {\cal L}_q}{\partial \dot g_{\mu\nu}} \nonumber\\ &=& - \frac{1}{4} \sqrt{-g} \xi \phi^2 \Bigl[ - g^{0 \lambda} g^{\mu\nu} g^{\sigma\tau} - g^{0 \tau} g^{\mu\lambda} g^{\nu\sigma} - g^{0 \sigma} g^{\mu\tau} g^{\nu\lambda} + g^{0 \lambda} g^{\mu\tau} g^{\nu\sigma} \nonumber\\ &+& g^{0 \tau} g^{\mu\nu} g^{\lambda\sigma} + \frac{1}{2} ( g^{0 \mu} g^{\nu\lambda} + g^{0 \nu} g^{\mu\lambda} ) g^{\sigma\tau} \Bigr] \partial_\lambda g_{\sigma\tau} \nonumber\\ &-& \sqrt{-g} \Bigl[ \frac{1}{2} ( g^{0 \mu} g^{\rho\nu} + g^{0 \nu} g^{\rho\mu} ) - g^{\mu\nu} g^{\rho 0} \Bigr] \xi \phi \partial_\rho \phi \nonumber\\ &-& \frac{1}{2} \sqrt{-g} ( g^{0 \mu} g^{\nu\rho} + g^{0 \nu} g^{\mu\rho} - g^{0 \rho} g^{\mu\nu} ) \phi^2 b_\rho, \nonumber\\ \pi_\phi &=& \frac{\partial {\cal L}_q}{\partial \dot \phi} = - \epsilon \tilde g^{0 \mu} D_\mu \phi + 2 \tilde g^{0 \mu} \phi b_\mu + \xi \phi ( - \tilde g^{\alpha\beta} \Gamma^0_{\alpha\beta} + \tilde g^{0 \alpha} \Gamma^\beta_{\alpha\beta} ) \nonumber\\ &+& 6 \xi \tilde g^{0 \mu} \phi S_\mu + \tilde g^{0 \mu} \phi \partial_\mu B, \nonumber\\ \pi_S^\mu &=& \frac{\partial {\cal L}_q}{\partial \dot S_\mu} = - \sqrt{-g} H^{0\mu}, \quad \pi_B = \frac{\partial {\cal L}_q}{\partial \dot B} = \tilde g^{0 \mu} \phi \partial_\mu \phi, \nonumber\\ \pi_{c \sigma} &=& \frac{\partial {\cal L}_q}{\partial \dot c^\sigma} = - i \tilde g^{0 \mu} \phi^2 \partial_\mu \bar c_\sigma, \quad \pi_{\bar c}^\sigma = \frac{\partial {\cal L}_q}{\partial \dot {\bar c}_\sigma} = i \tilde g^{0 \mu} \phi^2 \partial_\mu c^\sigma, \nonumber\\ \pi_c &=& \frac{\partial {\cal L}_q}{\partial \dot c} = - i \tilde g^{0 \mu} \phi^2 \partial_\mu \bar c, \quad \pi_{\bar c} = \frac{\partial {\cal L}_q}{\partial \dot {\bar c}} = i \tilde g^{0 \mu} \phi^2 \partial_\mu c, \label{CCM} \end{eqnarray} where we have defined the time derivative such as $\dot g_{\mu\nu} \equiv \frac{\partial g_{\mu\nu}}{\partial t} \equiv \partial_0 g_{\mu\nu}$, and differentiation of ghosts is taken from the right. It can be easily seen that we have a primary constraint: \begin{eqnarray} \Psi_1 \equiv \pi_S^0 \approx 0. \label{Primary} \end{eqnarray} Let us recall that a secondary constraint comes from the consistency under time evolution of the primary contraint: \begin{eqnarray} \Psi_2 \equiv \dot \pi_S^0 = \{ \pi_S^0, H_T \}_P \approx 0, \label{Second1} \end{eqnarray} where $H_T$ is the Hamiltonian of the system at hand, which is defined as \begin{eqnarray} H_T &\equiv& \int d^3 x \, {\cal{H}}_T \nonumber\\ &=& \int d^3 x \, ( \pi_g^{\mu\nu} \dot g_{\mu\nu} + \pi_\phi \dot \phi + \pi_S^\mu \dot S_\mu + \pi_B \dot B + \pi_{c \mu} \dot c^\mu + \pi_{\bar c}^\mu \dot {\bar c}_\mu \nonumber\\ &+& \pi_c \dot c + \pi_{\bar c} \dot{\bar c} - {\cal L}_q ). \label{Hamil} \end{eqnarray} In order to obtain the Hamiltonian, we have to express the time derivatives of the canonical variables in terms of the canonical conjugate momenta in (\ref{CCM}). To do that, let us first consider $\pi_B$, which gives us the expression of $\dot \phi$ as \begin{eqnarray} \dot \phi = \tilde f \left( \frac{1}{\phi} \pi_B - \tilde g^{0i} \partial_i \phi \right). \label{dot-phi} \end{eqnarray} Next, let us turn our attention to the $(kl)$-components of $\pi_g^{\mu\nu}$, which take the form: \begin{eqnarray} \pi_g^{kl} = \hat A^{kl} + \hat B^{kl \rho} b_\rho + \hat C^{klmn} \dot g_{mn} + \hat D^{kl} \dot \phi, \label{dot-pi-kl} \end{eqnarray} where $\hat A^{kl}, \hat B^{kl \rho}, \hat C^{klmn}$ and $\hat D^{kl}$ commute with $g_{mn}$ and are defined as \begin{eqnarray} \hat A^{kl} &=& -\frac{1}{4} \sqrt{-g} \phi^2 \Bigl[ - g^{0m} g^{kl} g^{\sigma\tau} - g^{0\tau} g^{km} g^{l\sigma} - g^{0\sigma} g^{k\tau} g^{lm} + g^{0m} g^{k\tau} g^{l\sigma} + g^{0\tau} g^{kl} g^{m\sigma} \nonumber\\ &+& \frac{1}{2} ( g^{0k} g^{lm} + g^{0l} g^{km} ) g^{\sigma\tau} \Bigr] \partial_m g_{\sigma\tau} - \sqrt{-g} \xi \phi \left[ \frac{1}{2} ( g^{0k} g^{lm} + g^{0l} g^{km} ) - g^{kl} g^{0m} \right] \partial_m \phi, \nonumber\\ \hat B^{kl \rho} &=& -\frac{1}{2} \sqrt{-g} \phi^2 ( g^{0k} g^{l\rho} + g^{0l} g^{k\rho} - g^{0\rho} g^{kl} ), \nonumber\\ \hat C^{klmn} &=& -\frac{1}{4} \sqrt{-g} \xi \phi^2 ( - g^{00} g^{kl} g^{mn} - g^{0n} g^{0k} g^{lm} - g^{0m} g^{kn} g^{0l} + g^{00} g^{kn} g^{lm} \nonumber\\ &+& g^{0n} g^{kl} g^{0m} + g^{0k} g^{0l} g^{mn} ), \nonumber\\ \hat D^{kl} &=& \sqrt{-g} \xi \phi ( g^{00} g^{kl} - g^{0k} g^{0l} ). \label{hat-ABCD} \end{eqnarray} Solving (\ref{dot-pi-kl}) with respect to $\dot g_{kl}$ together with Eq. (\ref{dot-phi}) leads to: \begin{eqnarray} \dot g_{kl} = \hat C_{klmn}^{-1} \left[ \pi_g^{mn} - \hat A^{mn} - \hat B^{mn \rho} b_\rho - \hat D^{mn} \tilde f \left( \frac{1}{\phi} \pi_B - \tilde g^{0i} \partial_i \phi \right) \right], \label{dot-g-kl} \end{eqnarray} where $\hat C_{klmn}^{-1}$ is the inverse matrix of $\hat C^{klmn}$ given by \begin{eqnarray} \hat C_{klmn}^{-1} &=& \frac{2}{\xi \phi^2} \tilde f ( g_{kl} g_{mn} - g_{km} g_{ln} - g_{kn} g_{lm} ), \nonumber\\ \hat C^{klmn} \hat C_{mnij}^{-1} &=& \frac{1}{2} ( \delta_i^k \delta_j^l + \delta_i^l \delta_j^k ). \label{Inv-C} \end{eqnarray} Using the extended de Donder gauge condition (\ref{Ext-de-Donder}), $\dot g_{00}$ and $\dot g_{0k}$ are described as \begin{eqnarray} \dot g_{00} &=& \frac{1}{g^{00}} \left( g^{ij} \dot g_{ij} - 2 g^{\alpha i} \partial_i g_{0 \alpha} + \frac{4}{\phi} \dot \phi \right) \nonumber\\ &=& \frac{1}{g^{00}} \Biggl\{ g^{kl} \hat C_{klmn}^{-1} \left[ \pi_g^{mn} - \hat A^{mn} - \hat B^{mn \rho} b_\rho - \hat D^{mn} \tilde f \left( \frac{1}{\phi} \pi_B - \tilde g^{0i} \partial_i \phi \right) \right] \nonumber\\ &-& 2 g^{\alpha i} \partial_i g_{0 \alpha} + \frac{4}{\phi} \tilde f \left( \frac{1}{\phi} \pi_B - \tilde g^{0i} \partial_i \phi \right) \Biggr\}, \nonumber\\ \dot g_{0k} &=& \frac{1}{g^{00}} \left( - g^{0j} \dot g_{jk} - g^{\alpha i} \partial_i g_{\alpha k} + \frac{1}{2} g^{\alpha\beta} \partial_k g_{\alpha\beta} + \frac{2}{\phi} \partial_k \phi \right) \nonumber\\ &=& \frac{1}{g^{00}} \Biggl\{ - g^{0j} \hat C_{jkmn}^{-1} \left[ \pi_g^{mn} - \hat A^{mn} - \hat B^{mn \rho} b_\rho - \hat D^{mn} \tilde f \left( \frac{1}{\phi} \pi_B - \tilde g^{0i} \partial_i \phi \right) \right] \nonumber\\ &-& g^{\alpha i} \partial_i g_{\alpha k} + \frac{1}{2} g^{\alpha\beta} \partial_k g_{\alpha\beta} + \frac{2}{\phi} \partial_k \phi \Biggr\}. \label{dot-g-0mu} \end{eqnarray} In a similar manner, based on $\pi_\phi, \pi_S^\mu, \pi_{c \sigma}, \pi_{\bar c}^\sigma, \pi_c$ and $\pi_{\bar c}$ in Eq. (\ref{CCM}), the time derivatives $\dot B, \dot S_k, \dot S_0, \dot{\bar c}_\sigma, \dot c^\sigma, \dot{\bar c}$ and $\dot c$ can be expressed in terms of the canonical conjugate momenta as follows: \begin{eqnarray} \dot B &=& \tilde f \frac{1}{\phi} \Bigl[ \pi_\phi + \epsilon \frac{1}{\phi} \pi_B - ( 6 \xi + \epsilon ) \tilde g^{0\mu} \phi S_\mu - 2 \tilde g^{0\mu} \phi b_\mu - \tilde g^{0i} \phi \partial_i B \Bigr] \nonumber\\ &-& \xi \tilde f \Biggl\{ ( \tilde g^{00} g^{ij} - \tilde g^{0i} g^{0j} ) \hat C_{ijmn}^{-1} \Bigl[ \pi_g^{mn} - \hat A^{mn} - \hat B^{mn \rho} b_\rho - \xi \tilde f ( \tilde g^{00} g^{mn} \nonumber\\ &-& \tilde g^{0m} g^{0n} ) ( \pi_B - \tilde g^{0k} \phi \partial_k \phi ) \Bigr] + ( \tilde g^{0i} g^{\alpha\beta} - \tilde g^{0\alpha} g^{i\beta} ) \partial_i g_{\alpha\beta} \Biggr\}, \nonumber\\ \dot S_k &=& \partial_k S_0 + \tilde f ( - g_{kj} \pi_S^j + \tilde g^{0j} H_{kj} ), \nonumber\\ \dot S_0 &=& - \tilde f \left\{ \tilde g^{0i} \left[ 2 \partial_i S_0 + \tilde f ( - g_{ij} \pi_S^j + \tilde g^{0j} H_{ij} ) \right] + \tilde g^{ij} \partial_i S_j \right\}, \nonumber\\ \dot {\bar c}_\sigma &=& i \tilde f \phi^{-2} \pi_{c \sigma} - \tilde f \tilde g^{0i} \partial_i \bar c_\sigma, \nonumber\\ \dot c^\sigma &=& - i \tilde f \phi^{-2} \pi_{\bar c}^\sigma - \tilde f \tilde g^{0i} \partial_i c^\sigma, \nonumber\\ \dot {\bar c} &=& i \tilde f \phi^{-2} \pi_c - \tilde f \tilde g^{0i} \partial_i \bar c, \nonumber\\ \dot c &=& - i \tilde f \phi^{-2} \pi_{\bar c} - \tilde f \tilde g^{0i} \partial_i c, \label{dot-many} \end{eqnarray} where we have used Eq. (\ref{S-eq}) in deriving $\dot S_0$. Finally, we can also express the $b_\mu$ field in terms of canonical conjugate momenta. Since the $b_\mu$ field is regarded as a conjugate momentum of $\tilde g^{0\mu}$, we begin with $\pi_g^{\alpha0}$ which has a structure: \begin{eqnarray} \pi_g^{\alpha0} = A^\alpha + B^{\alpha\beta} \partial_\beta \phi + C^{\alpha\beta} b_\beta, \label{pi-0alpha} \end{eqnarray} where $A^\alpha, B^{\alpha\beta}$ and $C^{\alpha\beta} = - \frac{1}{2} \tilde g^{00} g^{\alpha\beta} \phi^2$ do not include $\dot g_{\mu\nu}$, and $B^{\alpha\beta} \partial_\beta \phi$ does not have $\dot \phi$. Solving this equation with respect to $b_\mu$ leads to: \begin{eqnarray} b_\mu &=& - 2 \tilde f \phi^{-2} g_{\mu\alpha} \pi_g^{\alpha0} - \frac{1}{2} \xi \tilde f \biggl[ \delta_\mu^0 \left( \tilde g^{0\tau} g^{\lambda\sigma} - \frac{1}{2} \tilde g^{0\lambda} g^{\sigma\tau} \right) \partial_\lambda g_{\sigma\tau} \nonumber\\ &-& \left( \tilde g^{0\tau} g^{0\sigma} - \frac{1}{2} \tilde g^{00} g^{\sigma\tau} \right) \partial_\mu g_{\sigma\tau} \biggr] - \xi \phi^{-1} \left( \partial_\mu \phi - \delta_\mu^0 \tilde f \tilde g^{0\alpha} \partial_\alpha \phi \right). \label{b-mu} \end{eqnarray} Note that the RHS of this equation does not involve $\dot g_{\mu\nu}$ and $\dot \phi$ as can be verified explicitly. Incidentally, the relation (\ref{pi-0alpha}) is utilized to derive some useful Poisson brackets such as $\{ g_{\mu\nu}, b_\rho^\prime \}_P$ etc. Using the Hamiltonian $H_T$, Eq. (\ref{Second1}) provides us with a secondary constraint: \begin{eqnarray} \Psi_2 = \partial_i \pi_S^i + ( 6 \xi + \epsilon ) ( \pi_B - \tilde g^{0 \mu} \phi^2 S_\mu ) \approx 0, \label{Second2} \end{eqnarray} which is just the same as the $(0 \mu)$-components of the field equation for $S_\mu$ in (\ref{q-field-eq}) and there are no more constraints since we can show that\footnote{A derivation of constraints is exhibited in Appendix A.} \begin{eqnarray} \dot \Psi_2 = \{ \Psi_2, H_T \}_P = 0. \label{Tert} \end{eqnarray} The Poisson bracket between the constraints is evaluated to be: \begin{eqnarray} \{ \Psi_1, \Psi_2^\prime \}_P = ( 6 \xi + \epsilon ) \tilde g^{00} \phi^2 \delta^3 = ( 6 \xi + \epsilon ) \frac{1}{\tilde f} \phi^2 \delta^3, \label{PB-12C} \end{eqnarray} which implies that the constraints are the second-class constraint so that they can be treated by means of the Dirac bracket defined as \begin{eqnarray} \{ A, B^\prime \}_D \equiv \{ A, B^\prime \}_P - \{ A, \Psi_a^{\prime\prime} \}_P C_{ab}^{-1} \{ \Psi_b^{\prime\prime}, B^\prime \}_P, \label{DB} \end{eqnarray} where $\Psi_a ( a = 1, 2 )$ are the second-class constraints and $C_{ab}^{-1}$ is the inverse matrix of $C_{ab} = \{ \Psi_a, \Psi_b^\prime \}_P$. Concretely, the matrix elements, $C_{ab}^{-1}$, are given by \begin{eqnarray} C_{12}^{-1} = - C_{21}^{-1} = - \frac{1}{6 \xi + \epsilon} \tilde f \phi^{-2} \delta^3, \qquad C_{11}^{-1} = C_{22}^{-1} = 0. \label{C-1} \end{eqnarray} As is well known, the canonical quantization can be carried out by replacing $i \{ A, B^\prime \}_D$ with the equal-time commutation relation $[ A, B^\prime ]$. After some calculations, we can write down several important ETCRs, which are needed for later calculations: \begin{eqnarray} &{}& [ \dot g_{\rho\sigma}, g_{\mu\nu}^\prime ] = - \frac{2}{\xi} i \tilde f \phi^{-2} [ g_{\rho\sigma} g_{\mu\nu} - g_{\rho\mu} g_{\sigma\nu} - g_{\rho\nu} g_{\sigma\mu} + \sqrt{-g} \tilde f ( \delta_\rho^0 \delta_\mu^0 g_{\sigma\nu} \nonumber\\ &{}& + \delta_\rho^0 \delta_\nu^0 g_{\sigma\mu} + \delta_\sigma^0 \delta_\mu^0 g_{\rho\nu} + \delta_\sigma^0 \delta_\nu^0 g_{\rho\mu} ) ] \delta^3, \nonumber\\ &{}& [ \dot g_{\rho\sigma}, \phi^\prime ] = 0, \qquad [ \dot g_{\rho\sigma}, B^\prime ] = 2 i \tilde f \phi^{-2} g_{\rho\sigma} \delta^3, \nonumber\\ &{}& [ \phi, b_\rho^\prime ] = [ B, b_\rho^\prime ] = [ B, \dot B^\prime ] = [ \dot \phi, \phi^\prime ] = [ \dot \phi, S_\mu^\prime ] = 0, \qquad [ \dot \phi, B^\prime ] = - i \tilde f \phi^{-1} \delta^3, \nonumber\\ &{}& [ S_0, S_k^\prime ] = - \frac{1}{6 \xi + \epsilon} i \tilde f \phi^{-2} \partial_k \delta^3, \qquad [ S_k, S_l^\prime ] = 0, \nonumber\\ &{}& [ S_0, \dot S_k^\prime ] = - i \tilde f g_{0k} \delta^3, \qquad [ S_k, \dot S_l^\prime ] = - i \tilde f g_{kl} \delta^3 + \frac{1}{ 6 \xi + \epsilon } i \partial_k ( \tilde f \phi^{-2} \partial_l \delta^3 ), \nonumber\\ &{}& [ S_0, b_\mu^\prime ] = - i \tilde f \phi^{-2} S_\mu \delta^3, \qquad [ S_0, B^\prime ] = - i \tilde f \phi^{-2} \delta^3, \qquad [ S_k, b_\rho^\prime ] = [ S_k, B^\prime ] = 0, \nonumber\\ &{}& [ \dot S_0, B^\prime ] = 2 i \tilde f \tilde g^{0i} \partial_i ( \tilde f \phi^{-2} \delta^3 ), \qquad [ \dot S_k, B^\prime ] = - i \partial_k ( \tilde f \phi^{-2} \delta^3 ), \nonumber\\ &{}& \{ \dot{\bar c}_\lambda, c^{\sigma \prime} \} = - \{ \dot c^\sigma, \bar c_\lambda^\prime \} = - \tilde f \phi^{-2} \delta_\lambda^\sigma \delta^3, \qquad \{ \dot{\bar c}, c^\prime \} = - \{ \dot c, \bar c^\prime \} = - \tilde f \phi^{-2} \delta^3, \nonumber\\ &{}& [ g_{\mu\nu}, b_\rho^\prime ] = - i \tilde f \phi^{-2} ( \delta_\mu^0 g_{\rho\nu} + \delta_\nu^0 g_{\rho\mu} ) \delta^3, \quad [ \tilde g^{\mu\nu}, b_\rho^\prime ] = i \tilde f \phi^{-2} ( \tilde g^{\mu0} \delta_\rho^\nu + \tilde g^{\nu0} \delta_\rho^\mu - \tilde g^{\mu\nu} \delta_\rho^0 ) \delta^3, \nonumber\\ &{}& [ g_{\mu\nu}, \dot b_\rho^\prime ] = i \{ [ \tilde f \phi^{-2} \partial_\rho g_{\mu\nu} - \partial_0 ( \tilde f \phi^{-2} ) ( \delta_\mu^0 g_{\rho\nu} + \delta_\nu^0 g_{\rho\mu} ) ] \delta^3 + [ ( \delta_\mu^k - 2 \delta_\mu^0 \tilde f \tilde g^{0k} ) g_{\rho\nu} \nonumber\\ &{}& + ( \mu \leftrightarrow \nu ) ] \partial_k ( \tilde f \phi^{-2} \delta^3 ) \}, \nonumber\\ &{}& [ \dot \phi, b_\rho^\prime ] = - i \tilde f \phi^{-2} \partial_\rho \phi \delta^3, \qquad [ \dot B, b_\rho^\prime ] = - i \tilde f \phi^{-2} \partial_\rho B \delta^3, \nonumber\\ &{}& [ \dot S_0, b_\mu^\prime ] = - i \tilde f^2 \phi^{-2} [ \tilde g^{0\nu} ( \partial_\mu S_\nu + \partial_\nu S_\mu ) - \tilde g^{0i} H_{\mu i} ] \delta^3 + 2 i \tilde f \tilde g^{0i} \partial_i ( \tilde f \phi^{-2} S_\mu \delta^3 ), \nonumber\\ &{}& [ \dot S_k, b_\mu^\prime ] = i \tilde f \phi^{-2} H_{k \mu} \delta^3 - i \partial_k ( \tilde f \phi^{-2} S_\mu \delta^3 ), \nonumber\\ &{}& [ b_\mu, b_\nu^\prime ] = 0, \qquad [ b_\mu, \dot b_\nu^\prime ] = i \tilde f \phi^{-2} ( \partial_\mu b_\nu + \partial_\nu b_\mu ) \delta^3, \nonumber\\ &{}& [ b_\rho, c^{\sigma\prime} ] = [ b_\rho, \bar c_\lambda^\prime ] = [ b_\rho, c^\prime ] = [ b_\rho, \bar c^\prime ] = 0, \nonumber\\ &{}& [ \dot{\bar c}_\lambda, b_\rho^\prime ] = - i \tilde f \phi^{-2} \partial_\rho \bar c_\lambda \delta^3, \qquad [ \dot c^\sigma, b_\rho^\prime ] = - i \tilde f \phi^{-2} \partial_\rho c^\sigma \delta^3, \nonumber\\ &{}& [ \dot{\bar c}, b_\rho^\prime ] = - i \tilde f \phi^{-2} \partial_\rho \bar c \delta^3, \qquad [ \dot c, b_\rho^\prime ] = - i \tilde f \phi^{-2} \partial_\rho c \delta^3. \label{ETCRs} \end{eqnarray} These ETCRs can be obtained from the explicit calculations and/or the BRST transformations. For instance, we will present a derivation of $[ B, \dot B^\prime ] = 0$ by the both methods. First, let us focus on the explicit calculation via the Dirac bracket: \begin{eqnarray} \{ B, \dot B^\prime \}_D = \{ B, \dot B^\prime \}_P - \{ B, \Psi_2^{\prime\prime} \}_P C_{21}^{-1} \{ \Psi_1^{\prime\prime}, \dot B^\prime \}_P. \label{B-B'} \end{eqnarray} Since we can easily evaluate each Poisson bracket whose result reads: \begin{eqnarray} \{ B, \dot B^\prime \}_P &=& \{ B, ( 6 \xi + \epsilon ) \tilde f \phi^{-2} \pi_B^\prime \}_P = ( 6 \xi + \epsilon ) \tilde f \phi^{-2} \delta^3, \nonumber\\ \{ B, \Psi_2^\prime \}_P &=& \{ B, ( 6 \xi + \epsilon ) \pi_B^\prime \}_P = ( 6 \xi + \epsilon ) \delta^3, \nonumber\\ \{ \Psi_1, \dot B^\prime \}_P &=& \{ \pi_S^0, - ( 6 \xi + \epsilon ) \tilde f \tilde g^{0\mu} S_\mu^\prime \} = ( 6 \xi + \epsilon ) \delta^3, \label{B-B'2} \end{eqnarray} the Dirac bracket becomes: \begin{eqnarray} \{ B, \dot B^\prime \}_D &=& ( 6 \xi + \epsilon ) \tilde f \phi^{-2} \delta^3 - ( 6 \xi + \epsilon ) C_{21}^{-1} ( 6 \xi + \epsilon ) \delta^3 \nonumber\\ &=& 0. \label{B-B'3} \end{eqnarray} Second, we will present a derivation by means of the BRST transformation which is more general and elegant than the above explicit calculation. The ETCR, $[ B, \pi_c^\prime ] = 0$, leads to $[ B, \dot{\bar{c}}^\prime ] = 0$. Taking the Weyl BRST transformation of this ETCR yields the equation: \begin{eqnarray} \{ [ i \bar Q_B, B ], \dot{\bar{c}}^\prime \} + [ B, \{ i \bar Q_B, \dot{\bar{c}}^\prime \} ] = 0. \label{W-B-B'} \end{eqnarray} Then, the Weyl BRST transformation (\ref{Weyl-BRST}) immediately leads to $[ B, \dot B^\prime ] = 0$. \section{Unitarity analysis} As in the conventional BRST formalism, the physical state $| \rm{phys} \rangle$ is defined by imposing two subsidiary conditions \cite{Kugo-Ojima}: \begin{eqnarray} Q_B | \rm{phys} \rangle = \bar Q_B | \rm{phys} \rangle = 0. \label{Phys-state} \end{eqnarray} It is then well known that the physical S-matrix is unitary under the assumption that all BRST singlet states have positive norm. In this section, we would like to prove the unitarity of the physical S-matrix in the present theory. From the classical analysis we know that the gauge field becomes massive via the Higgs mechanism. Thus, we wish to understand how the Higgs mechanism is described in terms of the BRST formalism. In analysing the unitarity, it is enough to take account of asymptotic fields of all the fundamental fields and the free part of the Lagrangian. Let us first assume the asymptotic fields as \begin{eqnarray} g_{\mu\nu} &=& \eta_{\mu\nu} + \varphi_{\mu\nu}, \qquad \phi = \phi_0 + \tilde \phi, \qquad S_\mu = s_\mu, \qquad b_\mu = \beta_\mu, \qquad B = \beta, \nonumber\\ c^\mu &=& \gamma^\mu, \qquad \bar c_\mu = \bar \gamma_\mu, \qquad c = \gamma, \qquad \bar c = \bar \gamma, \label{Asmp-exp} \end{eqnarray} where $\eta_{\mu\nu} ( = \eta^{\mu\nu} )$ is the flat Minkowski metric with the mostly positive signature and $\phi_0$ is a non-zero constant. In this section, the Minkowski metric is used to lower or raise the Lorentz indices. Using these asymptotic fields, the free part of the Lagrangian reads: \begin{eqnarray} {\cal L}_q &=& \frac{1}{2} \xi \phi_0^2 \left( \frac{1}{4} \varphi_{\mu\nu} \Box \varphi^{\mu\nu} - \frac{1}{4} \varphi \Box \varphi - \frac{1}{2} \varphi^{\mu\nu} \partial_\mu \partial_\rho \varphi_\nu{}^\rho + \frac{1}{2} \varphi^{\mu\nu} \partial_\mu \partial_\nu \varphi \right) \nonumber\\ &+& \xi \phi_0 \tilde \phi \left( - \Box \varphi + \partial_\mu \partial_\nu \varphi^{\mu\nu} \right) + ( 6 \xi + \epsilon ) \left( \phi_0 S_\mu \partial^\mu \tilde \phi - \frac{1}{2} \phi_0^2 S_\mu S^\mu \right) \nonumber\\ &-& \frac{1}{4} h_{\mu\nu}^2 - \frac{1}{2} \epsilon \partial_\mu \tilde \phi \partial^\mu \tilde \phi - \left( 2 \eta^{\mu\nu} \phi_0 \tilde \phi - \phi_0^2 \varphi^{\mu\nu} + \frac{1}{2} \phi_0^2 \eta^{\mu\nu} \varphi \right) \partial_\mu \beta_\nu \nonumber\\ &-& i \phi_0^2 \partial_\mu \bar \gamma_\rho \partial^\mu \gamma^\rho + \phi_0 \partial_\mu \beta \partial^\mu \tilde \phi - i \phi_0^2 \partial_\mu \bar \gamma \partial^\mu \gamma, \label{Free-Lag} \end{eqnarray} where $\Box \equiv \eta^{\mu\nu} \partial_\mu \partial_\nu, \, \varphi \equiv \eta^{\mu\nu} \varphi_{\mu\nu}$ and $h_{\mu\nu} \equiv \partial_\mu s_\nu - \partial_\nu s_\mu$. Based on this Lagrangian, it is easy to derive the linearized field equations: \begin{eqnarray} &{}& \frac{1}{2} \xi \phi_0^2 \biggl( \frac{1}{2} \Box \varphi_{\mu\nu} - \frac{1}{2} \eta_{\mu\nu} \Box \varphi - \partial_\rho \partial_{(\mu} \varphi_{\nu)}{}^\rho + \frac{1}{2} \partial_\mu \partial_\nu \varphi + \frac{1}{2} \eta_{\mu\nu} \partial_\rho \partial_\sigma \varphi^{\rho\sigma} \biggr) \nonumber\\ &{}& + \xi \phi_0 \left( - \eta_{\mu\nu} \Box + \partial_\mu \partial_\nu \right) \tilde \phi +\phi_0^2 \partial_{(\mu} \beta_{\nu)} - \frac{1}{2} \phi_0^2 \eta_{\mu\nu} \partial_\rho \beta^\rho = 0. \label{Linear-Eq1} \\ &{}& \epsilon \Box \tilde \phi + \xi \phi_0 ( - \Box \varphi + \partial_\mu \partial_\nu \varphi^{\mu\nu} ) - ( 6 \xi + \epsilon ) \phi_0 \partial_\mu s^\mu - 2 \phi_0 \partial_\rho \beta^\rho \nonumber\\ &{}& - \phi_0 \Box \beta = 0. \label{Linear-Eq2} \\ &{}& \partial^\nu h_{\mu\nu} + ( 6 \xi + \epsilon ) \phi_0^2 \left( s_\mu - \frac{1}{\phi_0} \partial_\mu \tilde \phi \right) = 0. \label{Linear-Eq3} \\ &{}& \partial_\mu \tilde \phi - \frac{1}{2} \phi_0 \left( \partial^\nu \varphi_{\mu\nu} - \frac{1}{2} \partial_\mu \varphi \right) = 0. \label{Linear-Eq4} \\ &{}& \Box \tilde \phi = \Box \gamma^\mu = \Box \bar \gamma_\mu = \Box \gamma = \Box \bar \gamma = 0. \label{Linear-Eq5} \end{eqnarray} Here we have introduced the symmetrization notation $A_{(\mu} B_{\nu)} \equiv \frac{1}{2} ( A_\mu B_\nu + A_\nu B_\mu )$. Now, operating $\partial^\mu$ on Eq. (\ref{Linear-Eq4}) and using Eq. (\ref{Linear-Eq5}), we obtain: \begin{eqnarray} \partial_\mu \partial_\nu \varphi^{\mu\nu} - \frac{1}{2} \Box \varphi = 0. \label{Linear-Eq6} \end{eqnarray} Next, taking the trace of Eq. (\ref{Linear-Eq1}) with the help of Eqs. (\ref{Linear-Eq5}) and (\ref{Linear-Eq6}) leads to: \begin{eqnarray} \Box \varphi + \frac{4}{\xi} \partial_\rho \beta^\rho = 0. \label{Linear-Eq7} \end{eqnarray} Moreover, operating $\partial^\mu$ on Eq. (\ref{Linear-Eq3}), and using the identity $\partial^\mu \partial^\nu h_{\mu\nu} = 0$ and Eq. (\ref{Linear-Eq5}) yields the Lorenz condition: \begin{eqnarray} \partial_\mu s^\mu = 0. \label{Linear-Lorenz} \end{eqnarray} As can been seen in Eq. (\ref{Linear-Eq3}), it is more convenient to introduce $\hat s_\mu$ defined as \begin{eqnarray} \hat s_\mu = s_\mu - \frac{1}{\phi_0} \partial_\mu \tilde \phi, \label{hat-s} \end{eqnarray} which also obeys the Lorenz condition owing to Eqs. (\ref{Linear-Eq5}) and (\ref{Linear-Lorenz}): \begin{eqnarray} \partial_\mu \hat s^\mu = 0. \label{Linear-Lorenz2} \end{eqnarray} With the new gauge field $\hat s_\mu$ and the corresponding field strength $\hat h_{\mu\nu} \equiv \partial_\mu \hat s_\nu - \partial_\nu \hat s_\mu$, the ``Maxwell equation'' (\ref{Linear-Eq3}) can be cast to the form: \begin{eqnarray} \partial^\nu \hat h_{\mu\nu} + ( 6 \xi + \epsilon ) \phi_0^2 \hat s_\mu = 0, \label{Mass-Max} \end{eqnarray} which clearly shows that the Weyl gauge field absorbs the Nambu-Goldstone boson $\tilde \phi$ associated with spontaneous symmetry breakdown of the Weyl gauge symmetry, thereby becoming massive with the mass squared $( 6 \xi + \epsilon ) \phi_0^2$ (Here we assume $6 \xi + \epsilon > 0$, which is consistent with the positive Newton constant $\xi > 0$). To put it differently, after spontaneous symmetry breakdown of the Weyl gauge symmetry, the Weyl gauge field $\hat s_\mu$ satisfies not only the Lorenz condition (\ref{Linear-Lorenz2}) but also the massive Klein-Gordon equation: \begin{eqnarray} ( \Box - m^2 ) \hat s_\mu = 0, \label{Mass-KG} \end{eqnarray} where $m^2$ is defined by \begin{eqnarray} m^2 \equiv ( 6 \xi + \epsilon ) \phi_0^2. \label{Mass-squared} \end{eqnarray} Furthermore, with the help of Eqs. (\ref{Linear-Eq5}), (\ref{Linear-Eq6}), (\ref{Linear-Eq7}) and (\ref{Linear-Lorenz}), Eq. (\ref{Linear-Eq2}) can be rewritten as \begin{eqnarray} \Box \beta = 0. \label{Linear-Eq8} \end{eqnarray} Moreover, acting $\partial^\mu$ on Eq. (\ref{Linear-Eq1}) yields: \begin{eqnarray} \Box \beta_\mu = 0. \label{Linear-Eq9} \end{eqnarray} Finally, using various equations obtained thus far, the ``Einstein equation'' (\ref{Linear-Eq1}) is reduced to the form: \begin{eqnarray} \Box \varphi_{\mu\nu} + \frac{4}{\xi} \partial_{(\mu} \beta_{\nu)} = 0, \label{Linear-Eq10} \end{eqnarray} which means that the field $\varphi_{\mu\nu}$ is not a simple pole field but a dipole field: \begin{eqnarray} \Box^2 \varphi_{\mu\nu} = 0. \label{Linear-Eq11} \end{eqnarray} On the other hand, in addition to Eq. (\ref{Mass-KG}), the other fields are all simple pole fields: \begin{eqnarray} \Box \tilde \phi = \Box \beta_\mu = \Box \beta =\Box \gamma^\mu = \Box \bar \gamma_\mu = \Box \gamma = \Box \bar \gamma = 0. \label{Linear-Eq12} \end{eqnarray} Note that Eq. (\ref{Linear-Eq12}) corresponds to Eq. (\ref{X-M-eq}) in a curved space-time. Following the standard technique, we can calculate the four-dimensional (anti-)commutation relations (4D CRs) between asymptotic fields. The point is that the simple pole fields, for instance, the Nakanishi-Lautrup field $\beta_\mu (x)$ can be expressed in terms of the invariant delta function $D(x)$ as \begin{eqnarray} \beta_\mu (x) = - \int d^3 z D(x-z) \overleftrightarrow{\partial}_0^z \beta_\mu (z), \label{D-beta} \end{eqnarray} whereas the dipole field $\varphi_{\mu\nu}(x)$ takes the form: \begin{eqnarray} &{}& \varphi_{\mu\nu} (x) = - \int d^3 z \left[ D(x-z) \overleftrightarrow{\partial}_0^z \varphi_{\mu\nu} (z) + E(x-z) \overleftrightarrow{\partial}_0^z \Box \varphi_{\mu\nu} (z) \right] \nonumber\\ &{}& = - \int d^3 z \left[ D(x-z) \overleftrightarrow{\partial}_0^z \varphi_{\mu\nu} (z) - \frac{4}{\xi} E(x-z) \overleftrightarrow{\partial}_0^z \partial_{(\mu} \beta_{\nu)} (z) \right], \label{E-varphi} \end{eqnarray} where in the last equality we have used Eq. (\ref{Linear-Eq10}). Here the invariant delta function $D(x)$ for massless simple pole fields and its properties are described as \begin{eqnarray} &{}& D(x) = - \frac{i}{(2 \pi)^3} \int d^4 k \, \epsilon (k_0) \delta (k^2) e^{i k x}, \qquad \Box D(x) = 0, \nonumber\\ &{}& D(-x) = - D(x), \qquad D(0, \vec{x}) = 0, \qquad \partial_0 D(0, \vec{x}) = \delta^3 (x), \label{D-function} \end{eqnarray} where $\epsilon (k_0) \equiv \frac{k_0}{|k_0|}$. Similarly, the invariant delta function $E(x)$ for massless dipole fields and its properties are given by \begin{eqnarray} &{}& E(x) = - \frac{i}{(2 \pi)^3} \int d^4 k \, \epsilon (k_0) \delta^\prime (k^2) e^{i k x}, \qquad \Box E(x) = D(x), \nonumber\\ &{}& E(-x) = - E(x), \qquad E(0, \vec{x}) = \partial_0 E(0, \vec{x}) = \partial_0^2 E(0, \vec{x}) = 0, \nonumber\\ &{}& \partial_0^3 E(0, \vec{x}) = - \delta^3 (x), \label{E-function} \end{eqnarray} where $\delta^\prime (k^2) \equiv \frac{d \delta (k^2)}{d k^2}$. On the other hand, the Weyl gauge field $\hat s(x)$ obeys the massive Klein-Gordon equation (\ref{Mass-KG}), so it needs to be described in terms of the invariant delta function $\Delta(x; m^2)$ for massive simple pole fields as \begin{eqnarray} \hat s_\mu (x) = - \int d^3 z \Delta (x-z; m^2) \overleftrightarrow{\partial}_0^z \hat s_\mu (z), \label{s-Delta} \end{eqnarray} where $\Delta(x; m^2)$ is defined as \begin{eqnarray} &{}& \Delta(x; m^2) = - \frac{i}{(2 \pi)^3} \int d^4 k \, \epsilon (k_0) \delta (k^2 + m^2) e^{i k x}, \quad (\Box - m^2) \Delta(x; m^2) = 0, \nonumber\\ &{}& \Delta(-x; m^2) = - \Delta(x; m^2), \quad \Delta(0, \vec{x}; m^2) = 0, \nonumber\\ &{}& \partial_0 \Delta(0, \vec{x}; m^2) = \delta^3 (x), \qquad \Delta(x; 0) = D(x). \label{Delta-function} \end{eqnarray} It is easy to show that the RHS of Eqs. (\ref{D-beta}), (\ref{E-varphi}) and (\ref{s-Delta}) is independent of $z^0$. Thus, for instance, when we evaluate the four-dimensional commutation relation $[ \varphi_{\mu\nu} (x), \varphi_{\sigma\tau} (y) ]$, we can put $z^0 = y^0$ and use the three-dimensional commutation relations among asymptotic fields. After some manipulation, we find that the 4D CRs are given by \begin{eqnarray} &{}& [ \varphi_{\mu\nu} (x), \varphi_{\sigma\tau} (y) ] = - \frac{2}{\xi} i \phi_0^{-2} [ ( \eta_{\mu\nu} \eta_{\sigma\tau} - \eta_{\mu\sigma} \eta_{\nu\tau} - \eta_{\mu\tau} \eta_{\nu\sigma} ) D(x-y) \nonumber\\ &{}& + ( \eta_{\mu\sigma} \partial_\nu \partial_\tau + \eta_{\nu\sigma} \partial_\mu \partial_\tau + \eta_{\mu\tau} \partial_\nu \partial_\sigma + \eta_{\nu\tau} \partial_\mu \partial_\sigma ) E(x-y) ], \label{4D-CR1} \\ &{}& [ \varphi_{\mu\nu} (x), \beta_\rho (y) ] = - i \phi_0^{-2} ( \eta_{\mu\rho} \partial_\nu + \eta_{\nu\rho} \partial_\mu ) D(x-y). \label{4D-CR2} \\ &{}& [ \varphi_{\mu\nu} (x), \beta (y) ] = 2 i \phi_0^{-1} \eta_{\mu\nu} D(x-y). \label{4D-CR3} \\ &{}& [ \tilde \phi (x), \beta (y) ] = - i \phi_0^{-1} D(x-y). \label{4D-CR4} \\ &{}& [ \hat s_\mu (x), \hat s_\nu (y) ] = i \left( \eta_{\mu\nu} - \frac{1}{m^2} \partial_\mu \partial_\nu \right) \Delta (x-y; m^2). \label{4D-CRs1} \\ &{}& \{ \gamma^\sigma (x), \bar \gamma_\tau (y) \} = \phi_0^{-2} \delta_\tau^\sigma D(x-y). \label{4D-CR5} \\ &{}& \{ \gamma (x), \bar \gamma (y) \} = \phi_0^{-2} D(x-y). \label{4D-CR6} \end{eqnarray} The other 4D CRs vanish identically. Now we would like to discuss the issue of the unitarity of the physical S-matrix. To do that, it is convenient to perform the Fourier transformation of Eqs. (\ref{4D-CR1})-(\ref{4D-CR6}). However, for the dipole field we cannot use the three-dimensional Fourier expansion to define the creation and annihilation operators. We therefore make use of the four-dimensional Fourier expansion \cite{N-O-text}:\footnote{For simplicity, the Fourier transform of a field is denoted by the same field except for the argument $p$ instead of $x$.} \begin{eqnarray} \varphi_{\mu\nu} (x) = \frac{1}{(2 \pi)^{\frac{3}{2}}} \int d^4 p \, \theta (p_0) [ \varphi_{\mu\nu} (p) e^{i p x} + \varphi_{\mu\nu}^\dagger (p) e^{- i p x} ], \label{FT-varphi} \end{eqnarray} where $\theta (p_0)$ is the step function. For any simple pole fields, we adopt the same Fourier expansion, for instance, \begin{eqnarray} \beta_\mu (x) = \frac{1}{(2 \pi)^{\frac{3}{2}}} \int d^4 p \, \theta (p_0) [ \beta_\mu (p) e^{i p x} + \beta_\mu^\dagger (p) e^{- i p x} ]. \label{FT-beta} \end{eqnarray} Thus, using Eqs. (\ref{D-beta}), (\ref{E-varphi}), (\ref{FT-varphi}) and (\ref{FT-beta}), for instance, the Fourier transforms of, e.g., $\varphi_{\mu\nu} (x)$ and $\beta_\mu (x)$ take the following expression: \begin{eqnarray} \varphi_{\mu\nu} (p) &=& \frac{i}{(2 \pi)^{\frac{3}{2}}} \theta( p_0 ) \int d^3 z \, e^{-i p z} \overleftrightarrow{\partial}_0^z [ \delta(p^2) \varphi_{\mu\nu} (z) + \delta^\prime (p^2) \Box \varphi_{\mu\nu} (z) ], \nonumber\\ \beta_\mu (p) &=& \frac{i}{(2 \pi)^{\frac{3}{2}}} \theta( p_0 ) \delta(p^2) \int d^3 z \, e^{-i p z} \overleftrightarrow{\partial}_0^z \beta_\mu (z). \label{FT-fields} \end{eqnarray} Incidentally, for a generic simple pole field $\Phi$ with a mass $m$, the three-dimensional Fourier expansion is defined as \begin{eqnarray} \Phi (x) = \frac{1}{(2 \pi)^{\frac{3}{2}}} \int d^3 p \, \frac{1}{\sqrt{2 \omega_p}} [ \Phi (\vec{p}) e^{i p x} + \Phi^\dagger (\vec{p}) e^{ - i p x } ], \label{3D-FT} \end{eqnarray} with being $\omega_p = \sqrt{ \vec{p}^2 + m^2}$, whereas the four-dimensional Fourier expansion reads: \begin{eqnarray} \Phi (x) = \frac{1}{(2 \pi)^{\frac{3}{2}}} \int d^4 p \, \theta (p_0) [ \Phi (p) e^{i p x} + \Phi^\dagger (p) (p) e^{- i p x} ]. \label{4D-FT} \end{eqnarray} Thus, the annihilation operator $\Phi (p)$ in the four-dimensional Fourier expansion has connection with the annihilation operator $\Phi (\vec{p})$ in the three-dimensional Fourier expansion via \begin{eqnarray} \Phi (p) = \theta (p_0) \delta (p^2 + m^2) \sqrt{2 \omega_p} \Phi (\vec{p}). \label{3D-4D} \end{eqnarray} Based on these Fourier expansions, we can calculate the Fourier transform of Eqs. (\ref{4D-CR1})-(\ref{4D-CR6}): \begin{eqnarray} &{}& [ \varphi_{\mu\nu} (p), \varphi_{\sigma\tau}^\dagger (q) ] = - \frac{2}{\xi} \phi_0^{-2} \theta (p_0) \delta^4 (p-q) [ \delta(p^2) ( \eta_{\mu\nu} \eta_{\sigma\tau}- \eta_{\mu\sigma} \eta_{\nu\tau} - \eta_{\mu\tau} \eta_{\nu\sigma} ) \nonumber\\ &{}& - 3 \delta^\prime (p^2) ( \eta_{\mu\sigma} p_\nu p_\tau + \eta_{\nu\sigma} p_\mu p_\tau + \eta_{\mu\tau} p_\nu p_\sigma + \eta_{\nu\tau} p_\mu p_\sigma ) ]. \label{FT-4D-CR1} \\ &{}& [ \varphi_{\mu\nu} (p), \beta_\rho^\dagger (q) ] = - i \phi_0^{-2} ( \eta_{\mu\rho} p_\nu + \eta_{\nu\rho} p_\mu ) \theta (p_0) \delta(p^2) \delta^4 (p-q). \label{FT-4D-CR2} \\ &{}& [ \varphi_{\mu\nu} (p), \beta^\dagger (q) ] = 2 \phi_0^{-1} \eta_{\mu\nu} \theta (p_0) \delta(p^2) \delta^4 (p-q). \label{FT-4D-CR3} \\ &{}& [ \tilde \phi (p), \beta^\dagger (q) ] = - \phi_0^{-1} \theta (p_0) \delta(p^2) \delta^4 (p-q). \label{FT-4D-CR4} \\ &{}& [ \hat s_\mu (p), \hat s_\nu^\dagger (q) ] = + \left( \eta_{\mu\nu} - \frac{1}{m^2} p_\mu p_\nu \right) \theta(p_0) \delta(p^2 + m^2) \delta^4 (p-q). \label{FT-4D-CRs} \\ &{}& \{ \gamma^\sigma (p), \bar \gamma^\dagger_\tau (q) \} = - i \phi_0^{-2} \delta_\tau^\sigma \theta (p_0) \delta(p^2) \delta^4 (p-q). \label{FT-4D-CR5} \\ &{}& \{ \gamma (p), \bar \gamma^\dagger (q) \} = - i \phi_0^{-2} \theta (p_0) \delta(p^2) \delta^4 (p-q). \label{FT-4D-CR6} \end{eqnarray} Next, let us turn our attention to the linearized field equations. After Fourier transformation, Eq. (\ref{Linear-Eq4}) takes the form: \begin{eqnarray} p^\nu \varphi_{\mu\nu} - \frac{1}{2} p_\mu \varphi = 2 \phi_0^{-1} p_\mu \tilde \phi. \label{FT-Linear-Eq3} \end{eqnarray} If we fix the degree of freedom associated with $\tilde \phi$, which will be discussed later, this equation gives us four independent relations on ten components of $\varphi_{\mu\nu} (p)$, thereby reducing the independent components of $\varphi_{\mu\nu} (p)$ to be six. To deal with six independent components of $\varphi_{\mu\nu} (p)$, it is convenient to take a specific Lorentz frame such that $p_1 = p_2 = 0$ and $p_3 > 0$, and choose the six components as follows: \begin{eqnarray} &{}& \varphi_1 (p) = \frac{1}{2} [ \varphi_{11} (p) - \varphi_{22} (p) ], \qquad \varphi_2 (p) = \varphi_{12} (p), \qquad \omega_0 (p) = - \frac{1}{2 p_0} \varphi_{00} (p), \nonumber\\ &{}& \omega_I (p) = - \frac{1}{p_0} \varphi_{0I} (p), \qquad \omega_3 (p) = - \frac{1}{2 p_3} \varphi_{33} (p), \label{Lorentz} \end{eqnarray} where the index $I$ takes the transverse components $I = 1, 2$. In this respect, it is worthwhile to consider the GCT BRST transformation for these components. First, let us write down the GCT BRST transformation for the Fourier expansion of the asymptotic fields, which reads: \begin{eqnarray} &{}& \delta_B \varphi_{\mu\nu} (p) = - i [ p_\mu \gamma_\nu (p) + p_\nu \gamma_\mu (p) ], \quad \delta_B \gamma^\mu (p) = 0, \quad \delta_B \bar \gamma_\mu (p) = i \beta_\mu (p), \nonumber\\ &{}& \delta_B \tilde \phi (p) = \delta_B \beta_\mu (p) = \delta_B \beta (p) = \delta_B \gamma (p) = \delta_B \bar \gamma (p) = 0. \label{Q_B-FT} \end{eqnarray} Using this BRST transformation, the GCT BRST transformation for the components in (\ref{Lorentz}) takes the form: \begin{eqnarray} &{}& \delta_B \varphi_I (p) = 0, \qquad \delta_B \omega_\mu (p) = i \gamma_\mu (p), \nonumber\\ &{}& \delta_B \bar \gamma_\mu (p) = i \beta_\mu (p), \qquad \delta_B \gamma_\mu (p) = \delta_B \beta_\mu (p) = 0, \label{Q_B-Comp} \end{eqnarray} where $p_1 = p_2 = 0$ was used. This BRST transformation implies that $\varphi_I (p)$ could be the physical observable while a set of fields, $\{ \omega_\mu (p), \beta_\mu (p), \gamma_\mu (p), \bar \gamma_\mu (p) \}$ might belong to the BRST quartet and thus are dropped from the physical state by the Kugo-Ojima subsidiary condition, $Q_B | \rm{phys} \rangle = 0$ \cite{Kugo-Ojima}.\footnote{The situation is in fact a bit complicated since $\beta_\mu (p), \gamma_\mu (p)$ and $\bar \gamma_\mu (p)$ are simple pole fields obeying $p^2 \beta_\mu (p) = p^2 \gamma_\mu (p) = p^2 \bar \gamma_\mu (p) = 0$, while $\varphi_{\mu\nu} (p)$ is a dipole field satisfying $( p^2 )^2 \varphi_{\mu\nu} (p) = 0$, so that a naive Kugo-Ojima's quartet mechanism does not work in a direct way. But this problem can be remedied by introducing an operator which takes out a simple pole from a dipole field. The detail can be shown in Ref. \cite{Oda-W}.} Next, let us move on to the other BRST transformation, which is the BRST transformation for the Weyl transformation. The Weyl BRST transformation for the asymptotic fields is of form: \begin{eqnarray} &{}& \bar \delta_B \varphi_{\mu\nu} = 2 c \eta_{\mu\nu}, \quad \bar \delta_B \tilde \phi = - \phi_0 \gamma, \quad \bar \delta_B \gamma = 0, \quad \bar \delta_B \bar \gamma = i \beta, \nonumber\\ &{}& \bar \delta_B \beta = \bar \delta_B \beta_\mu = \bar \delta_B \gamma_\mu = \bar \delta_B \bar \gamma_\mu = 0. \label{W-Q_B-Asym} \end{eqnarray} The Weyl BRST transformation of $\varphi_I$ is vanishing: \begin{eqnarray} \bar \delta_B \varphi_I = 0, \label{W-Q_B-Obs} \end{eqnarray} which means that together with $\delta_B \varphi_I = 0$, $\varphi_I$ is truely the physical observable. The four-dimensional commutation relations among the fields $\{ \tilde \phi, \beta, \gamma, \bar \gamma \}$ read: \begin{eqnarray} &{}& [ \tilde \phi (p), \tilde \phi^\dagger (q) ] = 0, \nonumber\\ &{}& [ \tilde \phi (p), \beta^\dagger (q) ] = - \phi_0^{-1} \theta (p_0) \delta (p^2) \delta^4 (p-q), \nonumber\\ &{}& \{ \gamma (p), \bar \gamma^\dagger (q) ] = - i \phi_0^{-2} \theta (p_0) \delta (p^2) \delta^4 (p-q). \label{W-4D-CRs} \end{eqnarray} As can be also seen in these 4D CRs, all the fields $\{ \varphi_I, \tilde \phi, \beta, \gamma, \bar \gamma \}$ are massless simple pole fields. Via relation (\ref{3D-4D}) the three-dimensional commutation relations $[ \Phi (\vec{p}), \Phi^\dagger (\vec{q}) \}$ with $\Phi (\vec{p}) \equiv \{ \varphi_I (\vec{p}), \tilde \phi (\vec{p}), \beta (\vec{p}), \gamma (\vec{p}), \bar \gamma (\vec{p}) \}$, are of form: \begin{eqnarray} [ \Phi (\vec{p}), \Phi^\dagger (\vec{q}) \} &=& \left( \begin{array}{cc|cc|cc} \frac{2}{\xi} \phi_0^{-2} \delta_{IJ} & & & & \\ \hline & & 0 & - \phi_0^{-1} & \\ & & - \phi_0^{-1} & 0 & \\ \hline & & & & & -i \phi_0^{-2} \\ & & & & +i \phi_0^{-2} & \\ \end{array} \right) \nonumber\\ &\times& \delta ( \vec{p} - \vec{q} ). \label{W-3D-CRs} \end{eqnarray} Thus, $\varphi_I$ is the physical observable while the set of fields, $\{ \tilde \phi, \beta, \gamma, \bar \gamma \}$ consists of the BRST quartet and is the unphysical mode by the Kugo-Ojima's subsidiary condition \cite{Kugo-Ojima}. Here it is worth mentioning that the Nambu-Goldstone boson $\tilde \phi$ associated with spontaneous symmetry breaking of the Weyl gauge symmetry is an unphysical particle. In this context, let us recall that the Nambu-Goldstone theorem never tells us whether the Nambu-Goldstone boson is physical or unphysical. From our analysis at hand, we can conclude that the Nambu-Goldstone boson $\tilde \phi$ is the unphysical mode, which is absorbed into the longitudinal mode of the Weyl gauge field $s_\mu(x)$, thereby the gauge field becoming massive. Finally, let us focus on the Weyl gauge field $\hat s_\mu$, which satisfies the Lorenz condition (\ref{Linear-Lorenz2}) and the massive Klein-Gordon equation (\ref{Mass-KG}). In a specific Lorentz frame: \begin{eqnarray} p_\mu = ( m, 0, 0, 0), \label{Lor-frame} \end{eqnarray} the Lorenz condition (\ref{Linear-Lorenz2}) produces: \begin{eqnarray} \hat s_0 (p) = 0. \label{Zero-s0} \end{eqnarray} With the Lorentz frame (\ref{Lor-frame}), it turns out that the spacial components of $\hat s_\mu$ are invariant under both GCT and Weyl BRST transformations: \begin{eqnarray} \delta_B \hat s_i (p) = \bar \delta_B \hat s_i (p) = 0. \label{Zero-BRST} \end{eqnarray} Moreover, using the relation (\ref{3D-4D}) and Eq. (\ref{FT-4D-CRs}), the commutation relation between the three-dimensional annihilation and creation operators reads: \begin{eqnarray} [ \hat s_i (\vec{p}), \hat s_j^\dagger (\vec{q}) ] = \delta_{ij} \delta^3 ( \vec{p} - \vec{q} ). \label{3D-Rel} \end{eqnarray} Together with the BRST invariance in Eq. (\ref{Zero-BRST}), this equation clearly shows that the spacial components $\hat s_i (x)$ are really genuine physical massive modes belonging to BRST singlets with positive norm. \section{Choral symmetry} In the previous article \cite{Oda-W}, we have clarified the existence of a huge global symmetry called ``choral symmetry'', which is the $IOSp(10|10)$ symmetry, in Weyl invariant scalar-tensor gravity in Riemann geometry. We will show that the choral symmetry also exists in the theory at hand. The existence of the choral symmetry is expected from the fact that as shown in Section 4, a set of fields (including the space-time coordinates $x^\mu$) $X^M \equiv \{ x^\mu, b_\mu, \sigma, B, c^\mu, \bar c_\mu, c, \bar c \}$ obeys a very simple equation: \begin{eqnarray} g^{\mu\nu} \partial_\mu \partial_\nu X^M = 0. \label{d'Alemb-eq} \end{eqnarray} It is worthwhile to note that this equation holds if and only if we adopt the extended de Donder gauge condition (\ref{Ext-de-Donder}) for the GCT and the scalar gauge condition (\ref{Scalar-gauge}) for the Weyl gauge transformation. Furthermore, Eq. (\ref{d'Alemb-eq}) implies that there should be many conserved currents defined in Eq. (\ref{Cons-currents}) in the theory under consideration. In this section, along the same line of argument as that in the previous article \cite{Oda-W, Oda-V}, we will explicitly prove that there is the choral symmetry $IOSp(10|10)$ in Weyl conformal gravity in Weyl geometry. Let us start with the Lagrangian (\ref{ST-q-Lag}), which can be cast to the form: \begin{eqnarray} {\cal L}_q &=& \sqrt{- g} \biggl[ \frac{1}{2} \xi \phi^2 ( R - 6 \nabla_\mu S^\mu - 6 S_\mu S^\mu ) - \frac{1}{4} H_{\mu\nu} H^{\mu\nu} - \frac{1}{2} \epsilon g^{\mu\nu} ( - 2 \phi \partial_\mu \phi S_\nu \nonumber\\ &+& S_\mu S_\nu \phi^2 ) \biggr] - \frac{1}{2} \tilde g^{\mu\nu} \phi^2 \hat E_{\mu\nu}, \label{Choral-Lag} \end{eqnarray} where we have defined $\hat E_{\mu\nu}$ as \begin{eqnarray} \hat E_{\mu\nu} = \frac{1}{2} \epsilon \partial_\mu \sigma \partial_\nu \sigma + \partial_\mu b_\nu + i \partial_\mu \bar c_\lambda \partial_\nu c^\lambda - \partial_\mu B \partial_\nu \sigma + i \partial_\mu \bar c \partial_\nu c + ( \mu \leftrightarrow \nu ), \label{hat-E} \end{eqnarray} and used the relation (\ref{Dilaton}) between the scalar field $\phi$ and the dilaton $\sigma$. Next, let us focus our attention on the last term in Eq. (\ref{Choral-Lag}) and rewrite it into a more compact form: \begin{eqnarray} {\cal L}_q^{(E)} &\equiv& - \frac{1}{2} \tilde g^{\mu\nu} \phi^2 \hat E_{\mu\nu} = - \frac{1}{2} \tilde g^{\mu\nu} \phi^2 \eta_{NM} \partial_\mu X^M \partial_\nu X^N \nonumber\\ &=& - \frac{1}{2} \tilde g^{\mu\nu} \phi^2 \partial_\mu X^M \tilde \eta_{MN} \partial_\nu X^N. \label{E-Lag} \end{eqnarray} Here we have introduced an $IOSp(10|10)$ metric $\eta_{NM} = \eta_{MN}^T \equiv \tilde \eta_{MN}$ defined as \cite{Kugo} \begin{align} \eta_{NM} = \tilde \eta_{MN} = \begin{array}{c} x^\nu \\ b_\nu \\ \sigma \\ B \\ c^\nu \\ \bar c_\nu \\ c \\ \bar c \end{array} & \left( \begin{array}{cc|cc|cc|cc} & \delta_\mu^\nu & & & & \\ \delta^\mu_\nu & & & & & \\ \hline & & \epsilon & -1 & & & \\ & & -1 & 0 & & & \\ \hline & & & & & -i\delta_\mu^\nu & & \\ & & & & i\delta^\mu_\nu & & & \\ \hline & & & & & & & -i \\ & & & & & & i & \\ \end{array} \right)_. \label{OSp-metric} \\ & \quad \begin{array}{cccccccc} x^\mu & b_\mu & \;\;\sigma & \; B & \;\;\, c^\mu & \;\;\, \bar c_\mu & \;\, c & \;\: \bar c \end{array} \nonumber \end{align} Let us note that this $IOSp(10|10)$ metric $\eta_{NM}$, which is a c-number quantity, has the symmetry property such that \begin{eqnarray} \eta_{MN}=(-)^{|M| \cdot |N|} \eta_{NM} = (-)^{|M|} \eta_{NM}=(-)^{|N|} \eta_{NM}, \label{Prop-OSp-metric} \end{eqnarray} where the statistics index $|M|$ is 0 or 1 when $X^M$ is Grassmann-even or Grassmann-odd, respectively. This property comes from the fact that $\eta_{MN}$ is `diagonal' in the sense that its off-diagonal, Grassmann-even and Grassmann-odd, and vice versa, matrix elements vanish, i.e., $\eta_{MN} = 0$ when $|M| \neq |N|$, thereby being $|M| = |N| = |M| \cdot| N|$ in front of $\eta_{MN}$ \cite{Kugo}. Now that (\ref{E-Lag}) is expressed in a manifestly $IOSp(10|10)$ invariant form except for the Weyl invariant metric $\tilde g^{\mu\nu} \phi^2$, which will be discussed later, there could exist an $IOSp(10|10)$ as a global symmetry in our theory. Note that the infinitesimal $OSp$ rotation is defined by \begin{eqnarray} \delta X^M = \eta^{ML} \varepsilon_{LN} X^N \equiv \varepsilon^M{}_N X^N, \label{OSp-rot} \end{eqnarray} where $\eta^{MN}$ is the inverse matrix of $\eta_{MN}$, and the infinitesimal parameter $\varepsilon_{MN}$ has the following properties: \begin{eqnarray} \varepsilon_{MN} = (-)^{1 + |M| \cdot |N|} \varepsilon_{NM}, \qquad \varepsilon_{MN} X^L = (-)^{|L| (|M| + |N|)} X^L \varepsilon_{MN}. \label{varepsilon} \end{eqnarray} In order to find the conserved current, we assume that the infinitesimal parameter $\varepsilon_{MN}$ depends on the space-time coordinates $x^\mu$, i.e., $\varepsilon_{MN} = \varepsilon_{MN} (x^\mu)$. Assuming for a while that the metric $\tilde g^{\mu\nu} \phi^2$ is invariant under the $OSp$ rotation (\ref{OSp-rot}), we find that (\ref{E-Lag}) is transformed as \begin{eqnarray} \delta {\cal L}_q^{(E)} = - \tilde g^{\mu\nu} \phi^2 \left( \partial_\mu \varepsilon_{NM} X^M \partial_\nu X^N + \varepsilon_{NM} \partial_\mu X^M \partial_\nu X^N \right). \label{Var-E-Lag} \end{eqnarray} It is easy to prove that the second term on the RHS vanishes owing to the first property in Eq. (\ref{varepsilon}). Thus, ${\cal L}_q^{(E)}$ is invariant under the infinitesimal $OSp$ rotation. The conserved current is then calculated to be: \begin{eqnarray} \delta {\cal L}_q^{(E)} &=& - \tilde g^{\mu\nu} \phi^2 \partial_\mu \varepsilon_{NM} X^M \partial_\nu X^N \nonumber\\ &=& - \frac{1}{2} \tilde g^{\mu\nu} \phi^2 \partial_\mu \varepsilon_{NM} \left[ X^M \partial_\nu X^N - (-)^{|M| \cdot |N|} X^N \partial_\nu X^M \right] \nonumber\\ &=& - \frac{1}{2} \tilde g^{\mu\nu} \phi^2 \partial_\mu \varepsilon_{NM} \left( X^M \partial_\nu X^N - \partial_\nu X^M X^N \right) \nonumber\\ &=& - \frac{1}{2} \tilde g^{\mu\nu} \phi^2 \partial_\mu \varepsilon_{NM} X^M \overset{\leftrightarrow}{\partial}_\nu X^N \nonumber\\ &\equiv& - \frac{1}{2} \partial_\mu \varepsilon_{NM} {\cal M}^{\mu MN}, \label{OSp-current} \end{eqnarray} with the conserved current ${\cal M}^{\mu MN}$ for the $OSp$ rotation taking the form: \begin{eqnarray} {\cal M}^{\mu MN} = \tilde g^{\mu\nu} \phi^2 X^M \overset{\leftrightarrow}{\partial}_\nu X^N. \label{OSp-current-M} \end{eqnarray} The above proof makes sense only under the assumption that the metric $\tilde g^{\mu\nu} \phi^2$ and the other terms except for the last term in (\ref{Choral-Lag}) are invariant under the $OSp$ rotation, but it is obviously not the case. However, this problem is cured by noticing that the $OSp$ rotation includes a Weyl transformation on the dilaton: \begin{eqnarray} \delta \sigma = \eta^{\sigma L} \varepsilon_{LN} X^N = - \varepsilon_{BN} X^N \equiv - \varepsilon(x), \label{Dilaton-OSp} \end{eqnarray} where we have used (\ref{OSp-metric}) and \begin{eqnarray} \begin{pmatrix} \epsilon & -1 \\ -1 & 0 \end{pmatrix}^{-1} = \begin{pmatrix} 0 & -1 \\ -1 & -\epsilon \end{pmatrix}, \label{Matrix} \end{eqnarray} where recall that the matrix $\eta^{ML}$ is the inverse matrix of $\eta_{ML}$. As for the scalar field $\phi(x)$, this transformation for the dilaton can be interpreted as a Weyl transformation: \begin{eqnarray} \phi \rightarrow \phi^\prime = e^{- \varepsilon (x)} \phi. \label{Weyl-phi} \end{eqnarray} Thus, simultaneously with the $OSp$ rotation, if we perform a Weyl transformation given by \begin{eqnarray} \delta g_{\mu\nu} = 2 \varepsilon (x) g_{\mu\nu}, \qquad \delta S_\mu = - \partial_\mu \varepsilon (x), \label{Weyl-g&S} \end{eqnarray} and a local shift for the Nakanishi-Lautrup field $B$:\footnote{Under the $OSp$ rotation, the $B$ field is transformed as $\delta B = \eta^{B L} \varepsilon_{LN} X^N = - \varepsilon_{\sigma N} X^N + \epsilon \varepsilon$. The transformation (\ref{B-shift}) is carried out independently of this $OSp$ rotation.} \begin{eqnarray} \delta B = \epsilon \, \varepsilon (x), \label{B-shift} \end{eqnarray} it turns out that under the (local) $OSp$ rotation (\ref{OSp-rot}), the quantum Lagrangian ${\cal{L}}_q$ is transformed as \begin{eqnarray} \delta {\cal L}_q = - \frac{1}{2} \partial_\mu \varepsilon_{NM} {\cal M}^{\mu MN}. \label{Var-q-Lag} \end{eqnarray} As a result, the conserved current ${\cal M}^{\mu MN}$ for the $OSp$ rotation takes the form (\ref{OSp-current-M}). In a similar way, we can derive the conserved current for the infinitesimal translation: \begin{eqnarray} \delta X^M = \varepsilon^M, \label{transl} \end{eqnarray} and it turns out that the conserved current ${\cal P}^{\mu M}$ for the translation reads: \begin{eqnarray} {\cal P}^{\mu M} = \tilde g^{\mu\nu} \phi^2 \partial_\nu X^M = \tilde g^{\mu\nu} \phi^2 \left( 1 \overset{\leftrightarrow}{\partial}_\nu X^M \right). \label{transl-current-P} \end{eqnarray} From the conserved currents (\ref{OSp-current-M}) and (\ref{transl-current-P}), the corresponding conserved charges are given by \begin{eqnarray} M^{MN} &\equiv& \int d^3 x \, {\cal M}^{0 MN} = \int d^3 x \, \tilde g^{0 \nu} \phi^2 X^M \overset{\leftrightarrow}{\partial}_\nu X^N, \nonumber\\ P^M &\equiv& \int d^3 x \, {\cal P}^{0 M} = \int d^3 x \, \tilde g^{0 \nu} \phi^2 \partial_\nu X^M. \label{IOSp-charge} \end{eqnarray} For instance, the BRST charges for the GCT and Weyl transformation are respectively expressed as \begin{eqnarray} &{}& Q_B \equiv M (b_\rho, c^\rho) = \int d^3 x \, \tilde g^{0 \nu} \phi^2 b_\rho \overset{\leftrightarrow}{\partial}_\nu c^\rho, \nonumber\\ &{}& \bar Q_B \equiv M (B, c) = \int d^3 x \, \tilde g^{0 \nu} \phi^2 B \overset{\leftrightarrow}{\partial}_\nu c. \label{Choral-Symm} \end{eqnarray} We can then verify that using various ETCRs obtained so far, the $IOSp(10|10)$ generators $\{ M^{MN}, P^M \}$ generate an $IOSp(10|10)$ algebra: \begin{eqnarray} &{}& [ P^M, P^N \} = 0, \nonumber\\ &{}& [ M^{MN}, P^R \} = i \bigl[ P^M \tilde \eta^{NR} - (-)^{|N| |R|} P^N \tilde \eta^{MR} \bigr], \nonumber\\ &{}& [ M^{MN}, M^{RS} \} = i \bigl[ M^{MS} \tilde \eta^{NR} - (-)^{|N| |R|} M^{MR} \tilde \eta^{NS} - (-)^{|N| |R|} M^{NS} \tilde \eta^{MR} \nonumber\\ &{}& + (-)^{|M| |R| + |N| |S|} M^{NR} \tilde \eta^{MS} \bigr]. \label{IOSp-algebra} \end{eqnarray} Finally, it is useful to compare our extended choral symmetry $IOSp(10|10)$ with the original choral symmetry $IOSp(8|8)$ in Einstein's general relativity \cite{N-O-text}. In our case, the choral symmetry is extended in the sense that the GCT is replaced with a larger symmetry, which consists of both the GCT and the Weyl gauge transformation. Accordingly the dilaton $\sigma$, the Nakanishi-Lautrup field $B$, ghost $c$ and anti-ghost $\bar c$ are joined in the algebra. The choral symmetry $IOSp(10|10)$ therefore includes the dilaton, or equivalently, the scalar field, which exists in the classical Lagrangian and is closely related to a classical theory. In contrast, the original $IOSp(8|8)$ symmetry is purely a symmetry among quantum fields, which are the NL field and ghosts, so the symmetry is limited to the sector related to the gauge-fixing procedure. From this viewpoint, we expect that the extended $IOSp(10|10)$ choral symmetry might play an important role in clarifying the dynamics peculiar to the classical theory. \section{Gravitational conformal symmetry and spontaneous symmetry breakdown} One of the most interesting features in the formalism at hand is that as an analog of the well-known conformal symmetry in a flat Minkowski space-time, there is a gravitational conformal symmetry which is a subgroup of the choral symmetry, and its spontaneous symmetry breakdown down to the Poincar\'e symmetry guarantees that the graviton and the dilaton are exactly massless Nambu-Goldstone particles \cite{Oda-W}. This feature is so important for future developments of quantum gravity that we would like to explain the gravitational conformal symmetry and its spontaneous symmetry breakdown in detail. In particular, as already shown in Section 6, there is a $\it{massive}$ Weyl gauge field in the spectrum, so at first sight it appears to be strange that there is a conformal symmetry in the present theory since it is usually thought that conformal or scale symmetry exists in the theories with only massless particles. With regard to this, it is worthwhile to recall that the massless Weyl gauge field acquires the mass via spontaneous symmetry breakdown (SSB) of Weyl gauge symmetry and the SSB is the breakdown of symmetry at the level of not field operators but the representation of field operators in the sense that the symmetry cannot be realized by a unitary transformation in the state vector space. Thus, it is not strange that there is a conformal symmetry in the present theory with the massive gauge field if the mass is generated through the SSB. Moreover, this physical situation is also supported by the Zumino theorem \cite{Zumino} to some degree since the theorem insists that theories invariant under general coordinate transformation and Weyl transformation at the same time should possess conformal symmetry in a flat Minkowski background at least classically. As clarified in the previous paper \cite{Oda-W}, the extended de Donder gauge condition (\ref{Ext-de-Donder}) and the scalar gauge condition (\ref{Scalar-gauge}) have a residual symmetry which corresponds to the dilatation and the special conformal transformation in a flat Minkowski space-time. Indeed, the quantum Lagrangian (\ref{ST-q-Lag}) is still invariant under the restricted Weyl transformation \cite{Oda-R}: \begin{eqnarray} \delta g_{\mu\nu} &=& 2 \Lambda g_{\mu\nu}, \qquad \delta \phi = - \Lambda \phi, \nonumber\\ \delta S_\mu &=& - \partial_\mu \Lambda, \qquad \delta b_\mu = - \partial_\mu \Lambda B, \label{R-Weyl} \end{eqnarray} where the infinitesimal transformation parameter $\Lambda$ takes the form: \begin{eqnarray} \Lambda = \lambda - 2 k_\mu x^\mu, \label{R-Weyl-Lam} \end{eqnarray} with $\lambda$ and $k_\mu$ being infinitesimal constants corresponding to a global scale transformation and the special conformal transformation, respectively \cite{Oda-W}. Note that $\Lambda$ obeys the equation $g^{\mu\nu} \partial_\mu \partial_\nu \Lambda = 0$, which is a characteristic feature of the restricted Weyl transformation. The whole global symmetry in the theory under consideration should be included in the extended $IOSp(10|10)$ choral symmetry. Actually, we can construct the generators corresponding to the transformation parameters $\lambda$ and $k_\mu$ out of those of the choral symmetry as \begin{eqnarray} D_0 &\equiv& - P(B) = - \int d^3 x \, \tilde g^{0 \nu} \phi^2 \partial_\nu B, \nonumber\\ K^\mu &\equiv& 2 M^\mu (x, B) = 2 \int d^3 x \, \tilde g^{0 \nu} \phi^2 x^\mu \overset{\leftrightarrow}{\partial}_\nu B. \label{Res-gen} \end{eqnarray} It is easy to verify that these generators generate the symmetry (\ref{R-Weyl}) in terms of the ETCRs in (\ref{ETCRs}). Our theory is also invariant under the translation and the general linear transformation $GL(4)$. Actually, we can make the translation generator $P_\mu$ and $GL(4)$ generator $G^\mu{}_\nu$ from the choral symmetry as \begin{eqnarray} P_\mu &\equiv& P_\mu (b) = \int d^3 x \, \tilde g^{0 \nu} \phi^2 \partial_\nu b_\mu, \nonumber\\ G^\mu{}_\nu &\equiv& M^\mu{}_\nu (x, b) - i M^\mu{}_\nu (c^\tau, \bar c_\tau) \nonumber\\ &=& \int d^3 x \, \tilde g^{0 \lambda} \phi^2 ( x^\mu \overset{\leftrightarrow}{\partial}_\lambda b_\nu - i c^\mu \overset{\leftrightarrow}{\partial}_\lambda \bar c_\nu ). \label{Trans-GL} \end{eqnarray} For instance, based on the ETCRs in (\ref{ETCRs}), we can check that the $GL(4)$ generator $G^\mu{}_\nu$ correctly generates the $GL(4)$ transformation on the fields $\phi, S_\rho$ and $g_{\sigma\tau}$: \begin{eqnarray} &{}& [ i G^\mu{}_\nu, \phi ] = x^\mu \partial_\nu \phi, \qquad [ i G^\mu{}_\nu, S_\rho ] = x^\mu \partial_\nu S_\rho + \delta_\rho^\mu S_\nu, \nonumber\\ &{}& [ i G^\mu{}_\nu, g_{\sigma\tau} ] = x^\mu \partial_\nu g_{\sigma\tau} + \delta_\sigma^\mu g_{\nu\tau} + \delta_\tau^\mu g_{\nu\sigma}. \label{Ex-GL} \end{eqnarray} Finally, we can build a generator corresponding to the dilatation in a flat Minkowski space-time, which is closely related to the generator $D_0$ of the scale transformation in (\ref{Res-gen}). With this in mind, let us consider a set of generators, $\{ P_\mu, G^\mu{}_\nu, K^\mu, D_0 \}$. From these generators we wish to construct the generator $D$ for the dilatation. Let us recall that in conformal field theory in the four-dimensional Minkowski space-time, the dilatation generator obeys the following algebra for an local operator $O_i (x)$ of conformal dimension $\Delta_i$ \cite{Gross, Nakayama}: \begin{eqnarray} [ i D, O_i (x) ] = x^\mu \partial_\mu O_i (x) + \Delta_i O_i (x). \label{D-com} \end{eqnarray} The scalar field $\phi (x)$, for example, has conformal dimension $1$ and therefore satisfies the equation: \begin{eqnarray} [ i D, \phi (x) ] = x^\mu \partial_\mu \phi (x) + \phi (x). \label{D-phi-com} \end{eqnarray} With this knowledge, let us construct a generator for the dilatation in such a way that the transformation law on the scalar field satisfies this equation (\ref{D-phi-com}). From Eq. (\ref{Ex-GL}) and the definition of $D_0$ in (\ref{Res-gen}), we find that \begin{eqnarray} [ i G^\mu{}_\mu, \phi (x) ] = x^\mu \partial_\mu \phi (x), \qquad [ i D_0, \phi (x) ] = - \phi (x). \label{GD-phi-com} \end{eqnarray} It therefore turns out that the following linear combination of $G^\mu{}_\mu$ and $D_0$ does the job: \begin{eqnarray} D \equiv G^\mu{}_\mu - D_0. \label{D-def} \end{eqnarray} As a consistency check, it is useful to see how this operator $D$ acts on the metric field. The resulting expression is: \begin{eqnarray} &{}& [ i D, g_{\sigma\tau} ] = [ i G^\mu{}_\mu, g_{\sigma\tau} ] - [ i D_0, g_{\sigma\tau} ] \nonumber\\ &{}& = ( x^\mu \partial_\mu g_{\sigma\tau} + 2 g_{\sigma\tau} ) - 2 g_{\sigma\tau} = x^\mu \partial_\mu g_{\sigma\tau}, \label{D-g-com} \end{eqnarray} which implies that the metric field has conformal dimension $0$ as defined in conformal field theory. Further calculations reveal that the algebra among the generators $\{ P_\mu, G^\mu{}_\nu, K^\mu, D \}$ closes and takes the form: \begin{eqnarray} &{}& [ P_\mu, P_\nu ] = 0, \quad [ P_\mu, G^\rho{}_\sigma ] = i P_\sigma \delta^\rho_\mu, \quad [ P_\mu, K^\nu ] = - 2 i ( G^\rho{}_\rho - D ) \delta^\nu_\mu, \quad \nonumber\\ &{}& [ P_\mu, D ] = i P_\mu, \quad [ G^\mu{}_\nu, G^\rho{}_\sigma ] = i ( G^\mu{}_\sigma \delta^\rho_\nu - G^\rho{}_\nu \delta^\mu_\sigma), \nonumber\\ &{}& [ G^\mu{}_\nu, K^\rho ] = i K^\mu \delta^\rho_\nu, \quad [ G^\mu{}_\nu, D ] = [ K^\mu, K^\nu ] = 0, \nonumber\\ &{}& [ K^\mu, D ] = - i K^\mu, \quad [ D, D] = 0. \label{Grav-conf0} \end{eqnarray} To extract the gravitational conformal algebra in quantum gravity, it is necessary to introduce the ``Lorentz'' generator. It can be contructed from the $GL(4)$ generator and the flat Minkowski metric to be: \begin{eqnarray} M_{\mu\nu} \equiv - \eta_{\mu\rho} G^\rho{}_\nu + \eta_{\nu\rho} G^\rho{}_\mu. \label{Lor-gene} \end{eqnarray} In terms of the generator $M_{\mu\nu}$, the algebra (\ref{Grav-conf0}) can be cast to the form: \begin{eqnarray} &{}& [ P_\mu, P_\nu ] = 0, \quad [ P_\mu, M_{\rho\sigma} ] = i ( P_\rho \eta_{\mu\sigma} - P_\sigma \eta_{\mu\rho} ), \nonumber\\ &{}& [ P_\mu, K^\nu ] = - 2 i ( G^\rho{}_\rho - D ) \delta^\nu_\mu, \quad [ P_\mu, D ] = i P_\mu, \nonumber\\ &{}& [ M_{\mu\nu}, M_{\rho\sigma} ] = - i ( M_{\mu\sigma} \eta_{\nu\rho} - M_{\nu\sigma} \eta_{\mu\rho} + M_{\rho\mu} \eta_{\sigma\nu} - M_{\rho\nu} \eta_{\sigma\mu}), \nonumber\\ &{}& [ M_{\mu\nu}, K^\rho ] = i ( - K_\mu \delta^\rho_\nu + K_\nu \delta^\rho_\mu ), \quad [ M_{\mu\nu}, D ] = [ K^\mu, K^\nu ] = 0, \nonumber\\ &{}& [ K^\mu, D ] = - i K^\mu, \quad [ D, D] = 0. \label{Grav-conf} \end{eqnarray} where we have defined $K_\mu \equiv \eta_{\mu\nu} K^\nu$. It is of interest that the the algebra (\ref{Grav-conf}) in quantum gravity, which we call ``gravitational conformal algebra'', formally resembles conformal algebra in the flat Minkowski space-time except for the expression of $[ P_\mu, K^\nu ]$ \cite{Gross, Nakayama}.\footnote{In case of conformal algebra in the flat space-time, the expression is given by $[ P_\mu, K^\nu ] = - 2 i ( \delta^\nu_\mu D + M_\mu{}^\nu )$.} This difference stems from the difference of the definition of conformal dimension (or weight) in both gravity and conformal field theory, for which the metric tensor field $g_{\mu\nu}$ has weight $2$ in gravity while it has weight $0$ in conformal field theory. Now, on the basis of the gravitational conformal symmetry, we are able to show that $GL(4)$, special conformal symmetry and dilatation are spontaneously broken down to the Poincar\'e symmetry. To this end, we postulate the existence of a unique vacuum $| 0 \rangle$, which is normalized to be the unity: \begin{eqnarray} \langle 0 | 0 \rangle = 1. \label{Vac-norm} \end{eqnarray} Furthermore, we assume that the vacuum is translation invariant: \begin{eqnarray} P_\mu | 0 \rangle = 0, \label{Trans-Vac} \end{eqnarray} and the vacuum expectation values (VEVs) of the metric tensor $g_{\mu\nu}$ and the scalar field $\phi$ are respectively the Minkowski metric $\eta_{\mu\nu}$ and a non-zero constant $\phi_0 \neq 0$: \begin{eqnarray} \langle 0 | g_{\mu\nu} | 0 \rangle = \eta_{\mu\nu}, \qquad \langle 0 | \phi | 0 \rangle = \phi_0. \label{VEV-A} \end{eqnarray} From (\ref{Ex-GL}), we find that the VEV of an equal-time commutator between the $GL(4)$ generator and the metric field reads: \begin{eqnarray} \langle 0 | [ i G^\mu{}_\nu, g_{\sigma\tau} ] | 0 \rangle = \delta^\mu_\sigma \eta_{\nu\tau} + \delta^\mu_\tau \eta_{\nu\sigma}. \label{G-g-CM} \end{eqnarray} Thus, the Lorentz generator defined in Eq. (\ref{Lor-gene}) has a vanishing VEV: \begin{eqnarray} \langle 0 | [ i M_{\mu\nu}, g_{\sigma\tau} ] | 0 \rangle = 0. \label{M-g-CM} \end{eqnarray} On the other hand, the symmetric part defined as $\bar M_{\mu\nu} \equiv \eta_{\mu\rho} G^\rho{}_\nu + \eta_{\nu\rho} G^\rho{}_\mu$ has the non-vanishing VEV: \begin{eqnarray} \langle 0 | [ i \bar M_{\mu\nu}, g_{\sigma\tau} ] | 0 \rangle = 2 ( \eta_{\mu\sigma} \eta_{\nu\tau} + \eta_{\mu\tau} \eta_{\nu\sigma} ). \label{BM-g-CM} \end{eqnarray} Thus, the $GL(4)$ symmetry is spontaneously broken to the Lorentz symmetry where the corresponding Nambu-Goldstone boson with ten independent components is nothing but the massless graviton \cite{NO}. Here, it is interesting that in a sector of the scalar field, the $GL(4)$ symmetry and of course the Lorentz symmetry as well do not give rise to a symmetry breaking. This can be seen in the following commutators: \begin{eqnarray} \langle 0 | [ i G^\mu{}_\nu, \phi ] | 0 \rangle = \langle 0 | [ i M_{\mu\nu}, \phi ] | 0 \rangle = \langle 0 | [ i \bar M_{\mu\nu}, \phi ] | 0 \rangle = 0. \label{G-phi-CM} \end{eqnarray} Now we wish to clarify how the dilatation and special conformal symmetry are spontaneously broken and what the corresponding Nambu-Goldstone bosons are. As for the dilatation, we find that \begin{eqnarray} \langle 0 | [ i D, \sigma ] | 0 \rangle = 1, \label{VEV-D-sigma} \end{eqnarray} which elucidates the spontaneous symmetry breakdown of the dilatation whose Nambu-Goldstone boson is just the massless dilaton $\sigma(x)$. Regarding the special conformal symmetry, we find: \begin{eqnarray} \langle 0 | [ i K^\mu, \partial_\nu \sigma ] | 0 \rangle = 2 \delta^\mu_\nu. \label{VEV-K-phi} \end{eqnarray} This equation means that the special conformal symmetry is certainly broken spontaneously and its Nambu-Goldstone boson is the derivative of the dilaton. This interpretation can be also verified from the gravitational conformal algebra as follows: In the algebra (\ref{Grav-conf}) we have a commutator between $P_\mu$ and $K^\nu$: \begin{eqnarray} [ P_\mu, K^\nu ] = - 2 i ( G^\rho{}_\rho - D ) \delta^\nu_\mu. \label{P-K} \end{eqnarray} Let us consider the Jacobi identity: \begin{eqnarray} [ [ P_\mu, K^\nu ], \sigma ] + [ [ K^\nu, \sigma ], P_\mu ] + [ [ \sigma, P_\mu ], K^\nu ] = 0. \label{Jacobi} \end{eqnarray} Using the translational invariance of the vacuum in Eq. (\ref{Trans-Vac}) and the equation: \begin{eqnarray} [ P_\mu, \sigma ] = - i \partial_\mu \sigma, \label{Jacobi2} \end{eqnarray} and taking the VEV of the Jacobi identity (\ref{Jacobi}), we can obtain the VEV: \begin{eqnarray} \langle 0 | [ K^\nu, \partial_\mu \sigma ] | 0 \rangle &=& - 2 \delta^\nu_\mu \langle 0 | [ G^\rho{}_\rho - D, \sigma ] | 0 \rangle \nonumber\\ &=& - 2 i \delta^\nu_\mu, \label{VEV-Jacobi} \end{eqnarray} which coincides with Eq. (\ref{VEV-K-phi}) as promised. In summary, the $GL(4)$ symmetry is spontaneously broken to the Poincar\'e symmetry whose Nambu-Goldstone boson is the graviton. The dilatation symmetry and the special conformal symmetry are also spontaneously broken and the corresponding Nambu-Goldstone bosons are the dilaton and the derivative of the dilaton, respectively. Interest here is that the Nambu-Goldstone boson associated with the special conformal symmetry is not an independent field in quantum gravity as in conformal field theory \cite{Kobayashi}. \section{Conclusion} In this article, we have presented a BRST formalism of a Weyl conformal gravity in Weyl geometry. The essential ingredient in our formalism is choosing suitable gauge conditions for the general coordinate invariance and the Weyl invariance. To implement two independent BRST transformations $\delta_B, \bar \delta_B$ corresponding to the GCT and the Weyl transformation, respectively, i.e., $\{ \delta_B, \bar \delta_B \} = 0$, one has to select the gauge conditions in such a way that the gauge condition for the GCT must be invariant under the Weyl transformation and that for the Weyl transformation must be so under the GCT \cite{Oda-W}. In addition, both gauge conditions must give us a gauge invariant measure in place of the conventional measure $\sqrt{-g}$ and ensure the masslessness of the dilaton. Interestingly enough, such the gauge conditions are almost uniquely determined by the extended de Donder gauge condition (\ref{Ext-de-Donder}) for the GCT and the scalar gauge condition (\ref{Scalar-gauge}) for the Weyl transformation. With the other gauge conditions, we cannot construct the conserved currents for the extended choral symmetry, and without the choral symmetry we cannot ensure the gravitational conformal algebra such that we cannot prove the masslessness of the graviton and the dilaton. It is usually said that the gauge conditions do not change the physical content of a theory, but it is true that the existence of global symmetries seems to critically depend on the gauge choice as seen in the present study of Weyl conformal gravity. As for the future works, we would like to present a BRST formalism of quadratic conformal gravity (\ref{L-QG}) since this theory is the unique theory which is invariant under Weyl gauge transformation without matter fields.\footnote{Recently, spontaneous symmetry breakdown of conformal symmetry in quantum quadratic gravity in Riemann geometry has been investigated in \cite{Kubo}.} However, it is known that higher-derivative gravities such as quadratic gravity generally suffer from the existence of a massless or massive ghost which prevents a lower bound of energy at the classical level and violates the unitarity at the quantum level. Thus, we have to provide a recipe for nullifying such a ghost. Since our choral symmetry is a huge global symmetry including the gravitational conformal symmetry, it might give us a useful tool for attacking various important problems such as the ghost and renormalizability. The work is currently in progress with partial affirmative results. \section*{Acknowledgment} This work is supported in part by the JSPS Kakenhi Grant No. 21K03539 and P. S. is supported by the IMPRS-PTFS.

e30445ce354dcc6debe2ce6066992f9aaa2e38f6

\section{Introduction} We have developed biped robots with a passive dynamic walking mechanism. Bipedal walking is generated through the dynamic interactions among the body, neural system, and environment. McGeer~\cite{McGeer1990} developed a simple robot with passive legs attached to the hip and demonstrated stable walking while descending a slope without any energy input other than gravity. This suggested that passive locomotion plays a significant role in locomotion. Many studies have clarified the gait generation mechanism by using a simple compass model consisting of a body mass and two rigid bars as legs~\cite{Garcia1998,Goswami1998,Garcia2000,Collins2005,Obayashi2015,Okamoto2020}. In addition to completely passive models, models with controls derived from physiological studies to imitate the neural system have also been proposed~\cite{Rybak2006,Cappellini2006,Ivanenko2007}. For example, Aoi et al. \cite{Aoi2006,Aoi2007} applied control based on a central pattern generator (CPG), which involves a rhythm resetting (i.e., phase resetting) function of human neural systems, to the compass model, and revealed the mechanism through which stable limit cycles are generated. The dynamics of both the lower limbs and the upper body play an important role in walking. Therefore, many researchers have investigated the effect of the upper body on locomotion. For example, in modeling studies, a rigid upper body has been attached to simple models such as a compass model, and its effect on gait has been investigated~\cite{Wisse2004,Wisse2007,Collins2009,Ackerman2013,Ackerman2014,Honjo2013,Honjo2019}. Notably, the upper body is not a single rigid body but consists of multiple soft tissues. Therefore, many researchers have focused on the active or passive wobbling mass of the upper body, the flexible spine, and the pendulum-like oscillation of the arms~\cite{Rome2006,Arellano2014,Hanazawa2015, Hanazawa2019, Toda2020}. These studies showed that such elements compensated for the torque generated by legs, reduced the peak ground reaction force, increased the energy efficiency, and increased or decreased the stability. In our previous study, we showed that a passive wobbling mass connected to the upper body and oscillating in the vertical direction generates human-like time profiles of the ground reaction force while running using a simple model and a running biped robot~\cite{Kamimura2021}. While the vertical wobbling mass plays a significant role in running, as shown in our previous study, a horizontal wobbling mass is also expected to play an important role in walking. For example, the horizontal wobbling mass is assumed to affect acceleration and deceleration in the horizontal direction and thus impact the energy efficiency of gait. Moreover, the horizontal movement of the upper body is supposed to affect the stability. During walking, the zero moment point (ZMP) always exists inside the support polygon~\cite{Popovic2005}. If the ZMP moves outside the support polygon, the walker will fall down. The horizontal movement of the body is assumed to play a significant role in the dynamics of the ZMP. Therefore, in this study, we investigate the effect of the horizontal dynamics of the upper body on the walking performance in terms of the energy efficiency and stability. To clarify this effect, based on a compass model, we propose a simple walking model with a horizontally oscillating mass point (i.e., a wobbling mass). Furthermore, we discuss the relationship between our simple model and actual walking. \section{Model} \begin{figure} \centering \includegraphics[scale=1.0]{fig/model_hori.pdf} \caption{Compass model with elastically supported horizontal wobbling mass.} \label{fig:model} \end{figure} We proposed a model based on a compass model equipped with a wobbling mass, as shown in Fig.~\ref{fig:model}. The model consists of a hip, swing leg, stance leg, and wobbling mass. The legs have length $L$ and are connected at the hip. Each foot can detect the touchdown timing on the ground. The tip of the stance leg is constrained on the ground and it behaves as a frictionless pin joint. The body mass $m_1$ and leg mass $m_2$ are assumed to be concentrated at the hip and tip, respectively. The wobbling mass $m_3$ is connected to the hip through a prismatic spring $K$. It can move only in the horizontal direction, and its height always equals the hip height. The whole upper body mass is $M = m_1 + m_3$, where $m_3 = \alpha M$ and $m_1 = (1-\alpha)M$. The model is constrained on the $x$-$y$ plane, with the walking direction along the $x$-axis. The model has three degrees of freedom, $\theta_1$, $\theta_2$, and $X$, where $\theta_1$ is the angle of the stance leg with respect to the vertical direction, $\theta_2$ is the angle between the swing leg and the stance leg, and $X$ is the distance between the hip and the wobbling mass. In this model, $\theta_1$ and $X$ are not controlled directly, whereas $\theta_2$ is controlled by the actuator torque $U$. We derived the dimensionless governing equations using the characteristic length $L$ and characteristic time scale $\tau = \sqrt{L/g}$. The state variables of the model are defined as $q = [\theta_1,\theta_2,y]^\top$, where $x = X/L$. The dimensionless equations of motion for the swing phase are given by \begin{align} \mathcal{M}(q)\ddot{q} + h(q,\dot{q}) + v(q) = Q, \end{align} where $\mathcal{M}(q)$ is the inertia matrix, $h(q,\dot{q})$ is the Coriolis and centrifugal forces, $v(q)$ is the conservative force, and $Q = [0,u = U/MgL,0]^\top$ is the input torque. $\dot{*}$ and $\ddot{*}$ respectively indicate the first and second derivatives of variable $*$ with respect to $t/\tau$. The swing leg touches the ground when $2\theta_1 = \theta_2$ is satisfied, at which time the swing and stance legs are immediately switched. The impulsive force occurs at the tip and results in a discontinuous change in the velocities From the law of conservation of angular momentum, the relationship between the states immediately before and after touch down is given by \begin{align} [(q^+)^\top, (\dot{q}^+)^\top]^\top = H(q^-,\dot{q}^-), \end{align} where $*^-$ and $*^+$ indicate the states immediately before and after touchdown, respectively. The input torque $u$ is determined based on a rhythmic signal oscillator (CPG)~\cite{Aoi2006} \begin{align} u = -K_{\rm p}(\theta_2 - \theta_2^{\rm d}(\gamma,\phi)) - K_{\rm d}(\dot{\theta}_2 - \dot{\theta}_2^{\rm d}(\gamma,\phi)), \end{align} where the desired angle $\theta_2^{\rm d}(\gamma,\phi))$ is defined as \begin{align} \theta_2^{\rm d}(\gamma,\phi) = \gamma(1+ \cos \phi) - S, \end{align} where $\gamma$ and $\phi$ are respectively the amplitude and phase of the oscillator, which has constant angular velocity $\omega$. Further, $S$ is the stride angle. \subsection{Searching limit cycles} We defined the Poincar\'{e} section at the touchdown moment. A Poincar\'{e} map $P$ was defined as \begin{align} z_{n+1} = P(z_n), \end{align} where $z_n$ is the state variable at the $n$th intersection with the Poincar\'{e} section. For a periodic gait, $z^* = P(z^*)$ is satisfied, where $z^*$ is a fixed point on the Poincar\'{e} section. We numerically searched for fixed points for periodic walking by using the \texttt{fsolve} function in MATLAB. \subsection{Stability and risk of accidental falling} We used the eigenvalues of the linearized Poincar\'{e} map around the fixed point on the Poincar\'{e} section. The limit cycle is stable when all of the eigenvalues are inside the unit cycle in the complex plane (these magnitudes are less than 1). The ZMP is also an important measure to determine whether walking is stable or not. Because the moment is zero around the ZMP, the horizontal position $p$ of the ZMP is given by \begin{align} p = x_g - \frac{\ddot{x}_g}{g+\ddot{y}_g}y_g, \end{align} where $(x_g, y_g)$ are respectively the horizontal and vertical positions of the center of mass of the whole body, including the wobbling mass. Although our model does not have a support polygon because it has point tips, we used the distance $d$ between the ground point and the ZMP as a criterion of the risk of accidental falling while walking. \subsection{Energy efficiency} To evaluate the energy efficiency, we defined the cost of transport (CoT) as \begin{align} {\rm CoT} = \frac{W}{mg\bar{v}} \end{align} where $W$ is the work generated by the hip actuator during one stride, and $\bar{v}$ is the averaged horizontal velocity of the center of mass for one gait cycle. A smaller CoT indicates better energy efficiency because the actuator expends lesser energy in one stride. \section{Results} \subsection{Obtained solution groups} The periodic solutions obtained for various $k=KL/Mg$ and $\alpha$ values using the CPG frequency $\omega = 3$ are shown in Fig.~\ref{fig:soltuons_distribution}. The obtained solutions are divided into several qualitatively different discrete groups; these are denoted as solution groups A, B, C, D, and E according to the value ranges of $k$, as shown in Fig.~\ref{fig:soltuons_distribution}. Regardless of the $\omega$ value, the obtained solutions were classified into similar groups. \begin{figure} \centering \includegraphics[width=80mm]{fig/solutions_for_various_param.pdf} \caption{Distribution of limit cycles on $k$-$\alpha$ plane with $\omega = 3$. Blue and grey areas show stable and unstable limit cycles, respectively. No solution was found in the white area. Solutions exhibit qualitatively different behaviors and are accordingly classified into groups A, B, C, D, and E. Points a, b, and c indicate typical stable solutions, whose behaviors are shown in Fig.~\ref{fig:variables}.} \label{fig:soltuons_distribution} \end{figure} Figure~\ref{fig:variables} shows the behaviors of typical solutions of groups A, B, and C. The solutions exhibited qualitatively different behaviors, particularly in $x$ and $\dot{\theta}_1$. In the solutions of group A (Fig.~\ref{fig:variables}a), the position of the wobbling mass $x$ moves backward ($x < 0$) in the first half of the stance phase and forward ($x > 0$) in the second half. By contrast, in the solutions of group B (Fig.~\ref{fig:variables}b), $x$ moves forward in the first half of the stance phase and backward in the second half. In the solutions of group C (Fig.~\ref{fig:variables}c), $x$ oscillates two times during one gait cycle; this is regarded as a doubling of the oscillation of $x$ in the solutions of group B. Similarly, in the solutions of groups D and E, $x$ oscillates three and four times, respectively. Furthermore, $\dot{\theta}_1$ has one, two, and three peaks in the solutions of groups A, B, and C, respectively. The time responses of the state variables in limit cycles in the traditional compass model, which does not include a wobbling mass, applied with the same controller ($\omega = 3$) are shown in Fig.~\ref{fig:variables_compass}. The behavior of the solution is similar to that of the solutions of group A of the proposed model because $\dot{\theta}_1$ has only one peak in one gait cycle. The state variables $\theta_i, \dot{\theta}_i$ ($i = 1,2$) show approximately the same trajectory as that of the solutions of group A. No other qualitatively different solutions were obtained for the compass model without the wobbling mass. \begin{figure} \centering \includegraphics[width=80mm]{fig/variables.pdf} \caption{Time profiles of state variables of limit cycles of proposed model for three typical solutions a, b, and c in Fig.\ref{fig:soltuons_distribution}. (a) $k = 0.5$ (solution a), (b) $k = 6.5$ (solution b), and (c) $k = 22.5$ (solution c) for $\alpha = 0.3$ and $\omega = 3$.} \label{fig:variables} \end{figure} \begin{figure} \centering \includegraphics[width=80mm]{fig/variables_normalCompass.pdf} \caption{Time profiles of state variables of limit cycle of traditional compass model using same controller as that of the proposed model with $\omega = 3$.} \label{fig:variables_compass} \end{figure} \subsection{Walking performances} We evaluated the walking performances of the obtained solutions for various $k$, $\alpha$, and $\omega$ values. Figure~\ref{fig:eigenValues} shows the maximum eigenvalues of the obtained limit cycles for various parameters. In a single solution group, a smaller $k$ results in higher stability (Fig.~\ref{fig:soltuons_distribution}). Further, the solutions become unstable when $\alpha$ is too large. When $\omega$ is large, the range of stable $k$ widens. In comparison with the stability of the traditional compass model, the maximum eigenvalues of all solutions of the proposed model are larger (Fig.~\ref{fig:eigenValues}), indicating that the stability of the limit cycles is reduced by the effect of the wobbling mass. \begin{figure} \centering \includegraphics[width = 80mm]{fig/lambda_for_various_param_small.pdf} \caption{Maximum eigenvalues of obtained solutions for $\alpha = 0.25$ (left) and $\alpha = 0.5$ (right) for (a) $\omega = 2$ and (b) $\omega = 3$. A, B, C, and D indicate solution groups in Fig.~\ref{fig:soltuons_distribution}. Red and blue curves indicate stable and unstable solutions, respectively. Horizontal dashed lines indicate the maximum eigenvalue of the compass model without a wobbling mass.} \label{fig:eigenValues} \end{figure} Figure~\ref{fig:ZMP} shows the distance $d$ between the toe and ZMP for various parameters. $d$ is small for ($k,\alpha$), where the stability of the solutions is high (maximum eigenvalue is small). The minimum value of $d$ in the solutions of group A is larger than that in the solutions of groups B and C. Furthermore, $d$ is larger when $\alpha$ and $\omega$ are large. In the solutions of group A, $d$ is always larger than that of the traditional compass model. However, in some stable solutions of groups B and C at $\omega = 2$, $d$ is smaller than that of the traditional compass model. In addition, $d$ decreases rapidly with decreasing $k$. \begin{figure} \centering \includegraphics[scale=1.0]{fig/d_for_various_param_small.pdf} \caption{Distance $d$ between toe and ZMP of obtained solutions for $\alpha = 0.25$ (left) and $\alpha = 0.5$ (right) for (a) $\omega = 2$ and (b) $\omega = 3$. A, B, C, and D indicate solution groups in Fig.~\ref{fig:soltuons_distribution}. Red and blue curves indicate stable and unstable solutions, respectively. Horizontal dashed lines indicate $d$ of the compass model without a wobbling mass.} \label{fig:ZMP} \end{figure} Moreover, the time profiles of $d$ for three typical stable solutions in groups A, B, and C with $\alpha = 0.3$ and $\omega = 3$ are shown in Fig.~\ref{fig:time-d}. $d$ becomes positive in the first half of the stance phase and then slowly becomes negative. Although there is no qualitative difference in the time profiles of $d$ among the solution groups, the solution with $k = 0.5$, which is a solution of group A, exhibits the largest fluctuation. \begin{figure} \centering \includegraphics[width = 80mm]{fig/time-d.pdf} \caption{Time profiles of $d$ for $k = 0.5$ (red, solution a in Fig.~\ref{fig:soltuons_distribution}), $k = 5.5$ (green, solution b in Fig.~\ref{fig:soltuons_distribution}), and $k = 18.5$ (blue, solution c in Fig.~\ref{fig:soltuons_distribution}) for $\alpha = 0.3$, $\omega = 3$.} \label{fig:time-d} \end{figure} The CoT for various parameters is shown in Fig.~\ref{fig:CoT}. The smaller the CoT, the higher is the energy efficiency. The CoT is small for ($k,\alpha$), where the limit cycles show high stability. The minimum CoT value for solutions of group A is larger than that for solutions of groups B and C. Furthermore, for larger $\alpha$ and $\omega$, the CoT is larger, indicating that the energy efficiency is lower. Although the CoT of the proposed model is always larger than that of the compass model for solutions of group A, it is smaller than that of the compass model for some stable solutions of groups B and C. Further, the CoT decreases rapidly as $k$ decreases. \begin{figure} \centering \includegraphics[scale=1.0]{fig/CoT_for_various_param_small.pdf} \caption{Energy efficiency (CoT) of obtained solutions for $\alpha = 0.25$ (left) and $\alpha = 0.5$ (right) for (a) $\omega = 2$ and (b) $\omega = 3$. A, B, C, and D indicate solution groups in Fig.~\ref{fig:soltuons_distribution}. Red and blue curves indicate stable and unstable solutions, respectively. Horizontal dashed lines indicate the CoT of the compass model without a wobbling mass.} \label{fig:CoT} \end{figure} \section{Discussion} \subsection{Parameter dependency of the model} The obtained results revealed that the body spring constant $k$ plays an important role in determining the solution group of the proposed compass model equipped with a wobbling mass. Depending on the spring constant, the solutions were obtained as some discrete groups (A, B, C, D, and E). Solutions in different groups exhibit qualitatively different behaviors, particularly in the time profiles of the wobbling mass position $x$. In usual walking, the body receives a deceleration force from the ground in the first half of the stance phase and an acceleration force in the second half of the stance phase. In the solutions of group A, the wobbling mass moves backward in the first half of the stance phase and forward in the second half, as shown in Fig.~\ref{fig:variables}a. In other words, it decelerates the body in the first half of the stance phase and accelerates the body in the second half. By contrast, in the solutions of group B, the wobbling mass moves forward in the first half of the stance phase and backward in the second half, exerting force in a direction that cancels the acceleration/deceleration of the body. Therefore, the large change in acceleration and deceleration of the body is suppressed by the wobbling mass in the solutions of group B. The existence of solution groups is assumed to be associated with the natural frequencies of the body spring. The solutions of group C have double the oscillation of the body spring than the solutions of group B. When the spring constant is further increased, the solutions of groups D and E are seen to involve three- and four-fold oscillations, respectively. Because the solutions are distributed in the $k$-$\alpha$ plane with a repetitive structure, solutions with $n$-fold period are assumed to exist when the spring constant is further increased. Within a single group of solutions, a spring constant dependence was also observed. The smaller the spring constant and the more stable the solution, the better is the energy efficiency and the closer is the ZMP to the toes. When the solution is unstable, the actuator consumes a large amount of energy and undergoes large acceleration and deceleration, resulting in large ZMP fluctuations, poor energy efficiency, and poor stability. The mass ratio $\alpha$ was also shown to affect the walking performance. If $\alpha$ is too large or too small, few stable solutions are found, and the optimal value is approximately $0.4 < \alpha < 0.5$ (Fig.~\ref{fig:soltuons_distribution}). By contrast, the larger the $\alpha$ value, the larger is the distance $d$ between the ZMP and the toes and CoT (Figs.~\ref{fig:ZMP} and~\ref{fig:CoT}). The results indicate that $\alpha$ should not be too large when designing a robot with a wobbling mass. Furthermore, as the phase angular velocity $\omega$ of the CPG increases, stable solutions are obtained over a wider range of $k$ and $\alpha$ values (Fig.~\ref{fig:eigenValues}). However, CoT and $d$ also increase as $\omega$ increases (Figs.~\ref{fig:ZMP} and~\ref{fig:CoT}). Walking at a fast pace requires large amounts of energy, resulting in large acceleration, deceleration, and $d$. \subsection{Effect of wobbling mass} We obtained several qualitatively different solutions for the proposed model with a wobbling mass. The state variables in the limit cycles of solutions of group A showed a behavior similar to that of the traditional compass model (Figs.~\ref{fig:variables} and~\ref{fig:variables_compass}). By contrast, solutions of groups B and C exhibited behaviors different from those of the traditional compass model, with $\dot{\theta}_1$ having multiple peaks and the wobbling mass undergoing oscillations. This behavior arose owing to the effect of the body spring that supports the wobbling mass. Furthermore, solutions of groups B and C had lower stability but better energy efficiency and a shorter distance between the ZMP and toe position than those of the traditional compass model. The stability of a limit cycle is a measure of how quickly it converges when a perturbation is applied in the limit cycle. The distance between the ZMP and the toes is a measure of whether or not a walker falls. In solutions of groups B and C, the motion of the ZMP was suppressed by the motion of the wobbling mass. Therefore, the risk of accidental falling while walking is expected to be reduced by the wobbling mass. In addition, because the acceleration and deceleration of the body are suppressed to be small, the work exerted by the actuators is also small; this is thought to improve the energy efficiency. These results indicate that the solutions of groups B and C use the wobbling mass to increase the energy efficiency and reduce the risk of accidental falling while walking. \subsection{Relationship between model and actual walking} In this study, we investigated the effect of the dynamics of a horizontally wobbling mass on the walking performance. The results showed that passively oscillating elements of the upper body could reduce the risk of accidental falling and improve the energy efficiency. Therefore, the walking performance can be improved by attaching a horizontal wobbling mass to an actual walking biped robot. In human walking, the soft tissues of the upper body may behave as a wobbling mass. In the model solution, the displacement of the horizontal wobbling mass was approximately $|x| = 0.05$; in other words, this displacement was only approximately 5\% of the leg length. Because even such a small oscillation can affect the gait, soft tissues in humans may play the same role as the wobbling mass in our model. Furthermore, the results suggested the possibility of a passive walking assist device that oscillates back and forth. Some birds tend to swing their heads forward and backward while walking. Although birds reportedly perform such movements to stabilize their vision, it may have dynamic effects similar to those of the wobbling mass in the proposed model. These movements may also have the effect of increasing the efficiency and stability of gait. \section{Conclusion} This study proposes a compass model with a wobbling mass that oscillates in the horizontal direction to reveal the influence of the horizontal motion of the upper body on walking. We searched the limit cycles of the model and their stability, energy efficiency, and risk of falling were investigated. Several qualitatively different limit cycles were obtained depending mainly on the spring constant that supports the wobbling mass. Specific types of solutions reduced the risk of falling and improved the energy efficiency, although the stability was decreased compared to those of the traditional compass model. Such results were due to the wobbling mass moving in the opposite direction to the upper body, thereby preventing large changes in acceleration and deceleration while walking. The obtained results also suggest that humans can use the soft tissues of the upper body to improve the gait performance. \subsection{Limitations and future works} The proposed wobbling mass can move only in the horizontal direction. However, the human body has wobbling parts, including arms and internal organs, that can move like a pendulum or move vertically. Future works will aim to clarify the relationship between these different oscillating parts and the actual human body. We used a controller using a CPG for the proposed model, however, its effect of it has not been investigated except for the frequency. We would like to investigate the effect of the neural system on the body with a wobbling mass. Furthermore, the gait of the proposed model will be compared with that of humans, and changes in gait when the walking support device is actually attached to the human body will be investigated. Finally, the relationship with the gait of birds will be investigated, and the similarities and differences in the bipedal gait of humans and birds will be clarified. \section*{Acknowledgement} This work was supported in part by JSPS KAKENHI Grant Number JP21K14104 and JP22H01445. \small \bibliographystyle{IEEEtran}

5949745c2337b5e45930c50177cd6506c26995f3

\section{Introduction} Quandles are algebraic structures modeled on the \emph{three} Reidemeister moves in classical knot theory. They have been used extensively to construct invariants of knots and links, see for example \cites{EN, Joyce, Matveev}. A Topological quandle is a quandle with a topology such that the quandle binary operation is compatible with the topology. Precisely, the binary operation is continous and the right multiplications are \emph{homeomorphisms}. Topological quandles were introduced in \cite{Rubin} where it was shown that the set of homomorphisms from the fundamental quandle of the knot to a topological quandle (called also the set of colorings) is an invariant of the knot. Equipped with the compact-open topology, the set of colorings is a topological space. In \cite{EM} a foundational account about topological quandles was given. More precisely, the notions of ideals, kernels, units, and inner automorphism group in the context of topological quandle were introduced. Furthermore, modules and quandle group bundles over topological quandles were introduced with the purpose of studying central extensions of topological quandles. Continuous cohomology of topological quandles was introduced in \cite{ESZ} and compared to the algebraic theories. Extensions of topological quandles were studied with respect to continuous 2-cocycles, and used to show differences in second cohomology groups for some specific topological quandles. Nontriviality of continuous cohomology groups for some examples of topological quandles was shown. In in \cite{CES} the problem of classification of topological Alexander quandle structures, up to isomorphism, on the real line and the unit circle was investigated. In \cite{Gr} the author investigated quandle objects internal to groups and topological spaces, extending the well-known classification of quandles internal to abelian groups \cite{Szymik}. In \cite{Tak} quandle modules over quandles endowed with geometric structures were studied. The author also gave an infinitesimal description of certain modules in the case when the quandle is a regular s-manifold (smooth quandle with certain properties). Since any finite $T_1$-space is discrete, the category of finite $T_0$-spaces was considered in \cite{Stong}, where the point set topological properties of finite spaces were investigated. The homeomorphism classification of finite spaces was investigated and some representations of these spaces as certain classes of matrices was obtained. This article arose from a desire to better understand the analogy of the work given in \cite{Stong} in the context of \emph{finite topological} quandles. It turned out that: there is no $T_0$-topology on any finite connected (meaning one orbit under the action of the Inner group) quandle $X$ that makes $X$ into a topological quandle (Theorem~\ref{noT0}). Thus we were lead to consider topologies on quandles with more than \emph{one} orbit. It is well known \cite{Alex} that the category of Alexandroff $T_0$-spaces is equivalent to the category of \emph{partially ordered sets} (posets). In our context, we prove that for a quandle $X$ with more than one orbit, there exists a unique non trivial topology which makes right multiplications of $X$ continuous maps (Proposition~\ref{Prop}). Furthermore, we prove that if $X$ be a finite quandle with two orbits $X_1$ and $X_2$ then any continuous poset on $X$ is biparatite with vertex set $X_1$ and $X_2$ (Proposition~\ref{Bipar}). This article is organized as follows. In Section~\ref{Review} we review the basics of topological quandles. Section~\ref{Poset} reviews some basics of posets, graphs and some hierarchy of separation axioms. In Section~\ref{Main} the main results of the article are given. Section~\ref{Computations} gives some explicit computations based on some computer softwares (Maple and Python) of quandles up to order \emph{five}. \section{Review of Quandles and Topological Quandles}\label{Review} A quandle is a set $X$ with a binary operation $*$ satisfying the following three axioms: \begin{enumerate} \item For all $x$ in $X$, $x*x=x,$ \item For all $y,z \in X$, there exists a unique $x$ such that $x*y=z$, \item For all $x,y,z \in X$, $(x*y)*z=(x*z)*(y*z)$. \end{enumerate} These three conditions come from the axiomatization of the three Reidemeister moves on knot diagrams. The typical examples of quandles are: (i) Any Group $G$ with conjugation $x*y=y^{-1}xy$, is a quandle called the \emph{conjugation quandle} and (ii) Any group $G$ with operation given by $x*y=yx^{-1}y$, is a quandle called the \emph{core quandle}.\\ Let $X$ be a quandle. For an element $y \in X$, left multiplication $L_y$ and right multiplication $R_y$ by an element $y$ are the maps from $X$ to $X$ given respectively by $L_y(x):=y*x$ and $R_y(x)=x*y$. A function $f: (X,*) \rightarrow (X,*)$ is a quandle {\em homomorphism} if for all $x,y \in X, f(x * y)=f(x) * f(y)$. If furthermore $f$ is a bijection then it it is called an \emph{automorphism} of the quandke $X$. We will denote by {\rm Aut(X)} the automorphism group of $X$. The subgroup of {\rm Aut(X)}, generated by the automorphisms $R_x$, is called the {\em inner} automorphism group of $X$ and denoted by {\rm Inn}$(X)$. If the group {\rm Inn}$(X)$ acts \emph{transitively} on $X$, we then say that $X$ is connected quandle meaning it has only one orbit. Since we do not consider topological connectedness in this article, then through the whole article, the word connected quandle will stand for algebraic connectedness. For more on quandles refer to \cite{EN, Joyce, Matveev, E}. Topological quandles have been investigated in \cite{CES, EM, Rubin, ESZ}. Here we review some basics of topological quandles. \begin{definition} A \textit{topological quandle} is a quandle $X$ whith a topology such that the map $X\times X\ni (x,y)\mto x * y\in X$ is a continuous, the right multiplication $R_x:X\ni y\mto y* x\in X$ is a homeomorphism, for all $x\in X$, and $x* x=x$. \end{definition} It is clear that any finite quandle is automatically a topological quandle with respect to the discrete topology. \begin{example} \cite{CES} Let $(G, +)$ be a topological abelian group and let $\sigma$ be a continuous automorphism of $G$. The continuous binary operation on $G$ given by $x*y=\sigma(x)+(Id-\sigma)(y), \forall x,y \in G,$ makes $(G,*)$ a topological quandle called \textit{topological Alexander quandle}. In particular, if $G=\mathbb{R}$ and $\sigma(x)=tx$ for non-zero $t \in \mathbb{R}$, we have the topological Alexander structure on $\mathbb{R}$ given by $x*y=tx+(1-t)y$. \end{example} \begin{example} The following examples were given in \cites{Rubin,EM}. The unit sphere $\mathbb{S}^n \subset \mathbb{R}^{n+1}$ with the binary operation $x\ast y=2(x\cdot y)y-x$ is a topological quandle, where $\cdot$ denotes the inner product of $\mathbb{R}^{n+1}$. Now consider $\lambda$ and $\mu$ be real numbers, and let $x,y \in S^n$. Then $$\lambda x * \mu y=\lambda[2{\mu}^2(x\cdot y)y-x].$$ In particular, the operation $$\pm x * \pm y=\pm (x * y)$$ provides a structure of topological quandle on the quotient space that is the projective space $\mathbb{RP}^n$. \end{example} \section{Review of topologies on finite sets, Posets and Graphs}\label{Poset} Now we review some basics of directed graphs, posets and $T_0$ and $T_1$ topologies. \begin{definition} A \emph{directed} graph G is a pair $(V, E)$ where $V$ is the set of vertices and $E$ is a list of directed line segments called edges between pairs of vertices. \end{definition} An edge from a vertex $x$ to a vertex $y$ will be denoted symbolically by $x < y$ and we will say that $x$ and $y$ are \emph{adjacent}. The following is an example of a directed graph. \begin{example} Let $G=(V,E) $ where $V=\{a,b,c,d\}$ and $E=\{b<a,c<a ,a<d \}$. \end{example} \begin{center} \begin{figure}[h!] \begin{tikzpicture} [nodes={draw, circle}, <-] \node{d} child { node {a} child { node {b} } child { node {c} } } child [ missing ]; \end{tikzpicture} \end{figure} \end{center} \begin{definition} An \emph{idependent set} in a graph is a set of pairwise non-adjacent vertices. \end{definition} \begin{definition} A (directed) graph $G=(V,E)$ is called biparatite if $V$ is the union of two disjoint independent sets $V_1$ and $V_2$. \end{definition} \begin{definition} A (directed) graph $G$ is called \emph{complete biparatite} if $G$ is bipartite and for every $v_1 \in V_1$ and $v_2 \in V_2$ there is an edges in $G$ that joins $v_1$ and $v_2$ \end{definition} \begin{example} Let $V=V_1 \cup V_2$ where $V_1=\{4,5\}$ and $V_2=\{1,2,3\}$. Then the directed graph $G=(V,E)$ is complete biparatite graph. \begin{center} \includegraphics[]{biparatite.png} \end{center} \end{example} Now we recall the definition of partially ordered set. \begin{definition} A partially ordered set (poset) is a set $X$ with an order denoted $\leq$ that is reflexive, antisymmetric and transitive. \end{definition} \begin{example} For any set $X$, the power set of $X$ ordered by the set inclusion relation $\subseteq$ forms a poset $( \mathcal{P}(X),\subseteq)$ \end{example} \begin{definition} Two partially ordered sets $P=(X,\leq)$ and $Q=(X,\leq')$ are said to be isomorphic if there exist a bijection $f:X \rightarrow X'$ such that $x \leq y$ if and only if $f(x)\leq'f(y).$ \end{definition} \begin{definition}\label{ConnPoset} A poset $(X,\leq )$ is connected if for all $x,y\in X$, there exists sequence of elements $x=x_1,x_2, \ldots, x_n=y $ such that every two consecutive elements $x_i$ and $x_{i+1}$ are comparable (meaning $x_i < x_{i+1}$ or $x_{i+1} < x_i$). \end{definition} \noindent {\bf Notation:} Given an order $\leq$ on a set $X$, we will denote $x<y$ whenever $x \neq y$ and $x \leq y$. Finite posets $(X,\leq)$ can be drawn as directed graphs where the vertex set is $X$ and an arrow goes from $x$ to $y$ whenever $x \leq y$. For simplicity, we will not draw loops which correspond to $x \leq x$. We will then use the notation $(X,<)$ instead of $(X,\leq)$ whenever we want to ignore the reflexivity of the partial order \begin{example}. Let $X=\mathbb{Z}_8$ be the set of integers modulo $8$. The map $f:X \rightarrow X$ given by $f(x)=3x-2$ induces an isomorphism between the following two posets $(X,<)$ and $(X,<')$. \begin{center} \includegraphics{example1.png}, \qquad \includegraphics{example2.png} \end{center} \end{example} \begin{definition} A chain in a poset $(X, <)$ is a subset $C$ of $X$ such that the restriction of $<$ to $C$ is a total order (i.e. every two elements are comparable). \end{definition} Now we recall some basics about topological spaces called $T_0$ and $T_1$ spaces. \begin{definition} A topological space $X$ is said to have the property $T_0$ if for every pair of distinct points of $X$, at least one of them has a neighborhood not containing the other point. \end{definition} \begin{definition} A topological space $X$ is said to have the property $T_1$ if for every pair of distinct points of $X$, each point has a neighborhood not containing the other point. \end{definition} Obviously the property $T_1$ implies the property $T_0$. Notice also that this definition is equivalent to saying singletons are closed in $X$. Thus a $T_1$-topology on a \emph{finite} set is a discrete topology. Since any finite $T_1$-space is discrete, we will focus on the category of finite $T_0$-spaces. First we need some notations. Let $X$ be a finite topological space. For any $x \in X$, we denote \[ U_x:=\textit{the smallest open subset of $X$ containing $x$} \] It is well known \cite{Alex} that the category of $T_0$-spaces is isomorphic to the category of posets. We have $x \leq y$ if and only if $U_y \subseteq U_x$ which is equivalent to $C_x \subset C_y$, where $C_v$ is the complelement $U_v^{c}$ of $U_v$ in $X$. Thus one obtain that $U_x=\{w \in X;\; x \leq w\}$ and $C_x=\{v \in X; \;v < x\}$. Under this correspondence of categories, the subcategory of finite posets is equivalent to the category of finite $T_0$-spaces. \\ Through the rest of this article we will use the notation of $x<y$ in the poset whenever $x\neq y$ and $x \leq y$. \begin{comment} $$\begin{array}{|c|c| c| c| c| c|} \hline n \ & Distinct\\& Topologies & Distinct\\&& T_0-Topologies & inequivalent\\&&& topologies & T_0-inequivalent\\&&&& topologies \\ \hline \ 1 & 1 &1 & 1 & 1 \\ \hline \ 2 & 4 & 3 & 3 & 2 \\ \hline \ 3 & 29 & 19 & 9 & 5 \\ \hline \ 4 & 355 & 219 & 33 & 16 \\ \hline \ 5 & 6942 & 4231 & 139 & 63 \\ \hline \ 6 & 209527 & 130023 & 718 & 318 \\ \hline \end{array} $$ \end{comment} \section{Topologies on non- connected Quandles}\label{Main} As we mentioned earlier, since $T_1$-topologies on a finite set are discrete, we will focus in this article on $T_0$-topologies on \emph{finite quandles}. A map on finite spaces is continuous if and only if it preserves the order. It turned out that on a finite quandle with a $T_0$-topology, left multiplications can not be continuous as can be seen in the following theorem \begin{theorem}\label{left} Let $X$ be a finite quandle endowed with a $T_0$-topology. Assume that for all $z \in X$, the map $L_z$ is continuous, then $x \leq y$ implies $L_z(x)=L_z(y)$. \end{theorem} \begin{proof} We prove this theorem by contradiction. Let $X$ be a finite quandle endowed with a $T_0$-topology. Assume that $x \leq y$ and $L_z(x) \neq L_z(y)$. If $x=y$, then obviously $L_z(x)=L_z(y)$. Now assume $x <y$, then for all $a \in X$, the continuity of $L_a$ implies that $a*x \leq a*y$. Assume that there exist $a_1 \in X$ such that, $z_1:=a_1*x=L_{a_1}(x) < a_1*y=L_{a_1}(y)$. The invertibility of right multiplications in a quandle implies that there exist unique $a_2$ such that $a_2*x=a_1*y$ hence $a_1*x<a_2*x$ which implies $a_1 \neq a_2.$ Now we have $a_1*x<a_2*x \leq a_2 *y=z_2$. We claim that $a_2*x < a_2 *y$. if $a_2*y=a_2*x$ and since $a_2*x=a_1*y$ we will have $a_2*y=a_2*x=a_1*y$ hence $a_2*y=a_1*y$ but $a_1 \neq a_2$, thus contradiction. Now that we have proved $a_2*x < a_2 *y$, then there exists $a_3$ such that $a_2*y=a_3*x$ we get, $a_2*x<a_3*x$ repeating the above argument we get, $a_3*x<a_3*y$. Notice that $a_1,a_2$ and $a_3$ are all pairwise disjoint elements of $X$. Similarly, we construct an \emph{infinite} chain, $a_1 * x< a_2*x <a_3 *x < \cdots $, which is impossible since $X$ is a finite quandle. Thus we obtain a contradiction. \end{proof} \begin{comment} Theorem~\ref{left} states that we can not have a notion of \emph{semitopological quandles} (continuity with respect to each variable considered separately) on finite quandles. For more on semitopological groups see \cite{Husain}. Thus through the rest of the paper, the word continuity will mean continuity of \emph{right multiplications} in the finite quandle. We will then say that the quandle with its topology is \emph{right continuous}. We borrow the following terminologies from the book \cite{Rup}. \end{comment} We have the following Corollary \begin{corollary} Let $X$ be a finite quandle endowed with a $T_0$-topology. If $C$ is a chain of $X$ as a poset then any left continuous function $L_x$ on $X$ is a constant function on $C$. \end{corollary} \begin{definition} A quandle with a topology in which right multiplications (respectively left multiplications) are continuous is called \emph{right topological quandle} (respectively \emph{left topological quandle}). \end{definition} In other words, right topological quandle means that for all $x,y,z \in X$, \[ x<y \implies x*z < y*z. \] and, since left multiplications are not necessarly bijective maps, left topological quandle means that for all $x,y,z \in X$, \[ x<y \implies z*x \leq z*y. \] \begin{comment} \begin{definition} A quandle with a topology in which both right multiplications and left multiplications are continuous is called \emph{semitopological quandle}. \end{definition} Obviously topological quandle implies semitopological quandle which implies right topological quandles, but the converses are not true. \end{comment} \begin{theorem}\label{noT0} There is no $T_0$-topology on a finite connected quandle $X$ that makes $X$ into a right topological quandle. \end{theorem} \begin{proof} Let $x <y$. Since $X$ is connected quandle, there exists $\phi \in Inn(X)$ such that $y=\phi(x)$. Since $X$ is finite, $\phi$ has a finite order $m$ in the group $Inn(X)$. Since $\phi$ is a continous automorphism then $x<\phi(x)$ implies $x<\phi^m(x)$ giving a contradiction. \end{proof} \begin{corollary}\label{noodd} There is no $T_0$-topology on any latin quandle that makes it into a right topological quandle. \end{corollary} \begin{comment} \begin{proof} For Dihedral quandle $R_y(x)=2y-x$\\ If n is odd then 2 is invertible in $Z_n$, hence for every\\ $z \in \mathbb(z)_n$ there exist $y=\frac{z+x}{2}$ such that $R_y(x)=z$ hence $R_y(x)$ is both injective in $x$ and $y$ hence quandle will have only one orbit.\\ Hence it is connected. \end{proof} \end{comment} Thus Theorem~\ref{noT0} leads us to consider quandles $X$ that are not connected, that is $X=X_1 \cup X_2\cup \ldots X_k$ as orbit decomposition, search for $T_0$-topology on $X$ and investigate the continuity of the binary operation. \begin{proposition}\label{Prop} Let $X$ be a finite quandle with orbit decomposition $X=X_1 \cup \{a\}$, then there exist unique non trivial $T_0$-topology which makes $X$ right continuous. \end{proposition} \begin{proof} Let $X=X_1 \cup \{a\}$ be the orbit decomposition of the quandle $X$. For any $x,y \in X_1$, there exits $\phi \in Inn(X)$ such that $\phi(x)=y$ and $\phi(a)=a$. Declare that $x<a$, then $\phi(x)<a$. Thus for any $z\in X_1$ we have $z<a$. Uniqueness is obvious. \end{proof} The $T_0$-topology in Proposition~\ref{Prop} is precisely given by $x<a$ for all $x \in X_1.$ \begin{proposition} \label{Bipar} Let $X$ be a finite quandle with two orbits $X_1$ and $X_2$. Then any right continuous poset on $X$ is biparatite with vertex set $X_1$ and $X_2$. \end{proposition} \begin{proof} We prove this proposition by contradiction. For every $x_1, y_1 \in X_1$ such that $x_1 <y_1$. We know that there exist $\phi \in Inn(X)$ such that $\phi(x_1)=y_1$. Hence, $x_1< \phi(x_1)$ implies $x_1 <\phi^m(x_1)=x_1$, where $m$ is the order of $\phi$ in $Inn(X)$. Thus we have a contradiction. \end{proof} \begin{proposition}\label{biComp} Let $X$ be a finite quandle with two orbits $X_1$ and $X_2$. Then the complete bipartite graph with vertex set $X_1$ and $X_2$ forms a right continuous poset. \end{proposition} \begin{proof} Let $X$ be a finite quandle with two orbits $X_1$ and $X_2$. If $x \in X_1$ and $y \in X_2$ then for every $\phi \in Inn(X)$ we have $\phi(x) \in X_1$ and $\phi(y) \in X_2$. Proposition~\ref{Bipar} gives that the graph is bipartite and thus $x<y$. We then obtain $\phi(x) < \phi(y)$ giving the result. \end{proof} \begin{remark} By Proposition~\ref{biComp} and Theorem~\ref{left}, there is a non-trivial $T_0$-topology making $X$ right continuous if and only if the quandle has more than one orbit. \end{remark} Notice that Proposition~\ref{biComp} can be generalized to $n$-paratite complete graph. The following table gives the list of right continuous posets on some even dihedral quandles. In the table, the notation $(a,b)$ on the right column means $a<b.$ \begin{table}[H] \caption{Right continuous posets on dihedral quandles}\label{TTable1} \begin{center} \begin{tabular}{|c|c|} \hline Quandle & Posets \\ \hline $R_4$ & ((0,1),(2,1),(0,3),(2,3)) \\ \hline $R_6$ & ((0, 1), (0, 5), (2, 1), (2, 3), (4, 3), (4, 5))\;; \\ & ((0,3), (2, 5), (4, 1)). \\ \hline $R_8$ & ((2, 7), (4, 7), (6, 1), (6, 3), (0, 5), (2, 5), (4, 1), (0, 3))\;;\\ & (( 0, 1), (6, 7), (4, 5), (0, 7), (2, 1), (2, 3), (4, 3), (6, 5)). \\ \hline $R_{10}$ & ((0, 1), (6, 7), (4, 5), (2, 1), (8, 9), (2, 3), (4, 3), (8, 7), (0, 9), (6, 5)) \;;\\ & ((4, 7), (6, 9), (2, 9), (8, 1), (8, 5), (0, 7), (6, 3), (2, 5), (4, 1), (0, 3))\;;\\ & ((2, 7), (8, 3), (0, 5), (4, 9), (6, 1)).\\ \hline \end{tabular} \end{center} \end{table} Notice that in table~\ref{TTable1}, the dihedral quandle $R_4$ has only one right continuous poset $((0,1),(2,1),(0,3),(2,3))$ which is complete biparatite. While the dihedral quandle $R_6$ has two continuous posets $((0, 1), (0, 5), (2, 1), (2, 3), (4, 3), (4, 5))$ and $((0,3), (2, 5), (4, 1))$ illustrated below. \begin{center} \begin{tikzpicture} [scale=.8,auto=left,every node/.style={circle,fill=gray}] \node (n6) at (2,10) {3}; \node (n4) at (2,7) {0}; \node (n5) at (3,10) {5}; \node (n1) at (3,7) {2}; \node (n2) at (4,10) {1}; \node (n3) at (4,7) {4}; \foreach \from/\to in {n6/n4,n5/n1,n2/n3} \draw (\from) -> (\to); \end{tikzpicture} \qquad \qquad \qquad \begin{tikzpicture} [scale=.8,auto=left,every node/.style={circle,fill=gray}] \node (n6) at (14,10) {3}; \node (n1) at (13,7) {2}; \node (n5) at (13,10) {5}; \node (n4) at (14,7) {0}; \node (n2) at (15,10) {1}; \node (n3) at (15,7) {4}; \foreach \from/\to in {n4/n2,n2/n1, n5/n3,n6/n1,n5/n3,n4/n5, n3/n6} \draw (\from) -- (\to); \end{tikzpicture} \end{center} Moreover, in table~\ref{TTable1}, for $R_8$ the bijection $f$ given by $f(k)=3k-2$ makes the two posets isomorphic. The same bijection gives isomorphism between the first two posets of $R_{10}$. The following Theorem characterizes non complete biparatite posets on dihedral quandles. \begin{theorem} Let $R_{2n}$ be a dihedral quandle of even order. Then $R_{2n}$ has $s+1$ right continuous posets, where $s$ is number of odd natural numbers less than n and relatively non coprime with $n$ \end{theorem} \begin{proof} Let $X=R_{2n}$ be the dihedral quandle with orbits $X_1=\{0,2, \ldots, 2n-2\}$ and $X_2=\{1,3, \ldots ,2n-1\}$. For every $x \in X_2$, we construct a partial order $<_x$ on $R_{2n}$, such that for all $y \in X$, we have $2y<_x2y-x$ and $2y <_x 2y+x$. Then $<_x$ is clearly right continuous partial order since $2y<2y-x$ and $2y<2y+x$ for all $y$ imply that $2z-2y<2z-(2y-x)$. In other words we obtain $2y*z<(2y-x)*z$. From the definition of the order $<_x$ it is clear the two partial orders $<_x$ and $<_{2n-x} $ are the same. Hence we obtain the following distinct partial orders $<_1, <_3, \ldots $. Now we check which ones are isomorphic. If $m$ is odd and $gcd(n,m)=1$ then $f(k)=mk-2$ is a bijective function making $<_1$ and $<_m$ isomorphic. Now let $m$ be odd and $gcd(m,n)=k>1$. The two posets $<_1$ and $<_m$ are non isomorphic since $<_1$ is connected poset, as in Definition~\ref{ConnPoset}, and $<_m$ is not connected poset. We show that these are the only right continuous posets. Given a right continuous poset on $R_{2n}$ then $ a < b$ can be written as $a < a-(a-b)$ which implies that $a <_x b$ where $x=a-b.$ Now if $a < b$ then by Proposition~\ref{Bipar}, we have $a \in X_1$, $b \in X_2$. Now let $a = 2\alpha$ and $b = 2\beta+1$ then $a-b = 2(\alpha-\beta)-1 \in X_2$. This ends the proof. \end{proof} \begin{corollary} For the dihedral quandle $R_{2^n}$ with $2^n$ elements, there is a unique right continuous poset. \end{corollary} \vspace*{-7pt} \section{Some Computer Calculations}\label{Computations} In this section we give non-trivial right and left continuous posets on the finite quandles of order up to $5$ based on Maple and Python computations. In the following tables we have excluded the trivial and connected quandles \begin{table}[H] \caption{Continuous posets on quandles of order $3$} \label{Table2} \begin{center} \begin{tabular}{ |c|c|c|} \hline Quandle for n = 3 & Right continuous Posets & Left continuous poset \\ \hline \small{ $\left[ \begin{array}{c} 0 \;0 \;1 \\ 1 \;1 \;0 \\ 2 \;2\; 2 \end{array} \right] $} & ((0,2),(1,2)) & ((0,1))\\ \hline \end{tabular} \end{center} \end{table} As seen in Table~\ref{Table2} for $n=3$, there exist a unique right continuous poset and a unique left continuous poset. \begin{table}[H] \caption{Continuous posets on quandles of order 4} \label{Table3} \begin{center} \begin{tabular}{ |c|c|c|} \hline Quandles for n = 4 & Right continuous poset & Left continuous poset \\ \hline \small{ $\left[ \begin{array}{c} 0 \;0 \;0 \;0 \\ 1 \;1 \;1 \;2 \\ 2 \;2 \;2 \;1 \\ 3 \;3 \;3 \;3 \end{array} \right] $} & \makecell{ ((0, 3))\;;\\ ((0, 1), (0, 2), (0, 3))\;;\\ ((0, 1), (0, 3), (1, 2))\;;\\ ((0, 1), (0, 2), (1, 3), (2, 3))\;;\\ ((2, 3), (1, 3))\;;\\ ((2, 3), (1, 3), (0, 3)).\\ } & ((0,1),(1,2)) and ((1,2)) \\ \hline \small{ $\left[ \begin{array}{c} 0 \;0 \;0 \;1 \\ 1 \;1\; 1 \;2 \\ 2 \;2 \;2 \;0 \\ 3 \;3 \;3 \;3 \end{array} \right] $} & \makecell{ ((0,3),(1,3),(2,3))\\ } & ((0,1),(1,2)) and ((1,2)) \\ \hline \small{ $\left[ \begin{array}{c} 0 \;0 \;1 \;1 \\ 1 \;1 \;0 \;0 \\ 2 \;2 \;2 \;2 \\ 3 \;3 \;3 \;3 \end{array} \right] $} & \makecell{ ((0,2),(1,2),(0,3),(1,3),(2,3))\;; \\ ((0,2),(1,2),(0,3),(1,3)) \;;\\ ((0,2),(1,2)) \;;\\ ((2,3)).\\ } & ((0,1),(2,3)) and ((2,3))\\ \hline \small{ $\left[ \begin{array}{c} 0 \;0 \;0 \;0 \\ 1 \;1 \;3 \;2 \\ 2 \;3 \;2 \;1 \\ 3 \;2 \;1 \;3 \end{array} \right] $} & \makecell{ ((0,1),(0,2),(0,3)) \\ } & None \\ \hline \small{ $\left[ \begin{array}{c} 0 \;0 \;1 \;1 \\ 1 \;1 \;0 \;0 \\ 3 \;3 \;2 \;2 \\ 2 \;2 \;3 \;3 \end{array} \right] $} & \makecell{ ((0,2),(0,3),(1,2),(1,3)) \\ } & ((0,1),(2,3)) \\ \hline \end{tabular} \end{center} \end{table} \begin{table}[H] \caption{Continuous posets on quandles of order 5, Part I} \label{Table4} \begin{center} \begin{tabular}{ |c|c|c|} \hline Quandles for n = 5& Right continuous& Left continuous \\ \hline \small{ $\left[ \begin{array}{c} 0\; 0\; 0\; 0\; 0 \\ 1\; 1\; 1\; 1\; 1 \\ 2\; 2\; 2\; 2\; 3 \\ 3\; 3\; 3\; 3\; 2 \\ 4\; 4\; 4\; 4\; 4 \end{array} \right] $} & \makecell{\\ ((0,1),(1,2),(1,3),(0,4))\;;\\ ((0,2),(0,3),(1,2),(1,3),(4,2),(4,3))\;;\\ ((0,2),(0,3),(1,2),(1,3),(2,4),(3,4))\;;\\ ((0,1),(1,4),(4,2),(4,3)).\\ \\ } & \makecell{ ((0,1),(1,2),(2,3))\;;\\ ((0,1),(1,2))\;;\\ ((1,2)).} \\ \hline \small{ $\left[ \begin{array}{c} 0\; 0\; 0\; 0\; 0 \\ 1\; 1\; 1\; 1\; 2 \\ 2\; 2\; 2\; 2\; 3 \\ 3\; 3\; 3\; 3\; 1 \\ 4\; 4\; 4\; 4\; 4 \end{array} \right] $} & \makecell{ ((0,1), (0,2), (0,3), (2,4), (3,4), (1,4))\;; \\ ((0,4)) \;;\\ ((0,1), (0,2), (0,3))\;; \\ (( 0,4 ),(4,1 ),(4,2 ),(4,3)). } & \makecell{\\((0,1), (1,2), (2,3)) \;;\\ ((0,1), (1,2)) \;;\\ ((2,3)). \\ } \\ \hline \small{ $\left[ \begin{array}{c} 0\; 0\; 0\; 0\; 1 \\ 1\; 1\; 1\; 1\; 0 \\ 2\; 2\; 2\; 2\; 3 \\ 3\; 3\; 3\; 3\; 2 \\ 4\; 4\; 4\; 4\; 4 \end{array} \right] $} & \makecell{\\ ((1,2),(0,3),(2,4),(3,4))\;;\\ ((1,2),(0,2),(1,3),(0,3),(2,4),(3,4))\;;\\ ((1,4),(0,4))\;;\\ ((1,2),(0,2),(1,3),(0,3)).\\ \\} & \makecell{ ((1,2),(0,1),(2,3))\;; \\ ((0,1),(0,2))\;; \\ ((0,2),(1,2)).\\ \\}\\ \hline \small{ $\left[ \begin{array}{c} 0\; 0\; 0\; 0\; 1 \\ 1\; 1\; 1\; 1\; 2 \\ 2\; 2\; 2\; 2\; 3 \\ 3\; 3\; 3\; 3\; 0 \\ 4\; 4\; 4\; 4\; 4 \end{array} \right] $} & \makecell{((0,4),(1,4),(2,4),(3,4)).} & \makecell{\\ ((1,2),(0,1),(2,3))\;; \\ ((0,1),(0,2))\;; \\ ((0,2),(1,2)).\\ \\ \\ } \\ \hline \small{ $\left[ \begin{array}{c} 0\; 0\; 0\; 0\; 0 \\ 1\; 1\; 1\; 1\; 1 \\ 2\; 2\; 2\; 4\; 3 \\ 3\; 3\; 4\; 3\; 2 \\ 4\; 4\; 3\; 2\; 4 \end{array} \right] $} &\makecell{\\ ((0,2),(0,3),(0,4))\;;\\ ((0,2),(0,3),(0,4),(1,2),(1,3),(1,4))\;;\\ ((0,1))\;;\\ ((0,2),(0,3),(0,4),(0,1),(1,2),(1,3),(1,4)).\\ \\ \\ } & \makecell{((0,1),(0,2),(0,3))\;;\\ ((0,1),(0,2))\;;\\ ((0,1)). } \\ \hline \small{ $\left[ \begin{array}{c} 0\; 0\; 0\; 0\; 0 \\ 1\; 1\; 1\; 2\; 2 \\ 2\; 2\; 2\; 1\; 1 \\ 3\; 3\; 3\; 3\; 3 \\ 4\; 4\; 4\; 4\; 4 \end{array} \right] $} & \makecell{\\ ((1,3),(2,3))\;;\\ ((1,3),(2,3),(1,4),(2,4))\;;\\ ((0,4))\;;\\ ((3,2),(3,1))\;;\\ ((1,3),(2,3),(4,1),(4,2)).\\ \\ } & \makecell{((0,1),(1,2),(3,4))\;; \\ ((0,1),(1,2))\;;\\ ((0,1),(3,4))\;;\\ ((0,1)).\\} \\ \hline \small{ $\left[ \begin{array}{c} 0\; 0\; 0\; 0\; 0 \\ 1\; 1\; 1\; 2\; 2 \\ 2\; 2\; 2\; 1\; 1 \\ 3\; 4\; 4\; 3\; 3 \\ 4\; 3\; 3\; 4\; 4 \end{array} \right] $} & \makecell{((0,1),(0,2),(0,3),(0,4))\;;\\ ((0,1),(0,2))\;;\\ ((0,1),(0,2),(2,3),(2,4),(1,3),(1,4))\;;\\ ((1,3),(1,4),(2,3),(2,4)).} & \makecell{((0,1),(1,2),(3,4))\;; \\ ((0,1),(1,2))\;;\\ ((0,1),(3,4))\;;\\ ((0,1)).\\} \\ \hline \end{tabular} \end{center} \end{table} \begin{table}[H] \caption{Continuous posets on quandles of order 5, Part II} \label{Table1} \begin{center} \begin{tabular}{ |c|c|c|} \hline Quandles for n = 5 & Right continuous & Left continuous \\ \hline \small{ $\left[ \begin{array}{c} 0\; 0\; 0\; 1\; 1 \\ 1\; 1\; 1\; 0\; 0 \\ 2\; 2\; 2\; 2\; 2 \\ 3\; 3\; 4\; 3\; 3 \\ 4\; 4\; 3\; 4\; 4 \end{array} \right] $} & \makecell{ ((0,2), (1,2) (2,3),(2,4))\;;\\ ((0,2), (1,2))\;;\\ ((2,3 ),(2,4))\;;\\ ((0,3),(0,4),(1,3),(1,4)).} & \makecell { ((0,1 ),(1,2 ),(3, 4))\;;\\ ((3,4))\;;\\ ((0,1),(1,2)).\\ } \\ \hline \small{ $\left[ \begin{array}{c} 0\; 0\; 0\; 1\; 1 \\ 1\; 1\; 1\; 2\; 2 \\ 2\; 2\; 2\; 0\; 0 \\ 3\; 3\; 3\; 3\; 3 \\ 4\; 4\; 4\; 4\; 4 \end{array} \right] $} & \makecell{ ((0, 3),(1,3), (2,3), (3,4))\;;\\ ((0,3),(1,3),(2, 3))\;;\\ ((3,4)).} & \makecell{ ((0,1),(1,2),(3,4))\;;\\ ((3,4))\;;\\ ((0,1),(1,2)).} \\ \hline \small{ $\left[ \begin{array}{c} 0\; 0\; 0\; 1\; 2 \\ 1\; 1\; 1\; 2\; 0 \\ 2\; 2\; 2\; 0\; 1 \\ 3\; 3\; 3\; 3\; 3 \\ 4\; 4\; 4\; 4\; 4 \end{array} \right] $} & \makecell{ ((0,3),(1,3),(2,3),(0,4),(1,4),(2,4))\;;\\ ((0,3),(1,3),(2,3))\;;\\ ((3,4)).} & \makecell{ ((0,1),(1,2))\;;\\ ((0,1)).} \\ \hline \small{ $\left[ \begin{array}{c} 0\; 0\; 0\; 0\; 0\\ 1\; 1\; 1\; 1\; 1\\ 2\; 2\; 2\; 2\; 2\\ 4\; 4\; 4\; 3\; 3\\ 3\; 3\; 3\; 4\; 4 \end{array} \right] $} & \makecell{\\ ((0,1),(0,2))\;;\\ ((0,1),(1,3),(1,4))\;;\\ (0,1),(1,2),(2,3),(2,4)).\\ \\ \\ } & \makecell{\\ ((0,1),(0,2))\;;\\ ((0,1),(1,2),(3,4))\;;\\ ((0,1),(3,4))\;;\\ ((3,4)).\\ }\\ \hline \small{ $\left[ \begin{array}{c} 0\; 0\; 0\; 0\; 0\\ 1\; 1\; 1\; 2\; 2\\ 2\; 2\; 2\; 1\; 1\\ 4\; 4\; 4\; 3\; 3\\ 3\; 3\; 3\; 4\; 4 \end{array} \right] $} & \makecell{((0,1),(0,2))\;;\\ ((0,1),(1,3),(1,4))\;;\\ ((1,3),(1,4),(2,3),(2,4)).} & \makecell{\\ ((0,1),(0,2))\;;\\ ((0,1),(1,2),(3,4))\;;\\ ((0,1),(3,4))\;;\\ ((3,4)).\\ \\ }\\ \hline \small{ $\left[ \begin{array}{c} 0\; 0\; 0\; 1\; 1\\ 1\; 1\; 1\; 2\; 2\\ 2\; 2\; 2\; 0\; 0\\ 4\; 4\; 4\; 3\; 3\\ 3\; 3\; 3\; 4\; 4 \end{array} \right] $} & \makecell{((0,4),(1,4),(2,4),(0,3),(1,3),(2,3)).} & \makecell{\\ ((0,1),(0,2))\;;\\ ((0,1),(1,2),(3,4))\;;\\ ((0,1),(3,4))\;;\\ ((3,4)).\\ \\ } \\ \hline \small{ $\left[ \begin{array}{c} 0\; 0\; 0\; 0\; 0\\ 1\; 1\; 4\; 2\; 3\\ 2\; 3\; 2\; 4\; 1\\ 3\; 4\; 1\; 3\; 2\\ 4\; 2\; 3\; 1\; 4 \end{array} \right] $} & \makecell{\\ \\ ((0,1),(0,2),(0,3),(0,4)).\\ \\ \\ } & None \\ \hline \small{ $\left[ \begin{array}{c} 0\; 0\; 1\; 1\; 1 \\ 1\; 1\; 0\; 0\; 0 \\ 2\; 2\; 2\; 2\; 3 \\ 3\; 3\; 3\; 3\; 2 \\ 4\; 4\; 4\; 4\; 4 \end{array} \right] $} & \makecell{\\ ((1,2),(0,2),(1,3),(0,3),(2,4),(3,4))\;;\\ ((1,4),(0,4))\;;\\ ((1,2),(0,2),(1,3),(0,3)). \\ \\ \\ } & \makecell{((0,1),(2,3))\;; \\ ((0,1)).}\\ \hline \end{tabular} \end{center} \end{table} \begin{table}[H] \caption{Continuous posets on quandles of order 5, Part III} \label{Table1} \begin{center} \begin{tabular}{ |c|c|c|} \hline Quandles for n = 5 & Right continuous & Left continuous \\ \hline \small{ $\left[ \begin{array}{c} 0\; 0\; 1\; 1\; 1\\ 1\; 1\; 0\; 0\; 0\\ 2\; 2\; 2\; 4\; 3\\ 3\; 3\; 4\; 3\; 2\\ 4\; 4\; 3\; 2\; 4 \end{array} \right] $} & \makecell{\\ \\ ((0,2),(0,3),(0,4),(1,2),(1,3),(1,4)).\\ \\ \\ \\} & \makecell{((0,1),(2,3))\;; \\ ((0,1)).} \\ \hline \small{ $\left[ \begin{array}{c} 0\; 0\; 1\; 1\; 1\\ 1\; 1\; 0\; 0\; 0\\ 2\; 2\; 2\; 2\; 2\\ 4\; 4\; 4\; 3\; 3\\ 3\; 3\; 3\; 4\; 4 \end{array} \right] $} & \makecell{\\ ((0,2),(0,3),(0,4),(1,2),(1,3),(1,4))\;;\\ ((0,2),(1,2))\;;\\ ((2,3),(2,4))\;;\\ ((0,3),(0,4),(1,3),(1,4)).\\ \\ } & \makecell{((0,1),(2,3)) \;;\\ ((0,1)).}\\ \hline \small{ $\left[ \begin{array}{c} 0\; 0\; 1\; 1\; 1\\ 1\; 1\; 0\; 0\; 0\\ 3\; 4\; 2\; 4\; 3\\ 4\; 2\; 4\; 3\; 2\\ 2\; 3\; 3\; 2\; 4 \end{array} \right] $} & \makecell{\\ \\ ((0,2),(0,3),(0,4),(1,2),(1,3),(1,4)).\\ \\ \\ \\} & None \\ \hline \end{tabular} \end{center} \end{table} \section*{Acknowledgement} Mohamed Elhamdadi was partially supported by Simons Foundation collaboration grant 712462. \section{Introduction} Quandles are algebraic structures modeled on the \emph{three} Reidemeister moves in classical knot theory. They have been used extensively to construct invariants of knots and links, see for example \cites{EN, Joyce, Matveev}. A Topological quandle is a quandle with a topology such that the quandle binary operation is compatible with the topology. Precisely, the binary operation is continous and the right multiplications are \emph{homeomorphisms}. Topological quandles were introduced in \cite{Rubin} where it was shown that the set of homomorphisms from the fundamental quandle of the knot to a topological quandle (called also the set of colorings) is an invariant of the knot. Equipped with the compact-open topology, the set of colorings is a topological space. In \cite{EM} a foundational account about topological quandles was given. More precisely, the notions of ideals, kernels, units, and inner automorphism group in the context of topological quandle were introduced. Furthermore, modules and quandle group bundles over topological quandles were introduced with the purpose of studying central extensions of topological quandles. Continuous cohomology of topological quandles was introduced in \cite{ESZ} and compared to the algebraic theories. Extensions of topological quandles were studied with respect to continuous 2-cocycles, and used to show differences in second cohomology groups for some specific topological quandles. Nontriviality of continuous cohomology groups for some examples of topological quandles was shown. In in \cite{CES} the problem of classification of topological Alexander quandle structures, up to isomorphism, on the real line and the unit circle was investigated. In \cite{Gr} the author investigated quandle objects internal to groups and topological spaces, extending the well-known classification of quandles internal to abelian groups \cite{Szymik}. In \cite{Tak} quandle modules over quandles endowed with geometric structures were studied. The author also gave an infinitesimal description of certain modules in the case when the quandle is a regular s-manifold (smooth quandle with certain properties). Since any finite $T_1$-space is discrete, the category of finite $T_0$-spaces was considered in \cite{Stong}, where the point set topological properties of finite spaces were investigated. The homeomorphism classification of finite spaces was investigated and some representations of these spaces as certain classes of matrices was obtained. This article arose from a desire to better understand the analogy of the work given in \cite{Stong} in the context of \emph{finite topological} quandles. It turned out that: there is no $T_0$-topology on any finite connected (meaning one orbit under the action of the Inner group) quandle $X$ that makes $X$ into a topological quandle (Theorem~\ref{noT0}). Thus we were lead to consider topologies on quandles with more than \emph{one} orbit. It is well known \cite{Alex} that the category of Alexandroff $T_0$-spaces is equivalent to the category of \emph{partially ordered sets} (posets). In our context, we prove that for a quandle $X$ with more than one orbit, there exists a unique non trivial topology which makes right multiplications of $X$ continuous maps (Proposition~\ref{Prop}). Furthermore, we prove that if $X$ be a finite quandle with two orbits $X_1$ and $X_2$ then any continuous poset on $X$ is biparatite with vertex set $X_1$ and $X_2$ (Proposition~\ref{Bipar}). This article is organized as follows. In Section~\ref{Review} we review the basics of topological quandles. Section~\ref{Poset} reviews some basics of posets, graphs and some hierarchy of separation axioms. In Section~\ref{Main} the main results of the article are given. Section~\ref{Computations} gives some explicit computations based on some computer softwares (Maple and Python) of quandles up to order \emph{five}. \section{Review of Quandles and Topological Quandles}\label{Review} A quandle is a set $X$ with a binary operation $*$ satisfying the following three axioms: \begin{enumerate} \item For all $x$ in $X$, $x*x=x,$ \item For all $y,z \in X$, there exists a unique $x$ such that $x*y=z$, \item For all $x,y,z \in X$, $(x*y)*z=(x*z)*(y*z)$. \end{enumerate} These three conditions come from the axiomatization of the three Reidemeister moves on knot diagrams. The typical examples of quandles are: (i) Any Group $G$ with conjugation $x*y=y^{-1}xy$, is a quandle called the \emph{conjugation quandle} and (ii) Any group $G$ with operation given by $x*y=yx^{-1}y$, is a quandle called the \emph{core quandle}.\\ Let $X$ be a quandle. For an element $y \in X$, left multiplication $L_y$ and right multiplication $R_y$ by an element $y$ are the maps from $X$ to $X$ given respectively by $L_y(x):=y*x$ and $R_y(x)=x*y$. A function $f: (X,*) \rightarrow (X,*)$ is a quandle {\em homomorphism} if for all $x,y \in X, f(x * y)=f(x) * f(y)$. If furthermore $f$ is a bijection then it it is called an \emph{automorphism} of the quandke $X$. We will denote by {\rm Aut(X)} the automorphism group of $X$. The subgroup of {\rm Aut(X)}, generated by the automorphisms $R_x$, is called the {\em inner} automorphism group of $X$ and denoted by {\rm Inn}$(X)$. If the group {\rm Inn}$(X)$ acts \emph{transitively} on $X$, we then say that $X$ is connected quandle meaning it has only one orbit. Since we do not consider topological connectedness in this article, then through the whole article, the word connected quandle will stand for algebraic connectedness. For more on quandles refer to \cite{EN, Joyce, Matveev, E}. Topological quandles have been investigated in \cite{CES, EM, Rubin, ESZ}. Here we review some basics of topological quandles. \begin{definition} A \textit{topological quandle} is a quandle $X$ whith a topology such that the map $X\times X\ni (x,y)\mto x * y\in X$ is a continuous, the right multiplication $R_x:X\ni y\mto y* x\in X$ is a homeomorphism, for all $x\in X$, and $x* x=x$. \end{definition} It is clear that any finite quandle is automatically a topological quandle with respect to the discrete topology. \begin{example} \cite{CES} Let $(G, +)$ be a topological abelian group and let $\sigma$ be a continuous automorphism of $G$. The continuous binary operation on $G$ given by $x*y=\sigma(x)+(Id-\sigma)(y), \forall x,y \in G,$ makes $(G,*)$ a topological quandle called \textit{topological Alexander quandle}. In particular, if $G=\mathbb{R}$ and $\sigma(x)=tx$ for non-zero $t \in \mathbb{R}$, we have the topological Alexander structure on $\mathbb{R}$ given by $x*y=tx+(1-t)y$. \end{example} \begin{example} The following examples were given in \cites{Rubin,EM}. The unit sphere $\mathbb{S}^n \subset \mathbb{R}^{n+1}$ with the binary operation $x\ast y=2(x\cdot y)y-x$ is a topological quandle, where $\cdot$ denotes the inner product of $\mathbb{R}^{n+1}$. Now consider $\lambda$ and $\mu$ be real numbers, and let $x,y \in S^n$. Then $$\lambda x * \mu y=\lambda[2{\mu}^2(x\cdot y)y-x].$$ In particular, the operation $$\pm x * \pm y=\pm (x * y)$$ provides a structure of topological quandle on the quotient space that is the projective space $\mathbb{RP}^n$. \end{example} \section{Review of topologies on finite sets, Posets and Graphs}\label{Poset} Now we review some basics of directed graphs, posets and $T_0$ and $T_1$ topologies. \begin{definition} A \emph{directed} graph G is a pair $(V, E)$ where $V$ is the set of vertices and $E$ is a list of directed line segments called edges between pairs of vertices. \end{definition} An edge from a vertex $x$ to a vertex $y$ will be denoted symbolically by $x < y$ and we will say that $x$ and $y$ are \emph{adjacent}. The following is an example of a directed graph. \begin{example} Let $G=(V,E) $ where $V=\{a,b,c,d\}$ and $E=\{b<a,c<a ,a<d \}$. \end{example} \begin{center} \begin{figure}[h!] \begin{tikzpicture} [nodes={draw, circle}, <-] \node{d} child { node {a} child { node {b} } child { node {c} } } child [ missing ]; \end{tikzpicture} \end{figure} \end{center} \begin{definition} An \emph{idependent set} in a graph is a set of pairwise non-adjacent vertices. \end{definition} \begin{definition} A (directed) graph $G=(V,E)$ is called biparatite if $V$ is the union of two disjoint independent sets $V_1$ and $V_2$. \end{definition} \begin{definition} A (directed) graph $G$ is called \emph{complete biparatite} if $G$ is bipartite and for every $v_1 \in V_1$ and $v_2 \in V_2$ there is an edges in $G$ that joins $v_1$ and $v_2$ \end{definition} \begin{example} Let $V=V_1 \cup V_2$ where $V_1=\{4,5\}$ and $V_2=\{1,2,3\}$. Then the directed graph $G=(V,E)$ is complete biparatite graph. \begin{center} \includegraphics[]{biparatite.png} \end{center} \end{example} Now we recall the definition of partially ordered set. \begin{definition} A partially ordered set (poset) is a set $X$ with an order denoted $\leq$ that is reflexive, antisymmetric and transitive. \end{definition} \begin{example} For any set $X$, the power set of $X$ ordered by the set inclusion relation $\subseteq$ forms a poset $( \mathcal{P}(X),\subseteq)$ \end{example} \begin{definition} Two partially ordered sets $P=(X,\leq)$ and $Q=(X,\leq')$ are said to be isomorphic if there exist a bijection $f:X \rightarrow X'$ such that $x \leq y$ if and only if $f(x)\leq'f(y).$ \end{definition} \begin{definition}\label{ConnPoset} A poset $(X,\leq )$ is connected if for all $x,y\in X$, there exists sequence of elements $x=x_1,x_2, \ldots, x_n=y $ such that every two consecutive elements $x_i$ and $x_{i+1}$ are comparable (meaning $x_i < x_{i+1}$ or $x_{i+1} < x_i$). \end{definition} \noindent {\bf Notation:} Given an order $\leq$ on a set $X$, we will denote $x<y$ whenever $x \neq y$ and $x \leq y$. Finite posets $(X,\leq)$ can be drawn as directed graphs where the vertex set is $X$ and an arrow goes from $x$ to $y$ whenever $x \leq y$. For simplicity, we will not draw loops which correspond to $x \leq x$. We will then use the notation $(X,<)$ instead of $(X,\leq)$ whenever we want to ignore the reflexivity of the partial order \begin{example}. Let $X=\mathbb{Z}_8$ be the set of integers modulo $8$. The map $f:X \rightarrow X$ given by $f(x)=3x-2$ induces an isomorphism between the following two posets $(X,<)$ and $(X,<')$. \begin{center} \includegraphics{example1.png}, \qquad \includegraphics{example2.png} \end{center} \end{example} \begin{definition} A chain in a poset $(X, <)$ is a subset $C$ of $X$ such that the restriction of $<$ to $C$ is a total order (i.e. every two elements are comparable). \end{definition} Now we recall some basics about topological spaces called $T_0$ and $T_1$ spaces. \begin{definition} A topological space $X$ is said to have the property $T_0$ if for every pair of distinct points of $X$, at least one of them has a neighborhood not containing the other point. \end{definition} \begin{definition} A topological space $X$ is said to have the property $T_1$ if for every pair of distinct points of $X$, each point has a neighborhood not containing the other point. \end{definition} Obviously the property $T_1$ implies the property $T_0$. Notice also that this definition is equivalent to saying singletons are closed in $X$. Thus a $T_1$-topology on a \emph{finite} set is a discrete topology. Since any finite $T_1$-space is discrete, we will focus on the category of finite $T_0$-spaces. First we need some notations. Let $X$ be a finite topological space. For any $x \in X$, we denote \[ U_x:=\textit{the smallest open subset of $X$ containing $x$} \] It is well known \cite{Alex} that the category of $T_0$-spaces is isomorphic to the category of posets. We have $x \leq y$ if and only if $U_y \subseteq U_x$ which is equivalent to $C_x \subset C_y$, where $C_v$ is the complelement $U_v^{c}$ of $U_v$ in $X$. Thus one obtain that $U_x=\{w \in X;\; x \leq w\}$ and $C_x=\{v \in X; \;v < x\}$. Under this correspondence of categories, the subcategory of finite posets is equivalent to the category of finite $T_0$-spaces. \\ Through the rest of this article we will use the notation of $x<y$ in the poset whenever $x\neq y$ and $x \leq y$. \begin{comment} $$\begin{array}{|c|c| c| c| c| c|} \hline n \ & Distinct\\& Topologies & Distinct\\&& T_0-Topologies & inequivalent\\&&& topologies & T_0-inequivalent\\&&&& topologies \\ \hline \ 1 & 1 &1 & 1 & 1 \\ \hline \ 2 & 4 & 3 & 3 & 2 \\ \hline \ 3 & 29 & 19 & 9 & 5 \\ \hline \ 4 & 355 & 219 & 33 & 16 \\ \hline \ 5 & 6942 & 4231 & 139 & 63 \\ \hline \ 6 & 209527 & 130023 & 718 & 318 \\ \hline \end{array} $$ \end{comment} \section{Topologies on non- connected Quandles}\label{Main} As we mentioned earlier, since $T_1$-topologies on a finite set are discrete, we will focus in this article on $T_0$-topologies on \emph{finite quandles}. A map on finite spaces is continuous if and only if it preserves the order. It turned out that on a finite quandle with a $T_0$-topology, left multiplications can not be continuous as can be seen in the following theorem \begin{theorem}\label{left} Let $X$ be a finite quandle endowed with a $T_0$-topology. Assume that for all $z \in X$, the map $L_z$ is continuous, then $x \leq y$ implies $L_z(x)=L_z(y)$. \end{theorem} \begin{proof} We prove this theorem by contradiction. Let $X$ be a finite quandle endowed with a $T_0$-topology. Assume that $x \leq y$ and $L_z(x) \neq L_z(y)$. If $x=y$, then obviously $L_z(x)=L_z(y)$. Now assume $x <y$, then for all $a \in X$, the continuity of $L_a$ implies that $a*x \leq a*y$. Assume that there exist $a_1 \in X$ such that, $z_1:=a_1*x=L_{a_1}(x) < a_1*y=L_{a_1}(y)$. The invertibility of right multiplications in a quandle implies that there exist unique $a_2$ such that $a_2*x=a_1*y$ hence $a_1*x<a_2*x$ which implies $a_1 \neq a_2.$ Now we have $a_1*x<a_2*x \leq a_2 *y=z_2$. We claim that $a_2*x < a_2 *y$. if $a_2*y=a_2*x$ and since $a_2*x=a_1*y$ we will have $a_2*y=a_2*x=a_1*y$ hence $a_2*y=a_1*y$ but $a_1 \neq a_2$, thus contradiction. Now that we have proved $a_2*x < a_2 *y$, then there exists $a_3$ such that $a_2*y=a_3*x$ we get, $a_2*x<a_3*x$ repeating the above argument we get, $a_3*x<a_3*y$. Notice that $a_1,a_2$ and $a_3$ are all pairwise disjoint elements of $X$. Similarly, we construct an \emph{infinite} chain, $a_1 * x< a_2*x <a_3 *x < \cdots $, which is impossible since $X$ is a finite quandle. Thus we obtain a contradiction. \end{proof} \begin{comment} Theorem~\ref{left} states that we can not have a notion of \emph{semitopological quandles} (continuity with respect to each variable considered separately) on finite quandles. For more on semitopological groups see \cite{Husain}. Thus through the rest of the paper, the word continuity will mean continuity of \emph{right multiplications} in the finite quandle. We will then say that the quandle with its topology is \emph{right continuous}. We borrow the following terminologies from the book \cite{Rup}. \end{comment} We have the following Corollary \begin{corollary} Let $X$ be a finite quandle endowed with a $T_0$-topology. If $C$ is a chain of $X$ as a poset then any left continuous function $L_x$ on $X$ is a constant function on $C$. \end{corollary} \begin{definition} A quandle with a topology in which right multiplications (respectively left multiplications) are continuous is called \emph{right topological quandle} (respectively \emph{left topological quandle}). \end{definition} In other words, right topological quandle means that for all $x,y,z \in X$, \[ x<y \implies x*z < y*z. \] and, since left multiplications are not necessarly bijective maps, left topological quandle means that for all $x,y,z \in X$, \[ x<y \implies z*x \leq z*y. \] \begin{comment} \begin{definition} A quandle with a topology in which both right multiplications and left multiplications are continuous is called \emph{semitopological quandle}. \end{definition} Obviously topological quandle implies semitopological quandle which implies right topological quandles, but the converses are not true. \end{comment} \begin{theorem}\label{noT0} There is no $T_0$-topology on a finite connected quandle $X$ that makes $X$ into a right topological quandle. \end{theorem} \begin{proof} Let $x <y$. Since $X$ is connected quandle, there exists $\phi \in Inn(X)$ such that $y=\phi(x)$. Since $X$ is finite, $\phi$ has a finite order $m$ in the group $Inn(X)$. Since $\phi$ is a continous automorphism then $x<\phi(x)$ implies $x<\phi^m(x)$ giving a contradiction. \end{proof} \begin{corollary}\label{noodd} There is no $T_0$-topology on any latin quandle that makes it into a right topological quandle. \end{corollary} \begin{comment} \begin{proof} For Dihedral quandle $R_y(x)=2y-x$\\ If n is odd then 2 is invertible in $Z_n$, hence for every\\ $z \in \mathbb(z)_n$ there exist $y=\frac{z+x}{2}$ such that $R_y(x)=z$ hence $R_y(x)$ is both injective in $x$ and $y$ hence quandle will have only one orbit.\\ Hence it is connected. \end{proof} \end{comment} Thus Theorem~\ref{noT0} leads us to consider quandles $X$ that are not connected, that is $X=X_1 \cup X_2\cup \ldots X_k$ as orbit decomposition, search for $T_0$-topology on $X$ and investigate the continuity of the binary operation. \begin{proposition}\label{Prop} Let $X$ be a finite quandle with orbit decomposition $X=X_1 \cup \{a\}$, then there exist unique non trivial $T_0$-topology which makes $X$ right continuous. \end{proposition} \begin{proof} Let $X=X_1 \cup \{a\}$ be the orbit decomposition of the quandle $X$. For any $x,y \in X_1$, there exits $\phi \in Inn(X)$ such that $\phi(x)=y$ and $\phi(a)=a$. Declare that $x<a$, then $\phi(x)<a$. Thus for any $z\in X_1$ we have $z<a$. Uniqueness is obvious. \end{proof} The $T_0$-topology in Proposition~\ref{Prop} is precisely given by $x<a$ for all $x \in X_1.$ \begin{proposition} \label{Bipar} Let $X$ be a finite quandle with two orbits $X_1$ and $X_2$. Then any right continuous poset on $X$ is biparatite with vertex set $X_1$ and $X_2$. \end{proposition} \begin{proof} We prove this proposition by contradiction. For every $x_1, y_1 \in X_1$ such that $x_1 <y_1$. We know that there exist $\phi \in Inn(X)$ such that $\phi(x_1)=y_1$. Hence, $x_1< \phi(x_1)$ implies $x_1 <\phi^m(x_1)=x_1$, where $m$ is the order of $\phi$ in $Inn(X)$. Thus we have a contradiction. \end{proof} \begin{proposition}\label{biComp} Let $X$ be a finite quandle with two orbits $X_1$ and $X_2$. Then the complete bipartite graph with vertex set $X_1$ and $X_2$ forms a right continuous poset. \end{proposition} \begin{proof} Let $X$ be a finite quandle with two orbits $X_1$ and $X_2$. If $x \in X_1$ and $y \in X_2$ then for every $\phi \in Inn(X)$ we have $\phi(x) \in X_1$ and $\phi(y) \in X_2$. Proposition~\ref{Bipar} gives that the graph is bipartite and thus $x<y$. We then obtain $\phi(x) < \phi(y)$ giving the result. \end{proof} \begin{remark} By Proposition~\ref{biComp} and Theorem~\ref{left}, there is a non-trivial $T_0$-topology making $X$ right continuous if and only if the quandle has more than one orbit. \end{remark} Notice that Proposition~\ref{biComp} can be generalized to $n$-paratite complete graph. The following table gives the list of right continuous posets on some even dihedral quandles. In the table, the notation $(a,b)$ on the right column means $a<b.$ \begin{table}[H] \caption{Right continuous posets on dihedral quandles}\label{TTable1} \begin{center} \begin{tabular}{|c|c|} \hline Quandle & Posets \\ \hline $R_4$ & ((0,1),(2,1),(0,3),(2,3)) \\ \hline $R_6$ & ((0, 1), (0, 5), (2, 1), (2, 3), (4, 3), (4, 5))\;; \\ & ((0,3), (2, 5), (4, 1)). \\ \hline $R_8$ & ((2, 7), (4, 7), (6, 1), (6, 3), (0, 5), (2, 5), (4, 1), (0, 3))\;;\\ & (( 0, 1), (6, 7), (4, 5), (0, 7), (2, 1), (2, 3), (4, 3), (6, 5)). \\ \hline $R_{10}$ & ((0, 1), (6, 7), (4, 5), (2, 1), (8, 9), (2, 3), (4, 3), (8, 7), (0, 9), (6, 5)) \;;\\ & ((4, 7), (6, 9), (2, 9), (8, 1), (8, 5), (0, 7), (6, 3), (2, 5), (4, 1), (0, 3))\;;\\ & ((2, 7), (8, 3), (0, 5), (4, 9), (6, 1)).\\ \hline \end{tabular} \end{center} \end{table} Notice that in table~\ref{TTable1}, the dihedral quandle $R_4$ has only one right continuous poset $((0,1),(2,1),(0,3),(2,3))$ which is complete biparatite. While the dihedral quandle $R_6$ has two continuous posets $((0, 1), (0, 5), (2, 1), (2, 3), (4, 3), (4, 5))$ and $((0,3), (2, 5), (4, 1))$ illustrated below. \begin{center} \begin{tikzpicture} [scale=.8,auto=left,every node/.style={circle,fill=gray}] \node (n6) at (2,10) {3}; \node (n4) at (2,7) {0}; \node (n5) at (3,10) {5}; \node (n1) at (3,7) {2}; \node (n2) at (4,10) {1}; \node (n3) at (4,7) {4}; \foreach \from/\to in {n6/n4,n5/n1,n2/n3} \draw (\from) -> (\to); \end{tikzpicture} \qquad \qquad \qquad \begin{tikzpicture} [scale=.8,auto=left,every node/.style={circle,fill=gray}] \node (n6) at (14,10) {3}; \node (n1) at (13,7) {2}; \node (n5) at (13,10) {5}; \node (n4) at (14,7) {0}; \node (n2) at (15,10) {1}; \node (n3) at (15,7) {4}; \foreach \from/\to in {n4/n2,n2/n1, n5/n3,n6/n1,n5/n3,n4/n5, n3/n6} \draw (\from) -- (\to); \end{tikzpicture} \end{center} Moreover, in table~\ref{TTable1}, for $R_8$ the bijection $f$ given by $f(k)=3k-2$ makes the two posets isomorphic. The same bijection gives isomorphism between the first two posets of $R_{10}$. The following Theorem characterizes non complete biparatite posets on dihedral quandles. \begin{theorem} Let $R_{2n}$ be a dihedral quandle of even order. Then $R_{2n}$ has $s+1$ right continuous posets, where $s$ is number of odd natural numbers less than n and relatively non coprime with $n$ \end{theorem} \begin{proof} Let $X=R_{2n}$ be the dihedral quandle with orbits $X_1=\{0,2, \ldots, 2n-2\}$ and $X_2=\{1,3, \ldots ,2n-1\}$. For every $x \in X_2$, we construct a partial order $<_x$ on $R_{2n}$, such that for all $y \in X$, we have $2y<_x2y-x$ and $2y <_x 2y+x$. Then $<_x$ is clearly right continuous partial order since $2y<2y-x$ and $2y<2y+x$ for all $y$ imply that $2z-2y<2z-(2y-x)$. In other words we obtain $2y*z<(2y-x)*z$. From the definition of the order $<_x$ it is clear the two partial orders $<_x$ and $<_{2n-x} $ are the same. Hence we obtain the following distinct partial orders $<_1, <_3, \ldots $. Now we check which ones are isomorphic. If $m$ is odd and $gcd(n,m)=1$ then $f(k)=mk-2$ is a bijective function making $<_1$ and $<_m$ isomorphic. Now let $m$ be odd and $gcd(m,n)=k>1$. The two posets $<_1$ and $<_m$ are non isomorphic since $<_1$ is connected poset, as in Definition~\ref{ConnPoset}, and $<_m$ is not connected poset. We show that these are the only right continuous posets. Given a right continuous poset on $R_{2n}$ then $ a < b$ can be written as $a < a-(a-b)$ which implies that $a <_x b$ where $x=a-b.$ Now if $a < b$ then by Proposition~\ref{Bipar}, we have $a \in X_1$, $b \in X_2$. Now let $a = 2\alpha$ and $b = 2\beta+1$ then $a-b = 2(\alpha-\beta)-1 \in X_2$. This ends the proof. \end{proof} \begin{corollary} For the dihedral quandle $R_{2^n}$ with $2^n$ elements, there is a unique right continuous poset. \end{corollary} \vspace*{-7pt} \section{Some Computer Calculations}\label{Computations} In this section we give non-trivial right and left continuous posets on the finite quandles of order up to $5$ based on Maple and Python computations. In the following tables we have excluded the trivial and connected quandles \begin{table}[H] \caption{Continuous posets on quandles of order $3$} \label{Table2} \begin{center} \begin{tabular}{ |c|c|c|} \hline Quandle for n = 3 & Right continuous Posets & Left continuous poset \\ \hline \small{ $\left[ \begin{array}{c} 0 \;0 \;1 \\ 1 \;1 \;0 \\ 2 \;2\; 2 \end{array} \right] $} & ((0,2),(1,2)) & ((0,1))\\ \hline \end{tabular} \end{center} \end{table} As seen in Table~\ref{Table2} for $n=3$, there exist a unique right continuous poset and a unique left continuous poset. \begin{table}[H] \caption{Continuous posets on quandles of order 4} \label{Table3} \begin{center} \begin{tabular}{ |c|c|c|} \hline Quandles for n = 4 & Right continuous poset & Left continuous poset \\ \hline \small{ $\left[ \begin{array}{c} 0 \;0 \;0 \;0 \\ 1 \;1 \;1 \;2 \\ 2 \;2 \;2 \;1 \\ 3 \;3 \;3 \;3 \end{array} \right] $} & \makecell{ ((0, 3))\;;\\ ((0, 1), (0, 2), (0, 3))\;;\\ ((0, 1), (0, 3), (1, 2))\;;\\ ((0, 1), (0, 2), (1, 3), (2, 3))\;;\\ ((2, 3), (1, 3))\;;\\ ((2, 3), (1, 3), (0, 3)).\\ } & ((0,1),(1,2)) and ((1,2)) \\ \hline \small{ $\left[ \begin{array}{c} 0 \;0 \;0 \;1 \\ 1 \;1\; 1 \;2 \\ 2 \;2 \;2 \;0 \\ 3 \;3 \;3 \;3 \end{array} \right] $} & \makecell{ ((0,3),(1,3),(2,3))\\ } & ((0,1),(1,2)) and ((1,2)) \\ \hline \small{ $\left[ \begin{array}{c} 0 \;0 \;1 \;1 \\ 1 \;1 \;0 \;0 \\ 2 \;2 \;2 \;2 \\ 3 \;3 \;3 \;3 \end{array} \right] $} & \makecell{ ((0,2),(1,2),(0,3),(1,3),(2,3))\;; \\ ((0,2),(1,2),(0,3),(1,3)) \;;\\ ((0,2),(1,2)) \;;\\ ((2,3)).\\ } & ((0,1),(2,3)) and ((2,3))\\ \hline \small{ $\left[ \begin{array}{c} 0 \;0 \;0 \;0 \\ 1 \;1 \;3 \;2 \\ 2 \;3 \;2 \;1 \\ 3 \;2 \;1 \;3 \end{array} \right] $} & \makecell{ ((0,1),(0,2),(0,3)) \\ } & None \\ \hline \small{ $\left[ \begin{array}{c} 0 \;0 \;1 \;1 \\ 1 \;1 \;0 \;0 \\ 3 \;3 \;2 \;2 \\ 2 \;2 \;3 \;3 \end{array} \right] $} & \makecell{ ((0,2),(0,3),(1,2),(1,3)) \\ } & ((0,1),(2,3)) \\ \hline \end{tabular} \end{center} \end{table} \begin{table}[H] \caption{Continuous posets on quandles of order 5, Part I} \label{Table4} \begin{center} \begin{tabular}{ |c|c|c|} \hline Quandles for n = 5& Right continuous& Left continuous \\ \hline \small{ $\left[ \begin{array}{c} 0\; 0\; 0\; 0\; 0 \\ 1\; 1\; 1\; 1\; 1 \\ 2\; 2\; 2\; 2\; 3 \\ 3\; 3\; 3\; 3\; 2 \\ 4\; 4\; 4\; 4\; 4 \end{array} \right] $} & \makecell{\\ ((0,1),(1,2),(1,3),(0,4))\;;\\ ((0,2),(0,3),(1,2),(1,3),(4,2),(4,3))\;;\\ ((0,2),(0,3),(1,2),(1,3),(2,4),(3,4))\;;\\ ((0,1),(1,4),(4,2),(4,3)).\\ \\ } & \makecell{ ((0,1),(1,2),(2,3))\;;\\ ((0,1),(1,2))\;;\\ ((1,2)).} \\ \hline \small{ $\left[ \begin{array}{c} 0\; 0\; 0\; 0\; 0 \\ 1\; 1\; 1\; 1\; 2 \\ 2\; 2\; 2\; 2\; 3 \\ 3\; 3\; 3\; 3\; 1 \\ 4\; 4\; 4\; 4\; 4 \end{array} \right] $} & \makecell{ ((0,1), (0,2), (0,3), (2,4), (3,4), (1,4))\;; \\ ((0,4)) \;;\\ ((0,1), (0,2), (0,3))\;; \\ (( 0,4 ),(4,1 ),(4,2 ),(4,3)). } & \makecell{\\((0,1), (1,2), (2,3)) \;;\\ ((0,1), (1,2)) \;;\\ ((2,3)). \\ } \\ \hline \small{ $\left[ \begin{array}{c} 0\; 0\; 0\; 0\; 1 \\ 1\; 1\; 1\; 1\; 0 \\ 2\; 2\; 2\; 2\; 3 \\ 3\; 3\; 3\; 3\; 2 \\ 4\; 4\; 4\; 4\; 4 \end{array} \right] $} & \makecell{\\ ((1,2),(0,3),(2,4),(3,4))\;;\\ ((1,2),(0,2),(1,3),(0,3),(2,4),(3,4))\;;\\ ((1,4),(0,4))\;;\\ ((1,2),(0,2),(1,3),(0,3)).\\ \\} & \makecell{ ((1,2),(0,1),(2,3))\;; \\ ((0,1),(0,2))\;; \\ ((0,2),(1,2)).\\ \\}\\ \hline \small{ $\left[ \begin{array}{c} 0\; 0\; 0\; 0\; 1 \\ 1\; 1\; 1\; 1\; 2 \\ 2\; 2\; 2\; 2\; 3 \\ 3\; 3\; 3\; 3\; 0 \\ 4\; 4\; 4\; 4\; 4 \end{array} \right] $} & \makecell{((0,4),(1,4),(2,4),(3,4)).} & \makecell{\\ ((1,2),(0,1),(2,3))\;; \\ ((0,1),(0,2))\;; \\ ((0,2),(1,2)).\\ \\ \\ } \\ \hline \small{ $\left[ \begin{array}{c} 0\; 0\; 0\; 0\; 0 \\ 1\; 1\; 1\; 1\; 1 \\ 2\; 2\; 2\; 4\; 3 \\ 3\; 3\; 4\; 3\; 2 \\ 4\; 4\; 3\; 2\; 4 \end{array} \right] $} &\makecell{\\ ((0,2),(0,3),(0,4))\;;\\ ((0,2),(0,3),(0,4),(1,2),(1,3),(1,4))\;;\\ ((0,1))\;;\\ ((0,2),(0,3),(0,4),(0,1),(1,2),(1,3),(1,4)).\\ \\ \\ } & \makecell{((0,1),(0,2),(0,3))\;;\\ ((0,1),(0,2))\;;\\ ((0,1)). } \\ \hline \small{ $\left[ \begin{array}{c} 0\; 0\; 0\; 0\; 0 \\ 1\; 1\; 1\; 2\; 2 \\ 2\; 2\; 2\; 1\; 1 \\ 3\; 3\; 3\; 3\; 3 \\ 4\; 4\; 4\; 4\; 4 \end{array} \right] $} & \makecell{\\ ((1,3),(2,3))\;;\\ ((1,3),(2,3),(1,4),(2,4))\;;\\ ((0,4))\;;\\ ((3,2),(3,1))\;;\\ ((1,3),(2,3),(4,1),(4,2)).\\ \\ } & \makecell{((0,1),(1,2),(3,4))\;; \\ ((0,1),(1,2))\;;\\ ((0,1),(3,4))\;;\\ ((0,1)).\\} \\ \hline \small{ $\left[ \begin{array}{c} 0\; 0\; 0\; 0\; 0 \\ 1\; 1\; 1\; 2\; 2 \\ 2\; 2\; 2\; 1\; 1 \\ 3\; 4\; 4\; 3\; 3 \\ 4\; 3\; 3\; 4\; 4 \end{array} \right] $} & \makecell{((0,1),(0,2),(0,3),(0,4))\;;\\ ((0,1),(0,2))\;;\\ ((0,1),(0,2),(2,3),(2,4),(1,3),(1,4))\;;\\ ((1,3),(1,4),(2,3),(2,4)).} & \makecell{((0,1),(1,2),(3,4))\;; \\ ((0,1),(1,2))\;;\\ ((0,1),(3,4))\;;\\ ((0,1)).\\} \\ \hline \end{tabular} \end{center} \end{table} \begin{table}[H] \caption{Continuous posets on quandles of order 5, Part II} \label{Table1} \begin{center} \begin{tabular}{ |c|c|c|} \hline Quandles for n = 5 & Right continuous & Left continuous \\ \hline \small{ $\left[ \begin{array}{c} 0\; 0\; 0\; 1\; 1 \\ 1\; 1\; 1\; 0\; 0 \\ 2\; 2\; 2\; 2\; 2 \\ 3\; 3\; 4\; 3\; 3 \\ 4\; 4\; 3\; 4\; 4 \end{array} \right] $} & \makecell{ ((0,2), (1,2) (2,3),(2,4))\;;\\ ((0,2), (1,2))\;;\\ ((2,3 ),(2,4))\;;\\ ((0,3),(0,4),(1,3),(1,4)).} & \makecell { ((0,1 ),(1,2 ),(3, 4))\;;\\ ((3,4))\;;\\ ((0,1),(1,2)).\\ } \\ \hline \small{ $\left[ \begin{array}{c} 0\; 0\; 0\; 1\; 1 \\ 1\; 1\; 1\; 2\; 2 \\ 2\; 2\; 2\; 0\; 0 \\ 3\; 3\; 3\; 3\; 3 \\ 4\; 4\; 4\; 4\; 4 \end{array} \right] $} & \makecell{ ((0, 3),(1,3), (2,3), (3,4))\;;\\ ((0,3),(1,3),(2, 3))\;;\\ ((3,4)).} & \makecell{ ((0,1),(1,2),(3,4))\;;\\ ((3,4))\;;\\ ((0,1),(1,2)).} \\ \hline \small{ $\left[ \begin{array}{c} 0\; 0\; 0\; 1\; 2 \\ 1\; 1\; 1\; 2\; 0 \\ 2\; 2\; 2\; 0\; 1 \\ 3\; 3\; 3\; 3\; 3 \\ 4\; 4\; 4\; 4\; 4 \end{array} \right] $} & \makecell{ ((0,3),(1,3),(2,3),(0,4),(1,4),(2,4))\;;\\ ((0,3),(1,3),(2,3))\;;\\ ((3,4)).} & \makecell{ ((0,1),(1,2))\;;\\ ((0,1)).} \\ \hline \small{ $\left[ \begin{array}{c} 0\; 0\; 0\; 0\; 0\\ 1\; 1\; 1\; 1\; 1\\ 2\; 2\; 2\; 2\; 2\\ 4\; 4\; 4\; 3\; 3\\ 3\; 3\; 3\; 4\; 4 \end{array} \right] $} & \makecell{\\ ((0,1),(0,2))\;;\\ ((0,1),(1,3),(1,4))\;;\\ (0,1),(1,2),(2,3),(2,4)).\\ \\ \\ } & \makecell{\\ ((0,1),(0,2))\;;\\ ((0,1),(1,2),(3,4))\;;\\ ((0,1),(3,4))\;;\\ ((3,4)).\\ }\\ \hline \small{ $\left[ \begin{array}{c} 0\; 0\; 0\; 0\; 0\\ 1\; 1\; 1\; 2\; 2\\ 2\; 2\; 2\; 1\; 1\\ 4\; 4\; 4\; 3\; 3\\ 3\; 3\; 3\; 4\; 4 \end{array} \right] $} & \makecell{((0,1),(0,2))\;;\\ ((0,1),(1,3),(1,4))\;;\\ ((1,3),(1,4),(2,3),(2,4)).} & \makecell{\\ ((0,1),(0,2))\;;\\ ((0,1),(1,2),(3,4))\;;\\ ((0,1),(3,4))\;;\\ ((3,4)).\\ \\ }\\ \hline \small{ $\left[ \begin{array}{c} 0\; 0\; 0\; 1\; 1\\ 1\; 1\; 1\; 2\; 2\\ 2\; 2\; 2\; 0\; 0\\ 4\; 4\; 4\; 3\; 3\\ 3\; 3\; 3\; 4\; 4 \end{array} \right] $} & \makecell{((0,4),(1,4),(2,4),(0,3),(1,3),(2,3)).} & \makecell{\\ ((0,1),(0,2))\;;\\ ((0,1),(1,2),(3,4))\;;\\ ((0,1),(3,4))\;;\\ ((3,4)).\\ \\ } \\ \hline \small{ $\left[ \begin{array}{c} 0\; 0\; 0\; 0\; 0\\ 1\; 1\; 4\; 2\; 3\\ 2\; 3\; 2\; 4\; 1\\ 3\; 4\; 1\; 3\; 2\\ 4\; 2\; 3\; 1\; 4 \end{array} \right] $} & \makecell{\\ \\ ((0,1),(0,2),(0,3),(0,4)).\\ \\ \\ } & None \\ \hline \small{ $\left[ \begin{array}{c} 0\; 0\; 1\; 1\; 1 \\ 1\; 1\; 0\; 0\; 0 \\ 2\; 2\; 2\; 2\; 3 \\ 3\; 3\; 3\; 3\; 2 \\ 4\; 4\; 4\; 4\; 4 \end{array} \right] $} & \makecell{\\ ((1,2),(0,2),(1,3),(0,3),(2,4),(3,4))\;;\\ ((1,4),(0,4))\;;\\ ((1,2),(0,2),(1,3),(0,3)). \\ \\ \\ } & \makecell{((0,1),(2,3))\;; \\ ((0,1)).}\\ \hline \end{tabular} \end{center} \end{table} \begin{table}[H] \caption{Continuous posets on quandles of order 5, Part III} \label{Table1} \begin{center} \begin{tabular}{ |c|c|c|} \hline Quandles for n = 5 & Right continuous & Left continuous \\ \hline \small{ $\left[ \begin{array}{c} 0\; 0\; 1\; 1\; 1\\ 1\; 1\; 0\; 0\; 0\\ 2\; 2\; 2\; 4\; 3\\ 3\; 3\; 4\; 3\; 2\\ 4\; 4\; 3\; 2\; 4 \end{array} \right] $} & \makecell{\\ \\ ((0,2),(0,3),(0,4),(1,2),(1,3),(1,4)).\\ \\ \\ \\} & \makecell{((0,1),(2,3))\;; \\ ((0,1)).} \\ \hline \small{ $\left[ \begin{array}{c} 0\; 0\; 1\; 1\; 1\\ 1\; 1\; 0\; 0\; 0\\ 2\; 2\; 2\; 2\; 2\\ 4\; 4\; 4\; 3\; 3\\ 3\; 3\; 3\; 4\; 4 \end{array} \right] $} & \makecell{\\ ((0,2),(0,3),(0,4),(1,2),(1,3),(1,4))\;;\\ ((0,2),(1,2))\;;\\ ((2,3),(2,4))\;;\\ ((0,3),(0,4),(1,3),(1,4)).\\ \\ } & \makecell{((0,1),(2,3)) \;;\\ ((0,1)).}\\ \hline \small{ $\left[ \begin{array}{c} 0\; 0\; 1\; 1\; 1\\ 1\; 1\; 0\; 0\; 0\\ 3\; 4\; 2\; 4\; 3\\ 4\; 2\; 4\; 3\; 2\\ 2\; 3\; 3\; 2\; 4 \end{array} \right] $} & \makecell{\\ \\ ((0,2),(0,3),(0,4),(1,2),(1,3),(1,4)).\\ \\ \\ \\} & None \\ \hline \end{tabular} \end{center} \end{table} \section*{Acknowledgement} Mohamed Elhamdadi was partially supported by Simons Foundation collaboration grant 712462.

5d748d2bfffb94fb2c94053afc465927f4919fe5

\section{Introduction} \label{S:intro} A decomposition of a graph $G$ is a set $\mathcal{H}=\{H_{1}, H_{2}, \ldots, H_{k}\}$ of subgraphs of $G$ such that $\bigcup_{i=1}^{k} E\left(H_{i}\right)=E(G)$ and $E(H_{i}) \cap E(H_{j})=\emptyset$ for $ i \neq j$. Such a decomposition is called an $\{H_{1}, H_{2}, \ldots, H_{k}\}$-decomposition of $G$. A factor in a graph $G$ is a spanning (not necessarily connected) subgraph of $G$. If a graph $G$ can be decomposed into $r_{i}$ factors isomorphic to the factor $F_{i}$ for $i\in [1, t]$, then we say that $G$ has a $\left\{F_{1}^{r_{1}}, F_{2}^{r_{2}}, \ldots, F_{t}^{r_{t}}\right\}$-factorization. When each $F_{i}$ factor consists of only $n_{i}$ cycles for $i\in [1, t]$, then we will call the $F_{i}$ factor as a $C_{n_{i}}$-factor and call this factorization as a $\left\{C_{n_{1}}^{r_{1}}, C_{n_{2}}^{r_{2}}, \ldots, C_{n_{t}}^{r_{t}}\right\}$-factorization where each $r_i$ is the number of $C_{n_{i}}$-factors. Graph factorizations constitute an important part of graph decomposition problems, especially when each factor is of regular degree. A $k$-regular spanning subgraph of $G$ is called a $k$-factor of $G$. It is easy to see that a 1-factor is a perfect matching in a graph and a 2-factor is either an Hamilton cycle or union of cycles. When it comes to 2-factorizations, there are two well-known graph factorization problems. One problem is the Oberwolfach Problem, which is posed by Ringel (see \cite{Ringel1971}) as a seating arrangement problem at a meeting in Oberwolfach. Given a conference venue with $k_{i}$ round tables, each of which has $m_{i}$ seats for $i \in [1, t]$, it asks whether it is possible that each participant of the conference (say $v$ many for odd $v$) sits next to (left or right) each other participant exactly once at the end of $\frac{v-1}{2}$ nights. In graph theory language, it asks whether the complete graph $K_v$ (or $K_v-I$ in the spouse avoiding version for even $v$) decomposes into isomorphic $2$-factors where each $2$-factor consists of $k_{i}$ $m_{i}$-cycles for each $i \in [1, t]$. This problem is denoted by OP$(m_{1}^{k_{1}}, m_{2}^{k_{2}}, \ldots, m_{t}^{k_{t}})$. If there is only one type of cycle, say of length $m$, in the factor, it can be denoted as OP$(m^{k})$, and its solution gives a $\{C_{m}^{\frac{v-1}{2}}\}$-factorization (or in short, a $C_m$-factorization) of $K_v$. The Hamilton-Waterloo Problem is a generalization of the Oberwolfach Problem where there are two conference venues (one in Hamilton and one in Waterloo as one may guess) with different seating arrangements. This time each $2$-factor can be isomorphic to one of the given two 2-factors, say $F_1$ or $F_2$. If $F_1$ consists of only $m$-cycles and $F_2$ consists of only $n$-cycles, then the corresponding Hamilton-Waterloo Problem is called as the uniform version, and it is denoted by HWP$(v; m^{r}, n^{s})$ where $r$ and $s$ are the number of $C_m$ and $C_n$-factors where $r+s= \frac{v-1}{2}$, respectively. Having a solution to HWP$(v; m^{r}, n^{s})$ means that $K_v$ has a $\left\{C_{m}^{r}, C_{n}^{s}\right\}$-factorization for all possible $r$ and $s$ in the range. The uniform versions of both problems are well-studied. In articles \cite{Alspach1985,Alspach1989,Hoffman1991}, authors solved completely the uniform version of Oberwolfach Problem. But the general case of the Oberwolfach Problem is still open. It is known that OP$\left(3^{2}\right)$, OP$\left(3^{4}\right)$, OP$(4,5)$ and OP$\left(3^{2}, 5\right)$ have no solution. In \cite{Adams2006,Deza2010}, it is shown that OP$(m_{1}^{k_{1}}, m_{2}^{k_{2}}, \ldots, m_{t}^{k_{t}})$ has a solution for all $n \leq 40$ with the above exceptions. As the first results on the uniform Hamilton-Waterloo Problem, Adams et al.\cite{Adams2002} showed that HWP$(v; m^{r},n^{s})$ has a solution for all $v\leq 16$ and gave solutions for the small cases where $(m, n)\in \{(4,6),(4,8)$,$(4,16),(8,16),(3,5),$ $(3,15),(5,15)\}$. Cycle sizes $(3,4)$ and in general $(4,m)$ for odd $m$ has been studied by several authors (see \cite{Bonvicini2017}, \cite{Danziger2009}, \cite{Keranen}, \cite{Odabas2016}). When $m$ and $n$ are odd, problem is almost completely solved in \cite{Burgess2017, Burgess2018} for odd $v$. In \cite{D.Bryant2011}, the problem is solved in the case of both $m$ and $n$ are even and $v \equiv 0 $ $(\bmod 4)$ except possibly when $r=s=1$. When $m$ and $n$ are both even and $v \equiv 2$ $(\bmod 4)$, this problem is solved by R. Haggkvist in \cite{Haggkvist1985} whenever $r$ and $s$ are both even. Also, if $m$ is even and $m\vert n$, the problem is completely solved in \cite{D.Bryant2013}. One generalization of these problems may be to consider sitting on the right and sitting on the left of a participant as separate entities. To represent such a sitting, one has to use directed cycles which led us to work on directed graphs. There are studies on the directed Oberwolfach Problem, and here we work on the directed version of the Hamilton-Waterloo Problem. We will denote a digraph $D$ as $D=(V(D),E(D))$, where $V(D)$ is the vertex set and $E(D)$ is the arc set. For clarity, edges and arcs are denoted by using curly braces and parentheses, respectively. For a simple graph $G$, we use $G^*$ to denote symmetric digraph with vertex set $V(G^*)=V(G)$ and arc set $E(G^*)=\bigcup_{\{x,y\}\in E(G)} \{(x,y),(y,x)\}$. Hence, $K_v^*$ and $K_{(x:y)}^*$ respectively denote the complete symmetric digraph of order $v$ and the complete symmetric equipartite digraph with $y$ parts of size $x$. Also, $\vv{C}_n$ will denote the directed cycle of order $n$. Similarly, a set $\{H_{1}, H_{2}, \ldots, H_{k}\}$ of arc-disjoint subdigraphs of a digraph $D$ is called a decomposition of $D$ if $\bigcup_{i=1}^{k} E\left(H_{i}\right)=E(D)$. If a symmetric digraph $G^*$ has decomposition which consists of $r_{i}$ factors having directed $n_{i}$ cycles for $i\in [1, t]$, then we say $G^*$ has a $\left\{\vv{C}_{n_{1}}^{r_{1}}, \vv{C}_{n_{2}}^{r_{2}}, \ldots, \vv{C}_{n_{t}}^{r_{t}}\right\}$-factorization. In the Directed versions of the Oberwolfach and the Hamilton-Waterloo Problems, $K_v^*$ is decomposed into factors of directed cycles. Hence, the seating arrangement is done over $v-1$ nights. If the sizes of directed cycles are $m_{1}, m_{2}, \ldots, m_{t}$ and the number of each directed cycle $m_i$ is $k_i$ for $i\in [1, t]$ where $\sum_{i=1}^{t} k_i m_{i}=v$, the Directed Oberwolfach Problem is denoted by OP$^{*}(m_{1}^{k_1}, \ldots, m_{t}^{k_t})$. Similarly, HWP$^{*}(v; m^{r}, n^{s})$ denotes the uniform directed Hamilton-Waterloo Problem with directed cycle sizes $m$ and $n$. Again, if HWP$^{*}(v; m^{r}, n^{s})$ has a solution, it means that $K_{v}^{*}$ has a $\{\vv{C}_{m}^{r}, \vv{C}_{n}^{s}\}$-factorization for all $r$ and $s$ with $r+s = v-1$. So far, the Directed Oberwolfach Problem has only partial results, but the Directed Hamilton-Waterloo Problem has not been studied yet up to our knowledge. As the first result on the Directed Oberwolfach Problem, OP$^{*}(3^k)$ with an exception $v= 6$ is solved by Bermond et al. \cite{Bermond1979}. In \cite{Bennett1990}, Bennett and Zhang solved OP$^{*}(4^k)$ except for $v=12$, and Adams and Bryant solved the remaining case OP$^{*}(4^3)$ (in an unpublished paper ``Resolvable directed cycle systems of all indices for cycle length 3 and 4''). In \cite{Alspach2003}, Alspach et al. showed that $K_{v}^{*}$ can be decomposed into $\vv{C}_{m}$ cycles with exceptions $(v, m) \neq(4,4),(6,3),(6,6)$ if and only if $m\vert v(v-1)$. They studied the problem in cases where $v$ and $m$ are even or odd, separately. Burgess and Sajna \cite{Sajna2014} investigated the necessary and sufficient conditions for the Directed Oberwolfach Problem with cycles of length $m$. In case $m$ is even, they obtained complete solution and presented a partial solution for odd cycle size. Also, they conjectured that $K_{2 m}^{*}$ admits a directed $m$-cycle factorization for odd $m$ if and only if $m \geq 5$. In \cite{Sajna2018}, Burgess et al. proved this conjecture for $m \leq 49$. The following theorem summarizes the results of Bermond et al. and Burgess and Sajna. \begin{theorem}\cite{Bermond1979, Sajna2014}\label{OP} Let $m$ be even, or $m$ and $k$ be odd integers with $m \geq 3$. Then $\mathrm{OP}^{*}(m^k)$ has a solution if and only if $(m, k) \neq (6,1)$. \end{theorem} In \cite{Sajna2020}, Shabani and Sajna proved that $K_{v}^{*}$ has a $\{\vv{C}_{2}, \vv{C}_{v-2}\}$-factorization for $v \geq 5$ and obtained the necessary and sufficient conditions for $K_{v}^{*}$ to admit a $\{\vv{C}_{m}, \vv{C}_{v-m}\}$-factorization for $2 \leq m \leq v-2$ and for odd $v$. Also they showed that if $v \geq 5$ and $v \equiv 1,3, \text{or} \, 7 \pmod 8$, then $K_{v}^{*}$ has a $\{\vv{C}_{2}, \vv{C}_{2}, \ldots, \vv{C}_{2}, \vv{C}_{3}\}$-factorization. In this paper, we follow the lead of the first results on the undirected Hamilton-Waterloo Problem and give solutions to the cases with directed cycle sizes $\{(4,6),(4,8),(4,12),\allowbreak(4,16),(6,12),(8,16),(3,5),(3,15),(5,15)\}$. We first give the necessary conditions for a solution to HWP$^{*}(v;m^{r}, n^{s})$ to exist. Second, we make the observation that for any given solution to HWP$(v; m^{r}, n^{s})$, one can construct a solution to HWP$^{*}(v; m^{2r}, n^{2s})$ for odd $v$. Then, we give two different constructions depending on the parity of the cycle sizes. For even cycle sizes, using our construction in Lemma \ref{mainlemma} and the preliminary Lemmata required in the construction, HWP$^{*}(v; m^{r}, n^{s})$ is solved for $(m, n)\in \{(4,6),(4,8),(4,12),\allowbreak(4,16),(6,12),(8,16)\}$ with $r+s=v-1$. For odd cycle sizes, we give new a construction in Lemma \ref{oddmainlemma} when $v$ is odd. Using this construction and the results required for this construction, we state that HWP$^{*}(v; m^{r}, n^{s})$ has a solution for $(m, n)\in \{(3,5),(3,15),(5,15)\}$ for odd $v$. Constructions given in Lemma \ref{mainlemma} and Lemma \ref{oddmainlemma} are general constructions and they can be used to solve the problem also for the other cycle sizes as long as the necessary small cases can be found. Let's first start with the necessary conditions and move to the preliminary results then. \begin{lemma}\label{necessary} If $\mathrm{HWP}^{*}(v; m^{r}, n^{s})$ has a solution then following statements hold, \begin{enumerate} \item if $r>0$, $v \equiv 0 \pmod m$, \item if $s>0$, $v \equiv 0 \pmod n$, \item $r+s=v-1$. \end{enumerate} \end{lemma} \section{Preliminary Results} If $G_{1}$ and $G_{2}$ are two edge disjoint graphs with $V(G_1)=V(G_2)$, then we use $G_{1} \oplus G_{2}$ to denote the graph on the same vertex set with $E\left(G_{1} \oplus G_{2}\right)=E\left(G_{1}\right) \cup E\left(G_{2}\right)$. We will denote the vertex disjoint union of $\alpha$ copies of $G$ by $\alpha G$. Finally, $\overline{K}_{n}$ denotes the empty graph on $n$ vertices. Let $G$ and $H$ be graphs, the wreath product of $G$ and $H$, denoted by $G \wr H$, is the graph obtained by replacing each vertex $x$ of $G$ with a copy of $H$, say $H_x$, and replacing each edge $\{x, y\}$ of $G$ with the edges joining every vertex of $H_{x}$ to every vertex of $H_{y}$. In case $G$ and $H$ are both digraphs, then the $G \wr H$ is the digraph obtained by replacing each vertex $x$ of $G$ with a copy of $H$, say $H_x$, and replacing each arc $(x, y)$ of $G$ by an arc pointing from every vertex of $H_{x}$ to every vertex of $H_{y}$. For example, $K_x^* \wr \overline{K}_{y} \cong K_{(y:x)}^*$, $\overline{K}_{x}\wr K_y^* \cong xK_y^*$ and $\overline{K}_{x} \wr \overline{K}_{y} \cong \overline{K}_{xy}$. If $G$ has a $\{H_1, H_2, \dots, H_k\}$-decomposition, then $G\wr \overline{K}_{n}$ has a $\{H_1\wr \overline{K}_{n}, H_2\wr \overline{K}_{n}, \dots, H_k\wr \overline{K}_{n}\}$-decomposition (see \cite{Alspach1989}). Also, for given three graphs $G$, $H$, and $J$, $(G\wr H) \wr J=G\wr (H \wr J)$, that is, the wreath product is associative (see \textit{p.} 185 of \cite{ProductGraphs}). Note that, the above properties of the wreath product extend to digraphs. Let $A$ be a finite additive group and let $S$ be a subset of $A$, where $S$ does not contain the identity of $A$. The Directed Cayley graph $\vv{X}(A ; S)$ on $A$ with connection set $S$ is digraph with $V(\vv{X}(A ; S))=A$ and $E(\vv{X}(A ; S))=\{(x,y):x,y\in A, y-x\in S\}$. The following observation is useful to reduce the number of cases when $v$ is odd. \begin{observation}\label{obs} If $\mathrm{HWP}(v; m^{r}, n^{s})$ has a solution for some $r$ and $s$ and $v$ is odd, then $\mathrm{HWP}^{*}(v; m^{2r},$ $ n^{2s})$ has a solution for the same $r$ and $s$. \end{observation} A solution for HWP$^{*}(v; m^{2r}, n^{2s})$ is obtained from a solution of HWP$(v;\allowbreak m^{r}, n^{s})$ by taking two copies of each 2-factor and replacing each edge $\{x, y\}$ with the arcs $(x, y)$ and $(y, x)$ in the two 2-factors. Similarly, we get an $H^{*}$-factorization of $G^{*}$ from an $H$-factorization of $G$. \begin{lemma}\label{lemma1.3} Let $G$ be a graph and $H$ be a subgraph of $G$. If $G$ has an $H$-factorization then, $G^{*}$ has an $H^{*}$-factorization. \end{lemma} The following lemma and theorem will be used in the solutions of even and odd cases of HWP$^{*}(v; m^{r}, n^{s})$, respectively. \begin{lemma}\cite{Sajna2014} \label{lemma2} Let $m \geq 4$ be an even integer and $x$ be a positive integer. Then $K_{(\frac{mx}{2}:2)}^{*}$ has a $\vv{C}_{m}$-factorization. \end{lemma} \begin{theorem}\cite{Liu2003} \label{liu} The complete equipartite graph $K_{(x: y)}$ has a $C_{m}$-fac\-tor\-i\-za\-tion for $m \geq 3$ and $x \geq 2$ if and only if $m\vert xy$, $x(y-1)$ is even, $m$ is even if $y=2$ and $(x, y, m) \neq(2,3,3),(6,3,3),(2,6,3),(6,2,6)$. \end{theorem} \section{Even Cycle Sizes} We will make use of the following lemma in the first main construction of this paper. \begin{lemma}\label{K_xWrK2} $K_{x}^{*} \wr \overline{K}_{2}$ has a $K_{2}^{*}$-factorization for every integer $x\geq 2$. \end{lemma} \begin{proof} Notice that $K_{x}^{*} \wr \overline{K}_{2}\cong K_{2x}^{*} -xK_{2}^{*}$. Using Kotzig's $1$-factorization of $K_{2x}$ and Lemma \ref{lemma1.3}, a decomposition of $K_{x}^{*} \wr \overline{K}_{2}$ into $2x-2$ $K_{2}^{*}$-factors is obtained. \end{proof} Here we give the main construction that is used to obtain solutions for the even cycle size cases. \begin{lemma}\label{mainlemma} Let $m\geq 4$ and $n\geq 4$ be even and $h=lcm(m,n)$. If $\mathrm{HWP}^{*}(h; m^{r'}, n^{s'})$ has a solution for all nonnegative integers $r'$, $s'$ satisfying $r'+s'=h-1$, then there is a solution to $\mathrm{HWP}^{*}(hx; m^r, n^s)$ for all nonnegative integers $r$, $s$, and $x$ with $r+s=hx-1$. \end{lemma} \begin{proof} We can decompose $K_{hx}^*$ as follows: \begin{eqnarray}\label{3.1} K_{hx}^{*} & \cong & xK_h^{*}\oplus \left(K_{x}^{*} \wr \overline{K}_{h}\right) \label{eq1} \end{eqnarray} Since $\overline{K}_{h}\cong \overline{K}_{2}\wr \overline{K}_{\frac{h}{2}}$, $K_{x}^{*} \wr \overline{K}_{h}$ is isomorphic to $(K_{x}^{*} \wr \overline{K}_{2})\wr \overline{K}_{\frac{h}{2}}$ by the associativity of the wreath product. Thus, by Lemma \ref{K_xWrK2}, $K_{x}^{*} \wr \overline{K}_{h}$ can be decomposed into factors each isomorphic to $K_{2}^{*} \wr \overline{K}_{\frac{h}{2}}$, and since $K_{2}^{*} \wr \overline{K}_{\frac{h}{2}}\cong K_{(\frac{h}{2}:2)}^{*}$, we have a decomposition of $K_{x}^{*} \wr \overline{K}_{h}$ into $2x-2$ $K_{(\frac{h}{2}:2)}^{*}$-factors. Now, let $F_0$ be the $K_{h}^{*}$-factor and $F_1,F_2,\dots,F_{2x-2}$ be the $K_{(\frac{h}{2}:2)}^*$-factors of $K_{hx}^{*}$. Since HWP$^{*}(h; m^{r'},n^{s'})$ is assumed to have a solution for all nonnegative integers $r'$ and $s'$, $F_0$ has a $\{\vv{C}_{m}^{r'},\vv{C}_{n}^{s'}\}$-factorization for all nonnegative integers $r'$ and $s'$ where $r'+s'=h-1$. Also, by Lemma \ref{lemma2} $K_{(\frac{h}{2}:2)}^*$ has a $\vv{C}_{m}$- and a $\vv{C}_{n}$-factorization for $m, n\geq 4$, so each $F_j$ has a $\{\vv{C}_{m}^{\frac{h}{2}r_j},\vv{C}_{n}^{\frac{h}{2}s_j}\}$-factorization for $j\in \{1,2, \dots, 2x-2\}$, where $r_j, s_j \in \{0, 1\}$ with $r_j+s_j=1$. Those factorizations give us a $\{\vv{C}_{m}^{r},\vv{C}_{n}^{s}\}$-factorization of $K_{hx}^{*}$ where $r=r'+\frac{h}{2}\sum_{j=1}^{2x-2} r_j$ and $s=s'+\frac{h}{2}\sum_{j=1}^{2x-2} s_j$ with $r+s=r'+s'+\frac{h}{2}\sum_{j=1}^{2x-2} (r_j+ s_j)=h-1+\frac{h}{2}(2x-2)=hx-1$. Since any nonnegative integer $0\leq r\leq hx-1$ can be written as $r=r'+\frac{h}{2}a$ for integers $0\leq r'\leq h-1$, $0\leq a\leq 2x-2$ and even $h$, a solution to $\mathrm{HWP}^{*}(hx; m^r, n^s)$ exists for each $r \geq0$ and $s \geq0$ satisfying $r+s=hx-1$. \end{proof} \begin{lemma}\label{new2} For every even integer $m\geq2$, $\vv{C}_{m} \wr \overline{K}_{2}$ has a $\vv{C}_{m}$- or $\vv{C}_{2m}$-factorization. \end{lemma} \begin{proof} Let $m \geq 2$ be an integer. We can represent $\vv{C}_{m} \wr \overline{K}_{2}$ as $\vv{X}\big(\mathbb{Z}_2\mathbb{\times Z}_m ; S_1\big)$, the directed Cayley graph over $\mathbb{Z}_2\mathbb{\times Z}_m$ with the connection set $S_1=\{(0,1),$ $(1,1)\}$. Let $\vv{C}_{(1)}=(v_0,v_1, \dots,\allowbreak v_{m-1})$ be a cycle of $\vv{C}_{m} \wr \overline{K}_{2}$, where $v_i=(0,i)$ for $0\leq i\leq m-1$, and it can be checked that $F_1=\vv{C}_{(1)}\cup (\vv{C}_{(1)}+(1,0))$ is a directed $m$-cycle factor of $\vv{C}_{m} \wr \overline{K}_{2}$. Also, let $\vv{C}_{(2)}=(u_0,u_1, \dots,u_{m-1})$ be a cycle of $\vv{C}_{m} \wr \overline{K}_{2}$, where $$ u_{i} = \begin{cases} (0,i) & if \,\ i \,\ is \,\ even\\ (1,i) & if \,\ i \,\ is \,\ odd \end{cases}$$ for $0\leq i\leq m-1$. It can be checked that $F_2=\vv{C}_{(2)}\cup (\vv{C}_{(2)}+(1,0))$ is a directed $m$-cycle factor of $\vv{C}_{m} \wr \overline{K}_{2}$. $F_1$ and $F_2$ are arc disjoint directed $m$-cycle factors of $\vv{C}_{m} \wr \overline{K}_{2}$. Thus $\{F_1,F_2\}$ is a $\vv{C}_{m}$-factorization of $\vv{C}_{m} \wr \overline{K}_{2}$. Let $\vv{C}_{(3)}=(v_0,v_1, \dots,\allowbreak v_{2m-1})$ be a cycle of $\vv{C}_{m} \wr \overline{K}_{2}$, where $$ v_i = \begin{cases} (0,i) & if \,\ 0\leq i\leq m-1\\ (1,i) & if \,\ m-2\leq i\leq 2m-1 \end{cases}$$ and it can be checked that $F_3=\vv{C}_{(3)}$ and $F_4=\vv{C}_{(3)}+(1,0)$ are arc disjoint directed $2m$-cycle factor of $\vv{C}_{m} \wr \overline{K}_{2}$. Thus $\{F_3,F_4\}$ is a $\vv{C}_{2m}$-factorization of $\vv{C}_{m} \wr \overline{K}_{2}$. \end{proof} For $m\geq 2$, we can represent $(\vv{C}_{m} \wr \overline{K}_{2}) \oplus mK_{2}^{*}$ as the directed Cayley graph over $\mathbb{Z}_2\mathbb{\times Z}_m$ with the connection set $S_2=\{(0,1),(1,0),(1,1)\}$ where $K_{2}^{*}$ consists of edges between $(0,i)$ and $(1,i)$ for $0\leq i \leq m-1$. For brevity, we will denote $(\vv{C}_{m} \wr \overline{K}_{2}) \oplus mK_{2}^{*}$ by $\Gamma_m$. \begin{lemma}\label{new} For every integer $m\geq2$, $\Gamma_m$ has a $\{\vv{C}_{m}^{1}, \vv{C}_{2m}^{2}\}$-factorization. \end{lemma} \begin{proof} Let $\vv{C}_{(1)}=(v_0,v_1, \dots,\allowbreak v_{m-1})$ be a cycle of $\Gamma_m$, where $v_i=(0,i)$ for $0\leq i\leq m-1$, and it can be checked that $F_1=\vv{C}_{(1)}\cup (\vv{C}_{(1)}+(1,0))$ is a directed $m$-cycle factor of $\Gamma_m$. Also, let $\vv{C}_{(2)}=(u_0,u_1, \dots,u_{2m-1})$ be a cycle of $\Gamma_m$, where $u_{2i} = (0,i)$, and $u_{2i+1} = (1,i)$ for $0\leq i\leq m-1$. Similarly, it can be checked that $F_2=\vv{C}_{(2)}$ and $F_3=\vv{C}_{(2)}+(1,0)$ are arc disjoint directed $2m$-cycle factors of $\Gamma_m$. Thus $\{F_1,F_2,F_3\}$ is a $\{\vv{C}_{m}^{1}, \vv{C}_{2m}^{2}\}$-factorization of $\Gamma_m$. \end{proof} Following Lemmata give the base blocks of our main construction. The cases when $r=0$ and $s=0$ of the Lemmata are obtained by Theorem \ref{OP} and the remaining factorizations for Lemma \ref{lemma4.8} and \ref{lemma3} are given in the Appendix. \begin{lemma}\label{lemma4.8} For nonnegative integers $r$ and $s$, $\mathrm{HWP}^{*}(8; 4^{r}, 8^{s})$ has a solution if and only if $r+s=7$. \end{lemma} \begin{lemma}\label{lemma3} For nonnegative integers $r$ and $s$, $\mathrm{HWP}^{*}(12; m^{r}, n^{s})$ has a solution for $(m,n)\in \{(4,6),(4,12)$, $(6,12)\}$ if and only if $r+s=11$. \end{lemma} \begin{lemma}\label{lemma6} For nonnegative integers $r$ and $s$, $\mathrm{HWP}^{*}(16; m^{r}, n^{s})$ has a solution for $(m,n)\in \{(4,16),(8,16)\}$ if and only if $r+s=15$. \end{lemma} \begin{proof} By Theorem \ref{OP}, the cases when $r=0$ and $s=0$ are obtained.\\ \textbf{Case 1 :} $(m,n)=(8,16)$: We will first analyse when $r$ is odd. We have that $K_{16}^{*} \cong (K_{8}^{*}\wr \overline{K}_{2})\oplus 8K_{2}^{*}$ by \eqref{3.1}, and $K_{8}^{*}$ have a $\vv{C}_{8}$-factorization by Lemma \ref{lemma4.8}. Then, we have a factorization of $K_{16}^{*}$ into six $\vv{C}_{8} \wr \overline{K}_{2}$ and a single $\Gamma_8$ factor. Also, each $\vv{C}_{8}\wr \overline{K}_{2}$ can be decomposed into two $\vv{C}_{8}$ or two $\vv{C}_{16}$-factors by Lemma \ref{new2}. By Lemma \ref{new}, $\Gamma_8$ has a $\{\vv{C}_{8}^{1}, \vv{C}_{16}^{2}\}$-factorization. Now, let $r_0$ and $s_0$ be nonnegative integers with $r_0+s_0=6$. Decomposing $r_0$ many $\vv{C}_{8}\wr \overline{K}_{2}$'s into $\vv{C}_{8}$-factors and remaining $s_0$ many $\vv{C}_{8}\wr \overline{K}_{2}$'s into $\vv{C}_{16}$-factors, as well as $\Gamma_8$ into a $\{\vv{C}_{8}^{1}, \vv{C}_{16}^{2}\}$-factor gives us a $\{\vv{C}_{8}^{2r_0+1}, \vv{C}_{16}^{2s_0+2}\}$-factorization of $K_{16}^{*}$. Since any odd integer $r$ can be written as $r=2r_0+1$ for nonnegative integer $r_0$, $\mathrm{HWP}^{*}(16; 8^{r}, 16^{s})$ has a solution for odd $r$ with $r+s=2r_0+1+2s_0+2=2(r_0+s_0)+3=15$. We list the solutions to the remaining even cases in the Appendix.\\ \textbf{Case 2 :} For $(m,n)=(4,16)$, solutions to all cases are given in the Appendix except for $r=0$ and $s=0$. \end{proof} \begin{theorem} For nonnegative integers $r$ and $s$, $\mathrm{HWP}^{*}(v; m^{r}, n^{s})$ has a solution for $(m,n)\in \{(4,6),(4,8),$ $(4,12),(4,16),(6,12),(8,16)\}$ if and only if $r+s=v-1$ and $lcm(m,n)\vert v$. \end{theorem} \begin{proof} If a solution to HWP$^{*}(v; m^{r}, n^{s})$ exists for $(m,n)\in \{(4,6),(4,8),\allowbreak(4,12),(4,16),(6,12),(8,16)\}$, then by Lemma \ref{necessary} we have $r+s=v-1$, and since $m\vert v$ and $n\vert v$ we have $h=lcm(m,n)\vert v$. For the sufficiency part, assume $h\vert v$ and $r+s=hx-1=v-1$ where $x$ is a nonnegative integer. For $(m,n)=(4,8)$, HWP$^{*}(8; 4^{r_0}, 8^{s_0})$ has a solution for all nonegative $r_0$ and $s_0$ with $r_0+s_0=7$ by Lemma \ref{lemma4.8}. Then, HWP$^{*}(v; 4^{r}, 8^{s})$ has a solution for $r+s=8x-1=v-1$ by Lemma \ref{mainlemma}. For $(m,n)\in\{(4,6),(4,12),(6,12)\}$, HWP$^{*}(12; m^{r_1}, n^{s_1})$ has a solution for all nonegative $r_1$ and $s_1$ with $r_1+s_1=11$ by Lemma \ref{lemma3}. Then, HWP$^{*}(v; m^{r}, n^{s})$ has a solution by Lemma \ref{mainlemma} for $(m,n)\in \{(4,6),(4,12)$, $(6,12)\}$ with $r+s=12x-1=v-1$. For $(m,n)\in \{(4,16),(8,16)\}$, HWP$^{*}(16; m^{r_2}, n^{s_2})$ has a solution for all nonegative $r_2$ and $s_2$ with $r_2+s_2=15$ by Lemma \ref{lemma6}. Then, by Lemma \ref{mainlemma}, HWP$^{*}(v; m^{r}, n^{s})$ has a solution for $(m,n)\in \{(4,16),(8,16)\}$ with $r+s=16x-1=v-1$. \end{proof} \section{Odd Cycle Sizes} Here we first give the following main construction, and using this construction we prove that HWP$^{*}(v; m^{r}, n^{s})$ has a solution for $(m,n)\in \{(3,5),(3,15),\allowbreak(5,15)\}$ with $r+s=v-1$, where $v$ is odd. \begin{lemma}\label{oddmainlemma} Let $m\geq3$ and $n\geq3$ be odd, $h=lcm(m,n)$ and $3\vert h$. If $\mathrm{HWP}^{*}(h; m^{r'}, n^{s'})$ has a solution for all $r'$, $s'$ satisfying $r'+s'=h-1$, then there is a solution to $\mathrm{HWP}^{*}(hx; m^{r}, n^{s})$ for all nonnegative $r$, $s$ and odd $x$ satisfying $r+s=hx-1$. \end{lemma} \begin{proof} By \eqref{3.1}, we have a decomposition of $K_{hx}^{*}$ into a $K_h^{*}$ and a $\left(K_{x}^{*} \wr \overline{K}_{h}\right)$-factor. Since $\overline{K}_{h}\cong \overline{K}_{3}\wr \overline{K}_{\frac{h}{3}}$, we have $K_{x}^{*} \wr \overline{K}_{h} \cong (K_{x}^{*} \wr \overline{K}_{3})\wr \overline{K}_{\frac{h}{3}}$. It is clear that $K_{x}^{*} \wr \overline{K}_{3}$ is isomorphic to $K_{3x}^{*} -xK_{3}^{*}$. Since Kirkman triple system of order $3x$ exists, we have a $C_3$-factorization of $K_{3x}$. Then, a $C_{3}^{*}\cong K_{3}^{*}$-factorization of $K_{3x}^{*} -xK_{3}^{*}$ is obtained by Lemma \ref{lemma1.3}. So, $K_{x}^{*} \wr \overline{K}_{3}$ has a decomposition into $\frac{3x-3}{2}$ $K_{3}^{*}$-factors. In $K_{x}^{*} \wr \overline{K}_{h}$, these $K_{3}^{*}$-factors form $K_{(\frac{h}{3}:3)}^{*}$-factors since $K_{3}^{*}\wr \overline{K}_{\frac{h}{3}} \cong K_{(\frac{h}{3}:3)}^{*}$. Let $F_0$ be the $K_{h}^*$-factor and $F_1,F_2,\dots,F_{\frac{3x-3}{2}}$ be the $K_{(\frac{h}{3}:3)}^*$-factors of $K_{hx}^{*}$. Since HWP$^{*}(h; m^{r'}, n^{s'})$ is assumed to have a solution for all nonnegative integers $r'$ and $s'$ where $r'+s'=h-1$, $F_0$ has a $\{\vv{C}_{m}^{r'},\vv{C}_{n}^{s'}\}$-factorization for all nonnegative integers $r'$ and $s'$ with $r'+s'=h-1$. Also, $K_{(\frac{h}{3}:3)}^*$ has a $\vv{C}_{m}$-factorization and a $\vv{C}_{n}$-factorization by Lemma \ref{lemma1.3} and Theorem \ref{liu}, so each $F_j$ has a $\{\vv{C}_{m}^{\frac{2h}{3}r_j},\vv{C}_{n}^{\frac{2h}{3}s_j}\}$-factorization for $j\in \{1,2, \dots,\frac{3x-3}{2}\}$, where $r_j, s_j \in \{0, 1 \}$ with $r_j+s_j=1$. These factorizations give us a $\{\vv{C}_{m}^{r},\vv{C}_{n}^{s}\}$-factorization of $K_{hx}^{*}$ where $r=r'+\sum_{i=0}^{\frac{3x-3}{2}} \frac{2h}{3}r_i$ and $s=s'+\sum_{i=0}^{\frac{3x-3}{2}} \frac{2h}{3} s_i$ with $r+s=r'+s'+\sum_{i=0}^{\frac{3x-3}{2}} \frac{2h}{3}(r_i+s_i)=h-1+hx-h=hx-1$. Since any nonnegative integer $0\leq r\leq hx-1$ can be written as $r=r'+\frac{2h}{3}a$ for integers $0\leq r'\leq h-1$ and $0\leq a\leq \frac{3x-3}{2}$, a solution to $\mathrm{HWP}^{*}(hx; m^r, n^s)$ exists for each $r \geq 0$ and $s \geq 0$ satisfying $r+s=hx-1$. \end{proof} \begin{lemma}\label{lemma3.2} For nonnegative integers $r$ and $s$, $\mathrm{HWP}^{*}(15; m^{r}, n^{s})$ has a solution for $(m,n)\in \{(3,5),(3,15)$, $(5,15)\}$ if and only if $r+s=14$ except possibly for $r\in \{11,12,13\}$ when $(m, n)=(3, 5)$ and for $r=13$ when $(m, n)=(3, 15)$. \end{lemma} \begin{proof} The cases when $r= 0$ and $s=0$ can be obtained by Theorem \ref{OP}. In \cite{Adams2002}, a solution to HWP$(15; m^{r_0}, n^{s_0})$ for $(m,n)\in \{(3,5),(3,15),(5,15)\}$ with the exception $(m,n,r_0,s_0)=(3,5,6,1)$ is given by Theorem 4.1. Thus, by Observation \ref{obs}, we have a solution to HWP$^{*}(15; m^{r}, n^{s})$ for $(m,n)\in \{(3,5),(3,15)$, $(5,15)\}$ with $r$ and $s$ are positive even integers except possibly for $(m,n,r,s)=(3,5,12,2)$. We list the solutions for the odd cases in the Appendix. \end{proof} Although there are missing cases in here, we can still find a solution of HWP$^{*}(v; m^{r}, n^{s})$ for $(m,n)\in \{(3,5),(3,15),(5,15)\}$ with $r+s=v-1$ and for odd $v>15$, using the main construction. \begin{theorem}\label{thm4.3} For all nonnegative integers $r$, $s$ and odd $v>15$, $\mathrm{HWP}^{*}(v;\allowbreak m^{r}, n^{s})$ has a solution for $(m,n)\in \{(3,5),(3,15),(5,15)\}$ if and only if $r+s=v-1$ and $15\vert v$. \end{theorem} \begin{proof} If a solution to HWP$^{*}(v; m^{r}, n^{s})$ exists for $(m,n)\in \{(3,5),(3,15),\allowbreak(5,15)\}$, then by Lemma \ref{necessary}, $r+s=v-1$ and $lcm(m,n)=15\vert v$. For the sufficiency part, assume $v=15x$ and $r+s=15x-1$ where $x>1$ is an odd integer. Using the main construction in Lemma \ref{oddmainlemma} for $v=15x$, we have a decomposition of $K_{v}^{*}$ into a $K_{15}^{*}$ and $\frac{3x-3}{2}$ $K_{(5:3)}^{*}$ factors. $K_{15}^{*}$ has a $\{\vv{C}_{m}^{r'},\vv{C}_{n}^{s'}\}$-factorization for $(m,n)\in \{(3,5),(3,15),(5,15)\}$ with $r'+s'=14$ except possibly for $r'\in \{11,12,13\}$ when $(m, n)=(3, 5)$ and for $r'=13$ when $(m, n)=(3, 15)$ by Lemma \ref{lemma3.2}. By Lemma \ref{lemma1.3} and Theorem \ref{liu}, each $K_{(5:3)}^{*}$ has a $\{\vv{C}_{m}^{10r_j},\vv{C}_{n}^{10s_j}\}$-factorization for $j\in \{1,2, \dots,\frac{3x-3}{2}\}$, where $r_j, s_j \in \{0, 1 \}$ with $r_j+s_j=1$. Let $r=r'+10a$ and $s=s'+10b$ for nonnegative $a$ and $b$ with $a+b=\frac{3x-3}{2}$, then we have $r+s=r'+s'+10(a+b)=14+5(3x-3)=15x-1=v-1$ with $0\leq r,s\leq 15x-1$. We obtain the requested integer $ r\in [0, 15x-1]$ from the sum of $r'$ and $10a$ for integers $0\leq r'\leq 14$ and $0\leq a\leq \frac{3x-3}{2}$. Therefore, $\mathrm{HWP}^{*}(v; m^r, n^s)$ has a solution with $r+s=15x-1=v-1$. \end{proof} According to our best knowledge, our results are the first findings for the directed version of the Hamilton-Waterloo Problem. We have first examined the cases $(m, n)\in \{(4,6),(4,8),(4,16),(8,16),(3,5),(3,15),(5,15)\}$, as done in the first paper on the undirected Hamilton-Waterloo Problem by Adams et al. \cite{Adams2002}. We have also solved the problem for the cases $(m, n)\in \{(4, 12), (6, 12)\}$. In addition to studying odd cycle cases $\{(3,5),(3,15),\allowbreak(5,15)\}$, we have also observed that if $\mathrm{HWP}(v; m^{r}, n^{s})$ has a solution for odd $v$, then $\mathrm{HWP}^{*}(v; m^{2r},n^{2s})$ has a solution for the same $r$ and $s$ as well. Since there is no $2$-factorizations of $K_v$ for even $v$, we cannot arrive the similar observation when $v$ is even and $m, n > 2$. Our constructions given in Lemma \ref{mainlemma} and Lemma \ref{oddmainlemma} can also be used to solve the problem for the other cycle sizes as long as the necessary small cases can be found. Now we can combine our results in the following main theorem. \begin{theorem} For nonnegative integers $r$ and $s$, $\mathrm{HWP}^{*}(v; m^{r}, n^{s})$ has a solution for \begin{enumerate} \item $(m,n)\in \{(4,6),(4,8),(4,12),(4,16),(6,12),(8,16)\}$ when $v$ is even, \item $(m,n)\in \{(3,5),(3,15),(5,15)\}$ when $v$ is odd \end{enumerate} if and only if $r+s=v-1$ and $lcm(m,n)\vert v$ except possibly for $r\in \{11,12,13\}$ when $(v, m, n)=(15, 3, 5)$ and for $r=13$ when $(v, m, n)=(15, 3, 15)$. \end{theorem}

456af2b28a488af16822bf277fab7f8f63a5a800

\section{Introduction} \label{sec:intro} The stabilizer formalism \cite{gottesman1997stabilizer} plays a central role in quantum information. Stabilizer states are states that lie in the intersection of the positive eigenspaces of $2^n$ commuting Pauli operators. Stabilizer states can be generated by Clifford circuits, which are the group of unitary transformations that normalize the Pauli group. Stabilizer states and the Clifford group have widespread applications in quantum error correction \cite{shor1995codes, calderbank1996codes}, measurement-based quantum computation \cite{raussendorf2000mbqc}, randomized benchmarking \cite{knill2008benchmarking}, and quantum learning algorithms \cite{huangkuengpresskill}. These applications are largely thanks to the rich algebraic structure afforded by the stabilizer formalism. Stabilizer states are also one of the few classes of states that admit efficient learning algorithms. Montanaro \cite{montanaro2017learning} gave an algorithm that takes $O(n)$ copies of an $n$-qubit stabilizer state and correctly identifies the state with high probability in time $O(n^3)$. Gross, Nezami, and Walter \cite{gross2021schur} gave an algorithm for \textit{property testing} stabilizer states, which is the task of distinguishing whether a state is a stabilizer state or is far from any stabilizer state. Remarkably, this algorithm requires only $6$ copes of the state. Despite finding numerous applications, Clifford circuits are not universal for quantum computation. Furthermore, in 1998, Gottesman and Knill showed that Clifford circuits acting on stabilizer states can be efficiently classically simulated \cite{gottesman1998heisenberg, aaronson2004stabilizer}. However, with the additional ability to apply a $T$-gate (the gate $\ket{0}\!\!\bra{0} + e^{i \pi / 4}\ket{1}\!\!\bra{1}$), the resulting gate set becomes universal. Therefore, efficient simulation of so-called Clifford+$T$ circuits would imply $\BPP = \BQP$, and a large line of work has been devoted to developing better simulation algorithms \cite{PashayanPhysRevLett.115.070501, BrayviPhysRevLett.116.250501, RallPhysRevA.99.062337, Bravyi2019simulationofquantum}. Currently, the best-performing simulation algorithms are based on modeling the output state of a quantum circuit as a decomposition of stabilizer states \cite{Bravyi2019simulationofquantum}. These decompositions give rise to simulation algorithms whose runtimes scale polynomially in the complexity of the decomposition. One such complexity measure is the \textit{stabilizer extent}. Consider the state $\ket{\psi} = \sum_i c_i \ket{\phi_i}$ for $c_i \in \mathbb{C}$ and stabilizer states $\ket{\phi_i}$. The stabilizer extent is the minimum $\left(\sum_i{\abs{c_i}}\right)^2$ over all such decompositions of $\ket{\psi}$, and scales exponentially in the number of $T$-gates in the circuit producing the state. A closely-related complexity measure is the \textit{stabilizer fidelity}, which is the maximum overlap between $\ket{\psi}$ and any stabilizer state. Indeed, the inverse of stabilizer fidelity lower bounds stabilizer extent \cite{Bravyi2019simulationofquantum}. Collectively, we informally refer to states with either low stabilizer extent or non-negligible stabilizer fidelity as states of low ``stabilizer complexity''. As a generalization of stabilizer states, it is natural to ask whether states of low stabilizer complexity are also efficiently learnable, and indeed a similar question has been raised before \cite{arunachalam2022phase}. Nevertheless, this problem remains largely open except in some highly restricted settings \cite{lai2022learning}. This could be in part because many of the useful properties of stabilizer states provably fail to generalize to states with low stabilizer complexity. For example, \cite{hinsche2022learning} observed that one can efficiently learn the output distribution of any Clifford circuit, given samples from this distribution.\footnote{Indeed, every such distribution is simply an affine subspace of $\mathbb{F}_2^n$.} However, this task already becomes intractable for circuits with a \textit{single} $T$-gate (producing a state of constant stabilizer extent), where \cite{hinsche2022learning} proved that learning the output distribution is as hard as the learning parities with noise problem. Furthermore, it is known that stabilizer states form a \textit{$t$-design} for $t = 3$, meaning that random stabilizer states duplicate the first 3 moments of the Haar measure \cite{ kuenghttps://doi.org/10.48550/arxiv.1510.02767, webb2016clifford}. By contrast, \cite{haferkamp2020homeopathy} showed that circuits with $\poly(t)$ non-Clifford gates are sufficient to generate approximate $t$-designs. Thus, for any constant $t$, states of constant stabilizer extent can form approximate $t$-designs. This suggests that states of low stabilizer complexity can give much stronger information-theoretic approximations to the Haar measure than ordinary stabilizer states, because stabilizer states fail to form a $t$-design for any $t > 3$ \cite{ZKGG16}. In this work, we investigate whether these properties that differentiate stabilizer states from low-stabilizer-complexity states can be pushed further, to prove hardness of learning low-stabilizer-complexity states. One natural approach towards proving that low-stabilizer-complexity states are hard to learn would be to show that they are \textit{pseudorandom}. Ji, Liu, and Song \cite{Ji10.1007/978-3-319-96878-0_5} define an ensemble of $n$-qubit states to be (computationally) pseudorandom if every $\poly(n)$-time quantum adversary has at most a negligible advantage in distinguishing copies of a state drawn randomly from the ensemble from copies of a Haar-random $n$-qubit state. Note that pseudorandom states are not efficiently learnable, as any algorithm for learning some set of quantum states gives an algorithm to distinguish those states from the Haar measure. Our main result is an efficient algorithm for distinguishing states of non-negligible stabilizer fidelity from Haar-random states, showing that such states \textit{cannot} be pseudorandom. This type of distinguishing is sometimes known as \textit{weak learning} in learning theory. \begin{theorem}[Informal version of \cref{thm:main-thm-alg}]\label{thm:main-theorem-informal} Let $\ket{\psi}$ be an unknown $n$-qubit pure state, and let $k \leq \frac{4}{5}2^{n/12}$. There is an efficient algorithm that distinguishes whether $\ket{\psi}$ is Haar-random or a state with stabilizer fidelity at least $\frac{1}{k}$, promised that one of these is the case. In particular, the algorithm uses $O(k^{12} \log(1/\delta))$ copies of $\ket{\psi}$ and $O(n k^{12} \log(1/\delta))$ time to succeed with probability at least $1 - \delta$. \end{theorem} \cref{thm:main-theorem-informal} also generalizes to distinguishing states with low stabilizer extent from Haar-random. To the best of our knowledge, prior to our work, it was even unknown whether states of stabilizer extent at most a \textit{constant} could be efficiently distinguished from Haar-random. We also emphasize that the contrast between our positive learning result and the hardness result of \cite{hinsche2022learning} stems in part from the differing access models: we assume access to copies of the quantum state, whereas \cite{hinsche2022learning} considers algorithms that only have outcomes of standard basis measurements of the state. As a simple corollary, we prove a first-of-its-kind lower bound on the number of $T$-gates required to prepare computationally pseudorandom quantum states. \begin{corollary}[\cref{cor:main-thm-pseudo}]\label{cor:main-thm-pseudo-informal} Any family of Clifford+$T$ circuits that produces an ensemble of $n$-qubit computationally pseudorandom quantum states must use at least $\omega(\log n)$ $T$-gates. \end{corollary} In some sense, \cref{cor:main-thm-pseudo-informal} contrasts sharply with the result of \cite{haferkamp2020homeopathy}, where circuits containing just a few non-Clifford gates are sufficient to produce strong information-theoretic approximations to the Haar measure (i.e. $t$-designs). Nevertheless, we emphasize that our result and \cite{haferkamp2020homeopathy} are formally incomparable, because computationally pseudorandom states need not form approximate $t$-designs for constant $t$, nor vice-versa. \subsection{Main Ideas} Let $x = (p,q) \in \mathbb{F}_2^{2n}$, where $p$ and $q$ are the first and last $n$ bits of $x$, respectively. Define $W_x \coloneqq i^{p\cdot q}X^p Z^q$ (a Pauli operator without phase), and let $\ket{\Phi^+} \coloneqq 2^{-n/2} \sum_{x \in \mathbb{F}_2^n} \ket{x, x}$ be a maximimally entangled state. Then, the set $\{\ket{W_x} \coloneqq (W_x \otimes I)\ket{\Phi^+} \mid x \in \mathbb{F}_2^{2n}\}$ is the \textit{Bell basis}, an orthonormal basis of $\mathbb{C}^{2^n} \otimes \mathbb{C}^{2^n}$. Our algorithm uses \textit{Bell difference sampling} \cite{montanaro2017learning, gross2021schur}, which works as follows (see \cref{ssec:weyl_and_bell} for more detail): Given four copies of an $n$-qubit pure state $\ket{\psi}$, perform a Bell-basis measurement on $\ket{\psi}^{\otimes 2}$ to get measurement outcome $x \in \mathbb{F}_2^{2n}$, repeat this on the remaining two copies to get measurement outcome $y \in \mathbb{F}_2^{2n}$, and return $z = x + y$. We refer to $p_\psi(x) \coloneqq 2^{-n} \abs{\braket{\psi|W_x|\psi}}^2$ as the \textit{characteristic distribution of} $\ket{\psi}$. To see that $p_\psi$ is a distribution, recall that since the Pauli operators form an orthonormal basis over Hermitian matrices, we can always decompose $\ketbra{\psi}{\psi} = \frac{1}{2^{n}}\sum_{x \in \mathbb{F}_2^n} \braket{\psi | W_x | \psi} \cdot W_x$. By assumption, $\abs{\braket{\psi|\psi}}^2 = 1$, so by Parseval's identity, \[\frac{1}{2^n}\sum_{x \in \mathbb{F}_2^n} \abs{\braket{\psi | W_x | \psi}}^2 = 1.\] Gross, Nezami, and Walter\ \cite{gross2021schur} showed that Bell difference sampling an arbitrary pure state $\ket{\psi}$ corresponds to sampling a random operator $W_x$ according to the following distribution: \[ q_\psi(x) =\!\!\! \sum_{y \in \mathbb{F}_2^{2n}} p_\psi (y) p_\psi (x + y). \] We call $q_\psi$ the \textit{Weyl distribution of} $\ket{\psi}$. Note that the Weyl distribution of $\ket{\psi}$ is the scaled convolution of the characteristic distribution with itself (i.e., $q_\psi = 4^n (p_\psi \ast p_\psi)$, where `$\ast$' is the convolution operator). Define the $\{\pm1\}$-outcome measurement $M_x \coloneqq \left\{\frac{I \pm W_x}{2}\right\}$ (projections onto the $\pm1$-eigenspaces of $W_x$). Our algorithm begins by repeating the following process $m$ times: sample a random Weyl operator $W_{x}$ (via Bell difference sampling) and perform the measurement $M_x^{\otimes 2}$ on $\ket{\psi^{\otimes 2}}$. Then, average all of the measurement outcomes. If the average is at least $1/\poly(k)$, we decide that $\ket{\psi}$ has stabilizer fidelity at least $\frac{1}{k}$. Otherwise, we decide that $\ket{\psi}$ is Haar-random. What statistic are we computing in our algorithm? Denote the measurement outcome on the $i$th iteration as $X_i \in \{\pm 1\}$. Observe that for all $X_i$, \[ \Es{}[X_i] =\!\!\! \sum_{x \in \mathbb{F}_2^{2n}} q_\psi(x) \abs{\braket{\psi|W_x|\psi}}^2 = 2^n \!\!\!\sum_{x \in \mathbb{F}_2^{2n}} q_\psi(x) p_\psi(x) = 2^n \Es{x \sim q_\psi}[p_\psi(x)], \] where the expectation $\mathop{\bf E\/}[X_i]$ is taken over sampling $x \sim q_\psi$ and the randomness from performing the measurement $M_x^{\otimes 2}$. Hence, for our algorithm to work, $\Es{x \sim q_\psi}[p_\psi(x)]$ must be ``different enough'' when $\ket{\psi}$ either is Haar-random or has low stabilizer complexity. Proving that this is the case is the main technical ingredient of our work: \begin{lemma}[Informal version of \cref{lem:expected-measurement}] Let $\ket{\psi}$ be an $n$-qubit pure state. Suppose the stabilizer fidelity of $\ket{\psi}$ is at least $\frac{1}{k}$. Then, $$ 2^n \mathop{\bf E\/}_{x \sim q_{\psi}}\left[ p_\psi(x) \right] \geq \dfrac{1}{k^6}. $$ In contrast, suppose $\ket{\psi}$ is a Haar-random quantum state. Then, with overwhelming probability over the Haar measure, $$ 2^n \mathop{\bf E\/}_{x \sim q_{\psi}}\left[ p_\psi(x) \right] \leq 2^{-n/2}. $$ \end{lemma} Our proof uses Fourier analysis of Boolean functions, and some parts of our proof are reminiscent of the celebrated Blum-Luby-Rubinfield linearity test \cite{BLRtest}. Intuitively, $q_\psi$ is significantly closer to linear when $\ket{\psi}$ has non-negligible stabilizer fidelity, as opposed to when $\ket{\psi}$ is a Haar-random state. With the above lemma, all that remains is ``merely'' a sample complexity analysis, namely: what $m$ is sufficient to distinguish whether the average is close to $0$ or $\Omega(1/k^6)$? In the simplest case, we show that $O(k^{12}\log(1/\delta))$ samples are sufficient by Hoeffding's inequality. However, this complexity can be improved if given access to a unitary that prepares $\ket{\psi}$ (and its inverse). In this model, we are able to achieve a quartic speedup in both sample and time complexity, which we explain in \cref{sec:qae}. \section{Preliminaries} \label{sec:preliminaries} First, we establish some notation used throughout this work. We denote $[n] \coloneqq \{1,\ldots,n\}$. For $v \in \mathbb{C}^n$, $\norm{v}_p \coloneqq (\sum_{i \in [n]}\abs{v_i}^p)^{1/p}$ is the $\ell_p$-norm. Logarithms are assumed to be in base $2$. For a probability distribution $P$ on a set $S$, we denote drawing a sample $s \in S$ according to $P$ by $s \sim P$. We denote drawing a sample $s \in S$ uniformly at random by $s \sim S$. \subsection{Stabilizer States and Stabilizer Complexity Measures}\label{subsec:stabilizer-prelim} We define the $1$-qubit Pauli group to be the collection of matrices $\{I, X, Y, Z\}$, where $$ I = \begin{pmatrix} 1 & 0 \\ 0 & 1 \end{pmatrix},\quad X = \begin{pmatrix} 0 & 1 \\ 1 & 0 \end{pmatrix},\quad Y = \begin{pmatrix} 0 & -i \\ i & 0 \end{pmatrix},\quad Z = \begin{pmatrix} 1 & 0 \\ 0 & -1 \end{pmatrix}. $$ The $n$-qubit Pauli group $\mathcal{P}_n$ is the set $\{\pm 1, \pm i\} \times \{I, X, Y, Z\}^{\otimes n}$. The Clifford group $\mathcal{C}_n$ is the group of unitary transformations generated by $H$, $S$, and $\mathrm{CNOT}$ gates, where $H$ is the Hadamard gate, $S \coloneqq \ket{0}\!\!\bra{0} + i \ket{1}\!\!\bra{1}$ is the phase gate, and $\mathrm{CNOT}$ is the controlled-not gate. We refer to unitary transformations in the Clifford group as Clifford circuits. Clifford circuits with the addition of the $T$-gate are universal, where the $T$-gate is defined by $T \coloneqq \ket{0}\!\!\bra{0} + e^{i \pi/4} \ket{1}\!\!\bra{1}$. A unitary transformation $U$ \textit{stabilizes} a state $\ket{\psi}$ when $U\ket{\psi} = \ket{\psi}$. It is folklore that if an $n$-qubit state can be reached from the $\ket{0^n}$ state by applying a Clifford circuit, then the state is stabilized by a group of $2^n$ commuting members of the subset $\{\pm 1\} \times \{I, X, Y, Z\}^{\otimes n} \subset \left(\mathcal{P}_n \setminus -I^{\otimes n}\right)$, called its \textit{stabilizer group}. Such states are called \textit{stabilizer states}, and we denote the set of stabilizer states by $\mathcal{S}_n$. For $\ket{\psi} \in \mathcal{S}_n$, we denote its stabilizer group as $\textrm{Stab}(\ket{\psi})$. For more background on stabilizer states, see, e.g., \cite{nielsen2002quantum}. We now define some complexity measures that characterize more general states in terms of stabilizer state decompositions. \begin{comment} \begin{definition}[approximate stabilizer rank \cite{Bravyi2019simulationofquantum}] Suppose $\ket{\psi}$ is a pure $n$-qubit state. Let $\delta > 0$ be a precision parameter. The \emph{approximate stabilizer rank} $\chi_\delta(\ket{\psi})$ is the smallest integer k such that $\norm{\ket{\psi} - \ket{\psi^\prime}}_2 \leq \delta$ for some state $\ket{\psi^\prime}$ with stabilizer rank $k$. \end{definition} \end{comment} \begin{definition}[stabilizer extent \cite{Bravyi2019simulationofquantum}] Suppose $\ket{\psi}$ is a pure $n$-qubit state. The \emph{stabilizer extent} of $\ket{\psi}$, denoted $(\ket{\psi})$, is the minimum of $\norm{c}_1^2$ over all decompositions $\ket{\psi} = \sum_i c_i \ket{\phi_i}$, where $\ket{\phi_i} \in \mathcal{S}_n$ and $c$ is some vector in $\mathbb{C}^{|\mathcal{S}_n|}$. \end{definition} \begin{definition}[stabilizer fidelity \cite{Bravyi2019simulationofquantum}] Suppose $\ket{\psi}$ is a pure $n$-qubit state. The \emph{stabilizer fidelity} of $\ket{\psi}$, denoted $F_{\mathcal{S}}$, is $$ F_{\mathcal{S}}(\ket\psi) \coloneqq \max_{\ket{\phi} \in \mathcal{S}_n}\left|\braket{\phi|\psi}\right|^2. $$ \end{definition} Below we give a useful relation between the complexity measures defined above. \begin{claim}\label{claim:metric-relations} Let $\ket{\psi}$ be an $n$-qubit pure state. Then, \[ \xi(\ket{\psi}) \geq \frac{1}{F_{\mathcal{S}}(\ket{\psi})} .\] \end{claim} \begin{proof} Let $\ket{\psi} = \sum_{\ket{\phi} \in \mathcal{S}_n} c_{\phi} \ket{\phi}$ be such that $\left(\sum_\phi \abs{c_\phi}\right)^2 = \xi(\ket{\psi})$. Suppose towards a contradiction that $F_{\mathcal{S}}(\ket{\psi}) < \frac{1}{\xi(\ket{\psi})}$ and therefore $\abs{\braket{\phi|\psi}} < \frac{1}{\xi(\ket{\psi})}$ for all $\ket{\phi} \in \mathcal{S}_n$. Then, \begin{align*} 1 = \abs{\braket{\psi|\psi}} = \left| \sum_{\ket{\phi_S} \in \mathcal{S}_n} c_{\phi}^\ast \braket{\phi|\psi} \right| & \leq \sum_{\ket{\phi_S} \in \mathcal{S}_n} \left| c_{\phi} \right| \left| \braket{\phi|\psi} \right| \\ & \leq \max_i \abs{\braket{\phi_i|\psi}} \sum_{\ket{\phi_S} \in \mathcal{S}_n} \left| c_{\phi} \right| \\ & \le F_{\mathcal{S}}(\ket{\psi})\sqrt{\xi(\ket{\psi})}\\ & < \sqrt{F_{\mathcal{S}}(\ket{\psi})} \leq 1 \end{align*} The last line follows from the fact that $F_{\mathcal{S}}(\ket{\psi}) \leq 1$ due to Cauchy-Schwarz and the definition of stabilizer fidelity. We then have $1 < 1$ as a clear contradiction. \end{proof} The claim above also follows as a special case of \cite[Theorem 4]{Bravyi2019simulationofquantum}, though its proof is more complicated. To prove lower bounds on the number of $T$-gates necessary to prepare pseudorandom quantum states, we need to upper bound the stabilizer extent of a quantum state prepared by a Clifford+$T$ circuit comprised of $t$ $T$-gates. \begin{claim}\label{claim:extent_linear_comb} For $\ket{\psi} = \alpha \ket{v} + \beta \ket{w}$, $$\xi(\ket{\psi}) \leq \left(\abs{\alpha} \sqrt{\xi(\ket{v})} + \abs{\beta} \sqrt{\xi(\ket{w})}\right)^2.$$ \end{claim} \begin{proof} Let $\ket{v} = \sum_i c_i \ket{\phi_i}$ and $\ket{w} = \sum_j d_j \ket{\varphi_j}$ be the minimal decompositions in terms of stabilizer extent (i.e., $\left(\sum_i \abs{c_i}\right)^2 = \xi(\ket{v})$). Since $\ket{\psi} = \alpha \ket{v} + \beta \ket{w} = \alpha \sum_i c \ket{\phi_i} + \beta \sum_j d \ket{\varphi_j}$, we have a stabilizer decomposition of $\ket{\psi}$. The stabilizer extent of this decomposition is at most \[ \left( \sum_i \abs{\alpha c_i + \beta d_i}\right)^2 \leq \left( \abs{\alpha} \sum_i \abs{c_i} + \abs{\beta}\sum_i \abs{d_i}\right)^2 \leq \left( \abs{\alpha} \sqrt{\xi(v)} + \abs{\beta}\sqrt{\xi(w)}\right)^2.\qedhere \] \end{proof} \begin{lemma}\label{lemma:clifford-t-extent} Let $C$ be any Clifford+$T$ circuit comprised of $t$ $T$-gates and $\ket\psi = C \ket{0^n}$. Then, $$\xi(\ket{\psi}) \leq \left(1 + \frac{1}{\sqrt{2}}\right)^t.$$ \end{lemma} \begin{proof} We note that a Clifford+$T$ circuit can be broken into layers of Clifford circuits, followed by a single $T$-gate, followed by more layers of Clifford circuits, and so on. Since Clifford circuits preserve stabilizer extent, we only need to show that the $T$-gate increases the stabilizer extent of any state by at most a constant multiplicative factor. Since the SWAP gate is a Clifford operation, we assume without loss of generality that each $T$-gate is applied to the first qubit. We proceed by induction on the layers of the circuit. In the first layer, when no $T$-gates have been applied, the bound is trivially true because the stabilizer extent of any stabilizer state is $1$. Now, assume that, after applying some portion of the circuit $C^\prime$ to $\ket{0^n}$ with $t-1$ $T$-gates, we get the state $\ket{\varphi}$. Observe that the $T$-gate can be expressed as $\cos(\pi/8)e^{i\pi/8} I + \sin(\pi/8)e^{i 13 \pi/8} Z$. Thus, $(T \otimes I^{\otimes n-1}) \ket{\varphi} = \cos(\pi/8)e^{i\pi/8} \ket{\varphi} + \sin(\pi/8)e^{i 13 \pi/8} \left(Z\otimes I^{\otimes n-1}\right)\ket{\varphi}$. Since $Z\otimes I^{\otimes n-1}$ is a Clifford operation, $\left(Z\otimes I^{\otimes n-1}\right) \ket{\varphi}$ has the same extent as $\ket{\varphi}$. Therefore, applying \cref{claim:extent_linear_comb}, \[ \xi(\ket\psi) \leq \left(\cos(\pi/8) + \sin(\pi/8)\right)^2\xi(\ket{\varphi}) \leq \left(1 + \frac{1}{\sqrt{2}}\right)^{t}. \qedhere \] \end{proof} \subsection{Boolean Fourier Analysis}\label{ssec:bool_fourier_analysis} We review the basics of Fourier analysis over the Boolean hypercube. \begin{definition} Let $S \subseteq [n]$ be an index of bits. Then the \emph{parity function} on $S$ is defined to be $$\chi_S(x) \coloneqq \prod_{i \in S}(-1)^{x_i}.$$ \end{definition} Alternatively, we can define $\chi_S(x) = (-1)^{x \cdot s}$ where $s_i = 1$ if and only if $i \in S$. This form will prove to be more natural for our purposes. The parity functions are orthonormal under the inner product $\langle f, g\rangle = \frac{1}{2^n} \sum_{x \in \mathbb{F}_2^n} f(x)g(x)$. Since there are $2^n$ distinct parity functions, this gives a complete basis. Given a function $f: \mathbb{F}_2^n \to \mathbb{R}$, we can then write $$ f(x) = \sum_{S \subseteq[n]}\hat{f}(S) \chi_S(x). $$ The $\hat{f}(S)$ are real numbers known as the \textit{Fourier coefficients} (collectively known as the \textit{Fourier spectrum}), and are equivalently given by the formula $$ \hat{f}(S) = \langle f(x), \chi_S(x) \rangle. $$ As a basis change, we can then rethink inner products to be over the Fourier coefficients as well. \begin{fact}[Plancherel's theorem]\label{fact:plancherel} $$ \langle f, g \rangle = \frac{1}{2^n}\sum_{S \subseteq [n]} \hat{f}(S)\hat{g}(S). $$ \end{fact} Finally, the convolution is an operation that appears frequently in Fourier analysis over the reals. We can similarly define it over Boolean inputs. \begin{definition} For functions $f, g: \mathbb{F}_2^n \to \mathbb{R}$, we define the \emph{convolution} $f \ast g$ as $$ (f \ast g)(x) \coloneqq \frac{1}{2^n}\sum_{t \in \mathbb{F}_2^n} f(t)g(x + t). $$ \end{definition} Much like Fourier transforms over the reals, convolution maps to multiplication in the Fourier domain. \begin{fact}[Convolution theorem]\label{fact:convolution_theorem} $\hat{f \ast g}(S) = \hat{f}(S)\hat{g}(S)$ \end{fact} For proofs of all of these facts, as well as for a comprehensive reference on analysis of Boolean functions, we recommend \cite{o2014analysis}. \subsection{Weyl Operators and Bell Difference Sampling}\label{ssec:weyl_and_bell} For $x = (p,q) \in \mathbb{F}_2^{2n}$, define the \textit{Weyl operator} as \[W_x \coloneqq i^{p \cdot q} (X^{p_1}Z^{q_1}) \otimes \ldots \otimes (X^{p_n}Z^{q_n}) = i^{p \cdot q}X^pZ^q. \] Each Weyl operator is a Pauli operator, and every Pauli operator is a Weyl operator (up to a phase). Note also that $W_x W_y = W_{x + y}$, up to a phase. We use Weyl operators (rather than Pauli operators) when it is convenient to identify members of the Pauli group with length-$2n$ bit strings. A critical subroutine in our work is \textit{Bell difference sampling}, which was introduced in \cite{montanaro2017learning, gross2021schur}. Let $\ket{\Phi^+} \coloneqq 2^{-n/2} \sum_{x \in \mathbb{F}_2^n} \ket{x, x}$. Then, the set of quantum states $\{\ket{W_x} \coloneqq (W_x \otimes I) \ket{\Phi^+} \mid x\in\mathbb{F}_2^{2n}\}$ forms an orthonormal basis of $\mathbb{C}^{2^n} \otimes \mathbb{C}^{2^n}$, which we call the \textit{Bell basis}. Bell sampling a state $\ket{\psi}$ refers to measuring $\ket{\psi}^{\otimes 2}$ in the Bell basis, and the measurement outcome is a length-$2n$ bit string $x$ that corresponds to a Weyl operator $W_x$. Bell difference sampling a state $\ket{\psi}$ refers to Bell sampling twice to get measurement outcomes $x, y \in \mathbb{F}_2^{2n}$ and returning $z = x + y$, which corresponds to a Weyl operator $W_z$ and uses four copies of $\ket{\psi}$. Montanaro showed Bell difference sampling can be performed in $O(n)$ time \cite{montanaro2017learning}. Bell difference sampling returns a random Weyl operator, but according to what distribution? Gross, Nezami, and Walter\ \cite{gross2021schur} showed that the underlying distribution depends on the so-called characteristic distribution of $\ket{\psi}$. \begin{definition}[characteristic distribution] The \emph{characteristic distribution} of $\ket{\psi}$ is defined as \[ p_{{\psi}}(x) \coloneqq 2^{-n} \abs{\braket{\psi|W_x|\psi}}^2. \] \end{definition} \begin{lemma}[{\cite[Theorem 3.2]{gross2021schur}}]\label{lem:bell_diff_sampling} Let $\ket{\psi}$ be an arbitrary $n$-qubit pure state. Bell difference sampling corresponds to drawing a sample from the following distribution: \[ q_{\psi}(x) \coloneqq 4^n (p_\psi \ast p_\psi)(x) = \sum_{y \in \mathbb{F}_2^{2n}} p_{\psi}(y) p_{{\psi}}(x + y). \] Additionally, if $\ket{\psi} \in \mathcal{S}_n$ is a stabilizer state, then \[ q_{\psi}(x) = p_{{\psi}}(x) = 2^{-n} \abs{\braket{\psi|W_x|\psi}}^2. \] \end{lemma} We refer to $q_\psi$ as the \textit{Weyl distribution}. Using this terminology, the characteristic distribution and Weyl distribution are equal only when $\ket{\psi}$ is a stabilizer state (i.e., when $4^n (p_\psi \ast p_\psi) = p_\psi$). \section{Certificate of Low Stabilizer Complexity}\label{sec:analysis} To efficiently distinguish a state with low stabilizer complexity (meaning, a state with low stabilizer extent or non-negligible stabilizer fidelity) from a Haar-random one, we require a property or statistic of the state that distinguishes it from Haar-random. As such, we present the following technical lemma, which forms the backbone of our algorithm. \begin{lemma} \label{lem:expected-measurement} Let $\ket\psi$ be an $n$-qubit pure state. If the stabilizer fidelity of $\ket{\psi}$ is at least $\frac{1}{k}$, then $$ \mathop{\bf E\/}_{x \sim q_{\psi}}\left[ \abs{\braket{\psi|W_x|\psi}}^2 \right] \geq \dfrac{1}{k^6}. $$ In contrast, if $\ket{\psi}$ is Haar-random and $n \geq 33$, then, with probability at least $1-\exp\left(-2^{n/2-15}\right)$ over the Haar measure, $$ \mathop{\bf E\/}_{x \sim q_{\psi}}\left[ \abs{\braket{\psi|W_x|\psi}}^2 \right] \leq 2^{-n/2}. $$ \end{lemma} Our algorithm then amounts to estimating the quantity $\mathop{\bf E\/}_{x \sim q_{\psi}}\left[ \abs{\braket{\psi|W_x|\psi}}^2 \right]$ via a procedure involving Bell difference sampling. To prove \cref{lem:expected-measurement}, as a first step, we relate $\mathop{\bf E\/}_{x \sim q_{\psi}}\left[ \abs{\braket{\psi|W_x|\psi}}^2 \right]$ to the Fourier coefficients of $p_\psi$. Note that this analysis closely resembles the BLR linearity test \cite{BLRtest} (see also \cite[Section 1.6]{o2014analysis}). \begin{fact}\label{fact:blr-similarity} Let $\ket{\psi}$ be an $n$-qubit pure state. Then, \[\mathop{\bf E\/}_{x \sim q_{\psi}}\left[ \abs{\braket{\psi|W_x|\psi}}^2 \right] = 32^n \!\!\!\sum_{x \in \mathbb{F}_2^{2n}} \wh{p_\psi}(x)^3. \] \end{fact} \begin{proof} \begin{align*} \mathop{\bf E\/}_{x \sim q_{\psi}}\left[ \abs{\braket{\psi|W_x|\psi}}^2 \right] &= 2^n \mathop{\bf E\/}_{x \sim q_{\psi}}\left[ p_\psi(x) \right] \\ &= 2^n \sum_{x \in \mathbb{F}_2^{2n}} p_\psi(x) q_\psi(x) \\ &= 8^n \sum_{x \in \mathbb{F}_2^{2n}} p_\psi(x) (p_\psi \ast p_\psi)(x) \\ &= 32^n \mathop{\bf E\/}_{x \sim \mathbb{F}_2^{2n}}[p_\psi(x) (p_\psi \ast p_\psi)(x)] \\ &= 32^n \sum_{x \in \mathbb{F}_2^{2n}} \wh{p_\psi}(x) \wh{p_\psi\ast p_\psi}(x)) && (\mathrm{\cref{fact:plancherel}})\\ &= 32^n \sum_{x \in \mathbb{F}_2^{2n}} \wh{p_\psi}(x)^3. && (\mathrm{\cref{fact:convolution_theorem}}) \qedhere \end{align*} \end{proof} For the remainder of this section, we use the following convention: when $x = (v, w) \in \mathbb{F}_2^{2n}$, $v$ and $w$ denote the first and last $n$ bits of $x$, respectively, and, we will sometimes write $p_\psi(v,w)$ and $q_\psi(v,w)$, rather than $p_\psi(x)$ and $q_\psi(x)$. \subsection{The Fourier Spectrum of the Characteristic Distribution} By \cref{fact:blr-similarity}, it is clear that understanding the Fourier spectrum of $p_\psi$ is one avenue to proving \cref{lem:expected-measurement}. \begin{proposition}\label{prop:fourier-coefficients} The Fourier coefficients of $p_\psi(v,w)$ are $\wh{p_\psi}(v,w) = \frac{1}{2^n} p_\psi(w,v)$. \end{proposition} \begin{proof} Define $f: \mathbb{F}_2^{2n} \xrightarrow[]{} [-1,1]$ as $f(v,w) \coloneqq \braket{\psi|i^{v\cdot w}X^v Z^w |\psi}$, where $v,w \in \mathbb{F}_2^n$. We begin by computing the Fourier expansion of $f$. \begin{align}\label{eq:expansion-of-f} f(v, w) &= \bra{\psi}i^{v\cdot w} X^v Z^{w} \ket{\psi} \nonumber\\ &= \left( \sum_{x \in \mathbb{F}_2^n} c^*_x \bra{x} \right)i^{v\cdot w} X^v Z^{w}\left( \sum_{x \in \mathbb{F}_2^n} c_x \ket{x} \right) \nonumber\\ &= i^{v\cdot w}\left( \sum_{x \in \mathbb{F}_2^n} c^*_x \bra{x + v} \right) \left( \sum_{x \in \mathbb{F}_2^n} (-1)^{x \cdot w} c_x \ket{x} \right) \nonumber\\ &= i^{v\cdot w}\sum_{x \in \mathbb{F}_2^n} c_{x+v}^* c_x (-1)^{w \cdot x}. \end{align} In the second line we are simply writing $\ket{\psi}$ in the computational basis. Observe now that $p_\psi(v,w) = \frac{1}{2^n} \abs{f(v,w)}^2$, which we can also treat as a function on Boolean variables. Hence, \begin{align*} p_\psi(v,w) &= \frac{1}{2^n}\left(i^{v\cdot w}\sum_{x \in \mathbb{F}_2^n} c_{x+v}^* c_x (-1)^{w \cdot x} \right)\left((-i)^{v\cdot w}\sum_{x \in \mathbb{F}_2^n} c_{x+v} c^*_x (-1)^{w \cdot x} \right) \nonumber\\ &= \frac{1}{2^n}\sum_{x,y \in \mathbb{F}_2^n} c^*_{v+y}c_y c_{v+x+y}c^*_{x+y} (-1)^{w \cdot x}, \end{align*} where the first equality follows by substituting in \cref{eq:expansion-of-f}. We can now compute the Fourier spectrum of $p_\psi$ by taking the inner product between $p_\psi$ and an arbitrary Fourier character (this is the simplest approach to computing Fourier coefficients). \begin{align*} \wh{p_\psi}(v,w) &= \frac{1}{4^n}\sum_{s, t \in \mathbb{F}_2^n} p_\psi(s, t) (-1)^{s\cdot v + t\cdot w}\\ &= \frac{1}{8^n}\sum_{s,t,x,y \in \mathbb{F}_2^n} c^*_{s+y}c_y c_{s+x+y}c^*_{x+y} (-1)^{t \cdot x + v \cdot s + w \cdot t}\\ &= \frac{1}{8^n}\sum_{s,x,y \in \mathbb{F}_2^n} c^*_{s+y}c_y c_{s+x+y}c^*_{x+y} (-1)^{v \cdot s} \sum_{t\in\mathbb{F}_2^n}(-1)^{t \cdot (x+w)}\\ &= \frac{1}{4^n}\sum_{s,y \in \mathbb{F}_2^n} c^*_{s+y}c_y c_{s+w+y}c^*_{w+y} (-1)^{v \cdot s} \\ &= \frac{1}{2^n}p_\psi(w,v).\qedhere \end{align*} \end{proof} \subsection{Low-Stabilizer-Complexity States} We prove the first part of \cref{lem:expected-measurement}; namely, that \[\mathop{\bf E\/}_{x \sim q_{\psi}}\left[ \abs{\braket{\psi|W_x|\psi}}^2 \right] \geq \dfrac{1}{k^6}\] when $\ket\psi$ has low stabilizer complexity. \begin{claim} \label{lem:sum_p-hat_cubed} For an $n$-qubit pure state $\ket{\psi} = \sum_{x \in \mathbb{F}_2^n} c_x \ket{x}$, \[ 32^n \sum_{x \in \mathbb{F}_2^{2n}} \wh{p_\psi}(x)^3 \geq \abs{c_0}^{12}. \] \end{claim} \begin{proof} \begin{align*} 32^n\!\! \sum_{v,w \in \mathbb{F}_2^{2n}} \wh{p_\psi}(v,w)^3 &= 4^n \sum_{v,w \in \mathbb{F}_2^{2n}}\!\! p_\psi(w,v)^3 && (\mathrm{\cref{prop:fourier-coefficients}}.) \\ &\geq 4^n \sum_{v \in \mathbb{F}_2^n}p_\psi(0,v)^3 && (\forall x, y, p_\psi(x,y) \geq 0.)\\ &= \frac{1}{2^{n}} \sum_{v \in \mathbb{F}_2^n} \abs{\braket{\psi | Z^v | \psi}}^{6}\\ & \geq \frac{1}{2^{6n}} \left(\sum_{v \in \mathbb{F}_2^n} \braket{\psi |Z^v| \psi}\right)^6 && \left(\sum_{i=1}^m \abs{a_i}^6 \geq \frac{1}{m^5}\left(\sum_{i = 1}^m \abs{a_i} \right)^6.\right)\\ & \geq \abs{c_0}^{12}. && \left(\sum_{v \in \mathbb{F}_2^n} Z^v = 2^n \ketbra{0^n}{0^n}.\right)\qedhere \end{align*} \end{proof} \begin{proof}[Proof of first part of \cref{lem:expected-measurement}] Let $\ket{\psi}$ be an $n$-qubit pure state, and suppose that the stabilizer fidelity of $\ket{\psi}$ is at least $\frac{1}{k}$. Then there exists a Clifford circuit $C \in \mathcal{C}_n$ such that $C\ket{\psi} = \sum_{x \in \mathbb{F}_2^n} c_x \ket{x}$ where $\abs{c_{0}}^2 \geq \frac{1}{k}$. Call $\ket{\phi} \coloneqq C\ket{\psi}$. By \cref{lem:sum_p-hat_cubed}, \[ 32^n \sum_{v,w \in \mathbb{F}_2^n} \wh{p_{\phi}}(v, w)^3\geq \abs{c_0}^{12} \geq \frac{1}{k^6}. \] A Clifford circuit $C$ is a permutation of the Pauli group under conjugation (i.e., $C^\dagger \mathcal{P}_n C = \mathcal{P}_n$ for any $C \in \mathcal{C}_n$). Hence, for all $C \in \mathcal{C}_n$ and $g:\mathcal{P}_n \rightarrow \mathbb{R}$, $$\sum_{x \in \mathbb{F}_2^{2n}} g(W_x) = \sum_{x \in \mathbb{F}_2^{2n}} g(C^\dagger W_x C).$$ Therefore, we conclude that \[ 32^n \sum_{v,w \in \mathbb{F}_2^n} \wh{p_{\psi}}(v, w)^3 \geq \frac{1}{k^6} \] as well. Combining this bound with \cref{fact:blr-similarity} completes the proof. \end{proof} \subsection{Haar-Random States} We complete the proof of \cref{lem:expected-measurement} by showing that $\mathop{\bf E\/}_{x \sim q_{\psi}}\left[ \abs{\braket{\psi|W_x|\psi}}^2 \right]$ is small when $\ket{\psi}$ is a Haar-random state. We begin by showing that, for a Haar-random state, all of the Weyl measurements (except $W_x = I$) are exponentially close to $0$ with overwhelming probability. \begin{lemma}[L\'{e}vy's Lemma, see e.g. \cite{Gerken13measureconcentration}] \label{lem:levy} Let $\mathbb{S}^N$ denote the set of all $N$-dimensional pure quantum states, and let $f: \mathbb{S}^N \to \mathbb{R}$ be $L$-Lipschitz, meaning that $\abs{f(\ket{\psi}) - f(\ket{\varphi})} \le L \cdot \norm{\ket{\psi} - \ket{\varphi}}_2$. Then: \[ \Pr_{\ket{\psi} \sim \mu_{\rm{Haar}}} \left[\abs{f(\ket{\psi}) - \mathop{\bf E\/}[f]} \ge \varepsilon \right] \le 2\exp\left(-\frac{N \varepsilon^2}{9\pi^3 L^2} \right). \] \end{lemma} \begin{lemma} \label{lem:weyl_lipschitz} For any $n$-qubit Weyl operator $W_x$, the function $f_x: \mathbb{S}^{2^n} \to \mathbb{R}$ defined by $f_x(\ket{\psi}) = \bra{\psi}W_x\ket{\psi}$ is $2$-Lipschitz. \end{lemma} \begin{proof} Write $W_x = \Pi_+ - \Pi_-$ where $\Pi_+$ and $\Pi_-$ are the projectors onto the positive and negative eigenspaces of $W_x$, respectively. Then, \begin{align*} \abs{f_x(\ket{\psi}) - f_x(\ket{\varphi})} &= \abs{\bra{\psi}W_x\ket{\psi} - \bra{\varphi}W_x\ket{\varphi}}\\ &= \abs{\bra{\psi}\Pi_+\ket{\psi} - \bra{\varphi}\Pi_+\ket{\varphi} - \bra{\psi}\Pi_-\ket{\psi} + \bra{\varphi}\Pi_-\ket{\varphi}}\\ &\le \abs{\bra{\psi}\Pi_+\ket{\psi} - \bra{\varphi}\Pi_+\ket{\varphi}} + \abs{\bra{\psi}\Pi_-\ket{\psi} + \bra{\varphi}\Pi_-\ket{\varphi}}\\ &= \abs{\ \norm{\Pi_+\ket{\psi}}_2 - \norm{\Pi_+\ket{\varphi}}}_2 + \abs{ \norm{\Pi_-\ket{\psi}}_2 - \norm{\Pi_-\ket{\varphi}}_2}\\ &\le \norm{\Pi_+(\ket{\psi} - \ket{\varphi})}_2 + \norm{\Pi_-(\ket{\psi} - \ket{\varphi})}_2\\ &\le 2\norm{\ket{\psi} - \ket{\varphi}}_2, \end{align*} where the third and fifth lines apply the triangle inequality, and the fourth and sixth lines use the fact that $\Pi_+$ and $\Pi_-$ are projectors. \end{proof} \begin{corollary} \label{cor:single_weyl_concentration} Let $W_x$ be any $n$-qubit Weyl operator in which $x \neq 0$ (i.e. $W_x \neq I$). Then: \[ \Pr_{\ket{\psi} \sim \mu_{\rm{Haar}}} \left[\abs{\braket{\psi | W_x | \psi}} \ge \varepsilon \right] \le 2\exp\left(-\frac{2^n \varepsilon^2}{36 \pi^3} \right). \] \end{corollary} \begin{proof} Define $f_x(\ket{\psi}) = \bra{\psi}W_x\ket{\psi}$ as in \Cref{lem:weyl_lipschitz}. By \Cref{lem:weyl_lipschitz}, we know that $f_x$ is $2$-Lipschitz. Additionally, observe that $\mathop{\bf E\/}[f] = 0$ over the Haar measure because exactly half of the eigenvalues of $W_x$ are $1$ and the other half are $-1$. Then the corollary follows from \Cref{lem:levy}. \end{proof} \begin{corollary} \label{cor:all_weyl_concentration} \[ \Pr_{\ket{\psi} \sim \mu_{\rm{Haar}}} \left[\exists x \neq 0 : \abs{\braket{\psi | W_x | \psi}} \ge \varepsilon \right] \le 2^{2n + 1}\exp\left(-\frac{2^n \varepsilon^2}{36 \pi^3} \right). \] \end{corollary} \begin{proof} This follows from \Cref{cor:single_weyl_concentration} and a union bound over all $2^{2n}$ possible Weyl operators. \end{proof} Note that if $\varepsilon \ge \frac{1}{\mathrm{poly}(n)}$, then the probability bound in \Cref{cor:all_weyl_concentration} is doubly-exponentially small. We have shown that, with high probability, all Weyl measurements (except $W_x = I$) are close to $0$. We use this to complete the proof of \cref{lem:expected-measurement}. \begin{proof}[Proof of second part of \cref{lem:expected-measurement}] Suppose $\ket{\psi}$ is a Haar-random state. By \cref{cor:all_weyl_concentration}, for all $W_x \neq I$, $\abs{\braket{\psi|W_x|\psi}}^2 = 2^n p(x) \leq \varepsilon^2$ with probability $1-2^{2n+1}\exp\left(-\frac{2^n\varepsilon^2}{36 \pi^3}\right)$. Therefore by \cref{fact:blr-similarity} and \cref{prop:fourier-coefficients}, \begin{align*} \mathop{\bf E\/}_{x \sim q_{\psi}}\left[ \abs{\braket{\psi|W_x|\psi}}^2 \right] &= 32^n \sum_{x, y \in \mathbb{F}_2^n} \widehat{p}(x, y)^3 \\ &= 4^n\sum_{w, v \in \mathbb{F}_2^n}p(v,w)^3\\ &= 4^n \left(\frac{1}{8^n} + \sum_{\substack{ w, v \in \mathbb{F}_2^n\\w,v \neq 0}} p(v,w)^3\right)\\ & \leq \frac{1 + (4^n-1)\varepsilon^6}{2^n}, \end{align*} with probability at least $1-2^{2n+1}\exp\left(-\frac{2^n\varepsilon^2}{36 \pi^3}\right)$. By setting $\epsilon^2 = \frac{1}{2^{n/6}}\left(\frac{2^n-2^{n/2}}{4^n-1}\right)^{1/3}$, we get \[ \mathop{\bf E\/}_{x \sim q_\psi}\left[ \abs{\braket{\psi|W_x|\psi}}^2 \right] \leq \frac{1}{2^{n/2}} \] with probability at least $1-2^{2n+1}\exp\left(-\frac{2^{5n/6}}{36 \pi^3}\left(\frac{2^n-2^{n/2}}{4^n-1}\right)^{1/3}\right)$, which is at least $1-\exp\left(-2^{n/2-15}\right)$ for $n \geq 33$. \end{proof} \section{Algorithm and Sample Complexity Analysis}\label{sec:main-result} We are now ready to state and analyze our algorithm that distinguishes between Haar-random states and states with low stabilizer complexity. Our algorithm uses the fact that we can efficiently sample from $q_\psi$ (via Bell difference sampling) and efficiently estimate $\abs{\braket{\psi|W_x|\psi}}^2$ for any given $x \in \mathbb{F}_2^{2n}$, using quantum measurements. By combining these subroutines, we construct an unbiased estimator for $\mathop{\bf E\/}_{x \sim q}\left[\abs{\braket{\psi| W_x |\psi}}^2\right]$. Motivated by \cref{lem:expected-measurement}, if our estimator exceeds a certain threshold we determine that the input state has low stabilizer complexity; otherwise, we determine that the state is Haar-random. For the remainder of this section, $\eta \coloneqq \mathop{\bf E\/}_{x \sim q}\left[\abs{\braket{\psi| W_x |\psi}}^2\right]$. \vspace{\baselineskip} \begin{algorithm}[H] \SetKwInOut{Promise}{Promise} \caption{Distinguishing Low-Stabilizer-Complexity States from Haar-Random} \label{alg:distinguisher} \DontPrintSemicolon \KwInput{Black-box access to copies of $\ket\psi$} \Promise{$\ket{\psi}$ is Haar-random or has stabilizer fidelity at least $\frac{1}{k}$} \KwOutput{$1$ if $\ket{\psi}$ has stabilizer fidelity at least $\frac{1}{k}$ and $0$ if $\ket{\psi}$ is Haar-random} Let $m = 60k^{12}\ln(1/\delta)$. \RepTimes{$m$}{ Perform Bell difference sampling to obtain $W_x \sim q_{\psi}$. Perform the measurement $W_x^{\otimes 2}$ on $\ket{\psi}^{\otimes 2}$. Let $X_i \in \{\pm 1\}$ denote the measurement outcome. } Set $\hat{\eta} = \frac{1}{m} \sum_i X_i$. Return $1$ if $\hat{\eta} \geq \frac{2}{3k^6}$, and $0$ otherwise. \end{algorithm} \vspace{\baselineskip} \begin{theorem} \label{thm:main-thm-alg} Let $\ket{\psi}$ be an unknown $n$-qubit pure state for some $n \geq 33$, and let $k \leq \frac{4}{5} 2^{n/12}$. \cref{alg:distinguisher} distinguishes whether $\ket{\psi}$ is Haar-random or a state with stabilizer fidelity at least $\frac{1}{k}$, promised that one of these is the case. The algorithm uses $O\left(k^{12}\log(1/\delta)\right)$ copies of $\ket{\psi}$ and $O(n k^{12} \log(1/\delta))$ time, and distinguishes the two cases with success probability at least $1-\delta$. \end{theorem} \begin{proof} Following the notation in \cref{alg:distinguisher}, $X_i$ is the outcome of the measurement on the $i$th iteration of the algorithm loop. Observe that for any $X_i$, \[ \Es{x\sim q_\psi,\\\text{meas. by $W_x^{\otimes 2}$}}[X_i] = \mathop{\bf E\/}_{x \sim q_\psi} \bra{\psi^{\otimes 2}}W_x^{\otimes 2}\ket{\psi^{\otimes 2}} = \mathop{\bf E\/}_{x \sim q_\psi} \abs{\braket{\psi|W_x|\psi}}^2= \eta. \] Therefore, $\hat{\eta} = \frac{1}{m} \sum_i X_i$ is an unbiased estimator of $\eta$ (i.e., $\mathop{\bf E\/}[\hat{\eta}] = \eta$). Suppose $\ket{\psi}$ has stabilizer fidelity at least $\frac{1}{k}$. Then, our algorithm fails when $\hat{\eta} < \frac{2}{3k^6}$. Hence, \[ \Pr[\text{\cref{alg:distinguisher} fails}] = \Pr\left[\hat{\eta} < \frac{2}{3k^6} \right] = \Pr\left[ \hat{\eta} - \eta < \frac{2}{3k^6} - \eta \right] \leq \Pr\left[\hat{\eta} - \eta \leq -\frac{1}{3k^6}\right], \] where the last inequality follows from \cref{lem:expected-measurement}. By Hoeffding's inequality, \[ \Pr\left[\hat{\eta} - \eta \leq -\frac{1}{3k^6}\right] \leq \exp\left(-\frac{m }{18 k^{12}} \right). \] Therefore, $m \geq 18k^{12} \ln(15) = 49k^{12}$ samples suffice for the failure probability to be at most $\frac{1}{15}$. Now suppose $\ket{\psi}$ is Haar-random. Then, our algorithm fails when $\hat{\eta} \geq \frac{2}{3k^6}$. By \cref{lem:expected-measurement}, $\eta \leq 2^{-n/2}$ with probability at least $1-\exp\left(-2^{n/2-15}\right) >= 1-e^{-2\sqrt{2}}$ for $n \geq 33$. Assuming that $\eta \leq 2^{-n/2}$, \[ \Pr[\text{\cref{alg:distinguisher} fails}] = \Pr\left[\hat{\eta} \geq \frac{2}{3k^6} \right] = \Pr\left[ \hat{\eta} - \eta \geq \frac{2}{3k^6} - \eta \right] \leq \Pr\left[\hat{\eta} - \eta \geq \frac{1}{2k^6} - 2^{-n/2} \right]. \] Once again, by Hoeffding's inequality, \begin{align*} \Pr\left[\hat{\eta} - \eta \geq \frac{1}{2k^6} - 2^{-n/2} \right] &\leq \exp\left(- \frac{m}{2} \left(\frac{2}{3k^6} - 2^{-n/2}\right)^2 \right) \\ &\leq \exp\left(- \frac{m}{2} \left(\frac{2}{3k^6} - \frac{1}{3k^6}\right)^2 \right) \\ &\leq \exp\left(- \frac{m}{18 k^{12}}\right). \end{align*} Therefore, $m \geq -18 k^{12} \ln\left(\frac{1}{15} - e^{-2\sqrt{2}}\right) \geq 88k^{12}$ samples suffice for the failure probability to be at most $\frac{1}{15} - e^{-2\sqrt{2}}$. By the union bound, the failure probability is at most $\frac{1}{15}$, where the randomness is over both the Haar measure and the quantum measurements. We have shown that in either case we output the correct answer with probability at least $\frac{14}{15}$. Using the Chernoff-Hoeffding theorem, the success probability can be boosted from $\frac{14}{15}$ to at least $1 - \delta$ by doing $\frac{2}{3}\ln (1/\delta)$ repetitions of \cref{alg:distinguisher} and taking the majority answer. Since each iteration of the algorithm loop uses $6$ copies of $\ket{\psi}$, \cref{alg:distinguisher} consumes $O\left(k^{12} \log(1/\delta)\right)$ copies in total. Finally, Bell difference sampling and performing the measurement $W_x^{\otimes 2}$ takes $O(n)$ time, so the total runtime is $O\left(n k^{12} \log(1/\delta)\right)$. \end{proof} All of these results also apply to states with stabilizer extent at most $k$, since by \cref{claim:metric-relations}, such states have stabilizer fidelity at least $\frac{1}{k}$. \begin{corollary} \label{cor:main-thm-alg-rank} Let $\ket{\psi}$ be an unknown $n$-qubit pure state for $n \geq 33$, and let $k \leq \frac{4}{5} 2^{n/12}$. \cref{alg:distinguisher} distinguishes whether $\ket{\psi}$ is Haar-random or a state with stabilizer extent at most $k$, promised that one of these is the case. The algorithm uses $O\left(k^{12}\log(1/\delta)\right)$ copies of $\ket{\psi}$ and distinguishes the two cases with success probability at least $1-\delta$. \end{corollary} The above result immediately implies that output states of Clifford+$T$ circuits with few $T$-gates cannot be computationally pseudorandom. \begin{corollary} \label{cor:main-thm-pseudo} Any family of Clifford+$T$ circuits that produces an ensemble of $n$-qubit computationally pseudorandom quantum states must use at least $\omega(\log n)$ $T$-gates. \end{corollary} \begin{proof} Consider any ensemble of states wherein each state in the ensemble is the output of some Clifford+$T$ circuit with at most $K \log n$ $T$-gates. By \cref{lemma:clifford-t-extent}, the stabilizer extent of any such state $\ket{\psi}$ is at most $n^{\alpha K}$ for $\alpha \leq 0.7716$. By \cref{cor:main-thm-alg-rank}, on input copies of $\ket{\psi}$, \cref{alg:distinguisher} takes $O(n^{12\alpha K+1}) \le \poly(n)$ time and outputs $1$ with probability at least $2/3$. On the other hand, if $\ket{\psi}$ is a Haar-random state then the same algorithm outputs $1$ with probability at most $\frac{1}{3}$. As such, the algorithm's distinguishing advantage between the ensemble and the Haar measure is non-negligible. This is to say that the ensemble cannot be pseudorandom under the definition of \cite{Ji10.1007/978-3-319-96878-0_5}. \end{proof} \section{Open Problems} \label{sec:discussion} An immediate direction for future work is to improve the sample complexity of our algorithm, or to prove sample complexity lower bounds. One can also endeavour to improve other features of our algorithm: Is it possible to remove the need for entangled measurements?\footnote{ The optimal algorithms for learning and testing stabilizer states use entangled measurements. So, a first step would be to understand how many separable measurements are required to separate stabilizer states from Haar-random.} Or, is it possible to show that entangled measurements are in any sense necessary? Are there quantum measurements that allow us to sample from $p_\psi$ directly (rather than $q_\psi$)? Beyond that, can Bell difference sampling be used for learning and/or property testing stabilizer-extent-$k$ states? For stabilizer states ($k=1$), a $6$-query property testing algorithm is given by \cite{gross2021schur} and a $\Theta(n)$-query learning algorithm is given by \cite{montanaro2017learning}. Both algorithms rely on Bell difference sampling and are asymptotically optimal. We ask if there are generalizations of these algorithms for states with higher stabilizer complexity, similar to the question that was raised in \cite{arunachalam2022phase}. \begin{question} Is there a $\poly(k)$-query algorithm for property testing stabilizer-extent-$k$ states? Likewise, is there a $\poly(n,k)$-time algorithm for learning stabilizer-extent-$k$ states? \end{question} Our results on stabilizer extent are due to the fact that extent and fidelity are inversely related. Is it possible to relate \textit{stabilizer rank} (a closely-related complexity measure, denoted by $\chi$) and stabilizer fidelity? For instance, proving that, for all states $\ket{\psi}$, \[ F_{\mathcal{S}}(\ket{\psi})^{-1} \leq \chi(\ket{\psi})^c,\quad\text{for any constant $c$,} \] would imply that our algorithm can distinguish low-stabilizer-rank states from Haar-random. However, proving such a relation for even $c < \frac{\alpha n}{\log n}$ for $\alpha \leq 0.2284$ would imply super-linear lower bounds on the stabilizer rank of Clifford magic states, a long-standing open problem. One can also ask if the lower bound on the number of $T$-gates necessary for computationally pseudorandom states can be improved. \begin{question} How many $T$-gates are necessary for a family of Clifford+$T$ circuits to produce an ensemble of $n$-qubit computationally pseudorandom states? \end{question} We remark that any improvements to our $\log n$ lower bound would require techniques beyond the ones used in our paper. Indeed, in \cref{sec:magic-extent} we show that one can hope for at most a quadratic improvement in the relationship between $\eta$ and stabilizer fidelity. Such an improvement would only yield constant-factor improvements on the number of $T$-gates necessary to prepare computationally pseudorandom states. On the other hand, we are not aware of any attempts to optimize the $T$-gate count for plausible constructions of $n$-qubit pseudorandom states. The best upper bound we know of is the essentially trivial bound of $O(n)$, based on constructions of with $O(n)$ general gates. This is because pseudorandom states can be constructed from pseudorandom functions (PRFs) with constant overhead \cite{brakerski10.1007/978-3-030-36030-6_10}, and PRFs are believed to be constructible in linear time \cite{ishai_10.1145/1374376.1374438, fan_10.1145/3519935.3520010}.\footnote{Technically, we are not sure whether the PRFs constructed in \cite{ishai_10.1145/1374376.1374438, fan_10.1145/3519935.3520010} are secure against quantum adversaries, which is necessary for instantiating \cite{brakerski10.1007/978-3-030-36030-6_10}'s construction, but we consider it reasonable to conjecture that linear-time quantum-secure PRFs exist.} \section*{Acknowledgments} We thank Scott Aaronson for helpful comments. SG, VI, DL are supported by Scott Aaronson's Vannevar Bush Fellowship from the US Department of Defense, the Berkeley NSF-QLCI CIQC Center, a Simons Investigator Award, and the Simons ``It from Qubit'' collaboration. WK is supported by an NDSEG Fellowship. \bibliographystyle{alphaurl}

4ef0eed79fefb8f793916f4d29488d1ff51d93fd

\section{Introduction} \label{sec:intro} Different from full reference image quality assessment (IQA) (e.g. PSNR, SSIM), blind IQA computes quality directly from the distorted image without a clean image, which makes it suitable to user generated content, communication and other scenarios where clean sources are inaccessible. Traditional blind IQA approach relies on the regularity of natural scene statistics. Hand-crafted features \cite{mittal2012no} are designed to measure the distortion of image from natural ones. More recently, deep CNN based blind IQA approaches have shown promising results as learned features are more powerful \cite{kang2014convolutional}. And a typical inference of such IQA approach involves several steps: (1). a CNN backbone (e.g. VGG16) is applied to extract deep features from distorted images. (2). the deep features are spatially average pooled (SAP). (3). the averaged features are passed through a regression head (e.g. fully connected layers) to predict image quality. In some works, the CNN backbone is pretrained on other tasks (e.g. ImageNet) and finetuned with regression head \cite{talebi2018nima}\cite{su2020blindly}\cite{chen2021nested}. While for other works auxiliary tasks such as predicting residual or noise type are added \cite{kim2018deep}\cite{chen2021nested}. Despite variation in the training methods, the inference procedures are similar. \begin{figure}[thb] \includegraphics[width=8.5cm]{fig1.PNG} \label{fig:res} \caption{(a). CNN based full reference IQA, (b). CNN based blind IQA and (c). our proposed SMP approach.} \end{figure} Formally, denote distorted image as $x' \in R^{H \times W}$, the feature from CNN as $f(x) \in R^{C \times H' \times W'}$, then the output of spatially average pooling $SAP(f(x'))$ can be seen as spatial mean of feature map. And the pooled feature map is further processed by head $g$ to obtain the predicted quality $d(x')$. \begin{equation} d(x') = g(SAP(f(x'))) = \\ g(E[f(x')]) \label{eq:1} \end{equation} On the other hand, CNN based full-reference IQA methods also show promising results. As a representative method, LPIPS \cite{zhang2018perceptual} has been widely adopted as perceptual loss \cite{mentzer2020high}. Similar to blind IQA, LPIPS first computes the feature maps of reference image $x$ and distorted image $x'$ from a CNN $f$. Then the square difference of those feature maps are computed and spatially averaged. Finally, the averaged results are processed by head $g$ to predict the distance metric $d(x, x')$. \begin{equation} \begin{array}{l} d(x, x') = g(SAP((f(x) - f(x'))^2)) \\ \hspace{3.5em} = g(Var[f(x) - f(x')]) \end{array} \label{eq:2} \end{equation} Though both Eq.~\ref{eq:1} and Eq.~\ref{eq:2} adopt SAP over features, the statistics computed are different (See Fig. 1). For blind IQA, the SAP is used to compute $1^{st}$ moment (mean) of feature map. For LPIPS, by assuming the residual has $0$ mean, the SAP computes the $2^{nd}$ central moment (variance) of feature map difference. For long, variance has been an important part of texture feature \cite{haralick1979statistical,sajjadi2017enhancenet}. Higher moments (e.g. skewness, kurtosis) are also applied in texture analysis \cite{shaban2001improvement}. So a natural question to ask is, is spatial variance and other moments more representative than mean alone for image quality? In this paper, we propose spatial moment pooling (SMP) as an extension of SAP for blind IQA. Our approach of adding higher moments absorbs the arts of full reference IQA methods and texture features. To be specific, for each channel of feature map, those higher central moments (e.g. variance) are also computed and concatenated with $1^{st}$ moment. It can be seen as a generalization of SAP ($1^{st}$ moment). Moreover, we also propose a normalization approach to avoid numeric issue when back-propagating through higher order moments. Experimental results show that by simply replacing SAP with SMP, the performance of many deep CNN based IQA methods is significantly improved. Our method outperforms previous CNN approaches. \section{Related Works} \label{sec:rw} \subsection{Poolings and Deep CNN based blind IQA} Despite the pioneer of CNN based blind IQA \cite{kang2014convolutional} adopts min-max-pooling, the majority of subsequent works \cite{su2020blindly}\cite{chen2021nested}\cite{chadha2021deep}\cite{kim2016fully} adopts SAP (include global average pooling) as it has become popular in CNN. \cite{ying2020patches} adopts a mixture of SAP and ROI Pooling. \cite{talebi2018nima} and \cite{ke2021musiq} resize input images or patches to fixed size instead, and do not use pooling for final features. On the other hand, covariance pooling \cite{wang2020deep} has been shown effective for high level vision tasks. And the $i = 2$ moment is the diagonal elements of a covariance matrix. However, to the best of our knowledge, neither moment pooling with $i \ge 3$ nor moment pooling with $i \ge 2$ for CNN based blind IQA has been studied. \section{spatial Moment Pooling} \subsection{Background} First, we will conduct a brief review of moments, spatial average pooling (SAP) and its relationship to convolution. Given a random variable $X \in R$, the $i^{th}$ moment is defined as $E[X^i], i = {1, ..., n}$, and the $i^{th}$ central moment is defined as $E[(X - E[X])^i], i = {2, ..., n}$. When $i = 1$, the moment is the mean $\mu$ of $X$. When $i = 2$, the central moment is the variance $\sigma^2$ of $X$. When $i = 3$, the central moment is the unnormalized skewness of $X$, which is defined as $E[(\frac{X - E[X]}{\sigma})^3]$. When $i = 4$, the central moment is the unnormalized kurtosis of $X$, which is defined as $E[(\frac{X - E[X]}{\sigma})^4]$. In practice, higher order moments with $i > 4$ are less commonly used. On the other hand, given a feature map $f \in R^{C \times H \times W}$, the SAP is the operation of computing the $i = 1$ moment of values inside a pooling window to output the pooled feature map $f_{SAP} \in R^{C \times H' \times W'}$ (See Fig. 2). Identical to the standard convolution operation, the outputted $H'\times W'$ are determined by kernel size, stride, dilation and padding. In fact, for any pooling, the pooling window is the same as convolution windows given same settings. Besides, it can also be implemented as \textit{im2col}. The difference between pooling and convolution is after extracting windows to row vectors. In convolution, the extracted rows $r_i^T$ is dot produced with the flattened kernel. In SAP, the mean of $r_i^T$ is computed. This makes SAP a special form of convolution with box filter kernel. However, the spatial moment pooling-n in Section 3.2 can not be generalized by convolution when $n \ge 2$, since the $i \ge 2$ moments are non-linear and convolution is linear. \subsection{Spatial Moment Pooling-n / SMP(n)} \begin{figure}[htb] \includegraphics[width=8cm]{fig3.png} \label{fig:smp} \caption{a. spatial average pooling/SAP illustrated. b. spatial moment pooling-4/SMP(4) illustrated.} \end{figure} Similarly, we define the spatial moment pooling-n/SMP(n) as the operation of computing the $i = {1, ..., n}$ central moments of a pooling window, concatenating the moments in channel dimension and outputting the pooled feature map $f_{SMP} \in R^{nC \times H' \times W'}$ (See Fig. 2). The outputted spatial dimensions $H'\times W'$ are the same as SAP. However, the channel size increases from $C$ to $nC$. SMP(n) is a generalization of SAP. In fact, SMP(1) is exactly the same as SAP. To intuitively show why SMP(n) might represent image quality better, we provide a toy example in Fig. 3. Two $3 \times 3$ feature maps with same mean are shown. One is checkerboard pattern, and the other is solid color. Despite the same means, the $i>1$ central moments differ. \begin{figure}[htb] \includegraphics[width=8cm]{fig2.PNG} \label{fig:example} \caption{The $1^{st}$ moment, $2^{nd}, 3^{rd}, 4^{th}$ central moment of a $3 \times 3$ checkerboard feature map and solid color feature map.} \end{figure} \subsection{Numerical Difficulties and Normalization} Simply elevating SMP(n) to $n \ge 3$ cases can cause severe difficulties when optimizing the network with back-propagation. As shown in Tab.~\ref{tab:ablation}, the training fails and produces NaN results without proper normalization. So we propose to add layer normalization after $i \ge 3$ central moments (See Fig. 2). The rationale of normalizing only $i \ge 3$ is: (1) $SMP(n), n \le 2$ works well without normalization. (2). only the statistical skewness and kurtosis are normalized with $\sigma$, the $\mu$ and $\sigma$ themselves are defined as unnormalized. The reason why to choose layer normalization over others is purely empirical. The comparison of normalization methods is detailed in Section 4.2.2. In the following sections, SMP(n) with $n \ge 3$ implies that $i \ge 3$ moments are layer normalized if no normalization method is specified. \section{Experiments} \subsection{Experiment Setup} The SMP(n) is implemented in Pytorch 1.8, all the experiments are conducted on a computer equipped with Intel Xeon E5-2620 v4 and 8 Nivida TitanXp GPU. Basically, there are two types of datasets for blind IQA. One is large scale, natural datasets without reference images (such as Koniq-10k \cite{hosu2020koniq}). The other is small scale synthetic datasets with reference images and distortion information (such as LIVE \cite{sheikh2006statistical}). Usually the methods designed for small dataset use reference and distortion as auxiliary task, which makes them not applicable to large datasets. Consequently, the sota methods for large and small dataset are different. For large dataset, we choose Koniq-10k as dataset, and HyperIQA \cite{su2020blindly} as current CNN-based sota method. For small dataset we choose LIVE \cite{sheikh2006statistical} and CSIQ \cite{larson2010most} as dataset, and NemgIQA \cite{chen2021nested} as sota method. For ablation study we use Koniq-10k as dataset, and a variant of NIMA \cite{chadha2021deep}\cite{talebi2018nima} as base method. The dataset splits are the same as \cite{su2020blindly} and \cite{chen2021nested}. \subsection{Ablation Study} For baseline, we use an contemporary variant of classical NIMA \cite{talebi2018nima}. As proposed and described in \cite{chadha2021deep}, this variant incorporates the multi-scale spatial average feature maps. To simplify notations in the tables below, we denote this variant as NIMA. We only conduct experiments on VGG16 backbone, as the selection of CNN backbone selection is beyond scope of this paper. \begin{table}[ht] \centering \caption{Ablation study results.} \vspace{2mm} \label{tab:ablation} \begin{tabular}{lcc} \toprule & \multicolumn{2}{c}{Koniq-10k} \\ \cmidrule(l){2-3} & SRCC & PLCC \\ \midrule NIMA \cite{talebi2018nima}\cite{chadha2021deep} / NIMA-SMP(1) & 0.864 & 0.879 \\ NIMA-Vanilla-Large & 0.864 & 0.874 \\ NIMA-SMP(2) & 0.886 & 0.896 \\ NIMA-SMP(4)-w/o-Norm & NaN & NaN \\ NIMA-SMP(4)-BatchNorm & 0.880 & 0.892 \\ NIMA-SMP(4)-MaxNorm & 0.888 & 0.899 \\ NIMA-SMP(4)-LayerNorm & \textbf{0.890} & \textbf{0.900} \\ \bottomrule \end{tabular} \end{table} \subsubsection{Effects of spatial moment pooling} Tab.~\ref{tab:ablation} shows that NIMA-SMP(2) evidently improves both PLCC and SRCC over NIMA. NIMA-SMP(4)-LayerNorm outperforms SMP(2) despite the gain is not as significant as NIMA-SMP(2) over NIMA. Therefore, we stop at $n=4$ considering that the moments higher than $4^{th}$ are also uncommon in statistics. \subsubsection{Effects of high moments normalization} Simply expanding NIMA-SMP(2) to NIMA-SMP(4)-w/o-Norm brings optimization difficulties. After several iterations of training, the network produces constant results regardless of input images. And such constant output has $0$ variance, which further leads to NaN in PLCC and SRCC (See Tab.~\ref{tab:ablation}). Although the NaNs can be eliminated by introducing batch normalization to $3^{rd}, 4^{th}$ moments, it also brings performance decay. The NIMA-SMP(4)-BatchNorm is outperformed by NIMA-SMP(2), this confirms the observations that batch normalization negatively effects scale sensitive low level computer vision tasks \cite{yu2018wide}. Empirically, we find that max and layer normalization enable NIMA-SMP(4) to produce a reasonably superior performance over NIMA-SMP(2). Moreover, NIMA-GMP(4)-LayerNorm outperforms all other methods. \subsubsection{Effects of network parameter increment} The SMP's increment of parameters is of no significance. For NIMA-SMP(4)-LayerNorm, the parameter is increased by $0.20\%$ compared with NIMA. And for NIMA-SMP(2), the parameter increase is only $0.068\%$. Moreover, the increment of MACs is also minimal. The base model NIMA has a MAC of 635.23G Mac (multiply-accumulate) when input image size is 3x1920x1080. Replacing last pooling with SMP-(2) brings 1.02M extra Macs. And replacing last pooling with SMP-(4) brings 3.03M extra Macs. To fully verify the performance improvement comes from SMP instead of parameter increase, we build NIMA-Vanillia-Large with much wider and deeper regression head. NIMA-Vanillia-Large increases the parameter by as much as $14.74\%$. However, Tab.~\ref{tab:ablation} shows that the performance does not benefit from na\"ive model size increase. \subsection{Results on Large Dataset} \begin{table}[ht] \centering \caption{Results on Koniq-10K datasets. Blue and black numbers in bold represent the best and second best respectively. Reference methods' results are sourced from \cite{chen2021nested}\cite{ke2021musiq}.} \vspace{2mm} \label{tab:large} \begin{tabular}{lcc} \toprule & \multicolumn{2}{l}{Koniq-10K} \\ \cmidrule(l){2-3} & SRCC & PLCC \\ \midrule \multicolumn{3}{l}{\textit{DCNN based methods}} \\ BIECON \cite{kim2016fully} & 0.618 & 0.651 \\ PQR \cite{zeng2018blind} & 0.880 & 0.884 \\ SFA \cite{li2018has} & 0.856 & 0.872 \\ NIMA \cite{talebi2018nima} & 0.864 & 0.879 \\ DBCNN \cite{zhang2018blind} & 0.875 & 0.884 \\ HyperIQA \cite{su2020blindly} & 0.906 & 0.917 \\ \midrule \multicolumn{3}{l}{\textit{DCNN + SMP based methods}} \\ HyperIQA-SMP(2) & \textbf{0.909} & \textbf{0.924} \\ HyperIQA-SMP(4) & \textbf{0.909} & 0.922 \\ \midrule \multicolumn{3}{l}{\textit{Transformer based methods}} \\ MUSIQ-Singlescale & 0.905 & 0.919 \\ MUSIQ-Multiscale & \textcolor{blue}{\textbf{0.916}} & \textcolor{blue}{\textbf{0.928}} \\ \bottomrule \end{tabular} \end{table} For HyperIQA on large dataset, we replace the global average pooling in local distortion aware modules. Moreover, we set the learning rate to $1 \times 10^{-5}$. The other settings are identical to original HyperIQA \cite{su2020blindly}. Results in Tab.~\ref{tab:large} show that such modification effectively improves HyperIQA, and a new sota of deep CNN based approaches is made. Notably, our HyperIQA-SMP(2) and SMP(4) are superior to the transformer with single scale image patch input. Although the multi-scale transformer outperforms SMP methods, it is not quite fair to draw comparison as MUSIQ-Multiscale requires multi-scale image inputs. We also note that it is possible to train HyperIQA-SMP(n) with multi-scale image crops, but the effects of single-scale and multi-scale input is beyond the scope of this paper. \subsection{Results on Small Dataset} \begin{table}[h] \centering \caption{Results on LIVE and CSIQ datasets. Blue and black numbers in bold represent the best and second best respectively. Reference methods' results are sourced from \cite{chen2021nested}\cite{ke2021musiq}.} \vspace{2mm} \label{tab:my-table} \begin{tabular}{lcccc} \toprule & \multicolumn{2}{c}{LIVE} & \multicolumn{2}{c}{CSIQ} \\ \cmidrule(l){2-3} \cmidrule(l){4-5} & SRCC & PLCC & SRCC & PLCC \\ \midrule \multicolumn{5}{l}{\textit{DCNN based methods}} \\ BIECON \cite{kim2016fully} & 0.963 & 0.965 & 0.837 & 0.858 \\ PQR \cite{zeng2018blind} & 0.965 & 0.971 & 0.873 & 0.901 \\ SGDNet \cite{yang2019sgdnet} & 0.961 & 0.964 & 0.878 & 0.909 \\ HyperIQA \cite{su2020blindly} & 0.961 & 0.963 & 0.914 & 0.927 \\ NemgIQA \cite{chen2021nested} & 0.971 & 0.975 & 0.923 & 0.934 \\ \midrule \multicolumn{5}{l}{\textit{DCNN + SMP based methods}} \\ NemgIQA-SMP(1) & 0.972 & 0.978 & 0.921 & 0.937 \\ NemgIQA-SMP(2) & \textbf{0.973} & \textbf{0.980} & \textcolor{blue}{\textbf{0.929}} & \textbf{0.942} \\ NemgIQA-SMP(4) & \textcolor{blue}{\textbf{0.976}} & \textcolor{blue}{\textbf{0.982}} & \textbf{0.924} & \textcolor{blue}{\textbf{0.945}} \\ \bottomrule \end{tabular} \end{table} For NemgIQA on small dataset, it is not possible to directly replace last global average pooling with SMP. Unlike other works with regression performed after pooling, NemgIQA uses a single channel output point wise convolution as regression head followed by pooling. To make SMP usable, we swap the order of pooling and regression back. To be speficic, we change last layers to: point wise regression with output channels 4, pooling, and an affine layer to produce score. The other settings are the same as original NemgIQA \cite{chen2021nested}. Due to such modifications NemgIQA-SMP(1) is not strictly equivalent to original NemgIQA. However, experimental results show that this does not effects its performance. Moreover, the NemgIQA-SMP(2) and NemgIQA-SMP(4) significantly improve PLCC and SRCC compared to both original NemgIQA and NemgIQA-SMP(1), and achieve sota performance on LIVE and CSIQ datasets. \section{Conclusion} In this paper, we extend spatial average pooling (SAP) into spatial moment pooling (SMP) by adding higher order moments. Moreover, we propose an optimization friendly normalization trick that makes training of network with SMP stable. Experimental results show that replacing SAP with SMP significantly improves deep CNN-based blind IQA approaches. \bibliographystyle{IEEEbib}

2ab8625751b739aa76ce82e8a2d1f671c3568783

\section{Introduction} The Weibull distribution, with probability density function \begin{align} \label{eqn:pdf} p(y | \bm{\theta}) = \left(\frac{k}{\lambda^k}\right) y^{k-1} \exp\left(-\left(\frac{y}{\lambda}\right)^k\right) , \end{align} where $\bm{\theta} = \{k,\lambda\}$ and $k>0$ is the shape parameter and $\lambda>0$ is the scale parameter, is a popular distribution in analysis of survival data. Given data ${\bf y} = (y_1, \ldots, y_n)^\prime$, a common approach to estimate Weibull parameters $\bm{\theta}$ is the method of maximum likelihood where the parameters are set to values that maximise the log-likelihood of the data \begin{equation} \label{eqn:complete:nll} \ell(\bm{\theta}) = -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} In this setting, the maximum likelihood (ML) estimates of $k,\lambda$ are \begin{equation} \label{eqn:ml:k} \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 \begin{equation} \label{eqn:ml:lambda} \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 therefore be obtained by numerical optimisation. It is well known that the maximum likelihood estimate of the shape parameter $k$ is biased for small sample sizes. Ross~\cite{Ross94} derived a simple adjustment formula for the ML estimate of $k$ \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} which aims to reduce the bias and later extended his approach to censored data~\cite{Ross96}. Hirose~\cite{Hirose99} proposed an alternative bias correction method for data with no censoring that was derived by fitting a non-linear function to simulation results. Teimouri and Nadarajah~\cite{TeimouriNadarajah13} develop improved maximum likelihood estimates for the Weibull distribution based on record statistics. In contrast, Yang and Xie~\cite{YangXie03} use the modified profile likelihood proposed by Cox and Reid~\cite{CoxReid87,CoxReid92} to derive an alternative maximum likelihood estimate of $k$ (MLC) from \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} Using simulations, Yang and Xie showed that their estimate of $k$ is less biased than the ML estimate and more efficient than the estimate (\ref{eqn:mle:kross}) proposed by Ross. In a follow-up paper, Shen and Yang~\cite{ShenYang15} developed a profile maximum likelihood estimate of $k$ for complete and censored samples and showed that it outperforms MLC in simulations with complete data. \subsection{Type I Censored Data} In survival analysis, one commonly does not observe complete data and instead has joint realisations of the random variables $(Y = y, \Delta = \delta)$ where $Y = \min (T, c)$ and \begin{eqnarray*} \Delta &=& {\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 variable $T$ denotes survival time and $c > 0$ is the fixed censoring time. The data comprises the survival time $T=t$ of an item if this is less than the corresponding censoring time $c$ (i.e., $T \leq c$); otherwise, we only know that the item survived beyond time $c$ (i.e., $T > c$). The log-likelihood of data $D = \{(y_1, \delta_1), \ldots, (y_n, \delta_n)\}$ under type I censoring is \begin{align} \ell(\bm{\theta}) &= d \log \left(\frac{k}{\lambda^k}\right) -\frac{1}{\lambda^k} \sum_{i=1}^n y_i^k + \sum_{i=1}^n \log y_i^{\delta_i (k-1)} \label{eqn:censored:nll} \end{align} where $d = \sum_{i=1}^n \delta_i$. 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 $\hat{k}({\bf y})$ is obtained from \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} As with the case of complete data, the maximum likelihood estimate of $k$ with type I censored data has large bias in small samples and for large amounts of censoring. Based on the modified profile likelihood, Yang and Xie~\cite{YangXie03} propose an alternative estimate of $k$ \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} We note that the above score function requires $d>1$ to yield a positive estimate for $k$. Yang and Xie demonstrate that the proposed estimate of $k$ is less biased and more efficient than the regular maximum likelihood estimate. Shen and Yang~\cite{ShenYang15} derived a new second- and third-order bias correction formula for the shape parameter of the Weibull distribution without censoring and with general right-censoring models. Although the new estimate is shown to be effective in correcting bias, it must be computed through bootstrap simulation. The same procedure was later extended to include Weibull regression with complete and general right censoring~\cite{ShenYang17}. More recently, Choi et al~\cite{ChoiEtAl20} examine a different problem of Weibull parameter overestimation caused by mass occurrences of (censored) events in the early time period and develop an expectation maximization (EM) algorithm to reduce bias. \section{A simple adjustment to maximum likelihood estimates to reduce estimation bias} In a landmark paper, Cox and Snell~\cite{CoxSnell68} derived an approximation to the finite sample bias of ML estimates for independent, but not necessarily identically distributed, data. Let $\bm{\theta} \in \mathbb{R}^p$, where $p > 0$ is the number of free parameters, which is $p=2$ in the case of the Weibull model. Cox and Snell showed that the bias for the $s$-th element of the ML estimate $\hat{\theta}_{\rm ML}$ can be written as \begin{equation} \left[ \text{Bias}(\hat{\theta}_{\rm ML}) \right]_{s} = \sum_{i=1}^p \sum_{j=1}^p \sum_{l=1}^p \kappa^{s,i} \kappa^{j,l} \left( \frac{1}{2} \kappa_{ijl} + \kappa_{ij,l}\right) + O(n^{-2}) \end{equation} for $s = 1,\ldots,p$, where the cumulants are \begin{align} \kappa_{ij} &= \mathbb{E}\left\{ \frac{\partial^2 \ell(\bm{\theta})}{\partial \theta_i \partial \theta_j} \right\}, \quad % \kappa_{ijl} = \mathbb{E}\left\{ \frac{\partial^3 \ell(\bm{\theta})}{\partial \theta_i \partial \theta_j \partial \theta_l} \right\}, \\ % \kappa_{ij,l} &= \mathbb{E}\left\{ \frac{\partial^2 \ell(\bm{\theta})}{\partial \theta_i \partial \theta_j} \frac{\partial \ell(\bm{\theta})}{\partial \theta_l}\right\}, \end{align} for $i,j = 1,\ldots, p$ and $\kappa^{i,j}$ is the $(i,j)$-th entry of the \emph{inverse} of the expected Fisher information matrix ${\bf K} = \{ -\kappa_{ij} \}$. Following Cordeiro and Klein~\cite{CordeiroKlein94}, we can compactly write this in matrix notation as \begin{equation} \label{eqn:CoxSnell} \text{Bias}(\hat{\theta}_{\rm ML}) = {\bf K}^{-1} {\bf A} \text{vec}({\bf K}^{-1}) + O(n^{-2}), \end{equation} where the matrix ${\bf A}$ is the $(p \times p^2)$ matrix given by \begin{align} \label{eq:Cordeiro:Klein:A} {\bf A} &= \left[ {\bf A}^{(1)} | {\bf A}^{(2)} | \cdots | {\bf A}^{(p)} \right], \quad {\bf A}^{(l)} = \{a_{ij}^{(l)} \} \\ \quad a_{ij}^{(l)} &= \kappa_{ij}^{(l)} - \frac{1}{2} \kappa_{ijl}, \quad \kappa_{ij}^{(l)} = \frac{\partial \kappa_{ij}}{ \partial \theta_l} \end{align} for $i,j,l = 1,\ldots, p$. The ML estimate with reduced bias $\bm{\tilde{\theta}}_{\rm ML}$ is then \begin{align} \nonumber \bm{\tilde{\theta}}_{\rm ML} &= \hat{\bm{\theta}}_{\rm ML} - \hat{\bf K}^{-1} \hat{\bf A} \text{vec}(\hat{\bf K}^{-1}), \\ \label{eqn:ml:bias:adjustment} &= \hat{\bm{\theta}}_{\rm ML} - \text{Bias}(\hat{\theta}_{\rm ML}) , \end{align} where $\hat{\bf K}$ and $\hat{\bf A}$ are evaluated at the usual maximum likelihood estimate $\hat{\bm{\theta}}_{\rm ML}$. A benefit of this bias approximation formula is that it can be computed even if the maximum likelihood estimate is not available in closed form. A similar approach to the above was used to derive bias-adjusted maximum likelihood estimates for the unit Weibull distribution~\cite{MenezesEtAl21} and the inverse Weibull distribution~\cite{MazucheliEtAl18} with complete data only. \begin{thm} \label{thm:jointpdf} The finite sample bias of the maximum likelihood estimate (\ref{eqn:ml:k}) for the Weibull distribution with complete data is \begin{align} \text{\rm Bias}(\hat{k}_{\rm ML}) &= k \left(\frac{18 \left(\pi ^2-2 \zeta (3)\right)}{n \pi ^4}\right) + O(n^{-2})\\ &\approx k \left(\frac{1.3795}{n}\right) \label{eqn:ml:k:adjusted} \end{align} where $\zeta(\cdot)$ is the Riemann zeta function and $\gamma \approx 0.5772$ is the Euler--Mascheroni constant. Maximum likelihood estimates of $k$ and $\lambda$ with reduced bias can be obtained from (\ref{eqn:ml:bias:adjustment}). \begin{proof} The proof involves the application of the Cordeiro and Klein~\cite{CordeiroKlein94} approach, given by (\ref{eqn:CoxSnell}) and (\ref{eq:Cordeiro:Klein:A}), to the Weibull distribution (\ref{eqn:pdf}). It is well known that expected Fisher information matrix for the Weibull distribution is \begin{align*} {\bf K} &= \left( \begin{array}{cc} \frac{\left(6 (\gamma -1)^2+\pi ^2\right) n}{6 k^2} & \frac{(\gamma -1) n}{\lambda } \\ \frac{(\gamma -1) n}{\lambda } & \frac{k^2 n}{\lambda ^2} \\ \end{array} \right), \\ {\bf K}^{-1} &= \left( \begin{array}{cc} \frac{6 k^2}{\pi ^2 n} & -\frac{6 (\gamma -1) \lambda }{\pi ^2 n} \\ -\frac{6 (\gamma -1) \lambda }{\pi ^2 n} & \frac{\left(6 (\gamma -1)^2+\pi ^2\right) \lambda ^2}{\pi ^2 k^2 n} \\ \end{array} \right) . \end{align*} By direct calculation, we can show that the matrix ${\bf A}$ is \begin{equation*} \left( \begin{array}{cccc} \frac{n \left(-12 \zeta (3)-3 \gamma \left(2 \gamma (\gamma -7)+\pi ^2+16\right)+7 \pi ^2+12\right)}{12 k^3} & -\frac{\left(6 \gamma (\gamma -4)+\pi ^2+12\right) n}{12 k \lambda } & -\frac{\left(6 \gamma (\gamma -4)+\pi ^2+12\right) n}{12 k \lambda } & \frac{(-\gamma k+3 k+\gamma -1) n}{2 \lambda ^2} \\ -\frac{\left(6 \gamma (\gamma -4)+\pi ^2+12\right) n}{12 k \lambda } & -\frac{(\gamma k+k+\gamma -1) n}{2 \lambda ^2} & \frac{(-\gamma k+3 k+\gamma -1) n}{2 \lambda ^2} & -\frac{(k-1) k^2 n}{2 \lambda ^3} \\ \end{array} \right) \end{equation*} Substituting ${\bf K}^{-1}$ and ${\bf A}$ into (\ref{eqn:CoxSnell}) and simplifying completes the proof. \end{proof} \end{thm} We observe that the maximum likelihood estimate of $k$ is biased upward for any finite $n$. An advantage of the proposed bias adjusted estimate is that it can readily be computed in any software that implements ML Weibull estimation. \begin{thm} \label{thm:typeI} The finite sample bias of the maximum likelihood estimate (\ref{eqn:ml:k}) for the Weibull distribution with type I censored data is \begin{align} \text{\rm Bias}(\hat{k}_{\rm ML}) &= k\left(\frac{f(p)}{n}\right) + O(n^{-2}) \end{align} where $f(p)$ is a somewhat lengthy and complicated function of the proportion of uncensored observations $p = 1-\exp(-\left(c/\lambda )\right)^k)$. A simple approximation to $f(p)$ based on rational functions is \begin{equation*} f(p) \approx \frac{-580.684 p^3+4690.74 p^2-20743.7 p+18830}{-17026.8 p^2+18804.5 p+1} \end{equation*} with the absolute approximation error being less than $0.003$ for all $0.05 \leq p \leq 0.95$. As with complete data, maximum likelihood estimate of $k$ with reduced bias can be obtained from (\ref{eqn:ml:bias:adjustment}). \end{thm} \begin{figure}[t] \begin{center} \includegraphics[width=6.0cm]{figs/bias_function.pdf} \end{center} \caption{Bias adjustment $f(p)$ for the maximum likelihood estimate of the Weibull distribution shape parameter $k$ of as a function of the proportion of uncensored observations $p=1-\exp(-(c/\lambda)^k)$. \label{fig:biasf}} \end{figure} The proof is straightforward involving the application of the Cordeiro and Klein~\cite{CordeiroKlein94} bias adjustment formulation to the Weibull distribution with type I censoring and is therefore omitted. Figure~\ref{fig:biasf} shows the bias adjustment $f(p)$ as a function of the proportion of uncensored observations $p$. As the proportion of uncensored observations $p \to 1$ (ie, no censoring), $f(p) \to (\approx) 1.3795$ as expected. Additionally, $f(p) \to \infty$ as the proportion of censored data is increased (ie, $p \to 0$). \subsection{Simulation} \label{eqn:sim} \begin{table*}[ht] \scriptsize \begin{center} \begin{tabular}{cccccccccccccccc} \toprule $n$ & $k^*$ & \multicolumn{4}{c}{Bias} & & \multicolumn{4}{c}{Mean Squared Error} & & \multicolumn{4}{c}{KL divergence}\\ & & ML & MLC & MLP & MMLE & ~ & ML & MLC & MLP & MMLE & ~ & ML & MLC & MLP & MMLE\\ \cmidrule{1-16} \multirow{4}{*}{10} & 0.5 & 0.085 & 0.008 & {\bf 0.001} & 0.004 & & 0.038 & 0.023 & {\bf 0.022} & 0.023 & & 0.268 & 0.169 & {\bf 0.164} & 0.165\\ & 1.0 & 0.170 & 0.017 & {\bf 0.002} & 0.009 & & 0.151 & 0.092 & {\bf 0.089} & 0.090 & & 0.272 & 0.169 & 0.164 & {\bf 0.158}\\ & 5.0 & 0.852 & 0.088 & {\bf 0.014} & 0.045 & & 3.775 & 2.305 & {\bf 2.241} & 2.268 & & 0.264 & 0.167 & 0.162 & {\bf 0.152}\\ & 10.0 & 1.701 & 0.172 & {\bf 0.026} & 0.087 & & 15.102 & 9.218 & {\bf 8.978} & 9.079 & & 0.266 & 0.168 & 0.164 & {\bf 0.153}\\ \vspace{-2mm} \\ \cmidrule{2-16} \vspace{-2mm} \\ \multirow{4}{*}{20} & 0.5 & 0.038 & 0.004 & {\bf 0.001} & 0.001 & & 0.012 & 0.009 & {\bf 0.009} & 0.009 & & 0.072 & 0.061 & {\bf 0.060} & 0.061\\ & 1.0 & 0.077 & 0.009 & {\bf 0.001} & 0.003 & & 0.048 & 0.037 & {\bf 0.037} & 0.037 & & 0.072 & 0.061 & 0.060 & {\bf 0.059}\\ & 5.0 & 0.382 & 0.042 & {\bf 0.004} & 0.011 & & 1.203 & 0.928 & {\bf 0.916} & 0.917 & & 0.072 & 0.061 & 0.060 & {\bf 0.058}\\ & 10.0 & 0.755 & 0.075 & {\bf 0.000} & 0.014 & & 4.769 & 3.686 & {\bf 3.636} & 3.639 & & 0.072 & 0.061 & 0.060 & {\bf 0.058}\\ \vspace{-2mm} \\ \cmidrule{2-16} \vspace{-2mm} \\ \multirow{4}{*}{40} & 0.5 & 0.014 & 0.002 & {\bf 0.000} & 0.000 & & 0.004 & 0.003 & 0.003 & {\bf 0.003} & & 0.023 & 0.022 & {\bf 0.022} & 0.022\\ & 1.0 & 0.029 & 0.003 & {\bf -0.000} & 0.000 & & 0.015 & 0.013 & 0.013 & {\bf 0.013} & & 0.023 & 0.021 & 0.021 & {\bf 0.021}\\ & 5.0 & 0.143 & 0.016 & {\bf 0.001} & 0.001 & & 0.367 & 0.329 & 0.328 & {\bf 0.327} & & 0.023 & 0.022 & 0.021 & {\bf 0.021}\\ & 10.0 & 0.290 & 0.036 & {\bf 0.005} & 0.006 & & 1.458 & 1.308 & 1.300 & {\bf 1.299} & & 0.023 & 0.022 & 0.022 & {\bf 0.021}\\ \vspace{-3mm} \\ \bottomrule \vspace{+1mm} \end{tabular} \caption{Bias, mean squared error and Kullback--Leibler (KL) divergence for maximum likelihood (ML), conditional maximum likelihood of Yang and Xie (MLC), profile maximum likelihood of Shen and Yang (MLP) and our bias-adjusted maximum likelihood (MMLE) estimates of $k^*$ computed over $10^5$ simulations runs with $\lambda^* = 1$.\label{tab:results:complete}} \end{center} \end{table*} We performed a simulation to examine the finite sample behaviour of the new bias-adjusted ML estimates of $k$ for both complete and type I censored data. \subsubsection{Complete data} For each run of the simulation, we generated $n$ data points from the model Weibull$(k^*, \lambda^* = 1)$ where $n = \{10, 20, 50\}$ and the shape parameter was set to $k^* \in \{0.5, 1, 5, 10\}$. Maximum likelihood estimates, our proposed bias-adjusted maximum likelihood estimates (MMLE), conditional maximum likelihood estimates (MLC) proposed by Yang and Xie~\cite{YangXie03}, and the profile maximum likelihood estimates of Shen and Yang (MLP)~\cite{ShenYang15} were then computed from the data. We used the second-order bias reduction of Shen and Yang as it was virtually indistinguishable from the third-order formula in our tests. We performed $10^5$ simulations for each combination of $(k^*, n)$ and recorded the average bias, mean squared error and Kullback--Leibler (KL) divergence~\cite{KullbackLeibler51} from the data generating model. In the case of the Weibull distribution, the KL divergence between the data generating model Weibull$(k_0, \lambda_0)$ and approximating model Weibull$(k_1, \lambda_1)$ is \begin{align*} {\rm KL}( k_0, \lambda_0 || k_1, \lambda_1) &= \left(\frac{\lambda_0 }{\lambda _1}\right)^{k_1} \left( \frac{k_1}{k_0} \right) \Gamma \left( \frac{k_1}{k_0} \right) + \left(\frac{k_1}{k_0}-1\right) \gamma \\ &+ \log \left(\frac{k_0}{k_1} \left(\frac{\lambda _1}{\lambda_0 }\right)^{k_1}\right)-1 . \end{align*} All simulation results are shown in Table~\ref{tab:results:complete}. All three bias-adjusted maximum likelihood estimates of $k$ result in a significant reduction in bias compared to the usual ML estimate. Compared to MLC, our estimate yields improved mean squared error and KL divergence, especially as $k$ increases. The profile maximum likelihood estimate has a slightly smaller bias than our estimate, while the mean squared error and the KL divergence for the two estimates are virtually identical. Unlike both the MLC and MLP estimates, our bias-adjusted maximum likelihood estimate of $k$ is simple to compute in software via existing Weibull ML estimation procedures and does not require the use of the parametric bootstrap. \subsubsection{Type I censored data} For each run of the simulation, we generated $n$ data points from the model Weibull$(k^*, \lambda^* = 1)$ where $n = \{10, 20, 50\}$ and the shape parameter was set to $k^* \in \{0.5, 1, 5, 10\}$. The proportion of uncensored observations was set to $p \in \{0.3, 0.5, 0.7, 0.9\}$. 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: the KL divergence between two Weibull models Weibull($k_0, \lambda_0$) and Weibull($k_1, \lambda_1$) is \begin{align*} \label{eqn:weibull:kl} {\rm KL}( k_0, \lambda_0 || k_1, \lambda_1) &= \exp(-\left(c/\lambda_0\right)^{k_0}) A_1 + \left(\frac{\lambda_0 }{\lambda _1}\right)^{k_1} A_2 \\ &+ \left(1-\frac{k_1}{k_0}\right) A_3 +\log \left(\frac{k_0}{k_1} \left(\frac{\lambda _1}{\lambda_0 }\right)^{k_1}\right)-1 , \end{align*} where \begin{align*} A_1 &= \log \left(\frac{k_1 }{k_0}c^{k_1-k_0} \lambda_0^{k_0} \lambda_1^{-k_1}\right)+\left(\frac{c}{\lambda _1}\right){}^{k_1}+1 ,\\ A_2 &= \Gamma \left(\frac{k_1}{k_0}+1\right)-\Gamma \left(\frac{k_1}{k_0}+1,\left(\frac{c}{\lambda }\right)^k\right) , \\ A_3 &= \text{Ei}\left(-\left(\frac{c}{\lambda_0 }\right)^{k_0}\right)-\gamma , \end{align*} and $\text{Ei}(\cdot)$ is the exponential integral function \begin{equation} \text{Ei}(z) = -\int_{-z}^\infty \frac{\exp(-t)}{t} \, dt . \end{equation} The newly proposed bias adjustment estimate of $k$ (MMLE) was again compared to the standard ML estimate, the conditional maximum likelihood estimate (MLC) proposed by Yang and Xie~\cite{YangXie03} and the profile maximum likelihood estimate (MLP) of Shen and Yang~\cite{ShenYang15}. The third-order profile maximum likelihood estimate had issues with numerical stability for small $n$ and large amounts of censoring sometimes resulting in a negative estimate of $k^*$, hence all the comparisons were made with the second-order variant. We restricted the experiments to exclude data sets where the number of uncensored observations $d (=\sum_i \delta_i) < 2$, as MLC may result in negative estimates of $k$ for $d < 2$. The results of these simulations, averaged over $10^5$ runs for each combination of $(n,p,k^*)$, are shown in Table~\ref{tab:results:censored}. We observe that our MMLE estimate of $k$ is more efficient and less biased than the standard ML of $k$ for all tested values of $(n,p,k^*)$. The conditional maximum likelihood estimate of $k$ is, in general, more biased and has higher mean squared error compared to the MLP and our MMLE estimates. In terms of bias reduction, the profile maximum likelihood estimate of $k$ is virtually identical to our MMLE for $n \geq 30$. For small sample sizes ($n = 20$) and higher levels of censoring ($p \leq 0.5$), the MMLE estimate appears superior to MLC and MLP in terms of bias, mean squared error and KL divergence. Additionally, our MMLE estimate is easily computed without the need for numerical simulation in any software that implements fitting of the Weibull distribution to complete and type I censored data. \begin{table*}[tbph] \scriptsize \begin{center} \begin{tabular}{ccccccccccccccccccccc} \toprule $n$ & $p$ & $k^*$ & \multicolumn{4}{c}{Bias} & & \multicolumn{4}{c}{Mean Squared Error} & & \multicolumn{4}{c}{KL Divergence} \\ & & & ML & MLC & MLP & MMLE & ~ & ML & MLC & MLP & MMLE & ~ & ML & MLC & MLP & MMLE \\ \multirow{16}{*}{20} & \multirow{4}{*}{ 0.3} & 0.5 & 0.115 & 0.021 & 0.004 & {\bf 0.002} & & 0.220 & 0.150 & 0.090 & {\bf 0.090} & & 0.070 & 0.060 & 0.062 & {\bf 0.053}\\ & & 1.0 & 0.228 & 0.040 & 0.005 & {\bf -0.001} & & 0.605 & 0.401 & 0.303 & {\bf 0.301} & & 0.070 & 0.060 & 0.061 & {\bf 0.053}\\ & & 5.0 & 1.156 & 0.214 & 0.033 & {\bf 0.008} & & 14.757 & 9.683 & 7.292 & {\bf 7.248} & & 0.070 & 0.061 & 0.061 & {\bf 0.053}\\ & & 10.0 & 2.251 & 0.374 & 0.058 & {\bf -0.007} & & 55.591 & 36.196 & 32.112 & {\bf 28.799} & & 0.070 & 0.062 & 0.061 & {\bf 0.054}\\ & \multirow{4}{*}{ 0.5} & 0.5 & 0.051 & {\bf 0.001} & 0.001 & -0.003 & & 0.037 & 0.028 & 0.028 & {\bf 0.028} & & 0.059 & 0.056 & 0.053 & {\bf 0.051}\\ & & 1.0 & 0.108 & 0.007 & 0.008 & {\bf 0.000} & & 0.144 & 0.109 & 0.110 & {\bf 0.108} & & 0.059 & 0.055 & 0.052 & {\bf 0.051}\\ & & 5.0 & 0.556 & 0.051 & 0.053 & {\bf 0.014} & & 3.672 & 2.785 & 2.785 & {\bf 2.738} & & 0.060 & 0.056 & 0.053 & {\bf 0.051}\\ & & 10.0 & 1.095 & 0.084 & 0.085 & {\bf 0.009} & & 15.172 & 11.561 & 11.571 & {\bf 11.377} & & 0.060 & 0.056 & 0.053 & {\bf 0.051}\\ & \multirow{4}{*}{ 0.7} & 0.5 & 0.034 & {\bf -0.002} & 0.003 & -0.003 & & 0.019 & {\bf 0.015} & 0.016 & 0.016 & & 0.056 & 0.054 & {\bf 0.051} & 0.051\\ & & 1.0 & 0.075 & 0.003 & 0.013 & {\bf 0.002} & & 0.081 & {\bf 0.065} & 0.068 & 0.066 & & 0.058 & 0.055 & {\bf 0.052} & 0.052\\ & & 5.0 & 0.381 & 0.023 & 0.075 & {\bf 0.017} & & 2.048 & {\bf 1.660} & 1.730 & 1.676 & & 0.058 & 0.055 & {\bf 0.052} & 0.052\\ & & 10.0 & 0.681 & {\bf -0.028} & 0.073 & -0.042 & & 7.474 & {\bf 6.122} & 6.363 & 6.181 & & 0.056 & 0.053 & {\bf 0.050} & 0.050\\ & \multirow{4}{*}{ 0.9} & 0.5 & 0.032 & {\bf 0.001} & 0.012 & 0.001 & & 0.013 & 0.011 & 0.012 & {\bf 0.011} & & 0.062 & 0.058 & 0.054 & {\bf 0.054}\\ & & 1.0 & 0.063 & {\bf 0.001} & 0.023 & 0.002 & & 0.051 & 0.042 & 0.045 & {\bf 0.042} & & 0.062 & 0.058 & 0.054 & {\bf 0.054}\\ & & 5.0 & 0.314 & {\bf 0.001} & 0.111 & 0.008 & & 1.320 & {\bf 1.082} & 1.165 & 1.083 & & 0.061 & 0.057 & 0.054 & {\bf 0.054}\\ & & 10.0 & 0.632 & {\bf 0.006} & 0.226 & 0.019 & & 5.306 & {\bf 4.346} & 4.680 & 4.349 & & 0.062 & 0.058 & 0.054 & {\bf 0.054}\\ \vspace{-2mm} \\ \cmidrule{2-17} \vspace{-2mm} \\ \multirow{16}{*}{20} & \multirow{4}{*}{ 0.3} & 0.5 & 0.114 & 0.020 & 0.003 & {\bf -0.000} & & 0.262 & 0.177 & 0.185 & {\bf 0.102} & & 0.070 & 0.061 & 0.061 & {\bf 0.053}\\ & & 1.0 & 0.231 & 0.042 & 0.006 & {\bf 0.001} & & 0.756 & 0.504 & 0.339 & {\bf 0.337} & & 0.071 & 0.061 & 0.062 & {\bf 0.053}\\ & & 5.0 & 1.149 & 0.207 & 0.036 & {\bf 0.002} & & 16.835 & 11.123 & 9.821 & {\bf 7.900} & & 0.070 & 0.061 & 0.061 & {\bf 0.053}\\ & & 10.0 & 2.312 & 0.427 & 0.108 & {\bf 0.014} & & 82.069 & 54.570 & 59.320 & {\bf 37.053} & & 0.070 & 0.062 & 0.061 & {\bf 0.053}\\ & \multirow{4}{*}{ 0.5} & 0.5 & 0.054 & 0.003 & 0.003 & {\bf -0.000} & & 0.037 & 0.028 & 0.028 & {\bf 0.028} & & 0.059 & 0.056 & 0.053 & {\bf 0.051}\\ & & 1.0 & 0.109 & 0.008 & 0.008 & {\bf 0.000} & & 0.151 & 0.115 & 0.114 & {\bf 0.112} & & 0.060 & 0.056 & 0.053 & {\bf 0.051}\\ & & 5.0 & 0.534 & 0.030 & 0.029 & {\bf -0.009} & & 3.720 & 2.843 & 2.826 & {\bf 2.780} & & 0.060 & 0.056 & 0.053 & {\bf 0.051}\\ & & 10.0 & 1.091 & 0.081 & 0.080 & {\bf 0.004} & & 15.178 & 11.578 & 11.507 & {\bf 11.321} & & 0.060 & 0.056 & 0.053 & {\bf 0.051}\\ & \multirow{4}{*}{ 0.7} & 0.5 & 0.036 & 0.001 & 0.006 & {\bf 0.000} & & 0.020 & {\bf 0.016} & 0.017 & 0.016 & & 0.057 & 0.055 & {\bf 0.051} & 0.052\\ & & 1.0 & 0.072 & 0.001 & 0.011 & {\bf -0.000} & & 0.078 & {\bf 0.064} & 0.067 & 0.065 & & 0.057 & 0.055 & {\bf 0.051} & 0.052\\ & & 5.0 & 0.366 & 0.009 & 0.061 & {\bf 0.003} & & 1.990 & {\bf 1.619} & 1.692 & 1.640 & & 0.058 & 0.055 & {\bf 0.052} & 0.052\\ & & 10.0 & 0.724 & 0.012 & 0.114 & {\bf -0.002} & & 7.781 & {\bf 6.334} & 6.608 & 6.408 & & 0.057 & 0.054 & {\bf 0.051} & 0.051\\ & \multirow{4}{*}{ 0.9} & 0.5 & 0.031 & {\bf 0.000} & 0.011 & 0.001 & & 0.013 & 0.011 & 0.012 & {\bf 0.011} & & 0.062 & 0.058 & 0.054 & {\bf 0.054}\\ & & 1.0 & 0.063 & {\bf 0.001} & 0.023 & 0.002 & & 0.053 & {\bf 0.043} & 0.047 & 0.043 & & 0.062 & 0.058 & 0.054 & {\bf 0.054}\\ & & 5.0 & 0.309 & -0.004 & 0.106 & {\bf 0.003} & & 1.298 & {\bf 1.066} & 1.146 & 1.066 & & 0.061 & 0.057 & 0.054 & {\bf 0.054}\\ & & 10.0 & 0.632 & {\bf 0.006} & 0.225 & 0.019 & & 5.283 & 4.332 & 4.662 & {\bf 4.332} & & 0.061 & 0.058 & 0.054 & {\bf 0.054}\\ \vspace{-2mm} \\ \cmidrule{2-17} \vspace{-2mm} \\ \multirow{16}{*}{30} & \multirow{4}{*}{ 0.3} & 0.5 & 0.065 & 0.007 & {\bf 0.001} & -0.001 & & 0.056 & 0.041 & 0.037 & {\bf 0.037} & & 0.042 & 0.038 & 0.038 & {\bf 0.034}\\ & & 1.0 & 0.133 & 0.016 & 0.004 & {\bf -0.000} & & 0.336 & 0.257 & 0.262 & {\bf 0.179} & & 0.042 & 0.039 & 0.038 & {\bf 0.034}\\ & & 5.0 & 0.653 & 0.068 & 0.009 & {\bf -0.009} & & 5.621 & 4.184 & 3.907 & {\bf 3.788} & & 0.041 & 0.038 & 0.038 & {\bf 0.034}\\ & & 10.0 & 1.334 & 0.161 & 0.045 & {\bf 0.008} & & 23.086 & 17.180 & 16.069 & {\bf 15.252} & & 0.042 & 0.038 & 0.038 & {\bf 0.034}\\ & \multirow{4}{*}{ 0.5} & 0.5 & 0.034 & 0.001 & 0.002 & {\bf -0.000} & & 0.020 & {\bf 0.017} & 0.017 & 0.017 & & 0.037 & 0.036 & 0.034 & {\bf 0.034}\\ & & 1.0 & 0.070 & 0.004 & 0.006 & {\bf 0.001} & & 0.081 & 0.067 & 0.068 & {\bf 0.067} & & 0.037 & 0.036 & 0.034 & {\bf 0.034}\\ & & 5.0 & 0.341 & 0.015 & 0.025 & {\bf 0.000} & & 2.019 & 1.680 & 1.698 & {\bf 1.679} & & 0.037 & 0.036 & 0.034 & {\bf 0.034}\\ & & 10.0 & 0.684 & 0.032 & 0.051 & {\bf 0.002} & & 8.095 & {\bf 6.732} & 6.815 & 6.740 & & 0.037 & 0.036 & 0.034 & {\bf 0.034}\\ & \multirow{4}{*}{ 0.7} & 0.5 & 0.024 & 0.001 & 0.004 & {\bf 0.000} & & 0.012 & {\bf 0.010} & 0.010 & 0.010 & & 0.036 & 0.035 & {\bf 0.034} & 0.034\\ & & 1.0 & 0.047 & 0.001 & 0.007 & {\bf 0.000} & & 0.046 & {\bf 0.040} & 0.041 & 0.041 & & 0.036 & 0.035 & {\bf 0.034} & 0.034\\ & & 5.0 & 0.235 & 0.004 & 0.038 & {\bf 0.001} & & 1.170 & {\bf 1.020} & 1.050 & 1.030 & & 0.036 & 0.035 & {\bf 0.034} & 0.034\\ & & 10.0 & 0.475 & 0.012 & 0.080 & {\bf 0.006} & & 4.662 & {\bf 4.058} & 4.179 & 4.096 & & 0.036 & 0.035 & {\bf 0.034} & 0.034\\ & \multirow{4}{*}{ 0.9} & 0.5 & 0.020 & -0.001 & 0.006 & {\bf -0.000} & & 0.008 & 0.007 & 0.007 & {\bf 0.007} & & 0.038 & 0.036 & 0.035 & {\bf 0.035}\\ & & 1.0 & 0.040 & {\bf -0.000} & 0.014 & 0.000 & & 0.031 & 0.027 & 0.029 & {\bf 0.027} & & 0.038 & 0.036 & 0.035 & {\bf 0.035}\\ & & 5.0 & 0.198 & -0.003 & 0.065 & {\bf -0.001} & & 0.778 & 0.684 & 0.716 & {\bf 0.684} & & 0.038 & 0.036 & 0.035 & {\bf 0.035}\\ & & 10.0 & 0.402 & {\bf -0.000} & 0.137 & 0.005 & & 3.133 & 2.750 & 2.881 & {\bf 2.748} & & 0.038 & 0.037 & 0.035 & {\bf 0.035}\\ \vspace{-2mm} \\ \cmidrule{2-17} \vspace{-2mm} \\ \multirow{16}{*}{40} & \multirow{4}{*}{ 0.3} & 0.5 & 0.047 & 0.004 & 0.002 & {\bf 0.001} & & 0.033 & 0.026 & 0.025 & {\bf 0.025} & & 0.029 & 0.028 & 0.027 & {\bf 0.025}\\ & & 1.0 & 0.094 & 0.009 & 0.004 & {\bf 0.001} & & 0.130 & 0.103 & 0.100 & {\bf 0.099} & & 0.029 & 0.028 & 0.027 & {\bf 0.025}\\ & & 5.0 & 0.465 & 0.039 & 0.014 & {\bf 0.001} & & 3.329 & 2.653 & 2.541 & {\bf 2.528} & & 0.029 & 0.028 & 0.028 & {\bf 0.025}\\ & & 10.0 & 0.937 & 0.084 & 0.035 & {\bf 0.009} & & 13.143 & 10.449 & 10.102 & {\bf 10.053} & & 0.029 & 0.028 & 0.027 & {\bf 0.025}\\ & \multirow{4}{*}{ 0.5} & 0.5 & 0.024 & {\bf 0.000} & 0.001 & -0.001 & & 0.014 & {\bf 0.012} & 0.012 & 0.012 & & 0.027 & 0.026 & 0.026 & {\bf 0.025}\\ & & 1.0 & 0.049 & 0.001 & 0.003 & {\bf -0.001} & & 0.055 & {\bf 0.048} & 0.048 & 0.048 & & 0.027 & 0.026 & 0.026 & {\bf 0.025}\\ & & 5.0 & 0.248 & 0.007 & 0.017 & {\bf -0.001} & & 1.382 & {\bf 1.204} & 1.217 & 1.208 & & 0.027 & 0.026 & 0.026 & {\bf 0.025}\\ & & 10.0 & 0.507 & 0.025 & 0.045 & {\bf 0.009} & & 5.550 & {\bf 4.826} & 4.882 & 4.842 & & 0.027 & 0.026 & 0.026 & {\bf 0.025}\\ & \multirow{4}{*}{ 0.7} & 0.5 & 0.018 & 0.001 & 0.003 & {\bf 0.000} & & 0.008 & {\bf 0.007} & 0.008 & 0.008 & & 0.027 & 0.026 & {\bf 0.025} & 0.025\\ & & 1.0 & 0.034 & {\bf -0.000} & 0.005 & -0.001 & & 0.033 & {\bf 0.030} & 0.030 & 0.030 & & 0.027 & 0.026 & {\bf 0.025} & 0.025\\ & & 5.0 & 0.180 & 0.008 & 0.033 & {\bf 0.006} & & 0.830 & {\bf 0.747} & 0.764 & 0.752 & & 0.027 & 0.026 & {\bf 0.025} & 0.025\\ & & 10.0 & 0.354 & 0.011 & 0.062 & {\bf 0.007} & & 3.306 & {\bf 2.977} & 3.043 & 2.998 & & 0.027 & 0.026 & {\bf 0.025} & 0.025\\ & \multirow{4}{*}{ 0.9} & 0.5 & 0.015 & {\bf 0.000} & 0.005 & 0.000 & & 0.006 & 0.005 & 0.005 & {\bf 0.005} & & 0.027 & 0.027 & 0.026 & {\bf 0.026}\\ & & 1.0 & 0.030 & {\bf 0.000} & 0.010 & 0.000 & & 0.022 & 0.020 & 0.021 & {\bf 0.020} & & 0.027 & 0.026 & 0.026 & {\bf 0.026}\\ & & 5.0 & 0.150 & {\bf 0.002} & 0.052 & 0.003 & & 0.558 & 0.505 & 0.523 & {\bf 0.505} & & 0.028 & 0.027 & 0.026 & {\bf 0.026}\\ & & 10.0 & 0.296 & {\bf -0.001} & 0.099 & 0.002 & & 2.222 & 2.017 & 2.085 & {\bf 2.015} & & 0.027 & 0.027 & 0.026 & {\bf 0.026}\\ \vspace{-3mm} \\ \bottomrule \vspace{+1mm} \end{tabular} \caption{Bias, mean squared error and Kullback--Leibler (KL) divergence for maximum likelihood (ML), conditional maximum likelihood of Yang and Xie (MLC), profile maximum likelihood of Shen and Yang (MLP) and our bias-adjusted maximum likelihood (MMLE) 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*} \bibliographystyle{unsrtnat}

cb107a143cfb02fffe73b76a9f8b282cb5ff018e

\section{Introduction} Regression analysis is a statistical method widely used in many research areas. It is often specified as the normal linear model, where coefficients are linear and the error term follows the normal distribution to simplify the analysis. This specification aims to approximate the state of nature and is often useful in prediction as well as discussing the causality. Among its specifications, variable selection is a central issue in the regression analysis. It is important to select an appropriate set of explanatory variables partly because of the cost of collecting variables. Many methods are proposed for the variable selection problem. In relation to the variable selection problem, this paper focuses on the superset model problem where the linear regression model selects a larger set of variables than the state of nature does (see the example provided in Section \ref{sec:Superset model problem}). The linear regression model tends to choose smaller set of variables when the state of nature is linear in variables due to its least squares loss. However, the nonlinear relationship is more likely and this case may lead to select a larger set of variables, which will be a deficiency of the linear regression model in terms of the data collection cost. To evaluate the superset model problem, this paper utilizes the Bayesian approach, which provides a measure of uncertainty in the form of the posterior probability, and proposes an alternative model that will select the true set of variables when the sample size is large. This paper is organized as follows. Section \ref{sec:Superset model problem} describes the superset model problem by providing an example. Bayes' Theorem is adopted to evaluate the problem in Section \ref{sec:Superset model probability} and two regression models for specification is explained in Section \ref{sec:Two regression models}. Section \ref{sec:Illustrative examples} illustrates the proposed method and discusses its robustness. \section{Superset model problem} \label{sec:Superset model problem} Suppose the continuous response $Y$ is associated with the set of explanatory variables $\bm{x}_{T}$. We are interested in its mean response conditional on $\bm{x}_{T}$. We often assume it to be linear in practice, although it is more likely to be nonlinear in reality. To this end, a regression model is specified as \begin{align*} Y = \phi (\bm{x}_{T}) + \epsilon, \end{align*} where $\epsilon$ is an additive error term with mean zero. The functional form $\phi (\cdot)$, the distribution of the error term, and the true set of explanatory variables $\bm{x}_{T}$ are all unknown. With this model, we make statistical inferences about the conditional mean response by estimating $\phi (\cdot)$ and the set of explanatory variables. Among problems about how this regression model should be specified, the variable selection problem focuses on the set of explanatory variables, based on the dataset $\{ y_{i}, \bm{x}_{i} \}_{i = 1}^{n}$. When the set of explanatory variables is known, the best fit is $E (Y \mid \bm{x}_{T})$ as an estimator of $\phi (\bm{x}_{T})$ when we use the squared loss. The linear regression model assumes $E (Y \mid \bm{x}_{T}) = \bm{x}_{T}^{\prime} \bm{\beta}$, where $\bm{\beta}$ is referred to as the regression coefficient vector. However, in general, $E (Y \mid \bm{x}_{T}) \neq \bm{x}_{T}^{\prime} \bm{\beta}$, contrary to the linearity assumption. For example, suppose \begin{align} E \left[ Y \mid x_{T} \right] = \alpha + \beta_{1} x_{T} + \beta_{2} x_{T}^{2}. \end{align} The linear regression with $(x_{T}, x_{U})$, where $x_{U} = x_{T}^{2}$, is better than the one with $x_{T}$, even though the latter selects the true explanatory variable. This is an example of the superset model problem. On the other hand, when $x_{U}^{\prime}$ is independent of $x_{T}$, $x_{U}^{\prime}$ should not be included in the regression to improve the fit. Above example suggests that the knowledge about association among variables is helpful to examine the superset model problem, and hence the variable selection problem. One approach is to estimate the conditional expectation without linearity and compare it with the one implied by the normal linear model. If they are different and the latter contains more explanatory variables, there exists the superset model problem. Because the dataset at hand is limited, it is difficult to determine whether the superset model problem exists or not. Rather, it is evaluated in a probability form, which is explained in the next section. \section{Superset model probability} \label{sec:Superset model probability} Suppose $\mathcal{M}^{\ast}$ is the set of explanatory variables in the state of nature, which is $\bm{x}_{T}$ when vectorized, and is known for a moment. The current dataset $\{ y_{i}, \bm{x}_{i} \}_{i = 1}^{n}$ is generated from this state of nature independently for each observation $i$ and is observed. Depending on a context, the normal linear model with $\mathcal{M}^{\ast}$ explanatory variables would be a choice if it approximates the state of nature well. In this case, we do not have the superset model problem. On the other hand, a normal linear model with a set of explanatory variables indexed by $\mathcal{M} (\neq \mathcal{M}^{\ast})$ is chosen independent of the state of nature in terms of, say, prediction, where the superset model problem arises when $\mathcal{M} \supset \mathcal{M}^{\ast}$. However, the state of nature is usually unknown and is inferred from the dataset. The uncertainty from inference is evaluated by the posterior probability over possible subsets of explanatory variables. This paper approximates it by assuming a flexible model (see Model \eqref{eq:alternative model} in Section \ref{sec:Two regression models}), which is denoted by $H_{0}$. Then, this posterior probability is calculated via Bayes' theorem, which is given by \begin{align} \Pr \left( \mathcal{M}^{\ast} \mid \{ y_{i}, \bm{x}_{i} \}_{i = 1}^{n}, H_{0} \right) = \frac{ \Pr \left( \{ Y_{i} \}_{i = 1}^{n} \mid \{ \bm{x}_{i} \}_{i = 1}^{n}, H_{0}, \mathcal{M}^{\ast} \right) \Pr \left( \mathcal{M}^{\ast}, H_{0} \right) }{ \sum_{ \tilde{\mathcal{M}} } \Pr \left( \{ Y_{i} \}_{i = 1}^{n} \mid \{ \bm{x}_{i} \}_{i = 1}^{n}, H_{0}, \tilde{\mathcal{M}} \right) \Pr \left( \tilde{\mathcal{M}}, H_{0} \right) }. \label{eq:conditional superset model probability} \end{align} The numerator is the cross product of the marginal likelihood and the prior belief about the set of explanatory variables. When the latter is uniform (which is assumed in the following empirical illustration), the posterior probability is proportional to the marginal likelihood under $H_{0}$. When uncertainty from inference about the normal linear model is evaluated from its posterior probability as well, the overall superset model probability is calculated as \begin{align} \sum_{ \mathcal{M} } \sum_{ \mathcal{M}^{\ast} } I \left( \mathcal{M} \supset \mathcal{M}^{\ast} \right) \Pr \left( \mathcal{M}^{\ast} \mid \{ y_{i}, \bm{x}_{i} \}_{i = 1}^{n}, H_{0} \right) \Pr \left( \mathcal{M} \mid \{ y_{i}, \bm{x}_{i} \}_{i = 1}^{n}, H_{1} \right), \label{eq:superset model probability} \end{align} where $H_{1}$ denotes the normal linear model (see Model \eqref{eq:normal linear regression model} in Section \ref{sec:Two regression models}). We note that the above expression is general enough to include variables (the response and explanatory variables) that are continuous or discrete. The next section specifies two regression models $H_{0}$ and $H_{1}$, where the response is assumed to be continuous for simplicity. \section{Two regression models} \label{sec:Two regression models} First, the linear regression model $H_{1}$ is specified as \begin{align} Y_{i} = \alpha + \bm{x}_{i}^{\prime} \bm{\beta} + \eta_{i}, \quad \eta_{i} \sim N \left( 0, \lambda^{2} \right), \label{eq:normal linear regression model} \end{align} where each of explanatory variables is standardized without loss of generality. To estimate model parameters $( \bm{\beta}, \lambda^{2} )$, the hyper-$g$ prior is assumed. Then, the marginal likelihood is analytically tractable (see \citet{miyawaki-maceachern-21} for example). Second, the model $H_{1}$ that is alternative to the normal linear regression model is specified as \begin{align} Y_{i} = \theta_{x} + \epsilon_{x}, \quad \epsilon_{x} \sim N \left( 0, \sigma_{x}^{2} \right), \label{eq:alternative model} \end{align} given $\bm{x}_{i} = \bm{x}$. The normal error assumption is made because we have no other knowledge on it. Further, it makes the conditional mean estimation simpler, in terms of the number of parameters as well as the computational burden. The main purpose of this semiparametric model is to estimate conditional means of $Y$ in a flexible manner, and to captures the association between $Y$ and $\bm{x}$ in the state of nature as the sample size increases, which can be viewed as as an extreme of the local constant estimation (see, e.g., \citet{fan-gijbels-03} for the local estimation). When the dataset is fixed, the covariate space becomes sparse as its dimension gets larger. Then, the marginal likelihood (hence, the superset model probability) under this alternative model is strongly dependent on the prior specification. To mitigate this influence, this paper takes the $m$-fold cross-validation approach, which is described in details below. The dataset is randomly divided into $m$ groups. One of them is used as the test set, while the remainings belong to the training set. Explanatory variables in the training set are standardized, and those in the test set are standardized by the mean and standard deviation of those in the training set. Let $\mathcal{D}_{0}$ and $\mathcal{D}_{1}$ be the sets of identification numbers of observations which belongs to the training and test sets. More precisely, $\mathcal{D}_{0} = \{ i \mid \text{the $i$-th observation is in the training set}, i = 1, \dots, n \}$ and $\mathcal{D}_{1} = \{ i \mid \text{the $i$-th observation is in the test set}, i = 1, \dots, n \}$. Given a choice of $\mathcal{D}_{0}$ and $\mathcal{D}_{1}$, we construct the prior and conditional marginal likelihood in the following manner. The prior for $\theta_{x}$ in the model \eqref{eq:alternative model} given $\bm{x}_{i} = \bm{x}$ and $i \in \mathcal{D}_{1}$ is assumed as \begin{align} \theta_{x} &\sim N \left( \hat{y}_{x}, t_{x}^{2} \right), \intertext{where $\hat{y}_{x} = \bar{y}_{0} + \bm{x}^{\prime} \hat{\bm{\beta}}$, $\bar{y}_{0}$ is the sample average of the response in $\mathcal{D}_{0}$,} \hat{\bm{\beta}} &= \left( \sum_{i \in \mathcal{D}_{0}} \bm{x}_{i} \bm{x}_{i}^{\prime} \right)^{-1} \sum_{i \in \mathcal{D}_{0}} \bm{x}_{i} y_{i}, \\ t_{x}^{2} &= s^{2} \left\{ \frac{1}{ | \mathcal{D}_{0} | } + \bm{x}^{\prime} \left( \sum_{i \in \mathcal{D}_{0}} \bm{x}_{i} \bm{x}_{i}^{\prime} \right)^{-1} \bm{x} \right\}, \\ \quad s^{2} &= \frac{1}{| \mathcal{D}_{0} | - k - 1} \sum_{i \in \mathcal{D}_{0}} \left( y_{i} - \bar{y}_{0} - \bm{x}_{i}^{\prime} \hat{\bm{\beta}} \right)^{2}. \end{align} When $\mathcal{A}$ is a set, $| \mathcal{A} |$ is the number of elements in the set. This prior is constructed from classical OLS estimates of mean and standard deviation of $Y_{i}$ at $\bm{x}_{i} = \bm{x}$. By using the prior that is obtained from the linear model and using the model that focuses on the local observation, we are able to combine local and global information. It is possible to use other estimates such as corresponding normal linear regression estimates under the hyper-g prior. However, to keep the methodology as simple as possible, we take the above prior specification. Then, we are able to derive the marginal likelihood conditional on the nuisance parameter $\sigma_{x}^{2}$ for each $\bm{x}_{i} = \bm{x}$ and $i \in \mathcal{D}_{1}$. Let $\mathcal{M}_{x} = \{ i \mid \bm{x}_{i} = \bm{x}, i \in \mathcal{D}_{1} \}$ and $n_{x} = | \mathcal{M}_{x} |$. Then, this conditional marginal likelihood is given by \begin{align*} &m^{\ast} \left( \{ Y_{i} \}_{i \in \mathcal{M}_{x} } \mid \{ \bm{x}_{i} \}_{i \in \mathcal{M}_{x} }, \sigma_{x}^{2}, \{ y_{i}, \bm{x}_{i} \}_{i \in \mathcal{D}_{0}} \right) \notag \\ &\hspace{100pt} = \frac{ \tau_{x} }{ \left( \sqrt{2 \pi} \sigma_{x} \right)^{n_{x}} t_{x} } \exp \left\{ -\frac{1}{2} \left( -\frac{\mu_{x}^{2}}{\tau_{x}^{2}} + \frac{ \sum_{i \in \mathcal{M}_{x}} y_{i}^{2} }{\sigma_{x}^{2}} + \frac{ \hat{y}_{x}^{2} }{ t_{x}^{2} } \right) \right\}, \\ &\mu_{x} = \tau_{x}^{2} \left( \frac{ \sum_{i \in \mathcal{M}_{x}} y_{i} }{\sigma_{x}^{2}} + \frac{\hat{y}_{x}}{t_{x}^{2}} \right), \quad \tau_{x}^{2} = \left( \frac{n_{x}}{\sigma_{x}^{2}} + \frac{1}{t_{x}^{2}} \right)^{-1}. \end{align*} The full Bayes analysis specifies a prior on the nuisance parameter $\sigma_{x}^{2}$ as well. However, because the data are sparse at $\bm{x}$, how we specify it affects the (unconditional) marginal likelihood much. To mitigate this problem, this paper takes the empirical Bayes approach. The marginal likelihood for each $\bm{x}_{i} = \bm{x}$ is the conditional marginal likelihood $m^{\ast} ( \{ Y_{i} \}_{i \in \mathcal{M}_{x} } \mid \{ \bm{x}_{i} \}_{i \in \mathcal{M}_{x} }, \sigma_{x}^{2}, \{ y_{i}, \bm{x}_{i} \}_{i \in \mathcal{D}_{0}} )$ maximized over $\sigma_{x}^{2}$. More precisely, \begin{align*} m \left( \{ Y_{i} \}_{i \in \mathcal{M}_{x} } \mid \{ \bm{x}_{i} \}_{i \in \mathcal{M}_{x} }, \{ y_{i}, \bm{x}_{i} \}_{i \in \mathcal{D}_{0}} \right) \equiv \max_{\sigma_{x}^{2}} m^{\ast} \left( \{ Y_{i} \}_{i \in \mathcal{M}_{x} } \mid \{ \bm{x}_{i} \}_{i \in \mathcal{M}_{x} }, \sigma_{x}^{2}, \{ y_{i}, \bm{x}_{i} \}_{i \in \mathcal{D}_{0}} \right). \end{align*} See the next section for this maximization in details. By multiplying it over all distinct $\bm{x}$ in $\mathcal{D}_{1}$ and taking the geometric mean, we have the marginal likelihood for $\mathcal{D}_{1}$ per one observation, which is given by \begin{align*} \left\{ \prod_{\bm{x}} m \left( \{ Y_{i} \}_{i \in \mathcal{M}_{x} } \mid \{ \bm{x}_{i} \}_{i \in \mathcal{M}_{x} }, \{ y_{i}, \bm{x}_{i} \}_{i \in \mathcal{D}_{0}} \right) \right\}^{1 / | \mathcal{D}_{1} |}. \end{align*} The geometric mean is to take care of different sample sizes in different test sets. Finally, we repeat above process until all $m$ groups are used as the test set and calculate above marginal likelihood for each test group selection. After averaging $m$ marginal likelihoods, we raise it to the power of $n$ to obtain the final marginal likelihood estimate for the model \eqref{eq:alternative model}. The robustness of this approach is of interest. The approach will be more useful if we know the upper and lower bounds of the superset model probability under $H_{0}$ when its specification changes. Two points are discussed regarding robustness. First, we consider the robustness to the number of folds in the cross-validation. In the following empirical analysis in Section \ref{sec:Illustrative examples}, we use the 10-fold cross-validation. When the number of folds changes from 2 to 15, we see the resulting probability does not change much with the diabetes dataset (see Figure \ref{fig:Superset model probability change as the number of folds increases.}). \begin{figure}[ht] \centering \includegraphics[width=12cm, clip, keepaspectratio]{robust_fold.pdf} \caption{Superset model probability change as the number of folds increases. (The dotted line indicates the 10-fold cross-validation, which will be used in Section \ref{sec:Illustrative examples})} \label{fig:Superset model probability change as the number of folds increases.} \end{figure} Second, it is important to check the robustness to the prior. One approach would be to use the $\epsilon$-contamination class prior (see Section 4.7.4 of \citet{berger-85}), and to show the sensitivity of the marginal likelihood, which will be our future analysis. \section{Maximize the marginal likelihood} \label{sec:Maximize the marginal likelihood} Letting \begin{align} \bar{y}_{x} &= \frac{1}{n_{x}} \sum_{i \in G_{x}} y_{i}, \\ s_{y}^{2} &= \frac{1}{n_{x}} \sum_{i \in G_{x}} \left( y_{i} - \bar{y}_{x} \right)^{2} = \frac{1}{n_{x}} \sum_{i \in G_{x}} y_{i}^{2} - \bar{y}_{x}^{2}, \end{align} the local marginal likelihood function is characterized by Theorem \ref{thm:maximization}. \begin{thm} The local marginal likelihood function has extremal values at positive solutions to the cubic equation \eqref{eq:cubic equation} given in the proof when $s_{y}^{2} > 0$; at 0 when $s_{y}^{2} = 0, t_{x}^{2} - ( \bar{y}_{x} - \hat{y}_{x} )^{2} \geq 0$; and at 0 and $- t_{x}^{2} + ( \bar{y}_{x} - \hat{y}_{x} )^{2}$ when $s_{y}^{2} = 0, t_{x}^{2} - ( \bar{y}_{x} - \hat{y}_{x} )^{2} < 0$. \label{thm:maximization} \end{thm} See Appendix \ref{sec:Proof of Theorem} for its proof. Table \ref{table:Local marginal likelihood function} summarizes the result. \begin{table} \centering \caption{Local marginal likelihood function} \label{table:Local marginal likelihood function} \begin{tabular}{lccc} \toprule & $\sigma_{x}^{2} \to 0$ & $0 < \sigma_{x}^{2} < \infty$ & $\sigma_{x}^{2} \to \infty$ \\ \midrule $s_{y}^{2} > 0$ & 0 & \multirow{2}{*}{See Lemma \ref{lem:extremal values}} & \multirow{2}{*}{0} \\ $s_{y}^{2} = 0$ & $\frac{1}{ \sqrt{2 \pi} t_{x}} \exp \left\{ -\frac{1}{2 t_{x}^{2}} \left( \bar{y}_{x} - \hat{y}_{x} \right)^{2} \right\}$ & & \\ \bottomrule \end{tabular} \end{table} By this theorem, we are able to set a value of $\sigma_{x}^{2}$ to maximize the marginal likelihood, instead of placing a prior on it. \section{Illustrative example} \label{sec:Illustrative examples} The diabetes data (see \citet{efron-etal-04}) are used to illustrate our method. This dataset contains 442 observations. For the analysis below, we use the logarithm of the diabetes progression measure as the response and use remaining 10 variables are used as exlanatory variables. The proposed method is applied, and the superset model probability for this dataset with 10-fold cross-validation is estimated to be 22.37\%. As discussed by \citet{maceachern-miyawaki-22}, the diabetes dataset seems to be collected from least two different sources. In particular, the precision of two explanatory variables (the blood pressure and fourth blood serum measurement) consists of a mix of finer and coarser observations. When the data are divided into two groups by this precision, \citet{maceachern-miyawaki-22} suggests these two datasets have different characteristics. This conclusion is also confirmed in terms of the superset model probability. When the dataset for observations with finer variables is used, it is 22.19\%. When, on the other hand, that for observations with coarser variables is used, it is 10.72\%. The superset model problem is more likely to occur with the former dataset than the latter one, which would be due to the difference in characteristics of these two datasets.

ed44e396ce5c3fe658e0018ad4f941454023985d

\section{introduction} The main results of this article extend principal results of \cite{HHLM08} on convex polynomials in freely noncommuting variables to the matrix-valued case and of \cite{JKMMP21a} on $xy$-convex polynomials to the matrix-valued setting in any finite number of freely noncommuting variables. Fix a positive integer ${\tt{g}}.$ Given a positive integer $d$ and $d\times d$ matrices, $A_0,A_1,\dots,A_{\tt{g}},$ the expression \[ L_A(x) = A_0 - \sum_{j=1}^{{\tt{g}}} A_j x_j \] is a \df{linear pencil}, where $A = (A_0,A_1,\dots,A_g).$ \index{$L_A$} In the case the $A_j$ are hermitian the pencil is \df{hermitian} and, in this case, it is typically assumed that $A_0$ is positive definite. When $L_A$ is hermitian and $x\in \mathbb{R}^g,$ the matrix $L_A(x)$ is hermitian and \[ L_A(x)\succeq 0 \] is a \df{linear matrix inequality (LMI)}. Here $T\succeq 0$ indicates that the hermitian matrix $T$ is positive semidefinite. The (scalar) solution, or \df{feasible}, set of a hermitian pencil $L_A,$ \[ \mathcal D_A[1] =\{x\in \mathbb{R}^{{\tt{g}}}: L_A(x) \succeq 0\}, \] is a \df{spectrahedron}. Because $L_A$ is affine linear, it is evident that \df{$\mathcal D_A[1]$} is convex. Spectrahedra figure prominently in numerous engineering applications. They are fundamental objects in semidefinite programming in convex optimization and in real algebraic geometry. Given $d\times d$ hermitian matrices $A_0,A_1,\dots,A_{{\tt{g}}},B_1, \dots, B_{{\tt{h}}}, C_{pq}, 1 \le p \le {\tt{g}}, 1 \le q \le {\tt{h}},$ the expression \begin{equation*} L(x,y) = A_0 - \sum_{j=1}^{{\tt{g}}} A_j x_j - \sum_{k=1}^{{\tt{h}}} B_k y_k - \sum_{p,q=1}^{{\tt{g}},{\tt{h}}} C_{pq} x_p y_q, \end{equation*} is an \df{$xy$-pencil}. When all the coefficient matrices are hermitian, $L$ is a \df{hermitian $xy$-pencil}. For a hermitian $xy$-pencil, the inequality $L(x,y)\succeq 0$ is a \df{Bilinear Matrix Inequality (BMI)}. Bilinear matrix inequalities appear in robust control. See for instance \cite{KSVdS,SGL,vAB} and the references therein and the MATLAB toolbox, \\ {\url{https://set.kuleuven.be/optec/ Software/bmisolver-a-matlab-package-for-solving-optimization-problems -with-bmi-constraints.}} It is natural from multiple perspectives to consider the fully matricial analogs of LMIs and BMIs. For positive integers $n,$ let \df{$\mathbb{S}_n({\mathbb C})$} denote the set of $n\times n$ hermitian matrices and let \df{$\mathbb{S}_n({\mathbb C}^\vg)$} denote the set of ${\tt{g}}$-tuples from $\mathbb{S}_n({\mathbb C}).$ Given $X=(X_1,\dots,X_{\tt{g}})\in \mathbb{S}_n({\mathbb C}^\vg),$ let \[ L_A(X) =A_0\otimes I_n -\sum A_j\otimes X_j \] and let \[ \mathcal D_A[n]=\{X\in \mathbb{S}_n({\mathbb C}^\vg): L_A(X)\succeq 0\}. \] The sequence $\mathcal D_A=(\mathcal D_A[n])_n$ is known as a \index{$\mathcal D_A$} \df{free spectrahedron} or \df{LMI domain}. While $\mathcal D_A[1]$ does not determine $A,$ up to unitary equivalence, the free spectrahedra $\mathcal D_A$ does. Free spectrahedra are \df{matrix convex}, meaning \begin{enumerate}[(i)] \item $\mathcal D_A$ is closed with respect to isometric compressions: if $X\in \mathcal D_A[n]$ and $V:\mathbb{C}^m\to \mathbb{C}^n$ is an isometry, then $V^*XV\in \mathcal D_A[m],$ where \[ V^*XV = V^*(X_1,\dots,X_{\tt{g}})V = (V^*X_1V, \dots, V^*X_{\tt{g}} V); \] and \item $\mathcal D_A$ is closed under direct sums: if $X\in \mathcal D_A[n]$ and $Y\in \mathcal D_A[m],$ then $X\oplus Y\in \mathcal D_A[n+m],$ where \[ (X\oplus Y)_j =\begin{pmatrix} X_j&0\\0& Y_j \end{pmatrix}. \] \end{enumerate} In particular each $\mathcal D_A[n]$ is convex in the ordinary sense. Free spectrahedra appear in the theories of completely positive maps and operator systems and spaces \cite{paulsen,pisier}. They appear in systems engineering problems governed by a signal flow diagram as explained in \cite{convert-to-matin,IMA-survey,CHSY}. They also produce tractable natural relaxations for optimizing over spectrahedra; e.g., the matrix cube problem \cite{BtN,DDSS,matcube}, which can be NP hard, but whose canonical free spectrahedral relaxation is a semidefinite program (SDP). The fully matricial analog of BMIs is described below, after the introduction of polynomials in freely noncommuting variables. \subsection{Free polynomials} The two types of partial convexity considered in this article are described in terms of free polynomials. Fix freely noncommuting variables $\chi_1,\dots,\chi_k.$ Given a word \begin{equation} \label{d:w} w=\chi_{i_1} \cdots \chi_{i_\ell} \end{equation} in these variables and $T\in \mathbb{S}_n({\mathbb C}^k),$ let \[ w(T)=T^w = T_{i_1}\cdots T_{i_\ell}. \] Let $\mathcal W$ denote the collection of words in the variables $\chi.$ A \df{$d \times d$ matrix-valued free polynomial} is an expression of the form, \[ p(\chi)= \sum_{w\in \mathcal W} p_w w, \] where the sum is finite and the $p_w \in M_d({\mathbb C})$. The free polynomial $p$ is naturally \df{evaluated} at $T\in \mathbb{S}_n({\mathbb C}^k)$ as \[ p(T) =\sum p_w \otimes T^w. \] There is a natural \df{involution ${}^*$} on free polynomials that reverses the order of products in words so that, for $w$ in equation~\eqref{d:w}, \[ w^* = \chi_{i_\ell}\cdots \chi_{i_1}; \] and such that \[ p^* =\sum p_w^* w^*. \] This involution is compatible with the adjoint operation on matrices, \[ p(T)^* =p^*(T). \] A free polynomial $p$ is {\bf hermitian} \index{hermitian polynomial} if $p^*=p$; equivalently, if $p(T)^*=p(T)$ for all $n$ and $T\in \mathbb{S}_n({\mathbb C}^k).$ From here on we often omit the adjectives matrix and free and simply refer to matrix-valued free polynomials as polynomials, particularly when there is no possibility of confusion. Since the involution fixes the variables, $\chi_j^*=\chi_j,$ we refer to $\chi_1,\dots,\chi_k$ as \df{hermitian variables}. In Subsection~\ref{s:BVMM}, non-hermitian variables naturally appear. \subsection{Partial convexity} Both types of partial convexity considered in this article involve partitioning freely noncommuting variables into two classes $x_1,\dots,x_{\mu}$ and $y_1,\dots,y_{\mu}.\footnote{ For the results here, there is no loss in generality in assuming the number of $x$ and $y$ variables is the same.}$ \subsubsection{$xy$-convexity} Since matrix multiplication does not commute, we now update the definition of an $xy$-pencil as follows. A matrix-valued free polynomial of the form \begin{equation*} L(x,y) = A_0 - \sum_{j=1}^{\mu} A_j x_j - \sum_{k=1}^{\mu} B_k y_k - \sum_{p,q=1}^{\mu} C_{pq} x_p y_q - \sum_{p,q=1}^{\mu} D_{qp} y_q x_p, \end{equation*} where $A_j,B_k,C_{pq},D_{qp}$ are all matrices of the same size, is an \df{$xy$-pencil}. The pencil $L$ is naturally evaluated at a tuple $(X,Y)\in \mathbb S_n({\mathbb C}^\vmu) \times \mathbb S_n({\mathbb C}^\vmu)$ as \[ L(X,Y) = A_0 \otimes I_n - \sum_{j=1}^{\mu} A_j \otimes X_j - \sum_{k=1}^{\mu} B_k \otimes Y_k - \sum_{p,q=1}^{\mu} C_{pq} \otimes X_p Y_q - \sum_{p,q=1}^{\mu} D_{qp} \otimes Y_q X_p. \] When the $A_j$ and $B_k$ are hermitian and $D_{qp}=C_{pq}^*,$ the pencil $L$ is a \df{hermtian $xy$-pencil} and $L(X,Y)\succeq 0$ is the matricial analog of a BMI. Assuming, as we usually do, $A_0$ is positive definite, writing $\Sigma=(A_j,B_j, C_{ij})$ and $L_\Sigma=L,$ let \[ \mathcal D_\Sigma[n] =\{(X,Y)\in \mathbb{S}_n^{\mu}\times \mathbb{S}_m^{\mu}: L_\Sigma(X,Y)\succeq 0\} \] and let \df{$\mathcal D_\Sigma$} denote the sequence $(\mathcal D_\Sigma[n])_n.$ The set $\mathcal D_\Sigma$ is \df{$xy$-convex}, meaning $\mathcal D_\Sigma$ is \begin{enumerate}[(i)] \item closed under direct sums; and \item if $(X,Y)\in \mathcal D[n]$ and $V:\mathbb{C}^m\to \mathbb{C}^n$ is an isometry such that $V^*(X_iY_j)V=V^*X_iVV^*Y_jV,$ for all $i,j,$ then $V^*(X,Y)V\in \mathcal D_\Sigma[m].$ \end{enumerate} A tuple $((X,Y),V)$ where $(X,Y)\in \mathbb S_n({\mathbb C}^\vmu)\times \mathbb S_n({\mathbb C}^\vmu)$ and $V:\mathbb{C}^m\to \mathbb{C}^n$ is an isometry such that $V^*(X_iY_j)V=V^*X_iVV^*Y_jV,$ for all $i,j,$ is an \df{$xy$-pair}. A hermitian matrix-valued free polynomial $p(x,y)$ is \df{$xy$-convex} if \[ p(V^*(X,Y)V)\preceq (I_d \otimes V)^*p(X,Y) (I_d \otimes V) \] for all xy-pairs $((X,Y),V).$ It is nearly immediate that, if $p$ is $xy$-convex, then the positivity set of $-p,$ \[ \mathcal D_{-p}=\{(X,Y): p(X,Y)\preceq 0\}, \] is also $xy$-convex. \index{$\mathcal D_{-p}$} Theorem ~\ref{t:introxyconvexp} below provides an algebraic certificate characterizing $xy$-convex polynomials. When $d=\mu=1,$ it reduces to \cite[Theorem~1.4]{JKMMP21}. \begin{theorem} \label{t:introxyconvexp} Suppose $p(x,y)$ is a hermitian $d \times d$ matrix-valued polynomial. If $p$ is $xy$-convex, then there exists a hermitian $d \times d$ matrix-valued $xy$-pencil $\lambda,$ a positive integer $N$ and an $N \times d$ matrix-valued $xy$-pencil $\Lambda$ such that \begin{equation} \label{eq:xy-conv} p(x,y) = \lambda(x,y) +\Lambda(x,y)^* \Lambda(x,y). \end{equation} In particular, $-p$ is the Schur complement of a Hermitian $xy$-pencil and $\mathcal D_{-p}$ is the feasible set of the BMI, \[ \begin{pmatrix} I & \Lambda(x,y)\\\Lambda(x,y)^* & -\lambda(x,y) \end{pmatrix} \succeq 0. \] The converse is easily seen to be true. \end{theorem} A proof of Theorem \ref{t:introxyconvexp} is contained in the proof of Proposition \ref{p:more-is-true} given in Section 3. \subsubsection{$a^2$-convexity} To maintain consistency with the literature, we now switch to freely noncommuting variables $a_1,\dots,a_\mu$ and $x_1\, \dots,x_{\mu}.$ A $d\times d$ matrix-valued hermitian polynomial $p(a,x)$ is \df{convex in $x$} if for each positive integer $n,$ each $A \in \mathbb S_n({\mathbb C}^\mu),$ each $X,Y \in \mathbb S_n({\mathbb C}^\mu)$ and each $0 < t < 1,$ one has \[ p(A, tX + (1-t)Y) \preceq t p(A,X) + (1-t)p(A,Y). \] A canonical example of a convex in $x$ polynomial is a hermitian \df{linear in $x$ pencil}; that is, a hermitian polynomial that is affine linear in $x.$ There is a fruitful alternate characterization of convexity in $x.$ A tuple $((A,X),V)$ where $(A,X)\in \mathbb S_n({\mathbb C}^\vmu)\times\mathbb S_n({\mathbb C}^\vmu)$ and $V:\mathbb{C}^m\to\mathbb{C}^n$ is an isometry is an \df{$a^2$-pair} if $V^*A_i^2V= (V^*A_iV)^2$ for each $1\le i\le \mu.$ Equivalently $((A,X),V)$ is an $a^2$-pair if $\operatorname{ran} V$ reduces $A.$ As we will see in Proposition~\ref{thm:partialconvexity}, a hermitian polynomial $p$ is convex in $x,$ or \df{$a^2$-convex}, if and only if \[ p(V^*(A,X)V) \preceq (I_d\otimes V^*)p(A,X) (I_d\otimes V) \] for all $a^2$-pairs $((A,X),V).$ Theorem~\ref{t:introxyconvexp} and Theorem~\ref{thm:partialconvexpolys} below -- the latter of which is a matrix polynomial version of \cite[Theorem~1.5]{HHLM08} and \cite[Corollary~1.3]{JKMMP21a} -- are the main results of this article. \begin{theorem} \label{thm:partialconvexpolys} Suppose $p(a,x)$ is a $d\times d$ matrix-valued hermitian polynomial. If $p(a,x)$ is convex in $x,$ then there exists a $d\times d$ matrix-valued hermitian linear in $x$ pencil $L$ a positive integer $N$ and a $N\times d$ matrix-valued polynomial $\Lambda$ that is linear in $x$ such that \[ p(a,x) = L(a,x) + \Lambda(a,x)^* \Lambda(a,x). \] In particular, $p$ has degree at most two in $x$ and $\mathcal D_{-p}$ is the feasible set of the affine linear in $x$ matrix inequality \[ \begin{pmatrix} I & \Lambda(a,x)\\ \Lambda(a,x)^* & - L(a,x) \end{pmatrix} \succeq 0. \] The converse is evidently true. \end{theorem} A proof of Theorem \ref{thm:partialconvexpolys} is given in Section~\ref{s:a2-convexity}. Proposition~\ref{p:more-is-true} below describes the relationship between $xy$-convexity and separate convexity in $x$ and $y$. It also extends \cite[Theorem~1.4]{JKMMP21} to both several $x$ and $y$ variables and matrix-valued polynomials. \begin{proposition} \label{p:more-is-true} Let $p(x,y)$ be a $d \times d$ matrix-valued hermitian polynomial. The following statements are equivalent. \begin{enumerate}[(i)] \item \label{i:more1} $p$ is $xy$-convex \item \label{i:more2} $p$ is convex in $x$ and $y$ separately. \item \label{i:more3} $p$ has the form given in equation \eqref{eq:xy-conv}. \end{enumerate} In particular, $p$ is $xy$-convex if and only if $p$ is convex in $x$ and $y$ separately. \end{proposition} A proof of Proposition \ref{p:more-is-true} is given in Section 3. \begin{remark}\rm An example in the appendix of the arxiv version of \cite{JKMMP21a} shows that there is not a local version of Proposition~\ref{p:more-is-true}. That is, as a local statement, separate convexity need not imply $xy$-convexity. \end{remark} \section{Partially convex hermitian matrix-valued NC polynomials} \label{s:a2-convexity} This section contains a proof of Theorem~\ref{thm:partialconvexpolys} and is organized as follows. Subsection~\ref{s:a2convex} presents alternate formulations of $a^2$-convexity. Needed versions of Amitsur's no polynomial identities results are collected in Subsection~\ref{s:amitsur}. The border vector middle matrix representation for a type of Hessian for polynomials in $a,x$ of degree two in $x$ is reviewed in Subsection~\ref{s:BVMM}. The proof of Theorem~\ref{thm:partialconvexpolys} concludes in Subsection~\ref{s:a2endproof}. Subsection~\ref{s:a2cors} contains two corollaries that apply to $xy$-convex polynomials. \subsection{Alternate formulations of convexity} \label{s:a2convex} The proof of Theorem~\ref{thm:partialconvexpolys} makes use of the following characterization of $a^2$-convex polynomials. It parallels \cite[Proposition~4.1]{JKMMP21} for $xy$-convex polynomials for $\mu=1$ and, to some extent, appears as \cite[Proposition~1.5]{JKMMP21a}. It also borrows liberally from the ideas in \cite{royal}. \begin{proposition} \label{thm:partialconvexity} For a $d \times d$ matrix-valued hermitian polynomial $p(a,x),$ the following statements are equivalent. \begin{enumerate}[(i)] \item \label{i:a2convex1} The polynomial $p$ is convex in $x;$ \item \label{i:a2convex2} If $((A,X),V)$ is an $a^2$-pair, then \[ (I_d \otimes V)^*\, p(A,X) \, (I_d \otimes V) \succeq p(V^*(A, X)V); \] \item \label{i:a2convex3} For each tuple $(A,X) \in \mathbb S_n({\mathbb C}^\vmu)\times\mathbb S_n({\mathbb C}^\vmu),$ each positive integer $m$ and all tuples $\alpha,\delta\in \mathbb S_m({\mathbb C}^\mu)$ and $\beta \in M_{n,m}({\mathbb C}^\mu),$ \[ (I_d \otimes W)^* p(R,S) (I_d \otimes W) \succeq p \left(W^* (R,S) W \right) \] where $W^* = \begin{pmatrix} I_n & 0 \end{pmatrix} \in M_{n, n+m}({\mathbb C}),$ \[ R = \left(\begin{pmatrix} A_1 & 0 \\ 0 & \alpha_1 \end{pmatrix},\dots, \begin{pmatrix} A_\mu & 0 \\ 0 & \alpha_\mu \end{pmatrix}\right) \in \mathbb S_{n+m}({\mathbb C}^\mu) \] and \[ S = \left(\begin{pmatrix} X_1 & \beta_1 \\ \beta_1^* & \delta_1 \end{pmatrix},\dots, \begin{pmatrix} X_\mu & \beta_\mu \\ \beta_\mu^* & \delta_\mu \end{pmatrix}\right) \in \mathbb S_{n+m}({\mathbb C}^\mu). \] \end{enumerate} \end{proposition} \begin{proof} To prove \ref{i:a2convex2} implies \ref{i:a2convex1}, let $A, X,Y \in \mathbb S_n({\mathbb C}^\vmu)$ and $t \in [0,1]$ be given. Let \[ \widehat A = \begin{pmatrix} A & 0 \\ 0 & A \end{pmatrix}, \ \ \ \widehat X = \begin{pmatrix} X & 0 \\0 & Y \end{pmatrix}, \ \ \ V= \begin{pmatrix} \sqrt{t} \, I_n & \sqrt{1 - t} \, I_n \end{pmatrix}^*. \] In particular, $((\widehat A,\widehat X),V)$ is an $a^2$-pair. Thus, \begin{equation*} \begin{split} p(A, tX + (1-t) Y) & = p(V^*( \widehat A, \widehat X)V) \\ & \preceq (I_d \otimes V)^*\, p(\widehat A,\widehat X)\, (I_d \otimes V) \\ & = (I_d \otimes V)^* \begin{pmatrix} p(A, X) & 0 \\ 0 & p(A, Y) \end{pmatrix} (I_d \otimes V) \\ & = t p(A,X) + (1-t) p(A, Y), \end{split} \end{equation*} where the inequality is a consequence of the hypothesis. Hence $p$ is convex in $x.$ Now suppose item~\ref{i:a2convex3} holds and let an $a^2$-pair $((A,X),V)$ be given. Since $V: {\mathbb C}^m \rightarrow {\mathbb C}^n$ is an isometry whose range $M$ reduces $A_j,$ the matrix representations of $V,$ $A_j$ and $X_j$ with respect to the decomposition ${\mathbb C}^n = M \oplus M^{\perp}$ take the forms \[ \begin{pmatrix} I_M \\ 0 \end{pmatrix}, \begin{pmatrix} A_j\restriction_M & 0 \\ 0 & A_j\restriction_{M^\perp} \end{pmatrix} \text{ and } \begin{pmatrix} P_{M} X_j\restriction_M & Y_j \\ Y_j^* & P_{M^{\perp}} X_j\restriction_{M^\perp} \end{pmatrix} \] respectively, where $P_M$ denotes the orthogonal projection of ${\mathbb C}^n$ onto $M.$ The conclusion of item~\ref{i:a2convex2} now follows by identifying $M$ with ${\mathbb C}^m$ and observing that, under this identification, the operators $V,$ $A$ and $X$ have the same form as $W,$ $R$ and $S$ in the hypothesis. Hence item~\ref{i:a2convex3} implies item~\ref{i:a2convex2}. It remains to prove \ref{i:a2convex1} implies \ref{i:a2convex3}. To this end, let \[ \widehat {S} = \left(\begin{pmatrix} X_1 &- \beta_1 \\ -\beta_1^* & \delta_1 \end{pmatrix},\dots, \begin{pmatrix} X_\mu & -\beta_\mu \\ -\beta_\mu^* & \delta_\mu \end{pmatrix}\right). \] By the convex in $x$ hypothesis, it follows that \begin{equation} \label{e:a2convex1} \begin{pmatrix} p(A, X) & 0 \\ 0 & p(\alpha, \delta) \end{pmatrix} = p \left(R, \frac{1}{2} (S + \widehat{S})\right) \preceq \frac1 2 (p(R,S) + p(R, \widehat{S})). \end{equation} Multiplying the inequality of equation~\eqref{e:a2convex1} by $(I_d \otimes W)^*$ on the left and $(I_d \otimes W)$ on the right gives \[ p(A,X) = p(W^*(R,S)W) \preceq \frac{1}{2} (I_d \otimes W)^* \, [p(R,S) + p(R, \widehat{S})] \, (I_d \otimes W). \] Thus, to complete the proof, it suffices to show. \begin{equation} \label{e:complete} (I_d \otimes W)^*\, p(R,\widehat{S}) \, (I_d \otimes W) = (I_d \otimes W)^* \, p(R,S) \, (I_d \otimes W). \end{equation} To this end, let \[ U=\begin{pmatrix} I & 0 \\0 & -I \end{pmatrix} \] and note $(R,\widehat{S}) = U^* (R,S)U.$ Consequently, \[ p(R,\widehat{S}) = (I_d \otimes U^*) p(R,S) (I_d \otimes U), \] and equation~\eqref{e:complete} follows. \end{proof} \subsection{Faithful representations} \label{s:amitsur} \begin{proposition} \label{p:a2convex-deg2} Suppose $p(a,x)$ is a hermitian polynomial. If $p$ is convex in $x,$ then the degree of $p$ in $x$ is at most two. \end{proposition} \begin{proof} Let $d$ denote the size of $p.$ Thus $p = \sum_w p_w w$ for some $p_w\in M_d({\mathbb C}).$ For $\gamma \in {\mathbb C}^d,$ define the polynomial $p_{\gamma}$ by $p_\gamma = \sum_w (\gamma^* p_w \gamma ) \, w.$ Since $p$ is hermitian, it follows that $p_\gamma$ is a hermitian polynomial with scalar coefficients. Also convexity of $p$ in $x$ implies the convexity of $p_{\gamma}$ in $x.$ Hence, by \cite[Corollary~1.3]{JKMMP21a},\footnote{The same result, but with real scalars, appears as \cite[Theorem~1.4]{HHLM08}} for each $\gamma\in {\mathbb C}^d,$ the degree of $p_\gamma$ in $x$ is at most two. Suppose the word $w = w(a,x)$ is such that $p_w \neq 0.$ Since the scalar field is ${\mathbb C},$ it follows that there exists a $\gamma\in {\mathbb C}^d$ such that $\gamma^*p_w\gamma\ne 0.$ Since $p_\gamma$ has degree at most two in $x,$ it follows that $w(a,x)$ has degree at most two in $x.$ Hence $p$ has degree at most two in $x.$ \end{proof} The following lemma is a variant of the Amitsur-Levitski Theorem. For the convenience of the reader, a sketch of a proof is included. \begin{lemma} \label{lem:vanishingpoly} If $p(a)$ is a polynomial of degree at most $m\ge 0$ in the freely noncommuting variables $a_1,\dots,a_\mu$ and if there is an $n\ge N(\mu,m) := \sum_{j=0}^{m} \mu^j$ and a nonempty open set $\mathcal U\subseteq \mathbb S_n({\mathbb C}^\vmu)$ such that $p(U)=0$ for all $U\in \mathcal U,$ then $p=0.$ \end{lemma} \begin{proof} Arguing by contradiction, suppose $p\ne 0,$ but there is an $n\ge N(\mu,m)$ and a nonempty open subset $\mathcal U\subseteq \mathbb S_n({\mathbb C}^\vmu)$ on which $p$ vanishes. In this case, there is no loss of generality assuming the degree of $p$ is $m.$ Since $p$ vanishes on an open subset of $\mathbb S_n({\mathbb C}^\vmu),$ it vanishes on all of $\mathbb S_n({\mathbb C}^\vmu).$ For the moment, assume $n=N.$ Let $H$ denote the Hilbert space with orthonormal basis $\mathcal W,$ the words of length at most $m$ in the variables $a.$ Hence $\dim H=N.$ Define linear maps $S_j,$ for $j=1,\dots,{\tt{g}},$ on $H$ by $S_j w= a_jw,$ if $w\in \mathcal W$ has length strictly less than $m,$ and $S_jw=0$ if $w$ has length $m.$ Observe that $S_j^*w$ has length strictly less than the length of $w\in \mathcal W.$ Let $T_j=S_j + S_j^*.$ Thus $T=(T_1,\dots,T_{\mu}) \in \mathbb{S}_N({\mathbb C}^\mu).$ % A straightforward computation shows, when $m\ge 1$ \[ p(T)\varnothing = \sum_{j=0}^{m-1} q_j + p_m, \] where $q_j$ are homogeneous polynomials of degree $j$ and $p_m$ is the homogeneous of degree $m$ part of $p.$ On the other hand, when $m=0,$ \[ p(T)\varnothing = p_\varnothing \varnothing. \] Since the set $\{q_0, q_1,\dots,q_{m-1},p_m\}\subseteq H$ is linearly independent and, by assumption, $p(T)\varnothing=0,$ it follows that $p_m=0,$ contradicting the assumption that the degree of $p$ is $m.$ To complete the proof, if $n>N,$ then replace the tuple $T$ by $R=T\oplus 0$ and $\varnothing$ with $\gamma =\varnothing \oplus 0,$ where the first $0$ is the zero tuple in $\mathbb{S}_{n-N}(\mathbb{C}^\mu)$ and the second $0$ is the zero vector in $\mathbb{C}^{n-N},$ and observe that $0=p(R)\gamma$ implies $p(T)\varnothing =0.$ \end{proof} \begin{lemma} \label{cor:gsforus} Let $q(a) = \sum_w q_w w(a)$ be a $d \times d$ matrix (not necessarily hermitian) polynomial in the freely noncommuting variables $a_1\dots,a_\mu.$ If $q(A) = 0$ for all $n \in {\mathbb N}$ and $A \in \mathbb S_n({\mathbb C}^\vmu),$ then $q = 0.$ \end{lemma} \begin{proof} Since $q(a)$ is a $d \times d$ matrix polynomial, i.e $q_w \in M_d({\mathbb C}),$ it can be viewed as a $d \times d$ matrix $(q^{i,j}(a))_{i,j=1}^d$ of scalar polynomials. Suppose that $q$ is nonzero. Choose $i,j$ such that $q^{i,j}(a)$ is nonzero. Since $q^{i,j}(A) = 0$ for all $n \in {\mathbb N}$ and $A \in \mathbb S_n({\mathbb C}^\mu),$ Lemma \ref{lem:vanishingpoly} implies $q^{i,j}$ is the zero polynomial, a contradiction. \end{proof} \begin{proposition} \label{rem:betterlinindep} For each positive integer $\kappa$ and each $n\ge N=\sum_{j=0}^\kappa \mu^j$ there exists $A \in \mathbb S_n({\mathbb C}^\vmu)$ and $v \in {\mathbb C}^n$ such that \[ \mathcal M_{A,v,\kappa}=\{w(A)v\,:\, \text{$w(a)$ is a word with degree at most $\kappa$} \} \] is linearly independent. In particular, in the case $\kappa=1,$ there is an $A\in \mathbb{S}_{\mu+1}({\mathbb C}^\mu)$ and a $v\in {\mathbb C}^{\mu+1}$ such that $\mathcal M_{A,v,1}$ is linearly independent. \end{proposition} \begin{proof} Fix $\kappa.$ Let $\mathcal W_\kappa$ denote the words in the (freely noncommuting) variables $a_1,\dots,a_\mu$ of degree at most $\kappa.$ The cardinality of $\mathcal W_\kappa$ is $N=\sum_{j=0}^\kappa \mu^j.$ Given $c:\mathcal W_\kappa\to {\mathbb C},$ let $c_w$ denote the value of $c$ at $w\in \mathcal W_\kappa.$ Let $\mathcal{C}$ denote the set of all functions $c:\mathcal W_\kappa\to {\mathbb C}$ such that $\sum_w |c_w|^2 =1.$ Thus $\mathcal{C}$ is identified with the unit sphere in $\mathbb{C}^N$ and is thus compact. Given $c\in \mathcal{C},$ let \[ q_c(a) = \sum_{w\in \mathcal W_\kappa} c_w w. \] Arguing by contradiction suppose, for each $n \ge N$, for each $C\in \mathbb S_n({\mathbb C}^\vmu)$ and each $\gamma\in {\mathbb C}^n,$ there exists a $c\in \mathcal{C}$ such that $q_c(C)\gamma=0.$ Given $n \ge N$ and $C\in \mathbb S_n({\mathbb C}^\vmu)$ and $\gamma\in {\mathbb C}^n,$ let \[ K_{C,\gamma} =\{c\in \mathcal{C}: q_c(C)\gamma=0\}. \] Thus $K_{C,\gamma}$ is nonempty for all $C$ and $\gamma.$ Likewise, since, for $C\in \mathbb S_n({\mathbb C}^\vmu)$ and $\gamma\in \mathbb{C}^n,$ the mapping \[ \mathcal{C}\ni c \mapsto q_c(C)\gamma\in \mathbb{C}^n \] is continuous, the sets $K_{C,\gamma}$ are compact. Given a positive integer $M$, positive integers $n_1,\dots, n_M \ge N$, $C^j\in \mathbb{S}_{n_j}({\mathbb C}^\mu)$ and $\gamma_j\in {\mathbb C}^{n_j}$ for $1\le j\le M,$ observe that \[ \cap_{j=1}^M K_{C^j,\gamma_j} = K_{\oplus C^j,\oplus \gamma_j} \ne \emptyset. \] Hence $\{K_{C,\gamma}: C,\gamma\}$ has the finite intersection property. It follows that \[ \cap_{C,\gamma} K_{C,\gamma} \ne \emptyset. \] Choosing any $\widetilde{c}$ in this intersection, \[ q_\widetilde{c}(C)\gamma=0 \] for all $C$ and $\gamma.$ Consequently $q_\widetilde{c}(C)=0$ for all $C \in \mathbb S_n({\mathbb C}^\vmu)$ and hence, by Lemma~\ref{lem:vanishingpoly}, $q_\widetilde{c}=0.$ Thus, $\widetilde{c}_w=0$ for all $\gamma,$ contradicting $\widetilde{c}\in\mathcal{C}.$ Hence, there exists an $C$ and $\gamma$ such that such that $\mathcal M_{C,\gamma,\kappa}$ is linearly independent. Let $\ell$ denote the size of $C;$ that is $C\in \mathbb{S}_\ell({\mathbb C}^\mu)$ and $\mathcal M_{C,\gamma,\kappa}$ is a subspace of $\mathbb{C}^\ell$ of dimension $N.$ Let $V$ denote the inclusion of $\mathcal M_{C,\gamma,\kappa}$ into $\mathbb{C}^\ell$ and let $B=V^*C V.$ Since $V^*A^\alpha V \gamma= A^\alpha \gamma\in \mathcal M_{C,\gamma,\kappa}$ for words $\alpha$ of length at most $\kappa,$ the set \[ \{w(B)\gamma: w \text{ is a word of length at most } \kappa\} \] is linearly independent. Given $m>N,$ let $A=B\oplus 0,$ where $0\in M_{m-N}({\mathbb C}^\mu).$ Likewise let $v=\gamma\oplus 0 \in{\mathbb C}^m = \mathbb{C}^N \oplus \mathbb{C}^{m-N}$ and note that $\mathcal M_{A,v,\kappa}$ is linearly independent. \end{proof} \subsection{The Border vector, middle matrix and non-hermitian variables} \label{s:BVMM} In this subsection, $q(a,x)$ denotes a fixed polynomial that is homogeneous of degree two in $x$ and $d_a$ denote its degree in $a.$ Enumerate the words in the variables $a_1,\dots,a_\mu$ of degree at most $d_a$ as $\{m_1,\dots,m_N\}.$ In particular, $N=\sum_{j=0}^{d_a} \mu^j.$ For $1\le j,k\le \mu$ and $1\le {\tt{r}},{\tt{t}} \le N$ there exists uniquely determined $d\times d$ matrix-valued polynomials $\mathfrak{Z}^{j,k}_{{\tt{r}},{\tt{t}}}(a)$ such that \begin{equation} \label{e:qax} q(a,x) = \sum_{j,k,{\tt{r}},{\tt{t}}} (I_d\otimes m_{{\tt{r}}}(a)^*x_j) \mathfrak{Z}^{j,k}_{{\tt{r}},{\tt{t}}}(a) (I_d\otimes x_km_{\tt{t}}(a)). \end{equation} In fact, \begin{equation} \label{e:qax-} (I_d\otimes m_{{\tt{r}}}(a)^*x_j) \mathfrak{Z}^{j,k}_{{\tt{r}},{\tt{t}}}(a) (I_d\otimes x_km_{\tt{t}}(a)) = \sum_{s=1}^N \{q_w w : w= m_{{\tt{r}}}(a)^* x_j m_s(a) x_k m_{{\tt{t}}}(a)\}. \end{equation} Letting $\mathcal{Z}$ denote the block matrix indexed by $((j,{\tt{r}}),(k,{\tt{t}}))$ with $d\times d$ polynomial entries $\mathfrak{Z}^{j,k}_{{\tt{r}},{\tt{t}}}(a)$ and letting $V(a)[x]$ the column vector with $(k,{\tt{t}})$ entry $I_d\otimes x_k m_{\tt{t}}(a),$ equation~\eqref{e:qax} becomes, \begin{equation} \label{e:qax+} q(a,x) = V(a)[x]^* \, \mathcal{Z}(a) \, V(a)[x]. \end{equation} The polynomial $V(a)[x]$ is the {\bf border vector} and $\mathcal{Z}(a)$ is the {\bf middle matrix} for $q.$ Equation~\eqref{e:qax+} is the \df{border vector-middle matrix} representation of $q.$ Before continuing, we pause to introduce non-hermitian freely noncommuting variables. Accordingly, let $\chi_1,\dots,\chi_k,z_1,\dots,z_\ell,w_1,\dots,w_\ell$ be freely noncommuting variables. Now let ${}^*$ denote an involution on words in these variables that reverses the order of products and satisfies $\chi_j^*=\chi_j$ and $z_j^*=w_j.$ Thus the $\chi$ variables are hermitian, but the $z,w$ variables are not. It is natural, and customary, to systematically use $z_j^*$ in place of $w_j.$ A polynomial in this mix of variables is now a linear combination of words with matrix coefficients. A word in these variables evaluates at a tuple $(X,Z) \in \mathbb{S}_n({\mathbb C}^k)\times M_n({\mathbb C}^\ell)$ in the natural way: replace $\chi_j$ with $X_j$ and similarly replace $z_j$ and $z_j^*$ with $Z_j$ and $Z_j^*.$ The involution extends in the evident fashion to this mixed variable setting. Namely, the coefficient matrices are replaced by their adjoints and the involution is applied to the words. Finally, a polynomial is hermitian if $p^*=p$; equivalently $p(X,Z)^*=p^*(X,Z)$ for all tuples $(X,Z).$ The definition of the border vector, as a polynomial, naturally extends to the case of non-hemitian $x$ variables. With this understanding, and given positive integers $m,n,$ a tuple $B\in \mathbb S_n({\mathbb C}^\vmu),$ a tuple $\beta\in M_{n,m}({\mathbb C}^\mu)$ and tuple $\alpha\in \mathbb{S}_m({\mathbb C}^\mu),$ \begin{equation} \label{e:qBba} \begin{split} \sum_{j,k,{\tt{r}},{\tt{t}}} & (I_d\otimes m_{{\tt{r}}}(B)^*\beta_j) \mathfrak{Z}^{j,k}_{{\tt{r}},{\tt{t}}}(\alpha) (I_d\otimes \beta_k^* m_{\tt{t}}(B))\\ &= V(B)[\beta^*]^* \mathcal{Z}(\alpha) V(B)[\beta^*]. \end{split} \end{equation} \begin{proposition} \label{p:BVMM} If $W,R,S$ are given as in Proposition~\ref{thm:partialconvexity} item~\ref{i:a2convex3}, then \begin{equation*} (I_d\otimes W)^* \, q(R,S)\, (I_d\otimes W) = q(A,X) + V(A)[\beta^*]^* \, \mathcal{Z}(\alpha)\, V(A)[\beta^*]. \end{equation*} \end{proposition} \begin{proof} Suppose $w = \ell(a)x_jc(a)x_kr(a),$ where $\ell(a), c(a), r(a)$ are words. Compute \begin{equation*} \begin{split} w(R,S) = & \begin{pmatrix} \ell(A) & 0 \\ 0 & \ell(\alpha) \end{pmatrix} \begin{pmatrix} X_j & \beta_j \\ \beta_j^* & \delta_j \end{pmatrix} \begin{pmatrix} c(A) & 0 \\ 0 & c(\alpha) \end{pmatrix} \begin{pmatrix} X_k & \beta_k \\ \beta_k^* & \delta_k \end{pmatrix} \begin{pmatrix} r(A) & 0 \\ 0 & r(\alpha)\end{pmatrix} \\ & = \begin{pmatrix} \ell(A)X_jc(A)X_kr(A) + \ell(A)\beta_jc(\alpha)\beta_k^*r(A) & \quad * \\ \quad * & \quad * \end{pmatrix}. \end{split} \end{equation*} Hence, % \begin{equation} \label{eq:imp1} W^*\, w(R,S)\, W = \ell(A)X_jc(A)X_kr(A) + \ell(A)\beta_j c(\alpha) \beta_k^*r(A). \end{equation} In particular, fixing ${\tt{r}},{\tt{t}},j,k$ and letting $Y=I_d\otimes W,$ equations~\eqref{e:qax-} and \eqref{eq:imp1} give \begin{equation} \label{e:ohmy} \begin{split} Y^* (I_d\otimes & S_j m_{{\tt{r}}}(R))^* \, \mathfrak{Z}^{j,k}_{{\tt{r}},{\tt{t}}}(R)\, (I_d\otimes S_k m_{{\tt{t}}}) (R))Y \\ = Y^* & \left (\sum_{s=1}^N \{ q_w \otimes w(R,S): w=m_{{\tt{r}}}(a)x_j m_s(a) x_k m_{{\tt{t}}}(a)\} \right ) Y \\ = & \sum_{s=1}^N \{ q_w \otimes [w(A,X) + m_{{\tt{r}}}(A)^*\beta_j m_s(\alpha) \beta_k^* m_{{\tt{t}}}(A)]: w=m_{{\tt{r}}}^*x_j m_s x_k m_{{\tt{t}}}\} \\ = & (I_d\otimes X_j m_{{\tt{r}}}(A))^* \, \mathfrak{Z}^{j,k}_{{\tt{r}},{\tt{t}}}(A) \, (I_d\otimes X_k m_{{\tt{t}}}(A)) \\ & + (I_d\otimes \beta_j^* m_{{\tt{r}}}(A))^* \, \mathfrak{Z}^{j,k}_{{\tt{r}},{\tt{t}}}(\alpha)\, (I_d\otimes \beta_k^* m_{{\tt{t}}}(A)). \end{split} \end{equation} Summing equation~\eqref{e:ohmy} over ${\tt{r}},{\tt{t}},j,k$ and using equations~\eqref{e:qax}, \eqref{e:qax+} and \eqref{e:qBba}, \begin{equation*} (I_d\otimes W)^* q(R,S) (I_d\otimes W) = q(A,X) + V(A)[\beta^*]^* \mathcal{Z}(\alpha) V(A)[\beta^*]. \qedhere \end{equation*} \end{proof} \subsection{Proof of Theorem \ref{thm:partialconvexpolys}} \label{s:a2endproof} \begin{proof}[Proof of Theorem \ref{thm:partialconvexpolys}] Since $p$ is convex in $x,$ Proposition~\ref{p:a2convex-deg2} says its degree in $x$ is at most two. Thus, \begin{equation*} p(a,x) = L(a,x) + q(a,x), \end{equation*} where $L(a,x)$ is affine linear in $x$ and \begin{equation*} q(a,x) =\sum_{w\in \Gamma} p_w w, \end{equation*} where $\Gamma$ denotes words in the variables $a,x$ that are homogeneous of degree two in $x.$ Since $p$ is convex in $x,$ so is $q,$ and it suffices to prove that there exists an $xy$-pencil $\Lambda$ such that $q=\Lambda^* \Lambda.$ Let $\kappa$ denote the degree of $q$ in $a.$ By Proposition~\ref{rem:betterlinindep}, there is an $\ell$ such that for all $n\ge \ell$ there exists an $A\in \mathbb S_n({\mathbb C}^\vmu)$ and a $v\in {\mathbb C}^n$ such that \[ \mathcal M_{A,v, \kappa} =\{w(A)v: w \mbox{ is a word of length at most } \kappa\} \] is linearly independent. Fix $n\ge \ell$ and choose $C \in \mathbb S_n({\mathbb C}^\vmu), v \in {\mathbb C}^n$ such that $\mathcal M_{C,v}$ is linearly independent. For this $C$ and a given $H\in M_n({\mathbb C}^\mu),$ the border vector evaluated at $(C,H^*)$ is \[ V(C)[H^*] = \displaystyle \bigoplus_{j=1}^\mu \begin{pmatrix} H^*_j m_1 (C) \\ \vdots \\ H_j^* m_N(C) \end{pmatrix}. \] By linear independence of $\mathcal M_{C,v},$ \begin{equation} \label{e:spanisall} \{V(C)[H^*]v:H\in M_n({\mathbb C}^\mu)\} = {\mathbb C}^{\mu n N}. \end{equation} Let $\alpha \in \mathbb S_n({\mathbb C}^\vmu)$ be given. To prove that $\mathcal{Z}(\alpha)\in M_d({\mathbb C})\otimes M_{\mu n N}({\mathbb C})$ is positive semidefinite, let $z\in {\mathbb C}^d\otimes {\mathbb C}^{\mu n N}$ be given. There exist $\gamma_1,\dots,\gamma_d\in {\mathbb C}^d$ and $u_1,\dots,u_d\in {\mathbb C}^{\mu n N}$ such that $z=\sum \gamma_{\tt{a}} \otimes u_{\tt{a}}.$ By equation \eqref{e:spanisall}, for each $1\le {\tt{a}}\le d,$ there exist $H^{\tt{a}} \in M_n({\mathbb C}^\mu)$ such that $u_{{\tt{a}}} = V(C)[(H^{\tt{a}})^*]v.$ Let $\beta_j$ denote the $d\times 1$ block matrix with $(a,1)$ entry $H_j^{\tt{a}}.$ Thus $\beta_j\in M_{dn,n}({\mathbb C})$ and $\beta \in M_{dn,n}({\mathbb C}^\mu).$ Let $v_{\tt{a}}=e_{\tt{a}}\otimes v \in {\mathbb C}^d\otimes {\mathbb C}^{n},$ where $\{e_1,\dots,e_d\}$ is the standard orthonormal basis for ${\mathbb C}^d.$ Set $A=I_d\otimes C\in \mathbb{S}_{dn}({\mathbb C}^\mu).$ Thus, $A$ is the direct sum of $C$ with itself $d$-times. Let $\Gamma=\sum_{{\tt{b}}=1}^d \gamma_{\tt{b}}\otimes v_{\tt{b}}$ and compute \begin{equation} \label{e:AtoC} \begin{split} (I_d\otimes V(A)[\beta^*]) & \Gamma = \sum_{{\tt{b}}=1}^d \gamma_{\tt{b}} \otimes V(A)[\beta^*](e_{\tt{b}}\otimes v)\\ = & \sum_{{\tt{b}}=1}^d \gamma_{\tt{b}} \otimes V(C)[(H^{\tt{b}})^*]v = \sum_{\tt{b}}^d \gamma_{\tt{b}} \otimes u_{\tt{b}} = z. \end{split} \end{equation} Let $\delta\in \mathbb S_n({\mathbb C}^\vmu)$ be given and let $W,R,S$ have the form given in Proposition~\ref{thm:partialconvexity} item~\ref{i:a2convex3}. Since $q$ is convex in $x,$ item~\ref{i:a2convex3} of Proposition~\ref{thm:partialconvexity} implies \begin{equation} \label{eq:imp2-alt} (I_d \otimes W)^*\, [q(R,S)]\, (I_d \otimes W) \succeq q(A,X). \end{equation} Proposition~\ref{p:BVMM} and equation~\eqref{eq:imp2-alt} give \begin{equation} \label{eq:imp3} (I_d \otimes V(A)[\beta^*])^* \, \mathcal Z(\alpha) \, (I_d \otimes V(A)[\beta^*]) \succeq 0. \end{equation} Combining equations~\eqref{eq:imp3} and \eqref{e:AtoC} gives, \[ 0 \le \langle \mathcal{Z}(\alpha) (I_d\otimes V(A)[\beta^*]) \Gamma, (I_d\otimes V(A)[\beta^*]) \Gamma\rangle = \langle \mathcal{Z}(\alpha) z,z\rangle \] and thus $\mathcal{Z}(\alpha)\succeq 0.$ At this point, it has been shown that there is an $\ell$ such that if $n\ge \ell$ and $\alpha\in \mathbb S_n({\mathbb C}^\vmu),$ then $\mathcal{Z}(\alpha)\succeq 0.$ Hence, by a standard direct sum argument, $\mathcal{Z}(\alpha)\succeq 0$ for all $n$ and $\alpha\in \mathbb S_n({\mathbb C}^\vmu);$ that is $\mathcal{Z}$ is a positive polynomial. Hence $\mathcal{Z}$ factors \cite{M} in the sense that there exists a (not necessarily square) matrix polynomial $F$ such that $\mathcal{Z}(a)=F(a)^*F(a).$ Consequently, \[ q(a,x) = V(a)[x]^* \mathcal{Z}(a) V(a)[x] = \Lambda(a,x)^* \Lambda(a,x), \] where $\Lambda(a,x) = F(a)V(a)[x]$ is linear in $x$ and the proof is complete. \end{proof} \subsection{Biconvexity} \label{s:a2cors} This section concludes by collecting consequences of Theorem~\ref{thm:partialconvexpolys} for later use. Let \df{$\mathcal{L}$} denote the set of words in $a,x$ of degree at most two in both $a$ and $x,$ but excluding those of the forms $a_ja_ix_kx_m$ and $x_mx_ka_ia_j.$ \begin{corollary} \label{c:preinfoabtp} Suppose $p(a,x)$ is a hermitian $d \times d$ matrix polynomial. If $p$ is convex in $x$ and has degree at most two in $a,$ then $p$ contains no words of the form $x_jx_\ell a_ka_m$ or $a_ma_kx_\ell x_j;$ that is $p(a,x) \in M_d \otimes \operatorname{span} \mathcal{L}.$ \end{corollary} \begin{proof} From Theorem~\ref{thm:partialconvexpolys}, \[ p(a,x) = L(a,x) + \Lambda(a,x)^*\Lambda(a,x), \] for matrix-valued polynomials $L$ and $\Lambda,$ where $L$ is affine linear in $x$ and $\Lambda$ is linear in $x.$ Since $p$ has degree at most two in $a,$ it is immediate that $L(a,x)$ has degree at most two in $a$ and thus is a (matrix-valued) linear combination of elements of $\mathcal{L}.$ Let $N$ denote the degree of $\Lambda$ in $a$ and, arguing by contradiction, suppose $N\ge 2.$ Write \[ \Lambda(a,x) =\sum_{u=0}^N \Lambda_u(a,x), \] where $\Lambda_u(a,x)$ is homogeneous of degree $u$ in $a.$ By assumption $\Lambda_N(a,x)\ne 0.$ Hence, by Lemma~\ref{cor:gsforus}, there exists $A,X$ such that $\Lambda_N(A,X)\ne 0.$ It follows that the matrix-valued polynomial of the single real variable $t,$ \[ F(t) =\Lambda(tA,X) =\sum_{u=0}^N t^u\Lambda_u(A,X) \] has degree $N.$ Hence \[ p(tA,X) = L(tA,X) + \Lambda(tA,X)^*\Lambda(tA,X) = L(tA,X) + F(t)^*F(t) \] has degree $2N\ge 4$ in $t,$ contradicting the assumption that $p$ has degree at most two in $A.$ We conclude that $\Lambda(a,x)$ has degree at most one in both $a$ and $x$ and the proof is complete. \end{proof} \begin{corollary} \label{prop:infoabtp} Suppose $p(a,x)$ is a hermitian $d \times d$ matrix polynomial. If $p$ is convex in both $a$ and $x$ (separately), then $p$ has degree at most two in both $a$ and $x$ and contains no words of the form $x_jx_\ell a_ka_m$ or $a_ma_kx_\ell x_j;$ that is $p(a,x) \in M_d \otimes \operatorname{span} \mathcal L.$ \end{corollary} \begin{proof} If the hermitian polynomial $p(a,x)$ is convex in both $a$ and $x,$ then Theorem~\ref{thm:partialconvexpolys} holds with the roles of $a$ and $x$ interchanged. In particular, if $p$ is convex in both $a$ and $x,$ then $p$ has degree at most two in both $a$ and $x$ and this result thus follows from Corollary~\ref{c:preinfoabtp}. \end{proof} \section{xy-convex hermitian polynomials} Proposition~\ref{p:more-is-true} and Theorem~\ref{t:introxyconvexp} are proved in this section. The proof strategy is to show $xy$-convexity here is to proceed directly from Corollary~\ref{prop:infoabtp}. For notational consistency with \cite{JKMMP21} we use $x=(x_1,\dots,x_\mu)$ and $y=(y_1,\dots,y_\mu)$ instead of $a,x$ for the two classes of variables. \begin{proposition}[[Proposition 4.1, JKMMP21] \label{p:xy-convexp-alt} A triple $((X,Y),V)$ is an $xy$-pair if and only if, up to unitary equivalence, it has the block form \begin{equation} \label{e:xypairform} X_j=\begin{pmatrix} X_{0j} & A_j&0\\ A_j^* & * & * \\ 0 & * & * \end{pmatrix}, \ \ Y_k= \begin{pmatrix} Y_{0k} & 0&C_k\\ 0& *&*\\ C_k^* & * & * \end{pmatrix}, \ \ V= \begin{pmatrix} I& 0&0\end{pmatrix}^*, \end{equation} $1 \le j,k \le \mu.$ Thus, a polynomial $p(x,y)\in M_d({\mathbb C} \langle x,y \rangle)$ is $xy$-convex if and only if \[ (I_d \otimes V)^*p(X,Y)( I_d \otimes V) - p(X_0,Y_0) \succeq 0 \] for each $xy$-pair $((X,Y),V)$ of the form of equation \eqref{e:xypairform}. \end{proposition} Recall the definition of $\mathcal{L}$ from Subsection~\ref{s:a2cors}. \begin{proof}[Proof of Proposition~\ref{p:more-is-true}] To show that $p$ is convex in $x$ and $y$ separately, simply replace $X_1,X_2,Y$ in the proof \cite[Lemma~4.3]{JKMMP21} with $X^1,X^2,Y\in \mathbb S_n({\mathbb C}^\mu).$ To prove item \ref{i:more2} implies item \ref{i:more3}, let $\mathcal W_1$ denote the words of degree at most one in each of $x$ and $y$ separately, and let $\mathcal W_2$ denote the set of words that have degree at least two, but no more than two in each of $x,y,$ but contains none of the words of the form $x_jx_\ell y_ky_m$ or $(x_jx_\ell y_ky_m)^*,$ for $1 \le j,k,\ell, m \le \mu$. Since $p(x,y)$ is convex in $x$ and $y$ separately, from Corollary~\ref{prop:infoabtp}, $p$ has the form, \[ p(x,y) = l(x,y) + q(x,y), \] where \[ \ell(x,y) =\sum_{w\in \mathcal W_1} p_w w, \ \ \ q(x,y)=\sum_{w\in \mathcal W_2} p_w w, \] for some $p_w\in M_d({\mathbb C}).$ Let $\mathcal W_{2,x}$ denote those words in $\mathcal W_2$ that have degree two in $x.$ Define $\mathcal W_{2,y}$ similarly. A computation shows, \[ \frac12 p_{x,x}(x,y)[x]=\frac12 q_{x,x}(x,y)[x] =\sum_{w\in \mathcal W_{2,x}} p_w w; \] that is, \begin{align*} \frac 1 2 p_{xx}(x,y)[x] & = \frac 1 2 q_{xx}(x,y)[x] = \sum_{j,k, \ell, m=1}^\mu [ p_{x_jx_\ell} x_j x_\ell\\ & + p_{x_j x_\ell y_k}x_jx_\ell y_k + p_{y_k x_\ell x_j} y_k x_\ell x_j + p_{x_jy_kx_\ell} x_j y_k x_\ell \\ & + p_{x_jy_ky_mx_\ell} x_jy_ky_mx_\ell + p_{y_k x_j x_\ell y_m} y_k x_j x_\ell y_m \\ & + p_{x_jy_kx_\ell y_m} x_jy_kx_\ell y_m + p_{y_kx_jy_m x_\ell} y_kx_jy_m x_\ell] \end{align*} Similarly, \[ \frac12 p_{y,y}(x,y)[y]=\frac12 q_{y,y}(x,y)[y] =\sum_ {w\in \mathcal W_{2,y} } p_w w \] Since $p(x,y)$ is convex in $x$ and $y$ separately, the partial Hessian of $p$ with respect to $x$ as well as $y$ is positive. In particular, \begin{equation} \label{e:partials-positive} p_{xx}(x,y)[x], \ \ p_{yy}(x,y)[y] \succeq 0. \end{equation} Let $\mathcal W_{1,x}$ denote those words in $\mathcal W_1$ that have degree one in $x.$ Define $\mathcal W_{1,y}$ similarly. By \cite[Theorem~0.2]{M}, the positivity condition in equation~\eqref{e:partials-positive} implies there exists an $N$ and $N \times d$ matrix-valued free polynomials $f(x,y)$ and $g(x,y)$ such that \[ p_{xx}(x,y)[x] = f(x,y)^*f(x,y) \,\, \, \, p_{yy}(x,y)[y] = g(x,y)^*g(x,y), \] where \[ f(x,y) = \sum_{w\in\mathcal W_{1,x}} f_w w = \sum_{j,k=1}^\mu f_{x_j} x_j + f_{x_jy_k}x_jy_k + f_{y_kx_j}y_kx_j \] and similarly, $g(x,y) =\sum_{w\in \mathcal W_{1,y}} g_w w.$ Let $\mathcal W_{1,x,y}$ denote the words of degree one in both $x$ and $y.$ Let $x$ and $y$ denote the column vectors \[ x=\begin{pmatrix} x_j \end{pmatrix}_{j=1}^\mu, \ \ \ y=\begin{pmatrix} y_j\end{pmatrix}_{j=1}^\mu \] and let {$v$} denote the column vector \[ v= \begin{pmatrix} w \end{pmatrix}_{w\in \mathcal W_{1,x,y}}. \] Let \[ \mathcal W_x=\begin{pmatrix} x \\ v \end{pmatrix}, \ \ \ \mathcal W_y =\begin{pmatrix} v \\ y \end{pmatrix}. \] Likewise, let $F_0,F_1$ and $F$ denote the row vectors, \[ F_0 =\begin{pmatrix} f_{x_j}\end{pmatrix}_{j=1}^\mu, \ \ \ F_1 =\begin{pmatrix} f_{w} \end{pmatrix}_{w\in \mathcal W_{1,x,y}}, \ \ \ F=\begin{pmatrix} F_0&F_1 \end{pmatrix} \] and similarly \[ G_0 = \begin{pmatrix} g_{x_j}\end{pmatrix}_{j=1}^\mu, \ \ \ G_1 =\begin{pmatrix} g_{w} \end{pmatrix}_{w\in \mathcal W_{1,x,y}}, \ \ \ G=\begin{pmatrix} G_1&G_0 \end{pmatrix}. \] Thus, \[ f = F\mathcal W_x, \ \ \ g=G\mathcal W_y \] and \[ \mathcal W_x^*F^*F \mathcal W_x = f^*f, \ \ \mathcal W_y^*G^*G \mathcal W_y = g^*g. \] Let \[ P =\begin{pmatrix}{p_{u^*w} }\end{pmatrix}_{w\in \mathcal W_{1,x,y}} \] and observe that \[ \begin{split} F^*F & =\begin{pmatrix} F_0^*F_0 & F_0^*F_1 \\ F_1^*F_0 & P \end{pmatrix}, \, G^*G =\begin{pmatrix} P & G_1^*G_0 \\ G_0^*G_1 & G_0^*G_0\end{pmatrix}. \end{split} \] Let \[ \mathcal M =\begin{pmatrix} F_0^*F_0 & F_0^*F_1 & 0 \\ F_1^*F_0 & P & G_1^*G_0\\ 0& G_0^*G_1 & G_0^*G_0 \end{pmatrix}, \ \ \mathcal W=\begin{pmatrix} x \\ v\\ y\end{pmatrix}, \] and observe \[ q(x,y)= \mathcal W^* \, \mathcal M \, \mathcal W. \] Since $F^*F$ and $G^*G$ are positive semidefinite, \cite[Proposition~1]{T} implies there is a $d \mu \times d \mu$ matrix $Q$ such that \[ \widehat{\mathcal M} = \mathcal M +\begin{pmatrix}0&0&Q\\0&0&0\\Q^* &0&0\end{pmatrix} = \begin{pmatrix} F_0^*F_0 & F_0^*F_1 & Q \\ F_1^*F_0 & P & G_1^*G_0\\ Q^* & G_0^*G_1 & G_0^*G_0 \end{pmatrix} \succeq 0. \] Letting $Q = \begin{pmatrix} Q_{j,k}\end{pmatrix}_{j.k=1}^\mu \in M_\mu \otimes M_d$, it follows that \[ p(x,y) = \lambda(x,y) + \mathcal W^* \widehat{\mathcal M} \mathcal W, \] where \begin{align*} \lambda(x,y) &= l(x,y) - \left \{\sum_{j,k=1}^\mu Q_{j,k} x_jy_k + Q_{j,k}^*y_kx_j \right\} \\ & = \sum_{j,k=1}^ \mu \left[p_{x_j} x_j + p_{y_j}y_j + (p_{x_jy_k} - Q_{j,k}) x_jy_k + (p_{y_kx_j} - Q_{j,k}^*) y_kx_j\right]. \end{align*} Since $\widehat{\mathcal M} \succeq 0$, there exists a matrix $\mathcal R$ such that $\widehat{\mathcal M} = \mathcal R^* \mathcal R$. Finally, letting $\Lambda(x,y) = \mathcal R \mathcal W$, it follows that $\Lambda(x,y)$ is a $d \times d$ matrix-valued $xy$-pencil and \[ p(x,y) = \lambda(x,y) + \Lambda(x,y)^*\Lambda(x,y). \qedhere \] To prove item~\ref{i:more3} implies item~\ref{i:more1}, let a triple $((X,Y),V)$ as in Proposition \ref{p:xy-convexp-alt} be given and observe, \begin{align*} p(V^*(X,Y)V) & = \lambda(V^*(X,Y)V) + \Lambda(V^*(X,Y)V)^* \Lambda (V^*(X,Y)V) \\ & = (I_d \otimes V)^*\lambda(X,Y)(I_d \otimes V) + (I_d \otimes V)^* \Lambda(X,Y)^* (I_d \otimes V V^*) \Lambda(X,Y) (I_d \otimes V) \\ & \preceq (I_d \otimes V)^*\lambda(X,Y)(I_d \otimes V) + (I_d \otimes V)^* \Lambda(X,Y)^* \Lambda(X,Y) (I_d \otimes V) \\ & = (I_d \otimes V)^*p(X,Y) (I_d \otimes V). \end{align*} It follows from Proposition \ref{p:xy-convexp-alt} that $p$ is $xy$-convex and the proof is complete. \end{proof}

fbe90c73d1d2d823af42fe51b4f69077a5844e9e

\section{\label{sec:Introduction}Introduction} Spin qubits in silicon QDs are a leading candidate for building a quantum processor due to their long coherence time~\cite{Dzurak2014AddressableQuantumDot, Dzurak2015TwoLogicGate}, potential scalability~\cite{Veldhorst2017InterfacingSpinQubits, Veldhorst2018CrossbarNetwork}, and compatibility with advanced semiconductor manufacturing technology~\cite{Camenzind2021FinFET, Vandersypen2022AdvancedSemiconductor, Zhang2018Semiconductor}. Nowadays, as an alternative to implementing electron spin resonance (ESR)~\cite{Koppens2006ESR, Huang2019FidelitySilicon, Veldhorst2020HotSilicon, Chan2021ExchangeSilicon, Morello2022PrecisionTomography, Gilbert2022OnDemandControl}, EDSR allows the single-qubit and two-qubit operation fidelities to achieve 99.9\%~\cite{Dzurak2018SiliconSpinQubit, Tarucha2018CoherenceLimit} and 99\%~\cite{Vandersypen2022QuantumLogic, Tarucha2022FastUniversal, Mills2022TwoQubitPetta}, respectively, and the qubit operation temperature to be higher than one kelvin~\cite{Dzurak2020AboveOneKelvin} To implement EDSR in Si-MOS QDs, a rectangular micromagnet is deployed to generate an inhomogeneous magnetic field and an oscillating electric field resonant with the Larmor frequency is coupled to drive the spin states~\cite{Rashba2003OrbitalMechnisms, Leon2020MultiElectronSiMOS, Zhang2021Synthetic, Hu2021OperationGuide}. During the conventional EDSR measurement, electrons in Si-MOS QD are confined in the quantum well, leading to a relatively small electric dipole~\cite{Pioro2008EDSR, Kawakami2014SiGeEDSR}. Driving single spin rotations in a DQD close to zero detuning where electron shuttles between two QDs, the "flopping-mode" EDSR increases the electric dipole in QDs~\cite{Benito2019FloppingTheory}. A longer coherent time with the same Rabi oscillation frequency ($f_\text{Rabi}$) has been achieved in Si/SiGe spin qubits by applying flopping-mode EDSR~\cite{Croot2020FloppingSiGe}. However, the small size and complicated distribution of Si-MOS QDs make cavity readout of a flopping-mode spin qubit in Si-MOS QDs difficult~\cite{Dzurak2014AddressableQuantumDot, Huang2019FidelitySilicon, Veldhorst2020HotSilicon, Dzurak2020AboveOneKelvin, Leon2020MultiElectronSiMOS, Chan2021ExchangeSilicon, Gilbert2022OnDemandControl}. Here, we demonstrate a flopping-mode single spin qubit in a Si-MOS QD via the Elzerman single-shot readout~\cite{Elzerman2004Single-shot}. By setting gate voltages carefully, a DQD with appropriate tunneling rates of an electron from QD to reservoir is formed underneath adjacent electrodes. Then, we measure the EDSR spectra, Rabi oscillation, and Ramsey fringes. Due to the large $2t_\text{c}$, an s-shape spin resonance frequency ($f_\text{q}$) as a function of the energy detuning ($\varepsilon$) is formed. We achieve an order of magnitude improvement in $f_\text{Rabi}$ around $\varepsilon=0$ with the spin dephasing times ($T_2^*$) virtually constant. Furthermore, we tune the gate voltage and reproduce the flopping-mode EDSR in another DQD. \section{\label{sec:Results and discussion}Results and discussion} \begin{figure*} \includegraphics{fig1.pdf \caption{\label{fig:1} Flopping-mode spin qubit device layout and operation. (a) False-color SEM image of the device. The left DQD is formed underneath gate LP (red) and BC (blue). Gate MC, BC (blue), LB, and M(yellow) form confinement barriers and laterally confine the QDs. LL (green) is the reservoir gate supplying electrons for the QDs. Gate SLB, SRB (yellow), SP (red), SLL, and SRL (green)define a single-electron transistor (SET), which is confined by gate TC and MC (blue). The white arrow above the SET indicates the SET current ($I_\text{S}$) direction. Gate LP (red) is connected to impedance-matched high-frequency lines via cryogenic bias-tees. And the left white arrow indicates that microwave is applied to gate LP. The right white arrow indicates the direction of the external magnetic field $B_\text{ext}$. The cobalt (Co) rectangular micromagnet at the bottom of the image is used to generate an inhomogeneous magnetic field. (b) Charge stability diagram of the left DQD is measured by differentiating $I_\text{S}$ as a function of gate voltages $V_\text{LP}$ and $V_\text{BC}$. The electron numbers in QD underneath gate LP and BC are labeled $(N_1, N_2)$ on the diagram. The direction of the energy detuning ($\varepsilon$) in the DQD is indicated by the black dashed arrow. (c) Charge stability diagram of the DQD as a function of gate voltages $V_\text{LP}$ and $V_\text{LB}$. The bias-tee connected to gate LP doesn't work well, so we have to apply the pulse to gate LB instead. The relative voltage magnitude at each step of the pulse sequence for qubit manipulation is illustrated by the black line between two circles. During the measurement, we calibrate $V_\text{LB}$ and $V_\text{BC}$ to maintain the tunneling rate of the electron from DQD to the reservoir at the transition line. (d) Cross-sectional schematics of the device, fabricated on a purified Silicon-28 epi-layer. Electrons confined in the left quantum well underneath gate SP are sensitive to charge movement in the QD region. The red ($N_1$) and blue ($N_2$) circles on the right side of the quantum wells represent the electron in the flopping-mode regime. (e) Cross-sectional schematics of the device along the perpendicular direction. The electron in the DQD tunnels across the barrier under gate LP to the reservoir under gate LL. The microwave is applied to gate LP to rotate the electron spin as mentioned in (a).} \end{figure*} \subsection{\label{sec:experimental setup}Experimental setup} Fig.~\ref{fig:1}(a) shows a scanning electron microscope (SEM) image of a typical Si-MOS DQD device~\cite{Hu2021OperationGuide}, nominally identical to the one measured in this paper. The device was fabricated on a natural silicon substrate with a 70 nm thick isotopically enriched $^{28}$Si epi layer which has a residual $^{29}$Si concentration of 60 ppm. The overlapping aluminum gate electrodes were fabricated using multi-layer gate stack technology~\cite{Zhang2020GiantAnisotropy}. The electrons are confined in the potential wells under gates LP and BC by selectively tuning gates LP, LB, and BC, as shown in Fig.~\ref{fig:1}(d). The DQD is formed under gates LP and BC, and the corresponding charge stability diagram is shown in Fig.~\ref{fig:1}(b). Gate MC and BC create a channel under gate LL for electrons to tunnel between the electron reservoir and the DQD. The tunneling rate can be modified by $V_\text{LB}$, as shown in Fig.~\ref{fig:1}(e). The energy difference induced by the external magnetic field, $B_\text{ext}=605\ \text{mT}$, is the main component of the Zeeman splitting between spin states. The micromagnet is fully magnetized, leading to a transverse magnetic field gradient of $\sim0.1\ \text{T}/\upmu\text{m}$~\cite{Zhang2021Synthetic}. The total magnetic field at the device $B_\text{tot}$ is the sum of $B_\text{ext}$ and the stray field from the micromagnet. As a result, $\sim20$ GHz microwave pulses are applied to the LP gate via a cryogenic bias-tee to manipulate the qubit.The device is in a dilution refrigerator at an electron temperature of $T_\text{e}=182.7\pm0.6$ mK (see Appendix~\ref{sec:lever arm and electron temperature}); the energy gap between $\ket{\uparrow}$ and $\ket{\downarrow}$ allows us to read the electron spin state via the Elzerman readout. \subsection{\label{sec:Elzerman readout}Elzerman readout} Fig.~\ref{fig:1}(b) is the charge stability diagram of the DQD underneath gate LP and BC, measured by differentiating $I_\text{S}$. $(N_1, N_2)$ on the diagram labels the corresponding number of electrons. The black arrow illustrates the direction of $\varepsilon$ between the DQD. Fig.~\ref{fig:1}(c) shows the charge stability diagram of the DQD as a function of $V_\text{LP}$ and $V_\text{LB}$. Here, we use gate LB to modify the tunneling rate of electrons from the DQD to the electron reservoir. Unfortunately, the bias-tee connected to gate LP doesn't work well, so we have to apply the two-step pulse sequence for the qubit operation to gate LB instead, as shown by points R (Read) and C (Control) in Fig.~\ref{fig:1}(c). We confirm the transitions between points R and C are adiabatic, as discussed in Appendix~\ref{sec: adiabatic}. The spin state is read out via state-to-charge conversion at point R, and a $\ket{\downarrow}$ electron is selectively loaded for initialization in the next pulse sequence~\cite{Hu2021OperationGuide}. By using sequences of selective EDSR pulses with microwave burst of frequency ($f_\text{s}$) at point C, we perform single-qubit operations on the electron. We then measure the single spin qubit along the charge transition line from (0, 0)--(0, 1) to (0, 0)--(1, 0); in the meanwhile, $\varepsilon$ increases from --4.5 to 4.5 meV (see Appendix~\ref{sec: calculation} for details). \subsection{\label{sec:detection of EDSR}Detection of EDSR} By selectively setting $V_\text{LP}$, $V_\text{LB}$, and $V_\text{BC}$, we perform the Elzerman readout with a fixed tunneling rate of around 150 Hz for the $\ket{\downarrow}$ at the exact transition line from (0, 0)--(0, 1) to (0, 0)--(1, 0). Then, we apply the two-step pulse sequence to gate LB. A frequency-chirped microwave pulse ($\pm2$ MHz around a microwave frequency ($f_\text{s}$) lasts $100\ \upmu$s) is applied to gate LP before the end of Control phase~\cite{Hu2021OperationGuide, Shafiei2013ChirpSiGe, Laucht2014ChirpPhosphorus, Sigillito2019ChirpSiGe}. If the frequency sweeps through $f_\text{q}$, the electron spin will end up in the excited state $\ket{\uparrow}$. So, it's convenient for us to identify $f_\text{q}$. We measure the probability of electron in the excited state ($P_\uparrow$) as a function of $f_\text{s}$ from 300 repeated single-shot readouts. For each $\varepsilon$, we repeat the measurement ten times, as mentioned in Ref.~\cite{Zhang2020GiantAnisotropy}. The EDSR spectra over $\varepsilon$ from --4.5 to 4.5 meV are shown in Fig.~\ref{fig:2}(b). There is an s-shape curve of increased $P_\uparrow$ with a wide of 4 MHz of $f_\text{s}$, where $f_\text{q}$ is located. We calibrate the peak of $P_\uparrow$ and extract $f_\text{q}$ as a function of $\varepsilon$ in Fig.~\ref{fig:2}(c). \begin{figure}[t] \includegraphics{fig2.pdf \caption{\label{fig:2} (a) Top: The eigenenergies calculated by diagonalizing the Hamiltonian in Eq.~\ref{eq:1}. Bottom: Schematics of potential well for far detuned (left and right) and flopping-mode (middle) regime. (b) EDSR spectra for the probability of $\ket{\uparrow}$ electron ($P_\uparrow$) as a function of both $\varepsilon$ and microwave frequency ($f_\text{s}$). The dashed curve marks the position of increased $P_\uparrow$, which represents coarse ranges of spin resonance frequency ($f_\text{q}$). (c) $f_\text{q}$ as a function of $\varepsilon$ for $B_\text{ext}=604\ \text{mT}$ (red) and $B_\text{ext}=605\ \text{mT}$ (blue). The left arrows indicate the Zeeman energy difference ($\delta f_\text{q}$) generated by longitudinal magnetic field difference ($b_\text{z}$) of micromagnet. The right arrows illustrate the splitting energy difference ($\Delta E_\text{z}$) for different $B_\text{ext}$.} \end{figure} To explain this s-shape feature, we focus on the Hamiltonian $H$ of a single-electron DQD system on the basis $(\ket{L\downarrow},\ \ket{L\uparrow},\ \ket{R\downarrow},\ \ket{R\uparrow})$~\cite{Benito2019FloppingTheory, Benito2017InputOutput}: \begin{equation} H=\frac{1}{2} \left( \begin{matrix} -\varepsilon-E_\text{z1} & -2t_\text{SO} & 2t_\text{c} & 0\\ -2t_\text{SO} & -\varepsilon+E_\text{z1} & 0 & 2t_\text{c}\\ 2t_\text{c} & 0 & \varepsilon-E_\text{z2} & 2t_\text{SO}\\ 0 & 2t_\text{c} & 2t_\text{SO} & \varepsilon+E_\text{z2} \end{matrix} \right). \label{eq:1} \end{equation} Here, $2t_\text{c}$ is the interdot tunnel coupling, $E_\text{z}=(E_\text{z1}+E_\text{z2})/2$ is the averaged Zeeman energy, $\delta E_\text{z}=(E_\text{z1}-E_\text{z2})/2=g\mu_\text{B} b_\text{z}$ is the Zeeman energy difference generated by longitudinal magnetic field difference ($b_\text{z}$) of the micromagnet, $2t_\text{SO}=g\mu_\text{B}b_\bot$ is the synthetic spin-orbit coupling induced by transverse magnetic field difference ($b_\bot$). The eigenenergies of this four-level system are shown in Fig.~\ref{fig:2}(a). The avoided crossings at $\varepsilon=0$ are generated by 2$t_\text{c}$. By diagonalizing the Hamiltonian in Eq.~(\ref{eq:1}), we calculate the energy splitting ($E_\text{s}$) between the lowest two energy levels as $f_\text{q}\equiv E_\text{s}/h$. $\delta f_\text{q}\equiv 2\delta E_\text{z}/h$ for far detuned limits is shown in Fig.~\ref{fig:2}(c). For the limitation of a small inhomogeneous field, i.e., $b_{\bot,z}\ll |\Omega-E_\text{z}|$, where $\Omega=\sqrt{\varepsilon^2+4t_\text{c}^2}$, $E_\text{s}$ is corrected by the transverse and longitude gradient to second and first order, respectively~\cite{Benito2019FloppingTheory}: \begin{equation} E_\text{s}\simeq E_\text{z}-\frac{E_\text{z}^2-\varepsilon^2}{2E_\text{z}(\Omega^2-E_\text{z}^2)}(g\mu_\text{B}b_\bot)^2-\frac{\varepsilon}{\Omega}g\mu_\text{B}b_\text{z}. \label{eq:energy splitting} \end{equation} \begin{figure}[b] \includegraphics{fig3.pdf \caption{\label{fig:3} (a) The Rabi chevron is acquired at $\varepsilon = 0.5\ \text{meV}$. Every point of $P_\uparrow$ is obtained by repeating the pulse sequence 300 times with $\uptau_\text{B}$ fixed. We repeat the measurement 10 times for every $\uptau_\text{B}$, and sum the results. (b) The Rabi frequency ($f_\text{Rabi}$) is plotted as a function of $\varepsilon$ with fixed microwave power. $f_\text{Rabi}$ is an order of magnitude larger at around $\varepsilon=0$ than the far detuned position. The solid curve is the fitting result of Eq.~\ref{eq:f_rabi}. (c) Rabi oscillations obtained at different $\varepsilon$. Top panel: $\varepsilon=0$ meV, $f_\text{Rabi}=1.262\pm0.002$ MHz and $T_2^\text{Rabi}=6.46\pm0.39$ $\upmu$s; Middle panel: $\varepsilon=1.5$ meV, $f_\text{Rabi}=0.429\pm0.003$ MHz and $T_2^\text{Rabi}=5.53\pm0.57$ $\upmu$s; Bottom panel: $\varepsilon=3.0$ meV, $f_\text{Rabi}=0.135\pm0.003$ MHz and $T_2^\text{Rabi}=7.01\pm0.82$ $\upmu$s.} \end{figure} We plot $f_\text{q}$ as a function of $\varepsilon$ for $B_\text{ext}=$ 605 and 604 mT in Fig.~\ref{fig:2}(c). By fitting $f_\text{q}$ with Eq~\ref{eq:energy splitting}, we obtain $2t_\text{c}=914\pm167$ and $705\pm40$ GHz for $B_\text{ext}=$ 605 and 604 mT, respectively. The difference between the fitted splitting energy ($\Delta E_\text{z}=19.790\pm0.002-19.760\pm0.001$ GHz) equals the difference between the external magnetic fields ($g\mu_\text{B}\Delta B_\text{ext}/h=28$ MHz). \subsection{\label{sec: Rabi oscillation}Rabi oscillation and Ramsey fringes} After calibrating $f_\text{q}$, we now use a microwave burst with a specific burst time ($\uptau_\text{B}$) to manipulate the spin qubit. First, we measure $P_\uparrow$ as an $f_\text{s}$ function with a fixed $\uptau_\text{B}$. Each point of $P_\uparrow$ in the curve is averaged from 300 repeated single-shot readouts. Then, we repeat the measurement ten times and sum $P_\uparrow$ with $\uptau_\text{B}$ changing from 0 to 4 $\upmu$s. The Rabi chevron is plotted in Fig.~\ref{fig:3}(a). Fig.~\ref{fig:3}(b) illustrates $f_\text{Rabi}$ as a function of $\varepsilon$. $f_\text{Rabi}$ is symmetric about $\varepsilon=0$ and is an order of magnitude larger at $\varepsilon=0$ than the far detuned position. For every $\varepsilon$, the corresponding $f_\text{Rabi}$ are extracted by fitting Rabi oscillation with the function $P_\uparrow(\uptau_\text{B})=A\cdot\text{exp}(-\uptau_\text{B}/T_2^\text{Rabi})\cdot\text{sin}(f_\text{Rabi}\uptau_\text{B})$, as shown in Fig.~\ref{fig:3}(c). \begin{figure}[t] \includegraphics{fig4.pdf \caption{\label{fig:4} (a) Ramsey fringes as a function of frequency detuning ($\Delta f_\text{s}$) and waiting time ($\uptau_\text{wait}$) is measured at $\varepsilon = 0.5\ \text{meV}$ by applying a standard Ramsey fringes sequence with two $\pi/2$ pulses. Every point of $P_\uparrow$ is obtained by repeating the pulse sequence 300 times. We repeat the measurement ten times for every $\uptau_\text{wait}$ and sum the results. (b) The spin dephasing times ($T_2^*$) were measured through Ramsey fringes for different $\varepsilon$. Top panel: $\varepsilon=0$ meV, $T_2^*=0.42\pm0.31$ $\upmu$s; Bottom panel: $\varepsilon=1.5$ meV, $T_2^*=0.43\pm0.02$ $\upmu$s.} \end{figure} For a typical flopping-mode EDSR process, Ref~\cite{Benito2019FloppingTheory} gives $f_\text{Rabi}$ as a function of $\varepsilon$ for small $b_\bot$: \begin{equation} f_\text{Rabi} = 4t_\text{c}^2g\mu_\text{B}b_\bot\Omega_\text{c}/\Omega|\Omega^2-E_\text{z}^2|. \label{eq:f_rabi} \end{equation} Here, $\Omega_\text{c}=edE_\text{ac}/\hbar$ is the Rabi frequency in the standard EDSR regime, proportional to the distance between the two QDs $d$, and the electric field with amplitude $E_{ac}$. $\Omega_\text{c}=15.8\pm0.8$ GHz can be obtained from the relevant result of $f_\text{Rabi}$ with a $g\mu_\text{B}b_\bot$ estimated as 0.232 $\upmu$eV~\cite{Zhang2021Synthetic}. We estimate $d\sim$ 0.02 $\upmu$m, thus $b_\bot\sim0.1$ T/$\upmu$m$\cdot 0.02$ $\upmu$m $=2$ mT. Fig.~\ref{fig:3}(c) shows details of Rabi oscillations for different $\varepsilon$. $f_\text{Rabi}=1.262\pm0.002$ MHz is achieved in the top panel. When $\varepsilon$ increases to 1.5 and 3 meV, the Rabi frequencies decrease to $f_\text{Rabi}=0.429\pm0.003$ and $f_\text{Rabi}=0.135\pm0.003$ MHz, respectively. By fitting Rabi oscillation to an exponentially decaying sinusoid, $T_2^\text{Rabi}=6.46\pm0.39$ $\upmu$s at $\varepsilon=0$, $T_2^\text{Rabi}=5.53\pm0.57$ $\upmu$s at $\varepsilon=1.5\ \text{meV}$ and $T_2^\text{Rabi}=7.01\pm0.82$ $\upmu$s at $\varepsilon=3\ \text{meV}$ are obtained. $T_2^\text{Rabi}$ is stable when $\varepsilon$ increases. We next measure $T_2^*$ for different $\varepsilon$ through Ramsey fringes. In Fig.~\ref{fig:4}(a), the Ramsey fringes are measured in the same way as the Rabi chevron. The averaged $T_2^*=0.42\ \upmu$s for $\varepsilon=$ 0 and 1.5 meV is acquired as shown in Fig.~\ref{fig:4}(b). In Ref.~\cite{Croot2020FloppingSiGe}, the lowest $E_\text{s}$ occurs near $\varepsilon=0$, leading to a sweet spot for spin dephasing. However, in our device $2t_\text{c}\gg E_\text{z}$, the second-order item $\frac{E_\text{z}^2-\varepsilon^2}{2E_\text{z}(\Omega^2-E_\text{z}^2)}(g\mu_\text{B}b_\bot)^2$ in Eq~\ref{eq:energy splitting} is suppressed, and there is no sweet spot around $\varepsilon=0$. Here, we attribute the enhancement of the quality factor ($Q\equiv2T_2^\text{Rabi}f_\text{Rabi}$) to the improvement of the electric dipole, since $b_\text{z}$ is constant during the measurement. As mentioned in Ref.~\cite{Zhang2021Synthetic}, longitudinal magnetic field difference $b_\text{z}$ is one of the most relevant sources for dephasing in our device. \subsection{\label{sec:Reproduce}Reproduce the results} \begin{figure}[b] \includegraphics{fig5.pdf \caption{\label{fig:5} (a) Charge stability diagram of the DQD as a function of gate voltages $V_\text{LP}$ and $V_\text{RP}$. The black dashed arrow illustrates the direction of $\varepsilon$ between the second DQD. The relative voltage magnitude at each step of the pulse sequence for qubit manipulation is represented by the black line between two circles. (b) EDSR spectra for the probability of $\ket{\uparrow}$ electron ($P_\uparrow$) as a function of $\varepsilon$ and $f_\text{s}$. The yellow triangles mark the regime of the increased $P_\uparrow$, which represents the approximate position of $f_\text{q}$. The solid curve fitting to the Eq.~\ref{eq:energy splitting} yields $2t_\text{c}=71\pm10$ GHz.} \end{figure} Finally, we try to reproduce the flopping-mode EDSR in another DQD. Fig.~\ref{fig:5}(a) shows the charge stability diagram of another DQD under gate LP and RP. We tune the tunneling rate of the electron from QDs to the reservoir under LL with gate LP and BC. By selectively tuning $V_\text{LB}$, $V_\text{LP}$ and $V_\text{BC}$, we can change $\varepsilon$ from --1 to 0.6 meV. Fig.~\ref{fig:5}(b) shows the corresponding EDSR spectra and $2t_\text{c}=71\pm10$ GHz is obtained by fitting the solid curve to Eq.~\ref{eq:energy splitting}. As shown in Fig.~\ref{fig:5}(b), $P_\uparrow$ background rises when $\varepsilon<0$, and the peak of the increased $P_\uparrow$ becomes invisible on the EDSR spectra. Here, we select $f_\text{q}$ manually and mark them with the yellow triangles. Some points around $\varepsilon=0$ deviate from the curve without affecting the overall s-shape feature. However, the weak visibility makes subsequent measurements around $\varepsilon=0$ of Rabi oscillation and Ramsey fringes impossible. \section{\label{sec:Conclusion}Conclusion} In summary, we demonstrate the flopping-mode EDSR in a Si-MOS quantum dot through the Elzerman single-shot readout. We construct a DQD with $2t_\text{c}\sim800$ GHz under adjacent electrodes by selectively setting gate voltages. We extract an s-shape $f_\text{q}$ as a function of $\varepsilon$ from the EDSR spectra. Then, we improve $f_\text{Rabi}$ an order of magnitude from $0.135\pm0.003$ to $1.262\pm0.002$ MHz by increasing the electric dipole. And $T_2^*$ and $T_2^\text{Rabi}$ remains around $0.42\pm0.03$ $\upmu$s and $6.46\pm0.39$ $\upmu$s, respectively. Moreover, we have shown that the s-shape EDSR spectra holds in another DQD with $2t_\text{c}\sim70$ GHz. We anticipate that flopping-mode EDSR will have better performance in the heavy hole regime~\cite{Mutter2021HeavyHoleFlopping} or phosphorus donor qubits~\cite{Krauth2022PhosphorusFlopping}, and will perform two-qubit operation~\cite{Cayao2020FloppingTwoQubits} in the future. \begin{acknowledgments} This work was supported by the National Natural Science Foundation of China (Grants No. 12074368, 92165207, 12034018 and 61922074), the Anhui Province Natural Science Foundation (Grants No. 2108085J03), the USTC Tang Scholarship, and this work was partially carried out at the USTC Center for Micro and Nanoscale Research and Fabrication. \end{acknowledgments} \section{\label{sec:Introduction}Introduction} Spin qubits in silicon QDs are a leading candidate for building a quantum processor due to their long coherence time~\cite{Dzurak2014AddressableQuantumDot, Dzurak2015TwoLogicGate}, potential scalability~\cite{Veldhorst2017InterfacingSpinQubits, Veldhorst2018CrossbarNetwork}, and compatibility with advanced semiconductor manufacturing technology~\cite{Camenzind2021FinFET, Vandersypen2022AdvancedSemiconductor, Zhang2018Semiconductor}. Nowadays, as an alternative to implementing electron spin resonance (ESR)~\cite{Koppens2006ESR, Huang2019FidelitySilicon, Veldhorst2020HotSilicon, Chan2021ExchangeSilicon, Morello2022PrecisionTomography, Gilbert2022OnDemandControl}, EDSR allows the single-qubit and two-qubit operation fidelities to achieve 99.9\%~\cite{Dzurak2018SiliconSpinQubit, Tarucha2018CoherenceLimit} and 99\%~\cite{Vandersypen2022QuantumLogic, Tarucha2022FastUniversal, Mills2022TwoQubitPetta}, respectively, and the qubit operation temperature to be higher than one kelvin~\cite{Dzurak2020AboveOneKelvin} To implement EDSR in Si-MOS QDs, a rectangular micromagnet is deployed to generate an inhomogeneous magnetic field and an oscillating electric field resonant with the Larmor frequency is coupled to drive the spin states~\cite{Rashba2003OrbitalMechnisms, Leon2020MultiElectronSiMOS, Zhang2021Synthetic, Hu2021OperationGuide}. During the conventional EDSR measurement, electrons in Si-MOS QD are confined in the quantum well, leading to a relatively small electric dipole~\cite{Pioro2008EDSR, Kawakami2014SiGeEDSR}. Driving single spin rotations in a DQD close to zero detuning where electron shuttles between two QDs, the "flopping-mode" EDSR increases the electric dipole in QDs~\cite{Benito2019FloppingTheory}. A longer coherent time with the same Rabi oscillation frequency ($f_\text{Rabi}$) has been achieved in Si/SiGe spin qubits by applying flopping-mode EDSR~\cite{Croot2020FloppingSiGe}. However, the small size and complicated distribution of Si-MOS QDs make cavity readout of a flopping-mode spin qubit in Si-MOS QDs difficult~\cite{Dzurak2014AddressableQuantumDot, Huang2019FidelitySilicon, Veldhorst2020HotSilicon, Dzurak2020AboveOneKelvin, Leon2020MultiElectronSiMOS, Chan2021ExchangeSilicon, Gilbert2022OnDemandControl}. Here, we demonstrate a flopping-mode single spin qubit in a Si-MOS QD via the Elzerman single-shot readout~\cite{Elzerman2004Single-shot}. By setting gate voltages carefully, a DQD with appropriate tunneling rates of an electron from QD to reservoir is formed underneath adjacent electrodes. Then, we measure the EDSR spectra, Rabi oscillation, and Ramsey fringes. Due to the large $2t_\text{c}$, an s-shape spin resonance frequency ($f_\text{q}$) as a function of the energy detuning ($\varepsilon$) is formed. We achieve an order of magnitude improvement in $f_\text{Rabi}$ around $\varepsilon=0$ with the spin dephasing times ($T_2^*$) virtually constant. Furthermore, we tune the gate voltage and reproduce the flopping-mode EDSR in another DQD. \section{\label{sec:Results and discussion}Results and discussion} \begin{figure*} \includegraphics{fig1.pdf \caption{\label{fig:1} Flopping-mode spin qubit device layout and operation. (a) False-color SEM image of the device. The left DQD is formed underneath gate LP (red) and BC (blue). Gate MC, BC (blue), LB, and M(yellow) form confinement barriers and laterally confine the QDs. LL (green) is the reservoir gate supplying electrons for the QDs. Gate SLB, SRB (yellow), SP (red), SLL, and SRL (green)define a single-electron transistor (SET), which is confined by gate TC and MC (blue). The white arrow above the SET indicates the SET current ($I_\text{S}$) direction. Gate LP (red) is connected to impedance-matched high-frequency lines via cryogenic bias-tees. And the left white arrow indicates that microwave is applied to gate LP. The right white arrow indicates the direction of the external magnetic field $B_\text{ext}$. The cobalt (Co) rectangular micromagnet at the bottom of the image is used to generate an inhomogeneous magnetic field. (b) Charge stability diagram of the left DQD is measured by differentiating $I_\text{S}$ as a function of gate voltages $V_\text{LP}$ and $V_\text{BC}$. The electron numbers in QD underneath gate LP and BC are labeled $(N_1, N_2)$ on the diagram. The direction of the energy detuning ($\varepsilon$) in the DQD is indicated by the black dashed arrow. (c) Charge stability diagram of the DQD as a function of gate voltages $V_\text{LP}$ and $V_\text{LB}$. The bias-tee connected to gate LP doesn't work well, so we have to apply the pulse to gate LB instead. The relative voltage magnitude at each step of the pulse sequence for qubit manipulation is illustrated by the black line between two circles. During the measurement, we calibrate $V_\text{LB}$ and $V_\text{BC}$ to maintain the tunneling rate of the electron from DQD to the reservoir at the transition line. (d) Cross-sectional schematics of the device, fabricated on a purified Silicon-28 epi-layer. Electrons confined in the left quantum well underneath gate SP are sensitive to charge movement in the QD region. The red ($N_1$) and blue ($N_2$) circles on the right side of the quantum wells represent the electron in the flopping-mode regime. (e) Cross-sectional schematics of the device along the perpendicular direction. The electron in the DQD tunnels across the barrier under gate LP to the reservoir under gate LL. The microwave is applied to gate LP to rotate the electron spin as mentioned in (a).} \end{figure*} \subsection{\label{sec:experimental setup}Experimental setup} Fig.~\ref{fig:1}(a) shows a scanning electron microscope (SEM) image of a typical Si-MOS DQD device~\cite{Hu2021OperationGuide}, nominally identical to the one measured in this paper. The device was fabricated on a natural silicon substrate with a 70 nm thick isotopically enriched $^{28}$Si epi layer which has a residual $^{29}$Si concentration of 60 ppm. The overlapping aluminum gate electrodes were fabricated using multi-layer gate stack technology~\cite{Zhang2020GiantAnisotropy}. The electrons are confined in the potential wells under gates LP and BC by selectively tuning gates LP, LB, and BC, as shown in Fig.~\ref{fig:1}(d). The DQD is formed under gates LP and BC, and the corresponding charge stability diagram is shown in Fig.~\ref{fig:1}(b). Gate MC and BC create a channel under gate LL for electrons to tunnel between the electron reservoir and the DQD. The tunneling rate can be modified by $V_\text{LB}$, as shown in Fig.~\ref{fig:1}(e). The energy difference induced by the external magnetic field, $B_\text{ext}=605\ \text{mT}$, is the main component of the Zeeman splitting between spin states. The micromagnet is fully magnetized, leading to a transverse magnetic field gradient of $\sim0.1\ \text{T}/\upmu\text{m}$~\cite{Zhang2021Synthetic}. The total magnetic field at the device $B_\text{tot}$ is the sum of $B_\text{ext}$ and the stray field from the micromagnet. As a result, $\sim20$ GHz microwave pulses are applied to the LP gate via a cryogenic bias-tee to manipulate the qubit.The device is in a dilution refrigerator at an electron temperature of $T_\text{e}=182.7\pm0.6$ mK (see Appendix~\ref{sec:lever arm and electron temperature}); the energy gap between $\ket{\uparrow}$ and $\ket{\downarrow}$ allows us to read the electron spin state via the Elzerman readout. \subsection{\label{sec:Elzerman readout}Elzerman readout} Fig.~\ref{fig:1}(b) is the charge stability diagram of the DQD underneath gate LP and BC, measured by differentiating $I_\text{S}$. $(N_1, N_2)$ on the diagram labels the corresponding number of electrons. The black arrow illustrates the direction of $\varepsilon$ between the DQD. Fig.~\ref{fig:1}(c) shows the charge stability diagram of the DQD as a function of $V_\text{LP}$ and $V_\text{LB}$. Here, we use gate LB to modify the tunneling rate of electrons from the DQD to the electron reservoir. Unfortunately, the bias-tee connected to gate LP doesn't work well, so we have to apply the two-step pulse sequence for the qubit operation to gate LB instead, as shown by points R (Read) and C (Control) in Fig.~\ref{fig:1}(c). We confirm the transitions between points R and C are adiabatic, as discussed in Appendix~\ref{sec: adiabatic}. The spin state is read out via state-to-charge conversion at point R, and a $\ket{\downarrow}$ electron is selectively loaded for initialization in the next pulse sequence~\cite{Hu2021OperationGuide}. By using sequences of selective EDSR pulses with microwave burst of frequency ($f_\text{s}$) at point C, we perform single-qubit operations on the electron. We then measure the single spin qubit along the charge transition line from (0, 0)--(0, 1) to (0, 0)--(1, 0); in the meanwhile, $\varepsilon$ increases from --4.5 to 4.5 meV (see Appendix~\ref{sec: calculation} for details). \subsection{\label{sec:detection of EDSR}Detection of EDSR} By selectively setting $V_\text{LP}$, $V_\text{LB}$, and $V_\text{BC}$, we perform the Elzerman readout with a fixed tunneling rate of around 150 Hz for the $\ket{\downarrow}$ at the exact transition line from (0, 0)--(0, 1) to (0, 0)--(1, 0). Then, we apply the two-step pulse sequence to gate LB. A frequency-chirped microwave pulse ($\pm2$ MHz around a microwave frequency ($f_\text{s}$) lasts $100\ \upmu$s) is applied to gate LP before the end of Control phase~\cite{Hu2021OperationGuide, Shafiei2013ChirpSiGe, Laucht2014ChirpPhosphorus, Sigillito2019ChirpSiGe}. If the frequency sweeps through $f_\text{q}$, the electron spin will end up in the excited state $\ket{\uparrow}$. So, it's convenient for us to identify $f_\text{q}$. We measure the probability of electron in the excited state ($P_\uparrow$) as a function of $f_\text{s}$ from 300 repeated single-shot readouts. For each $\varepsilon$, we repeat the measurement ten times, as mentioned in Ref.~\cite{Zhang2020GiantAnisotropy}. The EDSR spectra over $\varepsilon$ from --4.5 to 4.5 meV are shown in Fig.~\ref{fig:2}(b). There is an s-shape curve of increased $P_\uparrow$ with a wide of 4 MHz of $f_\text{s}$, where $f_\text{q}$ is located. We calibrate the peak of $P_\uparrow$ and extract $f_\text{q}$ as a function of $\varepsilon$ in Fig.~\ref{fig:2}(c). \begin{figure}[t] \includegraphics{fig2.pdf \caption{\label{fig:2} (a) Top: The eigenenergies calculated by diagonalizing the Hamiltonian in Eq.~\ref{eq:1}. Bottom: Schematics of potential well for far detuned (left and right) and flopping-mode (middle) regime. (b) EDSR spectra for the probability of $\ket{\uparrow}$ electron ($P_\uparrow$) as a function of both $\varepsilon$ and microwave frequency ($f_\text{s}$). The dashed curve marks the position of increased $P_\uparrow$, which represents coarse ranges of spin resonance frequency ($f_\text{q}$). (c) $f_\text{q}$ as a function of $\varepsilon$ for $B_\text{ext}=604\ \text{mT}$ (red) and $B_\text{ext}=605\ \text{mT}$ (blue). The left arrows indicate the Zeeman energy difference ($\delta f_\text{q}$) generated by longitudinal magnetic field difference ($b_\text{z}$) of micromagnet. The right arrows illustrate the splitting energy difference ($\Delta E_\text{z}$) for different $B_\text{ext}$.} \end{figure} To explain this s-shape feature, we focus on the Hamiltonian $H$ of a single-electron DQD system on the basis $(\ket{L\downarrow},\ \ket{L\uparrow},\ \ket{R\downarrow},\ \ket{R\uparrow})$~\cite{Benito2019FloppingTheory, Benito2017InputOutput}: \begin{equation} H=\frac{1}{2} \left( \begin{matrix} -\varepsilon-E_\text{z1} & -2t_\text{SO} & 2t_\text{c} & 0\\ -2t_\text{SO} & -\varepsilon+E_\text{z1} & 0 & 2t_\text{c}\\ 2t_\text{c} & 0 & \varepsilon-E_\text{z2} & 2t_\text{SO}\\ 0 & 2t_\text{c} & 2t_\text{SO} & \varepsilon+E_\text{z2} \end{matrix} \right). \label{eq:1} \end{equation} Here, $2t_\text{c}$ is the interdot tunnel coupling, $E_\text{z}=(E_\text{z1}+E_\text{z2})/2$ is the averaged Zeeman energy, $\delta E_\text{z}=(E_\text{z1}-E_\text{z2})/2=g\mu_\text{B} b_\text{z}$ is the Zeeman energy difference generated by longitudinal magnetic field difference ($b_\text{z}$) of the micromagnet, $2t_\text{SO}=g\mu_\text{B}b_\bot$ is the synthetic spin-orbit coupling induced by transverse magnetic field difference ($b_\bot$). The eigenenergies of this four-level system are shown in Fig.~\ref{fig:2}(a). The avoided crossings at $\varepsilon=0$ are generated by 2$t_\text{c}$. By diagonalizing the Hamiltonian in Eq.~(\ref{eq:1}), we calculate the energy splitting ($E_\text{s}$) between the lowest two energy levels as $f_\text{q}\equiv E_\text{s}/h$. $\delta f_\text{q}\equiv 2\delta E_\text{z}/h$ for far detuned limits is shown in Fig.~\ref{fig:2}(c). For the limitation of a small inhomogeneous field, i.e., $b_{\bot,z}\ll |\Omega-E_\text{z}|$, where $\Omega=\sqrt{\varepsilon^2+4t_\text{c}^2}$, $E_\text{s}$ is corrected by the transverse and longitude gradient to second and first order, respectively~\cite{Benito2019FloppingTheory}: \begin{equation} E_\text{s}\simeq E_\text{z}-\frac{E_\text{z}^2-\varepsilon^2}{2E_\text{z}(\Omega^2-E_\text{z}^2)}(g\mu_\text{B}b_\bot)^2-\frac{\varepsilon}{\Omega}g\mu_\text{B}b_\text{z}. \label{eq:energy splitting} \end{equation} \begin{figure}[b] \includegraphics{fig3.pdf \caption{\label{fig:3} (a) The Rabi chevron is acquired at $\varepsilon = 0.5\ \text{meV}$. Every point of $P_\uparrow$ is obtained by repeating the pulse sequence 300 times with $\uptau_\text{B}$ fixed. We repeat the measurement 10 times for every $\uptau_\text{B}$, and sum the results. (b) The Rabi frequency ($f_\text{Rabi}$) is plotted as a function of $\varepsilon$ with fixed microwave power. $f_\text{Rabi}$ is an order of magnitude larger at around $\varepsilon=0$ than the far detuned position. The solid curve is the fitting result of Eq.~\ref{eq:f_rabi}. (c) Rabi oscillations obtained at different $\varepsilon$. Top panel: $\varepsilon=0$ meV, $f_\text{Rabi}=1.262\pm0.002$ MHz and $T_2^\text{Rabi}=6.46\pm0.39$ $\upmu$s; Middle panel: $\varepsilon=1.5$ meV, $f_\text{Rabi}=0.429\pm0.003$ MHz and $T_2^\text{Rabi}=5.53\pm0.57$ $\upmu$s; Bottom panel: $\varepsilon=3.0$ meV, $f_\text{Rabi}=0.135\pm0.003$ MHz and $T_2^\text{Rabi}=7.01\pm0.82$ $\upmu$s.} \end{figure} We plot $f_\text{q}$ as a function of $\varepsilon$ for $B_\text{ext}=$ 605 and 604 mT in Fig.~\ref{fig:2}(c). By fitting $f_\text{q}$ with Eq~\ref{eq:energy splitting}, we obtain $2t_\text{c}=914\pm167$ and $705\pm40$ GHz for $B_\text{ext}=$ 605 and 604 mT, respectively. The difference between the fitted splitting energy ($\Delta E_\text{z}=19.790\pm0.002-19.760\pm0.001$ GHz) equals the difference between the external magnetic fields ($g\mu_\text{B}\Delta B_\text{ext}/h=28$ MHz). \subsection{\label{sec: Rabi oscillation}Rabi oscillation and Ramsey fringes} After calibrating $f_\text{q}$, we now use a microwave burst with a specific burst time ($\uptau_\text{B}$) to manipulate the spin qubit. First, we measure $P_\uparrow$ as an $f_\text{s}$ function with a fixed $\uptau_\text{B}$. Each point of $P_\uparrow$ in the curve is averaged from 300 repeated single-shot readouts. Then, we repeat the measurement ten times and sum $P_\uparrow$ with $\uptau_\text{B}$ changing from 0 to 4 $\upmu$s. The Rabi chevron is plotted in Fig.~\ref{fig:3}(a). Fig.~\ref{fig:3}(b) illustrates $f_\text{Rabi}$ as a function of $\varepsilon$. $f_\text{Rabi}$ is symmetric about $\varepsilon=0$ and is an order of magnitude larger at $\varepsilon=0$ than the far detuned position. For every $\varepsilon$, the corresponding $f_\text{Rabi}$ are extracted by fitting Rabi oscillation with the function $P_\uparrow(\uptau_\text{B})=A\cdot\text{exp}(-\uptau_\text{B}/T_2^\text{Rabi})\cdot\text{sin}(f_\text{Rabi}\uptau_\text{B})$, as shown in Fig.~\ref{fig:3}(c). \begin{figure}[t] \includegraphics{fig4.pdf \caption{\label{fig:4} (a) Ramsey fringes as a function of frequency detuning ($\Delta f_\text{s}$) and waiting time ($\uptau_\text{wait}$) is measured at $\varepsilon = 0.5\ \text{meV}$ by applying a standard Ramsey fringes sequence with two $\pi/2$ pulses. Every point of $P_\uparrow$ is obtained by repeating the pulse sequence 300 times. We repeat the measurement ten times for every $\uptau_\text{wait}$ and sum the results. (b) The spin dephasing times ($T_2^*$) were measured through Ramsey fringes for different $\varepsilon$. Top panel: $\varepsilon=0$ meV, $T_2^*=0.42\pm0.31$ $\upmu$s; Bottom panel: $\varepsilon=1.5$ meV, $T_2^*=0.43\pm0.02$ $\upmu$s.} \end{figure} For a typical flopping-mode EDSR process, Ref~\cite{Benito2019FloppingTheory} gives $f_\text{Rabi}$ as a function of $\varepsilon$ for small $b_\bot$: \begin{equation} f_\text{Rabi} = 4t_\text{c}^2g\mu_\text{B}b_\bot\Omega_\text{c}/\Omega|\Omega^2-E_\text{z}^2|. \label{eq:f_rabi} \end{equation} Here, $\Omega_\text{c}=edE_\text{ac}/\hbar$ is the Rabi frequency in the standard EDSR regime, proportional to the distance between the two QDs $d$, and the electric field with amplitude $E_{ac}$. $\Omega_\text{c}=15.8\pm0.8$ GHz can be obtained from the relevant result of $f_\text{Rabi}$ with a $g\mu_\text{B}b_\bot$ estimated as 0.232 $\upmu$eV~\cite{Zhang2021Synthetic}. We estimate $d\sim$ 0.02 $\upmu$m, thus $b_\bot\sim0.1$ T/$\upmu$m$\cdot 0.02$ $\upmu$m $=2$ mT. Fig.~\ref{fig:3}(c) shows details of Rabi oscillations for different $\varepsilon$. $f_\text{Rabi}=1.262\pm0.002$ MHz is achieved in the top panel. When $\varepsilon$ increases to 1.5 and 3 meV, the Rabi frequencies decrease to $f_\text{Rabi}=0.429\pm0.003$ and $f_\text{Rabi}=0.135\pm0.003$ MHz, respectively. By fitting Rabi oscillation to an exponentially decaying sinusoid, $T_2^\text{Rabi}=6.46\pm0.39$ $\upmu$s at $\varepsilon=0$, $T_2^\text{Rabi}=5.53\pm0.57$ $\upmu$s at $\varepsilon=1.5\ \text{meV}$ and $T_2^\text{Rabi}=7.01\pm0.82$ $\upmu$s at $\varepsilon=3\ \text{meV}$ are obtained. $T_2^\text{Rabi}$ is stable when $\varepsilon$ increases. We next measure $T_2^*$ for different $\varepsilon$ through Ramsey fringes. In Fig.~\ref{fig:4}(a), the Ramsey fringes are measured in the same way as the Rabi chevron. The averaged $T_2^*=0.42\ \upmu$s for $\varepsilon=$ 0 and 1.5 meV is acquired as shown in Fig.~\ref{fig:4}(b). In Ref.~\cite{Croot2020FloppingSiGe}, the lowest $E_\text{s}$ occurs near $\varepsilon=0$, leading to a sweet spot for spin dephasing. However, in our device $2t_\text{c}\gg E_\text{z}$, the second-order item $\frac{E_\text{z}^2-\varepsilon^2}{2E_\text{z}(\Omega^2-E_\text{z}^2)}(g\mu_\text{B}b_\bot)^2$ in Eq~\ref{eq:energy splitting} is suppressed, and there is no sweet spot around $\varepsilon=0$. Here, we attribute the enhancement of the quality factor ($Q\equiv2T_2^\text{Rabi}f_\text{Rabi}$) to the improvement of the electric dipole, since $b_\text{z}$ is constant during the measurement. As mentioned in Ref.~\cite{Zhang2021Synthetic}, longitudinal magnetic field difference $b_\text{z}$ is one of the most relevant sources for dephasing in our device. \subsection{\label{sec:Reproduce}Reproduce the results} \begin{figure}[b] \includegraphics{fig5.pdf \caption{\label{fig:5} (a) Charge stability diagram of the DQD as a function of gate voltages $V_\text{LP}$ and $V_\text{RP}$. The black dashed arrow illustrates the direction of $\varepsilon$ between the second DQD. The relative voltage magnitude at each step of the pulse sequence for qubit manipulation is represented by the black line between two circles. (b) EDSR spectra for the probability of $\ket{\uparrow}$ electron ($P_\uparrow$) as a function of $\varepsilon$ and $f_\text{s}$. The yellow triangles mark the regime of the increased $P_\uparrow$, which represents the approximate position of $f_\text{q}$. The solid curve fitting to the Eq.~\ref{eq:energy splitting} yields $2t_\text{c}=71\pm10$ GHz.} \end{figure} Finally, we try to reproduce the flopping-mode EDSR in another DQD. Fig.~\ref{fig:5}(a) shows the charge stability diagram of another DQD under gate LP and RP. We tune the tunneling rate of the electron from QDs to the reservoir under LL with gate LP and BC. By selectively tuning $V_\text{LB}$, $V_\text{LP}$ and $V_\text{BC}$, we can change $\varepsilon$ from --1 to 0.6 meV. Fig.~\ref{fig:5}(b) shows the corresponding EDSR spectra and $2t_\text{c}=71\pm10$ GHz is obtained by fitting the solid curve to Eq.~\ref{eq:energy splitting}. As shown in Fig.~\ref{fig:5}(b), $P_\uparrow$ background rises when $\varepsilon<0$, and the peak of the increased $P_\uparrow$ becomes invisible on the EDSR spectra. Here, we select $f_\text{q}$ manually and mark them with the yellow triangles. Some points around $\varepsilon=0$ deviate from the curve without affecting the overall s-shape feature. However, the weak visibility makes subsequent measurements around $\varepsilon=0$ of Rabi oscillation and Ramsey fringes impossible. \section{\label{sec:Conclusion}Conclusion} In summary, we demonstrate the flopping-mode EDSR in a Si-MOS quantum dot through the Elzerman single-shot readout. We construct a DQD with $2t_\text{c}\sim800$ GHz under adjacent electrodes by selectively setting gate voltages. We extract an s-shape $f_\text{q}$ as a function of $\varepsilon$ from the EDSR spectra. Then, we improve $f_\text{Rabi}$ an order of magnitude from $0.135\pm0.003$ to $1.262\pm0.002$ MHz by increasing the electric dipole. And $T_2^*$ and $T_2^\text{Rabi}$ remains around $0.42\pm0.03$ $\upmu$s and $6.46\pm0.39$ $\upmu$s, respectively. Moreover, we have shown that the s-shape EDSR spectra holds in another DQD with $2t_\text{c}\sim70$ GHz. We anticipate that flopping-mode EDSR will have better performance in the heavy hole regime~\cite{Mutter2021HeavyHoleFlopping} or phosphorus donor qubits~\cite{Krauth2022PhosphorusFlopping}, and will perform two-qubit operation~\cite{Cayao2020FloppingTwoQubits} in the future. \begin{acknowledgments} This work was supported by the National Natural Science Foundation of China (Grants No. 12074368, 92165207, 12034018 and 61922074), the Anhui Province Natural Science Foundation (Grants No. 2108085J03), the USTC Tang Scholarship, and this work was partially carried out at the USTC Center for Micro and Nanoscale Research and Fabrication. \end{acknowledgments}

35272ead7fa9edacfde3afe8113fc0c539ec8981

\section{Related Work} \label{sec:related} We give related works regarding direct volume rendering (DVR) of unstructured meshes, data-parallel rendering, and in~situ visualization. \subsection{Unstructured Volume Rendering} There are notable works proposed to render unstructured finite element meshes~\cite{nelson2006isosurface, Vo2007irun, marmitt2008cpuvolume, muigg2011interactivevol}. Two predominant strategies to render unstructured volumes are \emph{point-query sampling}, e.g.,~\cite{morrical2020rtxpointlocext} and \emph{ray-marching}~\cite{shirley1990rastertet}. In point-query sampling, zero-length rays are used to probe into an acceleration data structure, for example, a \emph{bounding volume hierarchy} (BVH), that was built over the elements to sample the volumetric data adaptively. These structures are traditionally used to find ray-particle collisions per frame for all pixels. Rathke et al.~\cite{rathke2015simd} propose a min/max BVH that speeds up the element look-up processes for samples and iso-surfaces. Wald et al.~\cite{wald2019rtxpointloc} use point location queries on tetrahedral meshes by utilizing NVIDIA's ray-tracing (RT) cores. This work is later extended by Morrical et al.~\cite{morrical2020rtxpointlocext} to include all unstructured elements. Due to the low number of samples taken per ray, these approaches can produce fast but noisy results, thus requiring many samples to be taken over time for a converged image. To further accelerate the process of convergence and sampling, empty space skipping~\cite{kruger2003accelerationgpu, hadwiger2018sparseleap} or adaptive sampling strategies~\cite{Wang2020noveladaptivesampling,szirmaykalos2011freeps} can be leveraged. The RTX hardware also can be exploited for empty space skipping and adaptive sampling~\cite{morrical2019spaceskip}. Standard ray-marching accumulates many samples while tracing rays without using external acceleration structures~\cite{marmitt2006fastraytet, weiler2003hardwarebased}. Usually, \emph{marching} is performed via visibility sorting or element connectivity. A well-known way to render tetrahedral meshes without connectivity information is by Shirley and Tuchmann~\cite{shirley1990rastertet}. Since visibility sorting tends to be very costly, several researchers turned their attention to connectivity storage, eliminating the need for sorting. Aman et al.~\cite{aman2021bth, aman2021compact} introduce a tetrahedra traversal algorithm that optimizes intersection tests by using 2D projection while still maintaining a connectivity list. However, most of these works are limited to pure tetrahedral meshes. Muigg et al.~\cite{muigg2011interactivevol} propose a ray-marching algorithm that can handle non-tetrahedral elements and non-convex bounding geometry by storing compact face-based connectivity lists and projecting vertices to a ray-centric coordinate system for intersections. When doing ray-marching, one also needs to find the first element where the ray first enters the volume; \cite{sahistan2021Shell} showed how this can be done with RTX hardware by building a BVH over the shell and tracing rays. Ray-marching techniques that utilize connectivity are particularly appealing to our purposes because many modern simulation systems already store connectivity data. Therefore they can be used to avoid worsening high memory pressure situations. \subsection{Data-parallel Rendering} Parallel approaches should be exploited to visualize massive simulation data sets with proper timings. There are various means to partition the workload and data among many compute nodes. Distributing data pieces (clusters) between nodes (i.e., data-parallel rendering or \emph{sort-last}) is a popular method employed by recent works~\cite{larsen2015raytracingdataparallel, castanie2006distributedsharedmemory}. Some works also propose an image-order partitioning (\emph{sort-first}) where work is distributed over pixel regions~\cite{brownlee2013imageparallel,biedert2018hwacceleratedmultitile}. There are also hybrid approaches~\cite{biedert2017taskbasedparallel, cao2019parallelvis}, which aim to address load-balancing issues by leveraging both perspectives. Sort-last algorithms allow for a static geometry assignment at the cost of exchanging intermediate images. Because of the static geometry assignment, sort-last is the most popular rendering algorithm on distributed memory systems. However, correctly and efficiently compositing intermediate images that generally overlap is challenging. Image-based compositors, such as IceT\cite{icet, moreland2011icet}, produce a single intermediate image per node, which is not suitable for clusters with non-convex domains. One of the recent works that tackle the sort-last compositing problem is by Grosset et al.~\cite{grosset2016imagecompositing}, which reduces delays and communications by implementing a spatiotemporally-aware compositor. Their approach uses ``chains'' that determine the blending order of each strip of the image. Usher et al.~\cite{usher2019distributedfb} introduce \textit{Distributed FrameBuffers}, a method that breaks the image processing operations into tiles of ranks with independent dependency trees. The Galaxy framework~\cite{abram2018galaxy} displays the idea of an asynchronous frame buffer, which leverages independent pixel updates sent from a server while allowing incremental refinements to the final image over time. Many recent works are proposed for optimizing workloads minimizing communication costs while generating images with the highest possible accuracy. Ma~\cite{liukwan1995parallelvolumedistributed} introduces a data-parallel unstructured volume rendering method with the ability to handle non-convex data boundaries properly. Similar to our shells (see \autoref{sec:shell-to-shell}), their technique makes use of a \emph{hierarchical data structure} that allows accessing the boundary faces and ray-casting operations from these faces. This work also describes how to do compositing in the correct order. However, unlike our deep compositing (\autoref{sec:deep-compositing}), they prefer sending smaller many messages between compute nodes during rendering. Some of these ideas are later extended to utilize asynchronous load balancing via object and image-order techniques~\cite{liukwan1997scalableparallelcell}. However, this work uses cell-projection techniques rather than ray-casting. Childs et al.~\cite{childs2006hybridmassive} layout a two-stage framework that first samples a $m \times n\times k$ view-aligned grid ---where m and n denote the pixel resolution and $k$ is the sample per pixel--- then composites these samples in the proper viewing order. In the sampling stage, first, they sample what they consider to be small-sized elements. Then, they distribute the large elements to processors, where they are sampled to balance the load. This work is later extended by Binyabib et al.~\cite{binyabib2019hybrid} by proposing a many-core hybrid scheme where they employ sampling over a similar view-aligned grid, but this scheme allows successive $k$ samples in the same pixel and node to be partially composited, reducing the memory footprint. Our deep compositing algorithm is an extension of the algorithms by Childs and Binyabib et al. in the regard that ours can also handle jagged cluster boundaries. However, the view-aligned grid-based sampling is infeasible for our purposes as, in theory, it will waste precious memory resources for large framebuffers. In theory, the 3D rasterization process required by Childs et al. will also be sensitive to overdraw when millions of elements fall within the same grid cell. Finally, both of these works' image-order load balancing method requires large elements to be either replicated or moved to some other nodes, thus requesting additional memory, which may not always be present given an in~situ scenario. We also acknowledge GPU architectures improve with new divergence handling methods and ray-tracing (RT) cores, so the adaptations we propose in this work are necessary. Unlike these prior works, our method is tailored for modern GPUs, which minimizes the costs of compositing and sorting operations. We also do not have to buffer every sample along the pixels since we ray-march through each segment to determine partial samples. Finally, unlike these works, our approach does not require re-distributing or replicating elements across nodes to render the data. \subsection{In~situ Visualization} File I/O has long been a bottleneck of high-performance computing. To overcome this hindrance, in~situ visualization couples computation and visualization together, thus, enabling the users to tap into a running simulation. In~situ visualization has many merits, such as examining the data, doing numerical queries, and generating graphical outputs while the simulation executes. Moreover, it allows verification so that the simulation may be stopped or modified, saving both time and computation resources~\cite{childs2020terminsitu}. We find our approach in line with in~situ applications because of our ability to generate correct images with little to no support from additional acceleration structures at interactive rates. There are notable in~situ applications used in industry~\cite{ayachit2015catalyst, kuhlen2011libsim} that render outputs generated by infrastructures like Strawman~\cite{Larsen2015strawman} or Ascent~\cite{larsen2017alpine}. In addition to standard systems, various recent algorithmic novelties have been proposed to handle time-varying data generated by the simulations. Yamoka et al.~\cite{yamaoka2019adaptivetimestepinsitu} illustrate a method that adapts the timestep sampling rate according to variations in the probability distribution function (PDF) estimation of the connected simulation. Aupy et al.~\cite{aupy2019highthroughputinsitu} give a model that allows them to analyze simulations, and then they use this model to formulate high-throughput scheduling. DeMarle and Bauer~\cite{demarle2021temporalinsitu} propose a temporal cache scheme that keeps much time-varying information produced from a running simulation, which can later be stored according to a pre-defined trigger. Marsaglia et al.~\cite{marsaglia2018explorativevis} introduce an error-bound in~situ compression scheme that allows saving complete spatiotemporal simulation data. Our proposed method only requires a couple of lightweight structures alongside what is already being kept in simulations. Furthermore, observing recent trends from these approaches, we see no major issues that stop our method from being used alongside current in~situ systems. \section{Problem Statement} \label{sec:problem} Modern simulation data is becoming more extensive and complex each day. With the unstructured volume data sets, like The Fun3D Mars Lander that contains many parts (i.e., \emph{clusters}) with non-convex boundaries (see \autoref{fig:overview} (b)), robust data-parallel solutions are needed. Moreover, the generation of such data sets require carefully tuned simulations. In~situ visualization can be utilized to verify the correctness of simulations by allowing visuals to be taken in simulation-time. However, due to the different requirements of the simulation and visualization algorithms, the volume rendering at interactive rates can be challenging. One major problem of such simulations is that the data distribution is generally unbalanced for visualization purposes. Due to time and memory costs, re-distributing the data does not offer a feasible option. Besides, allocating solely visualization-related acceleration structures over all elements may not be possible since nodes may not have enough space. The data-parallel rendering requires each partition to be on a separate computer node, where each node renders a portion of the final image. These portions are called \emph{fragments}, and in our approach, they are generated per \emph{ray segment}. Ray segments are defined between an entry and exit position of a \emph{shell} (faces defined by cluster boundaries), so for non-convex cluster boundaries, there might be more than one fragment since there might be more than one ray segment. Furthermore, these non-convex shells can be on different nodes and interleave each other, which makes the correct order compositing extremely difficult. \begin{figure*} \centering \resizebox{0.99\textwidth}{!}{ \textembeddedimg{drawings/shell-to-shell.pdf}{0.24\textwidth}{}{}{a}{black}\hfill \textembeddedimg{png/shellvis-10-56v2.png}{0.24\textwidth}{}{}{b}{white}\hfill \textembeddedimg{drawings/marching.pdf}{0.24\textwidth}{}{}{c}{black}\hfill \textembeddedimg{png/shellvisdvrsurf.png}{0.24\textwidth}{}{}{d}{black} } \caption{Overview of the volume integration process. (a) Shell-to-shell traversal for non-convex elements: rays with front-face culling are sent to find exit faces, and backward rays are cast from exit faces to find the entry faces over the shells. The blue and light blue clusters lie in the same MPI rank, but the yellow cluster is in another rank; hence, the shells of the two blue clusters are traced in order, but the yellow cluster's shells get traced in parallel. (b) The shells of the first 46 clusters of Small Lander. Each base color represents a different cluster's shell. (c) The ray-marching process is illustrated for the two blue clusters given in (a). (d) DVR of the same subset given in (b). } \label{fig:overview} \end{figure*} \section{Proposed Approach} \label{sec:method} We tackle the given problem by introducing a framework that supports interactive visualization of large, unstructured, and non-convexly partitioned volumetric data sets, animation of fixed topology data sets, and compatibility with the given data partitioning and the number of nodes in the simulations; i.e., native data, without re-partitioning or simplifications, and proper compositing even in the presence of non-convex data. Assuming that the data is natively pre-partitioned and distributed to different ranks, the proposed approach for rendering consists of four steps: \begin{enumerate} \item Each node generates connectivity information, shell-BVH, and XOR-compacted geometry representation required by our ray-marcher (\autoref{sec:data-preparation}). \item Rendering starts at each node by tracing two rays through each shell to create segments (\autoref{sec:shell-to-shell}). \autoref{fig:overview}~(a) illustrates the shell-to-shell traversal. \item Each node performs volume integration (cf.~\autoref{fig:overview}~(c)) via ray-marching, creating one RGBA-Z tuple; i.e., fragment per each segment, resulting in potentially multiple fragments for each pixel (\autoref{sec:integrate-segments}). \autoref{fig:overview} (d) depicts an integrated volume output using the shells from \autoref{fig:overview} (b). \item Finally, we apply a GPU-optimized ``deep compositing'' technique in which different ranks exchange their respective fragments and composite them in the proper order (\autoref{sec:deep-compositing}). \end{enumerate} In an offline rendering mode, each step is executed once in the given order; however, the last three steps are repeated repeatedly under interactive exploration. Besides, the first step's results can be cached for future use. \subsection{Data Preparation} \label{sec:data-preparation} We describe the connectivity information generation, shell-BVH construction and XOR-compacted geometry representation creation required by our ray-marcher. \subsubsection{Connectivity Generation} \label{sec:connectivity-gen} Our method needs to know the neighboring elements' indices to perform element marching. We generate the connectivity information by matching the element faces in the preprocessing step. We separate vertex and connectivity information and store the neighbor indices in an external buffer to keep the elements and neighbor indices aligned in memory. Although we picked this way of processing connectivity it should be noted that this buffer can be in any shape or form as long as one can access the next element from the current element using a face that is shared by both. Thus this part can be adapted to fit simulation or application's needs. \subsubsection{Per-Node Shell Generation} \label{sec:per-node-shell-gen} Our approach handles volumetric data that may contain convex and non-convex clusters. To this end, we identify boundary geometry for each cluster present per node by looking at elements which are missing a neighbour from the connectivity generation step. This boundary geometry comes in the form of triangles and quads, which we call \emph{shell-faces}. We utilize a method similar to the one described in Sahistan et al.~\cite{sahistan2021Shell}, where we identify each shell-face using connectivity. We mark the faces from elements with missing neighbors as shell-faces. We use triangles as provided and triangulate quadrilateral elements. We keep a list of triangle indices stored along the shell-BVH. We reserve four indices for each shell triangle: the first three are triangle indices, and the last one points to the volume element behind that triangle. The lower two bits of the fourth index signifies the element type (i.e., tetrahedron, pyramid, wedge, or hexahedron), and the remaining 30 bits is an index into the list of elements; this index is required to start marching. This encoding is similar to the BVH-node memory layout used by PBRT~\cite{pharr2021pbrtaccel}. We build our shell-BVH using OptiX~\cite{optix, wald2020owl} to exploit NVIDIA GPU's RTX cores for hardware-accelerated shell-to-shell traversal. \subsubsection{XOR-compaction} \label{sec:xor-comp} Our compaction scheme exploits the following property of exclusive-or (XOR) operations: $(a \oplus b) \oplus a = b$. We can generalize this property to $n$ numbers if we know the XOR of $n-1$ numbers. Let $X$ denote the XOR of $n$ terms and $Y$ denote the XOR of any subset with $n-1$ terms. Then, we can simply XOR $X$ and $Y$ to find the remaining term. We can exploit this idea on connected volume elements in a ray's path during ray-marching. Since we know that some of the vertices are shared between neighboring elements, previously calculated XOR fields can be employed to reduce index information per element. This compaction requires the first face to be known to start ray-marching since all the other steps depend on the information obtained from the previous step. To handle this initial case, we use our shell faces that we store explicitly. After that, each step utilizes the march state to access the previous step's information. For each element except hexahedra, we store a different XOR-compacted structure, illustrated in \autoref{fig:memorylayouts}. \begin{figure}[htbp] \begin{minipage}[t]{.22\columnwidth} {\small \begin{lstlisting}[language=c++] struct Tet{ uint vx; }; \end{lstlisting} } \end{minipage} \hfill \begin{minipage}[t]{.22\columnwidth} {\small \begin{lstlisting}[language=c++] struct Pyr{ uint dx; uint diag[2]; uint top; }; \end{lstlisting} } \end{minipage} \hfill \begin{minipage}[t]{.22\columnwidth} {\small \begin{lstlisting}[language=c++] struct Wed{ uint dx[2]; uint diag[2]; }; \end{lstlisting} } \end{minipage} \hfill \begin{minipage}[t]{.22\columnwidth} {\small \begin{lstlisting}[language=c++] struct Hex{ uint v[8]; }; \end{lstlisting} } \end{minipage} \begin{center} \fbox{% \includegraphics[width=0.28\columnwidth]{drawings/tet.pdf} } \fbox{% \includegraphics[width=0.28\columnwidth]{drawings/pyr.pdf} } \fbox{% \includegraphics[width=0.28\columnwidth]{drawings/wed.pdf} } \end{center} \caption{XOR-compacted memory layouts (top) and geometric illustrations of XOR calculations for \code{Tet}, \code{Pyr}, and \code{Wed} (bottom). \code{Hex} does not have a compaction scheme. \texttt{uint} stands for unsigned integer. The ``$\oplus$'' symbol indicates the XOR operation. The total sizes of each struct are 4, 16, 16, and 32 bytes for Tet, Pyr, Wed, and Hex, respectively. The vertices are in VTK~\cite{VTK} ordering.} \label{fig:memorylayouts} \end{figure} We apply the tetrahedron compaction method by Aman et al.~\cite{aman2021bth}. A tetrahedron shares three of its four vertices with any other neighboring element. For this reason, a single XOR field is enough to construct the unshared vertex. Given the entry face that contains vertex indices $v_0, v_1, v_3$ and a compacted tetrahedron with $vx = v_0 \oplus v_1 \oplus v_2 \oplus v_3$ one can XOR all four of the integers to get the missing vertex index $v_2$. For the 16 byte pyramid, we store one \code{dx} field that is the XOR of \nth{0} and \nth{2} vertex indices (according to VTK ordering), two vertex indices which happens to be the other diagonal of the quad (\nth{1} and \nth{3} vertices), and a top vertex index, which is always the \nth{4} vertex. During marching if the entry is from the quad face, the only missing index is the top vertex index, which is already explicitly stored. Otherwise, any triangle face should be composed of one of the explicitly stored diagonal vertices, the top vertex, and one vertex encoded in the \code{dx} field. Matching one of the vertices with one of the diagonal fields, we can decide which index to XOR with \code{dx}, thus obtaining one of the missing vertices. The other missing index for this case is the unmatched integer from \code{diag[2]}. Our 16 byte wedge structure is composed of two \code{dx} and two \code{diag} fields. \code{dx} fields contain two XORs: the first one is the XOR of \nth{2} and \nth{3} vertex indices; the second one is the XOR of \nth{1} and \nth{5} vertex indices. \code{diag} explicitly stores \nth{0} and \nth{4} vertex indices. Like pyramids, ray-marcher's wedge construction has two high-level cases. If the entry is from a triangular face, any triangle face should be composed of one of the explicitly stored diagonal vertices and two vertices encoded in the different \code{dx} fields. By matching one of the diagonal vertex indices to one of the diagonal fields, we can construct two missing vertices from \code{dx} fields. The other missing index for this case is the unmatched integer from \code{diag[2]}. If the ray enters from a quadrilateral, it must contain one or both of the indices stored in \code{diag}. By matching diagonal vertices, we can determine the entry quadrilateral. Then, we have two cases. The first case where two of the entry quadrilateral's indices match both of the indices is shown in \code{diag}. We can get the two missing indices by XOR'ing \code{dx} fields with unmatched indices of the entry quadrilateral separately. In the second case (either one of the diagonals match with one of the quadrilateral face indices), we can immediately get one of the missing vertices from unmatched \code{diag}. Finally, we can use one of the \code{dx} fields to get the other missing vertex. Hexahedra have the minimum shared vertices ratio (0.5) among all element types and demand that four vertices be obtained on entry. It is challenging to find an XOR-based hexahedra compaction scheme that reduces to the closest alignment of 16 bytes. Therefore, we store all hexahedra indices without compaction according to the VTK mesh ordering. \subsection{Per-Node Segment Generation via Shell Traversal} \label{sec:shell-to-shell} For each given shell-BVH per node, we initiate a step called \emph{shell-to-shell} traversal. This process is done for every ray per cluster and is fundamentally similar to~\cite{sahistan2021Shell}, which finds a \emph{segment} between entry and exit faces. \autoref{fig:overview} (a) illustrates this process, and the steps are as follows: \begin{enumerate} \item Trace the ray through the shell-BVH with front-face culling from the ray origin. \item If a ray hits a shell face, we then mark that face as the exit face and create a backward ray with the origin at the hit position. \item This backward ray is again traced using front-face culling to find an entry face. \item The found entry face contains four index values (\autoref{sec:per-node-shell-gen}), and the last one encodes the ID and the type of the element from where we start our ray-marcher (\autoref{sec:integrate-segments}). \end{enumerate} Some real-life data sets might have degenerate volume boundaries where instead of tightly interlocking, two neighboring faces might be intersecting or slightly apart. If we were to na\"{i}vely to find the closest hits for these degenerate boundaries, this would create incorrect ray segments, which might cause sampling and compositing errors ---casting two front-face culled rays to find an exit and entry point allows us to handle them robustly. Although we leverage the hardware acceleration of NVIDIA RTX GPUs in our work, this approach does not explicitly require the usage of OptiX/OWL frameworks or any specialized hardware. One can use any other framework or ray-tracing engine that supports these basic functionalities. \subsection{Per-Segment Volume Integration} \label{sec:integrate-segments} We use linear interpolation to sample elements at equidistant points in a segment. The coefficients to linear interpolation calculations are also utilized to check whether the current element contains the point that needs to be sampled. If not all of these coefficients are between 0 and 1, we keep marching until that becomes the case. When a sample is taken, a transfer function is used to look up its color and transparency, and then it is composited to that segment's color. Marching is terminated when a ray becomes opaque or the next sample position falls behind the exit face. Our marcher utilizes a method similar to ``Projected Tetrahedra''~\cite{shirley1990rastertet} to determine the exit face for a given element. We do not rasterize the elements directly to the screen, which is more in line with the methods described by Aman et al.~\cite{aman2021bth, aman2021compact} and Sahistan et al.~\cite{sahistan2021Shell}. However, we handle primitives other than tetrahedra as well. We employ XOR-based compaction schemes to reduce the memory footprint of the data while still allowing efficient traversal. Our compaction process reduces the vertex index storage per element, except hexahedra. We also address memory alignment with this scheme. We store connectivity information in a separate buffer to preserve memory alignment properties. Although reducing memory usage is usually helpful, this index removal may not be desirable for some in~situ scenarios; our marcher can perform without compaction. Ray-marching processes start from a cluster's shell that contains pbrt-style~\cite{pharr2021pbrtaccel} encoded element information. We can construct the first element behind the shell face using this information. After entering the shell, the connectivity buffer and compacted element information are enough to fetch and construct the next elements along the given ray segment. However, our compaction requires elements to be traversed in sequential order without skipping. When an element is reconstructed from XOR-compacted form, the vertices cannot be in any order because this introduces sampling artifacts. Therefore, our scheme not only re-obtains vertex indices but also consistently places them according to VTK mesh ordering for each element~\cite{VTK}. An exit face must be determined to select the next element on the ray segment. We project the element vertices to a ray-centric coordinate system to conduct 2D intersection tests to find the exit face. Finally, our method maintains a \emph{march state}, which book keeps the last intersected face type (triangle or quad), current element's type, index, and vertex indices for every marching step. Moreover, we follow a general rule while traversing the volume: placing the entry face indices into the same positions in the march state during marching. This rearrangement of the vertices allows us to ignore the entry face during exit face selection(since we now know which vertices belong to entry face). When the marcher leaps to another element, we update march state old entry face indices with exit face indices. Since each volume element is unique in terms of its geometry and face arrangement, it is hard to make a \emph{simple} algorithm that handles all possible combinations. For this reason, our element marching handles various elements in a case-by-case fashion. We handle tetrahedral elements similar to~\cite{sahistan2021Shell}. However, unlike Sahistan et al, we also allow intermediate points inside the elements to be sampled. We transform the vertices to the previously mentioned ray-centric coordinate system to determine the exit face for tetrahedral elements. After each vertex is transformed, we apply a maximum of two 2D left tests to determine the face containing point $(0,0)$. The exit face can again be found via 2D left tests when inside a pyramid. Due to the quad face that the pyramids have, we utilize the last intersected face type field stored in march state to simplify our left test cases. Using projected vertices, if the entry face is a quad face, we find the exit face among four triangles (similar to the tetrahedron case). Otherwise, we first check if the quad face contains the point $(0, 0)$, then test the remaining three triangles for the same condition. Finally, we update the last intersected face type accordingly. In terms of finding the exit face, wedges are similar to pyramids. It should be noted that wedges contain three quad faces; hence even if the intersected face type is a quad, the exit face might be another quad face. Like other element types, we ignore the entry face to avoid extra left tests. After the exit face is determined, we update intersected face type once more. Hexahedra are uniform like tetrahedra; however, they have more faces. Therefore finding the exit intersection requires the highest number of left tests. In the worst case, hexahedra need 13 left tests, whereas wedges, pyramids, and tetrahedra require 7, 5, and 2 left tests, respectively. \subsection{Deep Compositing} \label{sec:deep-compositing} The techniques described in the previous sections allow any rank to efficiently find, for a given ray, all the segments that overlap with that rank's part of the data (\autoref{sec:shell-to-shell}); and to efficiently integrate each of these segments (\autoref{sec:integrate-segments}). This integration step produces one RGB color and opacity value for each segment, plus the depth of the given segment. Borrowing terminology from triangle rasterization, we call each such tuple of color $C$, opacity $\alpha$, and depth $z$, a \emph{fragment} $F=({F_C,\;F_\alpha,\;F_z})$. Given all of a given pixel $P$'s fragments $F^{(P)}_{0},F^{(P)}_{1},\dots,F^{(P)}_{N^{(P)}}$, the correct final pixel color is the result of first sorting these fragments by their depth and compositing them usin $\widehat{O}(A,B)\;$/$\;\widehat{U}(A,B)$~\cite{OverOperator,Porter84}. The challenge is that any given pixel's fragments may get produced on many different ranks, requiring some merging of different ranks' results. Even worse, the irregular shape of the shells means that any ray can enter and leave the same shell multiple times at multiple distances, producing multiple---and in some cases, many---fragments for the same pixel. \autoref{fig:num-frags} illustrates the distribution of the average and the total number of fragments for a view of the Huge Lander for increasing rank counts. As it can be observed from the case where the rank count is 16 (cf.~\autoref{fig:num-frags-vis}), each pixel may have multiple fragments generated from multiple ranks. \begin{figure}[t] \centering \includegraphics[width=\columnwidth]{drawings/fragment_counts_plot.pdf} \caption{Box plots of average (left) and total (right) number of fragments generated by individual ranks while rendering the Huge Lander. We take averages over non-empty pixels where their opacity is greater than 0. The plots are for the rank counts of 16, 32, 48, 64, and 80. Scattered points signify individual ranks' average (left) and total (right) fragment counts at a given MPI size.} \label{fig:num-frags} \end{figure} \begin{figure}[tbp] \includegraphics[width=.50\columnwidth]{png/fragments16_v2.png} \hspace{-0.38em} \includegraphics[width=.50\columnwidth]{png/fragments16_combined_v2.png} \includegraphics[width=1.00\columnwidth]{png/fragments16_v2_colorbar.pdf} \caption{Heatmaps for the number of fragments for a view of the Huge Lander rendered with 16 MPI ranks (first box given in \autoref{fig:num-frags}): On the left, fragments for every 16 rank, and on the right, all heatmaps combined into one image.} \label{fig:num-frags-vis} \end{figure} The easiest approach to compositing this would be to first composite all ranks' fragments to a single fragment per pixel per rank and then use some optimized compositing library like IceT~\cite{moreland2011icet} to produce the final image. However, as neither $\widehat{O}$ nor $\widehat{U}$ are commutative, this will give wrong results every time a ray enters the same shell more than once. Fragments need to be composited in visibility order, considering that $N$ fragments along a ray are generally distributed unequally across the ranks. Let $\otimes$ denote compositing operation given any ray that produces two fragments $F^{(A)}_0$ and $F^{(A)}_1$ on the same rank must also have had at least one other fragment $F^{(B)}$ on at least one other rank. This requires compositing as $F^{(A)}_0 \otimes F^{(B)} \otimes F^{(A)}_1$, which in general is different from both $F^{(B)} \otimes (F^{(A)}_0 \otimes F^{(A)}_1)$ and $(F^{(A)}_0 \otimes F^{(A)}_1) \otimes F^{(B)}$. \subsubsection{Compositing with More than One Fragment/Pixel} To solve this compositing problem, we developed a new compositing framework that explicitly allows each rank to have multiple fragments per pixel. At an abstract level, our method expects each pixel to store one counter that specifies the number of fragments, $N$, plus an address (or offset) to a list of fragments, $F_0, \; \ldots\;, F_{N-1}$. Similar to parallel-direct-send~\cite{grosset2017todtree, favre2007directsend}, we then split the frame buffer into $R$ distinct \emph{regions} of pixels (where $R$ is the number of ranks); each rank will be responsible for receiving, compositing, and delivering the final composited results of one region of pixels. Compositing then works in the following steps: \def\compactstep#1{\par\medskip\noindent\textbf{#1}} \compactstep{1) Generating a contiguous send buffer.} Given each pixel's fragment lists, each rank computes a GPU-parallel prefix sum over all its fragment counts, which also yields the total number of fragments on this rank. We then allocate a single contiguous memory region for these fragments and compact the individual fragments into this buffer (using the prefix sum result as offsets). By design, this buffer will contain all fragments going to all other ranks in order. \compactstep{2) Exchanging per-pixel fragment counts ranges.} Given the assigned range of pixels, each rank computes which range of per-pixel counters it needs to send to any other rank. To this end, each rank allocates a per-rank counter buffer with size $R$ times the number of pixels in its region. Next, each rank computes the offsets to store the counters from other ranks. We then perform a collective \code{MPI\_Alltoallv} on these buffers, after which each rank has, for its assigned region of pixels, the fragment counts from every other rank. \compactstep{3) Exchanging Fragment Lists.} Having received all other ranks' per-pixel fragment counts for its range of pixels, each rank then performs a GPU prefix sum over those counters, the result of which can once again be seen as offsets into a compact buffer of all fragments for its range of pixels. Looking up the prefix sums at the correct offsets specifies how many fragments each rank will receive from any other rank and how many fragments it will receive altogether. We then allocate a receiving buffer of the required size, look up where each other rank's fragments will go in this buffer, and issue a second \code{MPI\_Alltoallv} that, in this case, collectively moves all fragments into the receive buffer of the rank assigned to that fragment's corresponding pixels. \compactstep{4) Local Compositing.} The result of the previous steps is that each rank now has two buffers containing all fragment lists for its assigned pixels. The first buffer ---\textit{fragment buffer}--- stores all fragments for that rank's pixels received from all other ranks, ordered by ranks and pixels within each rank. Given a specific MPI rank, this buffer stores all fragments for that rank's first pixel from rank 0, then all those for its second pixel from rank 0, and so on, followed by all fragments from rank 1, then all fragments from rank 2, and so on. The second buffer ---\textit{offset buffer}--- stores the results of prefix sum operations. It, by design, provides the offsets where the fragment lists start. For example, if $P$ is the number of pixels that this rank is responsible for, then the fragments from rank $r$ for pixel $j$ start at offset \code{offsets[r*P+j]}. Using this, we can now launch a CUDA kernel that, for each pixel $p$, looks up the $R$ different lists of fragments and composites them in the visibility order. \compactstep{5) Sending final results to master.} The output of the previous CUDA kernel is, on each rank, a fully composited RGBA value for each pixel in that rank's range of pixels. We send these to the master using a \code{MPI\_Send}; the master sets up $R$ matching \code{MPI\_Irecv}s, each of which uses the appropriate part of the final frame buffer as receive buffer. Once these are completed, the master has the final assembled frame buffer, and compositing is complete. This method is a natural extension of the parallel direct-send technique as described by Grosset et al.~\cite{grosset2017todtree} and Favre et al.~\cite{favre2007directsend}, with the main difference that we not only send one fragment per pixel but variable-sized lists of fragments. We term this method \emph{deep compositing} because it merged the concepts of image-based compositing with the orthogonal concept of \emph{deep frame buffers}~\cite{GershbeinHanrahan00}. \subsubsection{Fragment List Management} Though the compositing itself is easy to use from the host side, properly setting up the device-side inputs (fragment lists and counters) would require the renderer to handle what are akin to device-side dynamic memory allocations to manage those per-pixel variable-size fragment lists during rendering. To relieve the renderer of this low-level fragment list management, we also developed what we call a \emph{device interface} for this library, through which a renderer can simply \emph{write} new fragments into a pixel, with that interface then handling the proper storage of those fragments---which significantly simplifies the rendering code. \paragraph{Two-Pass, Flexible-length Fragment Lists} The main challenge for developing this interface was that we could not simply allocate more device memory during rendering, so we needed \emph{some} limit on how many fragments a renderer would be allowed to generate in any frame. We first developed a two-stage interface in which the renderer would be run twice: in the first stage, the interface would only count the fragments produced per pixel but not store any. After this stage, it would compute a prefix sum over those counters to allocate a big enough buffer, with the prefix sum values serving as offsets into this buffer. A second pass would then perform exactly the same rendering but store the fragments at the provided offsets. \paragraph{Single-Pass, Fixed-Length Fragment Lists} The two-pass method allows for arbitrary-sized fragment lists (up to device memory, obviously); but requires running at least the shell traversal twice, which may or may not be acceptable. We, therefore, also developed a second, single-pass device interface in which the renderer---upon initialization---specifies a maximum allowed number of fragments per pixel, which can then be used to pre-allocate lists to add fragments. Having a single pass is straightforward but requires some form of \emph{overflow}-handling if a render wants to submit fragments to a pixel whose list is already full. We currently implement two methods for this overflow handling: In the \emph{drop} method, we perform insertion sort into the existing list and simply drop the latest fragment. In \emph{merge}, we find the fragment with the lowest opacity and perform a \emph{over} compositing of this element onto the one in front of it (i.e., using the depth from the previous one), then insert the new fragment into the list. \subsubsection{Implementation Details} Though primarily developed for this particular application, we believe the method just described is also applicable to other, similar applications, and thus decided to implement this into a stand-alone \emph{deep compositing} library that the rest of our renderer then uses. The compositing itself uses MPI, for which in our application, we use a CUDA-aware version of OpenMPI 4.1. Using CUDA-aware MPI allows the compositing code to directly operate on device buffers, which means that the same library can work with both host and device-side renderers. We use CUDA for the device interface, with a simply host-side interface to initialize and trigger compositing. The bandwidth required for compositing is often a bottleneck in data-parallel parallel rendering, even with only a single fragment per pixel. One step we use to reduce bandwidth is that we allow the user to specify whether to use full \emph{float} precision for fragments (five floats total, for r, g, b, depth, and opacity); or to use a lower-precision encoding with 8-bit fixed-point for RGBA, and floats only for the depth value. The device interface in both cases is the same, but that interface encodes fragments as they get submitted. We also automatically discard fragments with zero opacity value, as these will not contribute to the image. Aside from the fragments, the per-pixel counters require a large bandwidth. To reduce that, we use specialized encodings with 2, 4, 8, or 32 bits for those counters, depending on the longest fragment list length. We use dedicated CUDA kernels for encoding and decoding the 32-bit counter arrays into this lower-precision representation before and after the \code{MPI\_Alltoallv} counter exchange; otherwise, perform the algorithm exactly as described above. \section{Experimental Results} \label{sec:evaluation} We conducted our experiments on Frontera RTX nodes of Texas Advanced Computing Center (TACC), where each of the 22 nodes had four NVIDIA Quadro RTX 5000 plugged into it. We utilize all four GPUs available per node for every data point of our experiments. \subsection{Evaluation of the Framework} We evaluate our rendering framework on Small Mars Lander and Huge Mars Lander data sets. \autoref{fig:teaser} show images of the Small Mars Lander rendered using our framework. Since Small Mars Lander has 72 clusters, we evaluate our framework using 72 GPUs distributed over 18 compute nodes where each cluster is loaded on a separate GPU. For Small Mars Lander, we achieve our peak performance using 72 GPUs yielding the average fps of 14.35. Since the TACC supercomputer does not have more than 22 RTX nodes, we could not test one cluster per GPU scenario for the Huge Mars Lander data set. Therefore, we scale up to a maximum GPU count of 88, yielding 9.83 fps. However, we observe our average peak performance of 10.25 fps for the Huge Mars Lander at 80 GPUs. Moreover, we evaluate our deep compositing scheme's correctness compared to a single fragment compositing technique. \autoref{fig:ice_t_diff} shows an image rendered by single fragment compositing and a heatmap that compares the difference between single fragment compositing and our deep compositing. The single fragment compositing method depicted is similar to the image-based single image per node compositing techniques such as IceT~\cite{moreland2011icet}. \begin{figure}[htbp] \begin{center} \includegraphics[width=0.44\columnwidth]{drawings/small_lander_frame19_ice_t.png} \includegraphics[width=0.44\columnwidth]{drawings/ice_t_diff.png} \includegraphics[height=0.44\columnwidth]{drawings/ice_t_diff_colorbar_vertical.pdf} \end{center} \caption{The left image is the rendering with single fragment compositing (similar to IceT \cite{moreland2011icet}). The right image is the heatmap showing the L2 difference between single fragment compositing and our deep compositing.} \label{fig:ice_t_diff} \end{figure} \subsection{Memory Overhead} We examine data distribution and memory footprints. \autoref{tab:element_shell_stats} displays the minimum, maximum and average counts per rank of volume elements and shell faces for our two data sets. The MPI sizes (rank counts) in the table are the sizes that experience the highest level of changes in terms of rendering times presented in \autoref{fig:timing_plots}. \autoref{fig:memory_plots} illustrates average memory footprints of our large data structures that may not be present in a simulation environment. Although connectivity information will likely be in most of the simulation systems, we wanted to include connectivity here for the sake of simulation systems like~\cite{ishii20194dtreebased}. \begin{table}[htbp] \centering \caption{Per rank statistics for selected MPI sizes for two test data sets. The first column gives the MPI size with a number and data set as ``small'' for Small Mars Lander and ``huge'' for Huge Mars Lander. Columns depict minimum, maximum, and average volume element counts and minimum, maximum, and average shell face counts per rank.} {\small \begin{tabular}{cccc|ccc} \hline \multicolumn{1}{c}{\multirow{2}{*}{\begin{tabular}[c]{@{}c@{}}MPI Size\\ \& data set\end{tabular}}} & \multicolumn{3}{c|}{Elements} & \multicolumn{3}{c}{Shell} \\ \cline{2-7} \multicolumn{1}{c}{} & \multicolumn{1}{c}{min.} & \multicolumn{1}{c}{max.} & \multicolumn{1}{c|}{avg.} & \multicolumn{1}{c}{min.} & \multicolumn{1}{c}{max.} & \multicolumn{1}{c}{avg.} \\ \hline 4-small & 192.2 M & 203.4 M & 199.6 M & 14.4 M & 16.2 M & 15.2 M \\ 24-small & 30.8 M & 34.9 M & 33.3 M & 2.0 M & 2.9 M & 2.5 M \\ 40-small & 10.0 M & 23.3 M & 20.0 M & 0.8 M & 1.9 M & 1.5 M \\ 72-small & 9.9 M & 11.9 M & 11.1 M & 0.6 M & 1.1 M & 0.8 M \\ \hline 16-huge & 365.2 M & 414.6 M & 399.2 M & 21.4 M & 25.1 M & 23.0 M \\ 32-huge & 177.8 M & 213.8 M & 200.0 M & 10.3 M & 13.2 M & 11.5 M \\ 56-huge & 93.8 M & 121.0 M & 114.1 M & 5.4 M & 7.3 M & 6.6 M \\ 88-huge & 60.9 M & 85.1 M & 72.6 M & 3.3 M & 5.3 M & 4.2 M \\ \hline \end{tabular} } \label{tab:element_shell_stats} \end{table} \begin{figure*}[htbp] \begin{center} \fbox{% \begin{minipage}{.48\textwidth} \includegraphics[align=c, width=.40\columnwidth]{png/offlineViewer_frame18.png} \hfill \includegraphics[align=c, width=.58\columnwidth]{drawings/small_plot_stack.pdf} \end{minipage} }% \hfill \fbox{% \begin{minipage}{.48\textwidth} \includegraphics[align=c, width=.40\columnwidth]{png/Huge-offlineViewer_frame18.png} \hfill \includegraphics[align=c, width=.58\columnwidth]{drawings/huge_plot_stack.pdf} \end{minipage} }% \end{center} \caption{Results of the scalability benchmarks for Small Mars Lander (left) and Huge Mars Lander (right) on RTX nodes of the Frontera system on TACC. Integration process timings are stacked over compositing process timings for the given number of ranks to form total rendering times. } \label{fig:timing_plots} \end{figure*} \begin{figure}[htbp] \begin{center} \includegraphics[width=.495\columnwidth]{drawings/small_lander_plot_memory.pdf} \includegraphics[width=.495\columnwidth]{drawings/huge_lander_plot_memory.pdf} \end{center} \caption{Per rank average memory consumption of various buffers for increasing MPI sizes: Small Mars Lander (left) and Huge Mars Lander (right). Average memory usage of XOR-compacted elements is stacked over average shell-BVH size, which again is stacked over average connectivity buffer size, providing the total memory usage introduced by these data. We also include a line that indicates the per rank average memory usage without XOR-based compaction (size of Shell-BVH + connectivity buffer + non-compact elements).} \label{fig:memory_plots} \end{figure} \subsection{Scalability} To assess the scalability, we test our approach for increasing the number of ranks (MPI sizes). We also measure sub-process timings, specifically the segment integration and compositing times. We compute them by using \code{MPI\_Barrier}s before and after integration calls to synchronize the processes before starting and ending the timers. Since these barriers stall early terminating integration processes, it causes the total rendering time to increase. For this reason, we measure the total rendering time and integration time in different runs and derive the compositing time by subtracting the total time from the integration time. We calculate the timings reported in \autoref{fig:timing_plots} as follows: we take an average for 20 sequential timesteps of the selected scalar field over 30 runs. Then we take the mean of these 20 average values to form the data points for the given MPI sizes. \subsection{Fragment Distribution} Since it is crucial to assess workload distribution across compute nodes for both compositing and integration steps, we measure fragment counts generated by each GPU during ray segment generation. To this end, we present a box plot in \autoref{fig:num-frags}, which depicts the total and average fragment counts for specific MPI sizes on the Huge Mars Lander. We also visualize a series of ``heat images'' for the case where the rank count is equal to 16, which illustrates the fragments composited for a view of the Huge Mars Lander. Each small image shows given nodes generated fragment counts, and the final image accumulates them on top of each other. \section{Discussion} \label{sec:discussion} We evaluate our approach regarding memory consumption, rendering correctness, and scalability. Among the data we precompute and store, the connectivity information takes up the lion's share with ratios around $\approx62.72\%$ for our test cases. We expect this since our approach stores one integer for all faces of a given volume element. XOR-compacted volume elements are the second-largest structure with ratios around $\approx17.11\%$, closely followed by Shell-BVH sizes ($\approx12.74\%$) of the total memory consumption. We observe that the total memory footprint of XOR-compacted representations is $\approx72.49\%$ smaller than their uncompacted versions. The presented framework is memory-wise compatible with in~situ scenarios because many modern simulation systems already store connectivity information. Our XOR-compaction reduces the space required for geometry information, and our shell-BVH sizes stay relatively small despite the large counts of shell faces. We observe that simple image compositing is not an option for non-trivially partitioned data sets like we present. As the error metric in \autoref{fig:ice_t_diff} confirms, single image compositing gives inaccurate results. Moreover, we found our compositing process to create low overheads even with the GPU counts going up to 88. Although the communication cost for compositing created an increasing trend in terms of time, it never surpassed 22.68\% of the total rendering time, which is an indication that our deep compositing method is highly scalable and generates correctly composited images interactively (see~\autoref{fig:timing_plots}). Although the proposed ray-marcher is suitable for the use-cases described, any other ray-marcher that can adequately handle non-convex boundaries can be utilized. For instance, point-query sampling techniques that can leverage adaptive sampling or space skipping~\cite{morrical2019spaceskip,Wang2020noveladaptivesampling} may produce much faster results. However, such methods rely heavily on hierarchical data structures to sample the volume, which would create more additional memory overhead. One could claim point-query sampling techniques negate this memory over-head by not storing connectivity; however, as pointed out before, many simulation environments have that data out of the box. Furthermore, point-query sampling techniques usually produce a noisy image that requires some convergence time to be passed, whereas our marching method generates deterministic noise-free images. For these reasons, we consider our marching algorithm to be more pragmatic in the context of data-parallel rendering and deep compositing. Finally, examining data distributions from \autoref{tab:element_shell_stats}, we see that directly utilizing native partitioning of the data causes uneven load balancing for some MPI sizes. \autoref{fig:num-frags} reveals this phenomenon where the average number of fragments per rank distribution varies. The effects of this phenomenon can also be observed in \autoref{fig:num-frags-vis}, where ranks 1, 4, and 14 have significantly fewer fragments than the others. Even though native partitioning causes uneven workloads, our timing experiments (cf.~\autoref{fig:timing_plots}) display decent scalability with the increasing number of GPUs we utilized. We smoothly achieve interactive rates with both of our data sets. For the small Mars Lander that has 72 clusters, we benchmark \emph{14.35 fps} using 72 GPUs, and for the 552-cluster huge Mars Lander, we measure \emph{10.27 fps} using 80 GPUs. We also observe an ongoing downwards trend for the timings with the increasing number of GPUs, so it is worth mentioning that our application can achieve even higher frame rates given a more extensive hardware setup. Furthermore, it is clear from \autoref{fig:timing_plots} that dominating term of rendering times is volume integration via ray-marching. Nevertheless, we observe a sharp increase in integration performance at $n=24$ and $n=32$ for Small and Huge Mars Lander, respectively. At the same time, it is expected for an embarrassingly parallel ray-casting algorithm to get faster with the increasing number of ranks; it is also likely for a compositing algorithm to slow down due to communication costs. We observe little to no increase in timings with our deep compositing, where it nearly behaves like a constant. \section{Conclusions and Future Work} \label{sec:conclusion} We introduce a GPU-based direct volume visualization framework that allows correct and interactive rendering even for non-convexly partitioned data. Our framework presents a mixed element ray-marching algorithm to integrate ray segments along the viewing direction. We achieved memory savings by exploiting XOR-based compaction schemes on our finite element data structures. Furthermore, we illustrate a deep compositing algorithm that allows proper order compositing of the RGBA-Z values obtained across multiple compute nodes. Our framework scales well for increasing GPU counts while using native partitioning of non-convex data sets. We consider our framework suitable for both in~situ and post~hoc applications. Possible areas for further research are as follows. While we allow visualizations of multiple scalar fields and timesteps, we do not use double/triple buffering techniques that can hide the buffer loading times. Our implementation naively takes one scalar set on request (i.e., does not pre-fetch anything). In order to improve the time steps loading performance, buffering and pre-fetching the time steps in GPU and main memory can be employed \cite{shih2014out}. Moreover, currently, we assume that the topology of the volumetric data does not change through time, yet this may not be the case. Furthermore, our sampling method does not support bilinear elements since determining vertex index order after element construction is difficult for them using our XOR-compaction. Also, we would utilize another compaction scheme over connectivity information as other works do \cite{muigg2011interactivevol, aman2021compact}. Image-based partitioning may further increase our method's efficiency; however, it can get challenging with the in~situ emphasis. Finally, integrating our approach into existing frameworks is another future work, field testing our claims. \acknowledgments{ Intentionally omitted for review.} \balance \bibliographystyle{abbrv-doi}

aaaf40128cd43f3edc08e41f55706704f58b86d8

\section{Introduction} We assume that $x = (x_1,x_2,...,x_d)$ and $y = (y_1,y_2,...,y_d)$ are distributed according to $$p(x\mid\lambda)=\prod\limits_{i=1}^d p(x_i \mid \lambda_i) =\exp\{-r(\lambda_1+\lambda_2+\cdots+\lambda_d)\}\ffrac{(r\lambda_1)^{x_1}}{x_1!}\cdots\ffrac{(r\lambda_d)^{x_d}}{x_d!}$$ and $$p(y\mid\lambda)=\prod\limits_{i=1}^d p(y_i \mid \lambda_i) =\exp\{-s(\lambda_1+\lambda_2+\cdots+\lambda_d)\}\ffrac{(s\lambda_1)^{y_1}}{y_1!}\cdots\ffrac{(s\lambda_d)^{y_d}}{y_d!},$$ respectively, where $r$ and $s$ are known positive real numbers. Here, $\lambda = (\lambda_1, \ldots, \lambda_d)$ is an unknown parameter. We consider the problem of predicting independent Poisson random variables $y = (y_1,y_2,...,y_d)$ using independent observations $x = (x_1,x_2,...,x_d)$. We adopt the Kullback--Leibler loss of predictive distribution $\hat{p}(y\mid x)$, which is $$D(p(y \mid \lambda),\hat{p}(y\mid x)) =\sum\limits_yp(y\mid\lambda) \log \ffrac{p(y\mid\lambda)}{\hat{p}(y\mid x)}.$$ \par There are numerous studies on the simultaneous estimation of Poisson parameters. \cite{clevenson1975simultaneous} proposed generalized Bayes estimators dominating the maximum likelihood estimator when $d\ge2$ under standardized squared error loss $\sum\lambda_i^{-1}(\hat\lambda_i-\lambda_i)^2.$ \cite{tsui1982simultaneous} studied the estimation under the generalized loss function $\sum(\hat\lambda_i-\lambda_i)^2/\lambda_i^{k},$ where $k$ is a given positive integer. \cite{ghosh1988simultaneous} characterized admissible linear estimators of multiple Poisson parameters under Kullback--Leibler loss. Estimation of parameters under Kullback--Leibler loss can be generalized to a predictive distribution problem, which is important for several statistical scenarios. Predictive method has been shown to be preferable in \cite{aitchison1975goodness}. Noninformative prior distributions or vague distributions are often used for constructing Bayesian predictive distributions. The Jeffreys prior is widely used in various problems under Kullback--Leibler loss, see \cite{akaike1978new} and \cite{clarke1994jeffreys}. In contrast to the large number of estimation studies, decision theory regarding predictive distributions on the Poisson model has been developed relatively recently. A class of prior distributions, $$\pi_{\alpha,\beta}(\lambda)\ensuremath{\mathrm{d}}\lambda_1\ensuremath{\mathrm{d}}\lambda_2\cdots \ensuremath{\mathrm{d}}\lambda_d\propto \ffrac{\lambda_1^{\beta_1-1}\lambda_2^{\beta_2-1}\cdots\lambda_d^{\beta_d-1}} {(\lambda_1+\lambda_2+\cdots+\lambda_d)^\alpha}\ensuremath{\mathrm{d}}\lambda_1\ensuremath{\mathrm{d}}\lambda_2\cdots \ensuremath{\mathrm{d}}\lambda_d$$ was proposed by \cite{komaki2004simultaneous}. The Bayesian predictive distribution based on $\pi_{\alpha=d/2-1,\beta=(1/2,\dots,1/2)}(\lambda)$ is shown to dominate that based on the Jeffreys prior when $d\ge3$. \cite{komaki2015simultaneous} considered the problem of independent Poisson processes with different durations and introduced a class of improper prior densities, which is a generalization of $\pi_{\alpha,\beta}(\lambda)$. The corresponding Bayesian predictive distribution is shown to dominate that based on the Jeffreys prior. A class of proper priors was proposed and Bayesian predictive densities and estimators based on the priors were shown to dominate the Bayesian predictive density and estimator based on the Jeffreys prior under Kullback--Leibler loss \citep{komaki2006class}. Recently, \cite{hamura2020bayesian} studied the Bayesian predictive distribution for a Poisson model with parametric restriction under Kullback--Leibler loss. \cite{yano2021minimax} presented a class of Bayes predictive densities that attain asymptotic minimaxity in sparse Poisson sequence models. From the viewpoint of a model manifold with Fisher metric, it is natural that similar results hold simultaneously for the multivariate normal model and the independent Poisson observables model \citep{komaki2006class}. There are several counterparts for these two models. For a multivariate normal model $N(\mu,\sigma^2I)$, the Bayesian predictive distribution based on Stein's harmonic prior \citep{stein1974estimation}, $$\pi(\mu)=\Vert\mu\Vert^{2-d},$$ dominates that based on the Jeffreys prior \citep{komaki2001shrinkage}, which is similar to the result reported in \cite{komaki2004simultaneous}. \cite{johnstone1984admissibility} studied the admissibility and recurrence in estimating a Poisson mean under standardized squared error loss, which is a counterpart to the diffusion characterization of admissibility in the estimation of a multivariate normal mean, introduced by \cite{brown1971admissible}. \par For the multivariate normal model, \cite{george2006improved} generalized the result in \cite{komaki2001shrinkage} and proved that Bayesian predictive densities based on superharmonic priors dominate those based on the Jeffreys prior. It is natural to speculate that there exists a similar result for independent Poisson observables model; this speculation is confirmed in this study. In Section 2, we show that Bayesian predictive densities based on priors constructed by superharmonic functions dominate those based on the Jeffreys prior. In Section 3, we show a similar result for independent Poisson processes with different durations. In Section 4 some examples are given, including point shrinkage priors and subspace shrinkage priors. We show that Bayesian predictive densities based on the example priors dominate those based on the Jeffreys prior. In Section 5, we conduct some numerical experiments. The lemmas used in Sections 2 and 3 are given in the appendix. \section{Improved Bayesian prediction using superharmonic function} We consider a class of improper prior densities, $$\pi_{\beta}(\lambda)\ensuremath{\mathrm{d}}\lambda_1\ensuremath{\mathrm{d}}\lambda_2\cdots \ensuremath{\mathrm{d}}\lambda_d\propto \lambda_1^{\beta_1-1}\lambda_2^{\beta_2-1}\cdots\lambda_d^{\beta_d-1}\ensuremath{\mathrm{d}}\lambda_1\ensuremath{\mathrm{d}}\lambda_2\cdots \ensuremath{\mathrm{d}}\lambda_d$$ with $\beta_i>0, i=1,2,\dots,d.$ Our goal is to find function $f(\lambda_1,\dots,\lambda_d)$ such that for prior $$\pi_{f,\beta}(\lambda)\ensuremath{\mathrm{d}}\lambda_1\ensuremath{\mathrm{d}}\lambda_2\cdots \ensuremath{\mathrm{d}}\lambda_d\propto f(\lambda_1,\dots,\lambda_d)\lambda_1^{\beta_1-1}\lambda_2^{\beta_2-1}\cdots\lambda_d^{\beta_d-1}\ensuremath{\mathrm{d}}\lambda_1\ensuremath{\mathrm{d}}\lambda_2\cdots \ensuremath{\mathrm{d}}\lambda_d,$$ the Bayesian predictive distribution $p_{f,\beta}(y\mid x)$ based on $\pi_{f,\beta}(\lambda)$ dominates the Bayesian predictive distribution $p_{\beta}(y\mid x)$ based on $\pi_{\beta}(\lambda)$ under Kullback--Leibler loss. In this section, we denote $\theta_i:=\sqrt{\lambda_i},\ i=1,\dots,d$. We also denote $g(\theta_1,\dots,\theta_d):=f(\lambda_1,\dots,\lambda_d)$ and $p(x\mid\theta):=p(x\mid\lambda)$. We show that if $g$ satisfies some conditions, then the Bayesian predictive distribution $p_{f,\beta}(y\mid x)$ dominates the Bayesian predictive distribution $p_{\beta}(y\mid x)$. As a corollary, if $g$ is a superharmonic function satisfying regularity conditions and the derivative of $g$ on the boundary is nonpositive, then $p_{f,\beta=(1/2,\dots,1/2)}(y\mid x)$ dominates $p_{J}(y\mid x)$ based on the Jeffreys prior.\\ \noindent \textbf{Theorem 1}. \begin{description} \item{~1)~\;} The Bayesian predictive distribution $p_{f,\beta}(y\mid x)$ dominates the Bayesian predictive distribution $p_{\beta}(y\mid x)$ if, for every $r>0$, the function $$F(x,r)=F(x_1,x_2,\dots,x_d,r):=\int f(\bar\lambda_1,\dots,\bar\lambda_d)\prod\limits_{i=1}^d\ffrac{r^{x_i+\beta_i}\bar\lambda_i^{x_i+\beta_i-1}\exp(-r\bar\lambda_i)}{\Gamma(x_i+\beta_i)}\ensuremath{\mathrm{d}}\bar\lambda$$ is not a constant function of $x$, there exists $\epsilon>0$ such that \begin{align} &\sum_x\Biggl[ \max_{t\in[r-\epsilon,r+\epsilon]}\biggl\{ \bigl| \log F(x,t) \bigr| (\sum_{i=1}^dx_i+1) \prod_{i=1}^d \frac{(t\lambda_i)^{x_i}\exp(-t\lambda_i)}{x_i!} \biggr\} \Biggr] <\infty, \label{F-differentiable} \end{align} and for every $x$, $r>0$, \begin{align} &\sum\limits_{i=1}^dx_i(F(x,r)-F(x-\delta_i,r))+\sum\limits_{i=1}^d(x_i+\beta_i)(F(x,r)-F(x+\delta_i,r))\ge 0, \label{Fineq} \end{align} where $F(x-\delta_i,r)$ is defined to be $1$ if $x_i=0$. \item{~2)~\;} Condition \eqref{Fineq} is satisfied if $g\in\mathbf{C}^2([0,\infty)^d)$, \begin{subequations} \begin{align} &\int g(\theta)\prod\limits_{j=1}^{d}\theta_j^{2\beta_j-1}p(x\mid\theta)\ensuremath{\mathrm{d}}\theta < \infty, \label{t1-condition-a}\\ &\int \Bigl\lvert\frac{\partial g}{\partial\theta_i}(\theta) \Bigr\rvert\prod\limits_{j=1}^{d} \theta_j^{2\beta_j-1}p(x\mid\theta)\ensuremath{\mathrm{d}}\theta < \infty, \label{t1-condition-b}\\ &\int \Bigl\lvert\ffrac{\partial}{\partial\theta_i} \Bigl\{ \frac{\partial g}{\partial\theta_i}(\theta)\prod\limits_{j=1}^d\theta_j^{2\beta_j-1} \Bigr\} \Bigl\rvert p(x\mid\theta)\ensuremath{\mathrm{d}}\theta < \infty, \label{t1-condition-c} \end{align} for every $x$, $i$, $r>0$, \begin{align} &\sum\limits_{i=1}^{d}\ffrac{\partial}{\partial \theta_i} \biggl( \ffrac{\partial g}{\partial \theta_i} \prod\limits_{j=1}^{d}\theta_j^{2\beta_j-1} \biggr) \le0, \label{t1-condition-d} \end{align} and \begin{align} &\lim_{\theta_i\to0} \frac{\partial g}{\partial\theta_i}(\theta) \, \theta_i^{2\beta_i-1}\le0, \label{t1-condition-e} \end{align} for every $i$. \end{subequations} \end{description} \vspace{0.5cm} \noindent \textbf{Proof}. 1) The Bayesian predictive distribution based on prior $\pi$ is $$p_{\pi}(y\mid x)=\ffrac{\int p(x,y\mid\lambda)\pi(\lambda) \ensuremath{\mathrm{d}}\lambda}{\int p(x\mid\lambda)\pi(\lambda) \ensuremath{\mathrm{d}}\lambda}.$$ The difference between the K-L risks of Bayesian predictive distributions based on $\pi_{f,\beta}$ and $\pi_{\beta}$ is \begin{align} &\text{E} \biggl( \log\ffrac{p_{\beta}(y\mid x)}{p_{f,\beta}(y\mid x)} \, \bigg| \, \lambda \biggr) \notag\\ &= \text{E} \Bigl(\log\int p(x,y\mid\bar\lambda)\pi_{\beta}(\bar\lambda) \ensuremath{\mathrm{d}}\bar\lambda \, \Big| \, \lambda \Bigr) -\text{E} \Bigl( \log\int p(x\mid\bar\lambda)\pi_{\beta}(\bar\lambda) \ensuremath{\mathrm{d}}\bar\lambda \, \Big| \, \lambda \Bigr) \notag\\ &\quad -\text{E} \Bigl( \log\int p(x,y\mid\bar\lambda)\pi_{f,\beta}(\bar\lambda) \ensuremath{\mathrm{d}}\bar\lambda \, \Big| \, \lambda \Bigr) + \text{E} \Bigl( \log\int p(x\mid\bar\lambda)\pi_{f,\beta}(\bar\lambda) \ensuremath{\mathrm{d}}\bar\lambda \,\Big| \, \lambda \Bigr) \notag\\ &=\text{E} \Biggl( \log\ffrac{\int p(x\mid\bar\lambda)\pi_{f,\beta}(\bar\lambda) \ensuremath{\mathrm{d}}\bar\lambda}{\int p(x\mid\bar\lambda)\pi_{\beta}(\bar\lambda) \ensuremath{\mathrm{d}}\bar\lambda} \, \Bigg| \, \lambda \Biggr) -\text{E} \Biggl( \log\ffrac{\int p(x,y\mid\bar\lambda)\pi_{f,\beta}(\bar\lambda) \ensuremath{\mathrm{d}}\bar\lambda}{\int p(x,y\mid\bar\lambda)\pi_{\beta}(\bar\lambda) \ensuremath{\mathrm{d}}\bar\lambda} \, \Bigg| \, \lambda \Biggr) \notag\\ &=\text{E} \Biggl( \log \frac{\int f(\bar\lambda_1,\dots,\bar\lambda_d)\bar\lambda_1^{\beta_1-1}\bar\lambda_2^{\beta_2-1} \cdots\bar\lambda_d^{\beta_d-1} \exp(-r\sum_{i=1}^d\bar\lambda_i) (r\bar\lambda_1)^{x_1}\cdots(r\bar\lambda_d)^{x_d} \ensuremath{\mathrm{d}}\bar\lambda} {\int \bar\lambda_1^{\beta_1-1}\bar\lambda_2^{\beta_2-1}\cdots\bar\lambda_d^{\beta_d-1} \exp(-r\sum_{i=1}^d\bar\lambda_i)(r\bar\lambda_1)^{x_1}\cdots(r\bar\lambda_d)^{x_d} \ensuremath{\mathrm{d}}\bar\lambda} \, \Bigg| \, \lambda \Biggr) \notag\\ &\quad -\text{E} \Biggl( \log\frac{\int f(\bar\lambda_1,\dots,\bar\lambda_d)\bar\lambda_1^{\beta_1-1}\bar\lambda_2^{\beta_2-1}\cdots \bar\lambda_d^{\beta_d-1}\exp\{-(r+s) \sum_{i=1}^d \bar\lambda_i\} \{(r+s)\bar\lambda_1\}^{x_1+y_1} \cdots\{(r+s)\bar \lambda_d \}^{x_d+y_d} \ensuremath{\mathrm{d}}\bar\lambda} {\int \bar\lambda_1^{\beta_1-1}\bar\lambda_2^{\beta_2-1}\cdots\bar\lambda_d^{\beta_d-1} \exp\{-(r+s)\sum_{i=1}^d\bar\lambda_i\}\{(r+s)\bar\lambda_1\}^{x_1+y_1} \cdots\{(r+s)\bar\lambda_d\}^{x_d+y_d} \ensuremath{\mathrm{d}}\bar\lambda} \, \Bigg| \, \lambda \Biggr) \notag\\ &=\text{E} \biggl( \log\int f(\bar\lambda_1,\dots,\bar\lambda_d)\prod_{i=1}^d \frac{r^{x_i+\beta_i}\bar\lambda_i^{x_i+\beta_i-1} \exp(-r\bar\lambda_i)}{\Gamma(x_i+\beta_i)}\ensuremath{\mathrm{d}}\bar\lambda \, \bigg| \, \lambda \biggr) \notag\\ & \quad -\text{E} \Biggl(\log\int f(\bar\lambda_1,\dots,\bar\lambda_d)\prod\limits_{i=1}^d\ffrac{(r+s)^{x_i+y_i+\beta_i}\bar\lambda_i^{x_i+y_i+\beta_i-1}\exp\{-(r+s)\bar\lambda_i\}}{\Gamma(x_i+y_i+\beta_i)}\ensuremath{\mathrm{d}}\bar\lambda \, \Bigg| \, \lambda \Biggr) \notag\\ & = \text{E} \biggl( \log(F(x,r)) \, \bigg| \, x_i\sim\text{Po}(r\lambda_i),i=1,\dots,d \biggr) -\text{E} \biggl( \log(F(x,r+s)) \, \bigg| \, x_i\sim\text{Po}((r+s)\lambda_i),i=1,\dots,d \biggr) \label{negativediff} \end{align} Using \eqref{F-differentiable}, we know that $$ \text{E} \biggl( \bigl| \log F(x,r) \bigr| \, \bigg| \, x_i\sim\text{Po}(r\lambda_i),i=1,\dots,d \biggr)<\infty $$ for every $r$. Thus, the risk difference \eqref{negativediff} is finite.\\ The risk difference \eqref{negativediff} is negative if \begin{align} \text{E} \biggl( \log F(x,r) \, \bigg| \, x_i\sim\text{Po}(r\lambda_i),i=1,\dots,d \biggr) =\sum_x\log(F(x,r))\Biggl\{ \prod_{i=1}^d \frac{(r\lambda_i)^{x_i}\exp(-r\lambda_i)}{x_i!} \Biggr\} \label{increasing} \end{align} is an increasing function of $r$. \vspace{0.5cm} If we exchange the integration and differentiation in $\ffrac{\partial F}{\partial r}(x,r)$, we have $$\ffrac{\partial F}{\partial r}(x,r)=\int f(\bar\lambda_1,\dots,\bar\lambda_d) \prod_{i=1}^d \frac{r^{x_i+\beta_i}\bar\lambda_i^{x_i+\beta_i-1}\exp(-r\bar\lambda_i)}{\Gamma(x_i+\beta_i)} \Bigl( \sum_{i=1}^d \frac{x_i+\beta_i}{r} - \sum_{i=1}^d \bar\lambda_i \Bigr) \ensuremath{\mathrm{d}}\bar\lambda.$$ From Lemma 1.1, we can exchange the integration and differentiation in $\ffrac{\partial F}{\partial r}(x,r)$. If we differentiate \eqref{increasing} item-by-item, the partial differential function of \eqref{increasing} wrt $r$ is \begin{align} &\sum_x \log(F(x,r)) \Biggl\{ \prod_{i=1}^d \frac{(r\lambda_i)^{x_i}\exp(-r\lambda_i)}{x_i!} \Biggr\} \Bigl(\sum_{i=1}^d \frac{x_i}{r} - \sum_{i=1}^d \lambda_i \Bigr) \notag \\ &\quad + \sum_x \Biggl\{ \frac{ \displaystyle \int f(\bar\lambda_1,\dots,\bar\lambda_d) \prod_{i=1}^d \frac{r^{x_i+\beta_i}\bar\lambda_i^{x_i+\beta_i-1}\exp(-r\bar\lambda_i)}{\Gamma(x_i+\beta_i)} \Bigl( \sum_{i=1}^d \frac{x_i+\beta_i}{r} - \sum_{i=1}^d \bar\lambda_i \Bigr) \ensuremath{\mathrm{d}}\bar\lambda} {F(x,r)} \Biggr\} \prod_{i=1}^d \ffrac{(r\lambda_i)^{x_i} \exp(-r\lambda_i)}{x_i!}. \label{differential} \end{align} From Lemma 1.2, we can differentiate \eqref{increasing} by terms under conditions \eqref{F-differentiable} and \eqref{Fineq}. \vspace{0.5cm} We notice that \begin{align} &\sum_x \Bigl\{ \log F(x_1,x_2,\dots,x_d,r) \Bigr\} \Biggl\{ \prod_{i=1}^d\ffrac{(r\lambda_i)^{x_i} \exp(-r\lambda_i)}{x_i!} \Biggr\} \sum_{i=1}^d \lambda_i \notag \\ &=\sum_{i=1}^d \sum_x \Bigl\{ \log F(x_1,x_2,\dots,x_d,r) \Bigr\} \Biggl\{ \prod_{j=1}^d\ffrac{(r\lambda_j)^{x_j} \exp(-r\lambda_j)}{x_j!} \Biggr\} \lambda_i \notag \\ &=\sum_{i=1}^d \sum_x \biggl[ \ffrac{x_i+1}{r} \Bigl\{ \log F(x_1,x_2,\dots,x_d,r) \Bigr\} \prod\limits_{j=1}^d\ffrac{(r\lambda_j)^{x_j+\delta_{ij}} \exp(-r\lambda_j)}{(x_j+\delta_{ij})!} \biggr] \notag\\ &= \sum_{i=1}^d \sum_x \biggl[ \ffrac{x_i}{r} \Bigl\{ \log F(x-\delta_i,r) \Bigr\} \prod\limits_{j=1}^d\ffrac{(r\lambda_j)^{x_j}\exp(-r\lambda_j)}{x_j!} \biggr], \label{transform1} \end{align} where $\delta_{ij}$ is defined to be 1 if $i = j$ and 0 if $i\neq j$, $\delta_i$ is defined to be the $d$-dimensional vector whose $i$-th element is 1 and all other elements are 0, and $F(x-\delta_i,r)$ is defined to be $1$ if $x_i=0$. Furthermore, we notice that \begin{align} \int &f(\bar\lambda_1,\dots,\bar\lambda_d) \prod_{j=1}^d\ffrac{r^{x_j+\beta_j}\bar\lambda_j^{x_j+\beta_j-1} \exp(-r\bar\lambda_j)}{\Gamma(x_j+\beta_j)} \bar\lambda_i \ensuremath{\mathrm{d}}\bar\lambda \notag\\ &= \int f(\bar\lambda_1,\dots,\bar\lambda_d) \prod_{j=1}^d \ffrac{r^{x_j+\delta_{ij}+\beta_j}\bar\lambda_j^{x_j+\delta_{ij}+\beta_j-1}\exp(-r\bar\lambda_j)} {\Gamma(x_j+\delta_{ij}+\beta_j)}\ffrac{x_i+\beta_i}{r} \ensuremath{\mathrm{d}}\bar\lambda \notag \\ &= F(x+\delta_i,r)\ffrac{x_i+\beta_i}{r} \label{transform2}. \end{align} \par Thus, from \eqref{transform1} and \eqref{transform2}, the partial differential function \eqref{differential} of \eqref{increasing} wrt $r$ is \begin{align} &\sum_x\sum_{i=1}^d \biggl[ \ffrac{x_i}{r} \biggl\{ \log F(x,r) \biggr\} \prod\limits_{j=1}^d\ffrac{(r\lambda_j)^{x_j}\exp(-r\lambda_j)}{x_j!} \biggr] -\sum_x \sum_{i=1}^d \biggl[ \ffrac{x_i}{r} \biggl\{ \log F(x-\delta_i,r) \biggr\} \prod_{j=1}^d \ffrac{(r\lambda_j)^{x_j} \exp(-r\lambda_j)}{x_j!} \biggr] \notag\\ &~~~+ \sum_x \Biggl\{ \frac{F(x,r) \sum_{i=1}^d (x_i+\beta_i)/r -\sum_{i=1}^d F(x+\delta_i,r) (x_i+\beta_i)/r}{F(x,r)} \Biggr\} \prod_{j=1}^d\ffrac{(r\lambda_j)^{x_j}\exp(-r\lambda_j)}{x_j!} \notag\\ &= \sum_x \Biggl[ \sum_{i=1}^d \ffrac{x_i}{r} \biggl\{ \log \ffrac{F(x,r)}{F(x-\delta_i,r)}\biggr\} \prod\limits_{j=1}^d\ffrac{(r\lambda_j)^{x_j}\exp(-r\lambda_j)}{x_j!} \Biggr] \notag \\ &~~~+\sum_x \Biggl[ \sum_{i=1}^d \ffrac{x_i+\beta_i}{r} \biggl\{ 1-\ffrac{F(x+\delta_i,r)}{F(x,r)} \biggr\} \prod_{j=1}^d \ffrac{(r\lambda_j)^{x_j}\exp(-r\lambda_j)}{x_j!} \Biggr]. \label{differential2} \end{align} By assumption, $F(x,r)$ is not a constant function of $x$, hence $F(x-\delta_i,r)\equiv F(x,r)$ does not hold. Therefore, from the inequality $\log \epsilon >1-\ffrac{1}{\epsilon}$, $\epsilon\neq1$, \eqref{differential2} is strictly greater than \begin{align} &\sum_x \Biggl[ \sum_{i=1}^d \ffrac{x_i}{r} \biggl\{ 1-\ffrac{F(x-\delta_i,r)}{F(x,r)} \biggr\} \prod_{j=1}^d\ffrac{(r\lambda_j)^{x_j}\exp(-r\lambda_j)}{x_j!} \Biggr] +\sum_x \Biggl[ \sum_{i=1}^d \ffrac{x_i+\beta_i}{r} \biggl\{ 1-\ffrac{F(x+\delta_i,r)}{F(x,r)} \biggr\} \prod\limits_{j=1}^d\ffrac{(r\lambda_j)^{x_j}\exp(-r\lambda_j)}{x_j!} \Biggr] \notag\\ &=\sum_x \Biggl[ \sum_{i=1}^d \ffrac{x_i}{r} \biggl\{ 1-\ffrac{F(x-\delta_i,r)}{F(x,r)} \biggr\} +\sum_{i=1}^d\ffrac{x_i+\beta_i}{r} \biggl\{ 1-\ffrac{F(x+\delta_i,r)}{F(x,r)} \biggr\}\Biggr] \prod\limits_{j=1}^d \ffrac{(r\lambda_j)^{x_j}\exp(-r\lambda_j)}{x_j!}. \label{Fpositive} \end{align} Using \eqref{Fineq}, we know that \eqref{Fpositive} is not negative. Thus, \eqref{differential2} is positive. Therefore, \eqref{increasing} is an increasing function and the first half of Theorem 1 is proved. \vspace{0.5cm} \noindent 2) We prove that \eqref{Fineq} is satisfied if $f$ satisfies the conditions of the second half of Theorem 1. Let $\theta_j := \sqrt{\lambda_j}$ $(j=1,\dots,d)$. Then, we have \begin{align*} &F(x+\delta_i,r)-F(x,r) \notag\\ &= \int f(\lambda) \prod_{j=1}^d \ffrac{r^{x_j+\delta_{ji}+\beta_j} \lambda_j^{x_j+\delta_{ji}+\beta_j-1} \exp (-r \lambda_j)}{\Gamma(x_j+\delta_{ji}+\beta_j)} \ensuremath{\mathrm{d}}\lambda -\int f(\lambda) \prod_{j=1}^d\ffrac{r^{x_j+\beta_j} \lambda_j^{x_j+\beta_j-1} \exp (-r \lambda_j)}{\Gamma(x_j+\beta_j)} \ensuremath{\mathrm{d}}\lambda \\ &=2^d\int g(\theta)\prod\limits_{j=1}^d\ffrac{r^{x_j+\delta_{ji}+\beta_j}\theta_j^{2x_j+2\delta_{ji}+2\beta_j-1} \exp(-r\theta_j^2)}{\Gamma(x_j+\delta_{ji}+\beta_j)}\ensuremath{\mathrm{d}}\theta-2^d \int g(\theta) \prod\limits_{j=1}^d \ffrac{r^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-r\theta_j^2)} {\Gamma(x_j+\beta_j)} \ensuremath{\mathrm{d}}\theta \\ &=-2^{d-1}\int\biggl\{ \int g(\theta)\prod\limits_{j\neq i}\ffrac{r^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-r\theta_j^2)} {\Gamma(x_j+\beta_j)} \ensuremath{\mathrm{d}} \theta_j \biggr\} \ffrac{r^{x_i+\beta_i}\theta_i^{2x_i+2\beta_i}\ffrac{\partial}{\partial\theta_i}\exp(-r\theta_i^2)} {\Gamma(x_i+1+\beta_i)}\ensuremath{\mathrm{d}}\theta_i \notag\\ & ~~ -2^{d-1}\int\biggl\{ \int g(\theta)\prod\limits_{j\neq i}\ffrac{r^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-r\theta_j^2)} {\Gamma(x_j+\beta_j)} \ensuremath{\mathrm{d}} \theta_j \biggr\} \ffrac{r^{x_i+\beta_i}(\frac{\partial}{\partial\theta_i}\theta_i^{2x_i+2\beta_i})\exp(-r\theta_i^2)}{\Gamma(x_i+1+\beta_i)}\ensuremath{\mathrm{d}}\theta_i. \end{align*} By integration by parts on $\theta_i$, we have \begin{align} &F(x+\delta_i,r)-F(x,r) \notag\\ &=\mathop{\lim}_{a\to0\atop b\to\infty}\mathop{\lim}_{u\to0\atop v\to\infty} \biggl[-2^{d-1}\int_u^v \biggl\{ \int_{[a,b]^{d-1}} g(\theta)\prod\limits_{j\neq i}\ffrac{r^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-r\theta_j^2)} {\Gamma(x_j+\beta_j)} \ensuremath{\mathrm{d}} \theta_j \biggr\} \ffrac{r^{x_i+\beta_i}\theta_i^{2x_i+2\beta_i}\ffrac{\partial}{\partial\theta_i}\exp(-r\theta_i^2)} {\Gamma(x_i+1+\beta_i)}\ensuremath{\mathrm{d}}\theta_i \notag\\ & ~~ -2^{d-1}\int_u^v \biggl\{ \int_{[a,b]^{d-1}} g(\theta)\prod\limits_{j\neq i}\ffrac{r^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-r\theta_j^2)} {\Gamma(x_j+\beta_j)} \ensuremath{\mathrm{d}} \theta_j \biggr\} \ffrac{r^{x_i+\beta_i}(\ffrac{\partial}{\partial\theta_i}\theta_i^{2x_i+2\beta_i})\exp(-r\theta_i^2)} {\Gamma(x_i+1+\beta_i)}\ensuremath{\mathrm{d}}\theta_i\biggl] \notag\\ &=\mathop{\lim}_{a\to0\atop b\to\infty}\mathop{\lim}_{u\to0\atop v\to\infty} \biggl[2^{d-1}\int_u^v \biggl\{\ffrac{\partial}{\partial\theta_i}\biggl\{ \int_{[a,b]^{d-1}} g(\theta)\prod\limits_{j\neq i}\ffrac{r^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-r\theta_j^2)} {\Gamma(x_j+\beta_j)} \ensuremath{\mathrm{d}} \theta_j \biggr\}\biggr\} \ffrac{r^{x_i+\beta_i}\theta_i^{2x_i+2\beta_i}\exp(-r\theta_i^2)}{\Gamma(x_i+1+\beta_i)}\ensuremath{\mathrm{d}}\theta_i \notag\\ & ~~~~ -2^{d-1} \Biggl[ \biggl\{ \int_{[a,b]^{d-1}} g(\theta)\prod\limits_{j\neq i}\ffrac{r^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-r\theta_j^2)} {\Gamma(x_j+\beta_j)} \ensuremath{\mathrm{d}} \theta_j \biggr\} \ffrac{r^{x_i+\beta_i}\theta_i^{2x_i+2\beta_i}\exp(-r\theta_i^2)} {\Gamma(x_i+1+\beta_i)} \Biggr]^{v}_{\theta_i = u}\biggl] \notag\\ &=\mathop{\lim}_{a\to0\atop b\to\infty}\mathop{\lim}_{u\to0\atop v\to\infty} \biggl[2^{d-1} \int_{[u,v]\times[a,b]^{d-1}} \biggl\{ \ffrac{\partial}{\partial\theta_i}g(\theta) \biggr\} \prod\limits_{j=1}^d\ffrac{r^{x_j+\beta_j} \theta_j^{2x_j+\delta_{ji}+2\beta_j-1}\exp(-r\theta_j^2)}{\Gamma(x_j+\delta_{ji}+\beta_j)} \ensuremath{\mathrm{d}}\theta \notag\\ & ~~~~ -2^{d-1} \Biggl[ \biggl\{ \int_{[a,b]^{d-1}} g(\theta)\prod\limits_{j\neq i}\ffrac{r^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-r\theta_j^2)} {\Gamma(x_j+\beta_j)} \ensuremath{\mathrm{d}} \theta_j \biggr\} \ffrac{r^{x_i+\beta_i}\theta_i^{2x_i+2\beta_i}\exp(-r\theta_i^2)} {\Gamma(x_i+1+\beta_i)} \Biggr]^{v}_{\theta_i = u}\biggl]. \label{F-diff-1} \end{align} Here, we use auxiliary variables $a,b,u,v$ and Lemma 1.3 to ensure the above equations hold. \vspace{0.2cm} \par Using Lemma 1.3, we have $$ \int \biggl\{ \int_{[a,b]^{d-1}} g(\theta)\prod\limits_{j\neq i} \ffrac{r^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-r\theta_j^2)} {\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\theta_j \biggr\} \ffrac{r^{x_i+\beta_i}\theta_i^{2x_i+2\beta_i} \exp(-r\theta_i^2)}{\Gamma(x_i+1+\beta_i)}\ensuremath{\mathrm{d}}\theta_i <\infty.$$ Because of Lemma 1.4, $\displaystyle \biggl\{ \int_{[a,b]^{d-1}} g(\theta)\prod\limits_{j\neq i} \ffrac{r^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-r\theta_j^2)} {\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\theta_j \biggr\} \ffrac{r^{x_i+\beta_i}\theta_i^{2x_i+2\beta_i} \exp(-r\theta_i^2)}{\Gamma(x_i+1+\beta_i)}$ converges as $\theta_i\to\infty$. Thus, \begin{equation} \lim_{\theta_i\to\infty} \biggl\{ \int_{[a,b]^{d-1}} g(\theta)\prod\limits_{j\neq i} \ffrac{r^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-r\theta_j^2)} {\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\theta_j \biggr\} \ffrac{r^{x_i+\beta_i}\theta_i^{2x_i+2\beta_i} \exp(-r\theta_i^2)}{\Gamma(x_i+1+\beta_i)} = 0. \label{1st-partial-1} \end{equation} Because $g\in \mathbf{C}^2([0,\infty)^d)$, $g$ is bounded on $[0,1]\times[a,b]^{d-1}$, then $$\biggl\{ \int_{[a,b]^{d-1}} g(\theta)\prod\limits_{j\neq i} \ffrac{r^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-r\theta_j^2)} {\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\theta_j \biggr\}$$ is bounded for $\theta_i\le1$. Thus, \begin{equation} \lim_{\theta_i\to0} \biggl\{ \int_{[a,b]^{d-1}} g(\theta)\prod\limits_{j\neq i} \ffrac{r^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-r\theta_j^2)} {\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\theta_j \biggr\} \ffrac{r^{x_i+\beta_i}\theta_i^{2x_i+2\beta_i} \exp(-r\theta_i^2)}{\Gamma(x_i+1+\beta_i)} = 0. \label{1st-partial-2} \end{equation} \vspace{0.3cm} Using \eqref{1st-partial-1} and \eqref{1st-partial-2}, we have \begin{equation} \mathop{\lim}_{u\to0\atop v\to\infty} -2^{d-1} \Biggl[ \biggl\{ \int_{[a,b]^{d-1}} g(\theta)\prod\limits_{j\neq i}\ffrac{r^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-r\theta_j^2)} {\Gamma(x_j+\beta_j)} \ensuremath{\mathrm{d}} \theta_j \biggr\} \ffrac{r^{x_i+\beta_i}\theta_i^{2x_i+2\beta_i}\exp(-r\theta_i^2)} {\Gamma(x_i+1+\beta_i)} \Biggr]^{v}_{\theta_i = u}=0. \label{1st-partial-3} \end{equation} Thus, using \eqref{F-diff-1}, Lemma 1.3, \eqref{1st-partial-3}, we have \begin{align} &F(x+\delta_i,r)-F(x,r) \notag\\ &=\mathop{\lim}_{a\to0\atop b\to\infty}\mathop{\lim}_{u\to0\atop v\to\infty} \biggl[2^{d-1} \int_{[u,v]\times[a,b]^{d-1}} \ffrac{\partial g }{\partial\theta_i}(\theta) \prod\limits_{j=1}^d\ffrac{r^{x_j+\beta_j} \theta_j^{2x_j+\delta_{ji}+2\beta_j-1}\exp(-r\theta_j^2)}{\Gamma(x_j+\delta_{ji}+\beta_j)} \ensuremath{\mathrm{d}}\theta\biggl] \notag\\ &= 2^{d-1}\int \ffrac{\partial g}{\partial\theta_i}(\theta) \prod\limits_{j=1}^d \ffrac{r^{x_j+\beta_j}\theta_j^{2x_j+\delta_{ji}+2\beta_j-1}\exp(-r\theta_j^2)} {\Gamma(x_j+\delta_{ji}+\beta_j)}\ensuremath{\mathrm{d}}\theta. \label{F-1st-diff} \end{align} \vspace{0.5cm} Using \eqref{F-1st-diff} and integration by parts on each parameter again, we have \begin{align} &\sum_{i=1}^d x_i \{ F(x,r)-F(x-\delta_i,r) \} +\sum_{i=1}^d (x_i+\beta_i) \{F(x,r)-F(x+\delta_i,r)\} \notag\\ &=\sum_{i=1}^d 2^{d-1} \int\biggl[ \ffrac{\partial g }{\partial\theta_i}(\theta) x_i \prod_{j=1}^d\ffrac{r^{x_j+\beta_j-\delta_{ji}}\theta_j^{2x_j-\delta_{ji}+2\beta_j-1} \exp(-r\theta_j^2)} {\Gamma(x_j+\beta_j)} \notag\\ & \qquad- \ffrac{\partial g }{\partial\theta_i} (\theta) (x_i+\beta_i) \prod_{j=1}^d \ffrac{r^{x_j+\beta_j}\theta_j^{2x_j+\delta_{ji}+2\beta_j-1}\exp(-r\theta_j^2)}{\Gamma(x_j+\delta_{ji}+\beta_j)} \biggr]\ensuremath{\mathrm{d}}\theta \notag\\ &=\sum_{i=1}^d \ffrac{2^{d-2}}{r} \int\biggl[ \ffrac{\partial g }{\partial\theta_i}(\theta) 2x_i \prod_{j=1}^d\ffrac{r^{x_j+\beta_j}\theta_j^{2x_j-\delta_{ji}+2\beta_j-1} \exp(-r\theta_j^2)} {\Gamma(x_j+\beta_j)} \notag\\ & \qquad- \ffrac{\partial g }{\partial\theta_i} (\theta) 2r \prod_{j=1}^d \ffrac{r^{x_j+\beta_j}\theta_j^{2x_j+\delta_{ji}+2\beta_j-1}\exp(-r\theta_j^2)}{\Gamma(x_j+\beta_j)} \biggr]\ensuremath{\mathrm{d}}\theta \notag\\ &=\sum_{i=1}^d \ffrac{2^{d-2}}{r} \int\biggl[ \ffrac{\partial g}{\partial\theta_i}(\theta)\theta_i^{2\beta_i-1} \biggl\{ \prod_{j\neq i}\ffrac{r^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-r\theta_j^2)}{\Gamma(x_j+\beta_j)}\biggr\} \ffrac{r^{x_i+\beta_i}\ffrac{\partial\theta_i^{2x_i}}{\partial\theta_i}\exp(-r\theta_i^2)}{\Gamma(x_i+\beta_i)} \notag\\ & \qquad+ \ffrac{\partial g}{\partial\theta_i}(\theta)\theta_i^{2\beta_i-1} \biggl\{\prod_{j\neq i}\ffrac{r^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-r\theta_j^2)}{\Gamma(x_j+\beta_j)}\biggr\} \ffrac{r^{x_i+\beta_i}\theta_i^{2x_i}\ffrac{\partial}{\partial\theta_i}\exp(-r\theta_i^2)} {\Gamma(x_i+\beta_i)} \biggr] \ensuremath{\mathrm{d}}\theta \notag\\ &=\sum_{i=1}^d \ffrac{2^{d-2}}{r} \mathop{\lim}_{a\to0\atop b\to\infty}\mathop{\lim}_{u\to0\atop v\to\infty} \Biggl\{ \int_u^v\int_{[a,b]^{d-1}} \ffrac{\partial g}{\partial\theta_i}(\theta)\theta_i^{2\beta_i-1} \biggl\{\prod_{j\neq i}\ffrac{r^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-r\theta_j^2)}{\Gamma(x_j+\beta_j)} \ensuremath{\mathrm{d}}\theta_j\biggr\} \ffrac{r^{x_i+\beta_i}\ffrac{\partial\theta_i^{2x_i}}{\partial\theta_i}\exp(-r\theta_i^2)}{\Gamma(x_i+\beta_i)}\ensuremath{\mathrm{d}}\theta_i \notag\\ & \qquad+ \int_u^v\int_{[a,b]^{d-1}}\ffrac{\partial g}{\partial\theta_i}(\theta)\theta_i^{2\beta_i-1} \biggl\{\prod_{j\neq i}\ffrac{r^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-r\theta_j^2)}{\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\theta_j\biggr\} \ffrac{r^{x_i+\beta_i}\theta_i^{2x_i}\ffrac{\partial}{\partial\theta_i}\exp(-r\theta_i^2)} {\Gamma(x_i+\beta_i)}\ensuremath{\mathrm{d}}\theta_i \Biggr\} \notag\\ &=\sum_{i=1}^d \ffrac{2^{d-2}}{r}\mathop{\lim}_{a\to0\atop b\to\infty}\mathop{\lim}_{u\to0\atop v\to\infty} \Biggl\{ -\int_u^v\Bigl[ \ffrac{\partial}{\partial\theta_i}\int_{[a,b]^{d-1}} \Bigl\{\ffrac{\partial g}{\partial\theta_i}(\theta)\theta_i^{2\beta_i-1}\Bigr\} \prod_{j\neq i}\ffrac{r^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-r\theta_j^2)}{\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\theta_j\Bigr] \ffrac{r^{x_i+\beta_i}\theta_i^{2x_i}\exp(-r\theta_i^2)}{\Gamma(x_i+\beta_i)}\ensuremath{\mathrm{d}}\theta_i \notag\\ &\qquad+ \biggl[\int_{[a,b]^{d-1}} \ffrac{\partial g}{\partial\theta_i}(\theta)\theta_i^{2\beta_i-1} \biggl\{\prod_{j\neq i}\ffrac{r^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-r\theta_j^2)}{\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\theta_j\biggr\} \ffrac{r^{x_i+\beta_i}\theta_i^{2x_i}\exp(-r\theta_i^2)}{\Gamma(x_i+\beta_i)} \biggr]^{v}_{\theta_i = u} \Biggr\} \notag\\ &=\sum_{i=1}^d \ffrac{2^{d-2}}{r}\mathop{\lim}_{a\to0\atop b\to\infty}\mathop{\lim}_{u\to0\atop v\to\infty} \Biggl\{ -\int_u^v\int_{[a,b]^{d-1}} \ffrac{\partial}{\partial\theta_i}\Bigl\{\ffrac{\partial g}{\partial\theta_i}(\theta)\theta_i^{2\beta_i-1}\Bigr\} \prod_{j\neq i}\ffrac{r^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-r\theta_j^2)}{\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\theta_j \ffrac{r^{x_i+\beta_i}\theta_i^{2x_i}\exp(-r\theta_i^2)}{\Gamma(x_i+\beta_i)}\ensuremath{\mathrm{d}}\theta_i \notag \\ &\qquad+ \biggl[\int_{[a,b]^{d-1}} \ffrac{\partial g}{\partial\theta_i}(\theta)\theta_i^{2\beta_i-1} \biggl\{\prod_{j\neq i}\ffrac{r^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-r\theta_j^2)}{\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\theta_j\biggr\} \ffrac{r^{x_i+\beta_i}\theta_i^{2x_i}\exp(-r\theta_i^2)}{\Gamma(x_i+\beta_i)} \biggr]^{v}_{\theta_i = u} \Biggr\}. \label{F-diff-2} \end{align} Here, we use auxiliary variables $a,b,u,v$ to ensure the above equations hold. \vspace{0.2cm} From \eqref{t1-condition-c}, we have $$\int \Biggl\lvert\ffrac{\partial}{\partial \theta_i} \Bigl\{ \ffrac{\partial g}{\partial\theta_i}(\theta) \prod_{j=1}^d\theta_j^{2\beta_j-1} \Bigr\} \Biggr\rvert \,\prod_{j=1}^d\theta_j^{2x_j}\exp(-r\theta_j^2)\ensuremath{\mathrm{d}}\theta<\infty.$$ Thus, \begin{equation} \int\biggl\lvert \ffrac{\partial}{\partial\theta_i} \Bigl\{ \ffrac{\partial g}{\partial\theta_i}(\theta) \prod\limits_{j=1}^d\theta_j^{2\beta_j-1} \Bigr\} \biggr\rvert \prod\limits_{j=1}^d\ffrac{r^{x_j+\beta_j}\theta_j^{2x_j} \exp(-r\theta_j^2)}{\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\theta <\infty. \label{2nd-diff-finite} \end{equation} \vspace{0.2cm} From \eqref{t1-condition-b}, we have $$\int \Bigl\lvert\ffrac{\partial g}{\partial\theta_i}(\theta)\Bigl\rvert\prod\limits_{j=1}^{d}\theta_j^{2\beta_j-1}\theta_j^{2x_j}\exp(-r\theta_j^2)\ensuremath{\mathrm{d}}\theta<\infty.$$ Thus, $$ \int \Bigl\lvert\ffrac{\partial g}{\partial\theta_i}(\theta)\Bigl\rvert\prod\limits_{j=1}^d \ffrac{r^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-r\theta_j^2)} {\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\theta_j<\infty, $$ and thus, \begin{equation} \int \int_{[a,b]^{d-1}} \ffrac{\partial g}{\partial\theta_i}(\theta)\theta_i^{2\beta_i-1} \biggl\{\prod_{j\neq i}\ffrac{r^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-r\theta_j^2)}{\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\theta_j\biggr\} \ffrac{r^{x_i+\beta_i}\theta_i^{2x_i}\exp(-r\theta_i^2)}{\Gamma(x_i+\beta_i)}\ensuremath{\mathrm{d}}\theta_i<\infty. \label{2st-partial-3} \end{equation} Because of Lemma 1.5, we know that $$\int_{[a,b]^{d-1}} \ffrac{\partial g}{\partial\theta_i}(\theta)\theta_i^{2\beta_i-1} \biggl\{\prod_{j\neq i}\ffrac{r^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-r\theta_j^2)}{\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\theta_j\biggr\} \ffrac{r^{x_i+\beta_i}\theta_i^{2x_i}\exp(-r\theta_i^2)}{\Gamma(x_i+\beta_i)}$$ converges as $\theta_i\to0$ or $\theta_i\to\infty$. Thus, using \eqref{2st-partial-3}, we have \begin{equation} \lim_{\theta_i\to\infty} \int_{[a,b]^{d-1}} \ffrac{\partial g}{\partial\theta_i}(\theta)\theta_i^{2\beta_i-1} \biggl\{\prod_{j\neq i}\ffrac{r^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-r\theta_j^2)}{\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\theta_j\biggr\} \ffrac{r^{x_i+\beta_i}\theta_i^{2x_i}\exp(-r\theta_i^2)}{\Gamma(x_i+\beta_i)} = 0. \label{2st-partial-1} \end{equation} \vspace{0.3cm} From \eqref{t1-condition-e}, we have \begin{equation} \lim_{\theta_i\to0} \int_{[a,b]^{d-1}} \ffrac{\partial g}{\partial\theta_i}(\theta)\theta_i^{2\beta_i-1} \biggl\{\prod_{j\neq i}\ffrac{r^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-r\theta_j^2)}{\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\theta_j\biggr\} \ffrac{r^{x_i+\beta_i}\theta_i^{2x_i}\exp(-r\theta_i^2)}{\Gamma(x_i+\beta_i)}\le0 \label{2st-partial-2} \end{equation} \vspace{0.3cm} Using \eqref{2st-partial-1} and \eqref{2st-partial-2}, we have \begin{align} &\mathop{\lim}_{u\to0\atop v\to\infty} \Biggl[\int_{[a,b]^{d-1}} \ffrac{\partial g}{\partial\theta_i}(\theta)\theta_i^{2\beta_i-1} \biggl\{\prod_{j\neq i}\ffrac{r^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-r\theta_j^2)}{\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\theta_j\biggr\} \ffrac{r^{x_i+\beta_i}\theta_i^{2x_i}\exp(-r\theta_i^2)}{\Gamma(x_i+\beta_i)} \Biggr]^{v}_{\theta_i = u} \ge0. \label{2st-partial-4} \end{align} Using \eqref{F-diff-2}, \eqref{2nd-diff-finite}, and \eqref{2st-partial-4}, we have \begin{align} &\sum_{i=1}^d x_i \{ F(x,r)-F(x-\delta_i,r) \} +\sum_{i=1}^d (x_i+\beta_i) \{F(x,r)-F(x+\delta_i,r)\} \notag\\ &\ge \sum_{i=1}^d \ffrac{2^{d-2}}{r}\mathop{\lim}_{a\to0\atop b\to\infty}\mathop{\lim}_{u\to0\atop v\to\infty} \Biggl\{ -\int_u^v\int_{[a,b]^{d-1}} \ffrac{\partial}{\partial\theta_i}\Bigl\{\ffrac{\partial g}{\partial\theta_i}(\theta)\theta_i^{2\beta_i-1}\Bigr\} \prod_{j\neq i}\ffrac{r^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-r\theta_j^2)}{\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\theta_j \ffrac{r^{x_i+\beta_i}\theta_i^{2x_i}\exp(-r\theta_i^2)}{\Gamma(x_i+\beta_i)}\ensuremath{\mathrm{d}}\theta_i \Biggl\} \notag\\ &=\sum_{i=1}^d-\ffrac{2^{d-2}}{r} \int \ffrac{\partial}{\partial\theta_i} \Bigl\{ \ffrac{\partial g}{\partial\theta_i}(\theta) \prod_{j=1}^d\theta_j^{2\beta_j-1} \Bigr\} \prod_{j=1}^d \ffrac{r^{x_j+\beta_j}\theta_j^{2x_j} \exp(-r\theta_j^2)}{\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\theta \notag\\ &=-\ffrac{2^{d-2}}{r} \int \sum_{i=1}^d \ffrac{\partial}{\partial\theta_i} \Bigl\{ \ffrac{\partial g}{\partial\theta_i}(\theta) \prod_{j=1}^d\theta_j^{2\beta_j-1} \Bigr\} \prod_{j=1}^d \ffrac{r^{x_j+\beta_j}\theta_j^{2x_j} \exp(-r\theta_j^2)}{\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\theta. \label{F-2nd-diff} \end{align} From \eqref{F-2nd-diff}, we know that \eqref{Fineq} holds if $g$ satisfies the conditions of the second half of Theorem 1. \qed \vspace{0.5cm} \par Considering the Jeffreys prior, $$\pi_J(\lambda)\propto\ffrac{1}{(\lambda_1\lambda_2\cdots\lambda_d)^{1/2}}\ensuremath{\mathrm{d}}\lambda_1\ensuremath{\mathrm{d}}\lambda_2\cdots \ensuremath{\mathrm{d}}\lambda_d,$$ we obtain the result below. \vspace{0.5cm} \noindent \textbf{Corollary 1}. \begin{description} \item{~1)~\;} The Bayesian predictive distribution $p_{f,\beta=(1/2,\dots,1/2)}(y\mid x)$ dominates the Bayesian predictive distribution $p_{J}(y\mid x)$ based on the Jeffreys prior if, for every $r>0$, the function $$F(x,r)=F(x_1,x_2,\dots,x_d,r):=\int f(\bar\lambda_1,\dots,\bar\lambda_d)\prod\limits_{i=1}^d\ffrac{r^{x_i+1/2}\bar\lambda_i^{x_i-1/2}\exp(-r\bar\lambda_i)}{\Gamma(x_i+1/2)}\ensuremath{\mathrm{d}}\bar\lambda$$ is not a constant function of $x$, there exists $\epsilon>0$ such that \begin{align} &\sum_x\Biggl[ \max_{t\in[r-\epsilon,r+\epsilon]}\biggl\{ \bigl| \log F(x,t) \bigr| (\sum_{i=1}^dx_i+1) \prod_{i=1}^d \frac{(t\lambda_i)^{x_i}\exp(-t\lambda_i)}{x_i!} \biggr\} \Biggr] <\infty, \label{F-differentiable-2} \end{align} and for every $x$, $r>0$, \begin{align} &\sum\limits_{i=1}^dx_i\bigl\{F(x,r)-F(x-\delta_i,r)\bigr\}+\sum\limits_{i=1}^d(x_i+1/2)\bigl\{F(x,r)-F(x+\delta_i,r)\bigr\}\ge 0, \label{Fineq2} \end{align} where $F(x-\delta_i,r)$ is defined to be $1$ if $x_i=0$. \item{~2)~\;} Condition \eqref{Fineq2} is satisfied if $g\in\mathbf{C}^2([0,\infty)^d)$, \begin{align*} &\int g(\theta) p(x\mid\theta)\ensuremath{\mathrm{d}}\theta<\infty,\\ &\int \Bigl\lvert \ffrac{\partial g}{\partial\theta_i}(\theta) \Bigr\rvert p(x\mid\theta)\ensuremath{\mathrm{d}}\theta<\infty,\\ &\int \Bigl\lvert \ffrac{\partial^2 g}{\partial\theta_i^2}(\theta) \Bigr\rvert p(x\mid\theta)\ensuremath{\mathrm{d}}\theta<\infty, \end{align*} for every $x$, $i$, $r>0$, and $g(\theta)$ is a superharmonic function, $\displaystyle\left.\ffrac{\partial g}{\partial\theta_i}(\theta)\right|_{\theta_i=0}\le0,\,\forall i$. \end{description} \vspace{0.5cm} \par The K-L risk of estimator $\hat\lambda$ is defined as the K-L risk of plug-in density $p(x\mid\hat\lambda)$, which is $$ \sum_x \biggl[ \sum_{i=1}^d \Bigl\{ r\lambda_i\log \Bigl(\ffrac{\lambda_i}{\hat\lambda_i} \Bigr) -r\lambda_i+r\hat\lambda_i \Bigr\} \biggr] \prod\limits_{j=1}^d\ffrac{(r\lambda_j)^{x_j}\exp(-r\lambda_j)}{x_j!}. $$ We know that the Bayesian estimator based on $\pi_{\beta}(\lambda)$ is $\displaystyle(\ffrac{x_1+\beta_1}{r},\ffrac{x_2+\beta_2}{r},\dots,\ffrac{x_d+\beta_d}{r})$ and the Bayesian estimator based on $\pi_{f,\beta}(\lambda)$ is $\displaystyle(\ffrac{x_1+\beta_1}{r}\ffrac{F(x+\delta_1,r)}{F(x,r)}, \ffrac{x_2+\beta_2}{r}\ffrac{F(x+\delta_2,r)}{F(x,r)},\dots,\ffrac{x_d+\beta_d}{r}\ffrac{F(x+\delta_d,r)}{F(x,r)})$. Therefore, the difference between the K-L risks of Bayesian estimators based on $\pi_{\beta}(\lambda)$ and $\pi_{f,\beta}(\lambda)$ has the same sign with \eqref{differential2}. Using the proof of Theorem 1, we have \vspace{0.5cm} \noindent \textbf{Corollary 2}. The Bayesian estimator based on $\pi_{f,\beta}$ dominates the Bayesian estimator based on $\pi_{\beta}$ if, for every $r>0$, the function $$ F(x,r)=F(x_1,x_2,\dots,x_d,r):=\int f(\bar\lambda_1,\dots,\bar\lambda_d)\prod\limits_{i=1}^d \ffrac{r^{x_i+\beta_i} \bar\lambda_i^{x_i+\beta_i-1}\exp(-r\bar\lambda_i)}{\Gamma(x_i+\beta_i)}\ensuremath{\mathrm{d}}\bar\lambda $$ is not a constant function of $x$, there exists $\epsilon>0$ such that \begin{align*} &\sum_x\Biggl[ \max_{t\in[r-\epsilon,r+\epsilon]}\biggl\{ \bigl| \log F(x,t) \bigr| (\sum_{i=1}^dx_i+1) \prod_{i=1}^d \frac{(t\lambda_i)^{x_i}\exp(-t\lambda_i)}{x_i!} \biggr\} \Biggr] <\infty, \end{align*} and for every $x$, $r>0$, \begin{equation*} \sum\limits_{i=1}^dx_i \bigl\{ F(x,r)-F(x-\delta_i,r) \bigr\} +\sum\limits_{i=1}^d(x_i+\beta_i) \bigl\{ F(x,r)-F(x+\delta_i,r) \bigr\} \ge 0, \end{equation*} where $F(x-\delta_i,r)$ is defined to be $1$ if $x_i=0$. \section{Improved prediction for independent Poisson processes with different durations} \par We consider the case of independent Poisson processes with different durations. Suppose that $x_i$ and $y_i\ (i = 1,\dots,d)$ are independently distributed according to Poisson distributions with mean $r_i\lambda_i$ and $s_i\lambda_i$, respectively, which means $$p(x\mid\lambda)=\prod\limits_{i=1}^d\ffrac{(r_i\lambda_i)^{x_i}}{x_i!}e^{-r_i\lambda_i},$$ $$p(y\mid\lambda)=\prod\limits_{i=1}^d\ffrac{(s_i\lambda_i)^{y_i}}{y_i!}e^{-s_i\lambda_i}.$$ In this section, we denote $\displaystyle\gamma_i:=\ffrac{1}{r_i}-\ffrac{1}{r_i+s_i},\ \theta_i:=\sqrt{\ffrac{\lambda_i}{\gamma_i}},\ i=1,\dots,d$ and $g(\theta_1,\dots,\theta_d):=f(\lambda_1,\dots,\lambda_d)$, $p(x\mid\theta):=p(x\mid\lambda)$. We show that if $g$ satisfies some conditions that are similar to the conditions in Section 2, then the Bayesian predictive distribution $p_{f,\beta}(y\mid x)$ dominates the Bayesian predictive distribution $p_{\beta}(y\mid x)$.\\ \noindent \textbf{Theorem 2.} \begin{description} \item{~1)~\;} The Bayesian predictive distribution $p_{f,\beta}(y\mid x)$ dominates the Bayesian predictive distribution $p_{\beta}(y\mid x)$ if the function $$F(x,r)=F(x_1,x_2,\dots,x_d,r):=\int f(\bar\lambda_1,\dots,\bar\lambda_d)\prod\limits_{i=1}^d\ffrac{r_i^{x_i+\beta_i}\bar\lambda_i^{x_i+\beta_i-1}\exp(-r_i\bar\lambda_i)}{\Gamma(x_i+\beta_i)}\ensuremath{\mathrm{d}}\bar\lambda$$ is not a constant function of $x$ for every $r>0$, there exists $\epsilon>0$ such that \begin{align} &\sum_x\Biggl[ \max_{t\in \bar B(r,\epsilon)}\biggl\{ \Bigl| \log F(x,t) \Bigr| (\sum_{i=1}^dx_i+1) \prod_{i=1}^d \frac{(t_i\lambda_i)^{x_i}\exp(-t_i\lambda_i)}{x_i!} \biggr\} \Biggr] <\infty, \label{F-differentiable-3} \end{align} where $\bar B(r,\epsilon)$ is the closed ball with center $r=(r_1,\dots,r_d)$ and radius $\epsilon$ for every $r>0$, and \begin{equation} \sum\limits_{i=1}^d\gamma_ir_ix_i \Bigl\{ F(x,r)-F(x-\delta_i,r)\Bigr\} +\sum\limits_{i=1}^d\gamma_ir_i(x_i+\beta_i) \Bigl\{F(x,r)-F(x+\delta_i,r) \Bigr\} \ge 0 \label{Fineq3} \end{equation} for every $x$, $r>0$, where $F(x-\delta_i,r):=1$ if $x_i=0$. \item{~2)~\;} Condition \eqref{Fineq3} is satisfied if $g\in\mathbf{C}^2([0,\infty)^d)$, \begin{subequations} \begin{align} &\int g(\theta)\prod\limits_{j=1}^{d}\theta_j^{2\beta_j-1}p(x\mid\theta)\ensuremath{\mathrm{d}}\theta < \infty, \label{t2-condition-a}\\ &\int \Bigl\lvert\frac{\partial g}{\partial\theta_i}(\theta) \Bigr\rvert\prod\limits_{j=1}^{d} \theta_j^{2\beta_j-1}p(x\mid\theta)\ensuremath{\mathrm{d}}\theta < \infty, \label{t2-condition-b}\\ &\int \Bigl\lvert\ffrac{\partial}{\partial\theta_i} \Bigl\{ \frac{\partial g}{\partial\theta_i}(\theta)\prod\limits_{j=1}^d\theta_j^{2\beta_j-1} \Bigr\} \Bigl\rvert p(x\mid\theta)\ensuremath{\mathrm{d}}\theta < \infty \label{t2-condition-c}, \end{align} for every $x$, $i$, $r>0$, \begin{align} &\sum\limits_{i=1}^{d}\ffrac{\partial}{\partial \theta_i} \biggl( \ffrac{\partial g}{\partial \theta_i}(\theta) \prod\limits_{j=1}^{d}\theta_j^{2\beta_j-1} \biggr) \le0, \label{t2-condition-d} \end{align} and \begin{align} &\lim_{\theta_i\to0} \frac{\partial g}{\partial\theta_i}(\theta) \, \theta_i^{2\beta_i-1}\le0, \label{t2-condition-e} \end{align} \end{subequations} for every $i$. \end{description} \noindent \textbf{Proof}. 1) The proof is a generalization of that of Theorem 1. \par For every $i$ and $\tau\in[0,1]$, let $\displaystyle\ffrac{1}{t_i(\tau)}:=\ffrac{1}{r_i}(1-\tau)+\ffrac{1}{r_i+s_i}\tau$. Then $t_i(\tau)$ is a smooth monotonically increasing function of $\tau\in[0, 1]$ satisfying $t_i(0) = r_i$, $t_i(1) = r_i + s_i$ and $\dot t_i/t_i=\gamma_it_i$. The difference between the K-L risks of Bayesian predictive distributions based on $\pi_{f,\beta}$ and $\pi_{\beta}$ is \begin{align} &\text{E} \biggl( \log\ffrac{p_{\beta}(y\mid x)}{p_{f,\beta}(y\mid x)} \, \bigg| \, \lambda \biggr) \notag\\ &= \text{E} \Bigl(\log\int p(x,y\mid\bar\lambda)\pi_{\beta}(\bar\lambda) \ensuremath{\mathrm{d}}\bar\lambda \, \Big| \, \lambda \Bigr) -\text{E} \Bigl( \log\int p(x\mid\bar\lambda)\pi_{\beta}(\bar\lambda) \ensuremath{\mathrm{d}}\bar\lambda \, \Big| \, \lambda \Bigr) \notag\\ &\quad -\text{E} \Bigl( \log\int p(x,y\mid\bar\lambda)\pi_{f,\beta}(\bar\lambda) \ensuremath{\mathrm{d}}\bar\lambda \, \Big| \, \lambda \Bigr) + \text{E} \Bigl( \log\int p(x\mid\bar\lambda)\pi_{f,\beta}(\bar\lambda) \ensuremath{\mathrm{d}}\bar\lambda \,\Big| \, \lambda \Bigr) \notag\\ &=\text{E} \Biggl( \log\ffrac{\int p(x\mid\bar\lambda)\pi_{f,\beta}(\bar\lambda) \ensuremath{\mathrm{d}}\bar\lambda}{\int p(x\mid\bar\lambda)\pi_{\beta}(\bar\lambda) \ensuremath{\mathrm{d}}\bar\lambda} \, \Bigg| \, \lambda \Biggr) -\text{E} \Biggl( \log\ffrac{\int p(x,y\mid\bar\lambda)\pi_{f,\beta}(\bar\lambda) \ensuremath{\mathrm{d}}\bar\lambda}{\int p(x,y\mid\bar\lambda)\pi_{\beta}(\bar\lambda) \ensuremath{\mathrm{d}}\bar\lambda} \, \Bigg| \, \lambda \Biggr) \notag\\ &=\text{E} \Biggl( \log \frac{\int f(\bar\lambda_1,\dots,\bar\lambda_d)\bar\lambda_1^{\beta_1-1}\bar\lambda_2^{\beta_2-1} \cdots\bar\lambda_d^{\beta_d-1} \exp(-\sum_{i=1}^dr_i\bar\lambda_i) (r_1\bar\lambda_1)^{x_1}\cdots(r_d\bar\lambda_d)^{x_d} \ensuremath{\mathrm{d}}\bar\lambda} {\int \bar\lambda_1^{\beta_1-1}\bar\lambda_2^{\beta_2-1}\cdots\bar\lambda_d^{\beta_d-1} \exp(-\sum_{i=1}^dr_i\bar\lambda_i)(r_1\bar\lambda_1)^{x_1}\cdots(r_d\bar\lambda_d)^{x_d} \ensuremath{\mathrm{d}}\bar\lambda} \, \Bigg| \, \lambda \Biggr) \notag\\ &\quad -\text{E} \Biggl( \log\frac{\int f(\bar\lambda_1,\dots,\bar\lambda_d)\bar\lambda_1^{\beta_1-1}\bar\lambda_2^{\beta_2-1}\cdots \bar\lambda_d^{\beta_d-1}\exp\{-\sum_{i=1}^d (r_i+s_i)\bar\lambda_i\} \{(r_1+s_1)\bar\lambda_1\}^{x_1+y_1} \cdots\{(r_d+s_d)\bar \lambda_d \}^{x_d+y_d} \ensuremath{\mathrm{d}}\bar\lambda} {\int \bar\lambda_1^{\beta_1-1}\bar\lambda_2^{\beta_2-1}\cdots\bar\lambda_d^{\beta_d-1} \exp\{-\sum_{i=1}^d(r_i+s_i)\bar\lambda_i\}\{(r_1+s_1)\bar\lambda_1\}^{x_1+y_1} \cdots\{(r_d+s_d)\bar\lambda_d\}^{x_d+y_d} \ensuremath{\mathrm{d}}\bar\lambda} \, \Bigg| \, \lambda \Biggr) \notag\\ &=\text{E} \biggl( \log\int f(\bar\lambda_1,\dots,\bar\lambda_d)\prod_{i=1}^d \frac{r_i^{x_i+\beta_i}\bar\lambda_i^{x_i+\beta_i-1} \exp(-r_i\bar\lambda_i)}{\Gamma(x_i+\beta_i)}\ensuremath{\mathrm{d}}\bar\lambda \, \bigg| \, \lambda \biggr) \notag\\ & \quad -\text{E} \Biggl(\log\int f(\bar\lambda_1,\dots,\bar\lambda_d)\prod\limits_{i=1}^d\ffrac{(r_i+s_i)^{x_i+y_i+\beta_i}\bar\lambda_i^{x_i+y_i+\beta_i-1}\exp\{-(r_i+s_i)\bar\lambda_i\}}{\Gamma(x_i+y_i+\beta_i)}\ensuremath{\mathrm{d}}\bar\lambda \, \Bigg| \, \lambda \Biggr) \notag\\ & = \text{E} \biggl( \log F(x,t(0)) \, \bigg| \, x_i\sim\text{Po}(t_i(0)\lambda_i),i=1,\dots,d \biggr) -\text{E} \biggl( \log F(x,t(1)) \, \bigg| \, x_i\sim\text{Po}(t_i(1)\lambda_i),i=1,\dots,d \biggr). \label{negativediff-2} \end{align} From \eqref{F-differentiable-3}, we have $$ \text{E} \biggl( \Bigl| \log F(x,r) \Bigr| \, \bigg| \, x_i\sim\text{Po}(r_i\lambda_i),i=1,\dots,d \biggr)<\infty $$ for every $r$. Thus, the risk difference \eqref{negativediff-2} is finite. The risk difference \eqref{negativediff-2} is negative if \begin{align} \text{E} \biggl( \log F(x,t(\tau)) \, \bigg| \, x_i\sim\text{Po}(t_i(\tau)\lambda_i),i=1,\dots,d \biggr) =\sum_x\Bigl\{ \log F(x,t(\tau)) \Bigr\} \prod_{i=1}^d \frac{(t_i(\tau)\lambda_i)^{x_i}\exp(-t_i(\tau)\lambda_i)}{x_i!} \label{increasing-t2} \end{align} is an increasing function of $\tau$. \vspace{0.5cm} Using $\dot t_i/t_i=\gamma_it_i$, we have \begin{equation} \ffrac{\partial F}{\partial \tau}(x,t(\tau))=\int f(\bar\lambda_1,\dots,\bar\lambda_d) \prod_{i=1}^d \frac{t_i(\tau)^{x_i+\beta_i}\bar\lambda_i^{x_i+\beta_i-1}\exp(-t_i(\tau)\bar\lambda_i)}{\Gamma(x_i+\beta_i)} \biggl\{ \sum_{j=1}^d (x_j+\beta_j)\gamma_jt_j(\tau) - \sum_{k=1}^d \bar\lambda_k\gamma_kt_k^2(\tau) \biggr\} \ensuremath{\mathrm{d}}\bar\lambda. \label{Fdiff-t2} \end{equation} From Lemma 2.1, we can exchange the integration and differentiation in $\ffrac{\partial F}{\partial \tau}(x,t(\tau))$. Using $\dot t_i/t_i=\gamma_it_i$ and \eqref{Fdiff-t2}, the partial differential function of \eqref{increasing-t2} with respect to $\tau$ is \begin{align} &\sum_x \Bigl\{ \log F(x,t(\tau)) \Bigr\} \Biggl\{ \prod_{i=1}^d \frac{(t_i(\tau)\lambda_i)^{x_i}\exp(-t_i(\tau)\lambda_i)}{x_i!} \Biggr\} \Bigl(\sum_{j=1}^d x_j\gamma_jt_j(\tau) - \sum_{k=1}^d \lambda_k\gamma_kt_k^2(\tau) \Bigr) \notag \notag\\ &\quad + \sum_x \Biggl[ \frac{ \displaystyle \int f(\bar\lambda_1,\dots,\bar\lambda_d) \prod_{i=1}^d \frac{t_i(\tau)^{x_i+\beta_i}\bar\lambda_i^{x_i+\beta_i-1}\exp(-t_i(\tau)\bar\lambda_i)}{\Gamma(x_i+\beta_i)} \Bigl\{ \sum_{j=1}^d (x_j+\beta_j)\gamma_jt_j(\tau) - \sum_{k=1}^d \bar\lambda_k\gamma_kt_k^2(\tau) \Bigr\} \ensuremath{\mathrm{d}}\bar\lambda} { F(x,t(\tau))} \Biggr] \notag\\ & ~~~~~ \Biggl\{ \prod_{l=1}^d \frac{(t_l(\tau)\lambda_l)^{x_l}\exp(-t_l(\tau)\lambda_l)}{x_l!} \Biggr\}. \label{differential-t2} \end{align} From Lemma 2.2, we can differentiate \eqref{increasing-t2} by terms under conditions \eqref{F-differentiable-3}, \eqref{Fineq3}. \vspace{0.5cm} We notice that \begin{align} &\sum_x \biggl\{ \log F(x,t(\tau)) \biggr\} \Biggl\{ \prod_{i=1}^d\ffrac{(t_i(\tau)\lambda_i)^{x_i} \exp(-t_i(\tau)\lambda_i)}{x_i!} \Biggr\} \sum_{j=1}^d \lambda_j\gamma_jt_j^2(\tau) \notag \\ &=\sum_{i=1}^d \sum_x \biggl\{ \log F(x,t(\tau)) \biggr\} \Biggl\{ \prod_{j=1}^d\ffrac{(t_j(\tau)\lambda_j)^{x_j} \exp(-t_j(\tau)\lambda_j)}{x_j!} \Biggr\} \lambda_i\gamma_it_i^2(\tau) \notag \\ &=\sum_{i=1}^d \sum_x \biggl[ (x_i+1)\gamma_it_i(\tau) \Bigl\{ \log F(x,t(\tau)) \Bigr\} \prod\limits_{j=1}^d\ffrac{(t_j(\tau)\lambda_j)^{x_j+\delta_{ij}} \exp(-t_j(\tau)\lambda_j)}{(x_j+\delta_{ij})!} \biggr] \notag\\ &= \sum_{i=1}^d \sum_x \biggl[ x_i\gamma_it_i(\tau) \Bigl\{ \log F(x-\delta_i,t(\tau)) \Bigr\} \prod\limits_{j=1}^d\ffrac{(t_j(\tau)\lambda_j)^{x_j}\exp(-t_j(\tau)\lambda_j)}{x_j!} \biggr], \label{transform1-t2} \end{align} where $\delta_{ij}$ is defined to be 1 if $i = j$ and 0 if $i\neq j$, $\delta_i$ is defined to be the $d$-dimensional vector whose $i$-th element is 1 and all other elements are 0, and $F(x-\delta_i,t(\tau))$ is defined to be $1$ if $x_i=0$. Furthermore, we notice that \begin{align} &\int f(\bar\lambda_1,\dots,\bar\lambda_d) \prod_{j=1}^d\ffrac{t_j(\tau)^{x_j+\beta_j}\bar\lambda_j^{x_j+\beta_j-1} \exp(-t_j(\tau)\bar\lambda_j)}{\Gamma(x_j+\beta_j)} \bar\lambda_i\gamma_it_i^2(\tau) \ensuremath{\mathrm{d}}\bar\lambda \notag\\ &= \int f(\bar\lambda_1,\dots,\bar\lambda_d) \prod_{j=1}^d \ffrac{t_j(\tau)^{x_j+\delta_{ij}+\beta_j}\bar\lambda_j^{x_j+\delta_{ij}+\beta_j-1}\exp(-t_j(\tau)\bar\lambda_j)} {\Gamma(x_j+\delta_{ij}+\beta_j)}(x_i+\beta_i)\gamma_it_i(\tau) \ensuremath{\mathrm{d}}\bar\lambda \notag \\ &= F(x+\delta_i,t(\tau))(x_i+\beta_i)\gamma_it_i(\tau) \label{transform2-t2}. \end{align} \par Thus, from \eqref{transform1-t2} and \eqref{transform2-t2}, the partial differential function \eqref{differential-t2} of \eqref{increasing-t2} with respect to $\tau$ is \begin{align} &\sum_x\sum_{i=1}^d \biggl[ x_i\gamma_it_i(\tau) \bigl\{ \log F(x,t(\tau)) \bigr\} \prod\limits_{j=1}^d\ffrac{(t_j(\tau)\lambda_j)^{x_j}\exp(-t_j(\tau)\lambda_j)}{x_j!} \biggr] \notag\\ &~~~-\sum_x \sum_{i=1}^d \biggl[ x_i\gamma_it_i(\tau) \bigl\{ \log F(x-\delta_i,t(\tau)) \bigr\} \prod_{j=1}^d \ffrac{(t_j(\tau)\lambda_j)^{x_j} \exp(-t_j(\tau)\lambda_j)}{x_j!} \biggr] \notag\\ &~~~+ \sum_x \Biggl\{ \frac{F(x,t(\tau)) \sum_{i=1}^d (x_i+\beta_i)\gamma_it_i(\tau) -\sum_{i=1}^d F(x+\delta_i,t(\tau)) (x_i+\beta_i)\gamma_it_i(\tau)}{F(x,t(\tau))} \Biggr\} \notag\\ &~~~~~~~ \biggl\{ \prod_{j=1}^d\ffrac{(t_j(\tau)\lambda_j)^{x_j}\exp(-t_j(\tau)\lambda_j)}{x_j!}\biggl\} \notag\\ &= \sum_x \Biggl[ \sum_{i=1}^d x_i\gamma_it_i(\tau) \biggl\{ \log \ffrac{F(x,t(\tau))}{F(x-\delta_i,t(\tau))}\biggr\} \prod\limits_{j=1}^d\ffrac{(t_j(\tau)\lambda_j)^{x_j}\exp(-t_j(\tau)\lambda_j)}{x_j!} \Biggr] \notag \\ &~~~+\sum_x \Biggl[ \sum_{i=1}^d (x_i+\beta_i)\gamma_it_i(\tau) \biggl\{ 1-\ffrac{F(x+\delta_i,t(\tau))}{F(x,t(\tau))} \biggr\} \prod_{j=1}^d \ffrac{(t_j(\tau)\lambda_j)^{x_j}\exp(-t_j(\tau)\lambda_j)}{x_j!} \Biggr]. \label{differential2-t2} \end{align} By assumption $F(x,t(\tau))$ is not a constant function of $x$, hence $F(x-\delta_i,t(\tau))\equiv F(x,t(\tau))$ does not hold. Therefore, from the inequality $\log \epsilon >1-\ffrac{1}{\epsilon}$, $\epsilon\neq1$, \eqref{differential2-t2} is strictly greater than \begin{align} &\sum_x \Biggl[ \sum_{i=1}^d x_i\gamma_it_i(\tau) \biggl\{ 1-\ffrac{F(x-\delta_i,t(\tau))}{F(x,t(\tau))} \biggr\} \prod\limits_{j=1}^d\ffrac{(t_j(\tau)\lambda_j)^{x_j}\exp(-t_j(\tau)\lambda_j)}{x_j!} \Biggr] \notag \\ &~~~+\sum_x \Biggl[ \sum_{i=1}^d (x_i+\beta_i)\gamma_it_i(\tau) \biggl\{ 1-\ffrac{F(x+\delta_i,t(\tau))}{F(x,t(\tau))} \biggr\} \prod_{j=1}^d \ffrac{(t_j(\tau)\lambda_j)^{x_j}\exp(-t_j(\tau)\lambda_j)}{x_j!} \Biggr] \notag\\ &=\sum_x \Biggl[ \sum_{i=1}^d x_i\gamma_it_i(\tau) \biggl\{ 1-\ffrac{F(x-\delta_i,t(\tau))}{F(x,t(\tau))} \biggr\} +\sum_{i=1}^d(x_i+\beta_i)\gamma_it_i(\tau) \biggl\{ 1-\ffrac{F(x+\delta_i,t(\tau))}{F(x,t(\tau))} \biggr\}\Biggr] \notag\\ &~~~~~~~ \biggl\{ \prod_{j=1}^d\ffrac{(t_j(\tau)\lambda_j)^{x_j}\exp(-t_j(\tau)\lambda_j)}{x_j!}\biggl\}. \label{Fpositive-t2} \end{align} Using \eqref{Fineq3}, we have $$\sum\limits_{i=1}^d\gamma_it_i(\tau)x_i\bigl\{F(x,t(\tau))-F(x-\delta_i,t(\tau))\bigr\}+\sum\limits_{i=1}^d\gamma_it_i(\tau)(x_i+\beta_i)\bigl\{F(x,t(\tau))-F(x+\delta_i,t(\tau))\bigr\}\ge 0.$$ Thus, \eqref{Fpositive-t2} is not negative, and \eqref{differential2-t2} is positive. Therefore, \eqref{increasing-t2} is an increasing function and the first half of Theorem 2 is proved. \vspace{0.5cm} \noindent 2) Next, we prove that if $f$ satisfies the conditions of the second half of Theorem 2, then \eqref{Fineq3} is satisfied. \par Let $\displaystyle\sqrt{\ffrac{\lambda_j}{\gamma_j}}=:\theta_j,\ j=1,\dots,d$. Then, we have $\lambda_j=\gamma_j\theta_j^2$, $\ensuremath{\mathrm{d}}\lambda_j=2\theta_j\gamma_j\ensuremath{\mathrm{d}}\theta_j$, \begin{align} &F(x+\delta_i,r)-F(x,r) \notag\\ &=\int f(\lambda)\prod\limits_{j=1}^d\ffrac{r_j^{x_j+\delta_{ji}+\beta_j}\lambda_j^{x_j+\delta_{ji}+\beta_j-1}\exp(-r_j\lambda_j)}{\Gamma(x_j+\delta_{ji}+\beta_j)}\ensuremath{\mathrm{d}}\lambda-\int f(\lambda)\prod\limits_{j=1}^d\ffrac{r_j^{x_j+\beta_j}\lambda_j^{x_j+\beta_j-1}\exp(-r_j\lambda_j)}{\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\lambda \notag\\ &=2^d\int g(\theta)\prod\limits_{j=1}^d\ffrac{(\gamma_jr_j)^{x_j+\delta_{ji}+\beta_j}\theta_j^{2x_j+2\delta_{ji}+2\beta_j-1}\exp(-\gamma_jr_j\theta_j^2)}{\Gamma(x_j+\delta_{ji}+\beta_j)}\ensuremath{\mathrm{d}}\theta \notag\\ &\qquad-2^d\int g(\theta)\prod\limits_{j=1}^d\ffrac{(\gamma_jr_j)^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-\gamma_jr_j\theta_j^2)}{\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\theta \notag\\ &=-2^{d-1}\int g(\theta)\prod_{j\neq i}\ffrac{(\gamma_jr_j)^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-\gamma_jr_j\theta_j^2)}{\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\theta_j \ffrac{(\gamma_ir_i)^{x_i+\beta_i}\theta_i^{2x_i+2\beta_i}\ffrac{\partial}{\partial\theta_i}\exp(-\gamma_ir_i\theta_i^2)}{\Gamma(x_i+1+\beta_i)}\ensuremath{\mathrm{d}}\theta_i \notag\\ &\qquad-2^{d-1}\int g(\theta)\prod_{j\neq i}\ffrac{(\gamma_jr_j)^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-\gamma_jr_j\theta_j^2)}{\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\theta_j \ffrac{(\gamma_ir_i)^{x_i+\beta_i}(\ffrac{\partial}{\partial\theta_i}\theta_i^{2x_i+2\beta_i}) \exp(-\gamma_ir_i\theta_i^2)}{\Gamma(x_i+1+\beta_i)}\ensuremath{\mathrm{d}}\theta_i. \notag \end{align} By integration by parts on $\theta_i$, we have \begin{align} &F(x+\delta_i,r)-F(x,r) \notag\\ &=\mathop{\lim}_{a\to0\atop b\to\infty}\mathop{\lim}_{u\to0\atop v\to\infty} \biggl[-2^{d-1}\int_u^v \biggl\{ \int_{[a,b]^{d-1}} g(\theta)\prod\limits_{j\neq i}\ffrac{(\gamma_jr_j)^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-\gamma_jr_j\theta_j^2)} {\Gamma(x_j+\beta_j)} \ensuremath{\mathrm{d}} \theta_j \biggr\} \ffrac{(\gamma_ir_i)^{x_i+\beta_i}\theta_i^{2x_i+2\beta_i}\ffrac{\partial}{\partial\theta_i}\exp(-\gamma_ir_i\theta_i^2)} {\Gamma(x_i+1+\beta_i)}\ensuremath{\mathrm{d}}\theta_i \notag\\ & ~~ -2^{d-1}\int_u^v \biggl\{ \int_{[a,b]^{d-1}} g(\theta)\prod\limits_{j\neq i}\ffrac{(\gamma_jr_j)^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-\gamma_jr_j\theta_j^2)} {\Gamma(x_j+\beta_j)} \ensuremath{\mathrm{d}} \theta_j \biggr\} \ffrac{(\gamma_ir_i)^{x_i+\beta_i}(\ffrac{\partial}{\partial\theta_i}\theta_i^{2x_i+2\beta_i})\exp(-\gamma_ir_i\theta_i^2)} {\Gamma(x_i+1+\beta_i)}\ensuremath{\mathrm{d}}\theta_i\biggl] \notag\\ &=\mathop{\lim}_{a\to0\atop b\to\infty}\mathop{\lim}_{u\to0\atop v\to\infty} \Biggl(2^{d-1}\int_u^v \biggl\{ \ffrac{\partial}{\partial\theta_i} \int_{[a,b]^{d-1}} g(\theta)\prod\limits_{j\neq i}\ffrac{(\gamma_jr_j)^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-\gamma_jr_j\theta_j^2)} {\Gamma(x_j+\beta_j)} \ensuremath{\mathrm{d}} \theta_j \biggr\} \ffrac{(\gamma_ir_i)^{x_i+\beta_i}\theta_i^{2x_i+2\beta_i}\exp(-\gamma_ir_i\theta_i^2)}{\Gamma(x_i+1+\beta_i)}\ensuremath{\mathrm{d}}\theta_i \notag\\ & ~~~~ -2^{d-1} \Biggl[ \biggl\{ \int_{[a,b]^{d-1}} g(\theta)\prod\limits_{j\neq i}\ffrac{(\gamma_jr_j)^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-\gamma_jr_j\theta_j^2)} {\Gamma(x_j+\beta_j)} \ensuremath{\mathrm{d}} \theta_j \biggr\} \ffrac{(\gamma_ir_i)^{x_i+\beta_i}\theta_i^{2x_i+2\beta_i}\exp(-\gamma_ir_i\theta_i^2)} {\Gamma(x_i+1+\beta_i)} \Biggr]^{v}_{\theta_i = u}\Biggl) \notag\\ &=\mathop{\lim}_{a\to0\atop b\to\infty}\mathop{\lim}_{u\to0\atop v\to\infty} \Biggl( 2^{d-1} \int_{[u,v]\times[a,b]^{d-1}} \ffrac{\partial g }{\partial\theta_i}(\theta) \prod\limits_{j=1}^d\ffrac{(\gamma_jr_j)^{x_j+\beta_j} \theta_j^{2x_j+\delta_{ji}+2\beta_j-1}\exp(-\gamma_jr_j\theta_j^2)}{\Gamma(x_j+\delta_{ji}+\beta_j)} \ensuremath{\mathrm{d}}\theta \notag\\ & ~~~~ -2^{d-1} \Biggl[ \biggl\{ \int_{[a,b]^{d-1}} g(\theta)\prod\limits_{j\neq i}\ffrac{(\gamma_jr_j)^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-\gamma_jr_j\theta_j^2)} {\Gamma(x_j+\beta_j)} \ensuremath{\mathrm{d}} \theta_j \biggr\} \ffrac{(\gamma_ir_i)^{x_i+\beta_i}\theta_i^{2x_i+2\beta_i}\exp(-\gamma_ir_i\theta_i^2)} {\Gamma(x_i+1+\beta_i)} \Biggr]^{v}_{\theta_i = u}\Biggl) . \label{F-diff-1-t2} \end{align} Here, we use auxiliary variables $a,b,u,v$ and Lemma 2.3 to ensure the above equations hold. \vspace{0.2cm} \par Using Lemma 2.3, we have $$ \int \biggl\{ \int_{[a,b]^{d-1}} g(\theta)\prod\limits_{j\neq i} \ffrac{(\gamma_jr_j)^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-\gamma_jr_j\theta_j^2)} {\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\theta_j \biggr\} \ffrac{(\gamma_ir_i)^{x_i+\beta_i}\theta_i^{2x_i+2\beta_i} \exp(-\gamma_ir_i\theta_i^2)}{\Gamma(x_i+1+\beta_i)}\ensuremath{\mathrm{d}}\theta_i <\infty.$$ Because of Lemma 2.4, $$\displaystyle \biggl\{ \int_{[a,b]^{d-1}} g(\theta)\prod\limits_{j\neq i} \ffrac{(\gamma_jr_j)^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-\gamma_jr_j\theta_j^2)} {\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\theta_j \biggr\} \ffrac{(\gamma_ir_i)^{x_i+\beta_i}\theta_i^{2x_i+2\beta_i} \exp(-\gamma_ir_i\theta_i^2)}{\Gamma(x_i+1+\beta_i)}$$ converges as $\theta_i\to\infty$. Thus, \begin{equation} \lim_{\theta_i\to\infty} \biggl\{ \int_{[a,b]^{d-1}} g(\theta)\prod\limits_{j\neq i} \ffrac{(\gamma_jr_j)^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-\gamma_jr_j\theta_j^2)} {\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\theta_j \biggr\} \ffrac{(\gamma_ir_i)^{x_i+\beta_i}\theta_i^{2x_i+2\beta_i} \exp(-\gamma_ir_i\theta_i^2)}{\Gamma(x_i+1+\beta_i)} = 0. \label{1st-partial-1-t2} \end{equation} Because $g\in \mathbf{C}^2([0,\infty)^d)$, $g$ is bounded on $[0,1]\times[a,b]^{d-1}$, then $$ \int_{[a,b]^{d-1}} g(\theta)\prod\limits_{j\neq i} \ffrac{(\gamma_jr_j)^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-\gamma_jr_j\theta_j^2)} {\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\theta_j $$ is bounded for $\theta_i\le1$. Thus, \begin{equation} \lim_{\theta_i\to0} \biggl\{ \int_{[a,b]^{d-1}} g(\theta)\prod\limits_{j\neq i} \ffrac{(\gamma_jr_j)^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-\gamma_jr_j\theta_j^2)} {\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\theta_j \biggr\} \ffrac{(\gamma_ir_i)^{x_i+\beta_i}\theta_i^{2x_i+2\beta_i} \exp(-\gamma_ir_i\theta_i^2)}{\Gamma(x_i+1+\beta_i)} = 0. \label{1st-partial-2-t2} \end{equation} \vspace{0.3cm} Consequently, using \eqref{1st-partial-1-t2} and \eqref{1st-partial-2-t2}, we have \begin{equation} \mathop{\lim}_{u\to0\atop v\to\infty} -2^{d-1} \Biggl[ \biggl\{ \int_{[a,b]^{d-1}} g(\theta)\prod\limits_{j\neq i}\ffrac{(\gamma_jr_j)^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-\gamma_jr_j\theta_j^2)} {\Gamma(x_j+\beta_j)} \ensuremath{\mathrm{d}} \theta_j \biggr\} \ffrac{(\gamma_ir_i)^{x_i+\beta_i}\theta_i^{2x_i+2\beta_i}\exp(-\gamma_ir_i\theta_i^2)} {\Gamma(x_i+1+\beta_i)} \Biggr]^{v}_{\theta_i = u}=0. \label{1st-partial-3-t2} \end{equation} Thus, using \eqref{F-diff-1-t2}, Lemma 2.3, \eqref{1st-partial-3-t2}, we have \begin{align} &F(x+\delta_i,r)-F(x,r) \notag\\ &=\mathop{\lim}_{a\to0\atop b\to\infty}\mathop{\lim}_{u\to0\atop v\to\infty} \biggl[2^{d-1} \int_{[u,v]\times[a,b]^{d-1}} \ffrac{\partial g}{\partial\theta_i}(\theta) \prod\limits_{j=1}^d\ffrac{(\gamma_jr_j)^{x_j+\beta_j} \theta_j^{2x_j+\delta_{ji}+2\beta_j-1}\exp(-\gamma_jr_j\theta_j^2)}{\Gamma(x_j+\delta_{ji}+\beta_j)} \ensuremath{\mathrm{d}}\theta\biggl] \notag\\ &= 2^{d-1}\int \ffrac{\partial g}{\partial\theta_i}(\theta) \prod\limits_{j=1}^d \ffrac{(\gamma_jr_j)^{x_j+\beta_j}\theta_j^{2x_j+\delta_{ji}+2\beta_j-1}\exp(-\gamma_jr_j\theta_j^2)} {\Gamma(x_j+\delta_{ji}+\beta_j)}\ensuremath{\mathrm{d}}\theta. \label{F-1st-diff-t2} \end{align} \vspace{0.5cm} Using \eqref{F-1st-diff-t2} and integration by parts on each parameter again, we have \begin{align} &\sum_{i=1}^d \gamma_ir_ix_i \{ F(x,r)-F(x-\delta_i,r) \} +\sum_{i=1}^d \gamma_ir_i(x_i+\beta_i) \{F(x,r)-F(x+\delta_i,r)\} \notag\\ &=\sum_{i=1}^d 2^{d-1} \int\biggl\{ \ffrac{\partial g}{\partial\theta_i}(\theta) \gamma_ir_ix_i \prod_{j=1}^d\ffrac{(\gamma_jr_j)^{x_j+\beta_j-\delta_{ji}}\theta_j^{2(x_j-\delta_{ji})+\delta_{ji}+2\beta_j-1} \exp(-\gamma_jr_j\theta_j^2)} {\Gamma(x_j+\beta_j)} \notag\\ & \qquad- \ffrac{\partial g}{\partial\theta_i} (\theta) \gamma_ir_i(x_i+\beta_i) \prod_{j=1}^d \ffrac{(\gamma_jr_j)^{x_j+\beta_j}\theta_j^{2x_j+\delta_{ji}+2\beta_j-1}\exp(-\gamma_jr_j\theta_j^2)}{\Gamma(x_j+\delta_{ji}+\beta_j)} \biggr\} \ensuremath{\mathrm{d}}\theta \notag\\ &=\sum_{i=1}^d 2^{d-2} \int\biggl\{ \ffrac{\partial g}{\partial\theta_i}(\theta) 2x_i \prod_{j=1}^d\ffrac{(\gamma_jr_j)^{x_j+\beta_j}\theta_j^{2x_j-\delta_{ji}+2\beta_j-1} \exp(-\gamma_jr_j\theta_j^2)} {\Gamma(x_j+\beta_j)} \notag\\ & \qquad- \ffrac{\partial g}{\partial\theta_i} (\theta) 2\gamma_ir_i \prod_{j=1}^d \ffrac{(\gamma_jr_j)^{x_j+\beta_j}\theta_j^{2x_j+\delta_{ji}+2\beta_j-1}\exp(-\gamma_jr_j\theta_j^2)}{\Gamma(x_j+\beta_j)} \biggr\} \ensuremath{\mathrm{d}}\theta \notag\\ &=\sum_{i=1}^d 2^{d-2} \int\biggl[ \ffrac{\partial g}{\partial\theta_i}(\theta)\theta_i^{2\beta_i-1} \biggl\{ \prod_{j\neq i}\ffrac{(\gamma_jr_j)^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-\gamma_jr_j\theta_j^2)} {\Gamma(x_j+\beta_j)}\biggr\} \ffrac{(\gamma_ir_i)^{x_i+\beta_i}\ffrac{\partial\theta_i^{2x_i}}{\partial\theta_i}\exp(-\gamma_ir_i\theta_i^2)}{\Gamma(x_i+\beta_i)} \notag\\ & \qquad+ \ffrac{\partial g}{\partial\theta_i}(\theta)\theta_i^{2\beta_i-1} \biggl\{ \prod_{j\neq i}\ffrac{(\gamma_jr_j)^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-\gamma_jr_j\theta_j^2)} {\Gamma(x_j+\beta_j)}\biggr\} \ffrac{(\gamma_ir_i)^{x_i+\beta_i}\theta_i^{2x_i}\ffrac{\partial}{\partial\theta_i}\exp(-\gamma_ir_i\theta_i^2)} {\Gamma(x_i+\beta_i)} \biggr] \ensuremath{\mathrm{d}}\theta \notag\\ &=\sum_{i=1}^d 2^{d-2} \mathop{\lim}_{a\to0\atop b\to\infty}\mathop{\lim}_{u\to0\atop v\to\infty} \Biggl\{ \int_u^v\int_{[a,b]^{d-1}} \ffrac{\partial g}{\partial\theta_i}(\theta)\theta_i^{2\beta_i-1} \prod_{j\neq i}\ffrac{(\gamma_jr_j)^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-\gamma_jr_j\theta_j^2)}{\Gamma(x_j+\beta_j)} \ensuremath{\mathrm{d}}\theta_j\notag\\ &~~~~~~~~~~~~~~~~~~ \ffrac{(\gamma_ir_i)^{x_i+\beta_i}\ffrac{\partial\theta_i^{2x_i}}{\partial\theta_i}\exp(-\gamma_ir_i\theta_i^2)}{\Gamma(x_i+\beta_i)}\ensuremath{\mathrm{d}}\theta_i \notag\\ & \qquad+ \int_u^v\int_{[a,b]^{d-1}} \ffrac{\partial g}{\partial\theta_i}(\theta)\theta_i^{2\beta_i-1} \biggl\{ \prod_{j\neq i}\ffrac{(\gamma_jr_j)^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-\gamma_jr_j\theta_j^2)}{\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\theta_j \biggr\} \ffrac{(\gamma_ir_i)^{x_i+\beta_i}\theta_i^{2x_i}\ffrac{\partial}{\partial\theta_i}\exp(-\gamma_ir_i\theta_i^2)} {\Gamma(x_i+\beta_i)}\ensuremath{\mathrm{d}}\theta_i \Biggr\} \notag\\ &=\sum_{i=1}^d 2^{d-2}\mathop{\lim}_{a\to0\atop b\to\infty}\mathop{\lim}_{u\to0\atop v\to\infty} \Biggl\{ -\int_u^v\biggl\{ \ffrac{\partial}{\partial\theta_i}\int_{[a,b]^{d-1}} \ffrac{\partial g}{\partial\theta_i}(\theta)\theta_i^{2\beta_i-1} \prod_{j\neq i}\ffrac{(\gamma_jr_j)^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-\gamma_jr_j\theta_j^2)}{\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\theta_j\biggr\} \notag\\ &~~~~~~~~~~~~~~~~~~ \ffrac{(\gamma_ir_i)^{x_i+\beta_i}\theta_i^{2x_i}\exp(-\gamma_ir_i\theta_i^2)}{\Gamma(x_i+\beta_i)}\ensuremath{\mathrm{d}}\theta_i \notag\\ &\qquad+ \biggl[ \int_{[a,b]^{d-1}} \ffrac{\partial g}{\partial\theta_i}(\theta)\theta_i^{2\beta_i-1} \biggl\{ \prod_{j\neq i}\ffrac{(\gamma_jr_j)^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-\gamma_jr_j\theta_j^2)} {\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\theta_j \biggr\} \ffrac{(\gamma_ir_i)^{x_i+\beta_i}\theta_i^{2x_i}\exp(-\gamma_ir_i\theta_i^2)}{\Gamma(x_i+\beta_i)} \biggr] ^{v}_{\theta_i = u} \Biggr\} \notag\\ &=\sum_{i=1}^d 2^{d-2}\mathop{\lim}_{a\to0\atop b\to\infty}\mathop{\lim}_{u\to0\atop v\to\infty} \Biggl\{ -\int_u^v\int_{[a,b]^{d-1}} \ffrac{\partial}{\partial\theta_i}\Bigl\{\ffrac{\partial g}{\partial\theta_i}(\theta)\theta_i^{2\beta_i-1}\Bigr\} \prod_{j\neq i}\ffrac{(\gamma_jr_j)^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-\gamma_jr_j\theta_j^2)}{\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\theta_j \notag\\ &~~~~~~~~~~~~~~~~~~ \ffrac{(\gamma_ir_i)^{x_i+\beta_i}\theta_i^{2x_i}\exp(-\gamma_ir_i\theta_i^2)}{\Gamma(x_i+\beta_i)}\ensuremath{\mathrm{d}}\theta_i \notag \\ &\qquad+ \biggl[ \int_{[a,b]^{d-1}} \ffrac{\partial g}{\partial\theta_i}(\theta)\theta_i^{2\beta_i-1} \biggl\{ \prod_{j\neq i}\ffrac{(\gamma_jr_j)^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-\gamma_jr_j\theta_j^2)}{\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\theta_j\biggr\} \ffrac{(\gamma_ir_i)^{x_i+\beta_i}\theta_i^{2x_i}\exp(-\gamma_ir_i\theta_i^2)}{\Gamma(x_i+\beta_i)} \biggr]^{v}_{\theta_i = u} \Biggr\} . \label{F-diff-2-t2} \end{align} Here, we use auxiliary variables $a,b,u,v$ to ensure the above equations hold. \vspace{0.2cm} From \eqref{t2-condition-c}, we have $$\int \Bigl\lvert\ffrac{\partial}{\partial\theta_i} \Bigl\{ \frac{\partial g}{\partial\theta_i}(\theta)\prod\limits_{j=1}^d\theta_j^{2\beta_j-1} \Bigr\} \Bigl\rvert \,\prod_{j=1}^d\theta_j^{2x_j}\exp(-\gamma_jr_j\theta_j^2)\ensuremath{\mathrm{d}}\theta<\infty.$$ Thus, \begin{equation} \int\Bigl\lvert\ffrac{\partial}{\partial\theta_i} \Bigl\{ \frac{\partial g}{\partial\theta_i}(\theta)\prod\limits_{j=1}^d\theta_j^{2\beta_j-1} \Bigr\} \Bigl\rvert \prod\limits_{j=1}^d\ffrac{(\gamma_jr_j)^{x_j+\beta_j}\theta_j^{2x_j} \exp(-\gamma_jr_j\theta_j^2)}{\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\theta <\infty. \label{2nd-diff-finite-t2} \end{equation} \vspace{0.2cm} From \eqref{t2-condition-b}, we have $$\int \Bigl\lvert\ffrac{\partial g}{\partial\theta_i}(\theta)\Bigl\rvert\prod\limits_{j=1}^{d}\theta_j^{2\beta_j-1}\theta_j^{2x_j}\exp(-\gamma_jr_j\theta_j^2)\ensuremath{\mathrm{d}}\theta<\infty.$$ Thus, \begin{equation} \int \Bigl\lvert\ffrac{\partial g}{\partial\theta_i}(\theta)\Bigl\rvert\prod\limits_{j=1}^d \ffrac{(\gamma_jr_j)^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-\gamma_jr_j\theta_j^2)} {\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\theta_j<\infty. \label{1nd-diff-t2} \end{equation} From \eqref{1nd-diff-t2}, we have \begin{equation} \int \int_{[a,b]^{d-1}} \ffrac{\partial g}{\partial\theta_i}(\theta)\theta_i^{2\beta_i-1} \biggl\{ \prod_{j\neq i}\ffrac{(\gamma_jr_j)^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-\gamma_jr_j\theta_j^2)} {\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\theta_j\biggr\} \ffrac{(\gamma_ir_i)^{x_i+\beta_i}\theta_i^{2x_i}\exp(-\gamma_ir_i\theta_i^2)}{\Gamma(x_i+\beta_i)}\ensuremath{\mathrm{d}}\theta_i<\infty. \label{2st-partial-3-t2} \end{equation} Because of Lemma 2.5, $$\int_{[a,b]^{d-1}} \ffrac{\partial g}{\partial\theta_i}(\theta)\theta_i^{2\beta_i-1} \biggl\{ \prod_{j\neq i}\ffrac{(\gamma_jr_j)^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-\gamma_jr_j\theta_j^2)} {\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\theta_j\biggr\} \ffrac{(\gamma_ir_i)^{x_i+\beta_i}\theta_i^{2x_i}\exp(-\gamma_ir_i\theta_i^2)}{\Gamma(x_i+\beta_i)}$$ converges as $\theta_i\to0$ or $\theta_i\to\infty$. Thus, using \eqref{2st-partial-3-t2}, we have \begin{equation} \lim_{\theta_i\to\infty} \int_{[a,b]^{d-1}} \ffrac{\partial g}{\partial\theta_i}(\theta)\theta_i^{2\beta_i-1} \biggl\{ \prod_{j\neq i}\ffrac{(\gamma_jr_j)^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-\gamma_jr_j\theta_j^2)}{\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\theta_j\biggr\} \ffrac{(\gamma_ir_i)^{x_i+\beta_i}\theta_i^{2x_i}\exp(-\gamma_ir_i\theta_i^2)}{\Gamma(x_i+\beta_i)} = 0. \label{2st-partial-1-t2} \end{equation} \vspace{0.3cm} From \eqref{t2-condition-e}, we have \begin{equation} \lim_{\theta_i\to0} \int_{[a,b]^{d-1}} \ffrac{\partial g}{\partial\theta_i}(\theta)\theta_i^{2\beta_i-1} \biggl\{ \prod_{j\neq i}\ffrac{(\gamma_jr_j)^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-\gamma_jr_j\theta_j^2)} {\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\theta_j \biggr\}\ffrac{(\gamma_ir_i)^{x_i+\beta_i}\theta_i^{2x_i}\exp(-\gamma_ir_i\theta_i^2)}{\Gamma(x_i+\beta_i)}\le0 \label{2st-partial-2-t2} \end{equation} \vspace{0.3cm} Using \eqref{2st-partial-1-t2} and \eqref{2st-partial-2-t2}, we have \begin{align} &\mathop{\lim}_{u\to0\atop v\to\infty} \Biggl[\int_{[a,b]^{d-1}} \ffrac{\partial g}{\partial\theta_i}(\theta)\theta_i^{2\beta_i-1} \biggl\{ \prod_{j\neq i}\ffrac{(\gamma_jr_j)^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-\gamma_jr_j\theta_j^2)} {\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\theta_j \biggr\} \ffrac{(\gamma_ir_i)^{x_i+\beta_i}\theta_i^{2x_i}\exp(-\gamma_ir_i\theta_i^2)}{\Gamma(x_i+\beta_i)} \Biggr]^{v}_{\theta_i = u} \ge0. \label{2st-partial-4-t2} \end{align} Using \eqref{F-diff-2-t2}, \eqref{2nd-diff-finite-t2}, and \eqref{2st-partial-4-t2}, we have \begin{align} & \sum_{i=1}^d \gamma_ir_ix_i \{ F(x,r)-F(x-\delta_i,r) \} +\sum_{i=1}^d \gamma_ir_i(x_i+\beta_i) \{F(x,r)-F(x+\delta_i,r)\} \notag\\ &\ge \sum_{i=1}^d 2^{d-2}\mathop{\lim}_{a\to0\atop b\to\infty}\mathop{\lim}_{u\to0\atop v\to\infty} \Biggl[ -\int_u^v\int_{[a,b]^{d-1}} \ffrac{\partial}{\partial\theta_i}\Bigl\{\ffrac{\partial g}{\partial\theta_i}(\theta)\theta_i^{2\beta_i-1}\Bigr\} \prod_{j\neq i}\ffrac{(\gamma_jr_j)^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-\gamma_jr_j\theta_j^2)}{\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\theta_j \notag\\ &~~~~~~~~~~~~~~~~~~ \ffrac{(\gamma_ir_i)^{x_i+\beta_i}\theta_i^{2x_i}\exp(-\gamma_ir_i\theta_i^2)}{\Gamma(x_i+\beta_i)}\ensuremath{\mathrm{d}}\theta_i \Biggl] \notag\\ &=\sum_{i=1}^d -2^{d-2} \int \ffrac{\partial}{\partial\theta_i} \biggl\{ \ffrac{\partial g}{\partial\theta_i}(\theta) \prod_{j=1}^d\theta_j^{2\beta_j-1} \biggr\} \prod_{j=1}^d \ffrac{(\gamma_jr_j)^{x_j+\beta_j}\theta_j^{2x_j} \exp(-\gamma_jr_j\theta_j^2)}{\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\theta \notag\\ &=-2^{d-2} \int \sum_{i=1}^d \ffrac{\partial}{\partial\theta_i} \biggl\{ \ffrac{\partial g}{\partial\theta_i}(\theta) \prod_{j=1}^d\theta_j^{2\beta_j-1} \biggr\} \prod_{j=1}^d \ffrac{(\gamma_jr_j)^{x_j+\beta_j}\theta_j^{2x_j} \exp(-\gamma_jr_j\theta_j^2)}{\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\theta, \label{F-2nd-diff-2} \end{align} From \eqref{F-2nd-diff-2}, we know that \eqref{Fineq3} holds if $g$ satisfies the conditions of the second half of Theorem 2. \qed \section{Examples} We show some examples that satisfy the conditions in Sections 2 and 3. These include point shrinkage priors and subspace shrinkage priors.\\ \noindent \textbf{Example 1}. We consider the class of prior in \cite{komaki2004simultaneous} and \cite{komaki2015simultaneous}, which is $$\pi_{\alpha,\beta}(\lambda)\ensuremath{\mathrm{d}}\lambda_1\ensuremath{\mathrm{d}}\lambda_2\cdots \ensuremath{\mathrm{d}}\lambda_d\propto\ffrac{\lambda_1^{\beta_1-1}\lambda_2^{\beta_2-1}\cdots\lambda_d^{\beta_d-1}}{(\lambda_1+\lambda_2+\cdots+\lambda_d)^{\alpha}}\ensuremath{\mathrm{d}}\lambda_1\ensuremath{\mathrm{d}}\lambda_2\cdots \ensuremath{\mathrm{d}}\lambda_d$$ for prediction for independent Poisson processes with the same duration, and $$\pi_{\alpha,\beta,\gamma}(\lambda)\ensuremath{\mathrm{d}}\lambda_1\ensuremath{\mathrm{d}}\lambda_2\cdots \ensuremath{\mathrm{d}}\lambda_d\propto\ffrac{\lambda_1^{\beta_1-1}\lambda_2^{\beta_2-1}\cdots\lambda_d^{\beta_d-1}}{(\lambda_1/\gamma_1+\lambda_2/\gamma_2+\cdots+\lambda_d/\gamma_d)^{\alpha}}\ensuremath{\mathrm{d}}\lambda_1\ensuremath{\mathrm{d}}\lambda_2\cdots \ensuremath{\mathrm{d}}\lambda_d$$ for prediction for independent Poisson processes with different durations, where $0<\alpha\le\sum\limits_{i=1}^d\beta_i-1$. The two priors are the same as $\pi_{f, \beta}(\lambda)$ in this study, where $f(\lambda)=g(\theta)=(\sum\limits_{i=1}^d\theta_i^2)^{-\alpha}$.\\ \noindent \textbf{Proposition 1}. When $f(\lambda)=g(\theta)=(\sum\limits_{i=1}^d\theta_i^2+\eta)^{-\alpha}$, $\eta\ge0$, $0<\alpha\le\sum\limits_{i=1}^d\beta_i-1$, the Bayesian predictive distribution $p_{f,\beta}(y\mid x)$ dominates the Bayesian predictive distribution $p_{\beta}(y\mid x)$, for prediction of independent Poisson processes with the same or different durations. \noindent \textbf{Proof}. \par Because independent Poisson processes with the same duration correspond to $r_1=r_2=\cdots=r_d$ and $s_1=s_2=\cdots=s_d$, we need only prove that $f(\lambda)=g(\theta)=(\sum\limits_{i=1}^d\theta_i^2+\eta)^{-\alpha}$ satisfies the conditions in the first half of Theorem 2. We show this in three parts. \vspace{0.3cm} \noindent \textbf{Part 1}. Condition \eqref{Fineq3} is satisfied. \par We first show that $g(\theta)=(\sum\limits_{i=1}^d\theta_i^2+\eta)^{-\alpha}$ satisfies the conditions in the second half of Theorem 2, when $\eta>0,\,0<\alpha\le\sum\limits_{i=1}^d\beta_i-1$. Here, $\displaystyle\theta_i=\sqrt{\ffrac{\lambda_i}{\gamma_i}},\ i=1,\dots,d$. \eqref{t2-condition-a} is satisfied because \begin{align} &\int g(\theta)\prod\limits_{j=1}^{d}\theta_j^{2\beta_j-1}p(x\mid\theta)\ensuremath{\mathrm{d}}\theta \notag\\ &= \int (\sum\limits_{j=1}^{d}\theta_j^2+\eta)^{-\alpha}\prod\limits_{j=1}^{d}\ffrac{\theta_j^{2\beta_j-1}(r_j\lambda_j)^{x_j}}{x_j!}e^{-r_j\lambda_j}\ensuremath{\mathrm{d}}\theta\notag\\ &\le \int \eta^{-\alpha}\prod\limits_{j=1}^{d}\ffrac{\theta_j^{2\beta_j-1}(r_j\lambda_j)^{x_j}}{x_j!}e^{-r_j\lambda_j}\ensuremath{\mathrm{d}}\theta\notag\\ &= 2^{-d}\eta^{-\alpha}\prod\limits_{j=1}^{d}\int\ffrac{\lambda_j^{x_j+\beta_j-1}r_j^{x_j}}{\gamma_j^{\beta_j}x_j!}e^{-r_j\lambda_j}\ensuremath{\mathrm{d}}\lambda_j\notag\\ &= 2^{-d}\eta^{-\alpha}\prod\limits_{j=1}^{d}\ffrac{\Gamma(x_j+\beta_j)}{(\gamma_jr_j)^{\beta_j}x_j!} \notag\\ &<\infty. \notag \end{align} \eqref{t2-condition-b} is satisfied because \begin{align} &\int \Bigl\lvert\ffrac{\partial g}{\partial\theta_i}(\theta) \Bigr\rvert\prod\limits_{j=1}^{d}\theta_j^{2\beta_j-1}p(x\mid\theta)\ensuremath{\mathrm{d}}\theta \notag\\ &= \int 2\alpha(\sum\limits_{j=1}^{d}\theta_j^2+\eta)^{-\alpha-1}\theta_i\prod\limits_{j=1}^{d}\ffrac{\theta_j^{2\beta_j-1}(r_j\lambda_j)^{x_j}}{x_j!}e^{-r_j\lambda_j}\ensuremath{\mathrm{d}}\theta\notag\\ &\le \int 2\alpha\eta^{-\alpha-1}\theta_i\prod\limits_{j=1}^{d}\ffrac{\theta_j^{2\beta_j-1}(r_j\lambda_j)^{x_j}}{x_j!}e^{-r_j\lambda_j}\ensuremath{\mathrm{d}}\theta\notag\\ &= 2^{-d+1}\alpha\eta^{-\alpha-1}\prod\limits_{j=1}^{d}\int\ffrac{\lambda_j^{x_j+\beta_j+\delta_{ji}/2-1}r_j^{x_j}}{\gamma_j^{\beta_j+\delta_{ji}/2}x_j!}e^{-r_j\lambda_j}\ensuremath{\mathrm{d}}\lambda_j\notag\\ &= 2^{-d+1}\alpha\eta^{-\alpha-1}\prod\limits_{j=1}^{d}\ffrac{\Gamma(x_j+\beta_j+\delta_{ji}/2)}{(\gamma_jr_j)^{\beta_j+\delta_{ji}/2}x_j!} \notag\\ &<\infty. \notag \end{align} \eqref{t2-condition-c} is satisfied because \begin{align} &\int \Bigl\lvert\ffrac{\partial}{\partial\theta_i} \Bigl\{ \frac{\partial g}{\partial\theta_i}(\theta)\prod\limits_{j=1}^d\theta_j^{2\beta_j-1} \Bigr\} \Bigl\rvert p(x\mid\theta)\ensuremath{\mathrm{d}}\theta \notag\\ &\le \int 4\alpha(\alpha+1)(\sum\limits_{j=1}^{d}\theta_j^2+\eta)^{-\alpha-2}\theta_i^2\prod\limits_{j=1}^{d}\theta_j^{2\beta_j-1}p(x\mid\theta)\ensuremath{\mathrm{d}}\theta \notag\\ &\qquad + \int 4\alpha\beta_i(\sum\limits_{j=1}^{d}\theta_j^2+\eta)^{-\alpha-1}\prod\limits_{j=1}^{d}\theta_j^{2\beta_j-1}p(x\mid\theta)\ensuremath{\mathrm{d}}\theta \notag\\ &\le \int 4\alpha(\alpha+1)\eta^{-\alpha-2}\theta_i^2\prod\limits_{j=1}^{d}\ffrac{\theta_j^{2\beta_j-1}(r_j\lambda_j)^{x_j}}{x_j!}e^{-r_j\lambda_j}\ensuremath{\mathrm{d}}\theta\notag\\ &\qquad + \int 4\alpha\beta_i\eta^{-\alpha-1}\prod\limits_{j=1}^{d}\ffrac{\theta_j^{2\beta_j-1}(r_j\lambda_j)^{x_j}}{x_j!}e^{-r_j\lambda_j}\ensuremath{\mathrm{d}}\theta\notag\\ &= 2^{-d+2}\alpha(\alpha+1)\eta^{-\alpha-2}\prod\limits_{j=1}^{d}\int\ffrac{\lambda_j^{x_j+\beta_j+\delta_{ji}-1}r_j^{x_j}}{\gamma_j^{\beta_j+\delta_{ji}}x_j!}e^{-r_j\lambda_j}\ensuremath{\mathrm{d}}\lambda_j\notag\\ &\qquad +2^{-d+2}\alpha\beta_i\eta^{-\alpha-1}\prod\limits_{j=1}^{d}\int\ffrac{\lambda_j^{x_j+\beta_j-1}r_j^{x_j}}{\gamma_j^{\beta_j}x_j!}e^{-r_j\lambda_j}\ensuremath{\mathrm{d}}\lambda_j\notag\\ &= 2^{-d+2}\alpha(\alpha+1)\eta^{-\alpha-2}\prod\limits_{j=1}^{d}\ffrac{\Gamma(x_j+\beta_j+\delta_{ji})}{(\gamma_jr_j)^{\beta_j+\delta_{ji}}x_j!} \notag\\ &\qquad +2^{-d+2}\alpha\beta_i\eta^{-\alpha-1}\prod\limits_{j=1}^{d}\ffrac{\Gamma(x_j+\beta_j)}{(\gamma_jr_j)^{\beta_j}x_j!} \notag\\ &<\infty. \notag \end{align} \eqref{t2-condition-d} is satisfied because \begin{align} &\sum\limits_{i=1}^{d}\ffrac{\partial}{\partial \theta_i} \biggl( \ffrac{\partial g}{\partial \theta_i}(\theta) \prod\limits_{j=1}^{d}\theta_j^{2\beta_j-1} \biggr) \notag\\ =& \sum\limits_{i=1}^{d}\ffrac{\partial}{\partial \theta_i}\Bigl\{-2\alpha\theta_i(\sum\limits_{j=1}^{d}\theta_j^2+\eta)^{-\alpha-1}\prod\limits_{j=1}^{d}\theta_j^{2\beta_j-1}\Bigr\} \notag\\ =& -2\alpha(\sum\limits_{j=1}^{d}\theta_j^2+\eta)^{-\alpha-2}\prod\limits_{j=1}^{d}\theta_j^{2\beta_j-1}\sum\limits_{i=1}^{d}\Bigl\{2\beta_i(\sum\limits_{j=1}^{d}\theta_j^2+\eta)+2(-\alpha-1)\theta_i^2\Bigr\} \notag\\ =& -4\alpha(\sum\limits_{j=1}^{d}\theta_j^2+\eta)^{-\alpha-2}\prod\limits_{j=1}^{d}\theta_j^{2\beta_j-1}\Bigl\{(\sum\limits_{i=1}^{d}\beta_i-\alpha-1)(\sum\limits_{j=1}^{d}\theta_j^2)+\sum\limits_{i=1}^{d}\beta_i\eta\Bigr\} \notag\\ \le& 0. \notag \end{align} \eqref{t2-condition-e} is satisfied because $$ \lim_{\theta_i\to0}\ffrac{\partial g}{\partial\theta_i}(\theta)\theta_i^{2\beta_i-1}=\lim_{\theta_i\to0}-2\alpha\theta_i^{2\beta_i}(\sum\limits_{j=1}^{d}\theta_j^2+\eta)^{-\alpha-1}\le0.$$ \par Using the statements above, we know that $g(\theta)=(\sum\limits_{i=1}^d\theta_i^2+\eta)^{-\alpha}$ satisfies the conditions of the second half of Theorem 2. Therefore, \eqref{Fineq3} holds when $g(\theta)=(\sum\limits_{i=1}^d\theta_i^2+\eta)^{-\alpha},\,\eta>0,\, 0<\alpha\le\sum\limits_{i=1}^d\beta_i-1$. \par Next, we prove that \eqref{Fineq3} still holds when $f(\lambda)=g(\theta)=(\sum\limits_{i=1}^d\theta_i^2)^{-\alpha}$. The key is to consider $\eta\to0$. Let $\zeta:=\min\{\gamma_1r_1,\dots,\gamma_dr_d\}$ and $\bar\lambda:=\theta^2$. Because $-\alpha+\sum\limits_{j=1}^{d}(\beta_j+x_j)\ge1$, we have \begin{align} &\int(\sum\limits_{j=1}^d\theta_j^2)^{-\alpha}\prod\limits_{j=1}^d\ffrac{r_j^{x_j+\beta_j}\lambda_j^{x_j+\beta_j-1}\exp(-r_j\lambda_j)}{\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\lambda \notag\\ &= 2^{d} \int (\sum\limits_{j=1}^{d}\theta_j^2)^{-\alpha}\prod\limits_{j=1}^{d}\ffrac{\theta_j^{2\beta_j-1+2x_j}(\gamma_jr_j)^{x_j+\beta_j}}{\Gamma(x_j+\beta_j)}e^{-\gamma_jr_j\theta_j^2}\ensuremath{\mathrm{d}}\theta \notag\\ &= \int (\sum\limits_{j=1}^{d}\bar\lambda_j)^{-\alpha}\prod\limits_{j=1}^{d}\ffrac{\bar\lambda_j^{\beta_j-1+x_j}(\gamma_jr_j)^{x_j+\beta_j}}{\Gamma(x_j+\beta_j)}e^{-\gamma_jr_j\bar\lambda_j}\ensuremath{\mathrm{d}}\bar\lambda \notag\\ &\le \int (\sum\limits_{j=1}^{d}\bar\lambda_j)^{-\alpha}\prod\limits_{j=1}^{d}\ffrac{\bar\lambda_j^{\beta_j-1+x_j}(\gamma_jr_j)^{x_j+\beta_j}}{\Gamma(x_j+\beta_j)}e^{-\zeta\bar\lambda_j}\ensuremath{\mathrm{d}}\bar\lambda \notag\\ &=\prod\limits_{j=1}^{d} \ffrac{(\gamma_jr_j)^{x_j+\beta_j}}{\zeta^{x_j+\beta_j}}\int (\sum\limits_{j=1}^{d}\bar\lambda_j)^{-\alpha}\prod\limits_{j=1}^{d}\ffrac{\bar\lambda_j^{\beta_j-1+x_j}\zeta^{x_j+\beta_j}}{\Gamma(x_j+\beta_j)}e^{-\zeta\bar\lambda_j}\ensuremath{\mathrm{d}}\bar\lambda \notag\\ &=\prod\limits_{j=1}^{d} \ffrac{(\gamma_jr_j)^{x_j+\beta_j}}{\zeta^{x_j+\beta_j}}\text{E}(\Lambda^{-\alpha}\mid\Lambda\sim\Gamma(\sum\limits_{j=1}^{d}(\beta_j+x_j),\zeta)) \label{t3-finite-1}\\ &<\infty. \notag \end{align} Therefore, using the dominated convergence theorem, we have $$\int (\sum\limits_{i=1}^d\theta_i^2+\eta)^{-\alpha}\prod\limits_{i=1}^d\ffrac{r_i^{x_i+\beta_i}\lambda_i^{x_i+\beta_i-1}\exp(-r_i\lambda_i)}{\Gamma(x_i+\beta_i)}\ensuremath{\mathrm{d}}\lambda \to \int (\sum\limits_{i=1}^d\theta_i^2)^{-\alpha}\prod\limits_{i=1}^d\ffrac{r_i^{x_i+\beta_i}\lambda_i^{x_i+\beta_i-1}\exp(-r_i\lambda_i)}{\Gamma(x_i+\beta_i)}\ensuremath{\mathrm{d}}\lambda$$ when $\eta\to0$. From the definition of $F$ in \eqref{Fineq3}, we know that \eqref{Fineq3} still holds when $f(\lambda)=g(\theta)=(\sum\limits_{i=1}^d\theta_i^2)^{-\alpha}$. \vspace{0.3cm} \noindent \textbf{Part 2}. $F(x,r)$ is not a constant function of $x$. $F(x,r)$ is not a constant function of $x$ because when $x_1\to\infty$, \begin{align} & F(x,r) \notag\\ &= \int (\lambda_1/\gamma_1+\lambda_2/\gamma_2+\cdots+\lambda_d/\gamma_d+\eta)^{-\alpha}\prod\limits_{i=1}^d\ffrac{\lambda_i^{x_i+\beta_i-1}r_i^{x_i+\beta_i}}{\Gamma(x_i+\beta_i)}e^{-r_i\lambda_i} \ensuremath{\mathrm{d}}\lambda \notag\\ &< \int (\lambda_1/\gamma_1)^{-\alpha}\prod\limits_{i=1}^d\ffrac{\lambda_i^{x_i+\beta_i-1}r_i^{x_i+\beta_i}}{\Gamma(x_i+\beta_i)}e^{-r_i\lambda_i} \ensuremath{\mathrm{d}}\lambda \notag\\ &= \int (\lambda_1/\gamma_1)^{-\alpha} \ffrac{\lambda_1^{x_1+\beta_1-1}r_1^{x_1+\beta_1}}{\Gamma(x_1+\beta_1)}e^{-r_1\lambda_1}\ensuremath{\mathrm{d}}\lambda_1\notag\\ &= \int \gamma_1^{\alpha} \ffrac{\lambda_1^{x_1+\beta_1-1-\alpha}r_1^{x_1+\beta_1}}{\Gamma(x_1+\beta_1)}e^{-r_1\lambda_1}\ensuremath{\mathrm{d}}\lambda_1\notag\\ &= \gamma_1^{\alpha}r_1^{\alpha} \ffrac{\Gamma(x_1+\beta_1-\alpha)}{\Gamma(x_1+\beta_1)}\to 0. \notag \end{align} \vspace{0.3cm} \noindent \textbf{Part 3}. Condition \eqref{F-differentiable-3} is satisfied. Let $\zeta:=\min\{\gamma_1r_1,\dots,\gamma_dr_d\}$ and $\kappa:=\max\{\gamma_1r_1,\dots,\gamma_dr_d\}$. Using \eqref{t3-finite-1} and $\sum\limits_{j=1}^{d}(\beta_j+x_j)-\alpha\ge1$, we have \begin{align} &F(x,r) \notag\\ &\le\prod\limits_{j=1}^{d} \ffrac{(\gamma_jr_j)^{x_j+\beta_j}}{\zeta^{x_j+\beta_j}}\text{E}(\Lambda^{-\alpha}\mid\Lambda\sim\Gamma(\sum\limits_{j=1}^{d}(\beta_j+x_j),\zeta)) \notag \\ &=\prod\limits_{j=1}^{d} \ffrac{(\gamma_jr_j)^{x_j+\beta_j}}{\zeta^{x_j+\beta_j}}\zeta^{\alpha}\ffrac{\Gamma(\sum_{j=1}^{d}(\beta_j+x_j)-\alpha)}{\Gamma(\sum_{j=1}^{d}(\beta_j+x_j))} \notag \\ &\le 2\zeta^{\alpha-\sum_{j=1}^{d}(\beta_j+x_j)}\prod\limits_{j=1}^{d} (\gamma_jr_j)^{x_j+\beta_j} \label{t3-F-upper} \end{align} and \begin{align} &F(x,r)\notag\\ &=\int(\sum\limits_{j=1}^d\theta_j^2+\eta)^{-\alpha}\prod\limits_{j=1}^d\ffrac{r_j^{x_j+\beta_j}\lambda_j^{x_j+\beta_j-1}\exp(-r_j\lambda_j)}{\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\lambda \notag\\ &= 2^{d} \int (\sum\limits_{j=1}^{d}\theta_j^2+\eta)^{-\alpha}\prod\limits_{j=1}^{d}\ffrac{\theta_j^{2\beta_j-1+2x_j}(\gamma_jr_j)^{x_j+\beta_j}}{\Gamma(x_j+\beta_j)}e^{-\gamma_jr_j\theta_j^2}\ensuremath{\mathrm{d}}\theta \notag\\ &= \int (\sum\limits_{j=1}^{d}\bar\lambda_j+\eta)^{-\alpha}\prod\limits_{j=1}^{d}\ffrac{\bar\lambda_j^{\beta_j-1+x_j}(\gamma_jr_j)^{x_j+\beta_j}}{\Gamma(x_j+\beta_j)}e^{-\gamma_jr_j\bar\lambda_j}\ensuremath{\mathrm{d}}\bar\lambda \notag\\ &\ge \int (\sum\limits_{j=1}^{d}\bar\lambda_j+\eta)^{-\alpha}\prod\limits_{j=1}^{d}\ffrac{\bar\lambda_j^{\beta_j-1+x_j}(\gamma_jr_j)^{x_j+\beta_j}}{\Gamma(x_j+\beta_j)}e^{-\kappa\bar\lambda_j}\ensuremath{\mathrm{d}}\bar\lambda \notag \\ &=\prod\limits_{j=1}^{d} \ffrac{(\gamma_jr_j)^{x_j+\beta_j}}{\kappa^{x_j+\beta_j}}\text{E}((\Lambda+\eta)^{-\alpha}\mid\Lambda\sim\Gamma(\sum\limits_{j=1}^{d}(\beta_j+x_j),\kappa)) \notag \end{align} It is simple to show that there exists a positive value $C_1$ such that for every $x$, $$\text{E}((\Lambda+\eta)^{-\alpha}\mid\Lambda\sim\Gamma(\sum\limits_{j=1}^{d}(\beta_j+x_j),\kappa))\ge C_1(\sum\limits_{j=1}^{d}(\beta_j+x_j))^{-\alpha-1}.$$ Thus, \begin{equation} F(x,r)\ge \prod\limits_{j=1}^{d} \ffrac{(\gamma_jr_j)^{x_j+\beta_j}}{\kappa^{x_j+\beta_j}}C_1(\sum\limits_{j=1}^{d}(\beta_j+x_j))^{-\alpha-1}. \label{t3-F-lower} \end{equation} \noindent Using \eqref{t3-F-upper} and \eqref{t3-F-lower}, for a given $r$, there exists a constant $C$ such that $$\max_{t\in \bar B(r,\epsilon)}\Bigl| \log F(x,t) \Bigr|\le C\sum_{j=1}^{d}x_j+C,\,\forall x,$$ where $\epsilon=\min\{r_1,\dots,r_d\}/2.$ Thus, \begin{align} &\sum_x\Biggl[ \max_{t\in \bar B(r,\epsilon)}\biggl\{ \Bigl| \log F(x,t) \Bigr| (\sum_{i=1}^dx_i+1) \prod_{i=1}^d \frac{(t_i\lambda_i)^{x_i}\exp(-t_i\lambda_i)}{x_i!} \biggr\} \Biggr] \notag\\ &\le\sum_x\biggl[ (C\sum_{j=1}^{d}x_j+C)(\sum_{i=1}^dx_i+1) \prod_{i=1}^d \frac{(2r_i\lambda_i)^{x_i}}{x_i!}\biggr] \notag \\ &=C\exp(2\sum_{i=1}^dr_i\lambda_i)\sum_x\biggl[ (\sum_{i=1}^dx_i+1)^2 \prod_{i=1}^d \frac{(2r_i\lambda_i)^{x_i}\exp(-2r_i\lambda_i)}{x_i!}\biggr] \notag\\ &=C\exp(2\sum_{i=1}^dr_i\lambda_i)\text{E}(\sum_{i=1}^dx_i+1)^2\mid x_i\sim \text{Po}(2r_i\lambda_i)) <\infty. \notag \end{align} Therefore, condition \eqref{F-differentiable-3} is satisfied. \par Therefore, $f(\lambda)=g(\theta)=(\sum\limits_{i=1}^d\theta_i^2+\eta)^{-\alpha},\ \eta\ge0,\ 0<\alpha\le\sum\limits_{i=1}^d\beta_i-1$ satisfies the conditions in the first half of Theorem 2. Thus, the Bayesian predictive distribution $p_{f,\beta}(y\mid x)$ dominates the Bayesian predictive distribution $p_{\beta}(y\mid x)$, for prediction of independent Poisson processes with the same or different durations. \qed \vspace{0.5cm} \noindent \textbf{Example 2}. We consider the subspace shrinkage prior for prediction of independent Poisson processes with the same duration. We denote $\theta_i:=\sqrt{\lambda_i},\ i=1,\dots,d$. In Example 1, the Bayesian predictions based on the prior with $f(\lambda)=g(\theta)=(\sum\limits_{i=1}^d\theta_i^2+\eta)^{-\alpha}$ shrink $\theta$ towards $\vec{0}$. Therefore, it is natural to investigate the subspace shrinkage prior $f(\lambda)=g(\theta)=(s_V(\theta))^{-\alpha}$, where $s_V(\theta)$ is the squared distance from $\theta$ to a linear subspace $V\subset \mathbb{R}^d$. \par We assume complementary space $V^\perp=\text{span}(v_1,v_2,\dots,v_{d-k})$, where $\{v_1,v_2,\dots,v_{d-k}\}$ is a standard orthonormal basis and $k$ is the dimension of $V$. We have $$s_V(\theta)=\sum\limits_{i=1}^{d-k}\langle \theta,v_i \rangle ^2.$$ Using Corollary 1, we have\\ \noindent \textbf{Proposition 2.} When $f(\lambda)=g(\theta)=(s_V(\theta))^{-\alpha}$, $0<\alpha\le (d-k-2)/2$, and if $\sum_{i=1}^{d-k}v_{i,j_1}v_{i,j_2}\ge0$ for every $1\le j_1,j_2\le d$, then the Bayesian predictive density based on $\pi_{f,\beta=(1/2,\dots,1/2)}(\lambda)$ dominates that based on the Jeffreys prior. \noindent \textbf{Proof.} We need only prove that $f(\lambda)=g(\theta)=(s_V(\theta))^{-\alpha}$ satisfies the conditions in the first half of Corollary 1. We show this in three parts. \vspace{0.3cm} \noindent \textbf{Part 1}. Condition \eqref{Fineq2} is satisfied. We construct a $\mathbf{C}^2$ prior $$f_{\epsilon}(\lambda)=g_{\epsilon}(\theta)=(s_V(\theta)+\epsilon)^{-\alpha},\ 0<\epsilon<1.$$ We first show that $g_{\epsilon}(\theta)$ satisfies the conditions of the second half of Corollary 1. \par First, we have \begin{align} \int g_{\epsilon}(\theta)p(x\mid\theta)\ensuremath{\mathrm{d}}\theta&\le\int \epsilon^{-\alpha}p(x\mid\theta)\ensuremath{\mathrm{d}}\theta \notag\\ &=2^{-d}\epsilon^{-\alpha}\prod\limits_{i=1}^{d}\int\ffrac{e^{-r\lambda_i}r^{x_i}\lambda_i^{x_i-1/2}}{x_i!}\ensuremath{\mathrm{d}}\lambda_i \notag\\ &=2^{-d}\epsilon^{-\alpha}\prod\limits_{i=1}^{d}r^{-1/2}\ffrac{\Gamma(x_i+1/2)}{x_i!}<\infty, \notag \end{align} and \begin{align} &\int \Bigl\lvert \ffrac{\partial g_{\epsilon}}{\partial\theta_i}(\theta) \Bigr\rvert p(x\mid\theta)\ensuremath{\mathrm{d}}\theta\notag\\ &= \int (\sum\limits_{j=1}^{d-k}\langle \theta,v_j \rangle^2+\epsilon)^{-\alpha-1}\alpha\Bigl\lvert\sum\limits_{j=1}^{d-k}2v_{ji}\langle \theta,v_j \rangle\Bigr\rvert p(x\mid\theta)\ensuremath{\mathrm{d}}\theta \notag\\ &\le \epsilon^{-\alpha-1}\alpha\int2(d-k)(\sum\limits_{j=1}^{d}\theta_j)p(x\mid\theta)\ensuremath{\mathrm{d}}\theta \notag\\ &= 2\epsilon^{-\alpha-1}\alpha(d-k)\sum\limits_{j=1}^{d}\int \theta_j p(x\mid\theta)\ensuremath{\mathrm{d}}\theta \notag\\ &= 2\epsilon^{-\alpha-1}\alpha(d-k)\sum\limits_{j=1}^{d}2^{-d}r^{-(d+1)/2}\prod\limits_{h=1}^{d}\ffrac{\Gamma(x_h+1/2+\delta_{hj}/2)}{x_h!}<\infty, \notag \end{align} and \begin{align} &\int \Bigl\lvert \ffrac{\partial^2 g_{\epsilon}}{\partial\theta_i^2}(\theta) \Bigr\rvert p(x\mid\theta)\ensuremath{\mathrm{d}}\theta\notag \\ &\le \int \biggl\{ (\sum\limits_{j=1}^{d-k}\langle \theta,v_j \rangle^2+\epsilon)^{-\alpha-2}\alpha(\alpha+1)(\sum\limits_{j=1}^{d-k}2v_{ji}\langle \theta,v_j \rangle)^2 +(\sum\limits_{j=1}^{d-k}\langle \theta,v_j \rangle^2+\epsilon)^{-\alpha-1}\alpha\sum\limits_{j=1}^{d-k}2v_{ji}^2 \biggr\} p(x\mid\theta)\ensuremath{\mathrm{d}}\theta \notag\\ &\le \epsilon^{-\alpha-2}\alpha(\alpha+1)4(d-k)^2\int(\sum\limits_{j=1}^{d}\theta_j^2)p(x\mid\theta)\ensuremath{\mathrm{d}}\theta+\epsilon^{-\alpha-1}\alpha2(d-k)\int p(x\mid\theta)\ensuremath{\mathrm{d}}\theta \notag\\ &= \epsilon^{-\alpha-2}\alpha(\alpha+1)4(d-k)^2\sum\limits_{j=1}^{d}\int\theta_j^2p(x\mid\theta)\ensuremath{\mathrm{d}}\theta+\epsilon^{-\alpha-1}\alpha2(d-k)\int p(x\mid\theta)\ensuremath{\mathrm{d}}\theta \notag\\ &= 4\epsilon^{-\alpha-2}\alpha(\alpha+1)(d-k)^2\sum\limits_{j=1}^{d}2^{-d}r^{-(d+2)/2}\prod\limits_{h=1}^{d}\ffrac{\Gamma(x_h+1/2+\delta_{hj})}{x_h!} \notag\\ & \qquad +2^{-d+1}\epsilon^{-\alpha-1}\alpha(d-k)\prod\limits_{h=1}^{d}r^{-1/2}\ffrac{\Gamma(x_h+1/2)}{x_h!}<\infty. \notag \end{align} \par Second, $g_{\epsilon}(\theta)$ is a superharmonic function because \begin{align} &\sum\limits_{i=1}^{d} \ffrac{\partial^2 g_{\epsilon}}{\partial\theta_i^2}(\theta)\notag \\ &= \sum\limits_{i=1}^{d}\biggl\{ (\sum\limits_{j=1}^{d-k}\langle \theta,v_j \rangle^2+\epsilon)^{-\alpha-2}\alpha(\alpha+1)(\sum\limits_{j=1}^{d-k}2v_{ji}\langle \theta,v_j \rangle)^2 -(\sum\limits_{j=1}^{d-k}\langle \theta,v_j \rangle^2+\epsilon)^{-\alpha-1}\alpha\sum\limits_{j=1}^{d-k}2v_{ji}^2 \biggr\} \notag\\ &= \alpha(\sum\limits_{j=1}^{d-k}\langle \theta,v_j \rangle^2+\epsilon)^{-\alpha-2}\biggl\{ 4(\alpha+1)\sum\limits_{j=1}^{d-k}\langle \theta,v_j \rangle^2-2(d-k)(\sum\limits_{j=1}^{d-k}\langle \theta,v_j \rangle^2+\epsilon)\biggr\} \notag\\ &= \alpha(\sum\limits_{j=1}^{d-k}\langle \theta,v_j \rangle^2+\epsilon)^{-\alpha-2}\biggl\{ (4(\alpha+1)-2(d-k))\sum\limits_{j=1}^{d-k}\langle \theta,v_j \rangle^2-2(d-k)\epsilon\biggr\} \notag\\ &<0. \notag \end{align} \par Third, we have \begin{align*} &\lim_{\theta_i\to0}\ffrac{\partial}{\partial\theta_i} g_{\epsilon}(\theta) \\ &=\lim_{\theta_i\to0}-(\sum\limits_{j=1}^{d-k}\langle \theta,v_j \rangle^2+\epsilon)^{-\alpha-1}\alpha\biggl\{ \sum\limits_{j=1}^{d-k}2v_{ji}\langle \theta,v_j \rangle\biggl\} \\ &=\lim_{\theta_i\to0}-(\sum\limits_{j=1}^{d-k}\langle \theta,v_j \rangle^2+\epsilon)^{-\alpha-1}\alpha\biggl\{ \sum\limits_{h=1}^{d}\theta_h( \sum\limits_{j=1}^{d-k}2v_{ji}v_{jh}) \biggl\}\\ &\le0. \end{align*} \par Therefore, using the second half of Corollary 1, we know that \eqref{Fineq2} holds when $f(\lambda)=g(\theta)=(s_V(\theta)+\epsilon)^{-\alpha}$. Next, we prove that \eqref{Fineq2} still holds when $f(\lambda)=g(\theta)=(s_V(\theta))^{-\alpha}$. The key is to consider $\epsilon\to0$. \par Because the rank of matrix $\{v_1,v_2,\dots,v_{d-k}\}$ is $d-k$, there exist $1\le j_1<\cdots<j_{d-k}\le d$ such that $\sum_{i=1}^{d-k}v_{i,j_{t}}^2>0,\ t=1,2,\dots,d-k.$ Without loss of generality, we assume $j_t=t$, and $\sum_{i=1}^{d-k}v_{it}^2\ge\zeta>0, \ t=1,2,\dots,d-k.$ Thus, we have \begin{align*} \sum\limits_{i=1}^{d-k}\langle \theta,v_i \rangle^2 &=\sum\limits_{i=1}^{d-k}(\sum\limits_{j=1}^{d}\theta_j^2v_{ij}^2+2\sum\limits_{j_1<j_2}\theta_{j_1}\theta_{j_2}v_{ij_1}v_{ij_2}) \\ &=\sum\limits_{j=1}^{d}\theta_j^2(\sum\limits_{i=1}^{d-k}v_{ij}^2)+\sum\limits_{j_1<j_2}2\theta_{j_1}\theta_{j_2}(\sum\limits_{i=1}^{d-k}v_{ij_1}v_{ij_2}) \\ &\ge \sum\limits_{t=1}^{d-k}\theta_t^2\zeta . \end{align*} Therefore, we have \begin{align} &\int(s_V(\theta))^{-\alpha}\prod\limits_{i=1}^d\ffrac{r^{x_i+1/2}\lambda_i^{x_i-1/2}\exp(-r\lambda_i)}{\Gamma(x_i+1/2)}\ensuremath{\mathrm{d}}\lambda \notag\\ &= \int(\sum\limits_{i=1}^{d-k}\langle \theta,v_i \rangle^2)^{-\alpha}\prod\limits_{i=1}^d\ffrac{r^{x_i+1/2}\lambda_i^{x_i-1/2}\exp(-r\lambda_i)}{\Gamma(x_i+1/2)}\ensuremath{\mathrm{d}}\lambda \notag\\ &\le \int(\sum\limits_{t=1}^{d-k}\theta_t^2\zeta)^{-\alpha}\prod\limits_{i=1}^d\ffrac{r^{x_i+1/2}\lambda_i^{x_i-1/2}\exp(-r\lambda_i)}{\Gamma(x_i+1/2)}\ensuremath{\mathrm{d}}\lambda \notag\\ &=\zeta^{-\alpha}\text{E}\bigl( (\sum_{i=1}^{d-k}\lambda_i)^{-\alpha}\mid \lambda_i\sim\Gamma(x_i+1/2,r) \bigl) \notag\\ &=\zeta^{-\alpha}\text{E}\bigl( \lambda^{-\alpha}\mid \lambda\sim\Gamma(\sum_{i=1}^{d-k} x_i+(d-k)/2,r) \bigl) \notag\\ &=\zeta^{-\alpha}r^{\alpha}\ffrac{\Gamma(\sum_{i=1}^{d-k} x_i+(d-k)/2-\alpha)}{\Gamma(\sum_{i=1}^{d-k} x_i+(d-k)/2)} <\infty. \label{p2-Fvalue} \end{align} \par Therefore, using the dominated convergence theorem, for every $x$ and $r>0$, $$\int (s_V(\theta)+\epsilon)^{-\alpha}\prod\limits_{i=1}^d\ffrac{r^{x_i+1/2}\lambda_i^{x_i-1/2}\exp(-r\lambda_i)}{\Gamma(x_i+1/2)}\ensuremath{\mathrm{d}}\lambda \to \int (s_V(\theta))^{-\alpha}\prod\limits_{i=1}^d\ffrac{r^{x_i+1/2}\lambda_i^{x_i-1/2}\exp(-r\lambda_i)}{\Gamma(x_i+1/2)}\ensuremath{\mathrm{d}}\lambda$$ when $\epsilon\to0$. From the definition of $F$ in \eqref{Fineq2}, \eqref{Fineq2} still holds when $f(\lambda)=g(\theta)=(s_V(\theta))^{-\alpha}$. \vspace{0.3cm} \noindent \textbf{Part 2}. Condition \eqref{F-differentiable-2} is satisfied. From \eqref{p2-Fvalue}, we have \begin{align} F(x,r)<2\zeta^{-\alpha}r^{\alpha}. \label{p2-Fvalue-upper} \end{align} We denote $\sum_{i=1}^{d-k}\Vert v_i \Vert^2=:\kappa$. Then, \begin{align} &F(x,r) \notag\\ &= \int(\sum\limits_{i=1}^{d-k}\langle \theta,v_i \rangle^2)^{-\alpha}\prod\limits_{i=1}^d\ffrac{r^{x_i+1/2}\lambda_i^{x_i-1/2}\exp(-r\lambda_i)}{\Gamma(x_i+1/2)}\ensuremath{\mathrm{d}}\lambda \notag\\ &\ge \int(\sum\limits_{t=1}^{d}\theta_t^2\kappa)^{-\alpha}\prod\limits_{i=1}^d\ffrac{r^{x_i+1/2}\lambda_i^{x_i-1/2}\exp(-r\lambda_i)}{\Gamma(x_i+1/2)}\ensuremath{\mathrm{d}}\lambda \notag\\ &=\kappa^{-\alpha}\text{E}\bigl( (\sum_{i=1}^d\lambda_i)^{-\alpha}\mid \lambda_i\sim\Gamma(x_i+1/2,r) \bigl) \notag\\ &=\kappa^{-\alpha}\text{E}\bigl( \lambda^{-\alpha}\mid \lambda\sim\Gamma(\sum_{i=1}^d x_i+d/2,r) \bigl) \notag\\ &=\kappa^{-\alpha}r^{\alpha}\ffrac{\Gamma(\sum_{i=1}^d x_i+d/2-\alpha)}{\Gamma(\sum_{i=1}^d x_i+d/2)} \ge \kappa^{-\alpha}r^{\alpha}\frac{1}{2}(\sum_{i=1}^d x_i+d/2)^{-\lceil\alpha\rceil}. \label{p2-Fvalue-lower} \end{align} Using \eqref{p2-Fvalue-upper} and \eqref{p2-Fvalue-lower}, for a given $r$, there exists a constant $C$ such that $$\max_{t\in [r/2,3r/2]}\Bigl| \log F(x,t) \Bigr|\le C\sum_{j=1}^{d}x_j+C,\,\forall x.$$ Thus, \begin{align} &\sum_x\Biggl[ \max_{t\in[r/2,3r/2]}\biggl\{ \bigl| \log F(x,t) \bigr| (\sum_{i=1}^dx_i+1) \prod_{i=1}^d \frac{(t\lambda_i)^{x_i}\exp(-t\lambda_i)}{x_i!} \biggr\} \Biggr] \notag\\ &\le\sum_x\biggl[ (C\sum_{j=1}^{d}x_j+C)(\sum_{i=1}^dx_i+1) \prod_{i=1}^d \frac{(2r\lambda_i)^{x_i}}{x_i!} \biggr] \notag\\ &=C\exp(2r\sum_{i=1}^d\lambda_i)\sum_x\biggl[ (\sum_{i=1}^dx_i+1)^2 \prod_{i=1}^d \frac{(2r\lambda_i)^{x_i}\exp(-2r\lambda_i)}{x_i!} \biggr] \notag\\ &=C\exp(2r\sum_{i=1}^d\lambda_i)\text{E}\bigl( (\sum_{i=1}^dx_i+1)^2 \mid x_i\sim\text{Po}(2r\lambda_i) \bigr) <\infty. \notag \end{align} Therefore, condition \eqref{F-differentiable-2} is satisfied. \vspace{0.3cm} \noindent \textbf{Part 3}. $F(x,r)$ is not a constant function of $x$. From \eqref{p2-Fvalue}, we notice that $F(x,r)\to0$ when $x_1\to\infty$. Therefore, $F$ is not a constant function of $x$. Using the first half of Corollary 1, we know that the Bayesian predictive density based on $\pi_{f,\beta=(1/2,\dots,1/2)}(\lambda)$ dominates that based on the Jeffreys prior. \qed \section{Numerical experiments} \par We use numerical experiments to show the difference between the risk of the Bayesian predictive density based on the Jeffreys prior and that of the Bayesian predictive density based on the priors in Section 4.\\ \noindent \textbf{Experiment 1.} We set $r=s=1$, $d=3$ and $\theta_i=\sqrt{\lambda_i},\ i=1,\dots,3$. The first prior is a point shrinkage prior with $$f(\lambda)=g(\theta)=(\sum\limits_{i=1}^{3}\theta_i^2)^{-(3-2)/2}.$$ The second prior is called the ``shift point shrinkage prior'' with $$f(\lambda)=g(\theta)=(\sum\limits_{i=1}^{3}\theta_i^2+1)^{-(3-2)/2}.$$ For each $\lambda$, we sampled $x$ and $y$ $10^5$ times and approximated the risk difference by the sample mean of $\displaystyle\log\ffrac{p_{f,\beta=(1/2,\dots,1/2)}(y\mid x)}{p_{J}(y\mid x)}$. Figure 1 shows the difference between the risks of the Bayesian predictive densities based on the two priors and the Jeffreys prior when $\lambda=\mu\times(1/3,1/3,1/3)$. When $\mu$ is small, the Bayesian predictive density based on the point shrinkage prior performs better. When $\mu$ is large, the Bayesian predictive density based on the shift point shrinkage prior performs better. \begin{figure}[H] \centering \includegraphics[width=0.6\textwidth]{point_shrinkage.pdf} \caption{Log value of the difference between the expected divergences, $\log\text{E}\Bigl[ D(p(y\mid\lambda),p_{J}(y\mid x))-D(p(y\mid\lambda),p_{f,\beta=(1/2,\dots,1/2)}(y\mid x))\,\Big|\,\lambda\Bigr] .$} \end{figure} \noindent \textbf{Experiment 2.} We set $r=s=1$, $d=4$ and $\theta_i=\sqrt{\lambda_i},\ i=1,\dots,4$. The first prior is a point shrinkage prior with $$f(\lambda)=g(\theta)=(\sum\limits_{i=1}^{4}\theta_i^2)^{-(4-2)/2}.$$ The second prior is called ``subspace shrinkage prior 1'' with $$f(\lambda)=g(\theta)=(\theta_1^2+\theta_2^2+\theta_3^2)^{-(3-2)/2},$$ which shrinks $\theta$ towards subspace $V=\text{span}((0,0,0,1))$. The third prior is called ``subspace shrinkage prior 2'' with $$f(\lambda)=g(\theta)=(\theta_1^2+\theta_2^2+\theta_4^2)^{-(3-2)/2},$$ which shrinks $\theta$ towards subspace $V=\text{span}((0,0,1,0))$. The fourth prior is called ``subspace shrinkage prior 3'' with $$f(\lambda)=g(\theta)=(\theta_1^2+\theta_2^2+(\theta_3+\theta_4)^2/2)^{-(3-2)/2},$$ which shrinks $\theta$ towards subspace $V=\text{span}((0,0,1,-1))$. For each $\lambda$, we sampled $x$ and $y$ $10^5$ times and approximated the risk difference by the sample mean of $\displaystyle\log\ffrac{p_{f,\beta=(1/2,\dots,1/2)}(y\mid x)}{p_{J}(y\mid x)}$. Figure 2 shows the difference between the risks of the Bayesian predictive densities based on the four priors and the Jeffreys prior when $\lambda=\mu\times(1,1,1,100)/20$. When $\mu$ is small, $\lambda$ is close to $\vec{0}$, and the Bayesian predictive densities based on point shrinkage prior and subspace shrinkage priors both perform well. Because $\theta_4$ is much larger than the others, the Bayesian predictive density based on subspace shrinkage prior 1 performs far better than the others, and the Bayesian predictive density based on subspace shrinkage prior 3 is the second best. \begin{figure}[H] \centering \includegraphics[width=0.6\textwidth]{subspace_shrinkage.pdf} \caption{Log value of the difference between the expected divergences, $\log\text{E}\Bigl[ D(p(y\mid\lambda),p_{J}(y\mid x))-D(p(y\mid\lambda),p_{f,\beta=(1/2,\dots,1/2)}(y\mid x))\,\Big|\,\lambda\Bigr] .$} \end{figure} The last example shows that the condition $\displaystyle\left.\ffrac{\partial g}{\partial\theta_i}(\theta)\right|_{\theta_i=0}\le0$ in Corollary 1 is important. We set $r=s=1$, $\lambda=(0.1,0.1,0.1)$. When $f(\lambda)=g(\theta)=(\sum_{i=1}^3(\theta_i-1)^2+0.01)^{-1/2}$, $$\text{E}\Bigl[ D(p(y\mid\lambda),p_{J}(y\mid x))-D(p(y\mid\lambda),p_{f,\beta=(1/2,\dots,1/2)}(y\mid x))\,\Big|\,\lambda\Bigr] =-0.09.$$ However, $g$ satisfies all conditions in Corollary 1 except $\displaystyle\left.\ffrac{\partial g}{\partial\theta_i}(\theta)\right|_{\theta_i=0}\le0$. \section*{Appendix} \noindent \textbf{Lemma 1.1} If the function $F(x,r)$ defined in Theorem 1 is finite, then for any given $r$ and $x$, there exists $\epsilon>0$ such that $$\int \max_{t\in[r-\epsilon,r+\epsilon]}\Biggl\{ f(\bar\lambda_1,\dots,\bar\lambda_d) \prod_{i=1}^d \frac{t^{x_i+\beta_i}\bar\lambda_i^{x_i+\beta_i-1}\exp(-t\bar\lambda_i)}{\Gamma(x_i+\beta_i)} \Bigl| \sum_{i=1}^d \frac{x_i+\beta_i}{t} - \sum_{i=1}^d \bar\lambda_i \Bigr| \Biggl\} \ensuremath{\mathrm{d}}\bar\lambda<\infty.$$ \noindent \textbf{Proof}\\ We choose $\epsilon=r/2$.\\ Then we have \begin{align*} &\int \max_{t\in[r-\epsilon,r+\epsilon]}\Biggl\{ f(\bar\lambda_1,\dots,\bar\lambda_d) \prod_{i=1}^d \frac{t^{x_i+\beta_i}\bar\lambda_i^{x_i+\beta_i-1}\exp(-t\bar\lambda_i)}{\Gamma(x_i+\beta_i)} \Bigl| \sum_{i=1}^d \frac{x_i+\beta_i}{t} - \sum_{i=1}^d \bar\lambda_i \Bigr| \Biggl\} \ensuremath{\mathrm{d}}\bar\lambda \\ &\le\int \max_{t\in[r/2,2r]} \Biggl\{ f(\bar\lambda_1,\dots,\bar\lambda_d) \prod_{i=1}^d \frac{t^{x_i+\beta_i}\bar\lambda_i^{x_i+\beta_i-1}\exp(-t\bar\lambda_i)}{\Gamma(x_i+\beta_i)} \Bigl| \sum_{i=1}^d \frac{x_i+\beta_i}{t} - \sum_{i=1}^d \bar\lambda_i \Bigr| \Biggl\} \ensuremath{\mathrm{d}}\bar\lambda\\ &\le\int f(\bar\lambda_1,\dots,\bar\lambda_d) \prod_{i=1}^d \ffrac{(2r)^{x_i+\beta_i}\bar\lambda_i^{x_i+\beta_i-1}\exp(-\ffrac{r}{2}\bar\lambda_i)}{\Gamma(x_i+\beta_i)} \Bigl( \sum_{i=1}^d \frac{2x_i+2\beta_i}{r} + \sum_{i=1}^d \bar\lambda_i \Bigr) \ensuremath{\mathrm{d}}\bar\lambda \\ &=\frac{ \sum_{i=1}^d(x_i+\beta_i)}{r}2^{2\sum_{i=1}^d(x_i+\beta_i)+1}\int f(\bar\lambda_1,\dots,\bar\lambda_d) \prod_{i=1}^d \ffrac{(r/2)^{x_i+\beta_i}\bar\lambda_i^{x_i+\beta_i-1}\exp(-\ffrac{r}{2}\bar\lambda_i)}{\Gamma(x_i+\beta_i)} \ensuremath{\mathrm{d}}\bar\lambda\\ &~~~~+\sum_{j=1}^{d}\int f(\bar\lambda_1,\dots,\bar\lambda_d) \prod_{i=1}^d \ffrac{(2r)^{x_i+\beta_i}\bar\lambda_i^{x_i+\beta_i-1}\exp(-\ffrac{r}{2}\bar\lambda_i)}{\Gamma(x_i+\beta_i)} \bar\lambda_j \ensuremath{\mathrm{d}}\bar\lambda \\ &=\frac{ \sum_{i=1}^d(x_i+\beta_i)}{r}2^{2\sum_{i=1}^d(x_i+\beta_i)+1}\int f(\bar\lambda_1,\dots,\bar\lambda_d) \prod_{i=1}^d \ffrac{(r/2)^{x_i+\beta_i}\bar\lambda_i^{x_i+\beta_i-1}\exp(-\ffrac{r}{2}\bar\lambda_i)}{\Gamma(x_i+\beta_i)} \ensuremath{\mathrm{d}}\bar\lambda\\ &~~~~+\ffrac{1}{r}2^{2\sum_{i=1}^d(x_i+\beta_i)+1}\sum_{j=1}^{d}\int f(\bar\lambda_1,\dots,\bar\lambda_d) \prod_{i=1}^d \ffrac{(r/2)^{x_i+\beta_i+\delta_{ij}}\bar\lambda_i^{x_i+\beta_i+\delta_{ij}-1}\exp(-\ffrac{r}{2}\bar\lambda_i)}{\Gamma(x_i+\beta_i)} \ensuremath{\mathrm{d}}\bar\lambda\\ &=\frac{\sum_{i=1}^d(x_i+\beta_i)}{r}2^{2\sum_{i=1}^d(x_i+\beta_i)+1}F(x,r/2)+\ffrac{1}{r}2^{2\sum_{i=1}^d(x_i+\beta_i)+1}\sum_{j=1}^{d}F(x+\delta_j,r/2)(x_j+\beta_j). \end{align*} Because $F(x,r/2)$ and $F(x+\delta_j,r/2)$ are finite, the proof is complete. \qed \vspace{0.5cm} \noindent \textbf{Lemma 1.2} Under conditions \eqref{F-differentiable} and \eqref{Fineq}, for a given $r$, there exists $\epsilon>0$ such that \begin{align*} &\sum_x \max_{t\in[r-\epsilon,r+\epsilon]} \Biggl[ \biggl| \log(F(x,t)) \biggr| \Bigl\{ \prod_{i=1}^d \frac{(t\lambda_i)^{x_i}\exp(-t\lambda_i)}{x_i!} \Bigr\} \Bigl| \sum_{i=1}^d \frac{x_i}{t} - \sum_{i=1}^d \lambda_i \Bigr| \Biggr] \notag \\ &\quad + \sum_x \max_{t\in[r-\epsilon,r+\epsilon]} \Biggl[ \frac{ \displaystyle \Biggl| \int f(\bar\lambda_1,\dots,\bar\lambda_d) \prod_{i=1}^d \frac{t^{x_i+\beta_i}\bar\lambda_i^{x_i+\beta_i-1}\exp(-t\bar\lambda_i)}{\Gamma(x_i+\beta_i)} \Bigl( \sum_{i=1}^d \frac{x_i+\beta_i}{t} - \sum_{i=1}^d \bar\lambda_i \Bigr) \ensuremath{\mathrm{d}}\bar\lambda \Biggl| } { F(x,t)} \prod_{i=1}^d \ffrac{(t\lambda_i)^{x_i} \exp(-t\lambda_i)}{x_i!} \Biggr] \notag \\ &<\infty. \end{align*} \noindent \textbf{Proof}\\ Using \eqref{transform2}, we know that \begin{align*} \int f(\bar\lambda_1,\dots,\bar\lambda_d) \prod_{j=1}^d\ffrac{t^{x_j+\beta_j}\bar\lambda_j^{x_j+\beta_j-1} \exp(-t\bar\lambda_j)}{\Gamma(x_j+\beta_j)} \bar\lambda_i \ensuremath{\mathrm{d}}\bar\lambda \notag= F(x+\delta_i,t)\ffrac{x_i+\beta_i}{t}. \end{align*} Therefore, the Lemma is equivalent to \begin{align} &\sum_x \max_{t\in[r-\epsilon,r+\epsilon]} \Biggl[ \biggl| \log F(x,t) \biggr| \Bigl\{ \prod_{i=1}^d \frac{(t\lambda_i)^{x_i}\exp(-t\lambda_i)}{x_i!} \Bigr\} \Bigl| \sum_{i=1}^d \frac{x_i}{t} - \sum_{i=1}^d \lambda_i \Bigr| \Biggr] \notag \\ &\quad + \sum_x \max_{t\in[r-\epsilon,r+\epsilon]} \Biggl[ \frac{ \displaystyle \Biggl| F(x,t)\sum_{i=1}^d \frac{x_i+\beta_i}{t}-\sum_{i=1}^dF(x+\delta_i,t)\ffrac{x_i+\beta_i}{t} \Biggl| } {F(x,t)} \prod_{i=1}^d \ffrac{(t\lambda_i)^{x_i} \exp(-t\lambda_i)}{x_i!} \Biggr] \notag \\ &<\infty. \label{finite-goal} \end{align} Using \eqref{F-differentiable}, we have $\epsilon\in(0,r)$ such that $$ \sum_x \max_{t\in[r-\epsilon,r+\epsilon]}\Biggl[ \biggl| \log F(x,t) \biggr|(\sum_{i=1}^dx_i+1)\biggl\{ \prod_{i=1}^d \frac{(t\lambda_i)^{x_i}\exp(-t\lambda_i)}{x_i!} \biggr\} \Biggr] <\infty, $$ thus, \begin{align} &\sum_x \max_{t\in[r-\epsilon,r+\epsilon]}\Biggl[ \biggl| \log F(x,t) \biggr| \biggl\{ \prod_{i=1}^d \frac{(t\lambda_i)^{x_i}\exp(-t\lambda_i)}{x_i!} \biggr\} \Bigl| \sum_{i=1}^d \frac{x_i}{t} - \sum_{i=1}^d \lambda_i \Bigr|\Biggl] \notag \\ &\le\sum_x \max_{t\in[r-\epsilon,r+\epsilon]}\Biggl[ \biggl| \log F(x,t) \biggr| \biggl\{ \prod_{i=1}^d \frac{(t\lambda_i)^{x_i}\exp(-t\lambda_i)}{x_i!} \biggr\}\ffrac{ \sum_{i=1}^d x_i +1}{\min\{ r-\epsilon,1/\sum_{i=1}^d \lambda_i \}} \Biggl] <\infty. \label{L2-finite1} \end{align} Using \eqref{Fineq}, we have $$\sum\limits_{i=1}^dx_i(F(x,t)-F(x-\delta_i,t))+\sum\limits_{i=1}^d(x_i+\beta_i)(F(x,t)-F(x+\delta_i,t))\ge 0,$$ thus, $$\sum_{i=1}^d (x_i+\beta_i)F(x+\delta_i,t) <2\sum_{i=1}^d(x_i+\beta_i)F(x,t).$$ Therefore, \begin{align} & \sum_x \max_{t\in[r-\epsilon,r+\epsilon]} \Biggl[ \ffrac{ \Biggl| F(x,t)\sum_{i=1}^d \ffrac{x_i+\beta_i}{t}-\sum_{i=1}^dF(x+\delta_i,t)\ffrac{x_i+\beta_i}{t} \Biggl| } { F(x,t)} \prod_{i=1}^d \ffrac{(t\lambda_i)^{x_i} \exp(-t\lambda_i)}{x_i!} \Biggr] \notag \\ &< \sum_x \max_{t\in[r-\epsilon,r+\epsilon]} \Biggl[ \frac{ \displaystyle F(x,t)\sum_{i=1}^d \frac{x_i+\beta_i}{t}} { F(x,t)} \prod_{i=1}^d \ffrac{(t\lambda_i)^{x_i} \exp(-t\lambda_i)}{x_i!} \Biggr] \notag \\ &\le \ffrac{1}{r-\epsilon}\sum_x \max_{t\in[r-\epsilon,r+\epsilon]} \biggl\{ \sum_{i=1}^d (x_i+\beta_i) \prod_{i=1}^d \ffrac{(t\lambda_i)^{x_i} \exp(-t\lambda_i)}{x_i!} \biggl\} \notag \\ &\le \ffrac{\exp(2\epsilon\sum_{i=1}^d\lambda_i)}{r-\epsilon}\sum_x \biggl\{ \sum_{i=1}^d (x_i+\beta_i) \prod_{i=1}^d \ffrac{((r+\epsilon)\lambda_i)^{x_i} \exp(-(r+\epsilon)\lambda_i)}{x_i!} \biggl\} \notag \\ &= \ffrac{\exp(2\epsilon\sum_{i=1}^d\lambda_i)}{r-\epsilon} \sum_{i=1}^d \biggl\{ (r+\epsilon)\lambda_i+\beta_i\biggl\} <\infty. \label{L2-finite2} \end{align} Therefore, using \eqref{L2-finite1} and \eqref{L2-finite2}, we obtain \eqref{finite-goal}. \qed \vspace{0.5cm} \noindent \textbf{Lemma 1.3} Under conditions \eqref{t1-condition-a} and \eqref{t1-condition-b}, we have \begin{align*} &~~~~~F(x,r)<\infty\\ &~~~~F(x+\delta_i,r)<\infty\\ \int \biggl\{& \int g(\theta)\prod\limits_{j\neq i} \ffrac{r^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-r\theta_j^2)} {\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\theta_j \biggr\} \ffrac{r^{x_i+\beta_i}\theta_i^{2x_i+2\beta_i} \exp(-r\theta_i^2)}{\Gamma(x_i+1+\beta_i)}\ensuremath{\mathrm{d}}\theta_i <\infty\\ &\int \Bigl\lvert\ffrac{\partial}{\partial\theta_i}g(\theta) \Bigl\rvert \prod\limits_{j=1}^d \ffrac{r^{x_j+\beta_j} \theta_j^{2x_j+\delta_{ji}+2\beta_j-1}\exp(-r\theta_j^2)}{\Gamma(x_j+\delta_{ji}+\beta_j)} \ensuremath{\mathrm{d}}\theta < \infty \end{align*} \noindent \textbf{Proof}\\ From \eqref{t1-condition-a}, we have $$\int g(\theta)\prod\limits_{j=1}^{d}\theta_j^{2\beta_j-1}\theta_j^{2x_j}\exp(-r\theta_j^2)\ensuremath{\mathrm{d}}\theta<\infty.$$ Thus, we have $F(x,r)<\infty$ and \begin{equation} \int g(\theta)\prod\limits_{j=1}^d \ffrac{r^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-r\theta_j^2)} {\Gamma(x_j+\beta_j+\delta_{ij})}\ensuremath{\mathrm{d}}\theta<\infty. \label{L3-infty-1} \end{equation} From \eqref{t1-condition-a}, we have $$\int g(\theta)\prod\limits_{j=1}^{d}\theta_j^{2\beta_j-1}p(x+\delta_i\mid\theta)\ensuremath{\mathrm{d}}\theta<\infty.$$ Thus, $$\int g(\theta)\prod\limits_{j=1}^{d}\theta_j^{2\beta_j-1}\theta_j^{2x_j+2\delta_{ij}}\exp(-r\theta_j^2)\ensuremath{\mathrm{d}}\theta<\infty.$$ Hence we have $F(x+\delta_i,r)<\infty$ and \begin{equation} \int g(\theta)\prod\limits_{j=1}^d \ffrac{r^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j+2\delta_{ij}-1}\exp(-r\theta_j^2)} {\Gamma(x_j+\beta_j+\delta_{ij})}\ensuremath{\mathrm{d}}\theta<\infty. \label{L3-infty-2} \end{equation} Using \eqref{L3-infty-1} and \eqref{L3-infty-2}, we know that $$ \int \biggl\{ \int g(\theta)\prod\limits_{j\neq i} \ffrac{r^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-r\theta_j^2)} {\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\theta_j \biggr\} \ffrac{r^{x_i+\beta_i}(\theta_i^{2x_i+2\beta_i-1}+\theta_i^{2x_i+2\beta_i+1}) \exp(-r\theta_i^2)}{\Gamma(x_i+1+\beta_i)}\ensuremath{\mathrm{d}}\theta_i <\infty. $$ Thus, $$ \int \biggl\{ \int g(\theta)\prod\limits_{j\neq i} \ffrac{r^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-r\theta_j^2)} {\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\theta_j \biggr\} \ffrac{r^{x_i+\beta_i}\theta_i^{2x_i+2\beta_i} \exp(-r\theta_i^2)}{\Gamma(x_i+1+\beta_i)}\ensuremath{\mathrm{d}}\theta_i <\infty. $$ From \eqref{t1-condition-b}, we have $$\int \Bigl\lvert\ffrac{\partial}{\partial\theta_i}g(\theta)\Bigl\rvert\prod\limits_{j=1}^{d}\theta_j^{2\beta_j-1}\theta_j^{2x_j}\exp(-r\theta_j^2)\ensuremath{\mathrm{d}}\theta<\infty.$$ Thus, \begin{equation} \int \Bigl\lvert\ffrac{\partial}{\partial\theta_i}g(\theta)\Bigl\rvert\prod\limits_{j=1}^d \ffrac{r^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-r\theta_j^2)} {\Gamma(x_j+\beta_j+\delta_{ij})}\ensuremath{\mathrm{d}}\theta<\infty. \label{L3-infty-3} \end{equation} From \eqref{t1-condition-b}, we have $$\int \Bigl\lvert\ffrac{\partial}{\partial\theta_i}g(\theta)\Bigl\rvert\prod\limits_{j=1}^{d}\theta_j^{2\beta_j-1}p(x+\delta_i\mid\theta)\ensuremath{\mathrm{d}}\theta<\infty.$$ Thus, $$\int \Bigl\lvert\ffrac{\partial}{\partial\theta_i}g(\theta)\Bigl\rvert\prod\limits_{j=1}^{d}\theta_j^{2\beta_j-1}\theta_j^{2x_j+2\delta_{ij}}\exp(-r\theta_j^2)\ensuremath{\mathrm{d}}\theta<\infty.$$ Hence \begin{equation} \int \Bigl\lvert\ffrac{\partial}{\partial\theta_i}g(\theta)\Bigl\rvert\prod\limits_{j=1}^d \ffrac{r^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j+2\delta_{ij}-1}\exp(-r\theta_j^2)} {\Gamma(x_j+\beta_j+\delta_{ij})}\ensuremath{\mathrm{d}}\theta<\infty. \label{L3-infty-4} \end{equation} Using \eqref{L3-infty-3} and \eqref{L3-infty-4}, we have \begin{equation*} \int \Bigl\lvert\ffrac{\partial}{\partial\theta_i}g(\theta) \Bigl\rvert \prod_{j=1}^d\ffrac{r^{x_j+\beta_j}\theta_j^{2x_j+\delta_{ji}+2\beta_j-1} \exp(-r\theta_j^2)}{\Gamma(x_j+\delta_{ji}+\beta_j)} \ensuremath{\mathrm{d}}\theta <\infty. \end{equation*} \qed \vspace{0.5cm} \noindent \textbf{Lemma 1.4} Under conditions \eqref{t1-condition-a} and \eqref{t1-condition-b}, $$\biggl\{ \int_{[a,b]^{d-1}} g(\theta)\prod\limits_{j\neq i} \ffrac{r^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-r\theta_j^2)} {\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\theta_j \biggr\} \ffrac{r^{x_i+\beta_i}\theta_i^{2x_i+2\beta_i} \exp(-r\theta_i^2)}{\Gamma(x_i+1+\beta_i)}$$ converges as $\theta_i\to\infty$.\\ \noindent \textbf{Proof} By integration by parts on $\theta_i$, we have \begin{align} &\int_u^v \biggl\{ \int_{[a,b]^{d-1}} g(\theta)\prod\limits_{j\neq i}\ffrac{r^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-r\theta_j^2)} {\Gamma(x_j+\beta_j)} \ensuremath{\mathrm{d}} \theta_j \biggr\} \ffrac{r^{x_i+\beta_i}\theta_i^{2x_i+2\beta_i}\ffrac{\partial}{\partial\theta_i}\exp(-r\theta_i^2)} {\Gamma(x_i+1+\beta_i)}\ensuremath{\mathrm{d}}\theta_i \notag\\ & ~~ +\int_u^v \biggl\{ \int_{[a,b]^{d-1}} g(\theta)\prod\limits_{j\neq i}\ffrac{r^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-r\theta_j^2)} {\Gamma(x_j+\beta_j)} \ensuremath{\mathrm{d}} \theta_j \biggr\} \ffrac{r^{x_i+\beta_i}(\ffrac{\partial}{\partial\theta_i}\theta_i^{2x_i+2\beta_i})\exp(-r\theta_i^2)} {\Gamma(x_i+1+\beta_i)}\ensuremath{\mathrm{d}}\theta_i \notag\\ & ~~ +\int_{[u,v]\times[a,b]^{d-1}} \biggl\{ \ffrac{\partial}{\partial\theta_i}g(\theta) \biggr\} \prod\limits_{j=1}^d\ffrac{r^{x_j+\beta_j} \theta_j^{2x_j+\delta_{ji}+2\beta_j-1}\exp(-r\theta_j^2)}{\Gamma(x_j+\delta_{ji}+\beta_j)} \ensuremath{\mathrm{d}}\theta \label{L4}\\ &=\Biggl[ \biggl\{ \int_{[a,b]^{d-1}} g(\theta)\prod\limits_{j\neq i}\ffrac{r^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-r\theta_j^2)} {\Gamma(x_j+\beta_j)} \ensuremath{\mathrm{d}} \theta_j \biggr\} \ffrac{r^{x_i+\beta_i}\theta_i^{2x_i+2\beta_i}\exp(-r\theta_i^2)} {\Gamma(x_i+1+\beta_i)} \Biggr]^{v}_{\theta_i = u}. \notag \end{align} Using Lemma 1.3, we have $F(x+\delta_i,r)<\infty$, thus $$ \int \biggl\{ \int g(\theta)\prod\limits_{j\neq i}\ffrac{r^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-r\theta_j^2)} {\Gamma(x_j+\beta_j)} \ensuremath{\mathrm{d}} \theta_j \biggr\} \ffrac{r^{x_i+\beta_i}\theta_i^{2x_i+2\beta_i}\ffrac{\partial}{\partial\theta_i}\exp(-r\theta_i^2)}{\Gamma(x_i+1+\beta_i)}\ensuremath{\mathrm{d}}\theta_i <\infty.$$ Using Lemma 1.3, we have $F(x,r)<\infty$, thus $$ \int \biggl\{ \int g(\theta)\prod\limits_{j\neq i}\ffrac{r^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-r\theta_j^2)} {\Gamma(x_j+\beta_j)} \ensuremath{\mathrm{d}} \theta_j \biggr\} \ffrac{r^{x_i+\beta_i}(\ffrac{\partial}{\partial\theta_i}\theta_i^{2x_i+2\beta_i})\exp(-r\theta_i^2)}{\Gamma(x_i+1+\beta_i)}\ensuremath{\mathrm{d}}\theta_i <\infty.$$ Using Lemma 1.3, we have $$\int \Bigl\lvert\ffrac{\partial}{\partial\theta_i}g(\theta) \Bigl\rvert \prod\limits_{j=1}^d \ffrac{r^{x_j+\beta_j} \theta_j^{2x_j+\delta_{ji}+2\beta_j-1}\exp(-r\theta_j^2)}{\Gamma(x_j+\delta_{ji}+\beta_j)} \ensuremath{\mathrm{d}}\theta < \infty. $$ Thus, all three terms in \eqref{L4} converge as $v\to\infty$. Hence $$\Biggl[ \biggl\{ \int_{[a,b]^{d-1}} g(\theta)\prod\limits_{j\neq i}\ffrac{r^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-r\theta_j^2)} {\Gamma(x_j+\beta_j)} \ensuremath{\mathrm{d}} \theta_j \biggr\} \ffrac{r^{x_i+\beta_i}\theta_i^{2x_i+2\beta_i}\exp(-r\theta_i^2)} {\Gamma(x_i+1+\beta_i)} \Biggr]^{v}_{\theta_i = u}$$ converges as $v\to\infty$, which completes the proof. \qed \vspace{0.5cm} \noindent \textbf{Lemma 1.5} If $g\in\mathbf{C}^2([0,\infty)^d)$ and conditions \eqref{t1-condition-a}, \eqref{t1-condition-b}, and \eqref{t1-condition-c} are satisfied, then $$\int_{[a,b]^{d-1}} \ffrac{\partial g}{\partial\theta_i}(\theta)\theta_i^{2\beta_i-1} \prod_{j\neq i}\ffrac{r^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-r\theta_j^2)} {\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\theta_j\ffrac{r^{x_i+\beta_i}\theta_i^{2x_i}\exp(-r\theta_i^2)}{\Gamma(x_i+\beta_i)}$$ converges as $\theta_i\to0$ or $\theta_i\to\infty$. \noindent \textbf{Proof}\\ By integration by parts on $\theta_i$, we have \begin{align} & \int_u^v\int_{[a,b]^{d-1}} \ffrac{\partial g}{\partial\theta_i}(\theta)\theta_i^{2\beta_i-1} \biggl\{\prod_{j\neq i}\ffrac{r^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-r\theta_j^2)}{\Gamma(x_j+\beta_j)} \ensuremath{\mathrm{d}}\theta_j\biggr\} \ffrac{r^{x_i+\beta_i}\ffrac{\partial\theta_i^{2x_i}}{\partial\theta_i}\exp(-r\theta_i^2)}{\Gamma(x_i+\beta_i)}\ensuremath{\mathrm{d}}\theta_i \notag\\ & \qquad+ \int_u^v\int_{[a,b]^{d-1}} \ffrac{\partial g}{\partial\theta_i}(\theta)\theta_i^{2\beta_i-1} \biggl\{\prod_{j\neq i}\ffrac{r^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-r\theta_j^2)}{\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\theta_j\biggr\} \ffrac{r^{x_i+\beta_i}\theta_i^{2x_i}\ffrac{\partial}{\partial\theta_i}\exp(-r\theta_i^2)} {\Gamma(x_i+\beta_i)}\ensuremath{\mathrm{d}}\theta_i \notag\\ & \qquad+\int_u^v\int_{[a,b]^{d-1}} \ffrac{\partial}{\partial\theta_i}\Bigl\{\ffrac{\partial g}{\partial\theta_i}(\theta)\theta_i^{2\beta_i-1}\Bigr\} \biggl\{\prod_{j\neq i}\ffrac{r^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-r\theta_j^2)}{\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\theta_j\biggr\} \ffrac{r^{x_i+\beta_i}\theta_i^{2x_i}\exp(-r\theta_i^2)}{\Gamma(x_i+\beta_i)}\ensuremath{\mathrm{d}}\theta_i \label{L5}\\ &= \biggl[\int_{[a,b]^{d-1}} \ffrac{\partial g}{\partial\theta_i}(\theta)\theta_i^{2\beta_i-1} \biggl\{\prod_{j\neq i}\ffrac{r^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-r\theta_j^2)}{\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\theta_j\biggr\} \ffrac{r^{x_i+\beta_i}\theta_i^{2x_i}\exp(-r\theta_i^2)}{\Gamma(x_i+\beta_i)} \biggr]^{v}_{\theta_i = u}. \notag \end{align} From \eqref{F-1st-diff} and $x_i \{ F(x,r)-F(x-\delta_i,r) \}<\infty$, we know that $$\int\int \ffrac{\partial g}{\partial\theta_i}(\theta)\theta_i^{2\beta_i-1} \biggl\{\prod_{j\neq i}\ffrac{r^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-r\theta_j^2)}{\Gamma(x_j+\beta_j)} \ensuremath{\mathrm{d}}\theta_j\biggr\} \ffrac{r^{x_i+\beta_i}\ffrac{\partial\theta_i^{2x_i}}{\partial\theta_i}\exp(-r\theta_i^2)}{\Gamma(x_i+\beta_i)}\ensuremath{\mathrm{d}}\theta_i<\infty.$$ From \eqref{F-1st-diff} and $(x_i+\beta_i) \{F(x,r)-F(x+\delta_i,r)\}<\infty$, we know that $$\int\int \ffrac{\partial g}{\partial\theta_i}(\theta)\theta_i^{2\beta_i-1} \biggl\{\prod_{j\neq i}\ffrac{r^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-r\theta_j^2)}{\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\theta_j\biggr\} \ffrac{r^{x_i+\beta_i}\theta_i^{2x_i}\ffrac{\partial}{\partial\theta_i}\exp(-r\theta_i^2)} {\Gamma(x_i+\beta_i)}\ensuremath{\mathrm{d}}\theta_i<\infty.$$ Using \eqref{2nd-diff-finite}, we have $$\int\int \ffrac{\partial}{\partial\theta_i}\Bigl\{\ffrac{\partial g}{\partial\theta_i}(\theta)\theta_i^{2\beta_i-1}\Bigr\} \biggl\{\prod_{j\neq i}\ffrac{r^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-r\theta_j^2)}{\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\theta_j\biggr\} \ffrac{r^{x_i+\beta_i}\theta_i^{2x_i}\exp(-r\theta_i^2)}{\Gamma(x_i+\beta_i)}\ensuremath{\mathrm{d}}\theta_i<\infty.$$ Thus, all three terms in \eqref{L5} converge as $u\to0$ or $v\to\infty$. Hence $$\biggl[\int_{[a,b]^{d-1}} \ffrac{\partial g}{\partial\theta_i}(\theta)\theta_i^{2\beta_i-1} \biggl\{\prod_{j\neq i}\ffrac{r^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-r\theta_j^2)}{\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\theta_j\biggr\} \ffrac{r^{x_i+\beta_i}\theta_i^{2x_i}\exp(-r\theta_i^2)}{\Gamma(x_i+\beta_i)} \biggr]^{v}_{\theta_i = u}$$ converges as $u\to0$ or $v\to\infty$, which completes the proof. \qed \vspace{0.5cm} \noindent \textbf{Lemma 2.1} If the function $F(x,r)$ defined in Theorem 2 is finite, then for any given $r=(r_1,\dots,r_d)$, $s=(s_1,\dots,s_d)$ and $x=(x_1,\dots,x_d)$, we have $$\int \max_{\tau\in[0,1]}\Biggl\{ f(\bar\lambda_1,\dots,\bar\lambda_d) \prod_{i=1}^d \frac{t_i(\tau)^{x_i+\beta_i}\bar\lambda_i^{x_i+\beta_i-1}\exp(-t_i(\tau)\bar\lambda_i)}{\Gamma(x_i+\beta_i)} \Bigl| \sum_{i=1}^d (x_i+\beta_i)\gamma_it_i(\tau) - \sum_{i=1}^d \bar\lambda_i\gamma_it_i^2(\tau) \Bigr| \Biggl\} \ensuremath{\mathrm{d}}\bar\lambda<\infty.$$ \noindent \textbf{Proof}\\ We denote $r_0=\min\{r_1,\dots,r_d\}$ and $s_0=\max\{r_1+s_1,\dots,r_d+s_d\}$. Because $t_i(\tau)\in[r_i,r_i+s_i]$, we have $t_i(\tau)\in[r_0,s_0]$. Then we have \begin{align*} &\int \max_{\tau\in[0,1]}\Biggl\{ f(\bar\lambda_1,\dots,\bar\lambda_d) \prod_{i=1}^d \frac{t_i(\tau)^{x_i+\beta_i}\bar\lambda_i^{x_i+\beta_i-1}\exp(-t_i(\tau)\bar\lambda_i)}{\Gamma(x_i+\beta_i)} \Bigl| \sum_{i=1}^d (x_i+\beta_i)\gamma_it_i(\tau) - \sum_{i=1}^d \bar\lambda_i\gamma_it_i^2(\tau) \Bigr| \Biggl\} \ensuremath{\mathrm{d}}\bar\lambda\\ &\le\int \Biggl[ f(\bar\lambda_1,\dots,\bar\lambda_d) \prod_{i=1}^d \frac{s_0^{x_i+\beta_i}\bar\lambda_i^{x_i+\beta_i-1}\exp(-r_0\bar\lambda_i)}{\Gamma(x_i+\beta_i)} \Bigl\{ \sum_{i=1}^d (x_i+\beta_i)\gamma_is_0 + \sum_{i=1}^d \bar\lambda_i\gamma_is_0^2 \Bigr\} \Biggl] \ensuremath{\mathrm{d}}\bar\lambda \\ &=\sum_{i=1}^d (x_i+\beta_i)\gamma_is_0\Bigl( \ffrac{s_0}{r_0}\Bigr) ^{\sum_{i=1}^d(x_i+\beta_i)}\int f(\bar\lambda_1,\dots,\bar\lambda_d) \prod_{i=1}^d \ffrac{r_0^{x_i+\beta_i}\bar\lambda_i^{x_i+\beta_i-1}\exp(-r_0\bar\lambda_i)}{\Gamma(x_i+\beta_i)} \ensuremath{\mathrm{d}}\bar\lambda\\ &~~~~+\sum_{j=1}^{d}\gamma_js_0^2\int f(\bar\lambda_1,\dots,\bar\lambda_d) \prod_{i=1}^d \frac{s_0^{x_i+\beta_i}\bar\lambda_i^{x_i+\beta_i-1}\exp(-r_0\bar\lambda_i)}{\Gamma(x_i+\beta_i)} \bar\lambda_j \ensuremath{\mathrm{d}}\bar\lambda \\ &=\sum_{i=1}^d (x_i+\beta_i)\gamma_is_0\Bigl( \ffrac{s_0}{r_0}\Bigr) ^{\sum_{i=1}^d(x_i+\beta_i)}\int f(\bar\lambda_1,\dots,\bar\lambda_d) \prod_{i=1}^d \ffrac{r_0^{x_i+\beta_i}\bar\lambda_i^{x_i+\beta_i-1}\exp(-r_0\bar\lambda_i)}{\Gamma(x_i+\beta_i)} \ensuremath{\mathrm{d}}\bar\lambda\\ &~~~~+\sum_{j=1}^{d}\gamma_js_0\Bigl( \ffrac{s_0}{r_0}\Bigr) ^{\sum_{i=1}^d(x_i+\beta_i)+1}\int f(\bar\lambda_1,\dots,\bar\lambda_d) \prod_{i=1}^d \frac{r_0^{x_i+\beta_i+\delta_{ij}}\bar\lambda_i^{x_i+\beta_i+\delta_{ij}-1}\exp(-r_0\bar\lambda_i)}{\Gamma(x_i+\beta_i)} \ensuremath{\mathrm{d}}\bar\lambda \\ &=\sum_{i=1}^d (x_i+\beta_i)\gamma_is_0\Bigl( \ffrac{s_0}{r_0}\Bigr) ^{\sum_{i=1}^d(x_i+\beta_i)}F(x,(r_0,\dots,r_0)) +\sum_{j=1}^{d}\gamma_js_0\Bigl(\ffrac{s_0}{r_0}\Bigr)^{\sum_{i=1}^d(x_i+\beta_i)+1}F(x+\delta_j,(r_0,\dots,r_0))(x_j+\beta_j). \end{align*} Because $F(x,(r_0,\dots,r_0))$ and $F(x+\delta_j,(r_0,\dots,r_0))$ are finite, the proof is complete. \qed \vspace{0.5cm} \noindent \textbf{Lemma 2.2} Under conditions \eqref{F-differentiable-3} and \eqref{Fineq3}, for a given $\tau_0$, there exists $\epsilon>0$ such that \begin{align*} &\sum_x \max_{|\tau-\tau_0|\le\epsilon} \Biggl[ \Biggl| \biggl\{ \log F(x,t(\tau)) \biggr\} \biggl\{ \prod_{i=1}^d \frac{(t_i(\tau)\lambda_i)^{x_i}\exp(-t_i(\tau)\lambda_i)}{x_i!} \biggr\} \Bigl(\sum_{i=1}^d x_i\gamma_it_i(\tau) - \sum_{i=1}^d \lambda_i\gamma_it_i^2(\tau) \Bigr) \Biggr| \Biggr] \notag \\ &\quad + \sum_x \max_{|\tau-\tau_0|\le\epsilon} \Biggl[ \frac{ \displaystyle \biggl| \int f(\bar\lambda_1,\dots,\bar\lambda_d) \prod_{i=1}^d \frac{t_i(\tau)^{x_i+\beta_i}\bar\lambda_i^{x_i+\beta_i-1}\exp(-t_i(\tau)\bar\lambda_i)}{\Gamma(x_i+\beta_i)} \Bigl( \sum_{i=1}^d (x_i+\beta_i)\gamma_it_i(\tau) - \sum_{i=1}^d \bar\lambda_i\gamma_it_i^2(\tau) \Bigr) \ensuremath{\mathrm{d}}\bar\lambda \Biggr| } { F(x,t(\tau))} \notag\\ & ~~~~~ \biggl\{ \prod_{i=1}^d \frac{(t_i(\tau)\lambda_i)^{x_i}\exp(-t_i(\tau)\lambda_i)}{x_i!} \biggr\} \Biggr] \notag \\ &<\infty. \end{align*} \noindent \textbf{Proof}\\ From \eqref{transform2-t2}, we know that \begin{align*} &\int f(\bar\lambda_1,\dots,\bar\lambda_d) \prod_{j=1}^d\ffrac{t_j(\tau)^{x_j+\beta_j}\bar\lambda_j^{x_j+\beta_j-1} \exp(-t_j(\tau)\bar\lambda_j)}{\Gamma(x_j+\beta_j)} \bar\lambda_i\gamma_it_i^2(\tau) \ensuremath{\mathrm{d}}\bar\lambda = F(x+\delta_i,t(\tau))(x_i+\beta_i)\gamma_it_i(\tau). \end{align*} Therefore, the Lemma is equivalent to \begin{align} &\sum_x \max_{|\tau-\tau_0|\le\epsilon} \Biggl[ \biggl| \log F(x,t(\tau)) \biggr| \Bigl\{ \prod_{i=1}^d \frac{(t_i(\tau)\lambda_i)^{x_i}\exp(-t_i(\tau)\lambda_i)}{x_i!} \Bigr\} \biggl| \sum_{i=1}^d x_i\gamma_it_i(\tau) - \sum_{i=1}^d \lambda_i\gamma_it_i^2(\tau) \biggr| \Biggl] \notag \\ &\quad + \sum_x \max_{|\tau-\tau_0|\le\epsilon} \Biggl\{ \frac{ \biggl| F(x,t(\tau))\sum_{i=1}^d (x_i+\beta_i)\gamma_it_i(\tau)-\sum_{i=1}^dF(x+\delta_i,t(\tau))(x_i+\beta_i)\gamma_it_i(\tau) \biggl| } {F(x,t(\tau))} \notag\\ &~~~~~~~~~~~~~~~~~\prod_{i=1}^d \ffrac{(t_i(\tau)\lambda_i)^{x_i} \exp(-t_i(\tau)\lambda_i)}{x_i!} \Biggr\} \notag \\ &<\infty. \label{finite-goal-t2} \end{align} Using \eqref{F-differentiable-3}, when we set $r=t(\tau_0)$, we know that there exists $\delta>0$ such that \begin{align} &\sum_x\Biggl[ \max_{t\in \bar B(t(\tau_0),\delta)}\biggl\{ \Bigl| \log F(x,t) \Bigr| (\sum_{i=1}^dx_i+1) \prod_{i=1}^d \frac{(t_i\lambda_i)^{x_i}\exp(-t_i\lambda_i)}{x_i!} \biggr\} \Biggr] <\infty. \notag \end{align} Because $t(\tau)$ is continuous, there exists $\epsilon>0$ such that for any $\tau\in[\tau_0-\epsilon,\tau_0+\epsilon]$ , $t(\tau)\in \bar B(t(\tau_0),\delta)$. Thus, we have \begin{align} &\sum_x \max_{|\tau-\tau_0|\le\epsilon} \Biggl[ \biggl| \log F(x,t(\tau)) \biggr| \Bigl\{ \prod_{i=1}^d \frac{(t_i(\tau)\lambda_i)^{x_i}\exp(-t_i(\tau)\lambda_i)}{x_i!} \Bigr\} \biggl| \sum_{i=1}^d x_i\gamma_it_i(\tau) - \sum_{i=1}^d \lambda_i\gamma_it_i^2(\tau) \biggr| \Biggl] \notag \\ &\le \sum_x \max_{t\in \bar B(t(\tau_0),\delta)} \Biggl[ \biggl| \log F(x,t) \biggr| \Bigl\{ \prod_{i=1}^d \frac{(t_i\lambda_i)^{x_i}\exp(-t_i\lambda_i)}{x_i!} \Bigr\} \biggl| \sum_{i=1}^d x_i\gamma_it_i - \sum_{i=1}^d \lambda_i\gamma_it_i^2 \biggr| \Biggl] \notag \\ &\le\sum_x \max_{t\in \bar B(t(\tau_0),\delta)}\Biggl[ \biggl| \log F(x,t) \biggr| \Bigl\{\prod_{i=1}^d \frac{(t_i\lambda_i)^{x_i}\exp(-t_i\lambda_i)}{x_i!}\Bigr\}( \sum_{i=1}^d x_i +1)\Bigl\{ \sum_{i=1}^d \gamma_i(\Vert t(\tau_0)\Vert+\delta)+\sum_{i=1}^d \lambda_i\gamma_i(\Vert t(\tau_0)\Vert+\delta)^2 \Bigr\} \Biggl] \notag\\ &=\sum_x \max_{t\in \bar B(t(\tau_0),\delta)}\Biggl[ \biggl| \log F(x,t) \biggr| \Bigl\{\prod_{i=1}^d \frac{(t_i\lambda_i)^{x_i}\exp(-t_i\lambda_i)}{x_i!}\Bigr\}( \sum_{i=1}^d x_i +1) \Biggl]\Bigl\{ \sum_{i=1}^d \gamma_i(\Vert t(\tau_0)\Vert+\delta)+\sum_{i=1}^d \lambda_i\gamma_i(\Vert t(\tau_0)\Vert+\delta)^2 \Bigr\} \notag\\ &<\infty. \label{L2-finite1-t2} \end{align} Using \eqref{Fineq3}, we have \begin{equation*} \sum\limits_{i=1}^d\gamma_ir_ix_i(F(x,r)-F(x-\delta_i,r))+\sum\limits_{i=1}^d\gamma_ir_i(x_i+\beta_i)(F(x,r)-F(x+\delta_i,r))\ge 0, \end{equation*} thus, when we set $r=t(\tau)$, we get $$\sum_{i=1}^d\gamma_it_i(\tau)(x_i+\beta_i)F(x+\delta_i,t(\tau)) <2\sum_{i=1}^d\gamma_it_i(\tau)(x_i+\beta_i)F(x,t(\tau)).$$ Therefore, \begin{align} & \sum_x \max_{|\tau-\tau_0|\le\epsilon} \Biggl\{ \frac{ \biggl| F(x,t(\tau))\sum_{i=1}^d (x_i+\beta_i)\gamma_it_i(\tau)-\sum_{i=1}^dF(x+\delta_i,t(\tau))(x_i+\beta_i)\gamma_it_i(\tau) \biggl| } {F(x,t(\tau))} \notag\\ &~~~~~~~~~~~\prod_{i=1}^d \ffrac{(t_i(\tau)\lambda_i)^{x_i} \exp(-t_i(\tau)\lambda_i)}{x_i!} \Biggr\} \notag \\ &< \sum_x \max_{|\tau-\tau_0|\le\epsilon} \Biggl\{ \frac{ F(x,t(\tau))\sum_{i=1}^d (x_i+\beta_i)\gamma_it_i(\tau) } {F(x,t(\tau))}\prod_{i=1}^d \ffrac{(t_i(\tau)\lambda_i)^{x_i} \exp(-t_i(\tau)\lambda_i)}{x_i!} \Biggr\}. \notag \end{align} We denote $r_0=\min\{r_1,\dots,r_d\}$ and $s_0=\max\{r_1+s_1,\dots,r_d+s_d\}$. Because $t_i(\tau)\in[r_i,r_i+s_i]$, we have $t_i(\tau)\in[r_0,s_0]$. Therefore, \begin{align} & \sum_x \max_{|\tau-\tau_0|\le\epsilon} \Biggl\{ \frac{ \biggl| F(x,t(\tau))\sum_{i=1}^d (x_i+\beta_i)\gamma_it_i(\tau)-\sum_{i=1}^dF(x+\delta_i,t(\tau))(x_i+\beta_i)\gamma_it_i(\tau) \biggl| } {F(x,t(\tau))} \notag\\ &~~~~~~~~~~~\prod_{i=1}^d \ffrac{(t_i(\tau)\lambda_i)^{x_i} \exp(-t_i(\tau)\lambda_i)}{x_i!} \Biggr\} \notag \\ &< \sum_x \max_{|\tau-\tau_0|\le\epsilon} \Biggl\{ \frac{ F(x,t(\tau))\sum_{i=1}^d (x_i+\beta_i)\gamma_it_i(\tau) } {F(x,t(\tau))}\prod_{i=1}^d \ffrac{(t_i(\tau)\lambda_i)^{x_i} \exp(-t_i(\tau)\lambda_i)}{x_i!} \Biggr\} \notag\\ &\le \sum_x \Biggl\{ \biggl( \sum_{i=1}^d (x_i+\beta_i)\gamma_is_0\biggl) \prod_{i=1}^d \ffrac{(s_0\lambda_i)^{x_i} \exp(-r_0\lambda_i)}{x_i!} \Biggr\} \notag \\ &\le \exp((s_0-r_0)\sum_{i=1}^d\lambda_i)\sum_x \Biggl\{ \biggl( \sum_{i=1}^d (x_i+\beta_i)\gamma_is_0\biggl) \prod_{i=1}^d \ffrac{(s_0\lambda_i)^{x_i} \exp(-s_0\lambda_i)}{x_i!} \Biggr\} \notag \\ &= \exp((s_0-r_0)\sum_{i=1}^d\lambda_i) \biggl( \sum_{i=1}^d (s_0\lambda_i+\beta_i)\gamma_is_0\biggl) \notag \\ &<\infty. \label{L2-finite2-t2} \end{align} Therefore, using \eqref{L2-finite1-t2} and \eqref{L2-finite2-t2}, we obtain \eqref{finite-goal-t2}. \qed \vspace{0.5cm} \noindent \textbf{Lemma 2.3} Under conditions \eqref{t2-condition-a} and \eqref{t2-condition-b}, we have \begin{align*} &~~~~~F(x,r)<\infty\\ &~~~~F(x+\delta_i,r)<\infty\\ \int \biggl\{& \int g(\theta)\prod\limits_{j\neq i} \ffrac{(\gamma_jr_j)^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-\gamma_jr_j\theta_j^2)} {\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\theta_j \biggr\} \ffrac{(\gamma_ir_i)^{x_i+\beta_i}\theta_i^{2x_i+2\beta_i} \exp(-\gamma_ir_i\theta_i^2)}{\Gamma(x_i+1+\beta_i)}\ensuremath{\mathrm{d}}\theta_i <\infty\\ &\int \Bigl\lvert\ffrac{\partial}{\partial\theta_i}g(\theta) \Bigl\rvert \prod\limits_{j=1}^d \ffrac{(\gamma_jr_j)^{x_j+\beta_j} \theta_j^{2x_j+\delta_{ji}+2\beta_j-1}\exp(-\gamma_jr_j\theta_j^2)}{\Gamma(x_j+\delta_{ji}+\beta_j)} \ensuremath{\mathrm{d}}\theta < \infty \end{align*} \noindent \textbf{Proof}\\ From \eqref{t2-condition-a}, we have $$\int g(\theta)\prod\limits_{j=1}^{d}\theta_j^{2\beta_j-1}\theta_j^{2x_j}\exp(-\gamma_jr_j\theta_j^2)\ensuremath{\mathrm{d}}\theta<\infty.$$ Thus, we have $F(x,r)<\infty$ and \begin{equation} \int g(\theta)\prod\limits_{j=1}^d \ffrac{(\gamma_jr_j)^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-\gamma_jr_j\theta_j^2)} {\Gamma(x_j+\beta_j+\delta_{ij})}\ensuremath{\mathrm{d}}\theta<\infty. \label{L3-infty-1-t2} \end{equation} From \eqref{t2-condition-a}, we have $$\int g(\theta)\prod\limits_{j=1}^{d}\theta_j^{2\beta_j-1}p(x+\delta_i\mid\theta)\ensuremath{\mathrm{d}}\theta<\infty.$$ Hence $$\int g(\theta)\prod\limits_{j=1}^{d}\theta_j^{2\beta_j-1}\theta_j^{2x_j+2\delta_{ij}}\exp(-\gamma_jr_j\theta_j^2)\ensuremath{\mathrm{d}}\theta<\infty.$$ Thus, we have $F(x+\delta_i,r)<\infty$ and \begin{equation} \int g(\theta)\prod\limits_{j=1}^d \ffrac{(\gamma_jr_j)^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j+2\delta_{ij}-1}\exp(-\gamma_jr_j\theta_j^2)} {\Gamma(x_j+\beta_j+\delta_{ij})}\ensuremath{\mathrm{d}}\theta<\infty. \label{L3-infty-2-t2} \end{equation} From \eqref{L3-infty-1-t2} and \eqref{L3-infty-2-t2}, we know that $$ \int \biggl\{ \int g(\theta)\prod\limits_{j\neq i} \ffrac{(\gamma_jr_j)^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-\gamma_jr_j\theta_j^2)} {\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\theta_j \biggr\} \ffrac{(\gamma_ir_i)^{x_i+\beta_i}(\theta_i^{2x_i+2\beta_i-1}+\theta_i^{2x_i+2\beta_i+1}) \exp(-\gamma_ir_i\theta_i^2)}{\Gamma(x_i+1+\beta_i)}\ensuremath{\mathrm{d}}\theta_i <\infty. $$ Thus, $$ \int \biggl\{ \int g(\theta)\prod\limits_{j\neq i} \ffrac{(\gamma_jr_j)^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-\gamma_jr_j\theta_j^2)} {\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\theta_j \biggr\} \ffrac{(\gamma_ir_i)^{x_i+\beta_i}\theta_i^{2x_i+2\beta_i} \exp(-\gamma_ir_i\theta_i^2)}{\Gamma(x_i+1+\beta_i)}\ensuremath{\mathrm{d}}\theta_i <\infty. $$ From \eqref{t2-condition-b}, we have $$\int \Bigl\lvert\ffrac{\partial}{\partial\theta_i}g(\theta)\Bigl\rvert\prod\limits_{j=1}^{d}\theta_j^{2\beta_j-1}\theta_j^{2x_j}\exp(-\gamma_jr_j\theta_j^2)\ensuremath{\mathrm{d}}\theta<\infty.$$ Thus, \begin{equation} \int \Bigl\lvert\ffrac{\partial}{\partial\theta_i}g(\theta)\Bigl\rvert\prod\limits_{j=1}^d \ffrac{(\gamma_jr_j)^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-\gamma_jr_j\theta_j^2)} {\Gamma(x_j+\beta_j+\delta_{ij})}\ensuremath{\mathrm{d}}\theta<\infty. \label{L3-infty-3-t2} \end{equation} From \eqref{t2-condition-b}, we have $$\int \Bigl\lvert\ffrac{\partial}{\partial\theta_i}g(\theta)\Bigl\rvert\prod\limits_{j=1}^{d}\theta_j^{2\beta_j-1}p(x+\delta_i\mid\theta)\ensuremath{\mathrm{d}}\theta<\infty.$$ Thus, $$\int \Bigl\lvert\ffrac{\partial}{\partial\theta_i}g(\theta)\Bigl\rvert\prod\limits_{j=1}^{d}\theta_j^{2\beta_j-1}\theta_j^{2x_j+2\delta_{ij}}\exp(-\gamma_jr_j\theta_j^2)\ensuremath{\mathrm{d}}\theta<\infty.$$ Hence \begin{equation} \int \Bigl\lvert\ffrac{\partial}{\partial\theta_i}g(\theta)\Bigl\rvert\prod\limits_{j=1}^d \ffrac{(\gamma_jr_j)^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j+2\delta_{ij}-1}\exp(-\gamma_jr_j\theta_j^2)} {\Gamma(x_j+\beta_j+\delta_{ij})}\ensuremath{\mathrm{d}}\theta<\infty. \label{L3-infty-4-t2} \end{equation} Using \eqref{L3-infty-3-t2} and \eqref{L3-infty-4-t2}, we have \begin{equation*} \int \Bigl\lvert\ffrac{\partial}{\partial\theta_i}g(\theta) \Bigl\rvert \prod_{j=1}^d\ffrac{(\gamma_jr_j)^{x_j+\beta_j}\theta_j^{2x_j+\delta_{ji}+2\beta_j-1} \exp(-\gamma_jr_j\theta_j^2)}{\Gamma(x_j+\delta_{ji}+\beta_j)} \ensuremath{\mathrm{d}}\theta <\infty. \end{equation*} \qed \vspace{0.5cm} \noindent \textbf{Lemma 2.4} Under conditions \eqref{t2-condition-a} and \eqref{t2-condition-b}, $$\displaystyle \biggl\{ \int_{[a,b]^{d-1}} g(\theta)\prod\limits_{j\neq i} \ffrac{(\gamma_jr_j)^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-\gamma_jr_j\theta_j^2)} {\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\theta_j \biggr\} \ffrac{(\gamma_ir_i)^{x_i+\beta_i}\theta_i^{2x_i+2\beta_i} \exp(-\gamma_ir_i\theta_i^2)}{\Gamma(x_i+1+\beta_i)}$$ converges as $\theta_i\to\infty$.\\ \noindent \textbf{Proof} By integration by parts on $\theta_i$, we have \begin{align} &\int_u^v \biggl\{ \int_{[a,b]^{d-1}} g(\theta)\prod\limits_{j\neq i}\ffrac{(\gamma_jr_j)^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-\gamma_jr_j\theta_j^2)} {\Gamma(x_j+\beta_j)} \ensuremath{\mathrm{d}} \theta_j \biggr\} \ffrac{(\gamma_ir_i)^{x_i+\beta_i}\theta_i^{2x_i+2\beta_i}\ffrac{\partial}{\partial\theta_i}\exp(-\gamma_ir_i\theta_i^2)} {\Gamma(x_i+1+\beta_i)}\ensuremath{\mathrm{d}}\theta_i \notag\\ & ~~ +\int_u^v \biggl\{ \int_{[a,b]^{d-1}} g(\theta)\prod\limits_{j\neq i}\ffrac{(\gamma_jr_j)^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-\gamma_jr_j\theta_j^2)} {\Gamma(x_j+\beta_j)} \ensuremath{\mathrm{d}} \theta_j \biggr\} \ffrac{(\gamma_ir_i)^{x_i+\beta_i}(\ffrac{\partial}{\partial\theta_i}\theta_i^{2x_i+2\beta_i})\exp(-\gamma_ir_i\theta_i^2)} {\Gamma(x_i+1+\beta_i)}\ensuremath{\mathrm{d}}\theta_i \notag\\ & ~~ +\int_{[u,v]\times[a,b]^{d-1}} \biggl\{ \ffrac{\partial}{\partial\theta_i}g(\theta) \biggr\} \prod\limits_{j=1}^d\ffrac{(\gamma_jr_j)^{x_j+\beta_j} \theta_j^{2x_j+\delta_{ji}+2\beta_j-1}\exp(-\gamma_jr_j\theta_j^2)}{\Gamma(x_j+\delta_{ji}+\beta_j)} \ensuremath{\mathrm{d}}\theta \label{L4-t2}\\ &=\Biggl[ \biggl\{ \int_{[a,b]^{d-1}} g(\theta)\prod\limits_{j\neq i}\ffrac{(\gamma_jr_j)^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-\gamma_jr_j\theta_j^2)} {\Gamma(x_j+\beta_j)} \ensuremath{\mathrm{d}} \theta_j \biggr\} \ffrac{(\gamma_ir_i)^{x_i+\beta_i}\theta_i^{2x_i+2\beta_i}\exp(-\gamma_ir_i\theta_i^2)} {\Gamma(x_i+1+\beta_i)} \Biggr]^{v}_{\theta_i = u}. \notag \end{align} Using Lemma 2.3, we have $F(x+\delta_i,r)<\infty$, thus $$ \int \biggl\{ \int g(\theta)\prod\limits_{j\neq i}\ffrac{(\gamma_jr_j)^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-\gamma_jr_j\theta_j^2)} {\Gamma(x_j+\beta_j)} \ensuremath{\mathrm{d}} \theta_j \biggr\} \ffrac{(\gamma_ir_i)^{x_i+\beta_i}\theta_i^{2x_i+2\beta_i}\ffrac{\partial}{\partial\theta_i}\exp(-\gamma_ir_i\theta_i^2)} {\Gamma(x_i+1+\beta_i)}\ensuremath{\mathrm{d}}\theta_i <\infty.$$ Using Lemma 2.3, we have $F(x,r)<\infty$, thus $$ \int \biggl\{ \int g(\theta)\prod\limits_{j\neq i}\ffrac{(\gamma_jr_j)^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-\gamma_jr_j\theta_j^2)} {\Gamma(x_j+\beta_j)} \ensuremath{\mathrm{d}} \theta_j \biggr\} \ffrac{(\gamma_ir_i)^{x_i+\beta_i}(\ffrac{\partial}{\partial\theta_i}\theta_i^{2x_i+2\beta_i})\exp(-\gamma_ir_i\theta_i^2)} {\Gamma(x_i+1+\beta_i)}\ensuremath{\mathrm{d}}\theta_i <\infty.$$ Using Lemma 2.3, we have $$\int \Bigl\lvert\ffrac{\partial}{\partial\theta_i}g(\theta) \Bigl\rvert \prod\limits_{j=1}^d \ffrac{(\gamma_jr_j)^{x_j+\beta_j} \theta_j^{2x_j+\delta_{ji}+2\beta_j-1}\exp(-\gamma_jr_j\theta_j^2)}{\Gamma(x_j+\delta_{ji}+\beta_j)} \ensuremath{\mathrm{d}}\theta < \infty. $$ Hence all three terms in \eqref{L4-t2} converge as $v\to\infty$. Thus, $$\Biggl[ \biggl\{ \int_{[a,b]^{d-1}} g(\theta)\prod\limits_{j\neq i}\ffrac{(\gamma_jr_j)^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-\gamma_jr_j\theta_j^2)} {\Gamma(x_j+\beta_j)} \ensuremath{\mathrm{d}} \theta_j \biggr\} \ffrac{(\gamma_ir_i)^{x_i+\beta_i}\theta_i^{2x_i+2\beta_i}\exp(-\gamma_ir_i\theta_i^2)} {\Gamma(x_i+1+\beta_i)} \Biggr]^{v}_{\theta_i = u}$$ converges as $v\to\infty$, which completes the proof. \qed \vspace{0.5cm} \noindent \textbf{Lemma 2.5} If $g\in\mathbf{C}^2([0,\infty)^d)$ and conditions \eqref{t2-condition-a}, \eqref{t2-condition-b}, and \eqref{t2-condition-c} are satisfied, then $$\int_{[a,b]^{d-1}} \ffrac{\partial g}{\partial\theta_i}(\theta)\theta_i^{2\beta_i-1} \prod_{j\neq i}\ffrac{(\gamma_jr_j)^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-\gamma_jr_j\theta_j^2)} {\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\theta_j\ffrac{(\gamma_ir_i)^{x_i+\beta_i}\theta_i^{2x_i}\exp(-\gamma_ir_i\theta_i^2)}{\Gamma(x_i+\beta_i)}$$ converges as $\theta_i\to0$ or $\theta_i\to\infty$. \noindent \textbf{Proof}\\ By integration by parts on $\theta_i$, we have \begin{align} & \int_u^v\int_{[a,b]^{d-1}} \ffrac{\partial g}{\partial\theta_i}(\theta)\theta_i^{2\beta_i-1} \biggl\{ \prod_{j\neq i}\ffrac{(\gamma_jr_j)^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-\gamma_jr_j\theta_j^2)}{\Gamma(x_j+\beta_j)} \ensuremath{\mathrm{d}}\theta_j \biggr\} \ffrac{(\gamma_ir_i)^{x_i+\beta_i}\ffrac{\partial\theta_i^{2x_i}}{\partial\theta_i}\exp(-\gamma_ir_i\theta_i^2)}{\Gamma(x_i+\beta_i)}\ensuremath{\mathrm{d}}\theta_i \notag\\ & \qquad+ \int_u^v\int_{[a,b]^{d-1}} \ffrac{\partial g}{\partial\theta_i}(\theta)\theta_i^{2\beta_i-1} \biggl\{\prod_{j\neq i}\ffrac{(\gamma_jr_j)^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-\gamma_jr_j\theta_j^2)}{\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\theta_j \biggr\} \ffrac{(\gamma_ir_i)^{x_i+\beta_i}\theta_i^{2x_i}\ffrac{\partial}{\partial\theta_i}\exp(-\gamma_ir_i\theta_i^2)} {\Gamma(x_i+\beta_i)}\ensuremath{\mathrm{d}}\theta_i \notag\\ & \qquad+\int_u^v\int_{[a,b]^{d-1}} \ffrac{\partial}{\partial\theta_i}\Bigl\{\ffrac{\partial g}{\partial\theta_i}(\theta)\theta_i^{2\beta_i-1}\Bigr\} \biggl\{ \prod_{j\neq i}\ffrac{(\gamma_jr_j)^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-\gamma_jr_j\theta_j^2)}{\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\theta_j\biggr\} \ffrac{(\gamma_ir_i)^{x_i+\beta_i}\theta_i^{2x_i}\exp(-\gamma_ir_i\theta_i^2)}{\Gamma(x_i+\beta_i)}\ensuremath{\mathrm{d}}\theta_i \label{L5-t2}\\ &= \biggl[\int_{[a,b]^{d-1}} \ffrac{\partial g}{\partial\theta_i}(\theta)\theta_i^{2\beta_i-1} \biggl\{ \prod_{j\neq i}\ffrac{(\gamma_jr_j)^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-\gamma_jr_j\theta_j^2)} {\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\theta_j \biggr\} \ffrac{(\gamma_ir_i)^{x_i+\beta_i}\theta_i^{2x_i}\exp(-\gamma_ir_i\theta_i^2)}{\Gamma(x_i+\beta_i)} \biggr]^{v}_{\theta_i = u}. \notag \end{align} Using \eqref{F-1st-diff-t2} and $\gamma_ir_ix_i \{ F(x,r)-F(x-\delta_i,r) \}<\infty$, we know that $$\int\int \ffrac{\partial g}{\partial\theta_i}(\theta)\theta_i^{2\beta_i-1} \biggl\{\prod_{j\neq i}\ffrac{(\gamma_jr_j)^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-\gamma_jr_j\theta_j^2)}{\Gamma(x_j+\beta_j)} \ensuremath{\mathrm{d}}\theta_j \biggr\} \ffrac{(\gamma_ir_i)^{x_i+\beta_i}\ffrac{\partial\theta_i^{2x_i}}{\partial\theta_i}\exp(-\gamma_ir_i\theta_i^2)}{\Gamma(x_i+\beta_i)}\ensuremath{\mathrm{d}}\theta_i<\infty.$$ Using \eqref{F-1st-diff-t2} and $\gamma_ir_i(x_i+\beta_i) \{F(x,r)-F(x+\delta_i,r)\}<\infty$, we know that $$\int\int \ffrac{\partial g}{\partial\theta_i}(\theta)\theta_i^{2\beta_i-1} \biggl\{ \prod_{j\neq i}\ffrac{(\gamma_jr_j)^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-\gamma_jr_j\theta_j^2)}{\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\theta_j\biggr\} \ffrac{(\gamma_ir_i)^{x_i+\beta_i}\theta_i^{2x_i}\ffrac{\partial}{\partial\theta_i}\exp(-\gamma_ir_i\theta_i^2)} {\Gamma(x_i+\beta_i)}\ensuremath{\mathrm{d}}\theta_i<\infty.$$ Using \eqref{2nd-diff-finite-t2}, we have $$\int\int \ffrac{\partial}{\partial\theta_i}\Bigl\{\ffrac{\partial g}{\partial\theta_i}(\theta)\theta_i^{2\beta_i-1}\Bigr\} \biggl\{ \prod_{j\neq i}\ffrac{(\gamma_jr_j)^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-\gamma_jr_j\theta_j^2)}{\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\theta_j \biggr\} \ffrac{(\gamma_ir_i)^{x_i+\beta_i}\theta_i^{2x_i}\exp(-\gamma_ir_i\theta_i^2)}{\Gamma(x_i+\beta_i)}\ensuremath{\mathrm{d}}\theta_i<\infty.$$ Hence all three terms in \eqref{L5-t2} converge as $u\to0$ or $v\to\infty$. Thus, $$\biggl[\int_{[a,b]^{d-1}} \ffrac{\partial g}{\partial\theta_i}(\theta)\theta_i^{2\beta_i-1} \biggl\{ \prod_{j\neq i}\ffrac{(\gamma_jr_j)^{x_j+\beta_j}\theta_j^{2x_j+2\beta_j-1}\exp(-\gamma_jr_j\theta_j^2)} {\Gamma(x_j+\beta_j)}\ensuremath{\mathrm{d}}\theta_j \biggr\} \ffrac{(\gamma_ir_i)^{x_i+\beta_i}\theta_i^{2x_i}\exp(-\gamma_ir_i\theta_i^2)}{\Gamma(x_i+\beta_i)} \biggr]^{v}_{\theta_i = u}$$ converges as $u\to0$ or $v\to\infty$, which completes the proof. \qed \section*{Acknowledgments} The first author is grateful for support from the China Scholarship Council. \bibliographystyle{acmtrans.bst}

d1a00db1183c92533e437e9cf2a9a6c9033dbb7d

\section{Introduction} \label{sec-int} Deep reinforcement learning (DRL) is a booming field of research in machine learning with diverse real-world applications \cite{ibarz2021}. In recent years, many \emph{model-free} DRL algorithms achieved cutting-edge performance in tackling various continuous reinforcement learning (RL) problems, including complex control tasks with high-dimensional state and action spaces \cite{liu2021}. These algorithms can effectively train \emph{deep neural networks} (DNNs) to precisely model high-quality control policies and are the central focus of this paper. Despite of widely reported success, a majority of existing \emph{actor-critic DRL algorithms}, such as DDPG \cite{lillicrap2015}, SAC \cite{haarnoja2018} and PPO \cite{schulman2017}, still suffer from some major limitations. Specifically, existing research works showed that the algorithm performance is highly sensitive to hyper-parameter settings and can vary substantially in different algorithm runs \cite{paine2020}. \emph{Ineffective exploration} is often considered as a major cause for the poor learning stability \cite{chan2019}, often resulting in overfitting and premature convergence to poor local optima \cite{kurutach2018}. Rather than relying on one learner (or DRL agent), an ensemble of base learners can be jointly utilized to boost exploration and stabilize the learning process \cite{osband2016deep,osband2017post}. For example, the \emph{ensemble deep deterministic policy gradient} (ED2) algorithm is a newly developed ensemble DRL method \cite{januszewski2021} that trains multiple DNN policies simultaneously using a shared \emph{experience replay buffer} (ERB), similar to several previously proposed parallel DRL algorithms \cite{barth2018,mnih2016asyn}. ED2 features a unique mixture of multiple well-studied tricks, including temporally-extended deep exploration, double Q-bias reduction, and target policy smoothing \cite{osband2016deep,osband2017post,van2016deep,fujimoto2018}. It was reported to outperform state-of-the-art ensemble DRL algorithms such as SUNRISE \cite{lee2021sunrise} on several difficult Mujoco benchmark control problems. As far as we know, most of the existing ensemble DRL algorithms are designed to train each base learner individually. For example, in ED2, every base learner trains its own DNN policy using its own critic, with the aim to improve its own performance without considering the impact of the trained policy on the ensemble. While sharing the same ERB, policy training is conducted largely independently by all base learners. This is shown to promote healthy exploration in \cite{januszewski2021}. However, there is no guarantee that the base learners will collaborate effectively such that the ensemble as a whole can achieve desirable performance. To address this limitation, we propose a new \emph{hierarchical approach} for training base learners in this paper. Specifically, we follow ED2 for \emph{low-level training} of DNN policies, which will be performed concurrently by all base learners. In the meantime, we construct a \emph{global critic}, which is trained constantly to predict the performance of the ensemble. Guided by the global critic, \emph{high-level training} of DNN policies will be performed regularly to strengthen cooperation among all the base learners. Since the ensemble is not used directly to collect state-transition samples from the learning environment, we must make sure that high-level training of the ensemble is not performed on \emph{out-of-distribution} data obtained by individual base learners. In view of this, it is important to encourage \emph{inter-learner parameter sharing} so that the DNN policy trained by one base learner can contribute directly to (or influence) the training of DNN policies by other base learners. For this purpose, we develop a new technique in this paper for high-level training of policies based on the \emph{multi-step integration methods} for solving \emph{ordinary differential equations} (ODEs) in \cite{scieur2017}. Our high-level policy training technique is theoretically justified as it guarantees stability for a wide range of optimization problems. Meanwhile, it can be shown analytically that the trained linear parametric policies (a special and important technique for policy approximation) of all base learners are expected to behave more consistently as the ensemble through high-level policy training, encouraging inter-learner collaboration and alleviating the data distribution issue. Driven by the hierarchical policy training method, we develop a new ensemble DRL algorithm called \emph{hierarchical ensemble deep deterministic policy gradient} (HED) in this paper. Experimental evaluation of HED has been conducted on a range of benchmark control problems, including the widely used Mujoco control tasks as well as the less popular and potentially more challenging PyBullet control problems. Our experiments clearly show that HED can outperform ED2, SUNRISE and several cutting-edge DRL algorithms on multiple benchmark problems. \iffalse The rest of this paper is organized as follows. We review the related research works in the recent literature in the subsequent section. Necessary technical background for developing the HED algorithm is presented next, followed by the detailed design of HED. The empirical performance of the new algorithm is further evaluated before the concluding remarks. \fi \section{Related Work} \label{sec-rw} Similar to ED2, HED trains an ensemble of policies using an \emph{off-policy} DRL algorithm to leverage on the algorithm's advantages in \emph{sample efficiency}. Recently, several off-policy DRL algorithms have been developed successfully for RL in continuous spaces, including DDPG \cite{lillicrap2015}, SAC \cite{haarnoja2018}, TD3 \cite{fujimoto2018}, and SOP \cite{wang2020}. These algorithms introduce a variety of tricks to stabilize the learning process. For example, TD3 extends the idea of double Q-network \cite{van2016deep} to a new double-Q bias reduction technique, which can effectively prevent over-optimistic training of DNN policies. In addition, empirical evidence showed that the learning process becomes more stable when the actor and critic in TD3 are trained with different frequencies or in different phases \cite{fujimoto2018,cobbe2021}. The base learners in our HED ensemble will adopt these tricks. The recent literature also provides some new tricks to stabilize learning. Specifically, various trust-region methods have been developed to prevent negative behavioral changes during policy training \cite{kurutach2018,schulman2015,shani2020,wu2017,schulman2017}. Meanwhile, entropy regularization techniques prohibit immature convergence of the trained policies and ensure prolonged profitable exploration \cite{chen2018,haarnoja2018}. However, these techniques are mainly applied to stochastic policies while we aim at learning an ensemble of deterministic policies. Previous research showed that deterministic policies can often be trained more efficiently than stochastic policies using the \emph{reparameterization trick} \cite{fujimoto2018,silver2014,baek2020}. The stability of a DRL algorithm depends critically on how the learner explores its environment. Besides the entropy regularization methods, curiosity metrics are popularly employed to encourage a learner to explore rarely visited states during RL \cite{reizinger2020,zhelo2018}. Meanwhile, many previous studies embraced the \emph{optimum in the face of uncertainty} (OFU) principle to design bonus rewards for actions with high potentials, thereby promoting exploration in promising areas of the learning environment \cite{bellemare2016}. One good example is the UCB exploration technique developed in \cite{chen2017ucb,lee2021sunrise}. However, in \cite{januszewski2021}, this technique was shown to be less effective than the bootstrap with random initialization trick adopted in ED2. Temporally-extended exploration on RL problems with continuous actions can also be achieved by adding a small amount of noise to DNN weights \cite{plappert2017}. This is directly related to the posterior sampling methods that are often used to select the best actions among a statistically plausible set of sampled actions \cite{osband2018}. Following the OFU principle, deep ensembles have been recently proposed to approximate Bayesian posteriors with high accuracy and efficiency \cite{lakshminarayanan2016}. They are subsequently used to approach deep exploration for reliable RL \cite{osband2016deep}. Several issues have been investigated under the context of ensemble DRL. For instance, the diversity of base learners is essential to the performance of the ensemble. To encourage diversity, either different DRL algorithms or the same algorithm with differed hyper-parameter settings have been adopted to train base learners \cite{huang2017,wiering2008}. Meanwhile, proper aggregation of the action outputs from all base learners in an ensemble poses another challenge. Typical approaches to tackle this issue include taking the mean action as the output of the ensemble and choosing the action with highest predicted cumulative rewards \cite{januszewski2021,chen2019}. As far as we know, few existing ensemble DRL algorithms in the literature have ever studied the important issue on how to effectively train all base learners to jointly improve the ensemble performance. This issue will be explored in-depth with the newly developed HED algorithm in this paper. \section{Background} \label{sec-back} An RL problem is modeled as a \emph{Markov Decision Process} (MDP) $(\mathcal{S},\mathcal{A},R,P,\gamma,p_0)$, where $\mathcal{S}$ and $\mathcal{A}$ refer respectively to the continuous multi-dimensional state space and action space. $P$ stands for the state-transition model that governs the probability of reaching any state $s_{t+1}\in\mathcal{S}$ at timestep $t+1$ upon performing any action $a_t\in\mathcal{A}$ in state $s_t\in\mathcal{S}$ at timestep $t$, with $t\in\mathbb{Z}^+$. Additionally, $\gamma\in[0,1)$ is the discount factor, $R$ is the reward function, and $p_0$ captures the initial state distribution. To solve any RL problem described above, we aim to learn an optimal \emph{deterministic ensemble policy} $\pi^e_*(s)$ that maps any state input $s\in\mathcal{S}$ to an action vector $a\in\mathcal{A}$ so as to maximize the \emph{cumulative rewards} defined below $$ \pi^e_*= \argmax_{\pi^e} J(\pi^e)=\argmax_{\pi^e}\mathbb{E}_{\tau\sim\pi^e} \left[ \sum_{t=1}^{\infty}\gamma^{t-1} R(s_t,a_t) \right], $$ where $\tau=[(s_t,a_t,r_t,s_{t+1})]_{t=1}^{\infty}$ contains a series of consecutive state-transition samples and is called a \emph{episode}, which can be obtained by following the ensemble policy $\pi^e$, and $r_t=R(s_t,a_t)$ is the immediate reward received at timestep $t$ in $\tau$. For an ensemble with $N$ base learners where each base learner $L_i$, $1\leq i\leq N$, maintains its own deterministic base policy $\pi^i$, the action output of $\pi^e$ is jointly determined by all the \emph{base policies} according to \begin{equation} \forall s\in\mathcal{S}, \pi^e(s)=\frac{1}{N}\sum_i^N \pi^i(s). \label{equ-pe} \end{equation} In order to train an ensemble to maximize the cumulative rewards, our baseline algorithm ED2 uses randomly selected base learners to sample a series of episodes $\{\tau_i\}$, which will be stored in the shared ERB. At regular time intervals, a mini-batch of state-transition samples will be retrieved from the ERB. Every base learner $L_i$ will then use the retrieved mini-batch to train its own actor $\pi^i$ and critic $Q^i$ individually. In other words, a base learner manages two separate DNNs, one models the deterministic policy $\pi^i$ and the other approximates the Q-function $Q^i$ of $\pi^i$. A base learner uses an existing actor-critic RL algorithm to train the two DNNs. In this paper, we choose TD3 for this purpose due to its proven effectiveness, high popularity and stable learning behavior \cite{fujimoto2018}. \section{Hierarchical Ensemble Deep Deterministic Policy Gradient} \label{sec-algo} The pseudo-code of the HED algorithm is presented in Algorithm \ref{alg-code}. HED follows many existing works including ED2 \cite{osband2016deep,januszewski2021} to achieve temporally-extended exploration through bootstrapping with random initialization of DNN policies. As clearly shown in \cite{januszewski2021}, this exploration technique is more effective than UCB and parameter randomization methods. Different from ED2 which completely eliminates the necessity of adding small random noises to the deterministic action outputs from the DNN policies, we keep a small level of action noise\footnote{The noise is sampled from the Normal distribution independently for each dimension of the action vector. The variance of the normal distribution is fixed at 0.01 during the learning process.} while using any chosen policy to explore the learning environment. We found empirically that this ensures coherent exploration, similar to \cite{osband2016deep}, while making the testing performance of the trained policies more stable. Different from ED2 and other ensemble algorithms for RL in continuous spaces, HED trains DNN policies at two separate levels. The low-level training of $\pi^i$ and $Q^i$ by each base learner $L_i$ is essentially the same as ED2 and TD3. Specifically, for any base learner $L_i$, $i\in\{1,\ldots,N\}$, $Q^i$ is trained by $L_i$ to minimize $MSE_i$ below \begin{equation} MSE_i=\frac{1}{|\mathcal{B}|} \sum_{(s,a,r,s')\in\mathcal{B}}\left(Q^i_{\phi_i}(s,a)-r-\gamma\min_{k=1,2}\hat{Q}^i_{k}(s',\pi^i(s')+\epsilon) \right)^2, \label{equ-mse} \end{equation} where $\phi_i$ represents the trainable parameters of the DNN that approximates $Q^i$. $\mathcal{B}$ is the random mini-batch retrieved from the ERB. $\hat{Q}^i_{k}$ with $k=1,2$ stands for the two target Q-networks of $L_i$ that together implement the double-Q bias reduction mechanism proposed in \cite{fujimoto2018}. Additionally, $\epsilon$ is a random noise sampled from a Normal distribution with zero mean and small variance\footnote{The variance for sampling $\epsilon$ is kept at a very small level of 0.01 in the experiments.}. Using the trained $Q^i$, the trainable parameters $\theta_i$ of the DNN that models policy $\pi^i$ is further updated by $L_i$ along the \emph{policy gradient} direction computed below \begin{equation} \nabla_{\theta_i}J(\pi^i_{\theta_i})=\frac{1}{|\mathcal{B}|}\sum_{s\in\mathcal{B}} \nabla_{a}Q^i(s,a)|_{a=\pi^i_{\theta_i}(s)}\nabla_{\theta_i}\pi^i_{\theta_i}(s). \label{equ-pg} \end{equation} Besides the above, HED constantly trains a separate high-level Q-function $Q^e$ to predict the performance of the ensemble policy $\pi^e$. Guided by the trained $Q^e$, high-level policy training is conducted regularly to update policy $\pi^i$ of all base learners so as to enhance their cooperation and performance. A new \emph{multi-step technique} is developed in HED to enable inter-learner parameter sharing during high-level policy training. To implement this technique, we keep track of a list of bootstrap policy parameters for the multi-step training process. More details can be found in the subsequent subsection. Theoretical justifications regarding the usefulness of the multi-step approach are also provided below. \begin{algorithm}[!ht] \begin{algorithmic} \STATE {\bf Input}: Ensemble size $N$; initial policy networks $\pi^i_{\theta_i}$ and Q-networks $Q^i_{\phi_i}$ for $i\in\{1,\ldots,N\}$; ERB; ensemble Q-network $Q^e_{\phi_e}$; target Q-networks for each base learner and the ensemble \STATE {\bf Output}: Trained ensemble policy $\pi^e$ \STATE {\bf While} the total number of sampled trajectories $<$ max number of trajectories: \STATE \ \ \ \ Randomly sample $i\in\{1,\ldots,N\}$ \STATE \ \ \ \ {\bf While} the current trajectory does not terminate: \STATE \ \ \ \ \ \ \ \ Use $\pi^i$ to perform the next action \STATE \ \ \ \ \ \ \ \ Store sampled state-transition in ERB \STATE \ \ \ \ \ \ \ \ Track number of steps sampled before critic training \STATE \ \ \ \ \ \ \ \ {\bf If} time for critic training: \STATE \ \ \ \ \ \ \ \ \ \ \ \ {\bf For} number of steps sampled: q\STATE \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ Sample a mini-batch $\mathcal{B}$ from ERB \STATE \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ Train $Q^i_{\phi_i}$ for $i\in\{1,\ldots,N\}$ using \eqref{equ-mse} \STATE \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ Train $Q^e_{\phi_e}$ using \eqref{equ-e-mse} \STATE \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ {\bf If} time for {\bf low-level} policy training: \STATE \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ Train $\pi^i_{\theta_i}$ for $i\in\{1,\ldots,N\}$ using \eqref{equ-pg} \STATE \ \ \ \ \ \ \ \ {\bf If} time for {\bf high-level} policy training: \STATE \ \ \ \ \ \ \ \ \ \ \ \ Set bootstrap list $\{x_j\}_{j=0}^{2}$ for each base learner \STATE \ \ \ \ \ \ \ \ \ \ \ \ {\bf For} a fraction of sampled steps: \STATE \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ Train $\pi^i_{\theta_i}$ for $i\in\{1,\ldots,N\}$ using \eqref{equ-mu-new} \STATE \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ Append trained $\theta_i$ for $i\in\{1,\ldots,N\}$ to the bootstrap lists of each base \STATE \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ learner for the next step of high-level policy training \end{algorithmic} \caption{The pseudo-code of the HED algorithm.} \label{alg-code} \end{algorithm} \subsection{A multi-step technique for high-level policy training} \label{subsec-multipol} In addition to $Q^i$ for each base learner $L_i$, $i\in\{1,\ldots,N\}$, HED maintains a separate Q-network to approximate $Q^e$ of the ensemble policy $\pi^e$. Similar to \eqref{equ-mse}, HED trains this central Q-network towards minimizing $MSE_e$ below \begin{equation} MSE_e=\frac{1}{|\mathcal{B}|} \sum_{(s,a,r,s')\in\mathcal{B}}\left(Q^e_{\phi_e}(s,a)-r-\gamma\hat{Q}^e(s',\pi^e(s')) \right)^2, \label{equ-e-mse} \end{equation} with $\phi_e$ representing the trainable parameters of the central Q-network. $\hat{Q}^e$ stands for the corresponding target Q-network that stabilizes the training process. For simplicity, we do not add random noise $\epsilon$ in \eqref{equ-mse} to the action outputs produced by the ensemble policy $\pi^e$ in \eqref{equ-e-mse}. Furthermore, following \cite{van2016deep}, one target Q-network instead of two is adopted in \eqref{equ-e-mse} to facilitate the training of $Q^e$. Building on the trained $Q^e$, we can calculate the \emph{ensemble policy gradient} with respect to $\theta_i$ of every base learner $L_i$ as follows \begin{equation} \nabla_{\theta_i}J(\pi^e)= \frac{1}{|\mathcal{B}|}\sum_{s\in\mathcal{B}} \nabla_{a}Q^e(s,a)|_{a=\pi^e(s)}\nabla_{a_i}\pi^e(s)|_{a_i=\pi^i_{\theta_i}(s)}\nabla_{\theta_i}\pi^i_{\theta_i}(s), \label{equ-e-pg} \end{equation} with $$ \nabla_{a_i}\pi^e(s)|_{a_i=\pi^i_{\theta_i}(s)}=\frac{1}{N} I, $$ according to \eqref{equ-pe}. $I$ stands for the $m\times m$ identity matrix where $m$ is the dimension of the action vector. One straightforward approach for high-level policy training is to update $\theta_i$ of every base learner $L_i$ in the direction of \eqref{equ-e-pg}. However, using \eqref{equ-e-pg} alone may not encourage any base learner $L_i$ to behave consistently with the ensemble (see Proposition \ref{the-2}). Consequently, high-level training of the ensemble policy may be performed on the out-of-distribution state-transition samples collected by the base learners, affecting the training effectiveness. Furthermore, ensembles are used mainly for temporally-extended exploration in the literature. The learning activity of one base learner can only influence other base learners indirectly through the shared ERB. Base learners do not explicitly share their learned policy parameters in an ensemble to strengthen inter-learner cooperation and boost the learning process. To address this limitation, we propose to promote inter-learner parameter sharing during high-level policy training, in order to achieve a desirable balance between exploration and inter-learner cooperation. Specifically, in addition to \eqref{equ-e-pg}, we randomly select \emph{two base learners} $L_p$ and $L_q$ and use their policy parameters to guide the training of policy $\pi^i$ of any base learner $L_i$. In comparison to \emph{selecting one base learner}, this allows more base learners to have the opportunity to share their parameters with the base learner $L_i$ during policy training. It is also possible to recruit more than two base learners. However, in this case, it is mathematically challenging to derive stable learning rules for high-level policy training. Motivated by the above discussion, a search through the literature leads us to the linear multi-step integration methods recently analyzed in \cite{scieur2017}. Consider a simple \emph{gradient flow equation} below \begin{equation} x(0)=\theta_i^0, \frac{\partial x(t)}{\partial t}=g(x(t))=\nabla_{\theta_i}J(\pi^e)|_{\theta_i=x(t)}, \label{equ-gfe} \end{equation} where $\theta_i^0$ refers to the initial policy parameter of base learner $L_i$. If $J(\pi^e)$ is strongly concave and Lipschitz continuous, the solution of \eqref{equ-gfe} allows us to obtain the optimal policy parameters $\theta_i^*$ when $x(t)$ approaches to $\infty$. Since $J(\pi^e)$ is not strongly concave for most of real-world RL problems, $x(t)$ in practice may only converge to a locally optimal policy, which is common among majority of the policy gradient DRL algorithms. Therefore high-level training of policy $\pi^e$ and hence $\pi^i$ can be approached by numerically solving \eqref{equ-gfe}. This can be achieved through a linear $\mu$-step method shown below \begin{equation} x_{k+\mu}=-\sum_{j=0}^{\mu-1}\rho_j x_{k+j}+h\sum_{j=0}^{\mu-1}\sigma_j g(x_{k+j}), \forall k\geq 0, \label{equ-mu-step} \end{equation} where $\rho_j,\sigma_j\in\mathbb{R}$ are the pre-defined coefficients of the multi-step method and $h$ is the learning rate. Clearly, each new point $x_{k+\mu}$ produced by the $\mu$-step method is a function of the preceding $\mu$ points. In this paper, we specifically consider the case when $\mu=3$. Meanwhile, let \begin{equation} x_0=\theta_p,x_1=\theta_q,x_2=\theta_i \label{equ-boot-list} \end{equation} where $p$ and $q$ are the randomly generated indices of two base learners and $i$ is the index of the base learner whose policy $\pi^i$ is being trained by the $\mu$-step method. Through this way, the training of policy $\pi^i$ is influenced directly by base learners $L_p$ and $L_q$ through explicit inter-learner parameter sharing. $x_i (i\geq 3)$ in \eqref{equ-mu-step} represents the trained policy parameters of $\pi^i$ in subsequent training steps. Although \eqref{equ-mu-step} allows us to use $\nabla_{\theta_p}J(\pi^e)$ and $\nabla_{\theta_q}J(\pi^e)$ to train $\theta_i$, they do not seem necessary for inter-learner parameter sharing. To simplify \eqref{equ-mu-step}, we set $\sigma_0=\sigma_1=0$ and $\sigma_2=1$. Hence only $g(x_{k+2})$, which is the ensemble policy gradient with respect to policy $\pi^i$ in \eqref{equ-e-pg}, is used to train $\pi^i$. With this simplification, we derive the new learning rule for high-level policy training below \begin{equation} \begin{split} & x_{k+3}= -\rho_2 x_{k+2}-\rho_1 x_{k+1}-\rho_0 x_k+h\cdot \nabla_{\theta_i}J(\pi^e)|_{\theta_i=x_{k+2}}, \forall k\geq 0 \\ & x_0=\theta_p,x_1=\theta_q,x_2=\theta_i \end{split} \label{equ-mu-new} \end{equation} To implement \eqref{equ-mu-new} in HED, before high-level policy training, every base learner $L_i$ must set up a \emph{bootstrap list} of policy parameters $\{x_0=\theta_p,x_1=\theta_q,x_2=\theta_i\}$. After the $k$-th training step ($k\geq 0$) based on \eqref{equ-mu-new}, $L_i$ appends the trained $\theta_i$ as $x_{k+3}$ to the bootstrap list, which will be utilized to train $\pi^i$ in the subsequent training steps. Reliable use of \eqref{equ-mu-new} demands for careful parameter settings of $\rho_0$, $\rho_1$, $\rho_2$ and $h$. Relevant theoretical analysis is presented below. \subsection{Theoretical analysis of the multi-step policy training technique} \label{subsec-multithe} In this subsection, a theoretical analysis is performed first to determine suitable settings of $\rho_0$, $\rho_1$, $\rho_2$ and $h$ for stable high-level policy training through \eqref{equ-mu-new}. To make the analysis feasible, besides the strongly concave and Lipschitz continuous conditions, we further assume that \begin{equation} \nabla_{\theta_i}J(\pi^e)\approx -A(\theta_i-\theta_i^*) \label{equ-grad-linear} \end{equation} where $A$ is a positive definite matrix whose eigenvalues are bounded positive real numbers. $\theta_i^*$ stands for the global-optimal (or local-optimal) policy parameters. Many strongly concave functions satisfy this assumption \cite{scieur2017}. Meanwhile, the attraction basin of the local optimum of many multi-modal optimization problems often satisfies this assumption too. Using this assumption, we can derive Proposition \ref{the-1} below. \begin{proposition} Upon using \eqref{equ-mu-new} to numerically solve \eqref{equ-gfe}, the following conditions must be satisfied for $x_k$ to converge to $\theta_i^*$ as $k$ approaches to $\infty$: \begin{enumerate} \item $\rho_2=\rho_0-1$, $\rho_1=-2\rho_0$; \item $0<\rho_0<\frac{1}{2}$; \item $h$ is reasonably small such that $0\leq \lambda h < 2-4\rho_0$, where $\lambda$ can take any real value between the smallest and the largest eigenvalues of the positive definite matrix $A$ in \eqref{equ-grad-linear}. \end{enumerate} \label{the-1} \end{proposition} The proof of Proposition \ref{the-1} can be found in Appendix A. Proposition \ref{the-1} provides suitable parameter settings for \eqref{equ-mu-new} and justifies its stable use for high-level policy training. We next show that \eqref{equ-mu-new} is also expected to make base learners behave more consistently with the ensemble, without affecting the behavior of the trained ensemble, when $\rho_0$ is sufficiently small. Consider specifically that each base learner $L_i$ trains a linear parametric policy of the form: \begin{equation} \pi^i(s)=\Phi(s)^T\cdot \theta_i \label{equ-lin-pol} \end{equation} where $\Phi(s)$ represents the \emph{state feature vector} with respect to any input state $s$. For simplicity, we study the special case of scalar actions. However, the analysis can be easily extended to high-dimensional action spaces. Meanwhile, we use $Sin()$ and $Mul()$ to represent respectively the action output of a policy trained for one iteration on the same state $s$ by using either the single-step method or the multi-step method in \eqref{equ-mu-new}. The single-step method can be considered as a special case of the multi-step method with $\rho_2=-1$ and $\rho_0=\rho_1=0$. Using these notations, Proposition \ref{the-2} is presented below. \begin{proposition} When each base learner $L_i$, $i\in\{1,\ldots,N\}$, trains its linear parametric policy $\pi^i$ with policy parameters $\theta_i$ on any state $s\in\mathcal{S}$ and when $0<\rho_0<\frac{1}{3}$, \begin{enumerate} \item $Sin(\pi^e(s))=\mathbb{E}\left[ Mul(\pi^e(s)) \right]$; \item $\sum_{i\in\{1,\ldots,N\}}\mathbb{E}\left[ \left( Mul(\pi^i(s))-\mathbb{E}\left[ Mul(\pi^e(s)) \right] \right)^2 \right] \\ < \sum_{i\in\{1,\ldots,N\}}\left( Sin(\pi^i(s))-Sin(\pi^e(s)) \right)^2 \\ =\sum_{i\in\{1,\ldots,N\}}\left( \pi^i(s)-\pi^e(s) \right)^2$ \end{enumerate} where the expectations above are taken with respect to any randomly selected $p,q\in\{1,\ldots,N\}$ in \eqref{equ-boot-list}. \label{the-2} \end{proposition} Proposition \ref{the-2} indicates that multi-step training in \eqref{equ-mu-new} is expected to reduce the difference between the action output of any base learner and that of the ensemble. Meanwhile the amount of action changes applied to $\pi^e$ remains identical to the single-step method. Therefore, using the multi-step policy training method developed in this section helps to enhance consistent behaviors among all base learners of the ensemble. \begin{table*}[!ht] \caption{Final performance of all competing algorithms on 9 benchmark problems. The results are shown with mean cumulative rewards and standard deviation over 10 independent algorithm runs. For each run, the cumulative rewards are obtained by averaging over 50 independent testing episodes.} \centering \resizebox{1.0\textwidth}{!}{ \begin{tabular}{c||cccc|c} \hline & TD3 & SAC & ED2 & SUNRISE & HED \\ \hline Hopper-v0 (PyBullet) & 917.38$\pm$178.46 & 1365.47$\pm$281.3 & 2095.54$\pm$148.86 & 1976.86$\pm$311.24 & \textbf{2359.63$\pm$50.28} \\ InvertedDoublePendulum-v0 (PyBullet) & 4394.07$\pm$558.8 & 8664.3$\pm$187.83 & 8733.3$\pm$1490.51 & 8746.82$\pm$58.68 & \textbf{9351.56$\pm$13.65} \\ InvertedPendulum-v0 (PyBullet) & 484.08$\pm$49.64 & 937.33$\pm$8.0 & \textbf{999.51$\pm$1.0} & 933.58$\pm$9.03 & 995.96$\pm$7.56 \\ Reacher-v0 (PyBullet) & 8.42$\pm$1.07 & 17.43$\pm$0.66 & 15.36$\pm$2.24 & \textbf{17.65$\pm$0.44} & 17.35$\pm$0.82 \\ Hopper-v3 (Mujoco) & 1399.54$\pm$250.32 & 2369.61$\pm$906.85 & 3043.19$\pm$971.53 & 2913.56$\pm$475.48 & \textbf{3503.08$\pm$83.35} \\ Humanoid-v3 (Mujoco) & 458.77$\pm$116.08 & 1720.54$\pm$525.07 & 868.17$\pm$384.58 & 3614.89$\pm$1402.7 & \textbf{3925.81$\pm$1029.59} \\ LunarLanderContinuous-v2 & 254.98$\pm$17.86 & 254.03$\pm$27.52 & 268.37$\pm$6.9 & \textbf{278.76$\pm$2.34} & 278.23$\pm$3.51 \\ Swimmer-v3 (Mujoco) & 68.52$\pm$33.54 & 40.37$\pm$1.01 & 87.81$\pm$28.79 & 49.45$\pm$0.47 & \textbf{89.23$\pm$26.78} \\ Walker2D-v3 (Mujoco) & 1965.25$\pm$248.21 & 1503.35$\pm$818.87 & 3298.73$\pm$1282.64 & 3766.55$\pm$1063.0 & \textbf{4442.8$\pm$408.88} \\ \hline \end{tabular}} \label{tab:final_perf_comp} \end{table*} \begin{figure*}[!hbt] \begin{center} \subfloat[Hopper-v0 (PyBullet)]{\label{fig-hc}\includegraphics[width=0.33\linewidth]{exp/HopperPyBulletEnv-v0_SR.pdf}} \subfloat[InvertedDoublePendulum-v0 (PyBullet)]{\label{fig-idpPB}\includegraphics[width=0.33\linewidth]{exp/InvertedDoublePendulumPyBulletEnv-v0_SR.pdf}} \subfloat[InvertedPendulum-v0 (PyBullet)]{\label{fig-ipPB}\includegraphics[width=0.33\linewidth]{exp/InvertedPendulumPyBulletEnv-v0_SR.pdf}}\\ \subfloat[Reacher-v0 (PyBullet)]{\label{fig-reacher}\includegraphics[width=0.33\linewidth]{exp/ReacherPyBulletEnv-v0_SR.pdf}} \subfloat[Hopper-v3 (Mujoco)]{\label{fig-hopper}\includegraphics[width=0.33\linewidth]{exp/Hopper-v3_SR.pdf}} \subfloat[Humanoid-v3 (Mujoco)]{\label{fig-humanoid}\includegraphics[width=0.33\linewidth]{exp/Humanoid-v3_SR.pdf}}\\ \subfloat[LunarLanderContinuous-v2]{\label{fig-lunarLander}\includegraphics[width=0.33\linewidth]{exp/LunarLanderContinuous-v2_SR.pdf}} \subfloat[Swimmer-v3 (Mujoco)]{\label{fig-swimmer}\includegraphics[width=0.33\linewidth]{exp/Swimmer-v3_SR.pdf}} \subfloat[Walker2D-v3 (Mujoco)]{\label{fig-walker}\includegraphics[width=0.33\linewidth]{exp/Walker2d-v3_SR.pdf}}\\ \end{center} \caption{Learning curves of HED and four baseline algorithms (i.e., TD3, SAC, ED2 and SUNRISE) on 9 benchmark RL problems.} \label{fig:training_perf} \end{figure*} \section{Experiment} \label{sec-exp} This section presents the experimental evaluation of HED, in comparison to several state-of-the-art DRL algorithms. The experiment setup is discussed first. Detailed experiment results are further presented and analyzed. \subsection{Experiment setting} We implement HED based on the high-quality implementation of TD3 provided by the publicly available OpenAI Spinning Up repository~\cite{SpinningUp2018}. We also follow closely the hyper-parameter settings of TD3 recommended in \cite{fujimoto2018} to build each base learner of HED. Specifically, based on~\cite{schulman2017}, a fully connected MLP with two hidden layers of 64 ReLU units is adopted to model all policy networks and Q-networks. Similar to \cite{januszewski2021,lee2021sunrise}, HED employs $5$ base learners, i.e., $N=5$. Each base learner has its own policy network and Q-network. Meanwhile, HED maintains and trains a separate ensemble Q-network with the same network architecture design. Each base learner trains its Q-network and also conducts the low-level training of the policy network repeatedly whenever HED collects 50 consecutive state-transition samples from the learning environment. Meanwhile, high-level policy training as well as the training of the ensemble Q-network is performed immediately after HED samples a full episode. HED adopts a separate Adam optimizer with the fixed learning rate of $5\mathrm{e}{-4}$ to train each Q-network and policy network. Furthermore, $\rho_0$ in \eqref{equ-mu-new} is set to 0.0001 for the main experiment results reported in Figure \ref{fig:training_perf}. The mini-batch size $|\mathcal{B}|$ is set to 100, following existing research \cite{januszewski2021} without any fine-tuning. HED is compared against four state-of-the-art DRL algorithms, including two Ensemble DRL algorithms, i.e., ED2~\cite{januszewski2021} and SUNRISE \cite{lee2021sunrise}), and two widely used off-policy DRL algorithms, i.e., SAC \cite{haarnoja2018} and TD3 \cite{fujimoto2018}. We evaluate their performance on 9 challenging continuous control benchmark problems, including four PyBullet benchmark problems \cite{benelot2018} (i.e., Hopper-v0, InvertedDoublePendulum-v0, InvertedPendulum-v0, and Reacher-v0), four Mujoco control tasks (i.e., Hopper-v3, Humanoid-v3, Swimmer-v3, and Walker2D-v3), and LunarLanderContinuous-v2 provided by OpenAI Gym \cite{openai_gym} . In the literature, PyBullet benchmarks are often considered to be more challenging than Mujoco benchmarks. Hence we decide to evaluate the performance of HED on both PyBullet and Mujoco benchmarks. The maximum episode length for each benchmark is fixed to 1000 timesteps. Each algorithm is run independently with 10 random seeds on all benchmarks. Besides the hyper-parameter settings of HED highlighted above, detailed hyper-parameter settings of all the competing algorithms have been summarized in Appendix~C. \subsection{Experiment result} \subsubsection{Performance comparison:} We compare HED against four cutting-edge DRL algorithms on 9 benchmark problems. Table~\ref{tab:final_perf_comp} presents the average cumulative rewards obtained by the policy networks (or policy ensembles for ensemble DRL algorithms) trained by all the competing algorithms across the same number of sampled episodes with respect to each benchmark. As evidenced in the table, HED achieved consistently the best performance\footnote{HED significantly outperformed ED2 on most benchmark problems, thanks to its use of the proposed high-level policy training technique.} on most of the benchmark problems except InvertedPendulum, Reacher, and LunarLander. Meanwhile, on InvertedPendulum, Reacher, and LunarLander, HED achieved very competitive performance with at least 98\% of the maximum cumulative rewards reached by the highest performing competing algorithms. Furthermore, on some problems such as Hopper-v0 and Walker2D-v3, HED outperformed the lowest performing algorithm by up to 150\% and the algorithm with the second highest performance by up to 18\%. In addition to Table~\ref{tab:final_perf_comp}, we also compared the learning curves of all the competing algorithms in Figure \ref{fig:training_perf}. As demonstrated in this figure, by explicitly strengthening inter-learner collaboration, HED converges clearly faster and is more stable during the learning process than other competing algorithms. Specifically, on several benchmark problems, such as Hopper-v0, InvertedDoublePendulum-v0, and Hopper-v3, HED achieved significantly lower variations in learning performance across 10 independent runs. In comparison to other ensemble DRL algorithms, the learning curves of HED also appear to be smoother on several benchmark problems, such as Walker2D-v3, suggesting that HED can achieve highly competitive stability during learning. \subsubsection{Impact of $\rho_0$:} To investigate the performance impact of $\rho_0$, we tested 4 different settings of $\rho_0$, ranging from $5\mbox{e$-$}05$ to $0.01$, on the InvertedPendulum-v0 and Hopper-v0 problems (similar observations can be found on other benchmark problems and are omitted in this paper). The learning curves are plotted in Figure~\ref{fig:rho_impact}. It is witnessed in the figure that HED can convergence faster under suitable settings of $\rho_0$. However, the ``best'' $\rho_0$ varies on different benchmark problems. For example, $\rho_0=0.005$ (green curve) converges slightly faster than other settings on InvertedPendulum-v0 while $\rho_0=5e-05$ (blue curve) converges slightly faster on Hopper-v0. Nevertheless, the impact of different $\rho_0$ on the final performance appears to be small as long as $\rho_0$ is reasonably small according to Proposition \ref{the-1}. \begin{figure} \begin{center} \subfloat[InvertedPendulum-v0]{\label{fig-invertedPendulum-rho}\includegraphics[width=0.4\linewidth]{exp/rho_impact/InvertedPendulumPyBulletEnv-v0_rho.pdf}} \subfloat[Hopper-v0]{\label{fig-hp-rho}\includegraphics[width=0.4\linewidth]{exp/rho_impact/HopperPyBulletEnv-v0_rho.pdf}} \end{center} \caption{The impact of using different $\rho_0$ in \eqref{equ-mu-new} on the performance of HED. $\rho_1$ and $\rho_2$ in \eqref{equ-mu-new} depend directly on $\rho_0$ according to Proposition \ref{the-1}.} \label{fig:rho_impact} \end{figure} \subsubsection{Ablation study on high-level policy training techniques:} High-level policy training can be conducted repeatedly whenever HED obtains either a new sampled episode or a fixed number of consecutive state-transition samples (e.g., samples collected from 50 consecutive timesteps). To understand which approach is more effective, experimental comparisons have been conducted in Appendix D with detailed performance results. According to the experiment results in Appendix D, on a majority of benchmark problems, episodic learning can produce more stable learning behavior and also makes HED converge faster. We also compared HED with its variation that performs high-level policy training by using the single-step method in \eqref{equ-e-pg} instead of the multi-step method in \eqref{equ-mu-new}. Detailed experiment results can be found in Appendix E. Our experiment results confirm that multi-step training in \eqref{equ-mu-new} enables HED to achieve significantly higher performance and learning stability than using the conventional single-step training technique in \eqref{equ-e-pg}. Hence, by explicitly sharing learned policy parameters among base learners in an ensemble through \eqref{equ-mu-new}, HED can effectively enhance inter-learner collaboration and boost the learning process. \section{Conclusions} \label{sec-con} In this paper, we conducted in-depth study of ensemble DRL algorithms, which have achieved cutting-edge performance on many benchmark RL problems in the recent literature. Different from existing research works that rely mainly on each base learner of an ensemble to train its policy network individually, we developed a new HED algorithm to explore the potential of training all base learners in a hierarchical manner in order to promote inter-learner collaboration and improve the collective performance of an ensemble of trained base learners. Specifically, we adopted existing ensemble DRL algorithms such as ED2 to perform low-level policy training. Meanwhile, a new multi-step training technique was developed for high-level policy training in HED to facilitate direct inter-learner parameter sharing. Both theoretical and empirical analysis showed that the HED algorithm can achieve stable learning behavior. It also outperformed several state-of-the-art DRL algorithms on multiple benchmark RL problems. \section*{Appendix A} This appendix presents a proof of Proposition \ref{the-1}. According to \cite{scieur2017}, any multi-step integration methods including \eqref{equ-mu-new} must satisfy three conditions to ensure its stability. They together guarantee that $x_k$ can converge to $\theta_i^*$ as $k$ approaches to $\infty$. We check each condition one-by-one below to derive the main conclusions in Proposition \ref{the-1}. \vspace{0.2cm} \noindent \textbf{Consistency condition}: We can re-write \eqref{equ-mu-new} as below $$ x_{k+3}+\rho_2 x_{k+2}+\rho_1 x_{k+1}+\rho_0 x_k= h\cdot g(x_{k+2}). $$ Define the \emph{shift operator} $F$, which maps $Fx_k\rightarrow x_{k+1}$. Furthermore, with $g(x_k)$ being simplified to $g_k$, $F$ also maps $Fg_k\rightarrow g_{k+1}$. Using $F$, \eqref{equ-mu-new} can be further written as $$ \rho(F)x_k=h\sigma(F)g_k, \forall k\geq 0, $$ where $$ \rho(F)=F^3+\rho_2 F^2+\rho_1 F+\rho_0, \sigma(F)=F^2. $$ The consistency condition requires $$ \rho(1)=0, \rho'(1)=\sigma(1). $$ This implies that \begin{equation*} \begin{split} & 1+\rho_2+\rho_1+\rho_0=0, \\ & 3+2\rho_2+\rho_1=1. \end{split} \end{equation*} Solving the above equations leads to $$ \rho_1=-2\rho_0,\ \rho_2=\rho_0-1. $$ Hence, \eqref{equ-mu-new} becomes $$ x_{k+3}=(1-\rho_0)x_{k+2}+2\rho_0 x_{k+1} -\rho_0 x_k+h\cdot g(x_{k+2}). $$ \vspace{0.2cm} \noindent \textbf{Zero-stability condition}: This condition requires all roots of $\rho(F)$ to be in the unit disk. Any roots on the unit circle must be simple. In other words, $$ \left| Roots(\rho(F)) \right|\leq 1. $$ In fact, $\rho(F)$ has three roots. They are $$ 1, \frac{1}{2}\left(-\rho_0 \pm \sqrt{\rho_0(\rho_0+4)} \right). $$ It is easy to verify that when $0<\rho_0<\frac{1}{2}$, $$ \left| \frac{1}{2}\left(-\rho_0 - \sqrt{\rho_0(\rho_0+4)} \right) \right|<1. $$ Meanwhile, when $\rho_0>0$, $$ \left| \frac{1}{2}\left(-\rho_0 + \sqrt{\rho_0(\rho_0+4)} \right) \right|<1. $$ In summary, the zero-stability condition requires $$ 0<\rho_0<\frac{1}{2}. $$ \vspace{0.2cm} \noindent \textbf{Absolute stability condition}: Define \begin{equation*} \begin{split} \Pi_{\lambda h} & \overset{\Delta}{=}\rho(F)+\lambda h\sigma(F) \\ &=F^3+\rho_2 F^2+\rho_1 F+\rho_0+\lambda h F^2 \\ &=F^3+(\rho_0-1) F^2 - 2\rho_0 F+\rho_0 +\lambda h F^2 \\ &=F^3 + (\rho_0-1+\lambda h)F^2-2\rho_0 F+\rho_0. \end{split} \end{equation*} Further define $$ r_{max}=\max_{\lambda\in [L,U]} \max_{r\in Roots(\Pi_{\lambda h}(F))}|r|, $$ where $L$ and $U$ in this appendix refer respectively to the smallest and the largest positive eigenvalues of matrix $A$ in \eqref{equ-grad-linear}. The absolute stability condition requires \begin{equation} r_{max}<1. \label{equ-abs-con} \end{equation} Let \begin{equation*} \begin{split} & B=\rho_0-1+\lambda h, \\ & C=-2\rho_0, \\ & D=\rho_0. \end{split} \end{equation*} Subsequently, define \begin{equation*} \begin{split} & A_0=1-B+C-D=2-\lambda h-4\rho_0, \\ & A_1=3-B-C-3D=4-\lambda h -2\rho_0, \\ & A_2=3+B-C-3D=2+\lambda h, \\ & A_3=1+B+C+D=\lambda h. \end{split} \end{equation*} According to the Routh-Hurwitz criterion \cite{nise2020}, the following two conditions jointly guarantee \eqref{equ-abs-con}: \begin{equation*} \begin{split} A_1,A_2,A_3,A_4>0, \\ A_1 A_2 > A_0 A_3. \end{split} \end{equation*} Specifically, the first condition above gives rise to the following: $$ \lambda h >0,\ \lambda h+2\rho_0<4,\ \lambda h+4\rho_0<2. $$ Following the second condition above, we can deduce the below: $$ \lambda h>2-\frac{4}{\rho_0}. $$ Given that $\lambda h>0$, we have \begin{equation*} \begin{split} & \lambda h > \max\left\{ 0, 2-\frac{4}{\rho_0} \right\}, \\ & \lambda h < \min\left\{ 2-4\rho_0, 4-2\rho_0 \right\}. \end{split} \end{equation*} Since $0<\rho_0<\frac{1}{2}$, $$ 2-4\rho_0<4-2\rho_0,\ 2-\frac{4}{\rho_0}<0. $$ Consequently $$ 0<\lambda h<2-4\rho_0. $$ Clearly, with sufficiently small $h$, the above condition on absolute stability can be easily satisfied. Hence, we can use \eqref{equ-mu-new} to perform high-level policy training stably in the HED algorithm. \section*{Appendix B} This appendix presents a proof of Proposition \ref{the-2}. Considering any specific state $s\in\mathcal{S}$, let $$ \nabla_a Q^e(s,a)|_{a=\pi^e(s)}=C, $$ where $C$ is an arbitrary scalar constant, in line with the assumption of scalar actions. Using \eqref{equ-pe} and \eqref{equ-lin-pol}, the ensemble policy gradient with respect to policy parameters $\theta_i$ of policy $\pi^i$, $i\in[1,\ldots,N]$, is $$ \nabla_{\theta_i} J(\pi^e)=\frac{C\Phi(s)}{N}. $$ According to the multi-step learning rule in \eqref{equ-mu-new}, updating $\theta_i$ for one iteration gives the updated $\theta_i$ as $$ (1-\rho_0)\theta_i+2\rho_0\theta_q-\rho_0\theta_p+h\frac{C\Phi(s)}{N}. $$ Therefore, \begin{equation*} Mul(\pi^i(s))=(1-\rho_0)\pi^i(s)+2\rho_0\pi^q(s)-\rho_0\pi^p(s) +\frac{hC}{N}\Phi(s)^T\Phi(s). \end{equation*} Hence $$ \mathbb{E}\left[ Mul(\pi^i(s)) \right]=(1-\rho_0)\pi^i(s)+\rho_0\pi^e(s)+\frac{h C}{N}\Phi(s)^T\Phi(s), $$ $$ \mathbb{E}\left[ Mul(\pi^e(s)) \right]=\pi^e(s)+\frac{h C}{N}\Phi(s)^T\Phi(s). $$ In comparison, upon using the single-step method, the updated $\theta_i$ becomes $$ \theta_i+h\frac{C\Phi(s)}{N}. $$ Subsequently, $$ Sin(\pi^i(s))=\pi^i(s)+\frac{h C}{N}\Phi(s)^T\Phi(s), $$ $$ Sin(\pi^e(s))=\pi^e(s)+\frac{h C}{N}\Phi(s)^T\Phi(s). $$ Clearly, $$ Sin(\pi^e(s))=\mathbb{E}\left[ Mul(\pi^e(s)) \right]. $$ Hence, the expected action changes applied to $\pi^e(s)$ are identical, regardless of whether single-step or multi-step method is used for high-level policy training\footnote{We assume in Proposition \ref{the-2} that high-level policy training is performed for one iteration on a specific state $s$.}. Define $$ \Delta=\sum_{i\in[1,\ldots,N]}(\pi^i(s)-\pi^e(s)). $$ For the single-step method, after all base learners trained their respective policies for one iteration on state $s$, it is easy to verify that \begin{equation*} \begin{split} & \sum_{i\in[1,\ldots,N]}\left( Sin(\pi^i(s))-Sin(\pi^e(s)) \right)^2\\ =& \sum_{i\in[1,\ldots,N]}\left( \pi^i(s)-\pi^e(s) \right)^2\\ =& \Delta. \end{split} \end{equation*} Meanwhile, \begin{equation*} \begin{split} & \left( Mul(\pi^i(s))-\mathbb{E}\left[ Mul(\pi^e(s)) \right] \right)^2 \\ =&\left( (1-\rho_0) (\pi^i(s)-\pi^e(s)) + 2\rho_0 (\pi^q(s)-\pi^e(s)) -\rho_0(\pi^p(s)-\pi^e(s)) \right)^2\\ =&(1-\rho_0)^2(\pi^i(s)-\pi^e(s))^2+4\rho_0^2(\pi^q(s)-\pi^e(s))^2+\rho^2(\pi^p(s)-\pi^e(s))^2\\ &+4(1-\rho_0)\rho_0(\pi^i-\pi^e(s))(\pi^q(s)-\pi^e(s)) \\ &-2(1-\rho_0)\rho_0(\pi^i(s)-\pi^e(s))(\pi^p(s)-\pi^e(s))\\ &-4\rho_0^2(\pi^q(s)-\pi^e(s))(\pi^p(s)-\pi^e(s)). \end{split} \end{equation*} Since the base learner indices $p$ and $q$ are randomly and independently selected, $$ \mathbb{E}\left[(\pi^i(s)-\pi^e(s))(\pi^p(s)-\pi^e(s))\right]=0, $$ $$ \mathbb{E}\left[(\pi^i(s)-\pi^e(s))(\pi^q(s)-\pi^e(s))\right]=0, $$ $$ \mathbb{E}\left[(\pi^q(s)-\pi^e(s))(\pi^p(s)-\pi^e(s))\right]=0. $$ Therefore \begin{equation*} \begin{split} & \sum_{i\in[1,\ldots,N]} \mathbb{E}\left[ \left( Mul(\pi^i(s))-\mathbb{E}\left[ Mul(\pi^e(s)) \right] \right)^2 \right] \\ =&(1-\rho_0)^2\Delta+4\rho_0^2\Delta+\rho_0^2\Delta \\ =&(1-2\rho_0+6\rho_0^2)\Delta. \end{split} \end{equation*} When $0<\rho_0<\frac{1}{3}$, $$ 1-2\rho_0+6\rho_0^2<1. $$ As a result, \begin{equation*} \begin{split} &\sum_{i\in[1,\ldots,N]}\mathbb{E}\left[ \left( Mul(\pi^i(s))-\mathbb{E}\left[ Mul(\pi^e(s)) \right] \right)^2 \right] \\ <&\sum_{i\in[1,\ldots,N]}\left( Sin(\pi^i(s))-Sin(\pi^e(s)) \right)^2. \end{split} \end{equation*} \section*{Appendix C} Table~\ref{tab:hyper-para} provides detailed hyper-parameter settings of all competing algorithms. Our hyper-parameter settings follow strictly the recommended settings in \cite{fujimoto2018,haarnoja2018,januszewski2021,lee2021sunrise}. \begin{table}[htb!] \caption{Hyper-parameter settings of competing algorithms.} \label{tab:hyper-para} \centering \resizebox{0.7\linewidth}{!}{ \begin{tabular}{l||llll} \hline Hyper-parameter & TD3 & SAC & ED2 & SUNRISE \\ \hline Num. episodes & 2500 & 2500 & 2500 & 3000 \\ Episode length & 1000 & 1000 & 1000 & 1000 \\ Minibatch size & 100 & 100 & 100 & 256 \\ Adam learning rate & 5e-4 & 3e-4 & 5e-4 & 3e-4 \\ Discount ($\gamma$) & 0.99 & 0.99 & 0.99 & 0.99 \\ GAE parameter ($\lambda$) & 0.995 & 0.995 & 0.995 & 0.995 \\ Replay buffer size & 1e6 & 1e6 & 1e6 & 1e6 \\ Update interval & 50 & 50 & 50 & 50 \\ Ensemble size & - & - & 5 & 5 \\ \hline \end{tabular}} \end{table} All experiments were run using a cluster of Linux computing nodes. Each node is equipped with 16 GB memory. The CPU specification is provided in Table~\ref{tab:cpu-para}. Each experiment was run in a Python virtual environment managed by Anaconda with Python packages specified in Table~\ref{tab:python-lib}. \begin{table}[htb!] \caption{CPU specification.} \label{tab:cpu-para} \centering \resizebox{0.6\linewidth}{!}{ \begin{tabular}{ll} \hline Architecture & x86\_64 \\ CPU op-mode(s) & 32-bit, 64-bit \\ CPU(s) & 16 \\ CPU family & 6 \\ Thread(s) per core & 2 \\ CPU max MHz & 4900.0000 \\ CPU min MHz & 800.0000 \\ Model name & 11th Gen Intel(R) Core(TM)\\ & i7-11700 @ 2.50GHz \\ \hline \end{tabular} } \end{table} \begin{table}[htb!] \caption{Python packages.} \label{tab:python-lib} \centering \begin{tabular}{ll} \hline Package name &Version \\ \hline cython &0.29.25 \\ gym &0.21.0 \\ keras &2.7.0 \\ mujoco-py &2.1.2.14 \\ numpy &1.21.4 \\ pybulletgym &0.1 \\ python &3.7.11 \\ scipy &1.7.3 \\ tensorflow &2.7.0 \\ \hline \end{tabular} \end{table} \begin{figure*}[!hbt] \begin{center} \subfloat[Hopper-v0 (PyBullet)]{\label{fig-hc-eq9}\includegraphics[width=0.33\linewidth]{exp/when2train_HighLevelPolicy/HopperPyBulletEnv-v0_when2HLtrain.pdf}} \subfloat[InvertedDoublePendulum-v0 (PyBullet)]{\label{fig-idpPB-when2HLtrain}\includegraphics[width=0.33\linewidth]{exp/when2train_HighLevelPolicy/InvertedDoublePendulumPyBulletEnv-v0_when2HLtrain.pdf}} \subfloat[InvertedPendulum-v0 (PyBullet)]{\label{fig-ipPB-when2HLtrain}\includegraphics[width=0.33\linewidth]{exp/when2train_HighLevelPolicy/InvertedPendulumPyBulletEnv-v0_when2HLtrain.pdf}}\\ \subfloat[Reacher-v0 (PyBullet)]{\label{fig-reacher-when2HLtrain}\includegraphics[width=0.33\linewidth]{exp/when2train_HighLevelPolicy/ReacherPyBulletEnv-v0_when2HLtrain.pdf}} \subfloat[Hopper-v3 (Mujoco)]{\label{fig-hopper-when2HLtrain}\includegraphics[width=0.33\linewidth]{exp/when2train_HighLevelPolicy/Hopper-v3_when2HLtrain.pdf}} \subfloat[Walker2D-v3 (Mujoco)]{\label{fig-walkerPB-when2HLtrain}\includegraphics[width=0.33\linewidth]{exp/when2train_HighLevelPolicy/Walker2d-v3_when2HLtrain.pdf}}\\ \end{center} \caption{Learning curves of HED with respect to two high-level policy training approaches. The method that conducts high-level policy training after every 50 timesteps is denoted as \emph{Fixed timestep}. The method that conducts high-level policy training at the end of each sampled episode is denoted as \emph{Each episode}.} \label{fig:when2HLtrain_impact} \end{figure*} \begin{figure*}[!hbt] \begin{center} \subfloat[Hopper-v0 (PyBullet)]{\label{fig-hc-eq9}\includegraphics[width=0.33\linewidth]{exp/eq9_impact/HopperPyBulletEnv-v0_EQ9.pdf}} \subfloat[InvertedDoublePendulum-v0 (PyBullet)]{\label{fig-idpPB-eq9}\includegraphics[width=0.33\linewidth]{exp/eq9_impact/InvertedDoublePendulumPyBulletEnv-v0_EQ9.pdf}} \subfloat[InvertedPendulum-v0 (PyBullet)]{\label{fig-ipPB-eq9}\includegraphics[width=0.33\linewidth]{exp/eq9_impact/InvertedPendulumPyBulletEnv-v0_EQ9.pdf}}\\ \subfloat[Reacher-v0 (PyBullet)]{\label{fig-reacher-eq9}\includegraphics[width=0.33\linewidth]{exp/eq9_impact/ReacherPyBulletEnv-v0_EQ9.pdf}} \subfloat[Hopper-v3 (Mujoco)]{\label{fig-hopper-eq9}\includegraphics[width=0.33\linewidth]{exp/eq9_impact/Hopper-v3_EQ9.pdf}} \subfloat[Walker2D-v3 (Mujoco)]{\label{fig-walkerPB-eq9}\includegraphics[width=0.33\linewidth]{exp/eq9_impact/Walker2d-v3_EQ9.pdf}}\\ \end{center} \caption{Learning curves of HED using the single-step high-level policy training technique in \eqref{equ-e-pg} vs. the proposed multi-step high-level policy training technique in \eqref{equ-mu-new}.} \label{fig:eq9_impact} \end{figure*} \section*{Appendix D} In this appendix, we study the effectiveness of conducting high-level policy training after HED obtains a full sampled episode. Figure~\ref{fig:when2HLtrain_impact} shows the performance comparison of HED with two different training frequencies: every 50 consecutive timesteps vs. every episode. It can be noticed that, on a majority benchmark problems (i.e., 5 out of 6), performing high-level policy training after every episode (orange curve) can significantly improve the HED algorithm in terms of both the final performance and convergence speed. For example, as shown in Figure~\ref{fig:when2HLtrain_impact}(e), the orange curve reaches 3500 after 1500 episodes while the blue curve converges to a lower cumulative reward (approx. 3000) after 2000 episodes. We also notice that episodic policy training is more robust to the randomness in the environment and less sensitive to the initialization of neural network weights. For example, in Figure~\ref{fig:when2HLtrain_impact}(e), episodic policy training produces a smaller confident interval (orange shaded area) compared to the fixed timestep training (blue shaded area) over 10 independent algorithm runs. Similar results can also be observed from Figures~\ref{fig:when2HLtrain_impact}(a) and (f). Note that in each algorithm run, both policy networks and Q-networks are initialized with different weights. The environment initial states also vary. \section*{Appendix E} This appendix investigates the effectiveness of multi-step policy training by using \eqref{equ-mu-new}. Specifically, we compare the performance of HED against its variant, which performs single-step high-level policy training by using \eqref{equ-e-pg}, on 6 problems that include both PyBullet and Mujoco benchmarks. As shown in Figure~\ref{fig:eq9_impact}, with the help of \eqref{equ-mu-new}, significant performance improvement can be observed (orange curve) on most benchmark problems. In particular, HED behaves more stably during the learning process. For example, in Figure~\ref{fig:eq9_impact}(c), the cumulative rewards obtained by the policy trained using~\eqref{equ-mu-new} (orange curve) remain stable at 1000 after 300 episodes. In comparison, the blue curve stays below 1000 and fluctuates between 800 and 1000 over the entire learning period. Similar trends can also be noticed in Figure~\ref{fig:eq9_impact}(b). The proposed multi-step policy training technique achieves clearly higher cumulative rewards. In the Hopper environment shown in Figure~\ref{fig:eq9_impact}(e), the orange curve outperforms the blue curve by up to 75\% after 2500 training episodes. Moreover, the orange curve converges to 3500 while the blue curve remains below 2000. The significant improvement in cumulative rewards can also be witnessed in Figure~\ref{fig:eq9_impact}(a) and (f). The shaded areas in Figure~\ref{fig:eq9_impact}(e) and (f) also show that the multi-step training technique is less sensitive to the environment randomness and neural network weight initialization, compared to using the conventional single-step training method in \eqref{equ-e-pg}. Hence, our experiment results confirm the importance of inter-learner collaboration. By enabling base learners in an ensemble to explicitly share their learned policy parameters, HED can achieve high learning stability and effectively boost the learning process. \bibliographystyle{IEEEtran} \section{Introduction} \label{sec-int} Deep reinforcement learning (DRL) is a booming field of research in machine learning with diverse real-world applications \cite{ibarz2021}. In recent years, many \emph{model-free} DRL algorithms achieved cutting-edge performance in tackling various continuous reinforcement learning (RL) problems, including complex control tasks with high-dimensional state and action spaces \cite{liu2021}. These algorithms can effectively train \emph{deep neural networks} (DNNs) to precisely model high-quality control policies and are the central focus of this paper. Despite of widely reported success, a majority of existing \emph{actor-critic DRL algorithms}, such as DDPG \cite{lillicrap2015}, SAC \cite{haarnoja2018} and PPO \cite{schulman2017}, still suffer from some major limitations. Specifically, existing research works showed that the algorithm performance is highly sensitive to hyper-parameter settings and can vary substantially in different algorithm runs \cite{paine2020}. \emph{Ineffective exploration} is often considered as a major cause for the poor learning stability \cite{chan2019}, often resulting in overfitting and premature convergence to poor local optima \cite{kurutach2018}. Rather than relying on one learner (or DRL agent), an ensemble of base learners can be jointly utilized to boost exploration and stabilize the learning process \cite{osband2016deep,osband2017post}. For example, the \emph{ensemble deep deterministic policy gradient} (ED2) algorithm is a newly developed ensemble DRL method \cite{januszewski2021} that trains multiple DNN policies simultaneously using a shared \emph{experience replay buffer} (ERB), similar to several previously proposed parallel DRL algorithms \cite{barth2018,mnih2016asyn}. ED2 features a unique mixture of multiple well-studied tricks, including temporally-extended deep exploration, double Q-bias reduction, and target policy smoothing \cite{osband2016deep,osband2017post,van2016deep,fujimoto2018}. It was reported to outperform state-of-the-art ensemble DRL algorithms such as SUNRISE \cite{lee2021sunrise} on several difficult Mujoco benchmark control problems. As far as we know, most of the existing ensemble DRL algorithms are designed to train each base learner individually. For example, in ED2, every base learner trains its own DNN policy using its own critic, with the aim to improve its own performance without considering the impact of the trained policy on the ensemble. While sharing the same ERB, policy training is conducted largely independently by all base learners. This is shown to promote healthy exploration in \cite{januszewski2021}. However, there is no guarantee that the base learners will collaborate effectively such that the ensemble as a whole can achieve desirable performance. To address this limitation, we propose a new \emph{hierarchical approach} for training base learners in this paper. Specifically, we follow ED2 for \emph{low-level training} of DNN policies, which will be performed concurrently by all base learners. In the meantime, we construct a \emph{global critic}, which is trained constantly to predict the performance of the ensemble. Guided by the global critic, \emph{high-level training} of DNN policies will be performed regularly to strengthen cooperation among all the base learners. Since the ensemble is not used directly to collect state-transition samples from the learning environment, we must make sure that high-level training of the ensemble is not performed on \emph{out-of-distribution} data obtained by individual base learners. In view of this, it is important to encourage \emph{inter-learner parameter sharing} so that the DNN policy trained by one base learner can contribute directly to (or influence) the training of DNN policies by other base learners. For this purpose, we develop a new technique in this paper for high-level training of policies based on the \emph{multi-step integration methods} for solving \emph{ordinary differential equations} (ODEs) in \cite{scieur2017}. Our high-level policy training technique is theoretically justified as it guarantees stability for a wide range of optimization problems. Meanwhile, it can be shown analytically that the trained linear parametric policies (a special and important technique for policy approximation) of all base learners are expected to behave more consistently as the ensemble through high-level policy training, encouraging inter-learner collaboration and alleviating the data distribution issue. Driven by the hierarchical policy training method, we develop a new ensemble DRL algorithm called \emph{hierarchical ensemble deep deterministic policy gradient} (HED) in this paper. Experimental evaluation of HED has been conducted on a range of benchmark control problems, including the widely used Mujoco control tasks as well as the less popular and potentially more challenging PyBullet control problems. Our experiments clearly show that HED can outperform ED2, SUNRISE and several cutting-edge DRL algorithms on multiple benchmark problems. \iffalse The rest of this paper is organized as follows. We review the related research works in the recent literature in the subsequent section. Necessary technical background for developing the HED algorithm is presented next, followed by the detailed design of HED. The empirical performance of the new algorithm is further evaluated before the concluding remarks. \fi \section{Related Work} \label{sec-rw} Similar to ED2, HED trains an ensemble of policies using an \emph{off-policy} DRL algorithm to leverage on the algorithm's advantages in \emph{sample efficiency}. Recently, several off-policy DRL algorithms have been developed successfully for RL in continuous spaces, including DDPG \cite{lillicrap2015}, SAC \cite{haarnoja2018}, TD3 \cite{fujimoto2018}, and SOP \cite{wang2020}. These algorithms introduce a variety of tricks to stabilize the learning process. For example, TD3 extends the idea of double Q-network \cite{van2016deep} to a new double-Q bias reduction technique, which can effectively prevent over-optimistic training of DNN policies. In addition, empirical evidence showed that the learning process becomes more stable when the actor and critic in TD3 are trained with different frequencies or in different phases \cite{fujimoto2018,cobbe2021}. The base learners in our HED ensemble will adopt these tricks. The recent literature also provides some new tricks to stabilize learning. Specifically, various trust-region methods have been developed to prevent negative behavioral changes during policy training \cite{kurutach2018,schulman2015,shani2020,wu2017,schulman2017}. Meanwhile, entropy regularization techniques prohibit immature convergence of the trained policies and ensure prolonged profitable exploration \cite{chen2018,haarnoja2018}. However, these techniques are mainly applied to stochastic policies while we aim at learning an ensemble of deterministic policies. Previous research showed that deterministic policies can often be trained more efficiently than stochastic policies using the \emph{reparameterization trick} \cite{fujimoto2018,silver2014,baek2020}. The stability of a DRL algorithm depends critically on how the learner explores its environment. Besides the entropy regularization methods, curiosity metrics are popularly employed to encourage a learner to explore rarely visited states during RL \cite{reizinger2020,zhelo2018}. Meanwhile, many previous studies embraced the \emph{optimum in the face of uncertainty} (OFU) principle to design bonus rewards for actions with high potentials, thereby promoting exploration in promising areas of the learning environment \cite{bellemare2016}. One good example is the UCB exploration technique developed in \cite{chen2017ucb,lee2021sunrise}. However, in \cite{januszewski2021}, this technique was shown to be less effective than the bootstrap with random initialization trick adopted in ED2. Temporally-extended exploration on RL problems with continuous actions can also be achieved by adding a small amount of noise to DNN weights \cite{plappert2017}. This is directly related to the posterior sampling methods that are often used to select the best actions among a statistically plausible set of sampled actions \cite{osband2018}. Following the OFU principle, deep ensembles have been recently proposed to approximate Bayesian posteriors with high accuracy and efficiency \cite{lakshminarayanan2016}. They are subsequently used to approach deep exploration for reliable RL \cite{osband2016deep}. Several issues have been investigated under the context of ensemble DRL. For instance, the diversity of base learners is essential to the performance of the ensemble. To encourage diversity, either different DRL algorithms or the same algorithm with differed hyper-parameter settings have been adopted to train base learners \cite{huang2017,wiering2008}. Meanwhile, proper aggregation of the action outputs from all base learners in an ensemble poses another challenge. Typical approaches to tackle this issue include taking the mean action as the output of the ensemble and choosing the action with highest predicted cumulative rewards \cite{januszewski2021,chen2019}. As far as we know, few existing ensemble DRL algorithms in the literature have ever studied the important issue on how to effectively train all base learners to jointly improve the ensemble performance. This issue will be explored in-depth with the newly developed HED algorithm in this paper. \section{Background} \label{sec-back} An RL problem is modeled as a \emph{Markov Decision Process} (MDP) $(\mathcal{S},\mathcal{A},R,P,\gamma,p_0)$, where $\mathcal{S}$ and $\mathcal{A}$ refer respectively to the continuous multi-dimensional state space and action space. $P$ stands for the state-transition model that governs the probability of reaching any state $s_{t+1}\in\mathcal{S}$ at timestep $t+1$ upon performing any action $a_t\in\mathcal{A}$ in state $s_t\in\mathcal{S}$ at timestep $t$, with $t\in\mathbb{Z}^+$. Additionally, $\gamma\in[0,1)$ is the discount factor, $R$ is the reward function, and $p_0$ captures the initial state distribution. To solve any RL problem described above, we aim to learn an optimal \emph{deterministic ensemble policy} $\pi^e_*(s)$ that maps any state input $s\in\mathcal{S}$ to an action vector $a\in\mathcal{A}$ so as to maximize the \emph{cumulative rewards} defined below $$ \pi^e_*= \argmax_{\pi^e} J(\pi^e)=\argmax_{\pi^e}\mathbb{E}_{\tau\sim\pi^e} \left[ \sum_{t=1}^{\infty}\gamma^{t-1} R(s_t,a_t) \right], $$ where $\tau=[(s_t,a_t,r_t,s_{t+1})]_{t=1}^{\infty}$ contains a series of consecutive state-transition samples and is called a \emph{episode}, which can be obtained by following the ensemble policy $\pi^e$, and $r_t=R(s_t,a_t)$ is the immediate reward received at timestep $t$ in $\tau$. For an ensemble with $N$ base learners where each base learner $L_i$, $1\leq i\leq N$, maintains its own deterministic base policy $\pi^i$, the action output of $\pi^e$ is jointly determined by all the \emph{base policies} according to \begin{equation} \forall s\in\mathcal{S}, \pi^e(s)=\frac{1}{N}\sum_i^N \pi^i(s). \label{equ-pe} \end{equation} In order to train an ensemble to maximize the cumulative rewards, our baseline algorithm ED2 uses randomly selected base learners to sample a series of episodes $\{\tau_i\}$, which will be stored in the shared ERB. At regular time intervals, a mini-batch of state-transition samples will be retrieved from the ERB. Every base learner $L_i$ will then use the retrieved mini-batch to train its own actor $\pi^i$ and critic $Q^i$ individually. In other words, a base learner manages two separate DNNs, one models the deterministic policy $\pi^i$ and the other approximates the Q-function $Q^i$ of $\pi^i$. A base learner uses an existing actor-critic RL algorithm to train the two DNNs. In this paper, we choose TD3 for this purpose due to its proven effectiveness, high popularity and stable learning behavior \cite{fujimoto2018}. \section{Hierarchical Ensemble Deep Deterministic Policy Gradient} \label{sec-algo} The pseudo-code of the HED algorithm is presented in Algorithm \ref{alg-code}. HED follows many existing works including ED2 \cite{osband2016deep,januszewski2021} to achieve temporally-extended exploration through bootstrapping with random initialization of DNN policies. As clearly shown in \cite{januszewski2021}, this exploration technique is more effective than UCB and parameter randomization methods. Different from ED2 which completely eliminates the necessity of adding small random noises to the deterministic action outputs from the DNN policies, we keep a small level of action noise\footnote{The noise is sampled from the Normal distribution independently for each dimension of the action vector. The variance of the normal distribution is fixed at 0.01 during the learning process.} while using any chosen policy to explore the learning environment. We found empirically that this ensures coherent exploration, similar to \cite{osband2016deep}, while making the testing performance of the trained policies more stable. Different from ED2 and other ensemble algorithms for RL in continuous spaces, HED trains DNN policies at two separate levels. The low-level training of $\pi^i$ and $Q^i$ by each base learner $L_i$ is essentially the same as ED2 and TD3. Specifically, for any base learner $L_i$, $i\in\{1,\ldots,N\}$, $Q^i$ is trained by $L_i$ to minimize $MSE_i$ below \begin{equation} MSE_i=\frac{1}{|\mathcal{B}|} \sum_{(s,a,r,s')\in\mathcal{B}}\left(Q^i_{\phi_i}(s,a)-r-\gamma\min_{k=1,2}\hat{Q}^i_{k}(s',\pi^i(s')+\epsilon) \right)^2, \label{equ-mse} \end{equation} where $\phi_i$ represents the trainable parameters of the DNN that approximates $Q^i$. $\mathcal{B}$ is the random mini-batch retrieved from the ERB. $\hat{Q}^i_{k}$ with $k=1,2$ stands for the two target Q-networks of $L_i$ that together implement the double-Q bias reduction mechanism proposed in \cite{fujimoto2018}. Additionally, $\epsilon$ is a random noise sampled from a Normal distribution with zero mean and small variance\footnote{The variance for sampling $\epsilon$ is kept at a very small level of 0.01 in the experiments.}. Using the trained $Q^i$, the trainable parameters $\theta_i$ of the DNN that models policy $\pi^i$ is further updated by $L_i$ along the \emph{policy gradient} direction computed below \begin{equation} \nabla_{\theta_i}J(\pi^i_{\theta_i})=\frac{1}{|\mathcal{B}|}\sum_{s\in\mathcal{B}} \nabla_{a}Q^i(s,a)|_{a=\pi^i_{\theta_i}(s)}\nabla_{\theta_i}\pi^i_{\theta_i}(s). \label{equ-pg} \end{equation} Besides the above, HED constantly trains a separate high-level Q-function $Q^e$ to predict the performance of the ensemble policy $\pi^e$. Guided by the trained $Q^e$, high-level policy training is conducted regularly to update policy $\pi^i$ of all base learners so as to enhance their cooperation and performance. A new \emph{multi-step technique} is developed in HED to enable inter-learner parameter sharing during high-level policy training. To implement this technique, we keep track of a list of bootstrap policy parameters for the multi-step training process. More details can be found in the subsequent subsection. Theoretical justifications regarding the usefulness of the multi-step approach are also provided below. \begin{algorithm}[!ht] \begin{algorithmic} \STATE {\bf Input}: Ensemble size $N$; initial policy networks $\pi^i_{\theta_i}$ and Q-networks $Q^i_{\phi_i}$ for $i\in\{1,\ldots,N\}$; ERB; ensemble Q-network $Q^e_{\phi_e}$; target Q-networks for each base learner and the ensemble \STATE {\bf Output}: Trained ensemble policy $\pi^e$ \STATE {\bf While} the total number of sampled trajectories $<$ max number of trajectories: \STATE \ \ \ \ Randomly sample $i\in\{1,\ldots,N\}$ \STATE \ \ \ \ {\bf While} the current trajectory does not terminate: \STATE \ \ \ \ \ \ \ \ Use $\pi^i$ to perform the next action \STATE \ \ \ \ \ \ \ \ Store sampled state-transition in ERB \STATE \ \ \ \ \ \ \ \ Track number of steps sampled before critic training \STATE \ \ \ \ \ \ \ \ {\bf If} time for critic training: \STATE \ \ \ \ \ \ \ \ \ \ \ \ {\bf For} number of steps sampled: q\STATE \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ Sample a mini-batch $\mathcal{B}$ from ERB \STATE \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ Train $Q^i_{\phi_i}$ for $i\in\{1,\ldots,N\}$ using \eqref{equ-mse} \STATE \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ Train $Q^e_{\phi_e}$ using \eqref{equ-e-mse} \STATE \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ {\bf If} time for {\bf low-level} policy training: \STATE \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ Train $\pi^i_{\theta_i}$ for $i\in\{1,\ldots,N\}$ using \eqref{equ-pg} \STATE \ \ \ \ \ \ \ \ {\bf If} time for {\bf high-level} policy training: \STATE \ \ \ \ \ \ \ \ \ \ \ \ Set bootstrap list $\{x_j\}_{j=0}^{2}$ for each base learner \STATE \ \ \ \ \ \ \ \ \ \ \ \ {\bf For} a fraction of sampled steps: \STATE \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ Train $\pi^i_{\theta_i}$ for $i\in\{1,\ldots,N\}$ using \eqref{equ-mu-new} \STATE \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ Append trained $\theta_i$ for $i\in\{1,\ldots,N\}$ to the bootstrap lists of each base \STATE \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ learner for the next step of high-level policy training \end{algorithmic} \caption{The pseudo-code of the HED algorithm.} \label{alg-code} \end{algorithm} \subsection{A multi-step technique for high-level policy training} \label{subsec-multipol} In addition to $Q^i$ for each base learner $L_i$, $i\in\{1,\ldots,N\}$, HED maintains a separate Q-network to approximate $Q^e$ of the ensemble policy $\pi^e$. Similar to \eqref{equ-mse}, HED trains this central Q-network towards minimizing $MSE_e$ below \begin{equation} MSE_e=\frac{1}{|\mathcal{B}|} \sum_{(s,a,r,s')\in\mathcal{B}}\left(Q^e_{\phi_e}(s,a)-r-\gamma\hat{Q}^e(s',\pi^e(s')) \right)^2, \label{equ-e-mse} \end{equation} with $\phi_e$ representing the trainable parameters of the central Q-network. $\hat{Q}^e$ stands for the corresponding target Q-network that stabilizes the training process. For simplicity, we do not add random noise $\epsilon$ in \eqref{equ-mse} to the action outputs produced by the ensemble policy $\pi^e$ in \eqref{equ-e-mse}. Furthermore, following \cite{van2016deep}, one target Q-network instead of two is adopted in \eqref{equ-e-mse} to facilitate the training of $Q^e$. Building on the trained $Q^e$, we can calculate the \emph{ensemble policy gradient} with respect to $\theta_i$ of every base learner $L_i$ as follows \begin{equation} \nabla_{\theta_i}J(\pi^e)= \frac{1}{|\mathcal{B}|}\sum_{s\in\mathcal{B}} \nabla_{a}Q^e(s,a)|_{a=\pi^e(s)}\nabla_{a_i}\pi^e(s)|_{a_i=\pi^i_{\theta_i}(s)}\nabla_{\theta_i}\pi^i_{\theta_i}(s), \label{equ-e-pg} \end{equation} with $$ \nabla_{a_i}\pi^e(s)|_{a_i=\pi^i_{\theta_i}(s)}=\frac{1}{N} I, $$ according to \eqref{equ-pe}. $I$ stands for the $m\times m$ identity matrix where $m$ is the dimension of the action vector. One straightforward approach for high-level policy training is to update $\theta_i$ of every base learner $L_i$ in the direction of \eqref{equ-e-pg}. However, using \eqref{equ-e-pg} alone may not encourage any base learner $L_i$ to behave consistently with the ensemble (see Proposition \ref{the-2}). Consequently, high-level training of the ensemble policy may be performed on the out-of-distribution state-transition samples collected by the base learners, affecting the training effectiveness. Furthermore, ensembles are used mainly for temporally-extended exploration in the literature. The learning activity of one base learner can only influence other base learners indirectly through the shared ERB. Base learners do not explicitly share their learned policy parameters in an ensemble to strengthen inter-learner cooperation and boost the learning process. To address this limitation, we propose to promote inter-learner parameter sharing during high-level policy training, in order to achieve a desirable balance between exploration and inter-learner cooperation. Specifically, in addition to \eqref{equ-e-pg}, we randomly select \emph{two base learners} $L_p$ and $L_q$ and use their policy parameters to guide the training of policy $\pi^i$ of any base learner $L_i$. In comparison to \emph{selecting one base learner}, this allows more base learners to have the opportunity to share their parameters with the base learner $L_i$ during policy training. It is also possible to recruit more than two base learners. However, in this case, it is mathematically challenging to derive stable learning rules for high-level policy training. Motivated by the above discussion, a search through the literature leads us to the linear multi-step integration methods recently analyzed in \cite{scieur2017}. Consider a simple \emph{gradient flow equation} below \begin{equation} x(0)=\theta_i^0, \frac{\partial x(t)}{\partial t}=g(x(t))=\nabla_{\theta_i}J(\pi^e)|_{\theta_i=x(t)}, \label{equ-gfe} \end{equation} where $\theta_i^0$ refers to the initial policy parameter of base learner $L_i$. If $J(\pi^e)$ is strongly concave and Lipschitz continuous, the solution of \eqref{equ-gfe} allows us to obtain the optimal policy parameters $\theta_i^*$ when $x(t)$ approaches to $\infty$. Since $J(\pi^e)$ is not strongly concave for most of real-world RL problems, $x(t)$ in practice may only converge to a locally optimal policy, which is common among majority of the policy gradient DRL algorithms. Therefore high-level training of policy $\pi^e$ and hence $\pi^i$ can be approached by numerically solving \eqref{equ-gfe}. This can be achieved through a linear $\mu$-step method shown below \begin{equation} x_{k+\mu}=-\sum_{j=0}^{\mu-1}\rho_j x_{k+j}+h\sum_{j=0}^{\mu-1}\sigma_j g(x_{k+j}), \forall k\geq 0, \label{equ-mu-step} \end{equation} where $\rho_j,\sigma_j\in\mathbb{R}$ are the pre-defined coefficients of the multi-step method and $h$ is the learning rate. Clearly, each new point $x_{k+\mu}$ produced by the $\mu$-step method is a function of the preceding $\mu$ points. In this paper, we specifically consider the case when $\mu=3$. Meanwhile, let \begin{equation} x_0=\theta_p,x_1=\theta_q,x_2=\theta_i \label{equ-boot-list} \end{equation} where $p$ and $q$ are the randomly generated indices of two base learners and $i$ is the index of the base learner whose policy $\pi^i$ is being trained by the $\mu$-step method. Through this way, the training of policy $\pi^i$ is influenced directly by base learners $L_p$ and $L_q$ through explicit inter-learner parameter sharing. $x_i (i\geq 3)$ in \eqref{equ-mu-step} represents the trained policy parameters of $\pi^i$ in subsequent training steps. Although \eqref{equ-mu-step} allows us to use $\nabla_{\theta_p}J(\pi^e)$ and $\nabla_{\theta_q}J(\pi^e)$ to train $\theta_i$, they do not seem necessary for inter-learner parameter sharing. To simplify \eqref{equ-mu-step}, we set $\sigma_0=\sigma_1=0$ and $\sigma_2=1$. Hence only $g(x_{k+2})$, which is the ensemble policy gradient with respect to policy $\pi^i$ in \eqref{equ-e-pg}, is used to train $\pi^i$. With this simplification, we derive the new learning rule for high-level policy training below \begin{equation} \begin{split} & x_{k+3}= -\rho_2 x_{k+2}-\rho_1 x_{k+1}-\rho_0 x_k+h\cdot \nabla_{\theta_i}J(\pi^e)|_{\theta_i=x_{k+2}}, \forall k\geq 0 \\ & x_0=\theta_p,x_1=\theta_q,x_2=\theta_i \end{split} \label{equ-mu-new} \end{equation} To implement \eqref{equ-mu-new} in HED, before high-level policy training, every base learner $L_i$ must set up a \emph{bootstrap list} of policy parameters $\{x_0=\theta_p,x_1=\theta_q,x_2=\theta_i\}$. After the $k$-th training step ($k\geq 0$) based on \eqref{equ-mu-new}, $L_i$ appends the trained $\theta_i$ as $x_{k+3}$ to the bootstrap list, which will be utilized to train $\pi^i$ in the subsequent training steps. Reliable use of \eqref{equ-mu-new} demands for careful parameter settings of $\rho_0$, $\rho_1$, $\rho_2$ and $h$. Relevant theoretical analysis is presented below. \subsection{Theoretical analysis of the multi-step policy training technique} \label{subsec-multithe} In this subsection, a theoretical analysis is performed first to determine suitable settings of $\rho_0$, $\rho_1$, $\rho_2$ and $h$ for stable high-level policy training through \eqref{equ-mu-new}. To make the analysis feasible, besides the strongly concave and Lipschitz continuous conditions, we further assume that \begin{equation} \nabla_{\theta_i}J(\pi^e)\approx -A(\theta_i-\theta_i^*) \label{equ-grad-linear} \end{equation} where $A$ is a positive definite matrix whose eigenvalues are bounded positive real numbers. $\theta_i^*$ stands for the global-optimal (or local-optimal) policy parameters. Many strongly concave functions satisfy this assumption \cite{scieur2017}. Meanwhile, the attraction basin of the local optimum of many multi-modal optimization problems often satisfies this assumption too. Using this assumption, we can derive Proposition \ref{the-1} below. \begin{proposition} Upon using \eqref{equ-mu-new} to numerically solve \eqref{equ-gfe}, the following conditions must be satisfied for $x_k$ to converge to $\theta_i^*$ as $k$ approaches to $\infty$: \begin{enumerate} \item $\rho_2=\rho_0-1$, $\rho_1=-2\rho_0$; \item $0<\rho_0<\frac{1}{2}$; \item $h$ is reasonably small such that $0\leq \lambda h < 2-4\rho_0$, where $\lambda$ can take any real value between the smallest and the largest eigenvalues of the positive definite matrix $A$ in \eqref{equ-grad-linear}. \end{enumerate} \label{the-1} \end{proposition} The proof of Proposition \ref{the-1} can be found in Appendix A. Proposition \ref{the-1} provides suitable parameter settings for \eqref{equ-mu-new} and justifies its stable use for high-level policy training. We next show that \eqref{equ-mu-new} is also expected to make base learners behave more consistently with the ensemble, without affecting the behavior of the trained ensemble, when $\rho_0$ is sufficiently small. Consider specifically that each base learner $L_i$ trains a linear parametric policy of the form: \begin{equation} \pi^i(s)=\Phi(s)^T\cdot \theta_i \label{equ-lin-pol} \end{equation} where $\Phi(s)$ represents the \emph{state feature vector} with respect to any input state $s$. For simplicity, we study the special case of scalar actions. However, the analysis can be easily extended to high-dimensional action spaces. Meanwhile, we use $Sin()$ and $Mul()$ to represent respectively the action output of a policy trained for one iteration on the same state $s$ by using either the single-step method or the multi-step method in \eqref{equ-mu-new}. The single-step method can be considered as a special case of the multi-step method with $\rho_2=-1$ and $\rho_0=\rho_1=0$. Using these notations, Proposition \ref{the-2} is presented below. \begin{proposition} When each base learner $L_i$, $i\in\{1,\ldots,N\}$, trains its linear parametric policy $\pi^i$ with policy parameters $\theta_i$ on any state $s\in\mathcal{S}$ and when $0<\rho_0<\frac{1}{3}$, \begin{enumerate} \item $Sin(\pi^e(s))=\mathbb{E}\left[ Mul(\pi^e(s)) \right]$; \item $\sum_{i\in\{1,\ldots,N\}}\mathbb{E}\left[ \left( Mul(\pi^i(s))-\mathbb{E}\left[ Mul(\pi^e(s)) \right] \right)^2 \right] \\ < \sum_{i\in\{1,\ldots,N\}}\left( Sin(\pi^i(s))-Sin(\pi^e(s)) \right)^2 \\ =\sum_{i\in\{1,\ldots,N\}}\left( \pi^i(s)-\pi^e(s) \right)^2$ \end{enumerate} where the expectations above are taken with respect to any randomly selected $p,q\in\{1,\ldots,N\}$ in \eqref{equ-boot-list}. \label{the-2} \end{proposition} Proposition \ref{the-2} indicates that multi-step training in \eqref{equ-mu-new} is expected to reduce the difference between the action output of any base learner and that of the ensemble. Meanwhile the amount of action changes applied to $\pi^e$ remains identical to the single-step method. Therefore, using the multi-step policy training method developed in this section helps to enhance consistent behaviors among all base learners of the ensemble. \begin{table*}[!ht] \caption{Final performance of all competing algorithms on 9 benchmark problems. The results are shown with mean cumulative rewards and standard deviation over 10 independent algorithm runs. For each run, the cumulative rewards are obtained by averaging over 50 independent testing episodes.} \centering \resizebox{1.0\textwidth}{!}{ \begin{tabular}{c||cccc|c} \hline & TD3 & SAC & ED2 & SUNRISE & HED \\ \hline Hopper-v0 (PyBullet) & 917.38$\pm$178.46 & 1365.47$\pm$281.3 & 2095.54$\pm$148.86 & 1976.86$\pm$311.24 & \textbf{2359.63$\pm$50.28} \\ InvertedDoublePendulum-v0 (PyBullet) & 4394.07$\pm$558.8 & 8664.3$\pm$187.83 & 8733.3$\pm$1490.51 & 8746.82$\pm$58.68 & \textbf{9351.56$\pm$13.65} \\ InvertedPendulum-v0 (PyBullet) & 484.08$\pm$49.64 & 937.33$\pm$8.0 & \textbf{999.51$\pm$1.0} & 933.58$\pm$9.03 & 995.96$\pm$7.56 \\ Reacher-v0 (PyBullet) & 8.42$\pm$1.07 & 17.43$\pm$0.66 & 15.36$\pm$2.24 & \textbf{17.65$\pm$0.44} & 17.35$\pm$0.82 \\ Hopper-v3 (Mujoco) & 1399.54$\pm$250.32 & 2369.61$\pm$906.85 & 3043.19$\pm$971.53 & 2913.56$\pm$475.48 & \textbf{3503.08$\pm$83.35} \\ Humanoid-v3 (Mujoco) & 458.77$\pm$116.08 & 1720.54$\pm$525.07 & 868.17$\pm$384.58 & 3614.89$\pm$1402.7 & \textbf{3925.81$\pm$1029.59} \\ LunarLanderContinuous-v2 & 254.98$\pm$17.86 & 254.03$\pm$27.52 & 268.37$\pm$6.9 & \textbf{278.76$\pm$2.34} & 278.23$\pm$3.51 \\ Swimmer-v3 (Mujoco) & 68.52$\pm$33.54 & 40.37$\pm$1.01 & 87.81$\pm$28.79 & 49.45$\pm$0.47 & \textbf{89.23$\pm$26.78} \\ Walker2D-v3 (Mujoco) & 1965.25$\pm$248.21 & 1503.35$\pm$818.87 & 3298.73$\pm$1282.64 & 3766.55$\pm$1063.0 & \textbf{4442.8$\pm$408.88} \\ \hline \end{tabular}} \label{tab:final_perf_comp} \end{table*} \begin{figure*}[!hbt] \begin{center} \subfloat[Hopper-v0 (PyBullet)]{\label{fig-hc}\includegraphics[width=0.33\linewidth]{exp/HopperPyBulletEnv-v0_SR.pdf}} \subfloat[InvertedDoublePendulum-v0 (PyBullet)]{\label{fig-idpPB}\includegraphics[width=0.33\linewidth]{exp/InvertedDoublePendulumPyBulletEnv-v0_SR.pdf}} \subfloat[InvertedPendulum-v0 (PyBullet)]{\label{fig-ipPB}\includegraphics[width=0.33\linewidth]{exp/InvertedPendulumPyBulletEnv-v0_SR.pdf}}\\ \subfloat[Reacher-v0 (PyBullet)]{\label{fig-reacher}\includegraphics[width=0.33\linewidth]{exp/ReacherPyBulletEnv-v0_SR.pdf}} \subfloat[Hopper-v3 (Mujoco)]{\label{fig-hopper}\includegraphics[width=0.33\linewidth]{exp/Hopper-v3_SR.pdf}} \subfloat[Humanoid-v3 (Mujoco)]{\label{fig-humanoid}\includegraphics[width=0.33\linewidth]{exp/Humanoid-v3_SR.pdf}}\\ \subfloat[LunarLanderContinuous-v2]{\label{fig-lunarLander}\includegraphics[width=0.33\linewidth]{exp/LunarLanderContinuous-v2_SR.pdf}} \subfloat[Swimmer-v3 (Mujoco)]{\label{fig-swimmer}\includegraphics[width=0.33\linewidth]{exp/Swimmer-v3_SR.pdf}} \subfloat[Walker2D-v3 (Mujoco)]{\label{fig-walker}\includegraphics[width=0.33\linewidth]{exp/Walker2d-v3_SR.pdf}}\\ \end{center} \caption{Learning curves of HED and four baseline algorithms (i.e., TD3, SAC, ED2 and SUNRISE) on 9 benchmark RL problems.} \label{fig:training_perf} \end{figure*} \section{Experiment} \label{sec-exp} This section presents the experimental evaluation of HED, in comparison to several state-of-the-art DRL algorithms. The experiment setup is discussed first. Detailed experiment results are further presented and analyzed. \subsection{Experiment setting} We implement HED based on the high-quality implementation of TD3 provided by the publicly available OpenAI Spinning Up repository~\cite{SpinningUp2018}. We also follow closely the hyper-parameter settings of TD3 recommended in \cite{fujimoto2018} to build each base learner of HED. Specifically, based on~\cite{schulman2017}, a fully connected MLP with two hidden layers of 64 ReLU units is adopted to model all policy networks and Q-networks. Similar to \cite{januszewski2021,lee2021sunrise}, HED employs $5$ base learners, i.e., $N=5$. Each base learner has its own policy network and Q-network. Meanwhile, HED maintains and trains a separate ensemble Q-network with the same network architecture design. Each base learner trains its Q-network and also conducts the low-level training of the policy network repeatedly whenever HED collects 50 consecutive state-transition samples from the learning environment. Meanwhile, high-level policy training as well as the training of the ensemble Q-network is performed immediately after HED samples a full episode. HED adopts a separate Adam optimizer with the fixed learning rate of $5\mathrm{e}{-4}$ to train each Q-network and policy network. Furthermore, $\rho_0$ in \eqref{equ-mu-new} is set to 0.0001 for the main experiment results reported in Figure \ref{fig:training_perf}. The mini-batch size $|\mathcal{B}|$ is set to 100, following existing research \cite{januszewski2021} without any fine-tuning. HED is compared against four state-of-the-art DRL algorithms, including two Ensemble DRL algorithms, i.e., ED2~\cite{januszewski2021} and SUNRISE \cite{lee2021sunrise}), and two widely used off-policy DRL algorithms, i.e., SAC \cite{haarnoja2018} and TD3 \cite{fujimoto2018}. We evaluate their performance on 9 challenging continuous control benchmark problems, including four PyBullet benchmark problems \cite{benelot2018} (i.e., Hopper-v0, InvertedDoublePendulum-v0, InvertedPendulum-v0, and Reacher-v0), four Mujoco control tasks (i.e., Hopper-v3, Humanoid-v3, Swimmer-v3, and Walker2D-v3), and LunarLanderContinuous-v2 provided by OpenAI Gym \cite{openai_gym} . In the literature, PyBullet benchmarks are often considered to be more challenging than Mujoco benchmarks. Hence we decide to evaluate the performance of HED on both PyBullet and Mujoco benchmarks. The maximum episode length for each benchmark is fixed to 1000 timesteps. Each algorithm is run independently with 10 random seeds on all benchmarks. Besides the hyper-parameter settings of HED highlighted above, detailed hyper-parameter settings of all the competing algorithms have been summarized in Appendix~C. \subsection{Experiment result} \subsubsection{Performance comparison:} We compare HED against four cutting-edge DRL algorithms on 9 benchmark problems. Table~\ref{tab:final_perf_comp} presents the average cumulative rewards obtained by the policy networks (or policy ensembles for ensemble DRL algorithms) trained by all the competing algorithms across the same number of sampled episodes with respect to each benchmark. As evidenced in the table, HED achieved consistently the best performance\footnote{HED significantly outperformed ED2 on most benchmark problems, thanks to its use of the proposed high-level policy training technique.} on most of the benchmark problems except InvertedPendulum, Reacher, and LunarLander. Meanwhile, on InvertedPendulum, Reacher, and LunarLander, HED achieved very competitive performance with at least 98\% of the maximum cumulative rewards reached by the highest performing competing algorithms. Furthermore, on some problems such as Hopper-v0 and Walker2D-v3, HED outperformed the lowest performing algorithm by up to 150\% and the algorithm with the second highest performance by up to 18\%. In addition to Table~\ref{tab:final_perf_comp}, we also compared the learning curves of all the competing algorithms in Figure \ref{fig:training_perf}. As demonstrated in this figure, by explicitly strengthening inter-learner collaboration, HED converges clearly faster and is more stable during the learning process than other competing algorithms. Specifically, on several benchmark problems, such as Hopper-v0, InvertedDoublePendulum-v0, and Hopper-v3, HED achieved significantly lower variations in learning performance across 10 independent runs. In comparison to other ensemble DRL algorithms, the learning curves of HED also appear to be smoother on several benchmark problems, such as Walker2D-v3, suggesting that HED can achieve highly competitive stability during learning. \subsubsection{Impact of $\rho_0$:} To investigate the performance impact of $\rho_0$, we tested 4 different settings of $\rho_0$, ranging from $5\mbox{e$-$}05$ to $0.01$, on the InvertedPendulum-v0 and Hopper-v0 problems (similar observations can be found on other benchmark problems and are omitted in this paper). The learning curves are plotted in Figure~\ref{fig:rho_impact}. It is witnessed in the figure that HED can convergence faster under suitable settings of $\rho_0$. However, the ``best'' $\rho_0$ varies on different benchmark problems. For example, $\rho_0=0.005$ (green curve) converges slightly faster than other settings on InvertedPendulum-v0 while $\rho_0=5e-05$ (blue curve) converges slightly faster on Hopper-v0. Nevertheless, the impact of different $\rho_0$ on the final performance appears to be small as long as $\rho_0$ is reasonably small according to Proposition \ref{the-1}. \begin{figure} \begin{center} \subfloat[InvertedPendulum-v0]{\label{fig-invertedPendulum-rho}\includegraphics[width=0.4\linewidth]{exp/rho_impact/InvertedPendulumPyBulletEnv-v0_rho.pdf}} \subfloat[Hopper-v0]{\label{fig-hp-rho}\includegraphics[width=0.4\linewidth]{exp/rho_impact/HopperPyBulletEnv-v0_rho.pdf}} \end{center} \caption{The impact of using different $\rho_0$ in \eqref{equ-mu-new} on the performance of HED. $\rho_1$ and $\rho_2$ in \eqref{equ-mu-new} depend directly on $\rho_0$ according to Proposition \ref{the-1}.} \label{fig:rho_impact} \end{figure} \subsubsection{Ablation study on high-level policy training techniques:} High-level policy training can be conducted repeatedly whenever HED obtains either a new sampled episode or a fixed number of consecutive state-transition samples (e.g., samples collected from 50 consecutive timesteps). To understand which approach is more effective, experimental comparisons have been conducted in Appendix D with detailed performance results. According to the experiment results in Appendix D, on a majority of benchmark problems, episodic learning can produce more stable learning behavior and also makes HED converge faster. We also compared HED with its variation that performs high-level policy training by using the single-step method in \eqref{equ-e-pg} instead of the multi-step method in \eqref{equ-mu-new}. Detailed experiment results can be found in Appendix E. Our experiment results confirm that multi-step training in \eqref{equ-mu-new} enables HED to achieve significantly higher performance and learning stability than using the conventional single-step training technique in \eqref{equ-e-pg}. Hence, by explicitly sharing learned policy parameters among base learners in an ensemble through \eqref{equ-mu-new}, HED can effectively enhance inter-learner collaboration and boost the learning process. \section{Conclusions} \label{sec-con} In this paper, we conducted in-depth study of ensemble DRL algorithms, which have achieved cutting-edge performance on many benchmark RL problems in the recent literature. Different from existing research works that rely mainly on each base learner of an ensemble to train its policy network individually, we developed a new HED algorithm to explore the potential of training all base learners in a hierarchical manner in order to promote inter-learner collaboration and improve the collective performance of an ensemble of trained base learners. Specifically, we adopted existing ensemble DRL algorithms such as ED2 to perform low-level policy training. Meanwhile, a new multi-step training technique was developed for high-level policy training in HED to facilitate direct inter-learner parameter sharing. Both theoretical and empirical analysis showed that the HED algorithm can achieve stable learning behavior. It also outperformed several state-of-the-art DRL algorithms on multiple benchmark RL problems. \section*{Appendix A} This appendix presents a proof of Proposition \ref{the-1}. According to \cite{scieur2017}, any multi-step integration methods including \eqref{equ-mu-new} must satisfy three conditions to ensure its stability. They together guarantee that $x_k$ can converge to $\theta_i^*$ as $k$ approaches to $\infty$. We check each condition one-by-one below to derive the main conclusions in Proposition \ref{the-1}. \vspace{0.2cm} \noindent \textbf{Consistency condition}: We can re-write \eqref{equ-mu-new} as below $$ x_{k+3}+\rho_2 x_{k+2}+\rho_1 x_{k+1}+\rho_0 x_k= h\cdot g(x_{k+2}). $$ Define the \emph{shift operator} $F$, which maps $Fx_k\rightarrow x_{k+1}$. Furthermore, with $g(x_k)$ being simplified to $g_k$, $F$ also maps $Fg_k\rightarrow g_{k+1}$. Using $F$, \eqref{equ-mu-new} can be further written as $$ \rho(F)x_k=h\sigma(F)g_k, \forall k\geq 0, $$ where $$ \rho(F)=F^3+\rho_2 F^2+\rho_1 F+\rho_0, \sigma(F)=F^2. $$ The consistency condition requires $$ \rho(1)=0, \rho'(1)=\sigma(1). $$ This implies that \begin{equation*} \begin{split} & 1+\rho_2+\rho_1+\rho_0=0, \\ & 3+2\rho_2+\rho_1=1. \end{split} \end{equation*} Solving the above equations leads to $$ \rho_1=-2\rho_0,\ \rho_2=\rho_0-1. $$ Hence, \eqref{equ-mu-new} becomes $$ x_{k+3}=(1-\rho_0)x_{k+2}+2\rho_0 x_{k+1} -\rho_0 x_k+h\cdot g(x_{k+2}). $$ \vspace{0.2cm} \noindent \textbf{Zero-stability condition}: This condition requires all roots of $\rho(F)$ to be in the unit disk. Any roots on the unit circle must be simple. In other words, $$ \left| Roots(\rho(F)) \right|\leq 1. $$ In fact, $\rho(F)$ has three roots. They are $$ 1, \frac{1}{2}\left(-\rho_0 \pm \sqrt{\rho_0(\rho_0+4)} \right). $$ It is easy to verify that when $0<\rho_0<\frac{1}{2}$, $$ \left| \frac{1}{2}\left(-\rho_0 - \sqrt{\rho_0(\rho_0+4)} \right) \right|<1. $$ Meanwhile, when $\rho_0>0$, $$ \left| \frac{1}{2}\left(-\rho_0 + \sqrt{\rho_0(\rho_0+4)} \right) \right|<1. $$ In summary, the zero-stability condition requires $$ 0<\rho_0<\frac{1}{2}. $$ \vspace{0.2cm} \noindent \textbf{Absolute stability condition}: Define \begin{equation*} \begin{split} \Pi_{\lambda h} & \overset{\Delta}{=}\rho(F)+\lambda h\sigma(F) \\ &=F^3+\rho_2 F^2+\rho_1 F+\rho_0+\lambda h F^2 \\ &=F^3+(\rho_0-1) F^2 - 2\rho_0 F+\rho_0 +\lambda h F^2 \\ &=F^3 + (\rho_0-1+\lambda h)F^2-2\rho_0 F+\rho_0. \end{split} \end{equation*} Further define $$ r_{max}=\max_{\lambda\in [L,U]} \max_{r\in Roots(\Pi_{\lambda h}(F))}|r|, $$ where $L$ and $U$ in this appendix refer respectively to the smallest and the largest positive eigenvalues of matrix $A$ in \eqref{equ-grad-linear}. The absolute stability condition requires \begin{equation} r_{max}<1. \label{equ-abs-con} \end{equation} Let \begin{equation*} \begin{split} & B=\rho_0-1+\lambda h, \\ & C=-2\rho_0, \\ & D=\rho_0. \end{split} \end{equation*} Subsequently, define \begin{equation*} \begin{split} & A_0=1-B+C-D=2-\lambda h-4\rho_0, \\ & A_1=3-B-C-3D=4-\lambda h -2\rho_0, \\ & A_2=3+B-C-3D=2+\lambda h, \\ & A_3=1+B+C+D=\lambda h. \end{split} \end{equation*} According to the Routh-Hurwitz criterion \cite{nise2020}, the following two conditions jointly guarantee \eqref{equ-abs-con}: \begin{equation*} \begin{split} A_1,A_2,A_3,A_4>0, \\ A_1 A_2 > A_0 A_3. \end{split} \end{equation*} Specifically, the first condition above gives rise to the following: $$ \lambda h >0,\ \lambda h+2\rho_0<4,\ \lambda h+4\rho_0<2. $$ Following the second condition above, we can deduce the below: $$ \lambda h>2-\frac{4}{\rho_0}. $$ Given that $\lambda h>0$, we have \begin{equation*} \begin{split} & \lambda h > \max\left\{ 0, 2-\frac{4}{\rho_0} \right\}, \\ & \lambda h < \min\left\{ 2-4\rho_0, 4-2\rho_0 \right\}. \end{split} \end{equation*} Since $0<\rho_0<\frac{1}{2}$, $$ 2-4\rho_0<4-2\rho_0,\ 2-\frac{4}{\rho_0}<0. $$ Consequently $$ 0<\lambda h<2-4\rho_0. $$ Clearly, with sufficiently small $h$, the above condition on absolute stability can be easily satisfied. Hence, we can use \eqref{equ-mu-new} to perform high-level policy training stably in the HED algorithm. \section*{Appendix B} This appendix presents a proof of Proposition \ref{the-2}. Considering any specific state $s\in\mathcal{S}$, let $$ \nabla_a Q^e(s,a)|_{a=\pi^e(s)}=C, $$ where $C$ is an arbitrary scalar constant, in line with the assumption of scalar actions. Using \eqref{equ-pe} and \eqref{equ-lin-pol}, the ensemble policy gradient with respect to policy parameters $\theta_i$ of policy $\pi^i$, $i\in[1,\ldots,N]$, is $$ \nabla_{\theta_i} J(\pi^e)=\frac{C\Phi(s)}{N}. $$ According to the multi-step learning rule in \eqref{equ-mu-new}, updating $\theta_i$ for one iteration gives the updated $\theta_i$ as $$ (1-\rho_0)\theta_i+2\rho_0\theta_q-\rho_0\theta_p+h\frac{C\Phi(s)}{N}. $$ Therefore, \begin{equation*} Mul(\pi^i(s))=(1-\rho_0)\pi^i(s)+2\rho_0\pi^q(s)-\rho_0\pi^p(s) +\frac{hC}{N}\Phi(s)^T\Phi(s). \end{equation*} Hence $$ \mathbb{E}\left[ Mul(\pi^i(s)) \right]=(1-\rho_0)\pi^i(s)+\rho_0\pi^e(s)+\frac{h C}{N}\Phi(s)^T\Phi(s), $$ $$ \mathbb{E}\left[ Mul(\pi^e(s)) \right]=\pi^e(s)+\frac{h C}{N}\Phi(s)^T\Phi(s). $$ In comparison, upon using the single-step method, the updated $\theta_i$ becomes $$ \theta_i+h\frac{C\Phi(s)}{N}. $$ Subsequently, $$ Sin(\pi^i(s))=\pi^i(s)+\frac{h C}{N}\Phi(s)^T\Phi(s), $$ $$ Sin(\pi^e(s))=\pi^e(s)+\frac{h C}{N}\Phi(s)^T\Phi(s). $$ Clearly, $$ Sin(\pi^e(s))=\mathbb{E}\left[ Mul(\pi^e(s)) \right]. $$ Hence, the expected action changes applied to $\pi^e(s)$ are identical, regardless of whether single-step or multi-step method is used for high-level policy training\footnote{We assume in Proposition \ref{the-2} that high-level policy training is performed for one iteration on a specific state $s$.}. Define $$ \Delta=\sum_{i\in[1,\ldots,N]}(\pi^i(s)-\pi^e(s)). $$ For the single-step method, after all base learners trained their respective policies for one iteration on state $s$, it is easy to verify that \begin{equation*} \begin{split} & \sum_{i\in[1,\ldots,N]}\left( Sin(\pi^i(s))-Sin(\pi^e(s)) \right)^2\\ =& \sum_{i\in[1,\ldots,N]}\left( \pi^i(s)-\pi^e(s) \right)^2\\ =& \Delta. \end{split} \end{equation*} Meanwhile, \begin{equation*} \begin{split} & \left( Mul(\pi^i(s))-\mathbb{E}\left[ Mul(\pi^e(s)) \right] \right)^2 \\ =&\left( (1-\rho_0) (\pi^i(s)-\pi^e(s)) + 2\rho_0 (\pi^q(s)-\pi^e(s)) -\rho_0(\pi^p(s)-\pi^e(s)) \right)^2\\ =&(1-\rho_0)^2(\pi^i(s)-\pi^e(s))^2+4\rho_0^2(\pi^q(s)-\pi^e(s))^2+\rho^2(\pi^p(s)-\pi^e(s))^2\\ &+4(1-\rho_0)\rho_0(\pi^i-\pi^e(s))(\pi^q(s)-\pi^e(s)) \\ &-2(1-\rho_0)\rho_0(\pi^i(s)-\pi^e(s))(\pi^p(s)-\pi^e(s))\\ &-4\rho_0^2(\pi^q(s)-\pi^e(s))(\pi^p(s)-\pi^e(s)). \end{split} \end{equation*} Since the base learner indices $p$ and $q$ are randomly and independently selected, $$ \mathbb{E}\left[(\pi^i(s)-\pi^e(s))(\pi^p(s)-\pi^e(s))\right]=0, $$ $$ \mathbb{E}\left[(\pi^i(s)-\pi^e(s))(\pi^q(s)-\pi^e(s))\right]=0, $$ $$ \mathbb{E}\left[(\pi^q(s)-\pi^e(s))(\pi^p(s)-\pi^e(s))\right]=0. $$ Therefore \begin{equation*} \begin{split} & \sum_{i\in[1,\ldots,N]} \mathbb{E}\left[ \left( Mul(\pi^i(s))-\mathbb{E}\left[ Mul(\pi^e(s)) \right] \right)^2 \right] \\ =&(1-\rho_0)^2\Delta+4\rho_0^2\Delta+\rho_0^2\Delta \\ =&(1-2\rho_0+6\rho_0^2)\Delta. \end{split} \end{equation*} When $0<\rho_0<\frac{1}{3}$, $$ 1-2\rho_0+6\rho_0^2<1. $$ As a result, \begin{equation*} \begin{split} &\sum_{i\in[1,\ldots,N]}\mathbb{E}\left[ \left( Mul(\pi^i(s))-\mathbb{E}\left[ Mul(\pi^e(s)) \right] \right)^2 \right] \\ <&\sum_{i\in[1,\ldots,N]}\left( Sin(\pi^i(s))-Sin(\pi^e(s)) \right)^2. \end{split} \end{equation*} \section*{Appendix C} Table~\ref{tab:hyper-para} provides detailed hyper-parameter settings of all competing algorithms. Our hyper-parameter settings follow strictly the recommended settings in \cite{fujimoto2018,haarnoja2018,januszewski2021,lee2021sunrise}. \begin{table}[htb!] \caption{Hyper-parameter settings of competing algorithms.} \label{tab:hyper-para} \centering \resizebox{0.7\linewidth}{!}{ \begin{tabular}{l||llll} \hline Hyper-parameter & TD3 & SAC & ED2 & SUNRISE \\ \hline Num. episodes & 2500 & 2500 & 2500 & 3000 \\ Episode length & 1000 & 1000 & 1000 & 1000 \\ Minibatch size & 100 & 100 & 100 & 256 \\ Adam learning rate & 5e-4 & 3e-4 & 5e-4 & 3e-4 \\ Discount ($\gamma$) & 0.99 & 0.99 & 0.99 & 0.99 \\ GAE parameter ($\lambda$) & 0.995 & 0.995 & 0.995 & 0.995 \\ Replay buffer size & 1e6 & 1e6 & 1e6 & 1e6 \\ Update interval & 50 & 50 & 50 & 50 \\ Ensemble size & - & - & 5 & 5 \\ \hline \end{tabular}} \end{table} All experiments were run using a cluster of Linux computing nodes. Each node is equipped with 16 GB memory. The CPU specification is provided in Table~\ref{tab:cpu-para}. Each experiment was run in a Python virtual environment managed by Anaconda with Python packages specified in Table~\ref{tab:python-lib}. \begin{table}[htb!] \caption{CPU specification.} \label{tab:cpu-para} \centering \resizebox{0.6\linewidth}{!}{ \begin{tabular}{ll} \hline Architecture & x86\_64 \\ CPU op-mode(s) & 32-bit, 64-bit \\ CPU(s) & 16 \\ CPU family & 6 \\ Thread(s) per core & 2 \\ CPU max MHz & 4900.0000 \\ CPU min MHz & 800.0000 \\ Model name & 11th Gen Intel(R) Core(TM)\\ & i7-11700 @ 2.50GHz \\ \hline \end{tabular} } \end{table} \begin{table}[htb!] \caption{Python packages.} \label{tab:python-lib} \centering \begin{tabular}{ll} \hline Package name &Version \\ \hline cython &0.29.25 \\ gym &0.21.0 \\ keras &2.7.0 \\ mujoco-py &2.1.2.14 \\ numpy &1.21.4 \\ pybulletgym &0.1 \\ python &3.7.11 \\ scipy &1.7.3 \\ tensorflow &2.7.0 \\ \hline \end{tabular} \end{table} \begin{figure*}[!hbt] \begin{center} \subfloat[Hopper-v0 (PyBullet)]{\label{fig-hc-eq9}\includegraphics[width=0.33\linewidth]{exp/when2train_HighLevelPolicy/HopperPyBulletEnv-v0_when2HLtrain.pdf}} \subfloat[InvertedDoublePendulum-v0 (PyBullet)]{\label{fig-idpPB-when2HLtrain}\includegraphics[width=0.33\linewidth]{exp/when2train_HighLevelPolicy/InvertedDoublePendulumPyBulletEnv-v0_when2HLtrain.pdf}} \subfloat[InvertedPendulum-v0 (PyBullet)]{\label{fig-ipPB-when2HLtrain}\includegraphics[width=0.33\linewidth]{exp/when2train_HighLevelPolicy/InvertedPendulumPyBulletEnv-v0_when2HLtrain.pdf}}\\ \subfloat[Reacher-v0 (PyBullet)]{\label{fig-reacher-when2HLtrain}\includegraphics[width=0.33\linewidth]{exp/when2train_HighLevelPolicy/ReacherPyBulletEnv-v0_when2HLtrain.pdf}} \subfloat[Hopper-v3 (Mujoco)]{\label{fig-hopper-when2HLtrain}\includegraphics[width=0.33\linewidth]{exp/when2train_HighLevelPolicy/Hopper-v3_when2HLtrain.pdf}} \subfloat[Walker2D-v3 (Mujoco)]{\label{fig-walkerPB-when2HLtrain}\includegraphics[width=0.33\linewidth]{exp/when2train_HighLevelPolicy/Walker2d-v3_when2HLtrain.pdf}}\\ \end{center} \caption{Learning curves of HED with respect to two high-level policy training approaches. The method that conducts high-level policy training after every 50 timesteps is denoted as \emph{Fixed timestep}. The method that conducts high-level policy training at the end of each sampled episode is denoted as \emph{Each episode}.} \label{fig:when2HLtrain_impact} \end{figure*} \begin{figure*}[!hbt] \begin{center} \subfloat[Hopper-v0 (PyBullet)]{\label{fig-hc-eq9}\includegraphics[width=0.33\linewidth]{exp/eq9_impact/HopperPyBulletEnv-v0_EQ9.pdf}} \subfloat[InvertedDoublePendulum-v0 (PyBullet)]{\label{fig-idpPB-eq9}\includegraphics[width=0.33\linewidth]{exp/eq9_impact/InvertedDoublePendulumPyBulletEnv-v0_EQ9.pdf}} \subfloat[InvertedPendulum-v0 (PyBullet)]{\label{fig-ipPB-eq9}\includegraphics[width=0.33\linewidth]{exp/eq9_impact/InvertedPendulumPyBulletEnv-v0_EQ9.pdf}}\\ \subfloat[Reacher-v0 (PyBullet)]{\label{fig-reacher-eq9}\includegraphics[width=0.33\linewidth]{exp/eq9_impact/ReacherPyBulletEnv-v0_EQ9.pdf}} \subfloat[Hopper-v3 (Mujoco)]{\label{fig-hopper-eq9}\includegraphics[width=0.33\linewidth]{exp/eq9_impact/Hopper-v3_EQ9.pdf}} \subfloat[Walker2D-v3 (Mujoco)]{\label{fig-walkerPB-eq9}\includegraphics[width=0.33\linewidth]{exp/eq9_impact/Walker2d-v3_EQ9.pdf}}\\ \end{center} \caption{Learning curves of HED using the single-step high-level policy training technique in \eqref{equ-e-pg} vs. the proposed multi-step high-level policy training technique in \eqref{equ-mu-new}.} \label{fig:eq9_impact} \end{figure*} \section*{Appendix D} In this appendix, we study the effectiveness of conducting high-level policy training after HED obtains a full sampled episode. Figure~\ref{fig:when2HLtrain_impact} shows the performance comparison of HED with two different training frequencies: every 50 consecutive timesteps vs. every episode. It can be noticed that, on a majority benchmark problems (i.e., 5 out of 6), performing high-level policy training after every episode (orange curve) can significantly improve the HED algorithm in terms of both the final performance and convergence speed. For example, as shown in Figure~\ref{fig:when2HLtrain_impact}(e), the orange curve reaches 3500 after 1500 episodes while the blue curve converges to a lower cumulative reward (approx. 3000) after 2000 episodes. We also notice that episodic policy training is more robust to the randomness in the environment and less sensitive to the initialization of neural network weights. For example, in Figure~\ref{fig:when2HLtrain_impact}(e), episodic policy training produces a smaller confident interval (orange shaded area) compared to the fixed timestep training (blue shaded area) over 10 independent algorithm runs. Similar results can also be observed from Figures~\ref{fig:when2HLtrain_impact}(a) and (f). Note that in each algorithm run, both policy networks and Q-networks are initialized with different weights. The environment initial states also vary. \section*{Appendix E} This appendix investigates the effectiveness of multi-step policy training by using \eqref{equ-mu-new}. Specifically, we compare the performance of HED against its variant, which performs single-step high-level policy training by using \eqref{equ-e-pg}, on 6 problems that include both PyBullet and Mujoco benchmarks. As shown in Figure~\ref{fig:eq9_impact}, with the help of \eqref{equ-mu-new}, significant performance improvement can be observed (orange curve) on most benchmark problems. In particular, HED behaves more stably during the learning process. For example, in Figure~\ref{fig:eq9_impact}(c), the cumulative rewards obtained by the policy trained using~\eqref{equ-mu-new} (orange curve) remain stable at 1000 after 300 episodes. In comparison, the blue curve stays below 1000 and fluctuates between 800 and 1000 over the entire learning period. Similar trends can also be noticed in Figure~\ref{fig:eq9_impact}(b). The proposed multi-step policy training technique achieves clearly higher cumulative rewards. In the Hopper environment shown in Figure~\ref{fig:eq9_impact}(e), the orange curve outperforms the blue curve by up to 75\% after 2500 training episodes. Moreover, the orange curve converges to 3500 while the blue curve remains below 2000. The significant improvement in cumulative rewards can also be witnessed in Figure~\ref{fig:eq9_impact}(a) and (f). The shaded areas in Figure~\ref{fig:eq9_impact}(e) and (f) also show that the multi-step training technique is less sensitive to the environment randomness and neural network weight initialization, compared to using the conventional single-step training method in \eqref{equ-e-pg}. Hence, our experiment results confirm the importance of inter-learner collaboration. By enabling base learners in an ensemble to explicitly share their learned policy parameters, HED can achieve high learning stability and effectively boost the learning process. \bibliographystyle{IEEEtran}

6bb442a5f974a957ff5c121a164eb72dc243cfba

\section{Introduction} Recently, bit allocation for Neural Video Compression (NVC) has drawn growing attention thanks to its great potential in boosting compression performance. Due to the frame reference structure in video coding, it is sub-optimal to use the same R-D (Rate-Distortion) trade-off parameter $\lambda$ for all frames. In bit allocation task, bitrate is allocated to different frames/regions to minimize R-D cost $R+\lambda D$, where $R$ is total bitrate, $D$ is total distortion, and $\lambda$ is the Lagrangian multiplier controlling R-D trade-off. \citet{li2022rate} are the pioneer of bit allocation for NVC, who improve the empirical R-D (Rate-Distortion) model from traditional video codec \citep{li2014lambda,li2016lambda} and solve the per-frame Lagrangian multiplier $\lambda$. Other concurrent works adopt simple heuristics for coarse bit allocation \citep{cetin2022flexible,hu2022coarse}. Most recently, BAO (Bit Allocation using Optimization) \citep{bao2022} proposes to formulate bit allocation as semi-amortized variational inference (SAVI) \citep{kim2018semi,marino2018iterative} and solves it by gradient-based optimization. Specifically, it directly optimizes the variational posterior parameter to be quantized and encoded by gradient ascent, aiming at maximizing the minus overall R-D cost, which is also the evident lowerbound (ELBO). BAO does not rely on any empirical R-D model and thus outperforms previous work. Further, BAO shows its optimality by proving its equivalence to bit allocation with precise R-D model. In this paper, we first show that BAO \citep{bao2022} is in fact, sub-optimal due to its implementation. Specifically, we find that it abuses SAVI \citep{kim2018semi,marino2018iterative} on latent with non-factorized variational posterior, which brings incorrect gradient signal during optimization. To solve this problem, we first extend SAVI to non-factorized latent by back-propagating through gradient ascent \citep{domke2012generic}. Then based on that, we correct the sub-optimal bit allocation in BAO to produce true optimal bit allocation for NVC. Furthermore, we propose a computational feasible approximation to such correct but intractable bit allocation method. And we show that our approximation outperforms the incorrect bit allocation (BAO) in terms of R-D performance and bitrate error, and performs better than all other bit allocation methods. To summarize, our contributions are as follows: \begin{itemize} \item We demonstrate that a previously claimed optimal bit allocation method is actually sub-optimal. We find that its sub-optimality comes from the improper application of SAVI to non-factorized latent. \item We present the correct way to conduct SAVI on non-factorized latent by recursively applying back-propagation through gradient ascent. Based on this, we derive the corrected optimal bit allocation algorithm for NVC. \item Furthermore, we propose a computational efficient approximation of the optimal bit allocation to make it feasible. Our proposed approach improves the R-D performance and bitrate error over the incorrect bit allocation (BAO), and outperforms all other bit allocation methods for NVC. \end{itemize} \section{Preliminaries} \subsection{Neural Video Compression} The input of NVC is a GoP (Group of Picture) $\bm{x}_{1:T}$, where $\bm{x}_i\in R^{H\times W}$ is the $i^{th}$ frame with $H\times W$ pixels, and $T$ is the number of frame inside the GoP. Most of the works in NVC follow a latent variable model with temporal autoregressive relationship \citep{yang2020hierarchical}. Specifically, to encode $\bm{x}_i$, we first extract the motion latent $\bm{w}_i=f^{w}_{\phi}(\bm{x}_i,\bm{x}'_i)$ from current frame $\bm{x}_i$ and previous reconstructed frame $\bm{x}'_{i-1}$, where $f^{w}_{\phi}(\cdot)$ is the motion encoder parameterized by $\phi$\footnote{Following previous works in deep generative modeling \citep{kingma2013auto,kim2018semi}, we denote all parameters related to encoder as $\phi$, and all parameters related to decoder and prior as $\theta$.}. Then, we encode the quantized latent $\bmt{w}_i=\lfloor\bm{w}_i\rceil$ with the probability mass function (pmf) estimator $P_{\theta}(\bmt{w}_i|\bmt{w}_{<i},\bmt{y}_{<i})$ parameterized by $\theta$, where $\lfloor\cdot\rceil$ is the rounding. Then, we obtain the residual latent $\bm{y}_i=f_{\phi}^y(\bm{x},\bm{x}',\bmt{w})$, where $f_{\phi}^y(\cdot)$ is the residual encoder. Then, similar to how we treat $\bm{w}_i$, we encode the quantized latent $\bmt{y}_i=\lfloor\bm{y}_i\rceil$ with pmf $P_{\theta}(\bmt{y}_i|\bmt{w}_{\le i},\bmt{y}_{<i})$. Finally, we obtain the reconstructed frame $\bm{x}'_i=g^{x}_{\theta}(\bm{x}'_{i-1},\bmt{w}_i,\bmt{y}_i)$, where $g^{x}_{\theta}(\cdot)$ is the decoder parameterized by $\theta$. As only the motion latent $\bmt{w}_i$ and residual latent $\bmt{y}_i$ exist in the bitstream, the above process can be simplified as Eq.~\ref{eq:enc} and Eq.~\ref{eq:dec}, where $f_{\phi}(\cdot)$ is the generalized encoder and $g_{\theta}(\cdot)$ is the generalized decoder. The target of NVC is to minimize the per-frame R-D cost $R_i+\lambda_i D_i$ (Eq.~\ref{eq:rd}), where $R_i$ is the bitrate, $D_i$ is the distortion and $\lambda_i$ is the Lagrangian multiplier controlling R-D trade-off. The bitrate $R_i$ and distortion $D_i$ is computed as Eq.~\ref{eq:dec}, where $d(\cdot,\cdot)$ is the distortion metric. And $\lambda_i D_i$ can be further interpreted as the data likelihood term $-\log p_{\theta}(\bm{x}_i|\bmt{w}_{\le i},\bmt{y}_{\le i})$ so long as we treat $\lambda_i D_i$ as the energy function of a Gibbs distribution \citep{minnen2018joint}. Specifically, when $d(\cdot,\cdot)$ is MSE, we can interpret $\lambda_iD_i=-\log p_{\theta}(\bm{x}_i|\bmt{w}_{\le i},\bmt{y}_{\le i})+const$, where $p_{\theta}(\bm{x}_i|\bmt{w}_{\le i},\bmt{y}_{\le i})$ is a Gaussian distribution $\mathcal{N}(\bm{\hat{x}}_i,1/2\lambda_i I)$. \begin{align} \bm{w}_i = f_{\phi}(\bm{x}_i, \bmt{w}_{<i},\bmt{y}_{<i}),\bm{y}_i = f_{\phi}(\bm{x}_i, \bmt{w}_{\le i}, \bmt{y}_{<i})&\textrm{, where } \bmt{w}_i=\lfloor\bm{w}_i\rceil\textrm{, }\bmt{y}_i=\lfloor\bm{y}_i\rceil \label{eq:enc}\\ R_i=\log P_{\theta}(\bmt{w}_i,\bmt{y}_i|\bmt{w}_{<i},\bmt{y}_{<i})\textrm{, }D_i=d&(\bm{x}_i,g_{\theta}(\bmt{w}_{\le i},\bmt{y}_{\le i})) \label{eq:dec}\\ \max -(R_i + \lambda_i D_i&) \label{eq:rd} \end{align} On the other hand, NVC is also closely related to Variational Autoencoder (VAE) \citep{kingma2013auto}. As the rounding $\lfloor\cdot\rceil$ is not differentiable, \citet{balle2016end,Theis17} propose to relax it by additive uniform noise (AUN), and replace $\bmt{w}_i=\lfloor\bm{w}_i\rceil$, $\bmt{y}_i=\lfloor\bm{y}_i\rceil$ with $\bmt{w}_i=\bm{w}_i+\mathcal{U}(-0.5,0.5)$, $\bmt{y}_i=\bm{y}_i+\mathcal{U}(-0.5,0.5)$. Under such formulation, the above encoding-decoding process becomes a VAE on graphic model $\bmt{w}_{\le i}, \bmt{y}_{\le i} \rightarrow \bm{x}_i$ with variational posterior as Eq.~\ref{eq:q}, where $\bm{w}_i,\bm{y}_i$ plays the role of variational posterior parameter. Then, minimizing the overall R-D cost (Eq.~\ref{eq:rd}) is equivalent to maximizing the evident lowerbound (ELBO) (Eq.~\ref{eq:elbo}). \begin{align} \hspace{-0.5em}q_{\phi}(\bmt{w}_i|\bm{x}_i, \bmt{w}_{<i}, \bmt{y}_{<i}) = \mathcal{U}(\bm{w}_i-0.5,\bm{w}_i+0.5), q_{\phi}(\bmt{y}_i&|\bm{x}_i, \bmt{w}_{\le i}, \bmt{y}_{<i}) = \mathcal{U}(\bm{y}_i-0.5,\bm{y}_i+0.5) \label{eq:q}\\ -(R_i + \lambda_i D_i) = \mathbb{E}_{q_{\phi}}[\underbrace{\log P_{\theta}(\bmt{w}_i,\bmt{y}_i|\bmt{w}_{<i},\bmt{y}_{<i})}_{-R_i}&+\underbrace{\log p_{\theta}(\bm{x}_i|\bmt{w}_{\le i},\bmt{y}_{\le i})}_{-\lambda_i D_i}\underbrace{\cancel{-\log q_{\phi}}}_{\textrm{bits-back bitrate: 0}}] \label{eq:elbo} \end{align} \subsection{Bit Allocation for Neural Video Compression} \label{sec:bgba} It is well known to video coding community that using the same R-D trade-off parameter $\lambda_i$ to optimize R-D cost in Eq.~\ref{eq:rd} for all $T$ frames inside a GoP is suboptimal \citep{li2014lambda,li2016lambda}. This sub-optimality comes from the frame reference structure and is explained in detail by \citet{li2022rate, bao2022}. The target of bit allocation is to maximize the minus of overall R-D cost (ELBO) $\mathcal{L}$ as Eq.~\ref{eq:l0} given the overall R-D trade-off parameter $\lambda_0$, instead of maximizing $\mathcal{L}_i$ of each frame $i$ separately. The pioneer work of bit allocation in NVC \citep{li2022rate} follows bit allocation for traditional video codec \citep{li2016lambda}. Specifically, it adopts empirical models to approximate the relationship of the rate dependency $\partial R_{i+1}/\partial R_{i}$ and distortion dependency $\partial D_{i+1}/\partial D_{i}$ between frames. Then it takes those models into Eq.~\ref{eq:l0} to solve $\lambda_{1:T}^*$ explicitly as Eq.~\ref{eq:rd1}.\textit{left}. However, its performance heavily relies on the accuracy of empirical models. \begin{align} &\max \mathcal{L} = \sum_{i=1}^{T} \mathcal{L}_{i}\textrm{, where } \mathcal{L}_i = -(R_i + \lambda_0 D_i)\label{eq:l0}\\ \lambda^{*}_{1:T} \leftarrow \arg \max_{\lambda_{1:T}}& \mathcal{L} (\lambda_{1:T})\textrm{, versus } \bm{w}^{*}_{1:T},\bm{y}^{*}_{1:T} \leftarrow \arg \max_{\bm{w}_{1:T},\bm{y}_{1:T}} \mathcal{L}(\bm{w}_{1:T},\bm{y}_{1:T}) \label{eq:rd1} \end{align} On the other hand, BAO \citep{bao2022} does not solve $\lambda_{1:T}^{*}$ explicitly. Instead, it adopts SAVI \citep{kim2018semi,marino2018iterative} to achieve implicit bit allocation. To be specific, it initializes the variational posterior parameter $\bm{w}_{1:T}^0, \bm{y}_{1:T}^0$ from fully amortized variational inference (FAVI) as Eq.~\ref{eq:enc}. Then, it optimizes $\bm{w}_{1:T},\bm{y}_{1:T}$ via gradient ascent to maximize $\mathcal{L}$ as Eq.~\ref{eq:rd1}.\textit{right}. During this procedure, no empirical model is required. BAO further proofs that optimizing Eq.~\ref{eq:rd1}.\textit{right} is equivalent to optimizing Eq.~\ref{eq:rd1}.\textit{left} with precise rate and distortion dependency model $\partial R_{i+1}/\partial R_i,\partial D_{i+1}/\partial D_i$ (See Thm.~1, Thm.~2 in \citet{bao2022}). Thus, BAO claims that it is optimal assuming gradient ascent achieves global maximum. However, in next section, we show that BAO \citep{bao2022} is in fact suboptimal due to its implementation. \section{Why BAO is Sup-optimal} \label{sec:baosub} BAO \citep{bao2022} achieves the SAVI \citep{kim2018semi,marino2018iterative} target in Eq.~\ref{eq:rd1}.\textit{right} by gradient-based optimization. More specifically, its update rule is described as Eq.~\ref{eq:fa_gradw} and Eq.~\ref{eq:fa_grady}, where $K$ is the total number of gradient ascent steps, and $\bm{w}_i^k,\bm{y}_i^k$ is the posterior parameter $\bm{w}_i,\bm{y}_i$ after $k$ steps of gradient ascent. In the original paper of BAO, the authors also find that directly optimizing $\bm{w}_i,\bm{y}_i$ simultaneously by Eq.~\ref{eq:fa_gradw} and Eq.~\ref{eq:fa_grady} performs worse than optimizing $\bm{y}_i$ alone using Eq.~\ref{eq:fa_grady}, but they have not offered any explanation. It is obvious that optimizing $\bm{y}_i$ alone is sub-optimal. However, it is not obvious why jointly optimizing $\bm{w}_i, \bm{y}_i$ with Eq.~\ref{eq:fa_gradw} and Eq.~\ref{eq:fa_grady} fails. \begin{align} \bm{w}^{k+1}_i \leftarrow \bm{w}^k_i + \alpha\frac{ d\mathcal{L}(\bm{w}^k_{1:T},\bm{y}^k_{1:T})}{d \bm{w}^k_i}\textrm{, where }\frac{d \mathcal{L}(\bm{w}^k_{1:T},\bm{y}^k_{1:T})}{d \bm{w}^k_i}=\sum_{j=i}^{T}\frac{\partial \mathcal{L}_{j}(\bm{w}^k_{1:j},\bm{y}^k_{1:j})}{\partial \bm{w}^k_i}\label{eq:fa_gradw}\\ \bm{y}^{k+1}_i \leftarrow \bm{y}^k_i + \alpha\frac{d \mathcal{L}(\bm{w}^k_{1:T},\bm{y}^k_{1:T})}{d \bm{y}^k_i}\textrm{, where }\frac{d \mathcal{L}(\bm{w}^k_{1:T},\bm{y}^k_{1:T})}{d \bm{y}^k_i}=\sum_{j=i}^{T}\frac{\partial \mathcal{L}_{j}(\bm{w}^k_{1:j},\bm{y}^k_{1:j})}{\partial\bm{y}^k_i} \label{eq:fa_grady} \end{align} In fact, the update rule in Eq.~\ref{eq:fa_gradw} and Eq.~\ref{eq:fa_grady} is exactly the SAVI \citep{kim2018semi,marino2018iterative} when $\bm{w}_i,\bm{y}_i$ fully factorizes (e.g. the full factorization used in mean-field \citep{blei2017variational}). However, in NVC the $\bm{w}_i,\bm{y}_i$ has complicated auto-regressive relationships (See Eq.~\ref{eq:enc} and Fig.~\ref{fig:grad}.(a)). Abusing SAVI on non-factorized latent causes gradient error in two aspects: (1). The total derivative $d \mathcal{L}/d\bm{w}_i,d \mathcal{L}/d\bm{y}_i$ is incomplete. (2). The total derivative $d \mathcal{L}/d\bm{w}_i,d \mathcal{L}/d\bm{y}_i$ and partial derivative $\partial \mathcal{L}_j/\partial\bm{w}_i,\partial\mathcal{L}_j/\partial\bm{y}_i$ is evaluated at wrong value. In next two sections, we elaborate those two issues with $\bm{w}_i$ related equations in main text and $\bm{y}_i$ related equations in Appendix.~\ref{app:cf}. \begin{figure}[thb] \centering \includegraphics[width=\linewidth]{fig_grad.PNG} \caption{(a). The gradient structure of NVC without SAVI. (b). After $k$ step of SAVI/gradient ascent, the gradient structure of NVC is broken. (c). The proposed approach using back-propagating through gradient ascent. We mark the difference between (b) and (c) in red.} \label{fig:grad} \end{figure} \subsection{Incomplete Total derivative evaluation} \label{sec:itde} According to the latent generation procedure described by Eq.~\ref{eq:enc} and Eq.~\ref{eq:dec}, we draw the computational graph to describe the latent dependency as Fig.~\ref{fig:grad}.(a). Based on that, we expand the total derivative $d\mathcal{L}/d\bm{w}_i,d\mathcal{L}/d\bm{y}_i$ as Eq.~\ref{eq:incwg} and Eq.~\ref{eq:incyg}. \begin{align} \frac{d \mathcal{L}(\bm{w}_{1:T},\bm{y}_{1:T})}{d \bm{w}_i}=&\sum_{j=i}^{T}\frac{d \mathcal{L}_j(\bm{w}_{1:j},\bm{y}_{1:j})}{d \bm{w}_i}\notag\\ \frac{d \mathcal{L}_j(\bm{w}_{1:j},\bm{y}_{1:j})}{d \bm{w}_i}=&\underbrace{\sum_{l=i+1}^{j}\frac{\partial \bm{w}_l}{\partial\bm{w}_i}\frac{d \mathcal{L}_{j}(\bm{w}_{1:j},\bm{y}_{1:j})}{d \bm{w}_l}+\sum_{l=i}^{j}\frac{\partial \bm{y}_l}{\partial\bm{w}_i}\frac{d \mathcal{L}_{j}(\bm{w}_{1:j},\bm{y}_{1:j})}{d \bm{y}_l}}_{\textrm{ignored by BAO}}+\underbrace{\frac{\partial \mathcal{L}_j(\bm{w}_{1:j},\bm{y}_{1:j})}{\partial \bm{w}_i}}_{\textrm{considered by BAO}}\label{eq:incwg} \end{align} As shown in Eq.~\ref{eq:fa_gradw}, Eq.~\ref{eq:fa_grady} and Fig.~\ref{fig:grad}.(b), BAO \citep{bao2022} treats the total derivative $d\mathcal{L}/d\bm{w}_i,d\mathcal{L}/d\bm{y}_i$ as the sum of the frame level partial derivative $\partial \mathcal{L}_j/\partial\bm{w}_i,\partial \mathcal{L}_j/\partial\bm{y}_i$, which is the direct contribution of frame $i^{th}$ latent $\bm{w}_i,\bm{y}_i$ to $j^{th}$ frame's R-D cost $\mathcal{L}_j$ (as marked in Eq.~\ref{eq:incwg} and Eq.~\ref{eq:incyg}). This incomplete evaluation of gradient signal brings sub-optimality. Further, it is not possible to correct BAO by simply including other parts of gradient into consideration. As BAO jointly updates all the latent $\bm{w}_{1:T},\bm{y}_{1:T}$, the relationship of Eq.~\ref{eq:dec} only holds for the initial latent parameters $\bm{w}_{1:T}^0,\bm{y}_{1:T}^0$ produced by FAVI. And this important relationship is broken for parameters $\bm{w}_{1:T}^k,\bm{y}_{1:T}^k$ after $k\ge1$ steps of update. \subsection{Incorrect Value to Evaluate Gradient} \label{sec:ipde} As shown in Eq.~\ref{eq:fa_gradw} and Eq.~\ref{eq:fa_grady}, BAO \citep{bao2022} simultaneously updates all the posterior parameter $\bm{w}_{1:T},\bm{y}_{1:T}$ with gradient evaluated at the same gradient ascent step $\bm{w}_{1:T}^k,\bm{y}_{1:T}^k$. However, as we show later in Sec.~\ref{sec:savi2} and Fig.~\ref{fig:grad}.(c), this is sub-optimal as all the descendant latent $\bm{w}_{>i},\bm{y}_{\ge i}$ of $\bm{w}_i$ should already complete all $K$ steps of gradient ascent before the gradient of $\bm{w}_i$ is evaluated. Moreover, $\bm{w}_{>i},\bm{y}_{\ge i}$ should be initialized by FAVI using precedents latent. Similar rule applies to $\bm{y}_i$. Specifically, the correct value to evaluate the gradient is as Eq.~\ref{eq:incygw} and Eq.~\ref{eq:incygy}, where $\bm{w}_i^{k_i}$ denotes the latent $\bm{w}_i$ after $k_i$ steps of update, and $\bm{y}_i^{k'_j}$ denotes the latent $\bm{y}_i$ after $k'_i$ steps of update. \begin{align} \bm{w}^{k_i+1}_i \leftarrow \bm{w}^{k_i}_i + \alpha\frac{ d\mathcal{L}(\bm{w}_1^{k_1},...,\bm{w}_i^{k_i},\bm{w}^K_{>i},\bm{y}_1^{k'_1},...,\bm{y}_{i-1}^{k'_{i-1}},\bm{y}^K_{\ge i})}{d \bm{w}^{k_i}_i},\notag\\\textrm{where }\bm{w}_{>i}^0,\bm{y}_{\ge i}^0=f(\bm{x},\bm{w}_1^{k_1},...,\bm{w}_i^{k_i},\bm{y}_1^{k'_1},...,\bm{y}_{i-1}^{k'_{i-1}})\label{eq:incygw} \end{align} Similar to the incomplete total derivative evaluation, this problem does not have a simple solution. In next section, we show how to correct both of the above-mentioned issues by recursively applying back-propagating through gradient ascent \citep{domke2012generic}. \section{Correcting the Sub-optimal Bit Allocation} \label{sec:cba} In this section, we first extend the generic SAVI \citet{kim2018semi,marino2018iterative} to 2-level non-factorized latent. Then we further extend this result to latent with any dependency that can be described by a DAG (Directed Acyclic Graph). And finally, we correct the sub-optimal bit allocation by applying the result in DAG latent to NVC. \subsection{SAVI on 2-level non-factorized latent} \label{sec:savi2} In this section, we extend the SAVI on 1-level latent \citep{kim2018semi} to 2-level non-factorized latent. We denote $\bm{x}$ as evidence, $\bm{w}$ as the variational posterior parameter of the first level latent $\bmt{w}$, $\bm{y}$ as the variational posterior parameter of the second level latent $\bmt{y}$, and the ELBO to maximize as $\mathcal{L}(\bm{w},\bm{y})$. The posterior $q(\bmt{w},\bmt{y}|\bm{x})$ factorizes as $q(\bmt{w}|\bm{x})q(\bmt{y}|\bmt{w},\bm{x})$, which means that $\bm{y}$ depends on $\bm{w}$. Given $\bm{w}$ is fixed, we can directly follow \citet{kim2018semi,marino2018iterative} to optimize $\bm{y}$ to maximize ELBO by SAVI. However, it requires some tricks to optimize $\bm{w}$. \begin{minipage}[t]{.42\textwidth} % \vspace{0pt} \IncMargin{1.0em} \begin{algorithm}[H] \DontPrintSemicolon \caption{SAVI on 2-level Latent}\label{alg:solve-2} \textbf{procedure} solve-2-level($\bm{x},\bm{w}^{k}$)\; $\quad$initialize $\bm{w}^0\leftarrow f(\bm{x})$ from FAVI\; $\quad$\textbf{for} $k=0,...,K-1$ \textbf{do}\; $\quad\quad \frac{d\mathcal{L}(\bm{w}^k,\bm{y}^K)}{d\bm{w}^k}=\textrm{grad-2-level}(\bm{x},\bm{w}^k)$\; $\quad\quad \bm{w}^{k+1}\leftarrow\bm{w}^k+\alpha\frac{d\mathcal{L}(\bm{w}^k,\bm{y}^K)}{d\bm{w}^k}$\; $\quad$\textbf{return} $\bm{w}^{K},\bm{y}^K$\; \BlankLine \textbf{procedure} grad-2-level($\bm{x},\bm{w}^{k}$)\; $\quad\bm{y}^0\leftarrow f(\bm{x},\bm{w}^{k})$ from FAVI\; $\quad$\textbf{for} $k'=0,...,K-1$ \textbf{do}\; $\quad\quad \bm{y}^{k'+1}\leftarrow\bm{y}^{k'}+\alpha\frac{d\mathcal{L}(\bm{w}^{k},\bm{y}^{k'})}{d\bm{y}^{k'}}$\; $\quad\overleftarrow{\bm{w}}\leftarrow\frac{\partial \mathcal{L}(\bm{w}^{k},\bm{y}^K)}{\partial \bm{w}^{k}}$\; $\quad\overleftarrow{\bm{y}^K}\leftarrow\frac{d\mathcal{L}(\bm{w}^{k},\bm{y}^K)}{d\bm{y}^K}$\; $\quad$\textbf{for} $k'=K-1,...,0$ \textbf{do}\; $\quad\quad\overleftarrow{\bm{w}}\leftarrow\overleftarrow{\bm{w}}+\alpha\frac{\partial^2 \mathcal{L}(\bm{w}^{k},\bm{y}^{k'})}{\partial \bm{w}^{k}\partial\bm{y}^{k'}}\overleftarrow{\bm{y}^{k'+1}}$\; $\quad\quad\overleftarrow{\bm{y}^{k'}}\leftarrow\overleftarrow{\bm{y}^{k'}}+\alpha\frac{\partial^2 \mathcal{L}(\bm{w}^{k},\bm{y}^{k'})}{\partial \bm{y}^{k'}\partial\bm{y}^{k'}}\overleftarrow{\bm{y}^{k'+1}}$\; $\quad\overleftarrow{\bm{w}}=\overleftarrow{\bm{w}}+\frac{\partial \bm{y}^0}{\partial\bm{w}^k}\overleftarrow{\bm{y}^{0}}$\; $\quad$\textbf{return} $\frac{d\mathcal{L}(\bm{w}^k,\bm{y}^K)}{d\bm{w}^k}=\overleftarrow{\bm{w}}$\; \end{algorithm} \end{minipage} \begin{minipage}[t]{.58\textwidth} % \vspace{0pt} \IncMargin{1.0em} \begin{algorithm}[H] \DontPrintSemicolon \caption{SAVI on DAG Latent}\label{alg:solve-dag} \textbf{procedure} solve-dag($\bm{x}$)\; $\quad$sort $\bm{y}_1,...,\bm{y}_N$ in topological order\; $\quad$\textbf{for} $\bm{y}_j$ with parent $\mathcal{P}(\bm{y}_j)=\varnothing$\; $\quad\quad$ add $\bm{y}_j$ to fake node $\bm{y}_0$'s children $\mathcal{C}(\bm{y}_0)$\; $\quad$grad-dag($\bm{x},\bm{y}_0^0$)\; $\quad$\textbf{return} $\bm{y}_1^K,...,\bm{y}_N^K$\; \BlankLine \textbf{procedure} grad-dag($\bm{x},\bm{y}_0^{k_0},...,\bm{y}_i^{k_i}$)\; $\quad$\textbf{for} $\bm{y}_{j}\in\mathcal{C}(\bm{y}_i)$ in topological order \textbf{do}\; $\quad\quad$ $\bm{y}_{j}^0\leftarrow f(\bm{x},\bm{y}_0^{k_0},...,\bm{y}_{<j}^{k_{<j}})$ from FAVI\; $\quad\quad$ \textbf{for} $k_j=0,...,K-1$ \textbf{do}\; $\quad\quad\quad \frac{d\mathcal{L}(\bm{y}_0^{k_0},...,\bm{y}_j^{k_j},\bm{y}_{>j}^K)}{d\bm{y}_{j}^{k_j}} \leftarrow \textrm{grad-dag}(\bm{x},\bm{y}_0^{k_0},...,\bm{y}_j^{k_j})$\; $\quad\quad\quad\bm{y}_{j}^{k_j+1}\leftarrow\bm{y}_{j}^{k_j}+ \alpha\frac{d\mathcal{L}(\bm{y}_0^{k_0},...,\bm{y}_j^{k_j},\bm{y}_{>j}^K)}{d\bm{y}_{j}^{k_j}}$\; $\quad\overleftarrow{\bm{y}_i}\leftarrow \frac{\partial\mathcal{L}(\bm{y}_0^{k_0},...,\bm{y}_i^{k_i},\bm{y}_{>i}^K)}{\partial \bm{y}_i^{k_i}}$\; $\quad$\textbf{for} $\bm{y}_{j}\in\mathcal{C}(\bm{y}_i)$ \textbf{do}\; $\quad\quad\overleftarrow{\bm{y}_j}\leftarrow\bm{0},\overleftarrow{\bm{y}_{j}^K}\leftarrow\frac{d \mathcal{L}(\bm{y}_0^{k_0},...,\bm{y}_i^{k_i},\bm{y}_{>i}^K)}{d\bm{y}_{j}^K}$\; $\quad\quad$\textbf{for} $k_j=K-1,...,0$ \textbf{do}\; $\quad\quad\quad\overleftarrow{\bm{y}_j}\leftarrow\overleftarrow{\bm{y}_j}+\alpha \frac{\partial^2 \mathcal{L}(\bm{y}_0^{k_0},...,\bm{y}_j^{k_j},\bm{y}_{>j}^K)}{\partial\bm{y}_i^{k_i}\partial\bm{y}_{j}^{k_j}}\overleftarrow{\bm{y}_{j}^{k_j+1}}$\; $\quad\quad\quad\overleftarrow{\bm{y}_j^{k_j}}\leftarrow\overleftarrow{\bm{y}_j^{k_j+1}}+\alpha \frac{\partial^2 \mathcal{L}(\bm{y}_0^{k_0},...,\bm{y}_j^{k_j},\bm{y}_{>j}^K)}{\partial\bm{y}_j^{k_j}\partial\bm{y}_{j}^{k_j}}\overleftarrow{\bm{y}_{j}^{k_j+1}}$\; $\quad\quad\overleftarrow{\bm{y}_i}\leftarrow \overleftarrow{\bm{y}_i} + \overleftarrow{\bm{y}_j}+\frac{\partial \bm{y}_j^0}{\partial \bm{y_i^{k_i}}}\overleftarrow{\bm{y}_j^{0}}$\; $\quad$\textbf{return}$\frac{d\mathcal{L}(\bm{y}_0^{k_0},...,\bm{y}_i^{k_i},\bm{y}_{>i}^K)}{d \bm{y}_i^{k_i}}=\overleftarrow{\bm{y}_i}$\; \end{algorithm} \end{minipage} The intuition is, we do not want to find a $\bm{w}$ that maximizes $\mathcal{L}(\bm{w},\bm{y})$ given a fixed $\bm{y}$ (or we have the gradient issue described in Sec.~\ref{sec:baosub}). Instead, we want to find a $\bm{w}$, whose $\max_{\bm{y}}\mathcal{L}(\bm{w},\bm{y})$ is maximum. This translates to the optimization problem as Eq.~\ref{eq:opt2}. In fact, Eq.~\ref{eq:opt2} is a variant of setup in back-propagating through gradient ascent \citep{samuel2009learning,domke2012generic}. The difference is, our $\bm{w}$ also contributes directly to optimization target $\mathcal{L}(\bm{w},\bm{y})$. From this perspective, Eq.~\ref{eq:opt2} is more closely connected to \citet{kim2018semi}, if we treat $\bm{w}$ as the model parameter and $\bm{y}$ as latent. \begin{align} \bm{w}\leftarrow \arg \max_{\bm{w}} \mathcal{L}(\bm{w},\bm{y}^*(\bm{w}))\textrm{, where } \bm{y}^*(\bm{w})\leftarrow \arg\max_{\bm{y}} \mathcal{L}(\bm{w},\bm{y}) \label{eq:opt2} \end{align} And as SAVI on 1-level latent \citep{kim2018semi,marino2018iterative}, we need to solve Eq.~\ref{eq:opt2} using gradient ascent. Specifically, denote $\alpha$ as step size (learning rate), $K$ as the total gradient ascent steps, $\bm{w}^k$ as the $\bm{w}$ after $k$ step update, $\bm{y}^{k'}$ as the $\bm{y}$ after $k'$ step update, and $f(.)$ as FAVI procedure generating initial posterior parameters $\bm{w}^0,\bm{y}^0$, the optimization problem as Eq.~\ref{eq:opt2} translates into the update rule as Eq.~\ref{eq:opt2grad}. Eq.~\ref{eq:opt2grad} is the guidance for designing optimization algorithm, and it also explains why the gradient of BAO \citep{bao2022} is evaluated at wrong value (See Sec.~\ref{sec:ipde}). \begin{align} \bm{w}^{k+1}\leftarrow \bm{w}^{k}+\alpha\frac{d\mathcal{L}(\bm{w}^k,\bm{y}^{K})}{d\bm{w}^k}, \bm{y}^{k'+1}\leftarrow \bm{y}^{k'}+\alpha\frac{d\mathcal{L}(\bm{w}^k,\bm{y}^{k'})}{d\bm{y}^{k'}}\textrm{, where } \bm{y}^0 = f(\bm{x},\bm{w}^k)\label{eq:opt2grad} \end{align} To solve Eq.~\ref{eq:opt2grad}, we note that although $d\mathcal{L}(\bm{w}^k,\bm{y}^{k'})/d\bm{y}^{k'}$ is directly computed, $d\mathcal{L}(\bm{w}^k,\bm{y}^{K})/d\bm{w}^{k}$ is not straightforward. Resorting to previous works \citep{samuel2009learning,domke2012generic} in implicit differentiation and extending the results in \citet{kim2018semi} from model parameters to variational posterior parameters, we implement Eq.~\ref{eq:opt2grad} as Alg.~\ref{alg:solve-2}. Specifically, we first initialize $\bm{w}^0$ from FAVI. Then we conduct gradient ascent on $\bm{w}$ with gradient $d\mathcal{L}(\bm{w}^k,\bm{y}^K)/d\bm{w}^{k}$ computed from the procedure grad-2-level($\bm{x},\bm{w}^k$). And inside grad-2-level($\bm{x},\bm{w}^k$), $\bm{y}$ is also updated by gradient ascent, the above procedure corresponds to Eq.~\ref{eq:opt2grad}. The key of Alg.~\ref{alg:solve-2} is the evaluation of gradient $d\mathcal{L}(\bm{w}^k,\bm{y}^K)/d\bm{w}^{k}$. Formally, we have: \begin{theorem} \label{th:2l} After \textup{grad-2-level($\bm{x},\bm{w}^k$)} of Alg.~\ref{alg:solve-2} executes, we have the return value $d \mathcal{L}(\bm{w}^k,\bm{y}^K)/d\bm{w}^k=\overleftarrow{\bm{w}}$. (See proof in Appendix.~\ref{app:pf}.) \end{theorem} \subsection{SAVI on DAG-defined Non-factorized Latent} \label{sec:savidag} In this section, we extend the result from previous section to SAVI on general non-factorized latent with dependency described by any DAG. This DAG is the computational graph during network inference, and it is also the directed graphical model (DGM) \citep{koller2009probabilistic} defining the factorization of latent variables during inference. This is the general case covering all dependency that can be described by DGM. This extension is necessary to perform SAVI on latent with complicated dependency (e.g. bit allocation of NVC). Similar to the 2-level latent setup, we consider performing SAVI on $N$ variational posterior parameter $\bm{y}_1,...,\bm{y}_N$ with their dependency defined by a computational graph $\mathcal{G}$, i.e., their corresponding latent variable $\bmt{y}_1,...,\bmt{y}_N$'s posterior distribution factorizes as $\mathcal{G}$. Specifically, we denote $\bm{y}_j\in\mathcal{C}(\bm{y}_i),\bm{y}_i\in\mathcal{P}(\bm{y}_j)$ if an edge exists from $\bm{y}_i$ to $\bm{y}_j$. This indicates that $\bmt{y}_j$ conditions on $\bmt{y}_i$. Without loss of generality, we assume $\bm{y}_1,...,\bm{y}_N$ is sorted in topological order. This means that if $\bm{y}_j\in\mathcal{C}(\bm{y}_i),\bm{y}_i\in\mathcal{P}(\bm{y}_j)$, then $i<j$. Each latent is optimized by $K$-step gradient ascent, and $\bm{y}_i^{k_i}$ denotes the latent $\bm{y}_i$ after $k_i$ steps of update. Then, similar to 2-level latent, we have the update rule as Eq.~\ref{eq:optdaggrad}: \begin{align} \bm{y}^{k_i+1}_i\leftarrow \bm{y}^{k_i}_i+\alpha\frac{d\mathcal{L}(\bm{y}_1^{k_1},...,\bm{y}_i^{k_i},\bm{y}_{>i}^K)}{d\bm{y}^{k_i}}\textrm{, where }\bm{y}_{>i}^0 = f(\bm{x},\bm{y}_1^{k_1},...,\bm{y}_i^{k_i}) \label{eq:optdaggrad} \end{align} , which can be translated into Alg.~\ref{alg:solve-dag}. Specifically, we first sort the latent in topological order. Then, we add a fake latent $\bm{y}_0$ to the front of all $\bm{y}$s. Its children are all the $\bm{y}s$ with 0 in-degree. Then, we can solve the SAVI on $\bm{y}_1,...,\bm{y}_N$ using gradient ascent by executing the procedure grad-dag($\bm{x},\bm{y}_0^{k_0},...,\bm{y}_i^{k_i}$) in Alg.~\ref{alg:solve-dag} recursively. Inside procedure grad-dag($\bm{x},\bm{y}_0^{k_0},...,\bm{y}_i^{k_i}$), the gradient to update $\bm{y}_i$ relies on the convergence of its children $\bm{y}_{j}\in\mathcal{C}(\bm{y}_i)$, which is implemented by the recursive depth-first search (DFS) in line 11. And upon the completion of procedure grad-dag($\bm{x},\bm{y}_0^0$), all the latent converges to $\bm{y}_1^K,...,\bm{y}_N^K$. Similar to the 2-level latent case, the key of Alg.~\ref{alg:solve-dag} is the evaluation of gradient $d\mathcal{L}(\bm{y}_0^{k_0},...,\bm{y}_i^{k_i},\bm{y}_{>i}^K)/d \bm{y}_i^{k_i}$. Formally, we have: \begin{theorem} \label{th:dag} After the procedure \textup{grad-dag($\bm{x},\bm{y}_0^{k_0},...,\bm{y}_i^{k_i}$)} in Alg.~\ref{alg:solve-dag} executes, we have the return value $d\mathcal{L}(\bm{y}_0^{k_0},...,\bm{y}_i^{k_i},\bm{y}_{>i}^K)/d \bm{y}_i^{k_i}=\overleftarrow{\bm{y}_i}$. (See proof in Appendix.~\ref{app:pf}.) \end{theorem} To better understand how Alg.~\ref{alg:solve-dag} works, we provide a detailed example in Fig.~\ref{fig:eg} of Appendix.~\ref{app:eg}. \subsection{Correcting the Sub-optimal Bit Allocation using SAVI on DAG} \label{sec:savinvc} With the result in previous section, correcting BAO \citep{bao2022} seems to be trivial. We only need to sort the latent in topological order as $\bm{w}_1,\bm{y}_1,...,\bm{w}_T,\bm{y}_T$, and run Alg.~\ref{alg:solve-dag} to obtain the optimized latent parameters $\bm{w}_1^K,\bm{y}_1^K,...,\bm{w}_T^K,\bm{y}_T^K$. And the gradient $d\mathcal{L}(\bm{y}_0^{k_0},...,\bm{y}_i^{k_i},\bm{y}_{>i}^K)/d \bm{y}_i^{k_i}$ computed in Alg.~\ref{alg:solve-dag} resolves the issue of BAO described in Sec.~\ref{sec:itde} and Sec.~\ref{sec:ipde}. However, an evident problem is the temporal complexity. Given the latent number $N$ and gradient ascent step number $K$, Alg.~\ref{alg:solve-dag} has temporal complexity of $\Theta(K^N)$. NVC with GoP size $10$ has approximately $N=20$ latent, and the SAVI on NVC \citep{bao2022} takes around $K=2000$ step to converge. For bit allocation, the complexity of Alg.~\ref{alg:solve-dag} is $\approx 2000^{20}$, which is intractable. On the other hand, BAO's complexity is reasonable ($\Theta(KN)\approx4\times10^{4}$). Thus, in next section, we provide a feasible approximation to such intractable corrected bit allocation. \subsection{Feasible Approximation to the Corrected Bit Allocation} \label{sec:approx} In order to solve problem with practical size such as bit allocation on NVC, we provide an approximation to the SAVI \citep{kim2018semi,marino2018iterative} on DAG described in Sec.~\ref{sec:savidag}. The general idea is that, when being applied to bit allocation of NVC, the accurate SAVI on DAG (Alg.~\ref{alg:solve-dag}) satisfies both requirement on gradient signal described in Sec.~\ref{sec:itde} and Sec.~\ref{sec:ipde}. We can not make it tractable without breaking them. Thus, we break one of them and achieve a reasonable complexity, while maintain a superior performance compared with BAO \citep{bao2022}. We consider the approximation in Eq.~\ref{eq:approx} which breaks the requirement for gradient evaluation in Sec.~\ref{sec:ipde}. Based on Eq.~\ref{eq:approx} and the requirement in Sec.~\ref{sec:itde}, we design an approximation of accurate SAVI as Alg.~\ref{alg:solve-adag}. When being applied to bit allocation in NVC, it satisfies the gradient requirement in Sec.~\ref{sec:itde} while maintaining a temporal complexity of $\Theta(KN)$ as BAO. \begin{align} \frac{d\mathcal{L}(\bm{y}_0^{k_0},...,\bm{y}_i^{k_i},\bm{y}_{>i}^K)}{d \bm{y}_i^{k_i}}\approx\frac{d\mathcal{L}(\bm{y}_0^{k_0},...,\bm{y}_i^{k_i},\bm{y}_{>i}^0)}{d \bm{y}_i^{k_i}} \label{eq:approx} \end{align} Specifically, with the approximation in Eq.~\ref{eq:approx}, the recurrent gradient computation in Alg.~\ref{alg:solve-dag} becomes unnecessary as the right hand side of Eq.~\ref{eq:approx} does not require $\bm{y}_{>i}^K$. However, to maintain the dependency of latent described in Sec.~\ref{sec:itde}, as Alg.~\ref{alg:solve-dag}, we still need to ensure that the children node $\bm{y}_{j}\in \mathcal{C}(\bm{y}_i)$ are re-initialized by FAVI every-time when $\bm{y}_i$ is updated. Therefore, a reasonable approach is to traverse the graph in topological order. We keep the children node $\bm{y}_j$ untouched until all its parent node $\bm{y}_i\in\mathcal{P}(\bm{y}_j)$'s gradient ascent is completed and $\bm{y}_i^K$ is known. And the resulting approximate SAVI algorithm is as Alg.~\ref{alg:solve-adag}. When applied to bit allocation, it satisfies the gradient requirement in Sec.~\ref{sec:itde}, and as BAO, its temporal complexity is $\Theta(KN)$. \begin{minipage}[t]{.44\textwidth} % \vspace{0pt} \IncMargin{1.0em} \begin{algorithm}[H] \DontPrintSemicolon \caption{BAO on DAG Latent}\label{alg:solve-bao} \textbf{procedure} solve-bao($\bm{x}$)\; $\quad\bm{y}_1^0,...,\bm{y}_N^0\leftarrow f(\bm{x})$ from FAVI\; $\quad$\textbf{for} $k =0,...,K-1$ \textbf{do}\; $\quad\quad$ \textbf{for} $i=1,...,N$ \textbf{do}\; $\quad\quad\quad \bm{y}_i^{k+1}\leftarrow\bm{y}_i^k+\alpha\frac{\partial\mathcal{L}(\bm{y}_1^k,...,\bm{y}_N^k)}{\partial\bm{y}_i^k}$\; $\quad$\textbf{return} $y_1^K,...,y_N^K$\; \end{algorithm} \end{minipage} \begin{minipage}[t]{.55\textwidth} % \vspace{0pt} \IncMargin{1.0em} \begin{algorithm}[H] \DontPrintSemicolon \caption{Approximate SAVI on DAG latent}\label{alg:solve-adag} \textbf{procedure} solve-approx-dag($\bm{x}$)\; $\quad$sort $\bm{y}_1,...,\bm{y}_N$ in topological order\; $\quad$\textbf{for} $i=1,...,N$ \textbf{do}\; $\quad\quad \bm{y}_i^0,...,\bm{y}_N^0\leftarrow f(\bm{x},\bm{y}_{<i}^K)$ from FAVI\; $\quad\quad$ \textbf{for} $k=0,...,K-1$ \textbf{do}\; $\quad\quad\quad\frac{d\mathcal{L}(\bm{y}_{<i}^K,\bm{y}_i^k,\bm{y}_{>i}^K)}{d\bm{y}_i^k}\approx\frac{d\mathcal{L}(\bm{y}_{<i}^K,\bm{y}_i^k,\bm{y}_{>i}^0)}{d\bm{y}_i^k}$\; $\quad\quad\quad \bm{y}_i^{k+1}\leftarrow\bm{y}_i^k+\alpha\frac{d\mathcal{L}(\bm{y}_{<i}^K,\bm{y}_i^k,\bm{y}_{>i}^K)}{d\bm{y}_i^k}$\; $\quad$\textbf{return} $y_1^K,...,y_N^K$\; \end{algorithm} \end{minipage} % To better understand BAO \citep{bao2022} in SAVI context, we rewrite it by general SAVI notation instead of NVC notation in Alg.~\ref{alg:solve-bao}. We highlight the difference between BAO (Alg.~\ref{alg:solve-bao}) \citep{bao2022}, the accurate SAVI on DAG latent (Alg.~\ref{alg:solve-dag}) and the approximate SAVI on DAG latent (Alg.~\ref{alg:solve-adag}) from several aspects: \begin{itemize} \item \textbf{Graph Traversal Order}: BAO performs gradient ascent on $\bm{y}_{1:T}$ all together. The accurate SAVI only updates $\bm{y}_i$ when $\bm{y}_{>i}$'s update is complete and $\bm{y}_{>i}^K$ is known. The approximate SAVI only updates $\bm{y}_i$ when $\bm{y}_{<i}$'s update is complete and $\bm{y}_{<i}^K$ is known. \item \textbf{Gradient Correctness}: When being applied to bit allocation in NVC, BAO violates the gradient rule in Sec.~\ref{sec:itde} and Sec.~\ref{sec:ipde}, accurate SAVI satisfies both rules, approximate SAVI satisfies Sec.~\ref{sec:itde} and violates Sec.~\ref{sec:ipde}. \item \textbf{Temporal Complexity}: With the latent number $N$ and steps of gradient ascent $K$, the complexity of BAO is $\Theta(KN)$, the complexity of accurate SAVI is $\Theta(K^N)$ and the complexity of approximate SAVI is $\Theta(KN)$. \end{itemize} Then we can simply apply Alg.~\ref{alg:solve-adag} to bit allocation in NVC to obtain a feasible approximation of the corrected optimal bit allocation. And in Sec.~\ref{sec:rd}, we empirically show that our approximation improves the R-D performance over BAO \citep{bao2022} with even smaller number of updates. \section{Related Work: Bit Allocation \& SAVI for Neural Compression} \citet{li2022rate} are the pioneer of bit allocation for NVC and their work is elaborated in Sec.~\ref{sec:bgba}. Other recent works that consider bit allocation for NVC only adopt simple heuristic such as inserting $1$ high quality frame per $4$ frames \citep{hu2022coarse,cetin2022flexible}. On the other hand, OEU \citep{lu2020content} is also recognised as frame-level bit allocation while its performance is inferior than BAO \citep{bao2022}. BAO is the most recent work with best R-D performance. It is elaborated in Sec.~\ref{sec:bgba} and Sec.~\ref{sec:baosub}, and corrected in the previous section. Semi-Amortized Variational Inference (SAVI) is proposed by \citet{kim2018semi,marino2018iterative}. The idea is that works following \citet{kingma2013auto} use fully amortized inference parameter $\phi$ for all data, which leads to the amortization gap \citep{cremer2018inference}. SAVI reduces this gap by optimizing the variational posterior parameter after initializing it with inference network. It adopts back-propagating through gradient ascent \citep{domke2012generic} to evaluate the gradient of model parameters. We adopt a similar method to extend SAVI to non-factorized latent. When applying SAVI to practical neural codec, researchers abandon the nested model parameter update for efficiency. Prior works \citep{djelouah2019content,yang2020improving,zhao2021universal,ce2022} adopt SAVI to boost R-D performance and achieve variable bitrate in image compression. And BAO \citep{bao2022} is the first to consider SAVI for bit allocation. \section{Experiments} \subsection{Experimental Settings} We implement our approach in PyTorch 1.9 with CUDA 11.2, and run the experiments on NVIDIA(R) A100 GPU. Most of the other settings are intentionally kept the same as BAO \citep{bao2022}. Specifically, we adopt HEVC Common Testing Condition (CTC) \citep{bossen2013common} and UVG dataset \citep{mercat2020uvg}. And we measure the R-D performance in Bjontegaard-Bitrate (BD-BR) and BD-PSNR \citep{bjontegaard2001calculation}. For baseline NVC \citep{lu2019dvc,li2021deep}, we adopt the official pre-trained models. And we select target $\lambda_0=\{256,512,1024,2048\}$. For gradient ascent, we adopt Adam \citep{kingma2014adam} optimizer with $lr=1\times10^{-3}$. We set the gradient ascent step $K=2000$ for the first frame and $K=400$ for other frames. More details are presented in Appendix.~\ref{app:impl}. \subsection{Quantitative Results} \label{sec:rd} As shown in Tab.~\ref{tab:bdbr}, our method consistently improves the R-D performance in terms of BD-BR over BAO \citep{bao2022} on both baseline methods and all datasets. Moreover, this improvement is especially significant (more than 10\% in BD-BR) when the baseline is DCVC \citep{li2021deep}. And both BAO and our proposed correction significantly outperform other approaches. It is also noteworthy that with our bit allocation, DVC (the SOTA method in 2019) already outperforms DCVC (the SOTA method in 2021) by large margin (See the red solid line and black dash line in Fig.~\ref{fig:bdbrd}). \begin{minipage}{\textwidth} \vspace{7pt} \begin{minipage}[t]{0.7\textwidth} \vspace{0pt} \centering \begin{small} \begin{tabular}{@{}llllll@{}} \toprule & \multicolumn{5}{c}{BD-BR (\%) $\downarrow$} \\ \cmidrule(rr){2-6} Method & Class B & Class C & Class D & Class E & UVG \\ \midrule \multicolumn{3}{@{}l@{}}{\textit{DVC \citep{lu2019dvc} as Baseline} } & & & \\ \citet{li2016lambda}$^1$ & 20.21 & 17.13 & 13.71 & 10.32 & 16.69 \\ \citet{li2022rate}$^1$ & -6.80 & -2.96 & 0.48 & -6.85 & -4.12 \\ OEU \citep{lu2020content}$^2$ & -13.57 & -11.29 & -18.97 & -12.43 & -13.78 \\ BAO \citep{bao2022}$^2$ & -28.55 & -26.82 & -25.37 & -32.54 & -27.68 \\ Proposed & -32.10 & -31.71 & -35.86 & -32.93 & -30.92 \\ \midrule \multicolumn{3}{@{}l@{}}{\textit{DCVC \citep{li2021deep} as Baseline} } & & & \\ OEU \citep{lu2020content}$^2$ & -10.75 & -14.34 & -16.30 & -7.15 & -16.07 \\ BAO \citep{bao2022}$^2$ & -20.59 & -19.69 & -20.60 & -23.33 & -25.22 \\ Proposed & -32.89 & -33.10 & -32.01 & -36.88 & -39.66 \\\bottomrule \end{tabular} \end{small} \captionof{table}{The BD-BR of our approach compared with others. $^1$ comes from \citet{li2022rate}. $^2$ comes from \citet{bao2022}.} \label{tab:bdbr} \end{minipage} \hfill \begin{minipage}[t]{0.295\textwidth} \vspace{0pt} \centering \includegraphics[width=\textwidth]{fig_rd_d_small.PNG} \captionof{figure}{The R-D curve on HEVC Class D.} \label{fig:bdbrd} \end{minipage} \end{minipage} Other than R-D performance, the bitrate error of our approach is also significantly smaller than BAO \citep{bao2022} (See Tab.~\ref{tab:bderr}). The bitrate error is measured as the relative bitrate difference before and after bit allocation. The smaller it is, the easier it is to achieve the desired bitrate accurately. For complexity, our approach only performs $920$ steps of gradient ascent per-frame, while BAO requires $2000$ steps. See more quantitative results (BD-PSNR \& R-D curves) in Appendix.~\ref{app:quant}. \subsection{Ablation Study, Analysis \& Qualitative Results } Tab.~\ref{tab:abl} shows that for BAO \citep{bao2022}, jointly optimizing $\bm{w}_{1:T},\bm{y}_{1:T}$ performs worse than optimizing $\bm{y}_{1:T}$ or $\bm{w}_{1:T}$ alone. This counter-intuitive phenomena comes from its incorrect estimation of gradient signal. For the proposed approach that corrects this, jointly optimizing $\bm{w}_{1:T},\bm{y}_{1:T}$ performs better than optimizing $\bm{y}_{1:T}$ or $\bm{w}_{1:T}$ alone, which is aligned with our intuition. \begin{minipage}{\textwidth} \vspace{7pt} \begin{minipage}[t]{0.68\textwidth} \vspace{0pt} \centering \begin{small} \begin{tabular}{@{}llllll@{}} \toprule & \multicolumn{5}{c}{Bitrate-Error (\%) $\downarrow$} \\ \cmidrule(rr){2-6} Method & Class B & Class C & Class D & Class E & UVG \\ \midrule \multicolumn{3}{@{}l@{}}{\textit{DVC \citep{lu2019dvc} as Baseline} } & & & \\ BAO \citep{bao2022}$^2$ & 8.41 & 12.86 & 21.39 & 5.94 & 3.73 \\ Proposed & 3.16 & 4.27 & 1.81 & 6.14 & 1.73 \\ \midrule \multicolumn{3}{@{}l@{}}{\textit{DCVC \citep{li2021deep} as Baseline} } & & & \\ BAO \citep{bao2022}$^2$ & 25.67 & 23.90 & 23.74 & 24.88 & 21.86 \\ Proposed & 4.27 & 7.29 & 5.73 & 8.03 & 3.06 \\\bottomrule \end{tabular} \end{small} \captionof{table}{The bitrate error of our approach compared with BAO.} \label{tab:bderr} \end{minipage} \hfill \begin{minipage}[t]{0.30\textwidth} \vspace{0pt} \centering \begin{small} \begin{tabular}{@{}ll@{}} \toprule Method & BD-BR (\%) $\downarrow$\\ \midrule BAO ($\bm{y}$) & -25.37 \\ BAO ($\bm{w}$) & -22.24 \\ BAO ($\bm{y},\bm{w}$) & -14.76 \\ Proposed ($\bm{y}$) & -32.60 \\ Proposed ($\bm{w}$) & -31.56 \\ Proposed ($\bm{y},\bm{w}$) & -35.86 \\ \bottomrule \end{tabular} \end{small} \captionof{table}{Ablation study with HEVC Class D and DVC \citep{lu2019dvc}.} \label{tab:abl} \end{minipage} \end{minipage} To better understand why our method works, we present the R-D cost, distortion and rate versus frame/latent index for different methods in Fig.~\ref{fig:anamain}: \textit{top-left} shows that the R-D cost of our approach consistently decreases according to SAVI stage. Moreover, it outperforms BAO after $4^{th}$ frame; \textit{top-right} shows that for each frame the R-D cost of our method is lower than BAO; \textit{bottom-left} shows that the distortion part of R-D cost of our approach is approximately the same as BAO. While \textit{bottom-right} shows that the advantage of our approach over BAO lies in the bitrate. More specifically, BAO increases the bitrate of $\bm{y}_i$s after SAVI, while our correction decreases it. See more analysis in Appendix.~\ref{app:ana} and qualitative results in Appendix.~\ref{app:qual}. \begin{figure}[h] \centering \includegraphics[width=\linewidth]{fig_ana_main.PNG} \caption{\textit{top-left}. R-D cost vs. SAVI stage. \textit{top-right}. R-D cost vs. frame index. \textit{bottom-left}. PSNR vs. frame index. \textit{bottom-right}. bpp vs. latent index. See enlarged-version in Appendix.~\ref{app:ana}.} \label{fig:anamain} \end{figure} \section{Discussion \& Conclusion} Despite our correction is already more efficient than original BAO \citep{bao2022}, its encoding speed remains far from real-time. Thus, it is limited to scenarios where R-D performance matters much more than encoding time (e.g. video on demand). See more discussion in Appendix.~\ref{app:disc}. To conclude, we show that a previous bit allocation method for NVC is sub-optimal as it abuses SAVI on non-factorized latent. Then, we propose the correct SAVI on general non-factorized latent by back-propagating through gradient ascent, and we further propose a feasible approximation to make it practical for bit allocation. Experimental results show that our correction significantly improves the R-D performance. \subsubsection*{Ethics Statement} Improving the R-D performance of NVC has positive social value, in terms of reducing carbon emission by saving the resources required to transfer and store videos. Moreover, unlike traditional codecs such as H.266 \citep{bross2021developments}, neural video codec does not require dedicated hardware. Instead, it can be deployed with general neural accelerators. Improving the R-D performance of NVC prompts the practical deployment of video codecs that are independent of dedicated hardware, and lowers the hardware-barrier of playing multi-media contents. \subsubsection*{Reproducibility Statement} For theoretical results, both of the two theorems are followed by proof in Appendix.~\ref{app:pf}. For a relatively complicated novel algorithm (Alg.~\ref{alg:solve-dag}), we provide an illustration of the step by step execution procedure in Appendix.~\ref{app:eg}. For experiment, both of the two datasets are publicly accessible. In Appendix.~\ref{app:impl}, we provide more implementation details including all the hyper-parameters. Moreover, we provide our source code for reproducing the empirical results in supplementary material.

517ee6531e27a049052c60af266fe02d558d60e0

\section{Introduction}\label{S:Introduction} \noindent A 2D vortex patch is a solution to the 2D Euler equations for which the vorticity is a constant multiplied by the characteristic function of a domain. We investigate the behavior of vortex patches in an infinite strip periodic in one direction, topologically $S^1 \times {\ensuremath{\BB{R}}}$, and the corresponding behavior of the vortex patch or layer in the full plane. Our main results are the extension of the $C^{1,\ensuremath{\varepsilon}}$ global regularity theory for the boundary of the vortex patch to this case, developing and using the appropriate contour dynamics equation for this purpose. Here, and throughout, we fix $\ensuremath{\varepsilon} \in (0, 1)$. \subsection{The Euler equations} We can write the 2D incompressible Euler equations (without forcing) on a domain $U$ in vorticity form as \begin{align}\label{e:EGen} \begin{cases} \ensuremath{\partial}_t \omega + \uu \cdot \ensuremath{\nabla} \omega = 0 &\text{in } {\ensuremath{\BB{R}}} \times U, \\ \uu = K[\omega] &\text{in } {\ensuremath{\BB{R}}} \times U, \\ \omega(0) = \omega^0 &\text{in } U. \end{cases} \end{align} Here, $\omega$ is the vorticity---the scalar curl of the velocity field $\uu$. The vorticity is transported by the velocity field as in \cref{e:EGen}$_1$, and the velocity field is recovered from the vorticity field by the constitutive law in \cref{e:EGen}$_2$ so as to be divergence-free and to satisfy any boundary conditions, decay at infinity, or periodicity that might be demanded based, in part, upon the nature of the domain $U$. Classically, if $U = {\ensuremath{\BB{R}}}^2$ and the solution has sufficient decay, one uses the Biot-Savart law as the constitutive law: \begin{align}\label{e:BSR2} K[\omega] := K * \omega, \quad K({\bm{\mathrm{x}}}) := \ensuremath{\nabla}^\perp \brac{\frac{1}{2 \pi} \log \abs{{\bm{\mathrm{x}}}}} = \frac{1}{2 \pi} \frac{{\bm{\mathrm{x}}}^\perp}{\abs{{\bm{\mathrm{x}}}}^2}. \end{align} Here, $K$ is the Biot-Savart kernel, which we note lies in $L^1_{loc}({\ensuremath{\BB{R}}}^2)$, though $K \notin L^p({\ensuremath{\BB{R}}}^2)$ for any $p \in [1, \ensuremath{\infty}]$. To handle solutions having insufficient spatial decay of the vorticity, we must either find an appropriate substitute for the Biot-Savart law or avoid it entirely by using a velocity, pressure formulation. \subsection{The plane and the cylinder} In this paper, we will consider two domains: $U = {\ensuremath{\BB{R}}}^2$ and $U = \Pi$, the infinite flat periodic strip, $S^1 \times {\ensuremath{\BB{R}}} \cong {\ensuremath{\BB{R}}}^2 / \ensuremath{\BB{Z}} \cong \C / \ensuremath{\BB{Z}}$, which we will most often treat in the form \begin{align}\label{e:Pi} \Pi := \brac{-\tfrac{1}{2}, \tfrac{1}{2}} \times {\ensuremath{\BB{R}}} \text{ with } \set{-\tfrac{1}{2}} \times {\ensuremath{\BB{R}}} \text{ identified with } \set{\tfrac{1}{2}} \times {\ensuremath{\BB{R}}}. \end{align} We will also find use for the same set as a subset of ${\ensuremath{\BB{R}}}^2$ or $\C$ without identifying its sides: \begin{align}\label{e:PiR} {\Pi_p} := \pr{-\tfrac{1}{2}, \tfrac{1}{2}} \times {\ensuremath{\BB{R}}} \subseteq {\ensuremath{\BB{R}}}^2. \end{align} Suppose we have an initial vorticity $\omega^0 = \CharFunc_\Omega$ for $\Omega$ a bounded domain in $\Pi$. We can periodize it to obtain an initial vorticity in ${\ensuremath{\BB{R}}}^2$ that is periodic in $x_1$. What results may consist of an infinite number of disconnected domains repeated periodically, one connected, $x_1$-periodic domain, or a combination of each. \Cref{f:VortexPatch} displays an example of a simply connected bounded domain in $\Pi$ yielding an infinite number of copies of the domain in ${\ensuremath{\BB{R}}}^2$. \Cref{f:VortexLayer} displays two examples of a non-simply connected domain in $\Pi$ producing one domain in ${\ensuremath{\BB{R}}}^2$ periodically repeating in $x_1$, a so-called \textit{vortex layer}. \begin{figure}[ht] \IncludeFigure{0.14}{FigPatch} \vspace{-1.2em} \caption{\textit{Example of a periodic vortex patch in ${\ensuremath{\BB{R}}}^2$ and in $\Pi$}}\label{f:VortexPatch} \end{figure} \begin{figure}[ht] \IncludeFigure{0.14}{FigLayer} \vspace{-1.5em} \caption{\textit{Two examples of a periodic vortex layer in ${\ensuremath{\BB{R}}}^2$ and in $\Pi$}}\label{f:VortexLayer} \end{figure} On the other hand, we could instead formulate the problem by starting with an initial vortex patch in ${\ensuremath{\BB{R}}}^2$ and periodize it in $x_1$. If we can translate the evolution of the patch in ${\ensuremath{\BB{R}}}^2$ to the evolution in $\Pi$ and back, we can use an understanding of patch behavior in $\Pi$ to gain an understanding of the periodic behavior in ${\ensuremath{\BB{R}}}^2$. The translation back and forth between $\Pi$ and ${\ensuremath{\BB{R}}}^2$ is best understood in the more general setting of weak solutions to the 2D Euler equations for bounded vorticity, which includes vortex patch data as a special case. \subsection{Three types of solutions} Toward this end, we consider three types of solution to the 2D Euler equations. We summarize the three types of solution briefly now, giving more complete descriptions in later sections. \begin{itemize}[leftmargin=5em,itemsep=0.5em] \item[\textbf{Type 1}] Assume that $\uu^0 \in L^\ensuremath{\infty}({\ensuremath{\BB{R}}}^2)$ is divergence-free with $\omega^0 := \curl \uu^0 \in L^\ensuremath{\infty}({\ensuremath{\BB{R}}}^2)$ as well. Obtain a bounded vorticity, bounded velocity solution to the the 2D Euler equations on all of ${\ensuremath{\BB{R}}}^2$ having initial velocity $\uu^0$ as done by Serfati in \cite{Serfati1995Bounded}. \item[\textbf{Type 2}] Assume $\uu^0 \in L^\ensuremath{\infty}(\Pi)$ is divergence-free with $\omega^0 := \curl \uu^0 \in L^\ensuremath{\infty}(\Pi)$ as well. Solve the 2D Euler equations in $\Pi$, as done in \cite{AfendikovMielke2005,GallaySlijepcevic2014,GallaySlijepcevic2015}. \item[\textbf{Type 3}] Let $\omega^0 \in L^\ensuremath{\infty}({\ensuremath{\BB{R}}}^2)$ be compactly supported. Solve the 2D Euler equations in vorticity form in all of ${\ensuremath{\BB{R}}}^2$ with initial vorticity $\omega^0$, but recovering the velocity by applying the Biot-Savart law symmetrically to pairs of the periodically extended copies of $\omega$. This leads to a replacement Biot-Savart kernel, $K_\ensuremath{\infty}$. \end{itemize} Type 1 and Type 2 solutions are for (potentially) non-decaying velocity and vorticity, but for Type 3 we restrict our attention to vertically decaying solutions, since our primary application is to vortex patch data. Moreover, the convolution $K_\ensuremath{\infty} * \omega$ cannot be easily defined without some decay assumption. We will find that all three types of solution are equivalent for a large class of initial data. Since our primary interest is in vortex patches and layers, we will keep things simple by assuming compact support in $\Pi$. Assuming, then, that $g \in L^\ensuremath{\infty}_c({\ensuremath{\BB{R}}}^2)$---the space of essentially bounded functions with compact support---we define $\Cal{P}er(g)$ on $\Pi$ by \begin{align*} \Cal{P}er(g)({\bm{\mathrm{x}}}) = \sum_{n \in \ensuremath{\BB{Z}}} g({\bm{\mathrm{x}}} - (n, 0)), \end{align*} noting that for each ${\bm{\mathrm{x}}}$ the sum has only finitely many nonzero terms. For any measurable function $f$ on $\Pi$ we define $\Cal{R}ep(f)$ on ${\ensuremath{\BB{R}}}^2$ by \begin{align*} \Cal{R}ep(f)({\bm{\mathrm{x}}}) := f(x_1 - \lfloor x_1 + \tfrac{1}{2} \rfloor, x_2). \end{align*} \begin{definition}\label{D:Equivalence} Two functions $g_1, g_2 \in L^\ensuremath{\infty}_c({\ensuremath{\BB{R}}}^2)$ are equivalent, $g_1 \sim g_2$, if $\Cal{P}er(g_1) = \Cal{P}er(g_2)$. \Cref{f:PropertyP} depicts the support of two functions in the same equivalence class. \end{definition} \begin{figure}[ht] \IncludeFigure{0.12}{FigPropertyP} \vspace{-1.2em} \caption{\textit{Support of two $L^\ensuremath{\infty}_c({\ensuremath{\BB{R}}}^2)$ functions in the same equivalence class}}\label{f:PropertyP} \end{figure} Suppose that $g \in L^\ensuremath{\infty}_c({\ensuremath{\BB{R}}}^2)$, and for purposes of illustration, let us treat it as the characteristic function of a bounded domain (our primary application), whose support is depicted as in either (a) or (b) of \Cref{f:PropertyP}. Below, we construct an initial vorticity from $g$ and depict the support of $\omega^0$ for each type of solution (the time-evolved vorticity being of a similar nature). \begin{itemize}[leftmargin=5em,itemsep=0.5em] \item[\textbf{Type 1}] Let $\omega^0 = \Cal{R}ep(\Cal{P}er(g))$. \IncludeFigure{0.12}{FigType1} \item[\textbf{Type 2}] Let Let $\omega^0 = \Cal{P}er(g)$. \vspace{-0.5em} \IncludeFigure{0.12}{FigType2} \item[\textbf{Type 3}] Let $\omega^0 = g$. The vorticity $\omega$ is transported by the flow from the single copy of $g$, and so is no longer the curl of $\uu$. There are, in effect, multiple phantom copies of $g$ matching those of Type 1. \IncludeFigure{0.12}{FigType3a} \begin{center} or \end{center} \IncludeFigure{0.12}{FigType3b} \end{itemize} The vorticity $\omega^0$ for Type 1 and 2 do not depend upon the representative for the equivalence class, though Type 3 does. We will find, nonetheless, that the velocity field for solutions of Type 3 is independent of the representative. It is mentioned in \cite{Gallay2017} that a Type 2 solution is equivalent to a Type 1 solution with periodic velocity and pressure. Following up on this comment, we will show that all three types of solution are equivalent. The equivalence of Type 1 and Type 2 solutions, which applies to a larger class of initial data than we have so far discussed, will rely upon the properties of the pressure required for uniqueness for those two types of solution. The equivalence of Type 3 and Type 2 (and so of Type 1) will rest primarily on showing that solutions of Type 2 reduce to those of Type 3 when the vorticity has sufficient vertical decay. A side benefit of this approach is that it will give the well-posedness of Type 3 solutions. Such a well-posedness result could be obtained by adapting in a fairly straightforward way the approach Marchioro and Pulvirenti take in \cite{MP1984,MP1994} for the 2D Euler equations, except for subtle points regarding the periodicity of the pressure. It is thus more efficient to leverage the technology developed in \cite{AfendikovMielke2005,GallaySlijepcevic2014,GallaySlijepcevic2015}, though it is more than is strictly needed to develop Type 3 solutions alone. Specializing to vortex patch data, we will then show how the contour dynamics equation (CDE) is adapted from the classical form, which allows the propagation of regularity of the boundary of a vortex patch to be proved, adapting the argument of Bertozzi and Constantin in \cite{ConstantinBertozzi1993}. \subsection{Prior work}\label{S:ToDo} Bounded vortex patches evolving under the two-dimensional Euler equations have been well-studied, with global regularity of the boundary being established by Chemin \cite{cheminCRAS} and by Bertozzi and Constantin \cite{ConstantinBertozzi1993}. Regularity of the vortex patch boundary can also be seen to follow from a more general approach studying level sets of the vorticity, establishing striated regularity, as in the work of Chemin \cite{cheminENS} and Serfati \cite{serfatiStratifee}. Regularity of bounded vortex patches and/or striated regularity have been established for solutions of related evolution equations as well, such as aggregation equations \cite{bertozziEtAl}, active transport equations \cite{baeKelliher}, and the surface quasi-geostrophic equation and related systems \cite{chaeEtal}, \cite{gancedo}, \cite{kiselev}. None of these problems consider unbounded vortex patches as in the present work. There are seemingly fewer papers on the evolution of vortex layers. An equation similar to our version of the contour dynamics equation for the motion of the patch/layer boundary was developed in \cite{pullin}, and was subsequently used in \cite{golubeva} for the study of complex singularities in vortex layers. (We mention that the version of the contour dynamics equation developed in the present work lends itself to the study of global regularity.) Atassi, Bernoff, and Lichter study the interaction of a point vortex with a vortex layer \cite{bernoff}. Crowdy gives some exact solutions of vortex layers interacting with solid boundaries \cite{crowdy}. Benedetto and Pulvirenti have shown that vortex layers rigorously approximate vortex sheets in analytic function spaces \cite{benedettoPulvirenti}. Caflisch, Sammartino, and collaborators have considered vortex layers which are not sharp fronts in a series of papers \cite{caflisch2}, \cite{caflisch1}, \cite{caflisch3}, considering how such flows behave in the zero viscosity limit and how such flows may approximate vortex sheets, which represent a more singular vorticity configuration. In these works, they take the vorticity to be exponentially decaying (in the vertical direction) away from a core region, rather than being an indicator function as in the present work. Despite the difference there are similarities to the present work, such as the development of velocity integrals similar to the spatially periodic contour dynamics equation we develop for the periodic patch/layer problem. Further background on vortex layers may be found in \cite{gargano}. While we are unaware of other works on the global regularity of unbounded vortex patches for the two-dimensional Euler equations, the situation is different for the quasi-geostrophic equation. Rodrigo developed existence theory for a patch which is spatially periodic and vertically unbounded in one direction (similarly to a half-space) \cite{rodrigo1}, \cite{rodrigo2}. More recently Hunter, Shu, and Zhang have studied the related front solutions of the surface quasi-geostrophic equation \cite{hunter1}, \cite{hunter2}, \cite{hunter3}. \subsection{Organization of this paper} We will find many of our calculations much more convenient to perform in the complex plane, yet our results are all real-valued. We describe how to translate back and forth between these settings, largely a matter of notation, in \cref{S:RealComplex}. In \cref{S:PeriodicSolutions} we describe the process of symmetrizing in pairs that is behind the Type 1 solutions, which we explore in \cref{S:Type1}. In \cref{S:Type2} we describe the results of \cite{AfendikovMielke2005,GallaySlijepcevic2014,GallaySlijepcevic2015} that yield Type 2 solutions, and we use those results in \cref{S:Type3} to obtain Type 3 solutions. We show the equivalence of the three types of solution in \cref{S:ThreeTypes}. In \cref{S:VelocityGradient} we give expressions for the velocity gradient in terms of the vorticity, deferring the proofs to \cref{A:VelocityGradient}. We then specialize to vortex patch solutions for Type 1, 2, and 3 solutions, obtaining their contour dynamics equation in \cref{S:CDEPeriodic}, and establishing the global-in-time propogation of the regularity of a vortex patch boundary in \cref{S:BoundaryRegularity}. \section{Preliminaries: ${\ensuremath{\BB{R}}}^2$ and $\C$}\label{S:RealComplex} \subsection{Real to complex translation} \noindent Some of our calculations will be more easily performed using complex analysis, though the end results are all real-valued functions. For this we need a means, and a corresponding notation, to switch back and forth between viewing points in the plane as vectors or points in $\C^2$. For this purpose, we will use bold-face letters, such as ${\bm{\mathrm{x}}}$ or $\uu$, for quantities that are intrinsically elements of ${\ensuremath{\BB{R}}}^2$ or vector-valued. We define maps, \begin{align*} \left\{ \begin{array}{ll} \overrightarrow{} \, \colon \C \to {\ensuremath{\BB{R}}}^2, \\[4pt] \overrightarrow{x + i y} = (x, y) \end{array} \right\} \text{ and } \left\{ \begin{array}{ll} \overleftarrow{} \colon {\ensuremath{\BB{R}}}^2 \to \C, \\[4pt] \overleftarrow{(x, y)} = x + i y \end{array} \right\}. \end{align*} For a vector ${\bm{\mathrm{x}}} = (x, y)$, we define \begin{align*} {\bm{\mathrm{x}}}^\perp := (-y, x). \end{align*} Hence, ${\bm{\mathrm{x}}}^\perp$ is ${\bm{\mathrm{x}}}$ rotated 90 degrees counterclockwise. \begin{lemma}\label{L:vec} Let $z, w \in \C$ and $\cdot$ be the usual dot (inner) product of Euclidean vectors. Then \begin{align}\label{e:vecIds} \begin{split} \RE \, (zw) &= \vec{\overline{z}} \cdot \vec{w}, \\ \IM \, (zw) &= -\vec{\overline{z}} \cdot \vec{w}^\perp. \end{split} \end{align} If $a \in {\ensuremath{\BB{R}}}$, $z \in \C$, \begin{align}\label{e:vecProp} \overrightarrow{a z} = a \vec{z}, \quad \overrightarrow{i z} = \vec{z}^\perp, \quad \overleftarrow{{\bm{\mathrm{v}}}^\perp} = i \overleftarrow{{\bm{\mathrm{v}}}}. \end{align} Also, $f$ is analytic in some domain $U$ if and only if $\dv \vec{\overline{f}} = \curl \vec{\overline{f}} = 0$ in $U$, where for any vector field ${\bm{\mathrm{v}}}$, \begin{align*} \dv {\bm{\mathrm{v}}} &:= \pdx{v^1}{x_1} + \pdx{v^2}{x_2}, \quad \curl {\bm{\mathrm{v}}} := \pdx{v^2}{x_1} - \pdx{v^1}{x_2} \end{align*} are the divergence and (scalar) curl of ${\bm{\mathrm{v}}}$. \end{lemma} The boundary integrals we encounter will be real path integrals, but we will sometimes find it useful to transform them to complex contour integrals as in the following lemma: \begin{lemma}\label{L:ComplexToRealContourIntegrals} Let $\bgamma \colon [a, b] \to \C$ be a Lipschitz-continuous path on which the complex-valued function $f$ is continuous. Let $\BoldTau$ be the unit tangent vector in the direction of $\bgamma$ and ${\bm{n}}$ the associated unit normal, with $({\bm{n}}, \BoldTau)$ in the standard orientation of $(\bm{\mathrm{e}}_1, \bm{\mathrm{e}}_2)$. Let $C = \image \bgamma$. Then \begin{align*} \Xint{\C}_\bgamma f &= \int_C \vec{\overline{f}} \cdot \BoldTau + i \int_C \vec{\overline{f}} \cdot {\bm{n}}. \end{align*} Here, $\Xint{\C}$ is a complex contour integral. \end{lemma} Using \cref{L:vec}, it is not hard to rewrite the classical Biot-Savart law in the following hybrid real-complex form: \begin{theorem}\label{T:BSLawNS2} Assume that $\omega \in L^1 \cap L^\ensuremath{\infty}({\ensuremath{\BB{R}}}^2)$. With $K$ as in \cref{e:BSR2}, \begin{align}\label{e:BSLawNS2} \uu({\bm{\mathrm{x}}}) &:= K * \omega({\bm{\mathrm{x}}}) = -\overrightarrow{\frac{i}{2 \pi} \int_{{\ensuremath{\BB{R}}}^2} \frac{\omega({\bm{\mathrm{y}}})}{\text{ } \overline{\overleftarrow{{\bm{\mathrm{y}}} - {\bm{\mathrm{x}}}}} \text{ }} \, d {\bm{\mathrm{y}}}} \end{align} is divergence-free with $\curl \uu = \omega$, and $\uu$ is the unique such velocity field in $L^\ensuremath{\infty} \cap H^1({\ensuremath{\BB{R}}}^2)$. \end{theorem} \subsection{The cotangent} \begin{lemma}\label{L:ML} For any $z \in \C$ that is not an integer, \begin{align*} \pi \cot \pi z &= \frac{1}{z} + 2 \sum_{n = 1}^\ensuremath{\infty} \frac{z}{z^2 - n^2} = \lim_{N \to \ensuremath{\infty}} \sum_{n = -N}^N \frac{1}{z + n}. \end{align*} \end{lemma} \begin{proof} For the first equality see, for instance, Equation (11) in Section 5.2.1 of \cite{Ahlfors}. The second equality then follows from \begin{align*} \frac{z}{z^2 - n^2} &= \frac{z}{(z - n) (z + n)} = \frac{1}{2} \brac{\frac{1}{z - n} + \frac{1}{z + n}} \end{align*} and summing in pairs, $n$ with $-n$. \end{proof} \begin{lemma}\label{L:SumBringsIncot} For any ${\bm{\mathrm{x}}}, {\bm{\mathrm{y}}} \in {\ensuremath{\BB{R}}}^2$, \begin{align*} \lim_{N \to \ensuremath{\infty}} \sum_{n = -N}^N \frac{{\bm{\mathrm{x}}} + (n, 0)}{\abs{{\bm{\mathrm{x}}} + (n, 0)}^2} \cdot {\bm{\mathrm{y}}} = \pi \overrightarrow{\cot (\pi \overline{\overleftarrow{{\bm{\mathrm{x}}}}})} \cdot {\bm{\mathrm{y}}}. \end{align*} \end{lemma} \begin{proof} Letting $z = \overleftarrow{{\bm{\mathrm{x}}}}$, $w = \overleftarrow{{\bm{\mathrm{y}}}}$, and using \cref{{e:vecIds}}$_1$, we have \begin{align*} \frac{{\bm{\mathrm{x}}} + (n, 0)}{\abs{{\bm{\mathrm{x}}} + (n, 0)}^2} \cdot {\bm{\mathrm{y}}} &= \frac{\RE ((\overline{z} + n) w)}{\abs{z + n}^2} = \RE \frac{(\overline{z} + n) w}{\abs{z + n}^2} = \RE \frac{w}{z + n} \end{align*} so \begin{align*} \lim_{N \to \ensuremath{\infty}} &\sum_{n = -N}^N \frac{{\bm{\mathrm{x}}} + (n, 0)}{\abs{{\bm{\mathrm{x}}} + (n, 0)}^2} \cdot {\bm{\mathrm{y}}} = \RE \brac{w \lim_{N \to \ensuremath{\infty}} \sum_{n = -N}^N \frac{1}{z + n}} = \pi \RE (w \cot \pi z) \\ &= \pi \overrightarrow{\overline{\cot (\pi \overleftarrow{{\bm{\mathrm{x}}}})}} \cdot {\bm{\mathrm{y}}} = \pi \overrightarrow{\cot (\pi \overline{\overleftarrow{{\bm{\mathrm{x}}}}})} \cdot {\bm{\mathrm{y}}}, \end{align*} where we again used \cref{{e:vecIds}}$_1$. \end{proof} \subsection{Useful identities} The identities in \cref{e:sinIdentity,e:TrigIdentities} are easily verifiable; \cref{e:AbramowitzStegun1964} is 4.3.58 of \cite{AbramowitzStegun1964}. \begin{align} \abs{\sin z}^2 &= \sin^2 x + \sinh^2 y, \label{e:sinIdentity} \\ \cosh 2 x &= 2 \sinh^2 x + 1, \quad \cos 2 x = 1 - 2 \sin^2 x, \label{e:TrigIdentities} \\ \cot z &= \frac{\sin 2 x - i \sinh 2 y}{\cosh 2y - \cos 2x}. \label{e:AbramowitzStegun1964} \end{align} \subsection{Lifting paths and domains}\label{S:Lifting} We will find the need, in the proof of \cref{T:CDEType2}, to apply \cref{L:ComplexToRealContourIntegrals} while integrating in $\Pi$ and apply Cauchy's residue theorem. This could be done directly by introducing a version of the residue theorem for $\Pi$, which is a (flat) analytic manifold. Alternately, we can transform integrals in $\Pi$ to integrals of $x_1$-periodic functions in $\C$ by \textit{lifting} the domain $\Omega$ in $\Pi$ to a suitable domain $\widetilde{\Omega}$ in $\C$. Our main tool for doing this is the lifting of paths from a topological space to a covering space. Defining \begin{align*} p \colon \C \to \Pi, \quad p(x_1 + i x_2) = x_1 - \lfloor x_1 + \tfrac{1}{2} \rfloor + i x_2, \end{align*} we see that $(\C, p)$ is a covering space of $\Pi$ (see Section IX.7 of \cite{Co1978}, for instance). This will allow us to lift a path in $\Pi$ to a path in ${\ensuremath{\BB{R}}}^2$ or $\C$. \begin{remark} $\Cal{R}ep(f)({\bm{\mathrm{x}}}) = f(p({\bm{\mathrm{x}}}))$, though we do not make direct use of this. \end{remark} \begin{definition}\label{D:Lifting} A path in the topological space $X$ is a continuous map from an interval $I$ to $X$. The path $\widetilde{\bgamma}$ in $\C$ is a \textit{lift} or \textit{lifting} of the path $\bgamma$ in $\Pi$ if $p \circ \widetilde{\bgamma} = \bgamma$. \end{definition} \begin{lemma}\label{L:Lifting} Let $\bgamma$ be a finite length continuous path in $\Pi$ with initial point ${\bm{\mathrm{x}}}_0$. For any $\widetilde{{\bm{\mathrm{x}}}}_0 \in p^{-1}({\bm{\mathrm{x}}}_0)$, there exists a unique lifting $\widetilde{\bgamma}$ with initial point $\widetilde{{\bm{\mathrm{x}}}}_0$. \end{lemma} \begin{proof} This is a classical result; see, for instance, Corollary IX.7.5 of \cite{Co1978}. \end{proof} This lifting allows us to relate path integrals in $\Pi$ to lifted path integrals in ${\ensuremath{\BB{R}}}^2$ or $\C$: \begin{lemma}\label{L:LiftedIntegral} Let $\bgamma$ be a Lipschitz-continuous path in $\Pi$ with a lift $\widetilde{\bgamma}$ as given by \cref{L:Lifting}. For any any continuous function $f$ on $\Pi$, \begin{align*} \int_\bgamma f = \int_{\widetilde{\bgamma}} f \circ p. \end{align*} Moreover, the normal vector field ${\bm{n}}$ on $\bgamma$ lifts to itself as does $\BoldTau$; that is, ${\bm{n}}(\bgamma(\ensuremath{\alpha})) = {\bm{n}}(\widetilde{\bgamma}(\ensuremath{\alpha}))$ for all $\ensuremath{\alpha}$ in the domain of $\bgamma$ (which is the same as the domain of $\widetilde{\bgamma}$). \end{lemma} \begin{proof} Suppose that $\bgamma \colon [a, b] \to \Pi$, in which case also $\widetilde{\bgamma} \colon [a, b] \to \C$ with $p \circ \widetilde{\bgamma} = \bgamma$. Then \begin{align*} \int_{\widetilde{\bgamma}} f \circ p &= \int_a^b f \circ p(\widetilde{\bgamma}(\ensuremath{\alpha})) \widetilde{\bgamma}'(\ensuremath{\alpha}) \, d \ensuremath{\alpha} = \int_a^b f (\bgamma(\ensuremath{\alpha})) \bgamma'(\ensuremath{\alpha}) \, d \ensuremath{\alpha} = \int_\bgamma f. \end{align*} We used that $\widetilde{\bgamma}'(\ensuremath{\alpha}) = \bgamma'(\ensuremath{\alpha})$, since locally $\widetilde{\bgamma}$ and $\bgamma$ differ by a constant (if we view $\bgamma$ as giving values in ${\Pi_p}$). This also gives that ${\bm{n}}$ and $\BoldTau$ lift to themselves. \end{proof} \cref{L:LiftedIntegral} is not, however, the entire story when we lift the entire boundary of a domain in $\Pi$. An immediate difficulty stems from the ambient space $\Pi$, which is topologically a cylinder, having nontrivial fundamental (and first homology) group $\ensuremath{\BB{Z}}$. Let us say that a closed curve on $\Pi$ wraps around the cylinder $n$ times if it crosses $\set{x_1 = 0}$ (any vertical slice would do) $n$ times counted with sign, positive in one direction, negative in the other (arbitrarily fixing which direction is positive). A closed path that wraps zero times around the cylinder is homotopic to a point and lifts to a closed path in $\C$. A closed path that wraps around the cylinder $n$ times, however, will lift by \cref{L:Lifting} to a non-closed path in $\C$ that contains $\abs{n} + 1$ points of ${\bm{\mathrm{x}}}_0 + \Cal{L}$, where we define here and for future use, \begin{align}\label{e:L} \Cal{L} := \set{\ensuremath{\BB{Z}}} \times \set{0}, \quad \Cal{L}^* := \Cal{L} \setminus (0, 0), \end{align} treated as subsets of ${\ensuremath{\BB{R}}}^2$ or of $\C$. Since we are lifting paths that are boundary components, they will always be closed in $\Pi$, but can wrap only $0$ or $\pm 1$ times around the cylinder else they would of necessity self-intersect. \begin{figure}[ht] \IncludeFigure{0.14}{FigBoundaryLift} \vspace{-1.5em} \caption{\textit{Lifting of $\ensuremath{\partial} \Omega$ with base points at $x_1 = - \tfrac{1}{2}$}}\label{f:BoundaryLift} \end{figure} \Cref{f:BoundaryLift} shows an example of a domain $\Omega$ in $\Pi$ having two boundary components $\Gamma_1$, $\Gamma_2$ which lift to non-closed paths $\widetilde{\Gamma}_1$, $\widetilde{\Gamma}_2$. To make a domain from these paths, we could connect $\widetilde{\Gamma}_1$, $\widetilde{\Gamma}_2$ with vertical paths at $x_1 = - \tfrac{1}{2}$ and $x_1 = \tfrac{1}{2}$, oppositely oriented, so that the four paths together form the boundary of a lifted domain $\widetilde{\Omega}$. Equivalently, and in a manner more easily generalizable, we cut the cylinder $\Pi$ vertically\footnote{In pathological cases, we would have to perturb this cut to avoid producing an infinite number of boundary components, but we will not explore this issue further.} along the line $\ell = \set{x_1 = \pm \tfrac{1}{2}}$, which in effect means we view $\Pi$ in the form suggested in \cref{e:Pi}. For any line segment formed by $\ell \cap \ensuremath{\partial} \Omega$ we introduce oppositely oriented paths; together, the lifted components of $\ensuremath{\partial} \Omega$ and these paths, properly oriented, form the boundary components of the lifted domain $\widetilde{\Omega}$. In lifting these components and paths, however, we need to insure compatible initial points for the paths. To do this, fix any ${\bm{\mathrm{x}}}_0$ in $\Omega$. Let ${\bm{\mathrm{y}}}$ be any point in $\Omega$ and let $\bgamma_{\bm{\mathrm{y}}}$ be a path connecting ${\bm{\mathrm{x}}}_0$ to ${\bm{\mathrm{y}}}$. Being a domain, $\Omega$ is path-connected so this is always possible. By \cref{L:Lifting}, there is a unique lifting $\widetilde{\bgamma}_{\bm{\mathrm{y}}}$ of $\bgamma_{\bm{\mathrm{y}}}$ with initial point $\widetilde{{\bm{\mathrm{x}}}}_0$. Then $\widetilde{\Omega} := \cup_{{\bm{\mathrm{y}}} \in \Omega} \, \widetilde{\bgamma}_{\bm{\mathrm{y}}}$ is the desired lifting of $\Omega$. Lifted in this way, we have the following lemma: \begin{lemma}\label{L:LiftedBoundaryIntegral} Let $\Omega$ be a bounded domain in $\Pi$ and let $\widetilde{\Omega}$ be the lifted domain as described above. Let $\bgamma$ be a parameterization of $\ensuremath{\partial} \Omega$ and $\widetilde{\bgamma}$ a parameterization of $\ensuremath{\partial} \widetilde{\Omega}$. Let $f$ be any continuous complex-valued function. Then \begin{align*} \Xint{\C}_\bgamma f = \Xint{\C}_{\widetilde{\bgamma}} f \circ p, \quad \int_{\ensuremath{\partial} \Omega} \vec{\overline{f}} \cdot \BoldTau = \int_{\ensuremath{\partial} \widetilde{\Omega}} (\vec{\overline{f}} \circ p) \cdot \BoldTau, \quad \int_{\ensuremath{\partial} \Omega} \vec{\overline{f}} \cdot {\bm{n}} = \int_{\ensuremath{\partial} \widetilde{\Omega}} (\vec{\overline{f}} \circ p) \cdot {\bm{n}}. \end{align*} \end{lemma} \begin{proof} Follows from \cref{L:LiftedIntegral}, since the cuts introduce integrals that cancel in pairs. \end{proof} \section{Periodized functions and Biot-Savart kernels}\label{S:PeriodicSolutions} \noindent \begin{definition}\label{D:SymVel} Let $\omega \in L^1 \cap L^\ensuremath{\infty}({\ensuremath{\BB{R}}}^2)$. We say that the velocity field $\uu$ is obtained by symmetrizing in pairs (about $0$) if, letting $\omega^{(n)}({\bm{\mathrm{x}}}) = \omega({\bm{\mathrm{x}}} + (n, 0))$, we have \begin{align*} \uu = K_{sym}[\omega] := K * \omega + \sum_{n = 1}^\ensuremath{\infty} K*\pr{\omega^{(-n)} + \omega^{(n)}}. \end{align*} \end{definition} \begin{definition}\label{D:S} Let $S = S({\ensuremath{\BB{R}}}^2)$ be the Serfati space of bounded, divergence-free vector fields on ${\ensuremath{\BB{R}}}^2$ having bounded vorticity with norm, \begin{align*} \norm{\uu}_S := \norm{\uu}_{L^\ensuremath{\infty}({\ensuremath{\BB{R}}}^2)} + \norm{\curl \uu}_{L^\ensuremath{\infty}({\ensuremath{\BB{R}}}^2)}. \end{align*} We define $S(\Pi)$ similarly. \end{definition} \begin{remark} As shown in (2.11) of \cite{GallaySlijepcevic2015}, for any $\omega \in L^\ensuremath{\infty}(\Pi)$ there is a divergence-free vector field $\uu$ in $L^\ensuremath{\infty}(\Pi)$ and so in $S(\Pi)$ for which $\curl \uu = \omega$. $S({\ensuremath{\BB{R}}}^2)$ is very different, for there is no known general condition on $\omega \in L^\ensuremath{\infty}({\ensuremath{\BB{R}}}^2)$ alone that guarantees a $\uu$ in $L^\ensuremath{\infty}({\ensuremath{\BB{R}}}^2)$. \end{remark} \begin{prop}\label{P:SymConv} For $\omega \in L^\ensuremath{\infty}_c({\ensuremath{\BB{R}}}^2)$, let $\uu = K_{sym}[\omega]$ as in \cref{D:SymVel}. Then $\uu \in S({\ensuremath{\BB{R}}}^2)$ with $\curl \uu = \curl K_{sym}[\omega] = \Cal{R}ep (\omega)$. Further, \begin{align}\label{e:KInf} \uu &= K_{sym}[\omega] = K_\ensuremath{\infty} * \omega, \quad K_\ensuremath{\infty}({\bm{\mathrm{x}}}) := \overrightarrow{-\frac{i}{2} \cot \pi \overline{\overleftarrow{{\bm{\mathrm{x}}}}}} = \frac{1}{2} \brac{\overrightarrow{\cot \pi \overline{\overleftarrow{{\bm{\mathrm{x}}}}}}}^\perp, \end{align} where we note that $K_\ensuremath{\infty}$ is periodic in $x_1$ with period $1$ as is $\uu$. We also have, \begin{align}\label{e:KH} K_\ensuremath{\infty}({\bm{\mathrm{x}}}) &= K({\bm{\mathrm{x}}}) + H({\bm{\mathrm{x}}}), \end{align} where $H$ is harmonic on ${\ensuremath{\BB{R}}}^2 \setminus \Cal{L}^*$, where $\Cal{L}^*$ is defined in \cref{e:L}. \end{prop} \begin{proof} Applying \cref{T:BSLawNS2}, we have \begin{align*} I_n &:= K*\pr{\omega^{(-n)} + \omega^{(n)}}({\bm{\mathrm{x}}}) = -\overrightarrow{\frac{i}{2 \pi} \int_{{\ensuremath{\BB{R}}}^2} \brac{ \frac{\omega({\bm{\mathrm{y}}})}{\overline{\overleftarrow{{\bm{\mathrm{y}}} - {\bm{\mathrm{x}}}}} - n} } + \brac{ \frac{\omega({\bm{\mathrm{y}}})}{\overline{\overleftarrow{{\bm{\mathrm{y}}} - {\bm{\mathrm{x}}}}} + n} } \, d {\bm{\mathrm{y}}} }, \end{align*} so \begin{align*} \overleftarrow{I_n} &= -\frac{i}{2 \pi} \int_{{\ensuremath{\BB{R}}}^2} 2 \frac{\overline{\overleftarrow{{\bm{\mathrm{y}}} - {\bm{\mathrm{x}}}}}} {(\overline{\overleftarrow{{\bm{\mathrm{y}}} - {\bm{\mathrm{x}}}}})^2 - n^2} \omega({\bm{\mathrm{y}}}) \, d {\bm{\mathrm{y}}}. \end{align*} From \cref{D:SymVel} with \cref{L:ML}, then (the compact support of $\omega$ allows us to interchange integration and summation), \begin{align*} \overleftarrow{\uu({\bm{\mathrm{x}}})} &= -\frac{i}{2} \int_{{\ensuremath{\BB{R}}}^2} \cot (\pi (\overline{\overleftarrow{{\bm{\mathrm{y}}} - {\bm{\mathrm{x}}}}})) \omega({\bm{\mathrm{y}}}) \, d {\bm{\mathrm{y}}}, \end{align*} and \cref{{e:KInf}} follows from \cref{e:vecProp}. Since the singularity of $\cot (\pi z)$ at $z = 0$ is like $1/(\pi z)$ and $\omega$ is compactly supported, we see that the above integral lies in $L^\ensuremath{\infty}({\ensuremath{\BB{R}}}^2)$. Since the curl of each $I_n$ is $\omega^{(-n)} + \omega^{(n)}$ while its divergence is zero and the sum converges absolutely and uniformly, we know that $\dv \uu = 0$ and $\curl \uu = \Cal{R}ep (\omega)$. But $ \cot z = \frac{1}{z} + h(z) $ on $\C \setminus \Cal{L}^*$, where $h$ is analytic. From this \cref{e:KH} follows. \end{proof} \begin{prop} If $\omega_1 \sim \omega_2$ in $L^\ensuremath{\infty}_c({\ensuremath{\BB{R}}}^2)$ as in \cref{D:Equivalence} then $K_\ensuremath{\infty} * \omega_1 = K_\ensuremath{\infty} * \omega_2$. \end{prop} \begin{proof} For any $\omega \in L^\ensuremath{\infty}_c({\ensuremath{\BB{R}}}^2)$, \begin{align*} K_\ensuremath{\infty} * &\Cal{P}er(\omega)({\bm{\mathrm{x}}}) = \int_\Pi K_\ensuremath{\infty}({\bm{\mathrm{x}}} - {\bm{\mathrm{y}}}) \Cal{P}er(\omega)({\bm{\mathrm{y}}}) \, d{\bm{\mathrm{y}}} = \int_{\Pi_p} K_\ensuremath{\infty}({\bm{\mathrm{x}}} - {\bm{\mathrm{y}}}) \sum_{n \in \ensuremath{\BB{Z}}} \omega({\bm{\mathrm{y}}} - (n, 0)) \, d{\bm{\mathrm{y}}} \\ &= \int_{\Pi_p} \sum_{n \in \ensuremath{\BB{Z}}} K_\ensuremath{\infty}({\bm{\mathrm{x}}} - ({\bm{\mathrm{y}}} - (n, 0))) \omega({\bm{\mathrm{y}}} - (n, 0)) \, d{\bm{\mathrm{y}}} \\ &= \sum_{n \in \ensuremath{\BB{Z}}} \int_{\Pi_p} K_\ensuremath{\infty}({\bm{\mathrm{x}}} - ({\bm{\mathrm{y}}} - (n, 0))) \omega({\bm{\mathrm{y}}} - (n, 0)) \, d{\bm{\mathrm{y}}} = \sum_{n \in \ensuremath{\BB{Z}}} \int_{{\Pi_p} - (n, 0)} K_\ensuremath{\infty}({\bm{\mathrm{x}}} - {\bm{\mathrm{y}}}) \omega({\bm{\mathrm{y}}}) \, d{\bm{\mathrm{y}}} \\ &= \int_{{\ensuremath{\BB{R}}}^2} K_\ensuremath{\infty}({\bm{\mathrm{x}}} - {\bm{\mathrm{y}}}) \omega({\bm{\mathrm{y}}}) \, d{\bm{\mathrm{y}}} = K_\ensuremath{\infty} * \omega({\bm{\mathrm{x}}}). \end{align*} We were able to interchange the integral and sum here because for any fixed ${\bm{\mathrm{x}}}$, the compact support of $\omega$ makes all but a finite number of terms in the sum zero. Hence, if $\omega_1 \sim \omega_2$ then $K_\ensuremath{\infty} * \omega_1 = K_\ensuremath{\infty} * \Cal{P}er(\omega) = K_\ensuremath{\infty} * \omega_2$. \end{proof} We will see in \cref{S:BSStrip} that $K_\ensuremath{\infty}$ also serves as the Biot-Savart kernel on $\Pi$. \section{Type 1: Periodized solutions}\label{S:Type1} \noindent We review here results, obtained variously in \cite{Serfati1995Bounded,AKLL2015,TaniuchiEtAl2010,Kelliher2015}, on bounded vorticity, bounded velocity solutions to the 2D Euler equations in ${\ensuremath{\BB{R}}}^2$. Let $a_R$ be a radial cutoff function: $a_R(\cdot) = a(\cdot/R)$ for any $ R > 0$, where $a \in C^\ensuremath{\infty}({\ensuremath{\BB{R}}}^2)$ is radially symmetric and equal to $1$ in a neighborhood of the origin. For definitiveness, we will assume that $a \equiv 1$ on $B_1(0)$, $a \equiv 0$ on $B_2(0)^C$, and $\abs{a} \le 1$ on ${\ensuremath{\BB{R}}}^2$. \begin{theorem}[\cite{Kelliher2015}]\label{T:Uinf} Any weak solution to the Euler equations (Eulerian or Lagrangian) with $\uu \in L^\ensuremath{\infty}(0, T; S) \cap C([0, T] \times {\ensuremath{\BB{R}}}^2)$ having vorticity $\omega$ with $\uu(0) = \uu^0$, $\omega(0) = \omega^0$, must satisfy, for some ${\bm{\mathrm{U}}}_\ensuremath{\infty} \in C([0, T])^2$, the Serfati identity, \begin{align}\label{e:SerfatiId} \begin{split} u^j(t&) - (u^0)^j = U^j_\ensuremath{\infty}(t) + (a K^j) *(\omega(t) - \omega^0) \\ &\qquad- \int_0^t \pr{\ensuremath{\nabla} \ensuremath{\nabla}^\perp \brac{(1 - a) K^j}} \mathop{* \cdot} (u \otimes u)(s) \, ds, \end{split} \end{align} $j = 1, 2$, and the renormalized Biot-Savart law, \begin{align}\label{e:BSRenorm} \uu(t) - \uu^0 &= {\bm{\mathrm{U}}}_\ensuremath{\infty}(t) + \lim_{R \to \ensuremath{\infty}} (a_R K) * (\omega(t) - \omega^0) \end{align} on $[0, T] \times {\ensuremath{\BB{R}}}^2$. Furthermore, the corresponding pressure is of the form, \begin{align}\label{e:pqPressure} p(t, {\bm{\mathrm{x}}}) = -{\bm{\mathrm{U}}}_\ensuremath{\infty}'(t) \cdot {\bm{\mathrm{x}}} + q(t, {\bm{\mathrm{x}}}), \end{align} where $q$ grows sublinearly at infinity. \end{theorem} \cref{T:Uinf} characterizes solutions to the 2D Euler equations that have bounded vorticity and bounded velocity: their existence and uniqueness under the condition that \cref{e:SerfatiId} holds is shown, for ${\bm{\mathrm{U}}}_\ensuremath{\infty} \equiv 0$, in \cite{Serfati1995Bounded} and elaborated on in \cite{AKLL2015}, their extension to a general ${\bm{\mathrm{U}}}_\ensuremath{\infty}$ being a simple matter. Uniqueness under the assumption of sublinear growth of the pressure is established in \cite{TaniuchiEtAl2010}. Combining these results leads to the following: \begin{theorem}\label{T:Type1} Let $\uu^0 \in S({\ensuremath{\BB{R}}}^2)$ and set $\omega^0 = \curl \uu^0$. There exists a solution $(\uu, p)$ to the 2D Euler equations with $\uu \in L^\ensuremath{\infty}(0, T; S) \cap C([0, T] \times {\ensuremath{\BB{R}}}^2)$ having initial velocity $\uu^0$. Existence and uniqueness hold if we require that the solution satisfy any one (and hence all) of \crefrange{e:SerfatiId}{e:pqPressure} with ${\bm{\mathrm{U}}}_\ensuremath{\infty} \equiv 0$. \end{theorem} \section{Type 2: Solutions in an infinite periodic strip}\label{S:Type2} \noindent Let $\BUC(\Pi)$ be the space of bounded, uniformly continuous functions, noting that any vector field in $S(\Pi)$ lies in $\BUC(\Pi)$. Well-posedness of solutions to the Navier-Stokes equations for initial velocity in $\BUC(\Pi)$ was established by Afendikov and Mielke in \cite{AfendikovMielke2005}. Building on this, Gallay and Slijep\v{c}evi\'{c} in \cite{GallaySlijepcevic2015} (and see the comments in \cite{Gallay2017}) obtained improved bounds for the case where the initial velocity lies in $S(\Pi)$, having established properties of the pressure in \cite{GallaySlijepcevic2014}. These works are for the Navier-Stokes equations, but as the authors point out, the pertinent estimates are uniform in small viscosity and hold for solutions to the Euler equations as well (by repeating the argument with the viscous terms missing or by using known vanishing viscosity results). In \cref{T:Type2} we give the well-posedness result as derived from \cite{AfendikovMielke2005,GallaySlijepcevic2014,GallaySlijepcevic2015}, but for this we need to first explore some aspects of the analysis in these references. \subsection{Biot-Savart kernels}\label{S:BSStrip} The authors of \cite{AfendikovMielke2005,GallaySlijepcevic2014,GallaySlijepcevic2015} orient their periodic strip (infinite cylinder) horizontally and $S^1$ is, in effect, parametrized from $0$ to $1$. Let $(x_1', x_2')$ be the coordinates for the horizontal strip of \cite{AfendikovMielke2005,GallaySlijepcevic2014,GallaySlijepcevic2015}, while we will keep $(x_1, x_2)$ for our vertical strip. Rotating the horizontal strip 90 degrees counterclockwise induces the change of variables, \begin{align*} x_1' \mapsto x_2, \quad x_2' \mapsto -x_1. \end{align*} The Biot-Savart kernel on $\Pi$ used in \cite{AfendikovMielke2005} and (2.7) of \cite{GallaySlijepcevic2014} is $\ensuremath{\nabla}^\perp G$, where \begin{align*} G(x_1', x_2') := \frac{1}{4 \pi} \log \pr{2 \cosh(2 \pi x_1') - 2 \cos(2 \pi x_2')} \end{align*} is the Green's function for the Dirichlet Laplacian on $\Pi$. In $(x_1, x_2)$ variables, \begin{align}\label{e:Green} G(x_1, x_2) := \frac{1}{4 \pi} \log \pr{2 \cosh(2 \pi x_2) - 2 \cos(2 \pi x_1)}. \end{align} \begin{lemma}\label{L:KInfgradperG} We have $K_\ensuremath{\infty} = \ensuremath{\nabla}^\perp G$. Moreover, $G({\bm{\mathrm{x}}}) = (2 \pi)^{-1} \log \rho({\bm{\mathrm{x}}}) $, where \begin{align}\label{e:rho} \rho({\bm{\mathrm{x}}}) &:= \pr{\sin^2(\pi x_1) + \sinh^2(\pi x_2)}^{\frac{1}{2}}. \end{align} \end{lemma} \begin{proof} From \cref{e:TrigIdentities}, $2 \cosh(2 \pi x_2) - 2 \cos(2 \pi x_1) = 4 \rho({\bm{\mathrm{x}}})^2$, gives our alternate expression for $G$ (noting that the Green's function on $\Pi$ is unique up to an additive constant). From \cref{e:Green,e:AbramowitzStegun1964}, we have \begin{align}\label{e:KInfRealForm} \ensuremath{\nabla}^\perp G(x_1, x_2) = \frac{1}{2 \pi} \frac{(- \pi \sinh (2 \pi x_2), \pi \sin(2 \pi x_1))} {\cosh(2 \pi x_2) - \cos(2 \pi x_1)} = \frac{1}{2} \overrightarrow{\cot(\pi \overline{z})}^\perp, \end{align} matching the expression for $K_\ensuremath{\infty}$ in \cref{e:KInf}. Here, we used \cref{e:AbramowitzStegun1964}. \end{proof} \begin{lemma}\label{L:loglogDiff} The function $\log \rho({\bm{\mathrm{x}}}) - \log \abs{{\bm{\mathrm{x}}}}$ is harmonic on ${\ensuremath{\BB{R}}}^2 \setminus \Cal{L}^*$, where $\rho$ is defined in \cref{e:rho}. \end{lemma} \begin{proof} Letting $z = \overleftarrow{{\bm{\mathrm{x}}}}$, we have, using \cref{e:sinIdentity}, \begin{align*} \log \rho&({\bm{\mathrm{x}}}) - \log \abs{{\bm{\mathrm{x}}}} = \frac{1}{2} \log \abs[\bigg]{\frac{\rho({\bm{\mathrm{x}}})^2}{\abs{{\bm{\mathrm{x}}}}^2}} = \frac{1}{2} \log \abs[\bigg]{\frac{\sin z}{z}}^2 = \log \abs[\bigg]{\frac{\sin z}{z}} = \RE \log \frac{\sin z}{z}, \end{align*} which is the real part of a function that is complex analytic on $\C \setminus \Cal{L}^*$. \end{proof} \subsection{Mean horizontal values}\label{S:MeanHorizontal} As observed below Lemma 2.2 of \cite{AfendikovMielke2005}, although $K_\ensuremath{\infty} \in L^1_{loc}(\Pi)$, $K_\ensuremath{\infty}^2 \in L^1(\Pi)$ (accounting for the different orientation of the strip). Moreover, convolution with $K_\ensuremath{\infty}^1$ can be handled by subtracting from $u^2$ its mean horizontal value to give it mean value zero. We summarize here this process as described on page 1748 of \cite{GallaySlijepcevic2014}. If ${\bm{\mathrm{v}}}(t) \in S(\Pi)$, the mean value of $v^2(t)$ along the horizontal line segment $x_2 = a$ is independent of $a \in {\ensuremath{\BB{R}}}$, and if $({\bm{\mathrm{v}}}, p)$ solves the Euler equations on $\Pi$ then it is independent of time as well. Hence, we can define \begin{align}\label{e:m2} m_2(t) &= m_2[{\bm{\mathrm{v}}}(t)] := \innp{v^2(t)}, \end{align} the mean value of $v^2(t)$ along any such horizontal line segment and we will have $\innp{v^2(t)} = \innp{v^2_0}$. The mean value of $v^1(t)$, however, will depend upon $x_2$, so we write \begin{align*} m_1(t, x_2) &= m_1[{\bm{\mathrm{v}}}(t)](x_2) := \int_{-\frac{1}{2}}^{\frac{1}{2}} v^1(t, x_1, x_2) \, d x_1. \end{align*} Similarly, we define \begin{align*} \innp{\omega}(t, x_2) := \int_{-\frac{1}{2}}^{\frac{1}{2}} \omega(t, x_1, x_2) \, d x_1 \end{align*} and $\widehat{\omega}(t, x_1, x_2) := \omega(t, x) - \innp{\omega}(t, x_2)$. Also, \begin{align}\label{e:omegaMeanprt2m1} \innp{\omega}(t, x_2) &= \innp{\ensuremath{\partial}_1 u^2 - \ensuremath{\partial}_2 u^1}(t, x_2) = -\innp{\ensuremath{\partial}_2 u^1}(t, x_2) = - \ensuremath{\partial}_2 m_1(t, x_2). \end{align} A form of the Biot-Savart law given in (2.5, 2.6) of \cite{GallaySlijepcevic2014} (suppressing the time variable) is \begin{align}\label{e:BSGS2014} {\bm{\mathrm{v}}}({\bm{\mathrm{x}}}) &= \begin{pmatrix} -m_1(x_2) \\ m_2 \end{pmatrix} + \int_{-\ensuremath{\infty}}^\ensuremath{\infty} \int_{-\frac{1}{2}}^{\frac{1}{2}} K_\ensuremath{\infty}({\bm{\mathrm{x}}} - {\bm{\mathrm{y}}}) \widehat{\omega}({\bm{\mathrm{y}}}) \, d y_1 \, d y_2. \end{align} We note here that in transforming from the expression as written in \cite{GallaySlijepcevic2014}, a velocity $(v^1, v^2)$ in $(x_1', x_2')$ becomes $(v^2, - v^1)$ in $(x_1, x_2)$, which accounts for the minus sign in $-m_1(x_2)$. \subsection{Type 2 solutions} We can now summarize the known result we need for Type 2 solutions: \begin{theorem}[\cite{AfendikovMielke2005,GallaySlijepcevic2014,GallaySlijepcevic2015}]\label{T:Type2} For ${\bm{\mathrm{v}}}^0 \in S(\Pi)$ with $\innp{v^2_0} = 0$ there exists a unique solution $({\bm{\mathrm{v}}}, q)$ to the Euler equations, \begin{align}\label{e:EType2} \begin{cases} \ensuremath{\partial}_t {\bm{\mathrm{v}}} + {\bm{\mathrm{v}}} \cdot \ensuremath{\nabla} {\bm{\mathrm{v}}} + \ensuremath{\nabla} q = 0 &\text{in } [0, \ensuremath{\infty}) \times \Pi, \\ \dv {\bm{\mathrm{v}}} = 0 &\text{in } [0, \ensuremath{\infty}) \times \Pi, \\ {\bm{\mathrm{v}}}(0) = {\bm{\mathrm{v}}}^0 &\text{in } \Pi \end{cases} \end{align} for which $m_2(t) \equiv 0$ with ${\bm{\mathrm{v}}} \in C([0, \ensuremath{\infty}); BUC(\Pi)) \cap L^\ensuremath{\infty}([0, \ensuremath{\infty}); S(\Pi))$ and pressure $q \in W^{1, \ensuremath{\infty}}([0, \ensuremath{\infty}) \times \Pi)$. The pressure is given by\footnote{$+ 2 K_\ensuremath{\infty}^2$ is $-\ensuremath{\partial}_2 G$ in (2.8) of \cite{GallaySlijepcevic2014}: we have made the transformation from a horizontal to a vertical strip.} \begin{align*} q = - (u^2)^2 + 2 K_\ensuremath{\infty}^2 * (\omega u^1). \end{align*} The solutions are Eulerian in velocity and satisfy the vorticity equation. Moreover, $\uu$ can be recovered from $\omega$ by the Biot-Savart law as in \cref{e:BSGS2014}. \end{theorem} \subsection{Compactly supported vorticity} As a prelude to obtaining Type 3 solutions, let us consider the special case of Type 2 solutions that we can obtain when the vorticity is compactly supported in $\Pi$. First, we specialize the Biot-Savart law in \cref{e:BSGS2014} to compactly supported vorticity. \begin{lemma}\label{L:KInfm2m1} Let ${\bm{\mathrm{v}}} \in S(\Pi)$ with $\omega := \curl {\bm{\mathrm{v}}}$ compactly supported in $\Pi$. Then $m_1(-\ensuremath{\infty}) + m_1(\ensuremath{\infty}) \equiv m_2 \equiv 0$ if and only if ${\bm{\mathrm{v}}} = K_\ensuremath{\infty} * \omega$. \end{lemma} \begin{proof} Since $\innp{\omega} = - \ensuremath{\partial}_2 m_1$, we have \begin{align*} I^j &:= \brac{\int_{-\ensuremath{\infty}}^\ensuremath{\infty} \int_{-\frac{1}{2}}^{\frac{1}{2}} K_\ensuremath{\infty}({\bm{\mathrm{x}}} - {\bm{\mathrm{y}}}) \innp{\omega}({\bm{\mathrm{y}}}) \, d y_1 \, d y_2}^j \\ &= \int_{-\frac{1}{2}}^{\frac{1}{2}} \int_{-\ensuremath{\infty}}^\ensuremath{\infty} K_\ensuremath{\infty}^j((x - x', y - y')) \ensuremath{\partial}_2 m_1(y') \, d y' \, d x'. \end{align*} \cref{L:IntK}, below, gives that $I^2 = 0$. We now consider $I^1$. Because $\omega$ is compactly supported within some $[-1/2, 1/2] \times [-R_0, R_0]$, so, too, are $\innp{\omega}$ and then, by \cref{e:omegaMeanprt2m1}, $\ensuremath{\partial}_2 m_1$. Choose $\varphi \in C_C^\ensuremath{\infty}({\ensuremath{\BB{R}}})$ equal to 1 on $[-R, R]$ and equal to zero outside $[-R + 1, R + 1]$ where we will choose $R \ge R_0$ more precisely later. Let $m_1^\ensuremath{\varepsilon} = \eta_\ensuremath{\varepsilon} * m_1$, where $\eta_\ensuremath{\varepsilon}$ is a (compactly supported) Friedrich's mollifier. As in \cite{AfendikovMielke2005}, we treat $K_\ensuremath{\infty}^1$ as a distribution on $\Pi$ with $\varphi m_1^\ensuremath{\varepsilon}$ a test function. Since also $K_\ensuremath{\infty}^1 \in L^1_{loc}(\Pi)$, we have, for fixed ${\bm{\mathrm{x}}}$, \begin{align*} I^1 &= \lim_{\ensuremath{\varepsilon} \to 0} \int_{-\frac{1}{2}}^{\frac{1}{2}} \int_{-\ensuremath{\infty}}^\ensuremath{\infty} K_\ensuremath{\infty}^1((x - x', y - y')) \varphi(y) \ensuremath{\partial}_2 m_1^\ensuremath{\varepsilon}(y') \, d y' \, d x' \\ &= \lim_{\ensuremath{\varepsilon} \to 0} K_\ensuremath{\infty}^1 * (\varphi \ensuremath{\partial}_2 m_1^\ensuremath{\varepsilon}) = \lim_{\ensuremath{\varepsilon} \to 0} K_\ensuremath{\infty}^1 * \ensuremath{\partial}_2 (\varphi m_1^\ensuremath{\varepsilon}) - \lim_{\ensuremath{\varepsilon} \to 0} K_\ensuremath{\infty}^1 * (\ensuremath{\partial}_2 \varphi \, m_1^\ensuremath{\varepsilon}). \end{align*} Now, \begin{align*} \ensuremath{\partial}_2 K_\ensuremath{\infty}^1 &= - \ensuremath{\partial}_2^2 G = - \Delta G + \ensuremath{\partial}_1^2 G = - \delta + \ensuremath{\partial}_1^2 G, \end{align*} where $G$ is the Green's function for the Dirichlet Laplacian on $\Pi$ as in \cref{e:Green} and $\delta$ is the Dirac delta function on $\Pi$. Hence, \begin{align*} \ensuremath{\partial}_2 (\varphi m_1^\ensuremath{\varepsilon}) &= m_1^\ensuremath{\varepsilon}(x_2) - \int_{-\ensuremath{\infty}}^\ensuremath{\infty} \int_{-\frac{1}{2}}^{\frac{1}{2}} \ensuremath{\partial}_1^2 G((x - x', y - y')) \, dx' m_1^\ensuremath{\varepsilon}(y') \, d y' = m_1^\ensuremath{\varepsilon}(x_2), \end{align*} where the integral vanishes after integrating by parts, since $G$ is periodic in $x_1$. Hence, \begin{align*} I^1 &= m_1(x_2) - \lim_{\ensuremath{\varepsilon} \to 0} K_\ensuremath{\infty}^1 * (\ensuremath{\partial}_2 \varphi m_1^\ensuremath{\varepsilon}), \end{align*} and this equality holds regardless of our choice of $R \ge R_0$. Therefore, if we can evaluate $K_\ensuremath{\infty}^1 * (\ensuremath{\partial}_2 \varphi m_1^\ensuremath{\varepsilon})$ in the limit as $R \to \ensuremath{\infty}$, it will be its common value for all $R \ge R_0$. We see from \cref{e:KInfRealForm} that $K_\ensuremath{\infty}^1(x - y) \to \pm 1/2$ as $y_2 \to \pm \ensuremath{\infty}$ and $\ensuremath{\partial}_2 K_\ensuremath{\infty}^1(x - y) \to 0$ as $y_2 \to \pm \ensuremath{\infty}$, so \begin{align*} \lim_{R \to \ensuremath{\infty}} &K_\ensuremath{\infty}^1 * (\ensuremath{\partial}_2 \varphi m_1^\ensuremath{\varepsilon}) = \lim_{R \to \ensuremath{\infty}} \pr{\int_{-R - 1}^{-R} + \int_R^{R + 1}} \ensuremath{\partial}_2 \varphi K_\ensuremath{\infty}^1(x - y) m_1^\ensuremath{\varepsilon} \\ &= \lim_{R \to \ensuremath{\infty}} \brac{ (K_1^\ensuremath{\infty} m_1^\ensuremath{\varepsilon})(-R) - (K_1^\ensuremath{\infty} m_1^\ensuremath{\varepsilon})(R) } - \lim_{R \to \ensuremath{\infty}} \pr{\int_{-R - 1}^{-R} + \int_R^{R + 1}} \varphi \ensuremath{\partial}_2 K_\ensuremath{\infty}^1(x - y) m_1^\ensuremath{\varepsilon} \\ &= - \frac{1}{2} \lim_{R \to \ensuremath{\infty}} \brac{m_1^\ensuremath{\varepsilon}(-R) + m_1^\ensuremath{\varepsilon}(R)}. \end{align*} We also used here that $\ensuremath{\partial}_2 m_1^\ensuremath{\varepsilon} = - \eta_\ensuremath{\varepsilon} * \ensuremath{\partial}_2 \innp{\omega} = 0$ for $R \ge R_0$. Since this limit gives the value for all $R \ge R_0$, we can take $\ensuremath{\varepsilon} \to 0$ to conclude that \begin{align*} I_1 &= m_1(x_2) + \frac{1}{2} \brac{m_1(-\ensuremath{\infty}) + m_1(\ensuremath{\infty})}. \end{align*} Returning to \cref{e:BSGS2014}, then, we see that \begin{align}\label{e:vIsKInfConvPlusTerm} {\bm{\mathrm{v}}}(t, {\bm{\mathrm{x}}}) &= \frac{1}{2} \begin{pmatrix} m_1(-\ensuremath{\infty}) + m_1(\ensuremath{\infty}) \\ 2 m_2 \end{pmatrix} + (K_\ensuremath{\infty} * \omega(t)({\bm{\mathrm{x}}}). \end{align} This shows that $m_1(-\ensuremath{\infty}) + m_1(\ensuremath{\infty}) \equiv 0$ and $m_2 \equiv 0$ if and only if ${\bm{\mathrm{v}}} = K_\ensuremath{\infty} * \omega$. \end{proof} \begin{cor}\label{C:BSonPi} Let $\omega \in L^\ensuremath{\infty}_c(\Pi)$. Then ${\bm{\mathrm{v}}} = K_\ensuremath{\infty} * \omega$ is the unique element in $S(\Pi)$ for which $\curl {\bm{\mathrm{v}}} = \omega$, $m_2[{\bm{\mathrm{v}}}] = 0$, and $m_1[{\bm{\mathrm{v}}}](-\ensuremath{\infty}) + m_1[{\bm{\mathrm{v}}}](\ensuremath{\infty}) = 0$. \end{cor} \begin{prop}\label{P:KInfBSIsSameAsType2} Assume that $\omega^0 \in L^\ensuremath{\infty}_c(\Pi)$, ${\bm{\mathrm{v}}}^0 = K_\ensuremath{\infty} * \omega^0$, and ${\bm{\mathrm{v}}}$ is a Type 2 solution as in \cref{T:Type2} with ${\bm{\mathrm{v}}}$ given by \cref{e:BSGS2014}. Then ${\bm{\mathrm{v}}}(t) = K_\ensuremath{\infty} * \omega(t)$ for all $t$. \end{prop} \begin{proof} It follows from \cref{L:KInfm2m1} that $m_1(0, -\ensuremath{\infty}) + m_1(0, \ensuremath{\infty}) = 0$. But as observed following (2.11) of \cite{GallaySlijepcevic2014}, $ \ensuremath{\partial}_t m_1 = - \innp{u^2 \omega}, $ which we note vanishes for all sufficiently large $x_2$ because of the compact support of $\omega$. Hence, $m_1(t, -\ensuremath{\infty}) + m_1(t, \ensuremath{\infty}) = 0$ for all $t$. We conclude from \cref{e:vIsKInfConvPlusTerm} that ${\bm{\mathrm{v}}}(t) = K_\ensuremath{\infty} * \omega(t)$ for all $t$. \end{proof} We used \cref{L:IntK} in the proof of \cref{L:KInfm2m1}, above. \begin{lemma}\label{L:IntK} For all $y \in {\ensuremath{\BB{R}}}$, $K_\ensuremath{\infty}^1(x_1, x_2)$ is even in $x_1$ and odd in $x_2$, while $K_\ensuremath{\infty}^2(x_1, x_2)$ is odd in $x_1$ and even in $x_2$. \end{lemma} \begin{proof} This follows directly from \cref{e:KInfRealForm}, since $\ensuremath{\nabla}^\perp G = K_\ensuremath{\infty}$. \end{proof} \section{Type 3: Solutions with a periodized kernel}\label{S:Type3} \noindent \begin{theorem}\label{T:Type3} Let $\omega^0 \in L^\ensuremath{\infty}_c({\ensuremath{\BB{R}}}^2)$. There exists a solution $\mu$ to \begin{align*} \begin{cases} \ensuremath{\partial}_t \mu + {\bm{\mathrm{w}}} \cdot \ensuremath{\nabla} \mu = 0 &\text{in } [0, \ensuremath{\infty}) \times {\ensuremath{\BB{R}}}^2, \\ {\bm{\mathrm{w}}} = K_\ensuremath{\infty}*\mu &\text{in } [0, \ensuremath{\infty}) \times {\ensuremath{\BB{R}}}^2, \\ \mu(0) = \omega^0 &\text{in } {\ensuremath{\BB{R}}}^2. \end{cases} \end{align*} Moreover, $\curl {\bm{\mathrm{w}}} = \Cal{P}er (\mu)$, and ${\bm{\mathrm{w}}} \in L^\ensuremath{\infty}(0, T; S) \cap C([0, T] \times {\ensuremath{\BB{R}}}^2)$ is the unique solution to \begin{align}\label{e:Type3Vel} \begin{cases} \ensuremath{\partial}_t {\bm{\mathrm{w}}} + {\bm{\mathrm{w}}} \cdot \ensuremath{\nabla} {\bm{\mathrm{w}}} + \ensuremath{\nabla} r = 0 &\text{in } [0, \ensuremath{\infty}) \times {\ensuremath{\BB{R}}}^2, \\ \dv {\bm{\mathrm{w}}} = 0 &\text{in } [0, \ensuremath{\infty}) \times {\ensuremath{\BB{R}}}^2, \\ {\bm{\mathrm{w}}}(0) = K_\ensuremath{\infty} * \mu^0 &\text{in } {\ensuremath{\BB{R}}}^2, \end{cases} \end{align} with the uniqueness criteria being that $r$ is periodic. Finally, $r \in L^\ensuremath{\infty}([0, T] \times {\ensuremath{\BB{R}}}^2)$. \end{theorem} \begin{proof} From \cref{P:SymConv} we know that $K_\ensuremath{\infty} * \omega^0 \in L^\ensuremath{\infty}({\ensuremath{\BB{R}}}^2)$ and is periodic in $x_1$ with period 1; hence, abusing notation, we can set ${\bm{\mathrm{v}}}^0 = K_\ensuremath{\infty} * \omega^0|_\Pi$ and obtain by \cref{T:Type2} a unique solution $({\bm{\mathrm{v}}}, q)$ to \cref{e:EType2} for which $q$ is periodic in $x_1$ and $m_2(t) \equiv 0$. Since $\curl {\bm{\mathrm{v}}}^0 = \omega^0|_\Pi$ is compactly supported and so $\curl {\bm{\mathrm{v}}}$ remains compactly supported for all time, we know from \cref{P:KInfBSIsSameAsType2} that ${\bm{\mathrm{v}}} = K_\ensuremath{\infty}*\curl {\bm{\mathrm{v}}}$. So letting $\zeta = \curl {\bm{\mathrm{v}}}$, we see that \begin{align*} \begin{cases} \ensuremath{\partial}_t \zeta + {\bm{\mathrm{v}}} \cdot \ensuremath{\nabla} \zeta = 0 &\text{in } [0, \ensuremath{\infty}) \times \Pi, \\ {\bm{\mathrm{v}}} = K_\ensuremath{\infty} * \zeta &\text{in } [0, \ensuremath{\infty}) \times \Pi, \\ \zeta(0) = \omega^0 &\text{in } \Pi. \end{cases} \end{align*} Setting ${\bm{\mathrm{w}}} = {\bm{\mathrm{v}}}$, $\mu = \zeta$ gives the desired solution of Type 3. Moreover, since $q(t)$ is periodic, we can let $r = \Cal{R}ep(q)$, and we obtain a unique solution to \cref{e:Type3Vel}. \end{proof} \section{Three types of solution are equivalent}\label{S:ThreeTypes} \noindent For certain classes of initial data, our three types of solution are equivalent. The equivalence of Type 1 and Type 2 holds for a broader class, so we first prove it in \cref{T:TwoSolutions}. The equivalence of the third type holds for initial data in $L^\ensuremath{\infty}_c({\ensuremath{\BB{R}}}^2)$, as we show in \cref{T:ThreeSolutions}. This includes vortex patch data, our application in \cref{S:CDEPeriodic}. \begin{theorem}\label{T:TwoSolutions} Let ${\bm{\mathrm{v}}}^0 \in S(\Pi)$ and periodize it to give $\uu^0 = \Cal{R}ep({\bm{\mathrm{v}}}^0) \in S({\ensuremath{\BB{R}}}^2)$. Let $(\uu, p)$ be the solution of Type 1 with initial velocity $\uu^0$ given by \cref{T:Type1} and let $({\bm{\mathrm{v}}}, q)$ the solution of Type 2 with initial velocity ${\bm{\mathrm{v}}}^0$ given by \cref{T:Type2}. Then $\Cal{R}ep ({\bm{\mathrm{v}}}) = \uu$. \end{theorem} \begin{proof} We have $\curl {\bm{\mathrm{v}}}(0) = \curl \uu^0|_\Pi$, where we abuse notation somewhat. From \cref{T:Type2}, we have a pressure $q$ with $q(t) \in L^\ensuremath{\infty}(\Pi)$ for which \begin{align}\label{e:vPeriodicVelSol} \begin{cases} \ensuremath{\partial}_t {\bm{\mathrm{v}}} + {\bm{\mathrm{v}}} \cdot \ensuremath{\nabla} {\bm{\mathrm{v}}} + \ensuremath{\nabla} q = 0 &\text{in } [0, \ensuremath{\infty}) \times \Pi, \\ \dv {\bm{\mathrm{v}}} = 0 &\text{in } [0, \ensuremath{\infty}) \times \Pi, \\ {\bm{\mathrm{v}}}(0) = {\bm{\mathrm{v}}}^0 &\text{in } \Pi. \end{cases} \end{align} Since $\Cal{R}ep({\bm{\mathrm{v}}})$ and $\Cal{R}ep(q)$ are $x_1$-periodic with period 1, we can set $\widetilde{\vv} = \Cal{R}ep ({\bm{\mathrm{v}}})$ and $\widetilde{q} = \Cal{R}ep (q)$, and both will lie in $L^\ensuremath{\infty}([0, T] \times {\ensuremath{\BB{R}}}^2)$ with $\curl \widetilde{\vv}(t) = \Cal{R}ep (\curl {\bm{\mathrm{v}}}(t))$. Thus, $\widetilde{\vv}$ is ${\bm{\mathrm{v}}}$ periodized and $\curl \widetilde{\vv}$ is $\curl {\bm{\mathrm{v}}}$ periodized, meaning that \cref{e:vPeriodicVelSol} in effect holds on $\Pi_p$ translated by $(n, 0)$ for any integer $n$, so we see that \begin{align}\label{e:uR2VelSol} \begin{cases} \ensuremath{\partial}_t \widetilde{\vv} + \widetilde{\vv} \cdot \ensuremath{\nabla} \widetilde{\vv} + \ensuremath{\nabla} \widetilde{q} = 0 &\text{in } [0, \ensuremath{\infty}) \times {\ensuremath{\BB{R}}}^2, \\ \dv \widetilde{\vv} = 0 &\text{in } [0, \ensuremath{\infty}) \times {\ensuremath{\BB{R}}}^2, \\ \widetilde{\vv}(0) = \uu^0 &\text{ in } {\ensuremath{\BB{R}}}^2. \end{cases} \end{align} We see that $(\widetilde{\vv}, \widetilde{q})$ is a solution to the Euler equations on $[0, \ensuremath{\infty}) \times {\ensuremath{\BB{R}}}^2$. Manifestly, $\widetilde{\vv}$, $\curl \widetilde{\vv}$, and $\widetilde{q}$ each lie in $L^\ensuremath{\infty}([0, \ensuremath{\infty}) \times {\ensuremath{\BB{R}}}^2)$, being periodic in $x_1$. Hence, $\widetilde{\vv}$ is a bounded velocity, bounded vorticity solution to the Euler equations on $[0, \ensuremath{\infty}) \times {\ensuremath{\BB{R}}}^2$. Because the pressure $\widetilde{q}$ grows sublinearly it is, in fact, the (unique) Serfati solution (it satisfies the Serfati identity), as follows from \cref{T:Type1}. Therefore, $\uu = {\bm{\mathrm{v}}}$. \end{proof} \begin{theorem}\label{T:ThreeSolutions} For $\omega^0 \in L^\ensuremath{\infty}_c({\ensuremath{\BB{R}}}^2)$, let $\uu^0 = K_{sym}[\omega^0]$ be obtained by symmetrizing in pairs as in \cref{D:SymVel}, and let ${\bm{\mathrm{v}}}^0 = K_\ensuremath{\infty} * \Cal{P}er(\omega^0)$. Let $(\uu, p)$, $({\bm{\mathrm{v}}}, q)$ be the Type 1, 2 solutions with initial velocity $\uu^0$, ${\bm{\mathrm{v}}}^0$ and let ${\bm{\mathrm{w}}}^0$ be the velocity field for the Type 3 solution given by \cref{T:Type3}. Then $\Cal{R}ep ({\bm{\mathrm{v}}}) = \uu = {\bm{\mathrm{w}}}$. \end{theorem} \begin{proof} \cref{T:TwoSolutions} gives $\Cal{R}ep ({\bm{\mathrm{v}}}) = \uu$, while $\Cal{R}ep ({\bm{\mathrm{v}}}) = {\bm{\mathrm{w}}}$ is inherent in the proof of \cref{T:Type3}. \end{proof} \section{The velocity gradient}\label{S:VelocityGradient} \noindent The following expression for $\ensuremath{\nabla} (K * \omega)$ is classical (see, for instance, Proposition 2.20 of \cite{MB2002}): \begin{lemma}\label{L:graduClassical} Assume that $\omega \in L^\ensuremath{\infty}({\ensuremath{\BB{R}}}^2)$ is compactly supported and let $\uu = K * \omega$. Then \begin{align*} \ensuremath{\nabla} \uu({\bm{\mathrm{x}}}) &= \omega({\bm{\mathrm{x}}}) \begin{pmatrix} 0 & -1 \\ 1 & 0 \end{pmatrix} + \PV \int_{{\ensuremath{\BB{R}}}^2} \ensuremath{\nabla} K({\bm{\mathrm{x}}} - {\bm{\mathrm{y}}}) \omega({\bm{\mathrm{y}}}) \, d {\bm{\mathrm{y}}}, \end{align*} where we can write, \begin{align*} \ensuremath{\nabla} K({\bm{\mathrm{x}}}) &= \frac{1}{2 \pi} \frac{\sigma({\bm{\mathrm{x}}})}{\abs{{\bm{\mathrm{x}}}}^2}, \quad \sigma({\bm{\mathrm{x}}}) := \frac{1}{\abs{{\bm{\mathrm{x}}}}^2} \begin{pmatrix} 2 x_1 x_2 & x_2^2 - x_1^2 \\ x_2^2 - x_1^2 & - 2 x_1 x_2 \end{pmatrix}. \end{align*} \end{lemma} The analog for the $K_\ensuremath{\infty}$ kernel is \cref{L:graduTypeI}. \begin{lemma}\label{L:graduTypeI} Assume that $\omega \in L^\ensuremath{\infty}({\ensuremath{\BB{R}}}^2)$ is compactly supported and let $\uu = K_\ensuremath{\infty} * \omega$. Then \begin{align*} \ensuremath{\nabla} \uu({\bm{\mathrm{x}}}) &= \sum_{n \in \ensuremath{\BB{Z}}} \frac{\omega({\bm{\mathrm{x}}} + (n, 0))}{2} \begin{pmatrix} 0 & -1 \\ 1 & 0 \end{pmatrix} + \PV \int_{{\ensuremath{\BB{R}}}^2} \ensuremath{\nabla} K_\ensuremath{\infty}({\bm{\mathrm{x}}} - {\bm{\mathrm{y}}}) \omega({\bm{\mathrm{y}}}) \, d {\bm{\mathrm{y}}}, \end{align*} where $\rho$ is as in \cref{e:rho} and where we can write, \begin{align*} \ensuremath{\nabla} K_\ensuremath{\infty}({\bm{\mathrm{x}}}) &= \frac{\pi}{2} \frac{\beta({\bm{\mathrm{x}}})}{\rho({\bm{\mathrm{x}}})^2}, \end{align*} where \begin{align*} \beta({\bm{\mathrm{x}}}) &= \frac{1}{2 \rho({\bm{\mathrm{x}}})^2} \begin{pmatrix} \sin(2 \pi x_1) \sinh(2 \pi x_2) & \cos(2 \pi x_1) \cosh(2 \pi x_2) - 1 \\ \cos(2 \pi x_1) \cosh(2 \pi x_2) - 1 & - \sin(2 \pi x_1) \sinh(2 \pi x_2) \end{pmatrix}. \end{align*} \end{lemma} \begin{remark}\label{R:gradKInfSize} Like $\sigma$, the matrix $\beta$ is symmetric with trace zero. Near the origin, $\rho({\bm{\mathrm{x}}})^2 \approx \pi^2 \abs{{\bm{\mathrm{x}}}}^2$, and we can see that $\beta({\bm{\mathrm{x}}}) \approx 4 \pi^2 \abs{{\bm{\mathrm{x}}}}^2 / (2 \pi^2 \abs{{\bm{\mathrm{x}}}}^2) \approx 2 \approx \sigma({\bm{\mathrm{x}}})$, and so $\ensuremath{\nabla} K_\ensuremath{\infty}({\bm{\mathrm{x}}}) \approx 2 \pi / (2 \pi^2 \abs{{\bm{\mathrm{x}}}}^2) \approx 1/(\pi \abs{{\bm{\mathrm{x}}}}^2) \approx \ensuremath{\nabla} K({\bm{\mathrm{x}}})$. Also like $\sigma$, $\beta_{11}$ and $\beta_{22}$ integrate to zero over circles centered at the origin, but unlike $\sigma$, neither $\beta_{12}$ nor $\beta_{21}$ integrate to zero. \end{remark} We have the following immediate corollary of \cref{L:graduTypeI}: \begin{cor}\label{C:graduType} Let ${\bm{\mathrm{v}}} \in S(\Pi)$ with $\omega = \curl {\bm{\mathrm{v}}}$ compactly supported and let $\uu = K_\ensuremath{\infty} * \omega$. Then \begin{align*} \ensuremath{\nabla} \uu({\bm{\mathrm{x}}}) &= \frac{\omega({\bm{\mathrm{x}}})}{2} \begin{pmatrix} 0 & -1 \\ 1 & 0 \end{pmatrix} + \PV \int_\Pi \ensuremath{\nabla} K_\ensuremath{\infty}({\bm{\mathrm{x}}} - {\bm{\mathrm{y}}}) \omega({\bm{\mathrm{y}}}) \, d {\bm{\mathrm{y}}} \end{align*} and $\ensuremath{\nabla} K_\ensuremath{\infty}$ can be written as in \cref{L:graduTypeI}. \end{cor} \section{Contour Dynamics Equations}\label{S:CDEPeriodic} \noindent First we review the Contour Dynamics Equation (CDE) for a classical vortex patch---the characteristic function of a bounded, simply connected domain evolving under the vorticity equation for the Euler equations on all of ${\ensuremath{\BB{R}}}^2$---then turn to the CDE for Type 2 solutions. In what follows we use the Lipschitz space $Lip$ and homogeneous Lipschitz space $lip$. On $U \subseteq {\ensuremath{\BB{R}}}^d$ for $d \ge 1$, we define their semi-norm and norm, \begin{align*} \norm{f}_{lip(U)} := \sup_{x \ne y \in U} \frac{\abs{f(x) - f(y)}}{\abs{x - y}}, \quad \norm{f}_{Lip(U)} := \norm{f}_{L^\ensuremath{\infty}(U)} + \norm{f}_{lip(U)}. \end{align*} \subsection{Classical vortex patches}\label{S:ClassicalPatches} In the classical setting of a vortex patch in ${\ensuremath{\BB{R}}}^2$, we have \cref{T:ClassicalPatchBSLaw,T:ClassicalPatchCDE}, as in Proposition 8.6 of \cite{MB2002} and the derivation of the classical CDE that appears before it. \begin{theorem}\label{T:ClassicalPatchBSLaw} Let $\bgamma \colon [0, 2 \pi] \to {\ensuremath{\BB{R}}}^2$ be a $C^1$ counterclockwise\footnote{In \cite{MB2002}, the patch boundary is parameterized clockwise, but $(\BoldTau, {\bm{n}})$ is in the standard $(\bm{\mathrm{e}}_1, \bm{\mathrm{e}}_2)$ orientation; the two resulting sign changes between \cite{MB2002} and us cancel, so there is no sign change in our expressions.} parameterization of the boundary of a bounded, simply connected domain $\Omega$. Then \begin{align}\label{e:BSLawPatchR2} \uu({\bm{\mathrm{x}}}) = -\frac{\omega_0}{2 \pi} \int_0^{2 \pi} \log \abs{{\bm{\mathrm{x}}} - \bgamma(\ensuremath{\alpha})} \ensuremath{\partial}_\ensuremath{\alpha} \bgamma(\ensuremath{\alpha}) \, d \ensuremath{\alpha} \end{align} is the unique divergence-free vector field decaying at infinity for which $\curl \uu = \omega_0 \CharFunc_\Omega$. \end{theorem} Now let us suppose that $\Omega$ is a simply connected bounded domain in ${\ensuremath{\BB{R}}}^2$ with a $C^{1, \ensuremath{\varepsilon}}$ boundary. Let $\uu$ be the unique weak solution to the Euler equations with initial vorticity $\omega^0 := \omega_0 \CharFunc_\Omega$ and let $X$ be the flow map for $\uu$. Then we know that the vorticity $\omega(t) = \omega_0 \CharFunc_{\Omega_t}$, where $\Omega_t = X(t, \Omega)$. Let $\bgamma(0, \cdot)$ be a $C^1$-regular counterclockwise parameterization of $\Gamma = \ensuremath{\partial} \Omega$. Define a parameterization of $\ensuremath{\partial} \Omega_t = X(t, \Gamma)$ by $\bgamma(t, \cdot) := X(t, \bgamma(0, \cdot))$. The log-Lipschitz regularity of $\uu(t)$ induces $C^{c(t)}$-regularity of the flow map $X(t, \cdot)$ with $c(t) \in (0, 1)$ and decreasing with time, as in Lemma 8.2 of \cite{MB2002}. This is insufficient regularity to obtain a $C^1$-parameterization of $\ensuremath{\partial} \Omega_t$, so let us \textit{suppose} that our (classical) solution has $\uu \in C(0, T; lip)$. Then $\bgamma(t, \cdot)$ is a $C^1$-parameterization of $\ensuremath{\partial} \Omega_t$. Since we assumed $\ensuremath{\partial} \Omega$ is $C^{1, \ensuremath{\varepsilon}}$, we could give $\bgamma(0, \cdot)$ $C^{1, \ensuremath{\varepsilon}}$-regularity, but this does not itself ensure that $\bgamma(t, \cdot)$ is $C^{1, \ensuremath{\varepsilon}}$: proving that is tantamount to establishing the propagation of regularity of the vortex patch boundary. \begin{theorem}\label{T:ClassicalPatchCDE} Let $\uu(t, {\bm{\mathrm{x}}})$ be given by \cref{e:BSLawPatchR2} applied with $\bgamma(t, \cdot)$; that is, \begin{align*} \uu(t, {\bm{\mathrm{x}}}) := -\frac{\omega_0}{2 \pi} \int_0^{2 \pi} \log \abs{{\bm{\mathrm{x}}} - \bgamma(t, \ensuremath{\alpha})} \ensuremath{\partial}_\ensuremath{\alpha} \bgamma(t, \ensuremath{\alpha}) \, d \ensuremath{\alpha}. \end{align*} Then $\uu$ is a weak solution to the 2D Euler equations on $[0, T] \times {\ensuremath{\BB{R}}}^2$ with $\uu \in C(0, T; Lip)$ if and only if $\bgamma$ is a $C^1([-T, T]; C([0, 2 \pi])) \cap C([-T, T]; C^1([0, 2 \pi]))$ solution to the contour dynamics equations (CDE), \begin{align}\label{e:CDEClassical} \diff{}{t} \bgamma(t, \ensuremath{\alpha}) = -\frac{\omega_0}{2 \pi} \int_0^{2 \pi} \log \abs{\bgamma(t, \ensuremath{\alpha}) - \bgamma(t, \ensuremath{\alpha}')} \ensuremath{\partial}_{\ensuremath{\alpha}'} \bgamma(t, \ensuremath{\alpha}') \, d \ensuremath{\alpha}. \end{align} \end{theorem} \cref{T:ClassicalPatchBSLaw,T:ClassicalPatchCDE} were expressed for simply connected domains. As pointed out on page 330 of \cite{MB2002}, the only difference for multiply connected domains is that the integrals in \cref{e:BSLawPatchR2,e:CDEClassical} are summed over each component of the boundary. \begin{theorem}\label{T:ClassicalMultiplyConnected} \cref{T:ClassicalPatchBSLaw,T:ClassicalPatchCDE} hold for bounded, multiply connected domains if we evaluate and sum each of the boundary integrals over each boundary component. \end{theorem} We view \cref{e:CDEClassical} as a form of the Euler equations applying specifically to a vortex patch: it comes directly from \cref{e:BSLawPatchR2}, which we view as a form of the Biot-Savart law that recovers the velocity from the vorticity, as it is encoded by $\bgamma$. We work, now, to obtain replacements for these expressions that apply to periodized vortex patches. This is a matter of deriving the CDE for a solution to the Euler equations and showing, conversely, that any solution to the CDE satisfies the Euler equations. \subsection{Type 2 solutions} Turning to Type 2 solutions, we make the following assumptions on $\Omega$: \begin{assumption}\label{A:OmegaPi} Assume that $\Omega \subseteq \Pi$ is bounded with a finite number of boundary components, $\Gamma_1, \dots, \Gamma_J$, each $C^{1, \ensuremath{\varepsilon}}$ regular. \end{assumption} With $\Omega$ as in \cref{A:OmegaPi}, we let $\uu$ be the unique Type 2 solution having initial vorticity $\omega^0 := \omega_0 \CharFunc_\Omega$ with $m_2 \equiv m_1(t, -\ensuremath{\infty}) + m_2(t, \ensuremath{\infty}) \equiv 0$ given by \cref{T:Type2,P:KInfBSIsSameAsType2} ($m_1$, $m_2$ are defined in \cref{S:MeanHorizontal}). Set \begin{align*} \Omega_t := X(t, \Omega), \quad \Gamma_{t, j} := X(t, \Gamma_j), \end{align*} noting that because $X(t, \cdot)$ is a homeomorphism of ${\ensuremath{\BB{R}}}^2$ onto ${\ensuremath{\BB{R}}}^2$, $\Gamma_{t, j}$ is the $j^{th}$ of the $J$ components of $\ensuremath{\partial} \Omega_t$. We then define a parameterization $\bgamma_j$ of $\Gamma_{t, j}$ as we parameterized $\ensuremath{\partial} \Omega_t$ in \cref{S:ClassicalPatches}, setting $\bgamma_j(t, \cdot) := X(t, \bgamma_j(0, \cdot))$. As in that section, a priori, we do not even know that $\bgamma_j(t)$ has $C^1$ regularity for $t > 0$; proving that it has $C^{1, \ensuremath{\varepsilon}}$ regularity is the ultimate goal (of \cref{S:BoundaryRegularity}). We show in \cref{T:Type2PatchBSLaw,T:CDEType2} that the analog of \cref{T:ClassicalMultiplyConnected} holds for Type 2 solutions. \begin{theorem}\label{T:Type2PatchBSLaw} Let $\Omega$ be as in \cref{A:OmegaPi}, and for each $j$, let $\bgamma_j \colon [0, 2 \pi] \to {\ensuremath{\BB{R}}}^2$ be a $C^1$ counterclockwise parameterization of the boundary component $\Gamma_j$. With $\rho$ as in \cref{e:rho}, \begin{align}\label{e:BSlogsin} \begin{split} \uu({\bm{\mathrm{x}}}) &= -\frac{\omega_0}{2 \pi} \sum_{j = 1}^J \int_0^{2 \pi} \log \rho ( {\bm{\mathrm{x}}} - \bgamma_j(\ensuremath{\alpha}) ) \ensuremath{\partial}_\ensuremath{\alpha} \bgamma_j(\ensuremath{\alpha}) \, d \ensuremath{\alpha} \end{split} \end{align} is the unique divergence-free vector field in $S(\Pi)$ having curl equal to $\omega_0 \CharFunc_\Omega$ for which $m_2 = 0$ and $m_1(-\ensuremath{\infty}) + m_1(\ensuremath{\infty}) = 0$. \end{theorem} \begin{proof} By \cref{C:BSonPi}, we know that $\uu = K_\ensuremath{\infty} * \omega$ is the unique divergence-free vector field in $S(\Pi)$ having curl equal to $\omega_0 \CharFunc_\Omega$ for which $m_2 = 0$ and $m_1(-\ensuremath{\infty}) + m_1(\ensuremath{\infty}) = 0$. Then we have, using \cref{L:KInfgradperG} and parameterizing $\Gamma_{t, j}$ by arc length from 0 to $\ell_j$, setting ${\bm{\mathrm{y}}}(s) = \bgamma_j(\ensuremath{\alpha}(s))$, \begin{align*} \uu({\bm{\mathrm{x}}}) &= K_\ensuremath{\infty} * \omega({\bm{\mathrm{x}}}) = \ensuremath{\nabla}^\perp G * \omega({\bm{\mathrm{x}}}) = \frac{\omega_0}{2 \pi} \int_\Omega \ensuremath{\nabla}^\perp \log \rho({\bm{\mathrm{x}}} - {\bm{\mathrm{y}}}) \, d {\bm{\mathrm{y}}} \\ &= -\frac{\omega_0}{2 \pi} \int_\Omega \ensuremath{\nabla}^\perp_{\bm{\mathrm{y}}} \log \rho({\bm{\mathrm{x}}} - {\bm{\mathrm{y}}}) \, d {\bm{\mathrm{y}}} = -\frac{\omega_0}{2 \pi} \sum_{j = 1}^J \int_0^{\ell_j} \log \rho({\bm{\mathrm{x}}} - {\bm{\mathrm{y}}}(s)) (-n^2, n^1) \, d s \\ &= -\frac{\omega_0}{2 \pi} \sum_{j = 1}^J \int_0^{\ell_j} \log \rho({\bm{\mathrm{x}}} - {\bm{\mathrm{y}}}(s)) \BoldTau(s) \, d s = -\frac{\omega_0}{2 \pi} \sum_{j = 1}^J \int_0^{2 \pi} \log \rho ( {\bm{\mathrm{x}}} - \bgamma_j(\ensuremath{\alpha}) ) \ensuremath{\partial}_\ensuremath{\alpha} \bgamma_j(\ensuremath{\alpha}) \, d \ensuremath{\alpha}. \end{align*} Here $(n^1, n^2) = {\bm{n}}$ and $(-n^2, n^1) = \BoldTau$ (see \cref{L:ComplexToRealContourIntegrals}), and we used that \begin{align*} \ensuremath{\partial}_\ensuremath{\alpha} \bgamma_j(\ensuremath{\alpha}) \, d \ensuremath{\alpha} &= \frac{\ensuremath{\partial}_\ensuremath{\alpha} \bgamma_j(\ensuremath{\alpha})}{\abs{\ensuremath{\partial}_\ensuremath{\alpha} \bgamma_j(\ensuremath{\alpha})}} \abs{\ensuremath{\partial}_\ensuremath{\alpha} \bgamma_j(\ensuremath{\alpha})} \, d \ensuremath{\alpha} = \BoldTau(s) \, ds. \end{align*} From this, \cref{e:BSlogsin} follows. \end{proof} \begin{theorem}\label{T:CDEType2} Let $\uu$ be the Type 2 solution described above and assume that each $\bgamma_j$ is in $C^1([-T, T]; C([0, 2 \pi])) \cap C([-T, T]; C^1([0, 2 \pi]))$. Then \begin{align}\label{e:BSlogsintime} \uu(t, {\bm{\mathrm{x}}}) &= -\frac{\omega_0}{2 \pi} \sum_{j = 1}^J \int_0^{2 \pi} \log \rho ( {\bm{\mathrm{x}}} - \bgamma_j(t, \ensuremath{\alpha}) ) \ensuremath{\partial}_\ensuremath{\alpha} \bgamma_j(t, \ensuremath{\alpha}) \, d \ensuremath{\alpha} \end{align} and lies in $C(0, T; Lip)$. Moreover, each $\bgamma_k$ satisfies the CDE, \begin{align}\label{e:CDElogsin} \begin{split} \diff{}{t} \bgamma_k(t, \ensuremath{\alpha}) &= -\frac{\omega_0}{2 \pi} \sum_{j = 1}^J \int_0^{2 \pi} \log \rho ( \bgamma_k(t, \ensuremath{\alpha}) - \bgamma_j(t, \ensuremath{\alpha}') ) \ensuremath{\partial}_\ensuremath{\alpha} \bgamma_j(t, \ensuremath{\alpha}) \, d \ensuremath{\alpha}'. \end{split} \end{align} Conversely, if each $\bgamma_k$ in $C^1([-T, T]; C([0, 2 \pi]))$ $\cap$ $C([-T, T]; C^1([0, 2 \pi]))$ satisfies \cref{e:CDElogsin} then $\uu$ given by \cref{e:BSlogsintime} is a Type 2 solution with $\uu \in C(0, T; Lip)$ and $m_2 \equiv m_1(t, -\ensuremath{\infty}) + m_2(t, \ensuremath{\infty}) \equiv 0$. \end{theorem} \begin{proof} The forward direction follows directly from \cref{T:Type2PatchBSLaw}. For the converse, we parallel the proof of Proposition 8.6 of \cite{MB2002}, which consists of two steps: (1) Show that $\uu$ given by \cref{e:BSlogsintime} is divergence-free with $\curl \uu = \CharFunc \Omega_{0, t}$. (2) Show that $\uu$ solves the 2D Euler equations. To prove (1), let $\uu$ be given by \cref{e:BSlogsintime}. Reparameterizing by arc length as in the proof of \cref{T:Type2PatchBSLaw}, \begin{align*} \uu(t, {\bm{\mathrm{x}}}) &= -\frac{\omega_0}{2 \pi} \sum_{j = 1}^J \int_0^{\ell_j} \log \rho ({\bm{\mathrm{x}}} - {\bm{\mathrm{y}}}(s)) \BoldTau(s) ds = -\frac{\omega_0}{2 \pi} \sum_{j = 1}^J \int_{\Gamma_{t, j}} \log \rho ({\bm{\mathrm{x}}} - \cdot) \BoldTau. \end{align*} To apply $\dv$ and $\curl$ to this expression, we use that for a constant vector field ${\bm{\mathrm{w}}}$ and scalar function $g$, $\dv (g {\bm{\mathrm{w}}}) = \ensuremath{\nabla} g \cdot {\bm{\mathrm{w}}}$ and $\curl (g {\bm{\mathrm{w}}}) = \ensuremath{\nabla}^\perp g \cdot {\bm{\mathrm{w}}}$. Also, letting ${\bm{\mathrm{v}}} = (2 \pi)^{-1} \ensuremath{\nabla}^\perp \log \rho ({\bm{\mathrm{x}}} - \cdot)$ and $f = \overline{\overleftarrow{{\bm{\mathrm{v}}}}}$, we see that \begin{align*} \curl \uu(t, {\bm{\mathrm{x}}}) &= -\frac{\omega_0}{2 \pi} \sum_{j = 1}^J \int_{\Gamma_{t, j}} \ensuremath{\nabla}^\perp \log \rho ({\bm{\mathrm{x}}} - \cdot) \cdot \BoldTau = -\omega_0 \sum_{j = 1}^J \int_{\Gamma_{t, j}} \vec{\overline{f}} \cdot \BoldTau = -\omega_0 \int_{\ensuremath{\partial} \Omega_t} \vec{\overline{f}} \cdot \BoldTau, \\ \dv \uu(t, {\bm{\mathrm{x}}}) &= -\frac{\omega_0}{2 \pi} \sum_{j = 1}^J \int_{\Gamma_{t, j}} \ensuremath{\nabla} \log \rho ({\bm{\mathrm{x}}} - \cdot) \cdot \BoldTau = \frac{\omega_0}{2 \pi} \sum_{j = 1}^J \int_{\Gamma_{t, j}} \ensuremath{\nabla}^\perp \log \rho ({\bm{\mathrm{x}}} - \cdot) \cdot {\bm{n}} \\ &= \omega_0 \sum_{j = 1}^J \int_{\Gamma_{t, j}} \vec{\overline{f}} \cdot n = \omega_0 \int_{\ensuremath{\partial} \Omega_t} \vec{\overline{f}} \cdot n. \end{align*} Up to this point, we have been integrating over paths in $\Pi$ treated as ${\ensuremath{\BB{R}}}^2 / \Cal{L}$, but we wish to apply \cref{L:ComplexToRealContourIntegrals}, which obliges us to work in $\C$. To do this, we lift $\Omega_t$ to $\widetilde{\Omega}_t$ as described in \cref{S:Lifting}. Applying \cref{L:LiftedBoundaryIntegral,L:ComplexToRealContourIntegrals} (writing $f$ in place of $f \circ p$ by viewing $f$ as $x_1$-periodic with period $1$) gives for all ${\bm{\mathrm{x}}}$ not lying on $\ensuremath{\partial} \widetilde{\Omega}_t$ (a set of measure $0$), \begin{align*} \omega_0 \, \Xint{\C}_{\ensuremath{\partial} \Omega_t} f &= \omega_0 \, \Xint{\C}_{\ensuremath{\partial} \widetilde{\Omega}_t} f = \omega_0 \int_{\ensuremath{\partial} \widetilde{\Omega}_t} \vec{\overline{f}} \cdot \BoldTau + i \omega_0 \int_{\ensuremath{\partial} \widetilde{\Omega}_t} \vec{\overline{f}} \cdot n = -\curl \uu(t, {\bm{\mathrm{x}}}) + i \dv \uu(t, {\bm{\mathrm{x}}}). \end{align*} But we see from \cref{L:KInfgradperG} that ${\bm{\mathrm{v}}} = K_\ensuremath{\infty}({\bm{\mathrm{x}}} - \cdot)$ and that \begin{align*} f &= \frac{1}{2} \overline{\overleftarrow{\overrightarrow{\cot(\pi \overline{z})}^\perp}} = \frac{1}{2} \overline{\overleftarrow{\overrightarrow{i \overline{\cot(\pi z)}}}} = \frac{1}{2} \overline{i \overline{\cot(\pi z)}} = -\frac{i}{2} \cot(\pi z), \end{align*} where we used \cref{e:vecProp} and the identity $\overline{i \overline{z}} = - i z$. The complex meromorphic function $f$ has simple poles at each point in ${\bm{\mathrm{x}}} + \Cal{L}$ with residue $(-2 \pi)^{-1} i$. By the residue theorem, then, summing over all points of $\Cal{L}$ lying inside $\ensuremath{\partial} \widetilde{\Omega}_t$---that is, lying in $\widetilde{\Omega}_t$, \begin{align*} \omega_0 \, \Xint{\C}_{\ensuremath{\partial} \widetilde{\Omega}_t} f &= \RE \pr{ 2 \pi i \omega_0 \sum_n \Res (f, (n, 0)) } = \omega_0 \RE \pr{ \frac{2 \pi i}{-2 \pi i} \sum_n 1 } = -\omega_0 n. \end{align*} But $\widetilde{\Omega}_t$ can contain at most one point of ${\bm{\mathrm{x}}} + \Cal{L}$ else the lift given in \cref{S:Lifting} would map ${\bm{\mathrm{x}}}$ to more than one point in $\C$ (which would mean it is not a lift). We see, then, that \begin{align*} \curl \uu(t, {\bm{\mathrm{x}}}) &= -\omega_0 \, \Xint{\C}_{\ensuremath{\partial} \Omega_t} f = -\omega_0 \, \Xint{\C}_{\ensuremath{\partial} \widetilde{\Omega}_t} f = \omega_0 \CharFunc_{\Omega_t}(t, {\bm{\mathrm{x}}}) = \omega(t, {\bm{\mathrm{x}}}). \end{align*} We conclude that for all $t \in [0, T]$, $\dv \uu = 0$ and $\curl \uu = \omega = \omega_0 \CharFunc_{\Omega_t}$. Directly from \cref{e:BSlogsintime}, we know that $\uu \in L^\ensuremath{\infty}(\Pi)$ and hence $\uu \in S(\Pi)$. It follows from \cref{T:Type2PatchBSLaw} applied with $\gamma_j(t, \cdot)$ in place of $\gamma_j$ for any fixed $t$ that $m_2[\uu(t)] = 0$ and $m_1[\uu(t)](-\ensuremath{\infty}) + m_1[\uu](\ensuremath{\infty}) = 0$. Using (1), the proof of (2) that $\uu$ solves the 2D Euler equations on the time interval $[-T, T]$ proceeds just as it does in the proof of Proposition 8.6 on page 334 of \cite{MB2002}. \end{proof} \begin{remark}\label{R:Patch1and3} We can view Type 2 solutions as equivalent to Type 1 or 3 solutions by virtue of \cref{T:ThreeSolutions}. For vortex patches it is most natural to start with an $\Omega \in \Pi$ satisfying \cref{A:OmegaPi} and lift it to ${\ensuremath{\BB{R}}}^2$ as in \cref{S:Lifting} to give $\Omega_0$. It is also possible to start with a domain in ${\ensuremath{\BB{R}}}^2$, and use it to obtain via the $\Cal{P}er$ operator a domain in $\Pi$, but there are no simple general conditions to guarantee that the boundary of the domain in $\Pi$ is regular. \end{remark} \section{Regularity of a vortex patch boundary}\label{S:BoundaryRegularity} \noindent To prove the propagation of regularity of a vortex patch boundary for our Type 1, 2, or 3 solutions, it will be easiest to work with Type 2 solutions, the result then immediately following for the other two types by \cref{T:ThreeSolutions}. We will prove, in \cref{T:PatchType2}, that for Type 2 solutions, the regularity of the boundary of a periodic vortex patch is maintained for all time, as in the classical case. \begin{theorem}\label{T:PatchType2} Let $\Omega$ be as in \cref{A:OmegaPi} and let $\Omega_t = X(t, \Omega)$ for a Type 2 solution. Then $\ensuremath{\partial} \Omega_t$ is $C^{1, \ensuremath{\varepsilon}}$ for all time. The analogous result holds for Type 1 and 3 solutions. \end{theorem} \begin{proof} We describe only how the proof differs from the now classical proof as presented in Chapter 8 of \cite{MB2002}. There are two main steps to the proof given in \cite{MB2002}: First, show local-in-time existence of a $C^{1, \ensuremath{\varepsilon}}$ solution to the CDE (based on \cite{BertozziThesis1991}) then show that the solution extends globally in time (based on \cite{ConstantinBertozzi1993}). \bigskip\noindent\textbf{Local-in-time $C^{1, \ensuremath{\varepsilon}}$ solutions}: In brief, the first step is to define the function $F$ on the space $B^{1, \ensuremath{\varepsilon}}$ of closed $C^{1, \ensuremath{\varepsilon}}$ paths in $\Pi$ by (we have translated this to Type 2 solutions) by \begin{align*} F(\bgamma(\beta)) := \frac{\omega_0}{2 \pi} \int_0^{2 \pi} \log \rho ( \bgamma(\beta) - \bgamma(\ensuremath{\alpha}) ) \ensuremath{\partial}_\ensuremath{\alpha} \bgamma(\ensuremath{\alpha}) \, d \ensuremath{\alpha}. \end{align*} Here, $F$ is as defined for each boundary component separately, we suppress the sums over each boundary component for notational simplicity. First show that $F \colon \Cal{O}^M \to B^{1, \ensuremath{\varepsilon}}$ is Lipschitz-continuous on the open subset \begin{align*} \Cal{O}^M &:= \set {\bgamma \in B^{1, \ensuremath{\varepsilon}} \colon \abs{\bgamma}_* > M^{-1}, \, \norm{\bgamma'}_{L^\ensuremath{\infty}} < M}, \\ \abs{\bgamma}_* &:= \inf_{\ensuremath{\alpha} \ne \ensuremath{\alpha}'} \frac{\bgamma(\ensuremath{\alpha}) - \bgamma(\ensuremath{\alpha}')}{\abs{\ensuremath{\alpha} - \ensuremath{\alpha}'}} \end{align*} for some $M > 0$. A Picard fixed point theorem (Theorem 8.3 of \cite{MB2002}) then assures a local-in-time solution to the ODE, \begin{align*} \diff{\bgamma}{t} = F(\bgamma), \quad \bgamma(0) = \bgamma_0 \in \Cal{O}^M, \end{align*} with $\bgamma \in C^1([-T, T]; \Cal{O}^M)$ for a $T$ that depends upon $M$. To adapt the argument in \cite{MB2002} to Type 2 solutions, we decompose $\log \rho({\bm{\mathrm{x}}})$ as follows. Let $\varphi \in C_0^\ensuremath{\infty}(\Pi)$ be a radially symmetric cutoff function supported on $B_{1/4}(0)$ with $\varphi \equiv 1$ on $B_{1/8}(0)$. Then \begin{align*} \log \rho({\bm{\mathrm{x}}}) &= \varphi({\bm{\mathrm{x}}}) \log \abs{{\bm{\mathrm{x}}}} + R({\bm{\mathrm{x}}}), \\ R({\bm{\mathrm{x}}}) &:= \varphi({\bm{\mathrm{x}}}) \brac{\log \rho({\bm{\mathrm{x}}}) - \log \abs{{\bm{\mathrm{x}}}}} + (1 - \varphi({\bm{\mathrm{x}}})) \log \rho({\bm{\mathrm{x}}}). \end{align*} Recall that on $\Pi$, we use coordinates in which ${\bm{\mathrm{x}}} = (x_1, x_2)$ with $-1/2 \le x_1 < 1/2$. Because $\varphi({\bm{\mathrm{x}}}) = 0$ for $\abs{x_1} > 1/4$, the function $\varphi({\bm{\mathrm{x}}}) \log \abs{{\bm{\mathrm{x}}}}$ is in $C^\ensuremath{\infty}(\Pi \setminus (0, 0))$. Also, $\log \rho({\bm{\mathrm{x}}})$ is harmonic away from the origin, so $R({\bm{\mathrm{x}}}) \in C^\ensuremath{\infty}(\Pi)$, as follows from \cref{L:loglogDiff}. In particular, $\varphi({\bm{\mathrm{x}}}) \log \abs{{\bm{\mathrm{x}}}}$ and $R({\bm{\mathrm{x}}})$ are well-defined as functions on $\Pi$. It follows that for each component of $\ensuremath{\partial} \Omega_{t, 0}$, $F = F_1 + F_2$, where \begin{align*} F_1(\bgamma(\beta)) &:= \frac{\omega_0}{2 \pi} \int_0^{2 \pi} \varphi({\bgamma(\beta) - \bgamma(\ensuremath{\alpha})}) \log \abs{\bgamma(\beta) - \bgamma(\ensuremath{\alpha})} \ensuremath{\partial}_\ensuremath{\alpha} \bgamma(\ensuremath{\alpha}) \, d \ensuremath{\alpha}, \\ F_2(\bgamma(\beta)) &:= \frac{\omega_0}{2 \pi} \int_0^{2 \pi} R(\bgamma(\beta) - \bgamma(\ensuremath{\alpha})) \ensuremath{\partial}_\ensuremath{\alpha} \bgamma(\ensuremath{\alpha}) \, d \ensuremath{\alpha}. \end{align*} Other than the cutoff function, which introduces no real difficulties, $F_1$ is the same expression as in the classical setting and is estimated in $B^{1, \ensuremath{\varepsilon}}$ in the same manner. We note that applying $d/d \beta$ to $F_1(\bgamma(\beta))$ leads to a singularity in the integrand at $\ensuremath{\alpha} = \beta$. The key to estimating $F_1$ is treating $dF_1/d \beta$, beginning in Lemma 8.7 of \cite{MB2002}, as a principal value integral. The situation is no different here than in \cite{MB2002}. Similarly, for $F_2$, the key is bounding $dF_2/d \beta$ in $C^\ensuremath{\varepsilon}$. This is much simpler than bounding $d F_1/d \beta$, for we have \begin{align*} \diff{}{\beta} F_2(\bgamma(\beta)) &= \frac{\omega_0}{2 \pi} \int_0^{2 \pi} \ensuremath{\nabla} R(\bgamma(\beta) - \bgamma(\ensuremath{\alpha})) \cdot \ensuremath{\partial}_\beta \bgamma(\beta)) \, \ensuremath{\partial}_\ensuremath{\alpha} \bgamma(\ensuremath{\alpha}) \, d \ensuremath{\alpha}. \end{align*} Then for any $\ensuremath{\alpha}$, \begin{align*} &\norm{\ensuremath{\nabla} R(\bgamma(\beta) - \bgamma(\ensuremath{\alpha})) \cdot \ensuremath{\partial}_\beta \bgamma(\beta)) \, \ensuremath{\partial}_\ensuremath{\alpha} \bgamma(\ensuremath{\alpha})}_{C^\ensuremath{\varepsilon}} \\ &\qquad \le \abs{\ensuremath{\partial}_\ensuremath{\alpha} \bgamma(\ensuremath{\alpha})} \norm{\ensuremath{\nabla} R}_{C^\ensuremath{\varepsilon}(\Pi)} \norm{\bgamma(\beta) - \bgamma(\ensuremath{\alpha})}_{lip}^\ensuremath{\varepsilon} \norm{\ensuremath{\partial}_\beta \bgamma(\beta)}_{C^\ensuremath{\varepsilon}(0, 2 \pi)}. \end{align*} But, $\abs{\ensuremath{\partial}_\ensuremath{\alpha} \bgamma(\ensuremath{\alpha})} \le \norm{\bgamma}_{Lip} < M$ and $\norm{\bgamma(\beta) - \bgamma(\ensuremath{\alpha})}_{lip} = \norm{\bgamma}_{lip} < M$. Hence, \begin{align*} \norm[\bigg]{\diff{}{\beta} F_2(\bgamma(\beta))}_{C^\ensuremath{\varepsilon}(0, 2 \pi)} &\le C M^2 \abs{\omega_0} \norm{\bgamma}_{C^\ensuremath{\varepsilon}}. \end{align*} We see, then, that the bounds in Lemma 8.10 of \cite{MB2002} hold, and the proof of local-in-time existence is completed as in \cite{MB2002}. \bigskip\noindent\textbf{Global-in-time $C^{1, \ensuremath{\varepsilon}}$ solutions}: The proof of the global existence of a $C^{1, \ensuremath{\varepsilon}}$ solution to the CDE is the same as in Section 8.3.3 of \cite{MB2002}, except that \cref{C:graduType} is used to obtain $\ensuremath{\nabla} \uu$. By virtue of \cref{P:SymConv}, the estimates differ little from those for classical vortex patches. This completes the proof for Type 2 solutions. The result for Types 1 and 3 solutions then follows directly, exploiting the lifting of domains described in \cref{S:Lifting}. \end{proof}

1be23924e5c9d93508bf62e1b0f0db2d31e3f2ad

\section{Introduction} We consider the following sparse optimization problem: \begin{equation}\label{eq:optimizationproblem} (\text{P}): \quad \begin{array}{l} \min f(\mathbf{x},\Xi):=\frac{1}{N}\sum_{i=1}^{N}f^{(i)}(\mathbf{x}, \xi^{(i)}) \\ \text{s.t. } \mathbf{x} \in C_s \end{array} \end{equation} where $f^{(i)}: \mathbb{R}^n\times \Xi \rightarrow \mathbb{R}$ for $i=1,\dots, N$, $\Xi=\{\xi^{(1)}, \dots, \xi^{(N)}\}$, and $C_s=\{\mathbf{x} \in \mathbb{R}^n \mid \|\mathbf{x}\|_0 \leq s\}$ (sparsity constraint) is the union of finitely many subspaces whose dimension is less than or equal to the sparsity level $s$ such that $1 \leq s<n$. The importance of the Problem (P) is due to the fact that finding a sparse network whose accuracy is on a par with a dense network amounts to solving a bi-level, constrained, stochastic, nonconvex, and non-smooth sparse optimization problem \cite{damadi2022amenable}. Thus finding efficient algorithms that solve Problem (P) can be beneficial for addressing compression of deep neural networks. Among algorithms for solving sparse optimization the Iterative Hard Thresholding (IHT) algorithm has been a very successful one due to the simplicity of its implementation. The IHT algorithm not only has been practically efficient, but also shows theoretical promising results. It was originally devised for solving compressed sensing problems in 2008 \cite{blumensath2008iterative,blumensath2009iterative}. Since then, a large body of literature has been studying it from different perspectives. For example, \cite{beck2013sparsity,lu2014iterative, Lu2015OptimizationOS,pan2017convergent,zhou2021global} consider convergence of iterations, \cite{jain2014iterative, liu2020between} study the limit of the objective function value sequence, \cite{liu2017dual,zhu2018lagrange} address duality, \cite{zhou2020subspace, zhao2021lagrange} extend it to Newton's-type IHT, \cite{blumensath2012accelerated,khanna2018iht,vu2019accelerating,wu2020accelerated} address accelerated IHT, and \cite{wang2019fast, bahmani2013greedy} solve logistic regression problem using the IHT algorithm. Recently \cite{damadi2022gradient} introduced the concepts of HT-unstable stationary points (saddle points in the sense of sparse optimization) and showed the escapability property of the HT-unstable stationary points as one of the crucial properties of the IHT algorithm. Also, they showed Q-linearly convergence of the IHT algorithm towards strictly HT-stable stationary points. However, these desirable properties, requires to compute the batch (full) gradient at each iteration which is computationally expensive or impractical with current GPUs. On the other hand, almost all training for deep neural networks are done using the mini-batch stochastic gradient which is a combination of the stochastic approximation \cite{robbins1951stochastic} implemented by the backpropagation algorithm \cite{rumelhart1986learning}. By taking the mini-batch stochastic approximation, we consider solving Problem (P) using the mini-batch Stochastic Iterative Hard Thresholding algorithm outlined in Algorithm \ref{alg:siht}. Similar to practice where the mini-batch size is fixed beforehand, we fix the mini-batch size at the beginning which is different from previous work \cite{zhou2018efficient} in this area. Also, for showing our theoretical results we directly use the mini-batch stochastic gradient and derive our theoretical results which is different from previous works \cite{chen2016accelerated, li2016nonconvex} where the batch (full) gradient is used to show the theoretical results. As opposed to other works where restricted strong convexity is necessary for deriving convergence results \cite{liang2020effective, zhou2018efficient}, here the only assumption we make is the restricted strong smoothness on the objective function not on each individual one. Also, we assume that the objective function is a bounded below function which is the case for objective functions used in machine learning applications. Similar to practice where the mini-batch size is fixed beforehand, we fix the mini-batch size at the beginning which is different from previous works \cite{zhou2018efficient}. \subsection*{Summary of Contributions} By considering the mini-batch SIHT Algorithm \ref{alg:siht} for Problem (P), we develop the following results: \begin{itemize} \item We establish a new critical sparse stochastic gradient descent property of the hard thresholding (HT) operator that has not been found in the literature. \item For a given step-size $0 <\gamma < \frac{1}{L_s}$, we find a lower bound on the size of the mini-batch that guarantees the expected descent of the objective value function after hardthresholding. \item Using the sparse stochastic gradient descent property we show that the sequence generated by the mini-batch SIHT algorithm is supermartingale and converges with probability one. \item We show that for a certain class of functions in Problem (P) where $f(\mathbf{x},\xi^{i}):=f^{(i)}(\mathbf{V}_{i\bullet}\mathbf{x})$ $f^{(i)}: \mathbb{R}^n \rightarrow \mathbb{R}$, the sum of norm squared of individual gradients restricted to a set of some elements $\mathcal{J}$, i.e., $\sum_{i=1}^N \|\nabla_{\mathcal{J}} f^{(i)}\|_2^2$, evaluated at every point is proportionate to the norm of the batch gradient $\|\nabla_{\mathcal{J}} f\|_2^2$ where the proportionality constant only depends on the data. Moreover, dependency of the proportionality constant on the data is restricted to the set of $\mathcal{J}$ not the entire data. \end{itemize} \input{siht} \section{Related work} In order improve computational efficiency of the IHT algorithm, algorithms based on stochastic hard thresholding try to use the finite-sum structure of problem (P) \cite{nguyen2017linear, li2016nonconvex, shen2017tight}. The StoIHT algorithm is introduced in \cite{nguyen2017linear} where at each iteration a random element from the sum in Problem (P) is drawn and the associated gradient is calculated. Basically, the gradient is approximated by a mini-batch stochastic gradient with size one. The StoIHT algorithm defines a sparse subspace and then projects the updated vector into that. To show the theoretical results in \cite{nguyen2017linear}, the restricted strong smoothness condition for each individual function in Problem (P) is required as well as the restricted strong convexity for the objective function. In addition, the StoIHT algorithm needs the restricted condition number be to 4/3 which is hard to meet in practice. The stochastic variance reduced gradient hard thresholding (SVRG-HT) algorithm \cite{li2016nonconvex, shen2017tight} tries to mitigate the variance with a cost of calculating the (batch) full gradient at some stages. This information of the batch gradient is the key for reducing the variance. Similar to the StoIHT algorithm, the SVRG-HT algorithm requires the restricted strong smoothness condition for each individual function in Problem (P) as well as the restricted strong convexity for the objective function. The Accelerated Stochastic Block Coordinate Gradient Descent with Hard Thresholding (ASBCDHT) algorithm in \cite{chen2016accelerated} is a randomized version of the StoIHT algorithm which suffers the drawbacks of the StoIHT algorithm, i.e., calculating the full gradient and requirement of the restricted strong conditions. The Hybrid Stochastic Gradient Hard Thresholding (HSG-HT) algorithm in \cite{zhou2018efficient} is a variant of stochastic IHT algorithms that uses a mini-batch stochastic gradient at each step. However, from the theoretical perspective, the size of a mini-batch has to increase as the algorithm progresses. This makes the algorithm almost deterministic in calculating the gradient and defeats the purpose of using the mini-batch stochastic gradient. The stochastically controlled stochastic gradients (SCSG-HT) algorithm in \cite{liang2020effective} uses mini-batch stochastic gradients with large batch size as opposed to the SVRG-HT and the ASBCDHT algorithms to reduce the variance with less computation, i.e., not calculating the batch gradient at some steps. We present the mini-batch stochastic IHT algorithm and show that the stochastic sequence of the function value is a supermartingale sequence and it converges with probability one. To show our result, we assume the objective function has the restricted strong smoothness property and is bounded below which is the case for objective functions used machine learning applications. Also, to the best of our knowledge, in the regime of sparse optimization, this is the first time in the literature that it is shown that the sequence of the stochastic function values converges with probability one by fixing the mini-batch size for all steps. \section{Definitions} We provide some definitions that will be used throughout the paper. \begin{definition}[Restricted Strong Smoothness (RSS)]\label{def:rss} A differentiable function $f: \mathbb{R}^n \to \mathbb{R}$ is said to be restricted strongly smooth with modulus $L_s>0$ or is $L_s$-RSS if \begin{equation}\label{eq:rss} f(\mathbf{y}) \leq f(\mathbf{x}) + \langle \nabla f(\mathbf{x}) , \mathbf{y}-\mathbf{x} \rangle + \frac{L_{s}}{2}\|\mathbf{y}-\mathbf{x}\|_2^2 \quad \forall \mathbf{x},\mathbf{y} \in \mathbb{R}^n \text{ such that } \|\mathbf{x}\|_0 \leq s,\|\mathbf{y}\|_0\leq s. \end{equation} \end{definition} \begin{definition} [The HT operator] \label{def:hardthresholding} The HT operator $H_s(\cdot)$ denotes the orthogonal projection onto multiple subspaces of $\mathbb{R}^n$ with dimension $1 \leq s<n$, that is, \begin{equation}\label{eq:hardthreshold} H_s(\mathbf{x}) \in \arg\min_{\|\mathbf{z}\|_0\leq s }\|\mathbf{z}-\mathbf{x}\|_2. \end{equation} \end{definition} \begin{claim}\label{claim:tops} The HT operator keeps the $s$ largest entries of its input in absolute values. \end{claim} For a vector $\mathbf{x} \in \mathbb{R}^n$, $\mathcal{I}^{\mathbf{x}}_s \subset \{1,\dots, n\}$ denotes the set of indices corresponding to the first $s$ largest elements of $\mathbf{x}$ in absolute values. For example $H_2([1,-3,1]^{\top})$ is either $[0,-3,1]^{\top}$ or $[1,-3,0]^{\top}$ where $\mathcal{I}^{\mathbf{y}}_2=\{2,3\}$ and $\mathcal{I}^{\mathbf{y}}_2=\{1,2\}$, respectively. Therefore, the output of it may not be unique. This clearly shows why HTO is not a convex operator and why there is an inclusion in (\ref{eq:hardthreshold}) not an inequality. \begin{definition}[Convergence with probability one] A random sequence $(\mathbf{x}^k \in \mathbb{R}^n)$ in a sample space $\Omega$ converges to a random variable $\mathbf{x}^*$ with probability one if $$\mathbb{P}\Big[\omega \in \Omega: \displaystyle{\lim_{k \to \infty}}\|\mathbf{x}^k(\omega) - \mathbf{x}^*\|\Big]=0.$$ \end{definition} \section{Results} We consider solving Problem (\ref{eq:optimizationproblem}) using the mini-batch SIHT Algorithm \ref{alg:siht} and develop results that guarantee the convergence of the sequence of function values generated by the SIHT Algorithm. To do so, we present our results in two separate subsections. The first part provides stochastic results characterizing expectation of functions involving the sample average of given vectors. Then, in the subsequent subsection we use the aforementioned results to show Theorem \ref{theorem:stochasticdescent} which establishes a stochastic gradient result that is the foundation for the convergence of the function value sequence. \subsection{Stochastic results for sample average} In this subsection, we consider a sample average whose elements are drawn uniformly and without replacement. Then, we prove Lemma \ref{lemma:expectionofsampleaverage} that calculates the expected value of the norm squared of the sample average based on the covariance matrix of a random vector whose elements are Bernoulli random variable determining elements of the sample average. Next, in Corollary \ref{cor:distancetomean} using Lemma \ref{lemma:expectionofsampleaverage} we calculate the expected value of the squared distance between the sample average and the overall average. This result is extended in Theorem \ref{theorem:distancetoeach} where the expected value is calculated so that one is able to find the mentioned expectation based on each individual vector and the overall average. We start with the following well-known lemma. \begin{lemma}[\cite{mathai1992quadratic}]\label{lemma:randomquadratic} Let $\mathbf{\Lambda} \in \mathbb{R}^{n \times n}$ be a deterministic matrix and $\bm{\xi} \in \mathbb{R}^n$ be a random vector that is distributed according to some probability distribution $\mathcal{P}$. Then, \begin{equation}\label{eq:randomquadratic} \mathbb{E}_{\bm{\xi}}\Big[ \bm{\xi}^{\top} \mathbf{\Lambda} \bm{\xi} \Big] =\text{trace}(\mathbf{\Lambda} \text{Cov}(\bm{\xi})) + \mathbb{E}_{\bm{\xi}}^{\top}\Big[\bm{\xi}\Big]\mathbf{\Lambda} \mathbb{E}_{\bm{\xi}}\Big[\bm{\xi}\Big]. \end{equation} \end{lemma} To invoke the above lemma, notice that one can define a random vector whose elements are Bernoulli random variables determining whether the associated vector is in the sample average or not. Thus we prove the following lemma. \begin{lemma}\label{lemma:expectionofsampleaverage} Let $\mathbf{g}^{(1)}, \dots, \mathbf{g}^{(N)} \in \mathbb{R}^n$ be $N$ deterministic vectors and $\text{B} \subseteq \{1, \dots, N\}$ be a random set. Let $\bar{\mathbf{g}}:=\frac{1}{N}\sum_{i=1}^N \mathbf{g}^{(i)}$, $\mathcal{G}(\text{B}):=\frac{1}{|\text{B}|}\sum_{i \in \text{B}} \mathbf{g}^{(i)}$, $\mathbf{G}:=\Big[ \mathbf{g}^{(1)} \quad \dots \quad \mathbf{g}^{(N)} \Big] \in \mathbb{R}^{n \times N}$, and $\mathbf{z}(\text{B})=[z_1(\text{B}), \dots, z_N(\text{B})]^{\top}$ where $z_i(\text{B})$ is a Bernoulli random variable such that $z_i(\text{B})=1$ if $i \in \text{B}$ otherwise $z_i(\text{B})=0$ for $i=1, \dots, N$. Assume $\mathbb{E}_{\text{B}}\big[ \mathcal{G}(\text{B}) \big]=\bar{\mathbf{g}}$, then for any random set $\text{B}$ with fixed size $|\text{B}|$, the following holds: \begin{equation}\label{eq:expectionofsampleaverage} \mathbb{E}_{\text{B}}\big[ \| \mathcal{G}(\text{B})\|^2 \big] = \frac{1}{|\text{B}|^2}\text{trace}\Big(\mathbf{G}^{\top}\mathbf{G} \text{Cov}\big(Z(\text{B})\big)\Big) + \| \bar{\mathbf{g}}\|^2. \end{equation} \end{lemma} Once the above result is established, it is straightforward to show the following by observing the fact that the sample average is an unbiased estimator of the overall average, i.e., $\mathbb{E}_{\text{B}}\big[ \mathcal{G}(\text{B}) \big]=\bar{\mathbf{g}}$. \begin{corollary}\label{cor:distancetomean} Assume all the assumptions in Lemma \ref{cor:distancetomean} hold. Then for any random set $\text{B}$ with fixed size $|\text{B}|$, the following holds: \begin{equation}\label{eq:distancetomean} \mathbb{E}_{\text{B}}\big[ \| \mathcal{G}(\text{B}) - \bar{\mathbf{g}}\|^2 \big] = \frac{1}{|\text{B}|^2}\text{trace}\Big(\mathbf{G}^{\top}\mathbf{G} \text{Cov}\big(Z(\text{B})\big)\Big) \end{equation} \end{corollary} Finally, we use the above results to prove the following which calculates the expected squared distance between the sample average and the overall average based on individual vectors and the overall average. The following result is critical because later we will see that Equation (\ref{eq:distancetoeach}) connects the mini-batch stochastic gradient, the batch gradient, and individual gradients in Problem (P). \begin{theorem}\label{theorem:distancetoeach} Assume all the assumptions in Lemma \ref{cor:distancetomean} hold. If elements of the random set $\text{B}$ are drawn uniformly and without replacement, then \begin{equation}\label{eq:distancetoeach} \mathbb{E}_{\text{B}}\big[ \| \mathcal{G}(\text{B}) - \bar{\mathbf{g}}\|^2 \big] = \frac{N-|\text{B}|}{|\text{B}|N(N-1)} \Big( \sum_{i=1}^N \|\mathbf{g}^{(i)}\|_2^2 - N \|\bar{\mathbf{g}}\|^2\Big) = \frac{N-|\text{B}|}{|\text{B}|N} \frac{1}{N-1} \sum_{i=1}^N \|\mathbf{g}^{(i)}-\bar{\mathbf{g}}\|_2^2. \end{equation} \end{theorem} \subsection{Stochastic results for Hard Thresholding operator} The goal of this subsection is to show the random sequence $\big(f(\mathbf{x}^k)_{k \geq 1}\big)$ generated by the mini-batch SIHT algorithm converges with probability one. To show this we prove that the random sequence of the function value is a supermartingale sequence so the expected value of the function value sequence is decreasing. To achieve our goal, we prove the following lemma that provides an upper bound on the function value evaluated at a thresholded vector. Notice that the following result does not require the input be an updated vector by the gradient. \begin{lemma}\label{lemma:rsswithdelta} Let $f: \mathbb{R}^n \rightarrow \mathbb{R}$ be in $C^1$ and $L s$-RSS. Then for a fixed $\mathbf{x} \in C_s$ with any $\mathcal{I}_s^{\mathbf{x}}$, any $0 < \gamma \leq \frac{1}{L_s}$, and any given vector $\mathbf{g} \in \mathbb{R}^n$, either of the following holds for any $\mathbf{y} \in H_s(\mathbf{x}-\gamma \mathbf{g})$ with any $\mathcal{I}_s^{\mathbf{y}}$: \begin{equation} f(\mathbf{y}) \leq f(\mathbf{x}) - \frac{\gamma}{2}(1-L_s\gamma) \| \mathbf{g}_{\mathcal{I}_s^{\mathbf{y}}}\|^2_2 - \frac{\gamma}{2} \| \mathbf{g}_{\mathcal{I}_s^{\mathbf{x}} }\|^2_2 + \gamma \langle \mathbf{\delta}_{\mathcal{I}_s^{\mathbf{y}}}, \mathbf{g}_{\mathcal{I}_s^{\mathbf{y}}} \rangle + \gamma \langle \mathbf{\delta} _{\mathcal{I}\backslash \mathcal{I}_s^{\mathbf{y}}}, \mathbf{x}_{\mathcal{I}\backslash \mathcal{I}_s^{\mathbf{y}}} \rangle \label{eq:rsswithdeltaxandy} \end{equation} where $\mathcal{I} = \mathcal{I}_s^{\mathbf{x}} \cup \mathcal{I}_s^{\mathbf{y}}$ and $\mathbf{\delta}=\mathbf{g} - \nabla f(\mathbf{x})$. \end{lemma} Observe that in the above lemma the vector $\mathbf{g}$ can be any vector in $\mathbb{R}^n$. It need not be the gradient nor the mini-batch gradient. However, in the following lemma we prove that if $\mathbf{g}$ is designated to be an unbiased stochastic approximation of the gradient at an arbitrary point, then the following result holds. \begin{lemma}\label{lemma:unbiasedapproximation} Let $f: \mathbb{R}^n \rightarrow \mathbb{R}$ be in $C^1$ and $L s$-RSS. Assume $\mathbf{g}(\mathbf{x}, \omega)$ be an unbiased stochastic approximation of the gradient at $\mathbf{x} \in \mathbb{R}^n$ where $\omega \sim D$ for some distribution $D$, i.e., $\mathbb{E}_{\omega}[\mathbf{g}(\mathbf{x}, \omega)]=\nabla f(\mathbf{x})$. Then for a fixed $\mathbf{x} \in C_s$ with any $\mathcal{I}_s^{\mathbf{x}}$ and $0 < \gamma \leq \frac{1}{L_s}$, either of the following holds for any $\mathbf{y}(\omega) \in H_s(\mathbf{x}-\gamma \mathbf{g}(\mathbf{x}, \omega))$ with any $\mathcal{I}_s^{\mathbf{y}(\omega)}$: \begin{equation} \mathbb{E}_{\omega}[ f(\mathbb{Y}(\omega)) ] \leq f(\mathbf{x}) - \frac{\gamma}{2}(1-L_s\gamma) \mathbb{E}_{\omega}[ \| \mathbf{g}_{\mathcal{I}_s^{\mathbb{Y}(\omega)}}(\mathbf{x}, \omega)\|^2_2 ] - \frac{\gamma}{2} \| \nabla_{\mathcal{I}_s^{\mathbf{x}}} f(\mathbf{x})\|^2_2 + \gamma \mathbb{E}_{\omega}[ \|\mathbf{\delta}_{\mathcal{I}_s^{\mathbb{Y}(\omega)}}(\omega)\|_2^2 ] \label{eq:unbiasedwithxandy} \end{equation} where $\mathcal{I}(\omega) = \mathcal{I}_s^{\mathbf{x}} \cup \mathcal{I}_s^{\mathbb{Y}(\omega)}$ and $\mathbf{\delta}(\omega)=\mathbf{g}(\mathbf{x}, \omega) - \nabla f(\mathbf{x})$. \end{lemma} The following Theorem is the climax of our technical results because it establishes a stochastic gradient descent property for the expectation of the function value. Later we will see how Inequality (\ref{eq:generalminibatch}) is used in Theorem \ref{theorem:stochasticdescent} to show the sequence of the function values generated by the mini-batch SIHT is a supermartingale sequence. \begin{theorem}\label{theorem:sparseminibatch} Let $f^{(i)}: \mathbb{R}^n\times \Xi \rightarrow \mathbb{R}$ be in $C^1$ \footnote{The class consisting of all differentiable functions whose derivative is continuous.} for $i=1,\dots, N$ and $\Xi=\{\xi^{(1)}, \dots, \xi^{(N)}\}$ be a given set such that $f(\mathbf{x},\Xi)=\frac{1}{N}\sum_{i=1}^{N}f^{(i)}(\mathbf{x}, \xi^{(i)})$ be an $L_s$-RSS function. Assume there exists a $c>0$ \footnote{In Remark \ref{remark:1}, we explain why such a $c$ always exist for widespread objective functions in machine learning applications} such that \begin{equation}\label{eq:expectationindividualvsgradient} \mathbb{E}_{\mathcal{J}} \Big [\sum_{i=1}^N \|\nabla_{\mathcal{J}} f^{(i)}(\mathbf{x}, \xi^{(i)})\|_2^2 \Big ]\leq c \mathbb{E}_{\mathcal{J}} \Big [\|\nabla_{\mathcal{J}} f(\mathbf{x},\Xi)\|_2^2 \Big ] \end{equation} for all $\mathbf{x} \in \mathbb{R}^n$ and any random index set $\mathcal{J} \subseteq \{1, \dots, n\}$ with $|\mathcal{J}| \leq s$. Let $\mathcal{G}(\mathbf{x}, \Xi, B)=\frac{1}{|B|}\sum_{i \in B}\nabla f^{(i)}(\mathbf{x},\xi^{(i)})$ be the mini-batch stochastic gradient at any $\mathbf{x}\in \mathbb{R}^n$ where $B \subseteq \{1, \dots, N\}$ be a random set whose elements are drawn randomly and uniformly from $\{1, \dots, N\}$ without replacement and its size is $|B|$. For a fixed $0<\gamma < \frac{1}{L_s}$, assume the size of $B$ is fixed such that $|B| \geq N/\Big(1+\frac{1-L_s\gamma}{1+L_s\gamma}\frac{N-1}{\frac{c}{N}-1}\Big)$ and let $\zeta := \frac{N-|\text{B}|}{|B|(N-1)}$ for $N \geq 2$. Then for a fixed $\mathbf{x} \in C_s$ with any $\mathcal{I}_s^{\mathbf{x}}$ the following holds for any $\mathbb{Y}(B) \in H_s(\mathbf{x}-\gamma \mathbf{g}(\mathbf{x}, \Xi, B))$ with any $\mathcal{I}_s^{\mathbb{Y}(B)}$: \begin{equation}\label{eq:generalminibatch} \begin{aligned} \mathbb{E}_{B} \Big [ f(\mathbb{Y}(B),\Xi) \Big ] &\leq f(\mathbf{x}, \Xi) - \frac{\gamma}{2} \| \nabla_{\mathcal{I}_s^{\mathbf{x}}} f(\mathbf{x})\|^2_2 \\ &- \frac{\gamma}{2}(1+L_s\gamma)\zeta \Big( 1- \frac{c}{N} +\frac{1-L_s\gamma}{1+L_s\gamma}\frac{1}{\zeta} \Big) \mathbb{E}_{\mathcal{I}_s^{\mathbb{Y}(B)}}\Big[ \|\nabla_{\mathcal{I}_s^{\mathbb{Y}(B)}} f(\mathbf{x}, \Xi)\|^2 \Big] \end{aligned} \end{equation} where $1- \frac{c}{N} +\frac{1-L_s\gamma}{1+L_s\gamma}\frac{1}{\zeta} \geq 0$. \end{theorem} A crucial assumption for proving the results in Theorem (\ref{eq:generalminibatch}) is the assumption made in Inequality (\ref{eq:expectationindividualvsgradient}). In the following Claim we show that for a certain class of functions $c>0$ always exists and it does not depend on the function. We will prove that for these special classes of functions the value of $c$ only depends on the data. \begin{claim}\label{claim:individualvsgradient} Let the given set $\Xi$ in Problem (P) be defined such that $\Xi:=\{\mathbf{V}_{1\bullet}, \dots, \mathbf{V}_{N\bullet}\}$ where each $\mathbf{V}_{i\bullet}$ is the $i$-th row of a given matrix $\mathbf{V} \in \mathbb{R}^{N \times n}$. Then the objective function in Problem (P) can be defined as $f(\mathbf{x},\Xi):=\frac{1}{N}\sum_{i=1}^{N}f^{(i)}(\mathbf{V}_{i\bullet}\mathbf{x})$ $f^{(i)}: \mathbb{R}^n\times \Xi \rightarrow \mathbb{R}$ and the following holds: \begin{equation}\label{eq:individualvsgradient} \sum_{i=1}^N \|\nabla_{\mathcal{J}} f^{(i)}(\mathbf{V}_{i\bullet}\mathbf{x})\|_2^2 \leq \frac{N^2}{ \sigma_{min}^2(\mathbf{V}\mathbf{I}^{\top}_{\mathcal{J}\bullet}\mathbf{I}_{\mathcal{J}\bullet}\mathbf{V}^{\top}) } \Big( \max_{r=1, \dots, N} \Big\{ \|(\mathbf{V}_{r\bullet}^{\top})_{\mathcal{J}}\|_2^2 \Big\} \Big) \| \nabla_{\mathcal{J}} f(\mathbf{x},\mathbf{V})\|_2^2 \end{equation} where $\mathcal{J} \subseteq \{1, \dots, n\}$ with $|\mathcal{J}| \leq s$, $\mathbf{I}_{\mathcal{J}\bullet} \in \mathbb{R}^{|\mathcal{J}| \times n}$ is a restriction of the Identity matrix whose rows are associated with indices in $\mathcal{J}$, $\mathbf{V}\mathbf{I}^{\top}_{\mathcal{J}\bullet}\mathbf{I}_{\mathcal{J}\bullet}=\sum_{i=1}^{|\mathcal{J}|} \mathbf{V}_{\bullet i}\mathbf{V}_{\bullet i}^{\top}$, $\sigma_{min}(\cdot)$ is the smallest singular value, $\mathbf{V}_{\bullet i}$ is the $i$-th column of $\mathbf{V}$, and $(\cdot)\mathcal{J}$ is a vector restricted to indices in $\mathcal{J}$. \end{claim} \begin{remark}\label{remark:1} The above claim shows that for a class of functions $f(\mathbf{x},\Xi):=\frac{1}{N}\sum_{i=1}^{N}f^{(i)}(\mathbf{V}_{i\bullet}\mathbf{x})$ the constant $c>0$ in Theorem \ref{theorem:stochasticdescent} always exists and it does not depend on the value of $\mathbf{x}$ or its gradient whether it is batch (full) gradient or individual one. For an example of functions belonging to this class one can think of the mean square error loss used for linear regression as follows: $$ f(\mathbf{x}, \mathbf{V})=\frac{1}{N}\|\mathbf{V}\mathbf{x}-\mathbf{y}\|^2=\frac{1}{N}\sum_{i=1}^N(\mathbf{V}_{i\bullet}\mathbf{x}-y_i)^2 $$ where $\mathbf{V} \in \mathbb{R}^{N\times n}$, $\mathbf{V}_{i\bullet}$ is the $i$-th row of $\mathbf{V}$, $\mathbf{x} \in \mathbb{R}^n$ is the optimization variable, and $\mathbf{y} \in \mathbb{R}^N$ is the target. Also, the logistic regression loss (binary cross entropy) is a function for which $c>0$ in Inequality (\ref{eq:individualvsgradient}) always exists since it can be written as follows: $$ f(\mathbf{x}, \mathbf{V})=\frac{1}{N} \sum_{i=1}^{N}\Big( -y^{(i)}(\mathbf{V}_{i\bullet}\mathbf{x})+\log\big(1+e^{\mathbf{V}_{i\bullet}\mathbf{x}}\big)\Big) $$ where $\mathbf{V} \in \mathbb{R}^{N\times n}$ whose last column is all one, $\mathbf{V}_{i\bullet}$ is the $i$-th row of $\mathbf{V}$, $\mathbb{R}^n\ni \mathbf{x}=[\mathbf{w}, b]^{\top}$ such that $\mathbf{w} \in \mathbb{R}^{n-1}$ and $b \in \mathbb{R}$ are the optimization variables, and $y^{(i)} \in \{0, 1\}$ for $i=1, \dots, N$. \end{remark} Now we can provide a result showing that by fixing a sparse point, one can use the stochastic mini-batch gradient with a fixed mini-batch size determined in Theorem \ref{theorem:stochasticdescent} and decrease the function value in expectation. \begin{theorem}\label{theorem:stochasticdescent} Assume all the assumptions in Theorem \ref{theorem:sparseminibatch} hold. Then for a fixed $\mathbf{x} \in C_s$ with any $\mathcal{I}_s^{\mathbf{x}}$ the following holds for any $\mathbb{Y}(B) \in H_s(\mathbf{x}-\gamma \mathcal{G}(\mathbf{x}, \Xi, B))$: \begin{equation}\label{eq:generalminibatchwithx} \mathbb{E}_{B} \Big [ f(\mathbb{Y}(B),\Xi) \bigg\vert \mathbf{x} \Big ] \leq f(\mathbf{x}, \Xi) - \frac{\gamma}{2} \| \nabla_{\mathcal{I}_s^{\mathbf{x}}} f(\mathbf{x})\|^2_2 . \end{equation} \end{theorem} The above result is the analogue result to \cite[Corollary 1]{damadi2022gradient}. \begin{theorem}\label{theorem:functionconvergence} Assume all the assumptions in Theorem \ref{theorem:sparseminibatch} hold. Let $f$ be a bounded below differential function and $\big(\mathbb{X}^k \bigg\vert \mathbb{X}^{k-1})_{k\geq 0}$ be the stochastic IHT sequence. Then, $\Big(f(\mathbb{X}^{k}, \Xi, B) \bigg\vert \mathbb{X}^{k} \Big)_{k\geq 1}$ is a supermartingale sequence and converges to a random variable $f^*$ with probability one. \end{theorem} \section{Conclusion} We showed the stochastic sequence generated by the mini-batch stochastic IHT is a supermartingale sequence converging with probability one. To show this result we used the stochastic gradient descent property that we derived where we utilized the property of the mini-batch stochastic gradient as the sample sum of a finite sum. \newpage

2c1f4ced236a5278f83089ca6b1c111966d27efc

\section{Introduction} Since the beginning of the century we know that the observable Universe is in accelerated expansion which implies the existence of a {\em positive} cosmological constant $\Lambda >0$ \cite{Riess1998,Perlmutter1999}. It is also known that a positive $\Lambda$ imposes restrictions on the area of marginally (outer) future-trapped surfaces \cite{HSN,W} if these are stable in the sense of \cite{AMS,AMS1} --equivalently, `outer' in the sense of \cite{Hay,HSN}-- if the dominant energy condition holds. These limits can be generalized and strengthened by adding electromagnetic charge \cite{Simon}. The stability assumption can be understood as stating that the marginal trapped surfaces (MTS) enclose a black hole (BH) region. The area $A$ for (stable) MTS is limited by \begin{equation}\label{lim} A< \frac{4\pi}{\Lambda} . \end{equation} Taking into account the relationship between the area of MTS (and of black hole event horizons) and their mass, one may wonder which kind of mechanism, if any, prohibits a BH with $A$ near the limit \eqref{lim} to increase its area by simply receiving more mass-energy from its exterior. Observe that this is completely different from the known cases of over-spinning or over-charging BHs with the goal of producing naked singularities, since in those cases there exist repulsive forces and a struggle between the increase in charge and/or angular momentum and the associated increase of the mass \cite{SMilestone}: the theory seems to conspire so that cosmic censorship prevails \cite{Wald1}. In contrast, in principle a BH will simply become bigger by adding mass, and it is difficult to imagine what can prevent such physical process. In this paper, in order to understand this problem, I consider some simple models of spherical BHs that keep increasing their masses until the stable MTS of spherical topology reach the area-limit value and beyond. I will analyze the simplest possible models, first based on the Vaidya-de Sitter metric \cite{VS,Mallet} and then also in combination with the $\Lambda >0$ generalizations of the Oppenheimer-Snyder and Lema\^\i tre-Tolman collapses studied in \cite{GV,MS,L,DJCJ} --later re-discovered in \cite{SA,SA1}. In both cases we show that there is no problem in having larger and larger masses, but the dynamical horizon foliated by marginally trapped spheres then simply ends its existence. The cosmological horizon totally vanishes. The global structure of the resulting spacetimes is shown in convenient conformal diagrams. The global nature of event horizons is partly behind its dematerialization in these extreme spacetimes that I have called {\em ultra-massive}. However, that is not the main reason, or at least not only: the vanishing of future null infinity $\mathscr{J}^+$ is the basic fact. This absence leads to `frustrated event horzions' and, as was to be expected, the area limit \eqref{lim} is never surpassed by any stable MTS. The next section is devoted to understanding the basic properties of Vaidya-de Stitter spacetimes. They can be easily inferred from those of the Kottler metrics (also known as Schwarzschild-de Sitter), which are well known and thoroughly studied in the literature. However, for the benefit of the reader I have added a useful Appendix with the main properties and corresponding conformal diagrams of the Kottler metrics. They can help in better understanding the main text. In section \ref{sec:first} I present the first type of models, based on Vaidya-de Sitter exclusively. Section \ref{sec:second} is devoted to the second type of models, which combine the first type of models with black holes in formation by stellar collapse using the Oppenheimer-Snyder-de Sitter models and others. I end the paper with an extensive discussion. \section{The Vaidya-de Sitter metric} Using the advanced null coordinate $v$ the Vaidya-de Sitter metric takes precisely the form \eqref{kot2} with the mass parameter replaced by an arbitrary function of $v$: \begin{equation}\label{VdS} ds^2 = -\left(1-\frac{2m(v)}{r}-\frac{\Lambda}{3} r^2\right) dv^2 +2dv dr +r^2 d\Omega^2 \end{equation} where $d\Omega^2$ is the standard metric of the unit round sphere, $r$ is the areal coordinate (so that round spheres with $v$ and $r$ constant have area $4\pi r^2$) and the range of coordinates is $v\in(-\infty,\infty)$ and $r\in (0,\infty)$ (or $r\in (-\infty,0)$). The metric \eqref{VdS} is a solution of the Einstein field equations with cosmological constant for an energy-momentum tensor of null radiation $$ T_{\mu\nu}= \frac{2}{r^2}\frac{dm}{dv}k_\mu k_\nu $$ where the future pointing null one-form $k_\mu$ and vector field $k^\mu$ are given respectivly by $$ \mathbf{k}=-d v, \hspace{1cm} \vec k =-\frac{\partial}{\partial r}. $$ Thus, the massless particles of the `null dust' propagate along the null hypersurfaces $v=$ const.\ towards decreasing values of $r$, that is to say, towards round spheres of smaller areas. The dominant energy condition is satisfied whenever $m(v)$ is a non-decreasing function everywhere \begin{equation}\label{mdot} \frac{dm}{dv}\geq 0 \end{equation} which I assume from now on. I will also assume $m\geq 0$ everywhere. Kuroda \cite{K} proved that, under the above assumptions, a naked singularity would form in the Vaidya spacetime (with $\Lambda =0$) if the mass function initially increases slowly, that is if $m(v\rightarrow 0^+)/v\leq 1/16$ --see also \cite{FST,BeS}. This limit was confirmed for the Vaidya-de Sitter case in \cite{WM}, and thus for simplicity I am going to assume\footnote{If this condition does not hold, the main conclusions do not change: absence of $\mathscr{J}^+$ and event horizon. The only difference will be the existence of another, null and locally naked, singularity in addition to the universal one in the future.} \begin{equation}\label{precond} \lim_{v\rightarrow 0^+} \frac{m(v)}{v} > \frac{1}{16} . \end{equation} The hypersurfaces $r=$const.\ have a normal one-form that satisfies $$g^{\mu\nu}\partial_\mu r \partial_\nu r = 1-\frac{2m(v)}{r}-\frac{\Lambda}{3} r^2$$ so that they are always spacelike for large values of $r$ and for $r\longrightarrow 0$ if $m>0$. In those regions $r$ is a time coordinate. Those hypersurfaces can also be timelike if there are regions where the function above is positive. Fixing $v$, this can happen only if \begin{equation}\label{cond} \Lambda < \frac{1}{9m^2(v)} \end{equation} at that $v$. In particular this is always the case for $v$ such that $m(v)=0$. If condition \eqref{cond} holds at a given $v$, then there are two values of $r$, that I denote by $r_-(v)$ and $r_+(v)$, such that the round spheres with those values of $r$ at that value of $v$ are marginally trapped. It is easily seen that they satisfy $$0<r_-(v)\leq \frac{1}{\sqrt{\Lambda}}\leq r_+(v)<\frac{3}{\sqrt{\Lambda}}$$ and that $r_-(v)$ increases, while $r_+(v)$ decreases, with $v$, that is, as $m(v)$ increases. Equality here is only possible if there exists a value $\bar v$ of $v$ such that \begin{equation} \label{vbar} m (\bar v) = \frac{1}{3\sqrt{\Lambda}} . \end{equation} The two hypersurfaces defined by $r=r_\pm(v)$ have a normal one-form given by \begin{equation} \left( 1-\Lambda r_\pm^2(v) \right) dr -2 \frac{dm}{dv} dv \end{equation} whose norm is \cite{BeS} $$-4\frac{dm}{dv}(1-\Lambda r_\pm^2(v) )$$ so that these marginally trapped spheres pile up to form a (spacelike) dynamical horizon and a (timelike) marginally trapped tube (see \cite{AG,AK,BeS,Booth,S} for definitions) given by $r=r_-(v)$ and $r=r_+(v)$, respectively. I will denote by AH the former and by MTT the latter. (Of course, in open regions where $m(v)$ is a --non-zero-- constant, they become null Killing horizons as the metric is Kottler there). Observe that, in case the value $\bar v$ in \eqref{vbar} exists, then $$ r_- (\bar v) =r_+ (\bar v) = \frac{1}{\sqrt{\Lambda}} $$ so that AH and MTT merge at the special round sphere defined by $v=\bar v$ and $r=1/\sqrt{\Lambda}$, and they both become null and tangent to the $v=\bar v$ null hypersurface there. This special round sphere has precisely the maximum area $4\pi/\Lambda$. \section{First type of models}\label{sec:first} \begin{figure}[h!] \includegraphics[width=14cm]{VaidyadS1.pdf} \caption{{\footnotesize Conformal diagram of the Vaidya-de Sitter metric with $9 \mu^2 \Lambda <1$. Radial null geodesics are at 45$^o$ and future is upwards. Each point in the diagram represents a round sphere of area $4\pi r^2$ except the origin of coordinates on the left with $r=0$. Around this centre the metric is originally a portion of de Sitter, but at $v=0$ null matter collapses spherically towards that centre until the advanced time $v=v_1$. This is shown by the shadowed zone in the diagram, with the arrows pointing in the direction of propagation of the matter. The collapse produces a singularity and the appearance of future trapped surfaces enclosed by an apparent horizon AH, which is a dynamical horizon foliated by marginally future-trapped round spheres with $r=r_-(v)$ (see main text). At $v=v_1$ one has $m(v_1)=\mu$, and the resulting spacetime to the right of $v=v_1$ is a BH of Kottler type and mass parameter $\mu$ , as indicated. Therefore, there is also a cosmological horizon at $r=r_+(v_1)$ and a blue region with past trapped spheres to the past of future infinite $\mathscr{J}^+$. The AH merges with the event horizon EH at the sphere $v=v_1$ with $r=r_-(v_1)$. In the figure $r_\pm (v_1)$ are simply represented by $r_\pm$, and they correspond to the two values of $r$ at the Killing horizons of the Kottler part. The original Killing horizon with $r=\sqrt{3/\Lambda}$ in dS becomes a timelike MTT with $r=r_+(v)$ (see main text) when it is crossed by the in-falling matter until the inflow terminates, where it becomes the Killing horizon $r=r_+$ of the Kottler metric corresponding to the mass parameter $\mu$. The corresponding hypersurfaces with $r=r_\pm (v_1)$ in the dS and shadowed regions are also drawn. The hypersurface $r=\sqrt{3/\Lambda}$ is spacelike for $v>0$ as indicated. The hypersurfaces with constant $r$ that cross the AH become null there and then spacelike, as the one shown with $r<r_-$. The round spheres are future trapped in the zone above AH, and in the zone to the right of the blue line --which is an MTT with two Killing horizon portions-- and below the $r=r_+$ on the Kottler part. Not to overwhelm the picture, this is simply indicated with several `red' words. The metric is extendible towards the right, and the analytical extension is that shown in the Appendix for the Kottler metric.}} \label{fig:VdS1} \end{figure} The first type of models I am going to consider are defined by imploding null dust into an empty de Sitter universe. Thus, the mass function $m(v)$ is assumed to vanish at initial values of $v$. At a given advanced time (say $v=0$) $m(v)$ starts to increase until eventually reaches its maximum value $\mu$, say at $v=v_1$: \begin{equation} m(v\leq 0)=0, \hspace{1cm} m(v\geq v_1)=\mu \end{equation} where the condition \eqref{mdot} holds for $v\in (0,v_1)$, and \eqref{precond} is enforced too. There are two possibilities to be considered, depending on whether $9\mu^2 \Lambda <1$ or not. The case usually analyzed in the literature has $9\mu^2 \Lambda <1$, see e.g. \cite{ABK}, especially concerning BH evaporation \cite{M,Mallet1,H,ZYR} because it leads to a standard Kottler (or Schwarzschild-de Sitter) black hole. The conformal diagram is presented in Figure \ref{fig:VdS1}. However, in this paper I want to consider the other possibility, that is, when the final mass parameter satisfies $9\mu^2 \Lambda >1$. Therefore, now the value $\bar v$ in \eqref{vbar} does exist. In this situation a spherically symmetric dynamical horizon AH emerges and the area of the marginally trapped surfaces foliating AH increases with $m(v)$ until it reaches its maximum possible value \eqref{lim} at $v=\bar v$, but $m(v)$ keeps growing beyond that value until it reaches $m(v_1)=\mu > m(\bar v)$. This leads to the absence of future null infinity $\mathscr{J}^+$ and thereby to the nonexistence of any event horizon EH. What was going to become a BH actually grows up `too much' and ends up swallowing the entire spacetime that becomes a contracting universe of type \eqref{kot3} outside the matter. Now the singularity is universal and every single possible observer or photon will inevitably end up there in finite time (or affine parameter). The whole thing is explained, and perhaps better understood, in the corresponding conformal diagram of Figure \ref{fig:VdS2}. \begin{figure}[h!] \includegraphics[width=16cm]{VaidyadS2.pdf} \caption{{\footnotesize Conformal diagram of the Vaidya-de Sitter metric with $9 \mu^2 \Lambda >1$, same conventions as in the previous figure. The metric is originally a portion of de Sitter and at $v=0$ null matter coming from past null infinity collapses spherically towards the regular centre at $r=0$ until the advanced time $v=v_1$, where $m(v_1)=\mu$. This is shown by the shadowed zone in the diagram, with the arrows pointing in the direction of propagation of the matter. To the right of $v=v_1$ the spacetime is of Kottler type with mass parameter $\mu$, that is, of the type shown in Figure \ref{fig:kot2}. Therefore, there is no cosmological horizon (no blue region with past trapped spheres) nor future infinity $\mathscr{J}^+$ in this situation. As in the previous case, the collapse produces a singularity and the appearance of future trapped surfaces enclosed by an apparent horizon AH, which is a dynamical horizon foliated by marginally future-trapped round spheres with $r=r_-(v)$. However, at $v=\bar v$ one reaches the extreme value $m(\bar v) =1/(3\sqrt{\Lambda})$, hence the AH must end up there somehow. The original Killing horizon with $r=\sqrt{3/\Lambda}$ in dS becomes a timelike MTT with $r=r_+(v)$ (see main text) when it encounters the in-falling matter and then the area of their marginally trapped spheres decreases until asymptotically tends to $4\pi/\Lambda$ ($r=1/\sqrt{\Lambda}$) and to a merging with AH. Both AH and MTT become null at the crucial, marginally trapped, round sphere defined by $v=\bar v$ and $r=r_\pm(\bar v) =1/\sqrt{\Lambda}$. Some hypersurfaces with constant $r$ are also shown, they are timelike to the left of MTT (and below AH), become null at MTT and AH, and are spacelike to the right of the MTT (and above AH). All of them reach spacelike infinity $i^0$ asymptotically. Future trapped round spheres are abundant as they are un-trapped only in the zone to the left of MTT and below AH. This is a small portion of the spacetime, because a spacelike disk centered at $r=0$ and reaching up to the MTT has a finite volume, while spacelike cylinders starting anywhere in the diagram with $r>0$ and reaching $i^0$ have an infinite volume. Many observers, such as the $O$ showed, will never be able to see (or be influenced by) the matter creating the strong gravitational field, but they feel the latter. The $O$-particle horizon enclosing its complete past is shown by the dotted lines. Every possible observer in this spacetime ends up at the future singularity. The metric is globally hyperbolic and inextendible.}} \label{fig:VdS2} \end{figure} The null hypersurface $v=\bar v$ is a past horizon for the region with marginally trapped round spheres, and any event with $v>\bar v$ is unable to influence any such MTS. This region containing marginally trapped round spheres is actually a small (finite) portion of the entire spacetime, as can be proven by computing the volume of spacelike slices contained in the appropriate regions. On the one hand, the volume of spacelike spherically symmetric hypersurfaces orthogonal to the $r=$const.\ hypersurfaces contained in the region with $1-\frac{2m(v)}{r}-\frac{\Lambda}{3} r^2 >0$ have a volume $$ 4\pi \int_0^{r_\pm(v)} \frac{r^2}{\sqrt{1-\frac{2m(V(r))}{r} -\frac{\Lambda}{3} r^2}} dr =\mbox{finite} $$ where $v=V(r)$ is the function defining these hypersurfaces. On the other hand, the spacelike spherically symmetric hypersurfaces $r=$const.\ in the Kottler region $1-\frac{2\mu}{r}-\frac{\Lambda}{3} r^2 <0$ have an infinite volume $$ 4\pi r^2 \sqrt{\frac{\Lambda}{3} r^2+\frac{2\mu}{r} -1} \int_{v_1}^\infty dv = \infty . $$ \begin{figure}[h!] \includegraphics[width=16cm]{VaidyadS3.pdf} \caption{{\footnotesize Another conformal diagram of the Vaidya-de Sitter metric with $9 \mu^2 \Lambda >1$, same conventions as in the previous figure. Now the flow of incoming radiaton stops at $v=v_2$ until $v=v_3$, where more matter comes in until the advanced time $v=v_1$, where $m(v_1)=\mu$. The two zones with matter are shown by the shadowed strips in the diagram, with the arrows pointing in the direction of propagation of the matter. To the right of $v=v_1$ the spacetime is of Kottler type with mass parameter $\mu$, that is, the same portion as in Figure \ref{fig:VdS2}. Therefore, the main interesting features as in the previous figure remain: there is no cosmological horizon, no future infinity $\mathscr{J}^+$, and no EH. All features for $v<v_2$ and for $v\geq v_3$ are the same as in the spacetime of Figure \ref{fig:VdS2}. However, the key difference here is the existence of the region $v_2 < v < v_3$ where there is no matter and the metric is Kottler with two Killing horizons. One should observe that the region with $v< v_3$ is identical with a portion of the spacetime in Figure \ref{fig:VdS1}, where an EH had formed and a BH is settled down. Hence, one can argue that the region with $v<v_3$ is, in that region, a temporary BH in equilibrium with mass parameter $m(v_2)$. However, the null hypersurface destined to be the EH of such a BH in equilibrium never becomes an actual event horizon, here marked by `frustrated EH', as a second flow of matter turns the spacetime into another ultra-massive universe.}} \label{fig:VdS3} \end{figure} It is remarkable that the absence of future null infinity arises precisely because there is a positive cosmological constant. As proven in \cite{K}, the non-existence of $\mathscr{J}^+$ leading to the absence of the EH requires, for the Vaidya metric (that is, with $\Lambda =0$), that $$ \lim_{v\rightarrow \infty} \frac{m(v)}{v} > \frac{1}{16}. $$ In plain words, when $\Lambda =0$ one needs a very large {\em infinite} total mass. However, the existence of $\Lambda > 0$ changes this drastically and any finite $m(v)$ larger than $3/\sqrt{\Lambda}$ removes $\mathscr{J}^+$ and the EH. As the spacetime was locally creating a BH for a period of advanced time, these are somehow {\em frustrated black holes}, but the frustration arises simply because there are no observers reaching infinity, so the would-be BH ends up being a victim of its own success as its mass increases beyond the acceptable limit for the area of MTS. Therefore, recalling that the word `ultra' comes from the Greek `beyond', I think an acceptable name for these type of models is {\em ultra-massive spacetimes}. To make the features of these models more explicit and, perhaps, surprising, let us consider a mass function with the following properties $$ m(v\leq 0)=0, \hspace{1cm} m(v_2<v < v_3) =\mu_0 < \frac{3}{\sqrt{\Lambda}}, \hspace{1cm} m(v\geq v_1)=\mu >\frac{3}{\sqrt{\Lambda}} $$ with $0<v_2 <v_3<\bar v <v_1$ and $\mu_0$ a constant less than the critical value, while keeping \eqref{mdot} everywhere. This describes a situation identical to the first model, i.e. the creation of a BH of mass $\mu_0$, for all $v< v_3$, as the model settles down to a would-be EH, corresponding to the Killing horizon with $r=r_-(\mu_0)$ of the Kottler metric. Nevertheless, this will eventually become a frustrated EH due to the extra matter that falls into the would-be BH after $v=v_3$. Eventually, the BH never forms, again victim of its own success in accumulating matter. The corresponding conformal diagram is presented in Figure \ref{fig:VdS3}. I would like to remark that the zone $v_2<v<v_3$ can be made extremely large, so that such a `frustrated BH' can look like a real BH in equilibrium for a period of time that can be taken as large as desired. There is a limiting case when $v_1=\bar v$, that is to say, $m(\bar v) =m(v_1)=\mu =3/\sqrt{\Lambda}$. The corresponding conformal diagram can be easily drawn, by taking the required part of the extreme Kottler metric joined to a version of Figure \ref{fig:VdS2} with $v_1 =\bar v$, to the left of $\bar v$. In this case the AH and MTT both tend to merge at future infinity, arriving at the infinity `point P' of the type shown in Figure \ref{fig:kotext}. In this case, P is the only remaining vestige of the existence of future infinity, and there are some very special observers that can actually reach there. Therefore, for this special class of observers one can still define an EH. \section{Second type of models}\label{sec:second} The second type of models I am going to consider consist of BHs already formed (or better said in formation) by stellar collapse that, after having settled down to equilibrium, receive further matter that makes them grow beyond the limit \eqref{lim}. For illustration purposes I am going to use the generalization of the Oppenheimer-Snyder collapse \cite{OS,BJS} to the case with $\Lambda >0$ analyzed in \cite{GV,L,MS} years ago. Nevertheless, the construction of our models work for any other collapse that produces a Kottler (Schwarzschild-de Sitter) BH, such as those studied in \cite{DJCJ}. \begin{figure}[h!] \includegraphics[width=14cm]{OSdS.pdf} \caption{\footnotesize{Conformal diagram of the formation of a BH by collapsing homogeneous dust in the presence of $\Lambda >0$, same conventions as before (now the centre of the dust is the line $\chi =0$). One starts with initial conditions at an instant of time symmetry, a spacelike hypersurface with initial zero velocity $da/d\tau =0$ in the dust region. This dust portion is represented here by the shadowed zone. The timelike hypersurface $\chi=\chi_0$ is the boundary of the collapsing dust and the spacetime is matched at this hypersurface to the unique spherically symmetric vacuum exterior, which is Kottler metric with mass parameter $m$ such that \eqref{am}, or equivalently \eqref{am1}, holds. The collapse leads to the existence of a timelike MTT foliated by marginally trapped round spheres, which intersect the dust surface intersects at the time that the dust cloud has an area $4\pi r_-^2$. As usual, assuming that $9m^2 \Lambda <1$ there is also a cosmological horizon with $r=r_+$ and corresponding future null infinity $\mathscr{J}^+$. Hence thee is a BH with EH at $r=r_-$ in the Kottler part, as shown. The marginally trapped round spheres foliating the MTT are not stable in the sense of \cite{AMS}, the stable ones --that necessarily exist \cite{E,AM}-- are those foliating the EH in the Kottler region with area $4\pi r_-^2$. All round spheres in the red zones are future trapped, including the part of the dust cloud above MTT, indicated here with the word `red' not to overlap with the shadow of the dust. The blue region contains past-trapped round spheres. The metric is extendible towards the right. I have represented the null hypersurface $v=v_s$, defined as the limit of advanced times $v$ that reach the dust cloud. Therefore, at any $v>v_s$ one can match this spacetime to a Vaidya-dS metric in the same fashion as in Figure \ref{fig:VdS3} to produce another ultra-massive spacetime where the EH and the BH disappear. This is shown in the next diagram, Figure \ref{fig:OSVdS}.}} \label{fig:OSdS} \end{figure} The metric is described by the matching of the Kottler metric \eqref{kot} to a closed Robertson-Walker metric ($0<\chi<\pi$) \begin{equation} ds^2 =-d\tau^2 +a^2(\tau) \left(d\chi ^2 +\sin^2\chi d\Omega^2 \right) \end{equation} where the scale factor solves the Fridman-Lema\^\i tre equation for dust (i.e., pressure $p=0$) \begin{equation}\label{FLeq} \left(\frac{da}{d\tau}\right)^2 =\frac{a_m}{a} +\frac{\Lambda}{3} a^2 -1 . \end{equation} Here $a_m$ is a constant that represents the minimum value of the dust mass density and which, via the matching, can be related to the exterior (Kottler) constant mass parameter $m$ by \begin{equation}\label{am} a_m \sin^3\chi_0 =2m \end{equation} and the matching hypersurface is defined by $$ \chi =\chi_0 , \hspace{1cm} r=r_\Sigma = a(\tau) \sin\chi_0 $$ in the interior and exterior parts, respectively. The constant $\chi_0$ is choosable in principle, and one can easily check on using \eqref{am} the relation \begin{equation}\label{am1} \frac{4\pi}{3} \rho r_\Sigma^3 = \frac{c^2}{G} m (=M) \end{equation} where $\rho$ is the mass density of the dust cloud. The righthand side of equation \eqref{FLeq} never vanishes if $9m^2 \Lambda >1$, connecting with the conditions on the exterior for the absence of Killing horizons.\footnote{This situation was considered in \cite{MS} in their section V, and the exterior is of type \eqref{kot3} leading to a conformal diagram which has a portion of the Figure \ref{fig:kot2} as the exterior part. However, this was not presented in \cite{MS}, as the authors chose to place a second collapsing dust to the right of the diagram---see their figure 10. Nevertheless, these models do not have any MTS anywhere, and therefore they are not of interest in the present discussion.} However, I am going to consider the other possibility in which an MTT arises and a BH is formed. Thus choosing $9m^2 \Lambda <1$ there always exist values of $\chi_0 <\pi/2$ such that there is a bouncing time where $d a/d\tau=0$. This time is usually taken as the initial time of the dust collapse \cite{OS,MS}. The collapse leads to a BH of Kottler type and total mass $M$. The conformal diagram is presented in Figure \ref{fig:OSdS}. \begin{figure}[h!] \includegraphics[width=16cm]{OSVaidyadS.pdf} \caption{\footnotesize{Conformal diagram of a spacetime where collapsing homogeneous dust creates a temporary would-be BH that eventually is frustrated by the reception of extra matter in the form of null dust, such that $9\mu^2 \Lambda >1$. The dust could is represented by the shadowed zone to the left of the timelike hypersurface $\chi =\chi_0$ that represents the surface of the collapsing star. All conventions are as before. The part of this diagram to the left of $v=v_3> v_s$ is identical to that part of the diagram in Figure \ref{fig:OSdS}, while the portion to the right of $v=v_3 >v_s$ is the same as the corresponding portion in Figure \ref{fig:VdS3}. Hence, this is just another example of a frustrated BH, that might have been settled down and in equilibrium for very long ages, but eventually transforms into an ultra-massive spacetime with no $\mathscr{J}^+$ and no EH, producing a universal future singularity.}} \label{fig:OSVdS} \end{figure} Now, the idea is to throw matter into this `already formed' black hole in order to create an ultra-massive spacetime, so that no BH remains. To do that, one simply has to throw enough matter into the BH. This can be easily accomplished by using, as in previous cases, the Vaidya-dS metric starting at any $v> v_s$, where $v_s$ is defined as the limit of the advance times $v$ that reach the dust cloud (so that the entire matching is performed in the Kottler part, for simplicity). If the final total mass parameter $\mu$ is larger that $1/(3\sqrt{\Lambda})$ we again encounter a situation where something that looks like a BH in equilibrium for a long time, and was {\em formed by stellar collapse}, eventually becomes an ultra-massive spacetime with no $\mathscr{J}^+$ and no event horizon. Of course, these are idealized simple models, but the conclusion is robust in spherical symmetry\footnote{This follows from the uniqueness results of the Oppenheimer-Snyder-like models in spherical symmetry, which themselves follow, via the idea of complementary matchings \cite{FST}, from the uniqueness of the Einstein-Straus vacuoles, see \cite{MMV} for details.}: if one tries to increase the area of a MTS beyond the limit \eqref{lim} by throwing matter into its interior the outcome will be the end of the stable MTSs as there will be one with the maximum area \eqref{lim}. This entails the dematerialization of the EH and of $\mathscr{J}^+$, implying a general collapse into a future universal curvature singularity. I conjecture that the conclusion still holds without spherical symmetry, for instance, using the results in \cite{BP,PS} where the Robinson-Trautman metrics with $\Lambda$ are seen to possess properties similar to Vaidya-dS, and actually they all approach Vaidya-dS asymptotically. \section{Discussion}\label{sec:discussion} As we have seen, the limit \eqref{lim} is not violated in any of the models, even when increasing the total mass of the spacetime. Somehow, General Relativity is prepared to accept as much mass as one can imagine, nevertheless stable marginally trapped surfaces cannot increase its area indefinitely if there is a positive $\Lambda$. They simply approach an MTS with the maximum area that ceases to be stable. This can be better understood by noticing that AH, where the marginally trapped spheres are stable in some spacelike outward directions\footnote{The outward direction here is the null future direction with vanishing expansion, or equivalently the direction into which the (null) mean curvature vector of the MTS points \cite{S,BeS}} merges with an MTT where the marginally trapped spheres are not stable in any spacelike direction; and they merge becoming tangential to a null hypersurface. Therefore, the special round sphere where they merge cannot be deformed outwardly in any non-timelike direction without becoming a (weakly) trapped surface. It must be observed that the results of persistence of stable MTS \cite{AMMS} are not in conflict with the models we have presented because those results require the existence of an exterior untrapped barrier \cite{AM,E}, leading to stability, and this is precisely what is missing at the special MTS where AH and MTT merge, as there are no untrapped external spheres whatsoever. An important puzzling question that arises is that of BH evaporation via Hawking radiation. The usual picture cannot be applied here as there is no EH defining the black hole. Of course, it has long been argued that the origin of Hawking radiation may have a different origin \cite{Kod} such as dynamical horizons or marginally trapped tubes of the type AH and MTT \cite{Haj}. One can even argue that some kind of radiation can be associated to {\it any} MTS \cite{ST}. However, in the ultra-massive models herein presented, the question is where does any such radiation go. There is no infinity that allows the system to radiate (lose) energy away, and thus the already infinite curvatures at the singularity will become even larger if some energy arrives there from somewhere else. How quantum gravity might resolve this puzzle is uncertain. These results have also some implications on how to deal with BHs mergers and how to use numerical relativity to describe them. Because there seems to be a limit for the merger of apparent horizons and, if this limit is surpassed, the outerly stable MTS simply fail to exist: no numerical code will ever find them. Of course, there is also the query of how much mass is necessary to produce such ultra-massive spacetimes, and this depends on the value of the cosmological constant. If we accept the value that arises from the observed accelerated expansion of the visible Universe, which is about\cite{Planck} $$ \Lambda \simeq 1.1 \times 10^{-52} \mbox{m} , $$ the limit \eqref{lim} requires a gravitational radius $2m$ that should be greater than $$ 6.4 \times 10^{25} \, \mbox{m} $$ and this translates into a total mass of about $$ 2.2\times 10^{22} M_\odot\sim 2.97 \times 10^{52} \mbox{Kg} . $$ The estimated total mass of the observable universe now is about $$ 8.8\times 10^{52} - 1.0 \times 10^{54} \mbox{Kg} $$ so that one would need at least 3 times the total observed mass {\em now} to produce such ultra-massive objects. It does not seem they are going to be seen in the forseeable future! Still, the total mass of the entire Universe may well be much larger than that, hence these objects might be real somewhere, some time. And, in any case, there is a question of principle: if they may exist for any value of $\Lambda >0$, what is the relevant physics behind them and how to deal with the universal singularity? Finally, I would like to add a remark. The time reversals of ultra-massive spacetimes are also worth considering. One just has to look at the diagram upside down, so that the future direction is reversed. Then, for instance, the time reversal of the model represented in Figure \ref{fig:VdS2} will describe a universal big-bang singularity in the past and an expanding Universe of locally rotationally symmetric (Kantowski-Sachs) type \cite{Exact} in the Kottler region, but the mass-energy creating the gravitational field is radiated away towards $\mathscr{J}^+$ entirely, leaving behind a portion of de Sitter vacuum spacetime. And the model of Figure \ref{fig:OSVdS} will have two expanding regions coming from the big-bang singularity, one of them of FLRW type. This may lead to several interesting speculations. \section*{Acknowledgments} Research supported by the Basque Government grant number IT1628-22, and by Grant FIS2017-85076-P funded by the Spanish MCIN/AEI/10.13039/501100011033 together with ``ERDF A way of making Europe”. This research was carried out during a visitor professorship at Yukawa Institute for Theoretical Physics. \section*{Appendix: The Kottler metrics} \begin{figure}[h!] \includegraphics[width=10cm]{Kottler.pdf} \caption{{\footnotesize Conformal diagram of the Kottler metric covered by the coordinates in \eqref{kot2} when $9 m^2 \Lambda <1$. Radial null geodesics are at 45$^o$ and future is upwards. Each point in the diagram represents a round sphere of area $4\pi r^2$. The region near past infinity $\mathscr{J}^-$ is (close to) de Sitter, while the region close to the future singularity is similar to that part of Schwarzschild. The spheres in the red regions are future-trapped, while the white zone between $r_+$ and $r_-$ is static --and thus free of compact trapped surfaces \cite{MaSe}. The dotted lines with $r=r_\pm$ are Killing horizons foliated by marginally trapped round spheres. The metric is extendible across $v=\pm \infty$, where the geodesics arrive with $r\longrightarrow r_\pm$ respectively. The analytical extension can be easily built by gluing copies of this patch and its time reversal in the appropriate way, as shown in Figure \ref{fig:kot1}.}} \label{fig:kot} \end{figure} \begin{figure}[h!] \includegraphics[width=10cm]{dSpartial.pdf} \caption{{\footnotesize Conformal diagram of the portion of de Sitter spacetime covered by the coordinates in \eqref{kot2} with $m=0$, same conventions as in Fig.\ref{fig:kot} except that now $r=0$ is regular and can be seen as the origin of coordinates. The spheres in the red regions are future-trapped. The dotted line with $r=\sqrt{3/\Lambda}$ is a Killing horizon foliated by marginally trapped round spheres. The metric is extendible across $v= \infty$, where the geodesics arrive with $r\longrightarrow \sqrt{3/\Lambda}$. The analytical extension leads to the standard `square' diagram of de Sitter, shown here by the dashed lines. In the complete diagram the left ($r=0$) and right vertical lines represent the north and south pole of a 3-sphere. The topology of $\mathscr{J}^-$ is, for the case of the portion covered by the coordinates in \eqref{kot2}, $\mathbb{R}\times \mathbb{S}^2$ --the two extremes of the segment representing $\mathscr{J}^-$ are not part of the diagram. The slices of the spacetime have topology $\mathbb{R}^3$ if one includes the point $r=0$ at each instant of time.}} \label{fig:dS} \end{figure} The unique family of spherically symmetric vacuum solutions of the Einstein field equations --including a cosmological constant $\Lambda$-- is given by the Kottler (also known as Schwarzschild-de Sitter) metric \cite{Kot,Exact} \begin{equation}\label{kot} ds^2 = -c^2 \left(1-\frac{2m}{r}-\frac{\Lambda}{3} r^2\right) dT^2 + \left(1-\frac{2m}{r}-\frac{\Lambda}{3} r^2\right)^{-1} dr^2 +r^2 d\Omega^2 \end{equation} where $d\Omega^2$ is the standard metric of the unit round sphere, $r$ is the areal coordinate and $T$ is a fourth coordinate with range in $(-\infty,\infty)$. The `mass parameter' is given by $m:=GM/c^2$ where $M$ is interpreted as the total mass generating the spacetime. When $m=0$ it reduces to a (static) portion of de Sitter (dS) spacetime. In all cases, when $r\rightarrow \infty$, the metric tends to dS. \begin{figure}[h!] \includegraphics[width=16cm]{Kottler2.pdf} \caption{{\footnotesize Conformal diagram of the extended Kottler metric with $9m^2 \Lambda <1$, same conventions as in Fig.\ref{fig:kot}. The spheres in the red regions are future-trapped, and those in the blue regions are past-trapped, while the white zones between $r_+$ and $r_-$ are static --and thus free of compact trapped surfaces. The dotted lines with $r=r_\pm$ are Killing horizons foliated by marginally trapped round spheres. The blue and red regions on the left thus represent white and black hole regions, respectively, with $r=r_-$ as the hole bifurcate horizon \cite{Wald}. The red and blue regions on the right represent past and future cosmological zones approaching, at past and future infinity respectively, de Sitter spacetime. The metric is analytically extendible towards the left and the right by just adding copies of the same diagram, leading to many, or infinite, BH regions. The ``points'' where the singularities and infinities ``touch'' are not part of the diagram, and the topology of $\mathscr{J}$ is $\mathbb{R}\times \mathbb{S}^2$. This is also the topology of space sections of the spacetime (say horizontal lines in the diagram). An alternative is to identify the left and right vertical lines, which will produce just one BH and one asymptotic region, then changing the topology of the space sections to $\mathbb{S}\times \mathbb{S}^2$.}} \label{fig:kot1} \end{figure} As is well known, when $\Lambda >0$ there are three different possibilities for the metric \eqref{kot} depending on whether $9m^2$ is greater, equal, or smaller than $1/\Lambda$. The standard case, which includes a static region similar to that of the Schwarzschild metric, requires $$ \Lambda < \frac{1}{9m^2} $$ in which case the function $1-\frac{2m}{r}-\frac{\Lambda}{3} r^2$ has two positive zeros, $r_+$ and $r_-$ say. The particular values of $r_\pm$ can be found in \cite{LR,MS}. The two hypersurfaces defined by $r=r_\pm$ can be easily proven to define Killing horizons \cite{Wald} of the Killing vector $\partial_T$ through which the metric \eqref{kot} can be extended via the usual techniques. For instance, by using the advanced null coordinate $v$ defined by \begin{equation}\label{v} dv =cdT+\left(1-\frac{2m}{r}-\frac{\Lambda}{3} r^2\right)^{-1} dr \end{equation} the (extended) metric becomes \begin{equation}\label{kot2} ds^2 = -\left(1-\frac{2m}{r}-\frac{\Lambda}{3} r^2\right) dv^2 +2dv dr +r^2 d\Omega^2 \end{equation} with $r\in (0,\infty)$ (alternatively $r\in (-\infty,0)$). The hypersurface $r=r_+$ represents a cosmological horizon, and that with $r=r_-$ a black hole horizon, both of them null hypersurfaces foliated by marginally trapped round spheres. One has $$ 0< r_-< \frac{1}{\sqrt{\Lambda}} < r_+ < \sqrt{\frac{3}{\Lambda}} $$ the static region given by $r_- < r < r_+$, while the round spheres with $r<r_-$ and those with $r>r_+$ are trapped --see \cite{S,SMilestone} for the terminology and \cite{GH,GP} for further details. Thus, $r$ is a time coordinate in those two regions. The spacetime contains a curvature singularity at $r=0$ unless $m=0$. The conformal diagram of this spacetime is shown in Figures \ref{fig:kot}, \ref{fig:kot1} and \ref{fig:dS} \cite{LR,GH,GP}. \begin{figure}[h!] \includegraphics[width=10cm]{KottlerExtrem.pdf} \caption{{\footnotesize Conformal diagram of the Kottler metric with $9m^2\Lambda=1$ for the portion covered by the coordinates of \eqref{kot2}, same conventions as before. All round spheres are future trapped except those on the blue dotted line, which is a degenerate Killing horizon foliated by marginally trapped spheres. Thus, not to over-red the picture, I have signaled the `red' regions with the word `red'. There is a curvature singularity in the future and past null infinity $\mathscr{J}^-$. The `points' $P$ and $Q$ marked in the diagram are actually special regions at infinity, so that the null generators of the degenerate horizon are past and future complete. The metric is analytically extendible towards the left and the right by just adding copies of the same diagram, leading to many, or infinite, BH regions. (Again, there is the alternative of making identifications). The topology of $\mathscr{J}^-$ is $\mathbb{R}\times \mathbb{S}^2$. Observers starting from $\mathscr{J}^-$ can either end up at the singularity or try to avoid this by reaching $P$ with $r\longrightarrow 3m=1/\sqrt{\Lambda}$ as proper time goes to infinity. However, very few observers in free fall (geodesics) reach $P$ \cite{P}.}} \label{fig:kotext} \end{figure} The limiting case is when $$ \Lambda = \frac{1}{9m^2} $$ in which case there is only one (double) positive zero of the function $1-\frac{2m}{r}-\frac{r^2}{27m^2} $ given by $$ r = \frac{1}{\sqrt{\Lambda}} = 3m . $$ Now, the round spheres with constant $v$ and $r$ are always untrapped except for those with $r = \frac{1}{\sqrt{\Lambda}} = 3m $ which are marginally trapped. The hypersurface $r= \frac{1}{\sqrt{\Lambda}} = 3m $ is a degenerate Killing horizon, and infinity is only reachable for a tiny set of privileged observers \cite{P} ---a subset of the causal geodesics with $T=$const.\ in the original coordinates \eqref{kot} plus the lightlike geodesics on the horizon with $r=3m$. The global structure and general properties of this case were thoroughly analyzed in \cite{P}. The conformal diagram is shown in Figure \ref{fig:kotext} \begin{figure}[h!] \includegraphics[width=10cm]{KottlerRaro.pdf} \caption{{\footnotesize Conformal diagram of the Kottler metric with $9m^2\Lambda>1$ same conventions as before. In this case the original coordinates cover the entire spacetime. It is more visual to call $r \rightarrow t$, as in \eqref{kot3}, because now $r$ is a time coordinate everywhere. In the diagram $t$ decreases towards the future. All round spheres are now future trapped and there are no horizons whatsoever, so the entire diagram is now `red'. The spacetime represents a collapsing universe, with space topology $\mathbb{R}\times \mathbb{S}^2$. There is a curvature singularity in the future and past null infinity $\mathscr{J}^-$. There is no escape of the singularity, all observers will eventually end up there, so that there is no future null infinity nor asymptotic regions of any kind to the future. The metric is inextendible. The topology of $\mathscr{J}^-$ is $\mathbb{R}\times \mathbb{S}^2$. The `points' where the singularity and $\mathscr{J}^-$ touch are not part of the diagram.}} \label{fig:kot2} \end{figure} Finally, there is the case with $$ \Lambda > \frac{1}{9m^2} $$ so that in this situation the function $1-\frac{2m}{r}-\frac{\Lambda}{3} r^2$ has no real roots. In this case, there are no horizons and actually the original coordinates of \eqref{kot} cover the entire spacetime. However, as now the coordinate $r$ is a time everywhere, one should better write \eqref{kot} by renaming the coordinates so that is visually clearer: \begin{equation}\label{kot3} ds^2 = -\left(\frac{\Lambda}{3} t^2 +\frac{2m}{t} -1\right)^{-1} dt^2 +\left(\frac{\Lambda}{3} t^2 +\frac{2m}{t} -1\right) dX^2 +t^2d\Omega^2 . \end{equation} Notice that $\left(\frac{\Lambda}{3} t^2 +\frac{2m}{t} -1\right)$ is positive everywhere, $t\in(0,\infty)$ and, for compatibility with previous cases, I asusme that $-\partial_t$ is future pointing. Thus, there is a future curvature singularity at $t=0$. The spacetime represents a {\em locally rotationally symmetric} vacuum cosmological contracting model of `Kantowski-Sachs' type \cite{Exact}, included in the general solution first found in \cite{CD}. The metric is inextendible, and the conformal diagram \cite{G} is shown in Figure \ref{fig:kot2}.

0df23755c79c63e50e53a8534182aea239a43b06

\section{Introduction} Real-time object detection algorithms are getting a lot of attention in academia and industry for the broad set of applications that employ them, including autonomous driving and factory automation~\cite{its1,its5,its6,its7,its8}. Among those algorithms, the multi-object detection algorithm is widely used. The fundamental ideas of the algorithm mentioned above consist of two procedures; (i) \textit{bounding box} and \textit{confidence} are used for determining whether objects exist or not; then (ii) the concept of \textit{class probability map} is used for the classification of the detected objects~\cite{yolo}. The multi-object detection algorithms are based on convolutional neural networks (CNN). Nowadays, sophisticated algorithms are proposed to improve the accuracy of the network, such as YOLOv4 and DyHead~\cite{ YOLOv4, DyHead}. However, there is a tradeoff between inference time and accuracy. The object detection algorithm with many convolution layers performs better than others while increasing the computation time. In contrast, the simple network that requires less computation is suitable for real-time object detection applications while sacrificing specific detection accuracy. For autonomous driving applications, the major learning-based system research topic, it is necessary to ensure the system can use and relate to the driving environment. The driver can record the environment in a real-time video composed of continuous image arrivals. To use the continuous image arrivals, \textit{optical flow} is one of the best ways to model and process sequential continuous image arrivals~\cite{opticalflow}. There are two types of optical flow; (i) One is a sparse optical flow that presents a partial motion of pixels. (ii) The other is a dense optical flow that indicates the full motion of pixels. Dense optical flow has higher accuracy than sparse optical flow, while the speed is lower. In the recent decade, the neural network application on calculating dense optical flow (\textit{e.g.}, FlowNet) emerges to obtain a dense flow map with guarantying high accuracy and low speed~\cite{flownet}. This network is a CNN-based neural network that receives two consecutive images of a video as an input and returns the information of pixel displacement as an output. The optical flow estimation networks (OFEN), similar to object detection, have a tradeoff between the computation time (\textit{i.e.}, delay) and accuracy. Motivated by the \textit{``continuity"} property of videos in the object detection research domain, this paper investigates object detection and optical flow consolidation for the driving environment. As shown in Fig.~\ref{fig:flo_description}(a)--(c), we empirically find that leveraging optical flow can be a solution for improving the existing object detection networks (ODN). However, applying an optical flow on object detection directly is challenging due to following reasons. First, the driving environment is generally uncertain, \textit{e.g.}, the vehicle can be unexpectedly driving or stopping. Second, the driver state is time-varying. If the observer is static, the moving object is detected by optical flow. On the other hand, and according to Fig. \ref{fig:flo_description}(c), \textit{i.e.}, if the observer is moving, the moving objects may appear static when moving in the same direction. Additionally, we consider the computing capacities corresponding to object detection in the driving environment. The received images are time-varying. Thus, the number of objects in the changes every moment. However, the number of objects is proportional to the inference time of an object detection network~\cite{num_object}. In light of the issues above, we propose a novel real-time object detection model called \textit{Hybrid} by leveraging optical flow. Then, we design a self-configurable stabilized framework with Lyapunov optimization considering computational overheads \cite{book2010sno}. Our framework aims to guarantee time-average object detection performance maximization by deciding whether to use optical flow considering computation time (delay) based on the driving environment. \BfPara{Contributions} The key contributions of the proposed algorithm in this paper are as follows. \begin{itemize} \item We first propose a fusion of object detection and optical flow. We further suggest a novel flow map processing algorithm (see \textbf{Algorithm~\ref{alg:confidencemask}}), which is suitable for a time-varying driving environment. \item We propose a novel self-configurable framework for autonomous driving. Our architecture provides a time-average sequential optimal decision-making under the tradeoff between performance and delay. Furthermore, one of the main advantages of the Lyapunov optimization-based algorithm is low-complexity operation (see \textbf{Algorithm~\ref{alg:dpp2}}). Thus, our proposed algorithm is suitable for real-time computation in a fast-moving autonomous driving environment. \end{itemize} \BfPara{Organization} The rest of this paper is organized as follows. Sec.~\ref{sec:2} proposes a stabilized real-time object detection adaptation algorithm for autonomous driving. Sec.~\ref{sec:3} evaluates the performance of the proposed algorithm. Sec.~\ref{sec:4} concludes this paper and presents future research directions. The notations used in this paper are listed in Table~\ref{tab:notation}. \begin{figure}[t!] \centering \setlength{\tabcolsep}{2pt} \renewcommand{\arraystretch}{0.2} \begin{tabular}{cc} \includegraphics[page=1, width=0.45\linewidth]{fig1a.png} & \includegraphics[page=1, width=0.45\linewidth]{fig1b.png} \tabularnewline \tabularnewline \footnotesize (a) Input image & \footnotesize (b) Dense optical flow field \tabularnewline \tabularnewline \includegraphics[page=1, width=0.45\linewidth]{fig1c.png} & \includegraphics[page=1, width=0.45\linewidth]{fig1d.png} \tabularnewline \tabularnewline \footnotesize (c) Flow map & \footnotesize (d) The processed flow map \textbf{(Ours)} \end{tabular} \footnotesize \caption{A snapshot of optical flow in an actual driving environment.} \label{fig:flo_description} \end{figure} \begin{table}[t!] \caption{List of Notations} \label{tab:notation} \centering \footnotesize \begin{tabular}{c|l} \toprule[1pt] \textbf{Symbol} & \textbf{Description}\\\midrule[1pt] $K$ & The number of bounding boxes.\\ $\mathbf{F}_{M,N}$ & $M \times N$-sized flow map.\\ $\mathbf{B}_{M,N,K}$ & $K$ bounding boxes of size $M \times N$.\\ $Q$ & A queue-backlog.\\ $a[t]$ & An arrival process at time $t$.\\ $H$ & A hybrid model leveraging optical flow.\\ $T$ & An object detection model.\\ $\alpha[t]$ & A detection model, $\forall \alpha[t] \in \mathcal{A} \equiv \{H,T\}$.\\ $b(\alpha[t])$ & A service process with $\alpha[t]$ at time $t$.\\ $P(\alpha[t])$ & The object detection accuracy.\\ \bottomrule[1pt] \end{tabular} \end{table} \begin{figure*}[t!] \includegraphics[width=1.0\linewidth]{fig2.pdf} \caption{The system model consists of two components, \textit{i.e.}, (1) \textit{Hybrid model}, which utilizes optical flow estimation in object detection model, and (2) \textit{Model selection} which is based on Lyapunov optimization. Our proposed method determines whether the model uses optical flow estimation and flow map processing or not.} \label{fig:systemmodel} \end{figure*} \section{Stabilized Real-Time Object Detection for Autonomous Driving Applications}\label{sec:2} This section introduces our proposed stabilized real-time object detection for autonomous driving, consisting of two parts, \textit{i.e.,} the Hybrid model in Sec.~\ref{sec:2-1} and model selection in Sec.~\ref{sec:2-2}. Fig.~\ref{fig:systemmodel} briefly illustrates our proposed framework. \subsection{Hybrid Model}\label{sec:2-1} \subsubsection{Object Detection \& Optical Flow Estimation in the Nutshell} We present the object detection and optical flow estimation in the nutshell. ODN takes an image as an input and returns bounding boxes $\mathbf{B}_{M,N,K}$ where the component represents confidence score $c_{i,j,k} \in [0, 1]$: $\forall i \in \mathbb{N}[1,M]$, $\forall j \in \mathbb{N}[1,N]$, $\forall k \in \mathbb{N}[1,K]$. Among all bounding boxes, the highest scored bounding box of which the confidence score exceeds the confidence threshold, is regarded as the object detected~\cite{yolo}. As shown in Fig.~\ref{fig:flo_description}(a), the dense optical flow field is obtained by calculating pixel displacement of two consecutive images via OFEN. As shown in Fig.~\ref{fig:flo_description}(b), The flow map $\mathbf{F}_{M, N}$ indicates the magnitude of dense optical flow field, which is a matrix with the size of $ M \times N $ written as follows: \begin{equation} \mathbf{F}_{M,N} = \begin{pmatrix} e_{1,1} & e_{1,2} & \cdots & e_{1,N} \\ \vdots & \vdots & \ddots & \vdots \\ e_{M,1} & e_{M,2} & \cdots & e_{M,N} \end{pmatrix}, \end{equation} where $e_{i,j}$ stands for the magnitude of pixel motion of position $(i,j)$. Note that each element $e_{i,j}$ has a real value where $\forall e_{i,j} \in (-\infty, + \infty)$: $\forall i \in \mathbb{N}[1, M]$ and $\forall j \in \mathbb{N}[1, N]$. \subsubsection{Observation from Optical Flow According to Fig.~\ref{fig:flo_description}(c), the flow map has a linear increase/decrease pattern in each object (\textit{e.g.,} car or truck) and its background. Our initial insight is that if flow map and confidence are used as an additional condition for determining the existence of an object, we will be able to perform better than the \textit{de facto} ODN. Suppose that the confidence value of ODN is small, but there is an object detected in the flow map. If so, lowering the confidence threshold for the cell makes it possible to detect objects that ODN cannot detect. To enhance the performance of ODN, we design \textit{Hybrid}, which is the combination of optical flow and object detection. We elaborate on utilizing flow map into object detection next. \subsubsection{Flow Map Processing} This subsection introduces the flow map processing algorithm to apply the Hybrid to the road driving environment. All elements of $\mathbf{F}_{M,N}$ are in $(-\infty, + \infty)$, thus $\mathbf{F}_{M,N}$ should be min-max normalized. \begin{equation} \label{eq:normalize} \mathbf{F}_{M,N} \leftarrow \{ \mathbf{F}_{M,N} - e_{\min}\cdot\textbf{1}_{M,N} \}, \end{equation} where $\textbf{1}_{M,N}$ stands for $M\times N$-sized matrix of ones, and $e_{\min}$ denotes the minimum value of $\forall e_{i,j}$. The closer the values at both ends, \textit{i.e.}, $e_{\max}, e_{\min}$, the more dynamic pixel information is present, and the central value is static pixel information. As $e_{i,j}$ approaches the maximum value of $\mathbf{F}_{M,N}$, $e_{\max}$, or the minimum of $\mathbf{F}_{M,N}$, $e_{\min}$, dynamic pixel information exists. As it approaches the median value of $\mathbf{F}_{M,N}$, $e_{\mathrm{median}}$, static pixel information exists. This behavior is expressed as follows: \begin{equation} \label{eq:normalize2} \mathbf{F}_{M,N} \leftarrow |\mathbf{F}_{M,N} - e_{\mathrm{median}} \cdot\textbf{1}_{M,N} | \end{equation} After this procedure, the information of the moving pixel comes out closer to $e_{\max}$, and the information of the static pixel comes out closer to $e_{\min}$. Since the purpose of this algorithm is to consider only the moving pixel information, we present the process of removing the static pixel information; \begin{eqnarray} \mathbf{F}_{M,N} &\leftarrow& \frac{1}{1+\mathrm{exp}(-\mathbf{F}_{M,N})}. \label{eq:classify} \end{eqnarray} Then, flatten $\mathbf{F}_{M,N}$ to $\mathbf{f}_{MN}$, and duplicate $\mathbf{f_{MN}}$ as much as the number of bounding boxes $K$. Eventually, the vectorized flow map denoted as $\mathbf{f}_{MNK}$, is applied to the confidence threshold as follows: \begin{equation} \label{eq:vectorized_cth} \textbf{c}_{th} = \frac{c_{th} }{1+\mathrm{exp}\left(2\cdot \mathbf{f}_{MNK} \right)} \end{equation} where $\textbf{c}_{th}$, and $c_{th}$ stand for the vectorized confidence threshold, and the scalar confidence threshold, respectively. Finally, leveraging $\textbf{c}_{th}$ as a criterion for object detection, ODN utilizes optical flow. \begin{algorithm}[t!] \caption{Flow Map Processing} \label{alg:confidencemask} \small \begin{algorithmic}[1] \Statex $\hspace{-1.5em}\textbf{Input:}$ $\mathbf{F}_{M,N}$ // $M \times N$ flow map matrix \Statex $\hspace{-1.5em}\textbf{Output:}$ $\textbf{c}_{th}$ // ${M \times N \times K}$-sized confidence threshold vector \For{$\forall e_{i,j}$ in $\mathbf{F}_{M,N}$} \State $e_{i,j} \leftarrow e_{i,j}-\min(\mathbf{F}_{M,N})$; \EndFor \For{$\forall e_{i,j}$ in $\mathbf{F}_{M,N}$} \State $e_{i,j} \leftarrow |e_{i,j}-\textsf{median}(\mathbf{F}_{M,N})|$; \EndFor \For{$\forall e_{i,j}$ in $\mathbf{F}_{M,N}$} \State $e_{i,j} \leftarrow \textsf{sigmoid}(e_{i,j})$; // Equation 4 \EndFor \State Flatten $\mathbf{F}_{M,N}$ into vector $\mathbf{f}_{MN}$; \State Replicate $\mathbf{f}_{MN}$ $K$ times, then make $\mathbf{f}_{MNK}$; \State $\textbf{c}_{th} \leftarrow \frac{c_{th}}{1+\mathrm{exp} (2 \cdot \mathbf{f}_{MNK})}$; \end{algorithmic} \end{algorithm} The pseudo-code of the proposed flow map processing algorithm is presented in \textbf{Algorithm~\ref{alg:confidencemask}}. From (line 1) to (line 3), the elements of flow map $\mathbf{F}_{M,N}$ are normalized. From (line 4) to (line 9), the component of the static object in $\mathbf{F}_{M,N}$ is diminished. In (line 10), $\mathbf{F}_{M,N}$ is scaled to $\mathbf{F}_{S,S}$ using the bicubic interpolation~\cite{bicubic}. From (line 11) to (line 12), the processed flow is converted to the same size as the output size of YOLOv3-tiny. From (line 13) to (line 14), the process of making vectorized confidence threshold $\textbf{c}_{th}$ is expressed using \eqref{eq:vectorized_cth}. The computational complexity of \textbf{Algorithm~\ref{alg:confidencemask}} is $O(MN)$ for flow map with the size $M\times N$. \subsection{Lyapunov Optimization-based Model Selection}\label{sec:2-2} In the driving environment, a driving state and a stationary state exist. If the optical flow is used while driving, a Hybrid can improve the performance. However, when the optical flow is used in the stationary state, the pixel displacement does not exist, and the time to calculate it is wasted. Also, in the case of many objects, for example, there is the possibility of being pushed out in real-time suitability because the computation amount increases proportionally. In other words, Hybrid and ODN have a tradeoff between the computation time (\textit{i.e.}, delay) and object detection accuracy. Therefore, the Lyapunov optimization framework is designed to increase the system's stability by observing the queues and performance of the two networks and making the right decision. \subsubsection{Lyapunov Optimization Framework} This section introduces our proposed Lyapunov optimization framework, aiming at time-average detection performance maximization subject to system stability. We can design a time-average optimization framework considering stability by stabilizing the drift. In this case, and according to the Lyapunov optimization framework, the delay can be modeled by queue, where Lyapunov drifts can then again model the queue dynamics. By observing the queue-backlog and performance of every frame, the framework can use the Lyapunov optimization to select the next step deep learning object detection model, which is a sequential time-average optimal decision-making. The queue dynamics in the system $Q[t]$ are characterized as follows: \begin{equation} Q[t+1] \triangleq \max\{Q[t]+a[t]-b(\alpha[t]),0\}, \label{eq:queue} \end{equation} where $Q[t]$ is a queue-backlog size at time $t$ where $Q[0] = 0$ and $a[t]$ is an arrival process at $Q[t]$ at $t$. This arrival process is the received video streams in the system (i.i.d. random events). In \eqref{eq:queue}, $b(\alpha[t])$ is a service process at $Q[t]$ when our model selection decision is $\alpha[t]$ at $t$. With the Hybrid, the processing of $Q[t]$ will be relatively smaller compared to the case where ODN is used. The computational cost for ODN and Hybrid is used as the weights, as follows: \begin{equation} b(\alpha [t])= \begin{cases} w_1 , & \mbox{if }\alpha[t]=H \\ w_2 , & \mbox{if }\alpha[t]=T \end{cases}, \end{equation} where $H$ stands for the Hybrid and $T$ stands for ODN, respectively. Here, $a[t]$ is modeled as the ratio of fps per cycle to default fps: \begin{equation} a[t] = w_{fps} \cdot p[t] , \end{equation} where $p[t]$ and $w_{fps}$ stand for the time per cycle of the network and default fps, respectively. In addition, the object detection accuracy is defined based on the total number of objects detected, the number found correctly, the number found incorrectly, the number of overlapping objects, and the ratio of truly detected~\cite{performance}. Thus, the performance of the detection model is at time-step $t$ is modeled as follows: \begin{equation} P(\alpha [t])= \begin{cases} w_p \cdot \textsf{num}_H(\textsf{object}), & \mbox{if }\alpha[t]=H \\ \textsf{num}_T(\textsf{object}), & \mbox{if }\alpha[t]=T \end{cases}, \label{eq:exp2} \end{equation} where $w_p$, and $\textsf{num}_{(\cdot)}(\textsf{object})$ stand for the detection accuracy ratio of $H$ and $T$, and the number of detected object, respectively. In our reference system model, the model that processes video streams fast, whereas the object detection accuracy is relatively low, should be used if $Q[t]$ is near overflow. On the other hand, the model which processes video streams with high object detection accuracy while taking more time should be used if $Q[t]$ is near zero. Eventually, we can observe the tradeoff between our objective (\textit{i.e.}, object detection accuracy) and stability. This paper designs an object detection model selection algorithm for the time-average detection accuracy maximization while guaranteeing queue stability. The mathematical program for maximizing the time-average object detection accuracy, \textit{i.e.}, $P(\alpha[t])$, is as follows: \begin{eqnarray} \max: & & \lim_{t\rightarrow\infty}\sum_{\tau=0}^{t-1} P(\alpha[\tau]), \label{eq:opt} \\ \text{subject to} & & \lim_{t\rightarrow\infty}\frac{1}{t}\sum_{\tau=0}^{t-1} Q[\tau]<\infty \text{ (queue stability)}. \end{eqnarray} According to this tradeoff, the Lyapunov optimization theory-based drift-plus-penalty (DPP) algorithm~\cite{tvt2019minseok,ton2016joongheon,tmc2019jonghoe} maximizes the time-average utility subject to queue stability. Here, the Lyapunov function is defined as \begin{equation} L(Q[t]) \triangleq \frac{1}{2}Q^2[t] \end{equation} and $\Delta(.)$, the conditional quadratic Lyapunov function, is defined as \begin{equation} \mathbb{E}[L(Q[t+1])-L(Q[t])| Q[t]] \end{equation} also called the drift on $t$. According to \cite{book2010sno}, this dynamic policy is designed to achieve queue stability by minimizing an upper bound on DPP (\textit{i.e.}, minimizing the negative value of $P(\alpha[t])$), which is given by \begin{equation} \Delta(Q[t]) + V \mathbb{E} \Big[ -P(\alpha[t]) \Big], \end{equation} where $V$ is a tradeoff coefficient. The upper bound on the drift of the Lyapunov function at $t$ is derived as follows: \begin{align} &L(Q[t+1]) - L(Q[t]) = \frac{1}{2}\Big( Q([t+1]^2 - Q[t]^2 \Big) \\ &~\leq \frac{1}{2} \Big( a[t]^2 + b(\alpha[t])^2 \Big) + Q[t] (a[t] - b(\alpha[t])). \end{align} Therefore, the upper bound of the conditional Lyapunov drift can be derived as follows: \begin{align} \Delta(Q(t)) &= \mathbb{E}[L(Q[t+1]) - L(Q[t]) | Q[t]] \nonumber \\ &\leq C + \mathbb{E}\Big[ Q[t](a[t] - b(\alpha[t]) \Big| Q[t] \Big], \end{align} where $C$ is a constant given by \begin{equation} \frac{1}{2}\mathbb{E}\Big[ a[t]^2 + b(\alpha[t])^2 \Big| Q[t] \Big] \leq C, \end{equation} which assumes that the arrival and departure process rates are upper bounded. Due to the fact that $C$ is a constant and the arrival process $a[t]$ is not controllable, minimizing the upper bound on DPP becomes \begin{equation} V \mathbb{E}\Big[ -P(\alpha[t]) \Big] - \mathbb{E}\Big[ Q[t]\cdot b(\alpha[t]) \Big]. \end{equation} Thus, the time-average maximization problem can be re-formulated as \begin{equation} V \mathbb{E}\Big[ P(\alpha[t]) \Big] + \mathbb{E}\Big[ Q[t]\cdot b(\alpha[t]) \Big]. \end{equation} \begin{algorithm}[t] \caption{Stabilized Detection Accuracy Maximization} \label{alg:dpp2} \small \begin{algorithmic}[1] \Statex $\hspace{-1.5em}\textbf{Initialize:}$ $t\leftarrow 0$; $Q[t]\leftarrow 0$; \Statex $\hspace{-1.5em}\textbf{Decision Action:}$ $\forall \alpha[t]\in\mathcal{A}\equiv\{H,T\}$ \Statex $\hspace{-1.5em}\textbf{Stabilized Object Detection Accuracy Maximization:}$ \While{$t\leq T$} // $T$: operation time \State Observe $Q[t]$; \State $\mathcal{T}^{*} \leftarrow \infty$; \For{$\alpha[t]\in \mathcal{A}$} \State $\mathcal{T} \leftarrow V\cdot P(\alpha[t]) + Q[t]b(\alpha[t])$; \If {$\mathcal{T} \leq \mathcal{T}^{*}$} \State $\mathcal{T}^{*}\leftarrow\mathcal{T}$; \State $\alpha^{*}[t]\leftarrow \alpha[t]$; \EndIf \EndFor \EndWhile \end{algorithmic} \end{algorithm} Here, the concept of the maximization of the expectation is used; therefore, this result can be maximized by an algorithm that observes the current queue state $Q[t]$ and determines $\alpha[t]$ at every slot $t$, as follows: \begin{equation} \boxed{ \alpha^{*}[t+1]\leftarrow \arg\max_{\alpha[t]\in\mathcal{A}} \left[ V\cdot P(\alpha[t]) + Q[t]b(\alpha[t]) \right]} \label{eq:lyapunov-final} \end{equation} where $\mathcal{A}\equiv\{H,T\}$ is the set of all possible object detection models, $\alpha^{*}[t]$ is the optimal object detection model selection decision at $t$, and $V$ is the tradeoff coefficient between the processing accuracy and queue stability. To verify whether the \eqref{eq:lyapunov-final} works correctly or not, we provide the following two cases. \begin{itemize} \item \textit{Case 1:} Suppose that $Q[t]\approx \infty$. Then, \eqref{eq:lyapunov-final} tries to maximize $b(\alpha[t])$, thus the processing should be accelerated for satisfying the queue stability, and the object detection model at $t$ ($\alpha[t]$) is selected, which is the fastest one. \item \textit{Case 2:} Suppose that $Q[t] =0$. Then, \eqref{eq:lyapunov-final} tries to maximize $P(\alpha[t])$, thus the algorithm pursues the performance accuracy improvements, and the object detection model at $t$ ($\alpha[t]$) is selected, which is the most accurate one. \end{itemize} \subsubsection{Pseudo-Code and Complexity}\label{sec:2-2B} The pseudo-code of the proposed object detection model selection algorithm is presented in \textbf{Algorithm~\ref{alg:dpp2}}. All variables and parameters are initialized from (line 1) to (line 3). The algorithm works in each unit time as shown in (line 4). In (line 5), the current queue-backlog $Q[t]$ is observed to be used in \eqref{eq:lyapunov-final}. From (line 7) to (line 12), the main computation procedure for \eqref{eq:lyapunov-final} is described. Because our proposed algorithm solves a closed-form equation with the number of decision actions, \textit{i.e.}, the number of elements in $\mathcal{A}$, the run-time computational complexity is only $O(N)$. Thus, it is clear that our algorithm guarantees a low computational complexity. \section{Performance Evaluation}\label{sec:3} This section presents the implementation of Hybrid (refer to Sec.~\ref{sec:3-1}) and the performance of the proposed model selection (refer to Sec.~\ref{sec:3-2}), respectively. Note that the experiment settings are presented in Table~\ref{tab:settings}. Due to the real-time issue, we adopt YOLOv3-tiny and FlowNet2-S as the real-time ODN and OFEN, respectively. The demo video is available in~\cite{youtube}. \subsection{Implementation of Hybrid}\label{sec:3-1} \begin{table}[t!] \footnotesize \centering \caption{Experiment setting.} \label{tab:settings} \begin{tabular}{c|r} \toprule[1pt] \textbf{Specitications} & \textbf{Settings}\\\midrule \multicolumn{1}{l|}{Dataset} & nuScenes-mini~\cite{nuscenes}, 4K Driving~\cite{dataset} \\\midrule \multicolumn{1}{l|}{OS} & Windows 10 Pro \\\midrule \multicolumn{1}{l|}{Processor (CPU \& GPU)} & Intel Xeon E5-2638, Nvidia RTX 2080Ti\\\midrule \multicolumn{1}{l|}{Dev. environment} & Python 3.6, Pytorch 1.4, OpenCV 4.2\\\midrule \multicolumn{1}{l|}{Object detection networks} & YOLOv3/v3-tiny/v4/v7 pretrained \\ & with COCO data\\\midrule \multicolumn{1}{l|}{Optical flow } & FlowNet2/2C/2S pretrained\\ \multicolumn{1}{l|}{estimation networks} & with KITTI data\\ \midrule \multicolumn{1}{l|}{Lyapunov coefficient ($V$)} & 90\\ \midrule \multicolumn{1}{l|}{Confidence threshold ($c_{th}$)} & 0.5 \\ \midrule \multicolumn{1}{l|}{NMS threshold} & 0.2 \\ \midrule \multicolumn{1}{l|}{Weights ($w_1, w_2, w_{fps},w_p$)} & (3.64, 2.41, 30.0, 1.005)\\ \bottomrule[1pt] \end{tabular} \end{table} \begin{table}[t!] \footnotesize \centering \caption{Performance of Hybrid with various models and nuScenes-mini dataset (mAP50).} \label{tab:performance-general} \begin{tabular}{@{}l|cccc@{}} \toprule[1pt] \centering \hspace{1em}\textbf{ODN/OFEN} & w/o. OFEN & FlowNet2S & FlowNet2C & FlowNet2\\\midrule \multicolumn{1}{l|}{YOLOv3-tiny~\cite{yolov3}} & 2.40 & 2.45 & 2.42 & 2.47 \\ \multicolumn{1}{l|}{YOLOv3~\cite{yolov3}} & 13.65 & 13.66 & 13.66 & 14.20 \\ \multicolumn{1}{l|}{YOLOv4~\cite{YOLOv4}} & 15.62 & 15.65 & 15.66 & 16.07 \\ \multicolumn{1}{l|}{YOLOv7~\cite{YOLOv7}} & 14.45 & 14.48 & 14.47 & 14.54 \\ \bottomrule[1pt] \end{tabular} \end{table} We propose three main experiments to verify the performance of Hybrid. First, we investigate the impact of flow map on various ODNs with a public driving dataset (\textit{i.e.}, nuScenes-mini \cite{nuscenes}). Second, we investigate the performance of Hybrid with high-resolution driving dataset (\textit{i.e.}, Youtube 4K driving~\cite{dataset}) by comparing the performance of the real-time ODN (\textit{i.e.,} YOLOv3-tiny) and more complex ODN (\textit{i.e.,} YOLOv3). Finally, we conduct an ablation study of optical flow estimation. We measure the time taken for one cycle and the number of detected objects per image for performance evaluation.\\ \BfPara{Impact of flow map on detector networks} We conduct the performance evaluation to investigate the overall performance. In this paper, we adopt nuScenes-mini dataset for evaluation with various real-time ODN models, \textit{e.g.}, YOLOv3, YOLOv4, YOLOv5, and YOLOv7. In addition, we use various optical flow estimation networks, \textit{e.g.}, FlowNet2, FlowNet2C, and FlowNet2S. All ODN networks and OFEN networks are pretrained with COCO dataset and KITTI dataset, respectively. The overall performance is presented in Tab.~\ref{tab:performance-general}. As shown in Tab.~\ref{tab:performance-general}, all OFEN improve ODN networks. Especially, Hybrid composed of YOLOv3 and FlowNet2 achieves 4.02\% performance gain (\textit{i.e.}, 0.55\% mAP50 score) compared to YOLOv3 only. The state-of-the-art YOLOv7 shows the least performance gain 0.6\% (\textit{i.e.}, 0.09\% mAP50 score). In addition, more complex OFEN (\textit{i.e.}, FlowNet2) shows the highest performance gain, the lightest OFEN (\textit{i.e.}, FlowNet2S) shows the lowest performance gain. In summary, our flow map processing algorithm enhances the performance of various ODNs in the driving environment without training/fine-tuning ODNs nor OFENs.\\ \BfPara{Feasibility study of Hybrid} To figure out the feasibility of Hybrid concerning object detection, we test three models (\textit{i.e.}, Hybrid, YOLOv3-tiny, and YOLOv3) with a road driving dataset which consists of 27k frames~\cite{dataset}. We measure the inference time and number of objects per image. Fig.~\ref{fig:performance} shows the result in the GPU setting. We find a tradeoff between inference time (\textit{i.e.,} delay) and the number of the detected object (\textit{i.e.,} performance). In addition, the total number of detected objects is 46.3k for real-time ODN, 72.9k for Hybrid, and 164k for complex ODN, respectively. On average, the inference time in both GPU and CPU settings is high in the order of complex ODN, Hybrid, and real-time ODN as shown in Table~\ref{tab:inf_time}. All three models satisfy real-time in the GPU setting, whereas only the complex ODN does not in the CPU setting. In summary, the Hybrid significantly outperforms two comparison models in the mobile platforms where GPU is incapable.\\ \BfPara{Ablation study of optical flow estimation} To verify the accuracy of the Hybrid model, we design the experiment with humans observing and recording the results of two models (Hybrid and ODN). The road driving dataset, consisting of 2k frames, is used in the corresponding experiment~\cite{dataset}. We set the performance evaluation criteria for the total number of objects detected, the number found correctly, the number found incorrectly, the number of overlapping objects, and the ratio of truly detected. The only difference is whether the simulation leverages optical flow estimation. Tab.~\ref{tab:my_label} shows the performance evaluation results of the ablation study. The total number of detected objects is $1.64$x higher in Hybrid than ODN. Among them, the number of correctly detected objects is $1.596$x more for Hybrid than ODN. The ratio of detecting the same object overlapping is $8.82$\% for Hybrid and $3.12$\% for ODN. The percentage of falsely detecting objects is $6.77$\% for ODN and $3.42$\% for Hybrid, which is lower in Hybrid. In the total number of detected objects, compared to the number of objects excluding overlapping objects, the number of objects accurately detected is $93.22$\% for ODN and $96.24$\% for Hybrid, showing the superiority in the accuracy of $3.02$\% for Hybrid.\\ \begin{figure}[t!] \centering \setlength{\tabcolsep}{2pt} \renewcommand{\arraystretch}{0.2} \centering \includegraphics[page=1, width=\columnwidth]{fig3.pdf} \begin{tabular}{cc} \includegraphics[page=1, width=0.5\columnwidth]{fig3a.pdf} & \includegraphics[page=1, width=0.5\columnwidth]{fig3b.pdf} \\ \\ \\ \small (a) CDF of the inference time & \small (b) CDF of the number of objects\\ \end{tabular} \caption{The three models' cumulative density functions ($y$-axis values) depend on the inference time and the number of objects in the GPU setting.} \label{fig:performance} \end{figure} \begin{table}[t!] \footnotesize \centering \caption{The inference time of three models with various settings.} \label{tab:inf_time} \begin{tabular}{c|c|c|c} \toprule[1pt] \textbf{Settings} & \textbf{Hybrid} & \textbf{Real-time ODN} & \textbf{Complex ODN}\\\midrule \multicolumn{1}{c|}{GPU} & 83\,ms & 55\,ms & 102\,ms \\ \multicolumn{1}{c|}{CPU} &133\,ms & 67\,ms & 562\,ms \\ \bottomrule[1pt] \end{tabular} \end{table} \begin{table}[t!] \footnotesize \centering \caption{The performance comparison between two models for the number of the detected object.} \label{tab:my_label} \begin{tabular}{c|c|c} \toprule[1pt] \textbf{Metric} & \textbf{ODN} & \textbf{Hybrid (Ours)}\\\midrule \multicolumn{1}{l|}{Total object} & 4,603 & 7,562\\ \multicolumn{1}{l|}{Correctly Detected} & 4,157 & 6,636\\ \multicolumn{1}{l|}{Falsely Detected} & 312 & 259\\ \multicolumn{1}{l|}{Overlapped Detected} & 144 & 667\\ \multicolumn{1}{l|}{True Positive Rate(\%)} & 93.22 & 96.24\\ \bottomrule[1pt] \end{tabular} \end{table} \subsection{Implementation of Model Selection}\label{sec:3-2} \begin{figure}[t] \centering \setlength{\tabcolsep}{2pt} \renewcommand{\arraystretch}{0.2} \begin{tabular}{cc} \multicolumn{2}{c}{\includegraphics[page=1, width=0.5\textwidth]{fig4.pdf}}\\ \includegraphics[page=1,width=0.24\textwidth]{fig4a.pdf} & \includegraphics[page=1,width=0.24\textwidth]{fig4b.pdf} \\ \\ \small (a) Queue-backlog & \small (b) Accuracy on average \end{tabular} \caption{The results of \textit{model selection} } \label{fig:performance2} \end{figure} \begin{table}[t!] \footnotesize \centering \caption{The configuration of the Comp3 framework} \label{tab:comp3} \begin{tabular}{l|r} \toprule[1pt] \textbf{Element} & \textbf{Detail}\\\midrule[1pt] \multicolumn{1}{l|}{\multirow{2}{*}{State}} & $s_t=\{Q[t], a[t-1], b(\alpha[t-1]), b(\alpha[t]), $\\ & $P(\alpha[t]),~w_1,~w_2,~w_{fps},~w_p,~V\}$\\\midrule \multicolumn{1}{l|}{Action} & $a_t\in\mathcal{A}\equiv\{H,T\}$\\\midrule \multicolumn{1}{l|}{Reward} & $r_t = V\cdot P(\alpha[t])+Q[t]b(\alpha[t])$ \\\midrule \multicolumn{1}{l|}{Optimization method} & REINFORCE \cite{sutton1999policy} \\ \midrule \multicolumn{1}{l|}{Optimizer} & Adam optimizer \\ \midrule \multicolumn{1}{l|}{Learning rate} & 0.0002 \\ \midrule & 3 fully connected layers with ReLU function\\ \multicolumn{1}{l|}{\multirow{1}{*}{Neural Network}} & \multicolumn{1}{l}{~~- 1st layer: $10\times 128$ + ReLU} \\ \multicolumn{1}{l|}{\multirow{1}{*}{Architecture}} & \multicolumn{1}{l}{~~- 2nd layer: $128\times 128$ + ReLU}\\ & \multicolumn{1}{l}{~~- 3rd layer: $128\times 2$ + ReLU}\\\bottomrule[1pt] \end{tabular \end{table} We investigate the performance to verify the Lyapunov optimization framework, which chooses between ODN and the Hybrid in a finite time (\textit{e.g.}, 100 seconds) depending on the amount of queue-backlog (\textit{i.e.}, delay). For comparison with the proposed framework, we adopt three comparison frameworks, \textit{i.e.}, the policy with selecting only ODN, or only the Hybrid and the deep reinforcement learning (DRL)-based policy, which are denoted as `Comp1', `Comp2', and `Comp3', respectively. To configure Comp3, we design the state, action, reward, optimization methods, and neural network architecture depicted in Table~\ref{tab:comp3}. Note that Comp3 is the state-of-the-art stochastic controlling method. \BfPara{Numerical result of model selection} Fig.~\ref{fig:performance2} shows the result of the network in which Hybrid and ODN observe the queue and thereby select a model. In the cases of model selection based on the proposed model selection and Comp3 on the mobile platform, it is confirmed that neither transient stability nor queue overflow occurs. In contrast, transient stability occurs in Comp1. In addition, queue overflow occurs in Comp2. Thus, our proposed framework and Comp3 outperform Comp1 and Comp2, corresponding to the queue stability (see Fig.~\ref{fig:performance2}(a)) and average accuracy (see Fig.~\ref{fig:performance2}(b)). \BfPara{Computing cost for model selection} Comp3 consumes $35,582$\,FLOPS for every model selection regarding computing cost. Furthermore, Comp3 requires additional computing costs for training the policy. However, the computing cost of every model selection requires only $12$\,FLOPS for the proposed framework, whereas no computing cost is necessary for Comp1, Comp2. The proposed framework is most suitable for model selection for the real-time service and among all frameworks. Consequently, we corroborate that the proposed model selection framework guarantees time-average performance maximization and queue stability, which requires very low-computational cost for model selection. \section{Conclusions and Future Work}\label{sec:4} In this paper, we confirm the improved performance of ODN in the road driving environment by utilizing optical flow estimation via OFEN and flow map processing. The flow map processing algorithm can be applied to all ODNs that utilize bounding box detection. A Lyapunov optimization-based model selection framework is also constructed to select a model under autonomous driving environments' constraints automatically. In addition, it is known to be time-average optimal subject to system/queue stability. As a result of observing the queue-backlog, it is confirmed that the time-average performance maximization can be achieved, as theoretically expected. As future research directions, real-world experimental study results in an actual driving environment are meant to verify the novelty of this algorithm in practice. \bibliographystyle{IEEEtran} \balance

2e6029605f225b04310f4f1ebc7f9cee665af322

\section{Introduction} \IEEEPARstart{C}{oncerns} about climate change, resilience to hazardous events, and sustainability have shifted the electric power sector in the U.S. and elsewhere toward more involvement on the demand side to harness flexible distributed energy resources (DERs). More recently, the U.S. Federal Energy and Regulatory Commission (FERC) issued Order 2222, which removes barriers to the integration of DERs into wholesale electricity markets \cite{ferc2222}. More specifically, the ruling allows the integration of multiple DERs owned by different entities with different sizes and diverse technologies to participate in the regional organized wholesale capacity, energy, and ancillary services markets alongside traditional resources. This ruling offers incentives for households that own DERs (known as prosumers) to ``value-stack'' their assets to provide various types of energy-related commodities to the grid. Already, we are witnessing some activities in the marketplace in response to or anticipation of the order. For instance, OhmConnect, a clean tech company, recently announced a plan to link homes spread in California to form a 550 MW virtual power plant (VPP) of distributed energy resources. Naturally, prosumers are self-served for their own interests, and their behavior resulting from optimizing their private benefits is unlikely to be in the best interests of the energy market as a whole. Not surprisingly, how to compensate for the energy produced by prosumers has emerged as a critical issue that can facilitate or impair the deployment of DERs and is currently subject to contentious debates\cite{Costello14,bushnell18,chakraborty19}. The situation is also complicated by the existing formation of retail tariff. In general, a retail tariff consists of four core elements: (i) costs of electric energy; that is, wholesale locational marginal prices (LMPs), (ii) costs of other energy-related services, such as operating reserves or capacity costs, (iii) costs for network-related services, including investment and maintenance costs of transmission and distribution network assets, and (iv) charges to recover policy costs, such as procurement costs to support state's RPS (renewable portfolio standards) \cite{mit:16} The last two items, (iii) and (iv), are generally lumpy and non-convex as they are not directly tied to the level of energy consumption. These two elements are dubbed as ``residual costs'' in \cite{burger19}. The breakup of those four elements depends on specific markets; for example, the energy or LMP component can be as low as 10\% in the Netherlands or as high as 60\% in New Jersey \cite{mit:16}. \iffalse Currently, there are two commonly implemented systems in the power sector: net-billing and net-metering. Whereas the net-metering system is more common in the US, the net-billing is more popular in Europe. Under a net-billing system, two meters are installed, recording two quantities, i.e., power withdrawn from and power injected into the grid, where withdrawal and injection can be subject to difference prices. In contrast, a single meter can be installed under a net-metering system to record both transactions, subjecting them to the same price. \fi Recovering residual costs is a thorny political-economic endeavor, which may facilitate or impair the deployment of residential DERs, and has been subject to contentious debates \cite{Costello14,chakraborty19}. Two systems are of great interest. The first is referred to as net-metetering; that is, prosumers are only billed for their ``net" energy use. The energy they sell back to the grid will be paid at the same rate as buying from the grid. The second system is called net-billing, under which two meters are installed, recording two quantities: energy withdrawn from and energy injected into the grid. The withdrawal and injection can be subject to different prices. While the net-metering is the most common approach and has provided strong incentives for DER investment, it is also causing serious equity issues, as the more prosumers take advantage of net metering, the fewer residual costs are paid into the system, resulting in higher rates for non-net metering customers, likely those of low-income. Recently, the California Public Utility Commission (CPUC) engaged in a regulatory process to revamp its net-metering policy, as the CPUC is fully aware of its drawbacks.\footnote{\url{https://www.cpuc.ca.gov/-/media/cpuc-website/divisions/energy-division/documents/net-energy-metering-nem/nemrevisit/430903088.pdf}} Such issues have also been vetted by the academics, which have been described as ``revenue erosion'' or ``network defection''; that is, utilities are forced to increase the retail tariff to compensate for the revenue deficiency, further exacerbating the situation and leading to the so-called ``death spiral" \cite{Borenstein15,Picciariello15,Castaneda17,Kubli18,Darghouth16}. Some empirical evidence emerges: for instance, using data from three investor-owned utilities in California, Wolak \cite{wolak18} finds that two-thirds of the increases in residential distribution prices can be attributed to the growth of solar capacity. A number of recent studies have also addressed the issues of pricing energy produced by prosumers. Clastres et al. \cite{clastres19} estimate the extent of cross-subsidies between prosumers and conventional consumers in France. The authors also conclude that a demand charge may alleviate the network defection or death spiral problem facing distributed system operators. Using stylized models, Gautier \cite{Gautier18} concludes that net-metering decreases the payment from prosumers, which is cross-subsidized by the higher bills of conventional consumers. More recently, Gorman et al. \cite{gorman20} compare grid costs to off-grid costs of more than 2,000 utilities in the U.S. and find that network defection could increase from 1\% to 7\%, with 3\% in the Southwest region and California and 7\% in Hawaii. \textcolor{black}{However, little attention was given to examine the impact of retail tariffs on the energy equality among different income groups in the presence of prosumers. An exception is our earlier work that compares the energy expenditure incidence among different income groups when prosumers are subject to a net-metering and a net-billing policy \cite{chen21}. The paper concludes that net-metering is more regressive than net-billing under the volumetric tariff. A hybrid policy, which also features an income-based fixed charge may potentially improve energy equity.} \iffalse Other studies focus on cost allocations for transmission and/or distribution expansion when the system is subject to uncertain renewables or load growth. The models developed in these studies typically have multiple levels because they are interested in the effects of transmission planning and cost allocation on capacity expansion and generation operations. Examples include \cite{delgado15}, who proposes a multistage expansion planning problem jointly considering investment in distribution network and generation in distribution. Another study explores this issue using a tri-level model, where the first level represents transmission planning, the second is renewable energy expansion, and the third is operations \cite{wang18}. Different from the aforementioned studies, Kristiansen \cite{Kristiansen18} apply a cooperative-game-theoretical approach based on the Sharpley value to allocating the benefits and costs of international transmission investments, focusing on wind energy in the North Sea Offshore Grid. Other similar studies include \cite{zhao11,shen17}. Other relevant work also examines issues on recovering or incorporating fixed costs in pricing electricity. However, their focus is on the wholesale market. For instance, Gribik \cite{Gribik07} studies pricing rules that involve uplifts or make-whole payments to induce generators desirable market outcomes defined by unit commitment models. More recently, Mays \cite{mays21} investigates the impact of different pricing schemes can have on generator entry-and-exit decisions. The study finds that despite the presence of fixed production cost elements, prices derived from marginal costs support the optimal capacity mix. \fi The current study extends our previous work to offer a policy prescription on optimal retail tariff design in the face of a growing presence of prosumers. The problem is formulated as a bilevel optimization problem: the upper level represents a public utility commission (PUC)'s decision-making problem that has to decide a certain retail tariff structure to guarantee the recover utility's fixed costs as well as maintaining energy equity. The lower level represents an market equilibrium that consists of prosumers, consumers, producers, and an independent system operator (ISO), with the prosumer/consumers' retail rates set by the upper-level PUC. Although each prosumer may be relatively small, possessing limited ability to engage in the bulk energy market, we assume that an entity integrates a large number of prosumers and participates in the bulk energy market on their behalf; this is consistent with FERC Order 2222's requirements. The prosumers are endowed with renewables and decide on the amounts of self-consumption, dispatchable energy to produce, such as from back-up generators or energy storage, and energy to sell to or buy from the bulk energy market to maximize their net benefit. The ISO minimizes the generation costs while treating sales or purchases by the prosumers as exogenous. \textcolor{black}{Similar to the earlier work by Woo \cite{woo88}, we also explicitly consider the problem facing a PUC and distinguish the retail rate from the wholesale rate. However, unlike it, we extend the analysis to consider the energy expenditure incidence among income groups to address equity issues.} Our analysis of the theoretical properties of the bilevel is also worth noting. More specifically, we prove that \emph{laissez-faire} is socially optimal; that is, zero volumetric charge will maximize the social surplus, as it will not distort the equilibrium price in the wholesale market, and energy equity can be achieved through different fixed charges among different income groups. While a fixed-charge-only tariff is implausible in reality, our formulation is amenable to computing a second-best solution when a proportion of the utility's fixed cost is required to be recovered from volumetric charges. \textcolor{black}{While our lower-level problem} is related to \cite{clastres19,Gautier18}, it is different in significant ways. In particular, we consider the transmission network and market details, \textcolor{black}{e.g., pool-type market settlement, capacity ownership, generation capacity constraints, retail-wholesale market linkage}, which are crucial in determining realistic electricity market outcomes. \iffalse Third, we note that the size of the prosumers is exogenous since our interest is not the optimal proportion of prosumers. \textcolor{red}{Because, in reality, any changes to the retail tariff is a lengthy legal process, unlikely to be socially optimal, but is undoubtedly subject to various inputs from stakeholders with competing interests, e.g., utilities, the PUC, and consumer advocates. Thus, our intention herein is to understand how an income-based fixed-cost charge may improve the equity in energy expenditure among different income groups. } \fi The remainder of this paper is organized as follows. Section \ref{sec:mod} presents the lower- and upper-level models of the bilevel problem. Solution properties and theoretical results are shown in Section \ref{sec:Theo}. A numerical case study is presented in Section \ref{sec:case}. Finally, concluding remarks are provided in Section \ref{sec:con}. \section{Model}\label{sec:mod} We present the complete model in this section, starting with the lower-level market equilibrium formulation, followed by the upper-level problem to maximize social surplus and energy equity. The resulting problem can written as either a mathematical program with equilibrium constraints (MPECs) or a bilevel problem (BLP), with the former formulation amenable to computation and the latter one easier for theoretical analysis. \vspace{-1em} \subsection{Lower-Level Problem} The lower-level problem consists of problems faced by the consumers, prosumers, power plants, and the ISO. Throughout the paper, we make the blanket assumption that the market is perfectly competitive; that is, all market participants are price-takers of the market prices, without contemplating on how to manipulate the equilibrium prices through their unilateral actions. While market power abuse has always been a concern for wholesale energy markets, retail ratemaking usually lasts for a certain period of time; that is, retail rates do not change frequently. It is therefore unreasonable to assume that a wholesale market is subject to sustained market power abuse. \subsubsection{Consumers} Consider an energy market that has $N$ nodes and $K$ transmission lines that connect the nodes. Consumers at each node $i = 1, \ldots, N$ are grouped into two types, including conventional consumers and prosumers, whose marginal benefit functions (that is, their willingness-to-pay functions), denoted by $p_i^{con}$ and $p^{pro}_i$, respectively, are represented by the following linear inverse demand functions: \begin{align} p^{con}_i(d_i)& =P^0_i - \Big({P^0_i}/{\big((1-\alpha_i)Q^0_i\big)}\Big)d_i,\ \forall\ i \label{demand}\\ p_i^{pro}(l_i)& =P^0_i - \Big({P^0_i}/{\big(\alpha_i Q^0_i\big)}\Big)l_i,\ \forall\ i, \label{demandp} \end{align} where $P^0_i > 0$ and $Q^0_i >0$ respectively represent the vertical and horizontal intercepts of the ``horizontally aggregated'' retail inverse demand function: \textcolor{black}{$p_i^{r}(d_i+l_i)=P^0_i - \big({P^0_i}/{Q^0_i}\big)(d_i+l_i)$, as illustrated in Fig. \ref{fig:demand}}. The quantities demanded by conventional consumers and prosumers are denoted by $d_i$ and $l_i$, respectively. The parameter $\alpha_i$ is the fraction of prosumers at node $i$. Note that while $\alpha_i$ varies between 0 and 1, the aggregated demand does not change. \begin{figure}[!htbp] \centering \begin{tikzpicture}[scale=0.6] \begin{axis}[axis lines=center, xmin=0, xmax=22, ymin=0, ymax=22, xtick ={20/3,40/3,20}, xticklabels ={$\alpha_i Q^0_i$,$(1-\alpha_i) Q^0_i$,$Q^0_i$}, ytick ={20}, yticklabels ={$P^0_i$}] \addplot[domain=0:20]{ -x+20}; \addplot[dashed,domain=0:20]{ -3*x+20}; \addplot[dashed,domain=0:20]{ -1.5*x+20}; \node[left] at (axis cs:4,6) {(2)}; \node[left] at (axis cs:9,6) {(1)}; \node[right] at (axis cs:9.3,14) {$p_i^{r}(d_i+l_i)$}; \node[right] at (axis cs:9.3,12) {$=P^0_i - \big({P^0_i}/{Q^0_i}\big)(d_i+l_i)$}; \end{axis} \end{tikzpicture \caption{An illustration of retail demand curves} \label{fig:demand} \end{figure} Let $p_i$ denote the wholesale energy price at node $i$, and $\tau^b$ be the volumetric charge of energy purchase for all consumers/prosumers, which is a part of consumers' retail rates.\footnote{Note that retail rates are usually the same covering a broad area of customers, and hence, we do not have a node index of $\tau$, but can certainly do so.} The other part of the retail rates is the fixed charge. Since the fixed-charge rate will serve as a main tool to realize energy equity, we assume that conventional consumers and prosumers can be subject to different fixed-charge rates, and denote them as $\phi_i^{con}$ and $\phi_i^{pro}$, respectively. With the marginal benefits and costs defined, conventional consumers at each node $i = 1, \ldots, N$ maximize their net benefits (also referred to as surplus) by solving: \begin{maxi!}[3] {d_i\geq 0}{\int^{d_{i}}_{0} p^{con}_i(m_i)dm_i- (p_i + \tau^b)d_i - \phi_i^{con}.}{\label{eq:con}}{} \tag{3} \end{maxi!} Since $p_i$, $\tau^b$, and $\phi^{con}_i$ are all exogenous to consumers, the optimization problem is easily seen to be a strongly convex problem with the given linear inverse demand function as in\eqref{demand}. Hence, an optimal solution always exists with respect to any ($p_i$, $\tau^b$, $\phi^{con}_i$), and the first-order optimality conditions, aka the KKT conditions, are both necessary and sufficient for optimality. The collection of conventional consumers' KKT conditions are that for $i = 1, \ldots, N$: \begin{align} 0 \leq d_{i} \perp P_i^0 - \Big({P^0_i}/{\big((1-\alpha_i)Q^0_i\big)}\Big)d_i - (p_i + \tau^b) \leq 0, \label{eq:d_KKT} \end{align} where the `$\perp$' sign means that the product of the scalars or vectors is 0, and such a constraint is referred to as a complementarity constraint. The KKT conditions have intuitive economic interpretation: at an optimal solution (denoted with a `$*$' superscript), if $d_i^* > 0$, then consumers choose to purchase energy at the level where the marginal benefit, $P^0_i - \Big({P^0_i}/{\big((1-\alpha_i)Q^0_i\big)}\Big)d^*_i$ equals the marginal cost, which is the retail price $p^r_i:= p_i+\tau^b$. \subsubsection{Prosumers}\label{subsec:Prosumer} For prosumers, they pay the same volumetric charge $\tau^b$ when buying from the grid. However, when they sell to the grid, we assume that the rate they receive is $p_i + \tau^s$. If $\tau^s = \tau^b$, then it is the net-metering policy; otherwise, it is net billing. Note that while $\tau^b$ is always non-negative, $\tau^s$ can be positive or negative. When $\tau^s>0$, the prosumers effectively receive a ``subsidy'' in addition to the wholesale price $p_i$. In the case where $\tau^s<0$, it means that the prosumers are subject to a ``tax'' when selling power to the grid With the exogenous volumetric rates $\tau^b$, $\tau^s$ and fixed rate $\phi^{pro}$, we posit that a prosumer maximizes its surplus by deciding i) energy to buy ($z_i^b$) from or sell ($z_i^s$) to \textcolor{black}{the grid at} node $i$, ii) consumption level $l_i$ given renewable output $R_i>0$, and iii) generation $g_i$ from the backup dispatchable technology with a cost $C^g_i(g_i)$. The prosumer's problem at node $i$ is: \vspace{-1em} \iffalse \noindent \underline{\it{Net Metering}:} Under a net-metering system, there can be one meter recording the power injection ($z_i>0$) and withdrawal ($z_i<0$). Both types of transactions are subject to $\tau$, \textcolor{black}{which is regarded as a ``subsidy'' for sale ($z_i>0$) or an end-user transmission tariff for purchase ($z_i<0$)}. The optimization problem faced by the prosumer at node $i$ under a net-metering system is displayed as follows: \begin{maxi!}[3] {l_i,g_i \geq 0, z_{i}}{(p_i+\tau)z_{i}{+} {\int^{l_{i}}_{0} p^{pro}_i(m_i)dm_i}- C^g_i(g_i)-\phi^{pro}_i\protect\label{1a}}{\label{eq:nm-obj}}{} \addConstraint{z_{i} + l_i - R_i - g_i}{\leq 0}{\qquad(\delta_i)}{\label{eq:nm-con1}} \addConstraint{g_i}{\leq G_i}{\qquad(\kappa_i)}{\label{eq:nm-con2}} \end{maxi!} The corresponding KKT conditions of the prosumers' problem at node $i$ under the net-metering system are collected for all nodes as in (\ref{eq:nm1})--(\ref{eq:nmf}). \begin{subequations} \label{nm} \begin{align} & \textcolor{black}{p_i + \tau - \delta_i = 0}, \forall i \label{eq:nm1}\\ & 0 \leq l_i \perp P^0_i-\frac{P^0_i}{\alpha_i Q_i^0}l_i - \delta_i \leq 0, \forall i \label{eq:nm2}\\ & 0 \leq g_i \perp - C^{g\prime}_i(g_i) + \delta_i -\kappa_i \leq 0, \forall i \\ & 0 \leq \delta_i \perp z_i + l_i - K_i - g_i \leq 0, \forall i \\ & 0 \leq \kappa_i \perp g_i - G_i \leq 0, \forall i \label{eq:nmf} \end{align} \end{subequations} \fi \begin{subequations} \begin{align} \maximize{z^b_{i},z^s_{i},l_i,g_i \geq 0 }\ & \ (p_i+\tau^s)z^s_{i}-\left(p_i+\tau^b \right)z^b_{i}\nonumber \\[-5pt] & +\int^{l_{i}}_{0} p^{pro}_i(m_i)dm_i- C^g_i(g_i)- \phi^{pro}_i \label{eq:nb-obj} \\ \mathrm{subject\ to} & \ l_i + z^s_{i}-z^b_{i} - g_i - R_i = 0 \quad (\delta_i) \label{eq:nb-con1} \\ &\ g_i \leq G_i \quad (\kappa_i)\label{eq:nb-con2} \end{align} \end{subequations} In the constraint set, Eq. (\ref{eq:nb-con1}) ensures that \textcolor{black}{the net} demand, $l_i + z^s_{i}-z^b_{i}$, is balanced with the prosumer's own renewable and backup generation ($R_i + g_i$). Eq. (\ref{eq:nb-con2}) limits the backup output to be less than its capacity $G_i$. Two things to note about the above optimization problem. First, if $\tau^s > \tau^b$, the optimization problem is clearly unbounded. This is intuitive: if $p_i + \tau^s > p_i + \tau^b$, meaning that selling electricity back to the grid earns more than buying from the grid. Then a prosumer can simply arbitrage and earn an infinite amount of profit. To rule out this case, we make the no-arbitrage assumption that $\tau^s \leq \tau^b$. Second, under the no-arbitrage assumption, for the buy and sell decisions, $z_i^s$ and $z_i^b$ should not both be positive in an optimal solution, from a common sense perspective. Mathematically, however, this is not guaranteed, unless we have some explicit constraints such as $z_i^s \cdot z_i^b = 0$. In the following, we present the simple fact that such an explicit constraint is not necessary. To do so, we first write down the KKT conditions of the prosumers optimization problems (assuming that $C_i^g(\cdot)$ is differentiable). \begin{subequations} \begin{align} & 0 \leq l_i \perp P^0_i-\frac{P^0_i}{\alpha_i Q_i^0}l_i - \delta_i \leq 0,\ \forall i \label{eq:ProKKT_l}\\ & 0 \leq g_i \perp - C^{g\prime}_i(g_i) + \delta_i -\kappa_i \leq 0,\ \forall i \\ & 0 \leq \kappa_i \perp g_i - G_i \leq 0,\ \forall i\\ & 0 \leq z_i^s \perp (p_i + \tau^s) - \delta_i \leq 0,\ \forall i \label{eq:Pro_zs}\\ & 0 \leq z_i^b \perp - (p_i + \tau^b) + \delta_i \leq 0,\ \forall i \label{eq:Pro_zb}\\ & \delta_i \ \mathrm{free},\ l_i + z^s_{i}-z^b_{i} - g_i - R_i = 0,\ \forall i. \label{eq:ProKKT_delta} \end{align} \end{subequations} Since the constraints \eqref{eq:nb-con1} -- \eqref{eq:nb-con2} are all linear, and with a given tuple $(p_i, \tau^s, \tau^b, \phi_i^{pro})$, the objective function is concave (under the assumption that $C_i^g(\cdot)$ is a convex function), the KKT conditions are again necessary and sufficient optimality conditions. Then we have the following result. \begin{lemma}\label{lem:zszb} Assume that $C_i^g(\cdot)$ is continuously differentiable and convex. With a given tuple $(p_i, \tau^s, \tau^b, \phi_i^{pro}) \in \Re_+ \times \Re \times \Re_+ \times \Re$ in which $ \tau^s \leq \tau^b$, an optimal solution of the problem \eqref{eq:nb-obj} -- \eqref{eq:nb-con2}, denoted by $(l^*_i,g^*_i,z^{s^*}_{i},z^{b^*}_{i})$, exists. In addition, if $\tau^s \neq \tau^b$, then $z^{s^*}_{i} \cdot z^{b^*}_{i} = 0$. \end{lemma} \begin{proof} Let $X_i^{pro} \in \Re^4$ denote the feasible region of the prosumers' problem at node $i$. It is easy to see that $X_i^{pro} \neq \emptyset$ since the zero vector is always in $X_i^{pro}$. $X_i^{pro}$ is also clearly a closed set. When $\tau^s \leq \tau^b$, the objective function \eqref{eq:nb-obj} goes to $-\infty$ for any $(l_i,g_i,z^{s}_{i},z^{b}_{i}) \in X_i^{pro}$ with $|| (l_i,g_i,z^{s}_{i},z^{b}_{i})|| \rightarrow \infty$, which means that \eqref{eq:nb-obj} is coercive on $X_i^{pro}$. Since \eqref{eq:nb-obj} is also continuous, an optimal solution exists by a variant of the well-known Weierstrass' Theorem (such as Proposition A.8 in \cite{bertsekas99}). For the second part, when $\tau^s \neq \tau^b$, if $z^{s^*}_{i} \cdot z^{b^*}_{i} > 0$, the conditions \eqref{eq:Pro_zs} and \eqref{eq:Pro_zb} imply that $p_i + \tau^s = \delta_i^* = p_i + \tau^b$, a contradiction. \end{proof} When $\tau^s = \tau^b$, clearly the optimal solutions $z^{s^*}_{i}$ and $z^{b^*}_{i}$ are not unique and can be both positive. In this case, we can simply define $\tilde{z}_i^s := \max\{z^{s^*}_{i} - z^{b^*}_{i}, 0\}$ and $\tilde{z}_i^b := \max\{z^{b^*}_{i} - z^{s^*}_{i}, 0\}$; then at most one of them is nonzero in an optimal solution. In the following sections, we will see that it is always the net of a prosumer's decision, namely, $z^{s}_{i} - z^{b}_{i}$, that appears in the other part of the market equilibrium model. Hence, not including an explicit constraint such as $z_i^s \cdot z_i^b = 0$ will not affect the outcomes of a market equilibrium in any way, and omitting such combinatorial constraints will considerably simplify both theoretical analysis and computation. \subsubsection{Power Producers and The ISO} We assume that an ISO collects offers from electric power producers to minimize the total generation cost, while treating the demand and prosumers' buy and sell decisions as exogenous. The optimization problem is as follows: \vspace*{-5pt} \begin{subequations} \begin{align} \minimize{g_{ih}\geq 0,\ y_{i}\ \mathrm{free}}\ & \ \sum_{i=1}^N\sum_{h \in H_{i}} C_{ih}(g_{ih}) & \label{eq:iso} \\ \mathrm{subject\ to} & \ g_{ih} - G_{ih}\leq 0 \hspace*{10pt} (\rho_{ih}),\ \forall i, h \in H_i \label{eq:iso_con1} \\ & \ \sum_{i=1}^N y_{i}= 0 \hspace*{25pt} (\theta) \label{eq:sumY=0} \\ & \ \sum_{i=1}^N PTDF_{ki} y_i \leq T_k \hspace*{18pt} (\lambda_k^+),\ \forall k \label{eq:iso_con4}\\ & \ -\sum_{i=1}^N PTDF_{ki} y_i \leq T_k \hspace*{10pt} (\lambda_k^-),\ \forall k \label{eq:iso_con5} \\ & \ y_i-\sum_{h \in H_i}g_{ih}-(z^s_i-z^b_i) +d_i =0 \hspace*{10pt} (p_i),\ \forall i \label{eq:nb-price} \vspace{-2em} \end{align} \end{subequations} In the above problem, \eqref{eq:iso_con1} is the generation capacity constraint. The set $H_i$ represents all power plants at node $i$; therefore, we do not need to assume that there is only one power plant at each node $i$. Eq. (\ref{eq:sumY=0}) ensures that the total net injection/withdrawal in the system is equal to zero, where $y_i$ represents the energy flow from an arbitrarily assigned hub node to node $i$. We represent the transmission network as a hub-spoke system; that is, the energy flows from node $i$ to $j$ are considered as from $i$ to the hub, and from the hub to $j$. Eqs. (\ref{eq:iso_con4})--(\ref{eq:iso_con5}) describe that the flow in link $k$ is less than its transmission limit $T_k$. The term $PTDF_{ki}$ represents the power transfer distribution factors based on linearized DC flows. Eq. \eqref{eq:nb-price} is the mass balance constraint, whose shadow price, denoted by $p_i$, is exactly the wholesale energy price at node $i$ (that is, the LMP at $i$). To aid in model development and analysis, we make the following blanket assumption throughout the paper. \begin{assumption}\label{as:QuadraticCost} The generation cost function $C_{ih}(\cdot)$ is $C_{ih}(g_{ih}) = a_{ih}g_{ih} + \frac{1}{2}A_{ih}g_{ih}^2$, with the input parameters $a_{ih} > 0$ and $A_{ih} > 0$ for all $i = 1, \ldots, N$ and $h\in H_i$. \end{assumption} According to Eq. \eqref{eq:nb-price}, with a given $d_i$, the variable $y_i$ is implicitly bounded for $i = 1, \ldots, N$, since $g_{ih}$'s are bounded, and so is the quantity $z_i^s - z_i^b$ based on Eq. \eqref{eq:nb-con1}. Hence, with Assumption \ref{as:QuadraticCost}, if the feasible region is not empty with respect to a given demand vector $(d_i)_{i=1}^N$ and $(z_i^b, z_i^s)_{i=1}^N$, then an optimal solution of the above optimization problem exists, and the KKT conditions are necessary and sufficient of optimality, due to the all-linear constraints and the convex objective function. The detailed KKT conditions are: \begin{subequations}\label{e10} \begin{align} & 0 \leq g_{ih} \perp -C_{ih}^\prime(g_{ih})-\rho_{ih}+p_i \leq 0,\ \forall i, h \in H_i \label{eq:foc-iso-x}\\ & -\theta + \sum_{k=1}^K PTDF_{ki}(\lambda_{k}^- - \lambda_{k}^+) - p_i= 0,\ \forall i \\ & \textcolor{black}{0 \leq \rho_{ih} \perp g_{ih} - G_{ih} \leq 0,\ \forall i, h \in H_i} \\ & \theta\ \mathrm{free},\ \textcolor{black}{\sum_{i=1}^N y_{i}=0} \\ &0 \leq \lambda_{k}^+ \perp \sum_{i=1}^N PTDF_{ki}y_{i} - T_{k} \leq 0, \ \forall k \\ &0 \leq \lambda_{k}^- \perp -\sum_{i=1}^N PTDF_{ki}y_{i} - T_{k} \leq 0, \ \forall k \label{eq:foc-iso-f}\\ & p_i\ \mathrm{free},\ y_i-\sum_{h \in H_i}g_{ih}-(z^s_i-z^b_i) +d_i=0,\ \forall i. \label{eq:foc-iso-g} \end{align} \end{subequations} \vspace{-1em} \subsection{Upper-Level Problem} \subsubsection{Fixed Cost Recovery} First and foremost, revenue accrued through retail electricity rates (both volumetric and fixed rates) must cover utilities' fixed costs, which is ensured by the following constraint in the upper-level problem: \vspace{-5pt} \begin{align} \sum_{i=1}^N \big(\textcolor{black}{-z_i^s\tau^s} + z_i^b\tau^b +d_i\tau^b +\sum_{j \in {pro,con}} \phi^j_i \big)= \Tau, \label{eq:nb \end{align} where $\Tau$ represents the fixed cost to be recovered, which is exogenous to the model, and can be decided by utilities and approved by an energy commissioner. \subsubsection{Equity and Energy Expenditure Incidence} The upper-level decision-maker aims to strike the balance between maximizing the social surplus and maintaining energy equity when deciding on the retail electricity rates. To establish a model to aid decision making, we need quantitative measures of energy equity. In this work, we use the same measures (but with a small update) as developed in our earlier work \cite{chen21}. To make this paper self-contained, we reintroduce the measures here. The key concept is the energy expenditure incidence (EEI), defined for conventional consumers ($inc^{con}_i$) and prosumers ($inc^{pro}_i$) at node $i$ as follows: \begin{align} inc^{con}_i & = \frac{(p_i + \tau^b)d_i+\phi^{con}_i}{\textcolor{black}{I^{con}_i}}, \label{eq:inc_con} \\ inc^{pro}_i & = \frac{(p_i + \tau^b)z^b_i+ \phi^{pro}_i+C_i(g_i) + SC_i}{\textcolor{black}{I^{pro}_i}}. \label{eq:inc_pro} \end{align} The EEI measures the proportion of consumers' spending on electricity compared to total household income (denoted by $I^{con}_i$ and $I^{pro}_i$ for conventional consumers and prosumers, respectively). The conventional consumers' energy spending is easy to understand, which is the sum of the fixed charge $\phi_i^{con}$ and the cost of purchasing $d_i$ energy at the rate of $p_i + \tau^b$; the prosumers' spending includes the cost of backup generation ($C_i(g_i)$) and the sunk costs $(SC_i)$ spent on purchasing or renting renewable energy equipment. Note that we do not subtract the prosumers' earnings (i.e., $(p_i + \tau^s)z_i^s$) from the energy spending in the numerator (or add it to $I^{pro}_i$ in the denominator) in \eqref{eq:inc_pro}, as we believe that prosumers should not be penalized for selling energy to the grid. With the EEI defined, our idea of energy equity is to minimize the differences in EEI between conventional consumers and prosummers, that is, to minimize $\sum_{i}^N(inc^{con}_i-inc^{pro}_i)^2$. The definition of EEI in \eqref{eq:inc_con} and \eqref{eq:inc_pro} involves both lower-level variables $(d_i, z^b_i, g_i, p_i)_{i=1}^N$ and upper-level variables $(\tau^b, \bm{\phi}^{con}, \bm{\phi}^{pro})$, where $\bm{\phi}^{con} := (\phi_1^{con}, \ldots, \phi_N^{con})\in\Re^N$ and $\bm{\phi}^{pro} := (\phi_1^{pro}, \ldots, \phi_N^{pro})\in\Re^N$. To simplify the argument, we define a vector $\bm{\zeta}\equiv(\tau^b, \tau^s, \bm{\phi}^{con}, \bm{\phi}^{pro}) \in \Re^{2N + 2}$, and use the notation $d(\bm{\zeta})$ to denote the optimal solution mapping with a given $\bm{\zeta}$; that is, $d(\bm{\zeta})$ is a set of all optimal solutions of the consumers' optimization problem \eqref{eq:con} with a given $\bm{\zeta}$. The notation for all other optimal solution mappings is the same. Then we define the difference-of-incidence function as: \begin{equation}\label{eq:B_func} B\left[\bm{\zeta};d(\bm{\zeta}), z^b(\bm{\zeta}), g(\bm{\zeta}), p(\bm{\zeta})\right] := \sum_{i}^N\left(inc^{con}_i-inc^{pro}_i\right)^2. \end{equation} \subsubsection{The Complete Model} As the upper-level decision maker does not want to distort wholesale energy markets with their ratemaking, we include the social-surplus maximization problem for the wholesale energy market in the upper-level problem as well. Social surplus of the wholesale market is defined as the sum of all the market participants' surplus, which we denote as $\Pi(\bm{x}; (\tau^b, \tau^s))$, with $\bm{x}$ representing the collection of all lower-level variables: $(d, l, z^s, z^b, g, y)$: \begin{align} \Pi[\bm{x}; (\tau^b, \tau^s)]& := \sum_{i}^N\left[\int_{0}^{d_i}p^{con}_i(m_i)\, dm_i - \tau^b d_i \right] + \nonumber \\ & \sum_i^N\Bigg[\tau^s z^s_{i} - \tau^b z_i^b + \int_{0}^{l_i}p^{pro}_i(m_i)\, dm_i- C^g_i(g_i) \Bigg] \nonumber \\ & - \sum_{i=1}^N\sum_{h \in H_{i}} C_{ih}(g_{ih}). \label{eq:PI} \end{align} Note that by following the convention in economics, we do not include fixed charges $\bm{\phi}^{con}$ and $\bm{\phi}^{pro}$ in the social surplus, as the behavior of wholesale market participants is not affected by fixed charges in anyway. With the social surplus and EEI-difference function defined, the complete two-level problem can be written as follows: \begin{align} \maximize{\bm{\zeta},\ \bm{x}}\ & \ \Pi\left[\bm{x}; (\tau^b, \tau^s)\right] - M\cdot B(\bm{\zeta}; \bm{x}) \label{eq:MPEC}\\ \mathrm{subject\ to} & \ (\ref{eq:d_KKT}), (\ref{eq:ProKKT_l})-(\ref{eq:ProKKT_delta}), (\ref{eq:foc-iso-x})-(\ref{eq:foc-iso-g})\text{, and } (\ref{eq:nb}), \nonumber \\ & \tau^s \leq \tau^b, \ \tau^b, \bm{\phi}^{con}, \bm{\phi}^{pro} \geq 0, \tau^s \ \mathrm{free}, \nonumber \end{align} where the parameter $M > 0$ in the objective function is to balance between the two (possibly conflicting) objectives of maximizing the social surplus and minimizing the difference of energy expenditure incidence. Constraint (\ref{eq:nb}) is the upper-level constraint to ensure revenue adequacy for utilities; while the constraints (\ref{eq:d_KKT}), (\ref{eq:ProKKT_l})-(\ref{eq:ProKKT_delta}), (\ref{eq:foc-iso-x})-(\ref{eq:foc-iso-g}), the collection of the KKT systems, represent the market equilibrium of the wholesale energy market with a given retail rate $(\tau^b, \tau^s)$. Problem \eqref{eq:MPEC} is an MPEC, which can be solved by nonlinear programming (NLP) solvers capable of dealing with complementarity constraints, such as KNITRO \cite{Knitro} or FILTER \cite{Filter}. Granted that \eqref{eq:MPEC} is a nonconvex problem, only a locally optimal (or stationary) solution can be computed using an NLP solver. However, in the next section, we introduce a BLP formulation and establish the relationship between the MPEC and the BLP. The BLP will provide an approach to find a globally optimal solution of \eqref{eq:MPEC} by solving only convex problems. We also present some key theoretical results regarding the optimal solutions in the following section. \iffalse \begin{maxi!}[3] {d_i, \tau^s, \tau^b, \phi_i}{\sum_{i}\left(\int_{0}^{d_i}p^{con}_i(n_i)\, dn_i - (p_i+\tau^b)d_i-\phi^{con}_i\right)\nonumber} {\label{eq:MPEC}}{} \breakObjective {+ \sum_{i,h \in H_{i}} \Big(p_ig_{ih}- C_{ih}(g_{ih})\Big) \nonumber} \breakObjective{+\sum_i \Bigg((p_i+\tau^s)z^s_{i}-\left(p_i+\tau^b \right)z^b_{i}\nonumber} \breakObjective {+\int_{0}^{l_i}p^{pro}_i(m_i)\, dm_i- \sum_i C^g_i(g_i) -\phi^{pro}_i\ \Bigg)\nonumber} \addConstraint{(\ref{eq:nb1})-(\ref{eq:nbf}), (\ref{eq:foc-iso-x})-(\ref{eq:foc-iso-f}), (\ref{eq:nb-price},)\text{ and } (\ref{eq:nb})}{\nonumber} \end{maxi!} Consider the following formulation: \begin{maxi!}[3] {d_i, \tau^s, \tau^b, \phi_i}{\sum_{i}\left(\int_{0}^{d_i}p^{con}_i(n_i)\, dn_i - \tau^b d_i-\phi^{con}_i\right)\nonumber} {\label{eq:MPEC}}{} \breakObjective { \sum_{i,h \in H_{i}} - C_{ih}(g_{ih}) \nonumber} \breakObjective{+\sum_i \Bigg(\tau^s z^s_{i}- \tau^b z^b_{i}\nonumber} \breakObjective {+\int_{0}^{l_i}p^{pro}_i(m_i)\, dm_i- \sum_i C^g_i(g_i) -\phi^{pro}_i\ \Bigg)\nonumber} \addConstraint{l_i - g_i - K_i - z^b_{i} + z^s_{i} }{= 0,\ \forall i}{}{\nonumber} \addConstraint{ z^b_{i} \cdot z^s_{i} = 0 }{,\ \forall i}{}{\nonumber} \addConstraint{g_{ih} - G_{ih}}{\leq 0,\ \forall i, h \in H_i}{}{\nonumber} \addConstraint{y_i-\sum_{h \in H_i}g_{ih} - (z^s_i-z^b_i) +d_i}{=0, \ \forall i}{}{\nonumber} \addConstraint{\sum_{i} y_{i}}{= 0}{}{\nonumber} \addConstraint{\sum_i PTDF_{ki} y_i}{\leq T_k,\ \forall k}{}{\nonumber} \addConstraint{-\sum_i PTDF_{ki} y_i}{\leq T_k,\ \forall k}{}{\nonumber} \addConstraint{\sum_i \big(\textcolor{black}{-z_i^s\tau^s} + z_i^b\tau^b +d_i\tau^b\big)+\sum_{i, j \in {pro,con}} \phi^j_i \geq \Tau}{}{}{\nonumber} \addConstraint{d_i,\ l_{i},\ g_i,\ g_{ih}}{\geq 0,\ \forall i, h \in H_i}{}{\nonumber} \addConstraint{y_i,\ \tau^s,\ \tau^b,\ \phi^j_i }{\ \mathrm{free},\ \forall i, j.}{}{\nonumber} \end{maxi!} \fi \section{Theoretical Results} \label{sec:Theo} \subsection{Bilevel Reformulation} The idea of reformulating the MPEC into a BLP is simple: to replace the lower-level market equilibrium conditions with a centralized optimization problem. This is based on the well-known first fundamental theorem in welfare economics, which states that a competitive equilibrium leads to a Pareto efficient market outcome. More specifically, let the set $\bm{X}$ denote the feasible region of the lower-level variables $\bm{x}$; that is, $\bm{X} = \{\bm{x}: \eqref{eq:nb-con1} - \eqref{eq:nb-con2}, \eqref{eq:iso_con1} - \eqref{eq:nb-price}\}$. We can define the following optimal value function: \begin{equation}\label{eq:LowerNLP} V(\tau^b, \tau^s) := \underset{\bm{x} \in \bm{X}}{max} \ \Pi\left[\bm{x}; (\tau^b, \tau^s)\right]. \end{equation} Let $\bm{x}(\tau^b,\tau^s)$ denote the optimal solution of the above optimization problem with respect to a given $(\tau^b, \tau^s)$, which is a set-valued mapping in general. However, we state below that under mild conditions, the mapping is a singleton. \begin{lemma}\label{lem:unique} With the inverse demand functions as in \eqref{demand} and \eqref{demandp} and under Assumption \ref{as:QuadraticCost}, for any given $(\tau^b, \tau^s)$ with $\tau^b \geq 0$ and $\tau^s \leq \tau^b$, an optimal solution of the optimization problem in \eqref{eq:LowerNLP} exists. In addition, the vectors $d, l, g$ and the quantity $z^s - z^b$ in such an optimal solution are all unique. \end{lemma} The proof is relatively straightforward, with the existence following from the coercieveness of the objective function over the feasible region $\bm{X}$, and the uniqueness as the result of strict convexity of the objective function. Details are omitted here. The complete BLP model is as follows. \begin{subequations} \begin{align} \underset{\tilde{\tau}^b,\ \tilde{\tau}^s,\ \phi^{con},\ \phi^{pro}}{\mathrm{Minimize}} & B\left[\tilde{\tau}^b, \tilde{\tau}^s, \phi^j_i; d(\tilde{\tau}^b, \tilde{\tau}^s), z^b(\tilde{\tau}^b, \tilde{\tau}^s), p(\tilde{\tau}^b,\tilde{\tau}^s)\right] \label{eq:UpperObj} \\[-10pt] \mathrm{subject\ to} \quad & \sum_{i=1}^N \Big[-z_i^s(\tilde{\tau}^b, \tilde{\tau}^s)\tilde{\tau}^s + z_i^b(\tilde{\tau}^b, \tilde{\tau}^s)\tilde{\tau}^b + \nonumber \\ & \quad \quad d_i(\tilde{\tau}^b, \tilde{\tau}^s)\tau^b+\sum_{j \in {pro,con}} \phi^j_i \Big] = \Tau, \label{eq:BL_RevAdq} \\[-5pt] & \tilde{\tau}^s \leq \tilde{\tau}^b,\ \tilde{\tau}^b,\ \phi^{con},\ \phi^{pro} \geq 0,\ \\ & (\tilde{\tau}^b, \tilde{\tau}^s) \in \underset{(\tau^b, \tau^s)}{\arg\max}\ V(\tau^b, \tau^s), \label{eq:OptimalTau} \\ & (d_i(\tilde{\tau}^b, \tilde{\tau}^s), z_i^s(\tilde{\tau}^b, \tilde{\tau}^s), z_i^b(\tilde{\tau}^b, \tilde{\tau}^s), p(\tilde{\tau}^b, \tilde{\tau}^s)) \nonumber\\ & \in \bm{X}_{KKT}(\tilde{\tau}^b, \tilde{\tau}^s), \label{eq:UpperLastConstr} \end{align} \end{subequations} where the set $\bm{X}_{KKT}(\tilde{\tau}^b, \tilde{\tau}^s)$ represents the KKT system of the lower-level optimization problem \eqref{eq:LowerNLP} with respect to $(\tilde{\tau}^b, \tilde{\tau}^s)$. Under the assumptions of Lemma \ref{lem:unique}, the lower-level optimization problem is clearly a convex optimization problem with all linear constraints. Hence, the KKT conditions are necessary and sufficient optimality conditions. By Lemma \ref{lem:unique}, with the uniqueness of optimal solutions corresponding to a given $(\tau^b, \tau^s)$, constraint \eqref{eq:BL_RevAdq} is well defined. However, the objective function \eqref{eq:UpperObj} may not be. The issue is $p(\tilde{\tau}^b, \tilde{\tau}^s)$, which is the Lagrangian multiplier (i.e., the LMP) associated with the flow balancing constraint \eqref{eq:nb-price} may not be unique, even when the primal variables are unique. If the set of multipliers is not a singleton, it is understood that the optimization problem chooses a $p(\tilde{\tau}^b, \tilde{\tau}^s)$ to minimize the objective function \eqref{eq:UpperObj}. To further ensure the validity of the BLP, we need to ensure the attainability of maximum of the optimal value function $V(\tau^b, \tau^s)$. To do so, we first argue that $(\tau^b, \tau^s)$ should be in a bounded polyhedral set. We observe that, based on the KKT conditions \eqref{eq:d_KKT}, \eqref{eq:ProKKT_l}, and \eqref{eq:Pro_zb}, if $\tau^b > \max_{i=1, \ldots, N}\{P_i^0\}$, that is, if the volumetric charge is greater than the largest willingness-to-pay of any consumers/prosumers, then all the $d_i$'s and $l's$ will be 0 in an optimal solution, indicating no energy consumption at all due to the high costs. To avoid this, $\tau^b$ should be bounded above by $\max_{i=1, \ldots, N}\{P_i^0\}$, which we denote by $\widehat{\tau}^b$. Similarly, if $\tau^s$ is too negative, no prosumers would sell their self-produced energy, and all $z^s_i$ would be 0 in an optimal solution. Hence, we assume that $\tau^s$ is implicitly bounded below by $\underline{\tau^s} < 0$. Define the following set: \begin{equation} \label{eq:setT} \mathcal{T} := \{(\tau^b, \tau^s): 0\leq \tau^b \leq \widehat{\tau}^b,\ \underline{\tau}^s \leq \tau^s \leq \tau^b\} \in \Re^2, \end{equation} which is a bounded polyhedron. For the optimal value function $V(\tau^s, \tau^b)$, we want to show that it is continuous over $\mathcal{T}$. Under Assumption \ref{as:QuadraticCost}, $V(\tau^b, \tau^s)$ is the optimal value function of a convex quadratic program with parameterization only on the linear term in the objective function. Properties of such functions are well established, and in this specific case, it is known that $V(\tau^b, \tau^s)$ is a continuous function. (See Theorem 47 in \cite{TerlakyOptValueFunc}.) Therefore, an optimal solution of $\max_{(\tau^b, \tau^s)\in \mathcal{T}} V(\tau^b, \tau^s)$ exists and is attainable, again by the Weierstrass Extreme Value Theorem. \vspace{-1em} \subsection{Equivalence and Optimal Solutions} Note that the relationship between the MPEC model and the BLP model is different than their usual relationship. In a typical BLP, writing out the lower-level problem's KKT conditions will lead to an MPEC formulation. There, the upper-level decision maker does not optimize the lower-level objective function; that is, there is no optimization of the optimal value function $V(\tau^b, \tau^s)$ in \eqref{eq:OptimalTau}. It is needed here since the upper-level decision maker wants to maximize social surplus as well as maintaining equity among all energy consumers. Because of \eqref{eq:OptimalTau}, the BLP formulation \eqref{eq:UpperObj} -- \eqref{eq:UpperLastConstr} is not equivalent to the MPEC formulation \eqref{eq:MPEC} in general, due to the multi-objective decision-making in \eqref{eq:MPEC}; that is, there can be a solution that leads to a lower EEI-difference function $B$, but does not maximize the social surplus $\Pi$. In the following, we show that if an optimal solution exists for each model, the MPEC and the BLP formulations are indeed equivalent when there are no additional constraints on the volumetric and fixed charges. We then show that an optimal solution of the BLP does exist, with the optimal $\tau^{b^*} = \tau^{s^*} = 0$. \begin{proposition}\label{pr:Equiv} Assume that both the MPEC model \eqref{eq:MPEC} and the bilivel optimization model \eqref{eq:UpperObj} -- \eqref{eq:UpperLastConstr} have an optimal solution, which are denoted as $(\bm{x}^{MPEC}, \bm{\zeta}^{MPEC})$ and $\bm{\zeta}^{BLP}$, respectively. Under the assumptions in Lemma \ref{lem:unique}, and if the LICQ holds at $\bm{x}(\bm{\zeta}^{BLP})$ in the set $\bm{X}$, then the MPEC model \eqref{eq:MPEC} and the BLP model \eqref{eq:UpperObj} -- \eqref{eq:UpperLastConstr} are equivalent. \end{proposition} Due to the page limit, we can only show the sketch of the proof. The keys are: (i) the centralized optimization problem of $\max_{\bm{x} \in \bm{X}}\Pi\left[\bm{x}; (\tau^b, \tau^s)\right]$ and the KKT systems (\ref{eq:d_KKT}), (\ref{eq:ProKKT_l})-(\ref{eq:ProKKT_delta}), (\ref{eq:foc-iso-x})-(\ref{eq:foc-iso-g}) are equivalent, under the conditions in this proposition; and (ii) with an arbitrary $(\tau^b, \tau^s) \in \mathcal{T}$, the EEI-difference function $B$ can be minimized to 0 by equating the EEI of conventional consumers and prosumers, which, together with the constraint \eqref{eq:nb}, leads to the following linear system of equations with respect to $(\phi^{con}, \phi^{pro})$: \begin{align} \frac{(p_i + \tau^b)d_i+\phi^{con}_i}{\textcolor{black}{I^{con}_i}} = \frac{(p_i + \tau^b)z^b_i+ \phi^{pro}_i+C_i(g_i) + SC_i}{\textcolor{black}{I^{pro}_i}}, \nonumber\\ \mathrm{for\ all\ } i = 1, \ldots, N \nonumber\\ \sum_{i=1}^N \Bigg(\textcolor{black}{-z_i^s\tau^s} + z_i^b\tau^b +d_i\tau^b +\sum_{j \in {pro,con}} \phi^j_i \Bigg)= \Tau. \label{eq:EqualIncidence} \end{align} The above system has $N + 1$ equations but $2N$ variables. The coefficient matrix clearly has a full-row rank. Hence, for any given $(\tau^b, \tau^s)$, a solution of $(\phi^{con}, \phi^{pro})$ always exists. Proposition \ref{pr:Equiv} is established under the assumption that an optimal solution exists for both the MPEC and the BLP. In the following, we show that the BLP indeed has an optimal solution and in such an optimal solution, $\tau^{b^*} = \tau^{s^*} = 0$. This solution is trivially optimal for the MPEC model as well. \begin{proposition}\label{pr:OptimalTau} Under the assumptions of Lemma \ref{lem:unique} and Assumption \ref{as:QuadraticCost}, we have $\max_{(\tau^b, \tau^s) \in \mathcal{T}} V(\tau^b, \tau^s) = V(0, 0)$. \end{proposition} \begin{proof} See appendix. \end{proof} \begin{remark} With $\tau^{b^*} = \tau^{s^*} = 0$, as argued in the proof of Proposition \ref{pr:Equiv}, the system of equations \eqref{eq:EqualIncidence} always has a solution, which will make the $B$ function equal to zero; hence, they together form an optimal solution of the BLP formulation, which can be easily seen to be a globally optimal solution of the MPEC formulation. Obtaining such a solution can be done in two steps: first, solve the lower-level market equilibrium (either by solving a complementarity problem or an optimization problem as in \eqref{eq:LowerNLP}, with $(\tau^b,\tau^s) = (0,0)$; second, solve the linear system equations \eqref{eq:EqualIncidence}. Both steps can now be done with efficient algorithms since no non-convex problems are involved. One thing to note is that since the system equations have fewer equations than the variables when $N > 1$, the solution $\phi$'s are not unique. In this case, we can minimize $||\phi||_2^2$ over all solutions, which is a convex quadratic program and can be solved efficiently. \end{remark} \begin{remark} Proposition \ref{pr:OptimalTau} is consistent with a well-known fact in economics: under perfect competition, \emph{laissez-faire} is the first-best equilibrium. Any type of tax or subsidy can only reduce economic efficiency and social surplus. A fixed-charge-only tariff, however, would likely to encourage excessive energy use, as customers' energy bills are not tied to energy usage at all. Hence, in reality, almost all tariffs consist of both volumetric and fixed charges. In Section \ref{sec:case}, we numerically study the impact of different levels of volumetric charges on market outcomes and energy equity. However, one thing to note is that with additional requirements (aka constraints)on the fixed and volumetric charges, the system of equations \eqref{eq:EqualIncidence} (together with the additional constraints) may no longer have a solution, which means that the EEI difference function can no longer be zero. (This is indeed what we observe in certain numeric cases.) In this case, the equivalence between the MPEC and the BLP model may also break down. \vspace{-1.2em} \end{remark} \subsection{Model and Theoretical Results under Stochasticity} The above analyses are performed based on deterministic models. In this section, we show that the main result in Proposition \ref{pr:OptimalTau} still holds under uncertainties, regardless of the probability distributions. To do so, however, we need to present a formulation under stochasticity. Let $\bm{\xi} \in\Re^w$ denote the collection of all the random variables with a generic dimension of $w$ (such as prosumers' solar output $R_i$, generation capacity $G_h$, etc). It is defined in the probability space $(\Omega, \mathscr{A}, \mathcal{P})$, with $\Omega$ being the sample space of all the uncertainties, $\mathscr{A}$ the sigma field of $\Omega$ and $\mathcal{P}$ a probability measure in $(\Omega, \mathscr{A})$; that is, $\bm{\xi}$ is a measurable mapping from $\Omega$ to a set $\Xi \in \Re^w$. Let $\bm{x}^{\bm{\xi}}$ and $\bm{X}(\bm{\xi})$ respectively denote the lower-level variables and feasible region corresponding to a realization of the random vector $\bm{\xi}$. Furthermore, let $\Pi[\bm{x}^{\bm{\xi}}; (\tau^b, \tau^s), \bm{\xi}]$ represent the lower-level objective function under uncertainty, with a given volumetric rate $(\tau^b, \tau^s)$. Then we can define the expected optimal value function of the lower-level problem as follows: \begin{equation}\label{eq:EV} EV(\tau^b, \tau^s) := \mathbb{E}_{\bm{\xi}} \left\{\underset{\bm{x}^{\bm{\xi}} \in \bm{X}(\bm{\xi})}{max}\Pi\left[\bm{x}^{\bm{\xi}}; (\tau^b, \tau^s), \bm{\xi}\right]\right\}. \end{equation} To ensure that the above expectation is well-defined, we need the following assumption. \begin{assumption}\label{as:FiniteMoment} All the random functions in $\Pi\left[\bm{x}^{\bm{\xi}}; (\tau^b, \tau^s), \bm{\xi}\right]$ and $\bm{X}(\bm{\xi})$ have finite moments. \end{assumption} Under Assumption \ref{as:FiniteMoment}, for any $(\tau^b, \tau^s)\in\mathcal{T}$, with $\mathcal{T}$ defined in \eqref{eq:setT}, the expectation in \eqref{eq:EV} is well defined following the same argument as in \cite{PangSen}. With a finite expected value in \eqref{eq:EV}, since expectation preserves convexity (see, for example, Sec. 3.2.1 in \cite{Boyd_ConvexOpt}), we have that $EV(\tau^b, \tau^s)$ is also a convex function with respect to $(\tau^b, \tau^s) \in \mathcal{T}$. Hence, the extension of Proposition \ref{pr:OptimalTau} to the stochastic case is straightforward, and we only state the result below, omitting the proof. \begin{proposition}\label{pr:SOptimalTau} Under the assumptions of Lemma \ref{lem:unique}, and Assumption \ref{as:QuadraticCost} and \ref{as:FiniteMoment}, we have $\max_{(\tau^b, \tau^s) \in \mathcal{T}} EV(\tau^b, \tau^s) = EV(0, 0)$. \hfill$\Box$ \end{proposition} With the lower-level optimal value function defined, we can write out the stochastic version of the upper-level problem now. Note that the upper-level decisions on the retail rates have to be made before lower-level uncertainties are realized; that is, they are the here-and-now type of decisions. Since the lower-level optimal solutions $d$, $z^b$, and $p$ depend on the random vector $\bm{\xi}$, if we simply require that the revenue adequacy constraint \eqref{eq:nb} is held for all $\bm{\xi}$ (almost surely), it will likely to be always infeasible. To remedy this, a natural idea is to make constraint \eqref{eq:nb} a chance constraint as follows: \begin{align} \underset{\tilde{\tau},\ \phi}{min}\ & \mathbb{E}_{\bm{\xi}}\left[B\left(d_i^{\bm{\xi}}(\tilde{\tau}^b, \tilde{\tau}^s), \tilde{\tau}^b, \phi^j_i\right)\right] \label{eq:SBilevelObj}\\ \mathrm{s.t.} \quad \quad \ & \mathcal{P}\left\{\sum_i \Big[-z_i^{s^{\bm{\xi}}}(\tilde{\tau}^b, \tilde{\tau}^s)\tilde{\tau}^s + z_i^{b^{\bm{\xi}}}(\tilde{\tau}^b, \tilde{\tau}^s)\tilde{\tau}^b + \right.\nonumber \\ & \left. \quad d_i^{\bm{\xi}}(\tilde{\tau}^b, \tilde{\tau}^s)\tau^b+\sum_{j \in {pro,con}} \phi^j_i \Big] \geq \Tau \right\} \geq 1 - \epsilon \nonumber \\ & (\tilde{\tau}^s, \tilde{\tau}^b) \in \underset{(\tau^s, \tau^b)}{\arg\max}\ EV(\tau^s, \tau^b), \nonumber \end{align} where $\epsilon$ is a pre-specified parameter. Solving the chance-constrained stochastic program (SP) \eqref{eq:SBilevelObj} is much more involved than its deterministic counterpart, even when Proposition \ref{pr:SOptimalTau} holds. In this case, sample average approximation (SAA) methods developed for chance-constrained SPs, such as \cite{SAA_CC}, can be directly applicable here. However, if there are other requirements on the volumetric charges such that they cannot be zero, then specialized (and likely iterative) algorithms need to be developed to solve \eqref{eq:SBilevelObj}, since it also includes an optimal value function in its constraints. Developing such algorithms (and presenting numerical results under stochasticity) is deferred to future research. \vspace{-0.5em} \section{Numerical Case Study \vspace{-0.5em}} \label{sec:case} \subsection{Setup} To illustrate the effects of optimal pricing schemes, we apply the models developed in Section \ref{sec:mod} to a case study considering a three-node network with three firms, ten generating units, and three transmission lines. This setup is sufficiently general because it allows firms compete across different locations subject to transmission constraints. The three-node network is the simplest that allows for looped flows, which is important in modeling a power grid. We assume that a daily fixed cost of \$80k to be reimbursed to a utility. Consumers are grouped into three income levels: high, medium and low, residing at nodes A, B, and C, respectively. The baseline daily demand of the low-income group is assumed to be 20 kWh. The daily demand of the medium- and high-income groups are assumed to be 25\% and 50\% larger than the low-income group, broadly consistent with the data from the 2015 RECS survey \cite{recs15}. Given the assumed fixed demand in each node, we then recover the number of households in each income group. (Proportion of households among income groups are also compatible with the 2015 RECS results.) The income level is obtained by assuming that electricity expenditure is 1.5\% of the income in each group.\footnote{This represents the ``energy equity'' case at the baseline. The 1.5\% is at the lower end based on 2015 RECS. However, our interest lies on the relative changes of the energy incidence when the power sales by prosumers are subject to different tariff designs. This assumption is not essential.} Finally, based on RECS 2015, we assume that 20\% of the households in high-income group (or 3,067 households) own rooftop solar energy with a capacity of 8 KW each household or 25 MW in total. Thus, in the extreme case during a sunny summer day, we consider a daily solar output from prosumers to be $R = 150$ MWh; while in a cloudy/raining winter day, it generates only $R=25$ MWh. The numbers are carefully chosen with one corresponding to insufficient DER generation for prosumers, while the other with excess DER generation from prosumers. For the sunk cost $SC$ to be included in prosumers' energy expenditure in \eqref{eq:inc_pro}, we assume a daily cost \$5/day.\footnote{This is calculated approximately based on \$30K initial investment on solar panels (including installation) and an assumed break-even period of 15 years. This is also in the ballpark of rental costs of solar panels.} A 25 MW backup generator or energy storage is assumed for the prosumers. The MPEC formulation is used for computation, and it is written in AMPL and solved by Knitro solver version 12.4 on a Mac Book Pro with 2.8 GHz Quad-Core and Intel Core i7. \vspace{-1em} \subsection{Main Results} First, we add additional constraints to the upper-level problem and require that a certain percentage of utility's fixed cost $\Tau$ to be recovered through volumetric charge. More specifically, we consider two cases: 10\% and 90\% of $\Tau$ to be recovered by volumetric charges. Later we will compare such results with the case without the additional constraints (i.e., the original formulation as in \eqref{eq:MPEC}). The results of the volumetric charge $= 10\% \Tau$ and $90\%\Tau$ for the cases with renewable output equal to 25 MWh and 150 MWh are reported in Tables \ref{tab:25} and \ref{tab:150}, respectively. Table \ref{tab:25} indicates a significant increase in volumetric charge when a higher percentage of $\Tau$ is required to be recovered from the actual energy use. Specifically, $\tau^b$ increases from \$5.25/MWh under $10\%\Tau$ to \$49.18/MWh under $90\%\Tau$. In the case of 25 MWh of DER output, prosumers in both cases are in a net-buying position. The prosumers benefit from the case of requiring a higher percentage of volumetric charge, with an increase of surplus from \$73.47K to \$78.23K and a decline of energy expenditure incidence from 1.34\% to 1.29\%. Note that prosumers' energy incidence is less than that of consumers under the $90\%\Tau$ case; meaning that with this requirement, the system of equations \eqref{eq:EqualIncidence} no longer has a solution of $(\phi^{con}, \phi^{pro})$ that can lead to zero value of the EEI difference function $B$. This indicates that energy equity can no longer be maintained across different income groups. The generation from the wholesale market is 1,519.17 MWh and 1,464.09 MWh, under the 10\%- and 90\%-volumetric charge requirement, respectively. This is so because the higher energy prices faced by consumers in the $90\%\Tau$ case suppress demand, including that of prosumers, by a margin of 3.6\%. \begin{table}[!ht] \renewcommand{\arraystretch}{0.9} \caption{Results of 25 MWh Renewable case: 10\% vs. 90\% volumetric charge} \resizebox{\columnwidth}{!}{ \label{tab:pc} \centering \begin{tabular}{l|rrrrrr} \hline Variables\textbackslash Volumetric charge & \multicolumn{3}{r}{10\%} & \multicolumn{3}{r}{90\%} \\\hline Volumetric charge [\$/MWh] & \multicolumn{3}{r}{5.27} & \multicolumn{3}{r}{49.18} \\ Prosumer's sale(+)/buy(-) [MWh] & \multicolumn{3}{r}{-45.68 } & \multicolumn{3}{r}{-43.10} \\ Prosumer's load [MWh] & \multicolumn{3}{r}{95.68} & \multicolumn{3}{r}{93.10} \\ Backup generation [MWh] & \multicolumn{3}{r}{25.0} & \multicolumn{3}{r}{25.0} \\ Prosumer surplus [\$K] & \multicolumn{3}{r}{73.47} & \multicolumn{3}{r}{78.23} \\ Prosumer incidence [\%] & \multicolumn{3}{r}{1.34} & \multicolumn{3}{r}{1.29}\\ Prosumer fixed charge [\$/household] & \multicolumn{3}{r}{2.18} & \multicolumn{3}{r}{0.00031} \\\hline Variables\textbackslash Nodes & A & B & C & A & B & C \\ \hline Conventional demand [MWh] & 382.70 & 592.71 & 498.08 & 372.38 & 577.39 & 469.46 \\ Fixed Charge [\$/household] & 1.29 & 1.25& 0.93&0.00 &0.35 & 0.01\\ Power price [\$/MWh] & 77.93 & 52.63 & 37.84 & 77.66 & 40.20 & 37.50 \\ Energy incidence [\%] & 1.34 & 1.34 & 1.34 & 1.32 & 1.27 &1.28\\ Conventional generation [MWh] & \multicolumn{3}{r}{1,519.17} & \multicolumn{3}{r}{1,464.09} \\ \hline Total consumer surplus [\$K] & \multicolumn{3}{r}{806.51} & \multicolumn{3}{r}{807.80} \\ Producer surplus [\$K] & \multicolumn{3}{r}{11.51} & \multicolumn{3}{r}{4.38} \\ ISO's revenue [\$K] & \multicolumn{3}{r}{1.58} & \multicolumn{3}{r}{1.33} \\ Wholesale surplus [\$K] & \multicolumn{3}{r}{819.60} & \multicolumn{3}{r}{813.51} \\ \hline Total social surplus [\$K]& \multicolumn{3}{r}{893.07} & \multicolumn{3}{r}{891.74} \\ \hline \end{tabular}} \label{tab:25} \end{table} We now turn to discuss Table \ref{tab:150}, the case when the DER output equals 150 MWh. In this case, prosumers have excess output and are in a net-selling position, while subject to $\tau^s$ when selling to the grid. The fact that $\tau^s <0$ (-\$3.99/MWh and -\$13.97/MWh under the $10\%\Tau$ and $90\%\Tau$ cases, resp.) indicates that prosumers should contribute to recovering fixed costs when selling power to the grid (i.e., net billing with sales payment less than the utility retail rate is more optimal than net metering). This is consistent with the recommendation by the CPUC concerning the recent debate to revamp the net-metering policies in California \cite{borenstein22}. Overall, the same observations about the surplus distribution as in Table \ref{tab:25} emerge in Table \ref{tab:150}. More importantly, the $90\%\Tau$ requirement leaves the PUC with no adequate fund to maintain energy equity, leading to a divergence of energy incidence across different income groups. \vspace{-1em} \begin{table}[!ht] \renewcommand{\arraystretch}{0.9} \caption{Results of 150 MWh Renewable case: 10\% vs. 90\% volumetric charge} \resizebox{\columnwidth}{!}{ \label{tab:pc} \centering \begin{tabular}{l|rrrrrr} \hline Variables\textbackslash Volumetric Charges & \multicolumn{3}{r}{10\%} & \multicolumn{3}{r}{90\%} \\\hline Volumetric charge [\$/MWh] & \multicolumn{3}{r}{ 5.21 (-3.99)} & \multicolumn{3}{r}{\textcolor{black}{49.93 (-13.97)}} \\ Prosumer's sale(+)/buy(-) [MWh] & \multicolumn{3}{r}{ 78.70} & \multicolumn{3}{r}{78.10} \\ Prosumer's load [MWh] & \multicolumn{3}{r}{96.30} & \multicolumn{3}{r}{\textcolor{black}{96.90}} \\ Backup generation [MWh] & \multicolumn{3}{r}{25.0} & \multicolumn{3}{r}{25.0} \\ Prosumer surplus [\$K] & \multicolumn{3}{r}{79.57} & \multicolumn{3}{r}{\textcolor{black}{80.93}} \\ Prosumer incidence [\%] & \multicolumn{3}{r}{1.30} & \multicolumn{3}{r}{1.06} \\ Prosumer fixed charge [\$/household] & \multicolumn{3}{r}{3.31} & \multicolumn{3}{r}{2.61} \\ \hline Variables\textbackslash Nodes & A & B & C & A & B & C \\ \hline Conventional demand [MWh] & 383.02 & 593.02 & 498.11 & \textcolor{black}{372.49} & \textcolor{black}{577.02} & \textcolor{black}{470.76} \\ Fixed Charge [\$/household] & 1.12 & 1.19& 0.88&0.00 &0.00 & 0.00\\ Power price [\$/MWh] & 76.64 & 52.03 & 37.85 & \textcolor{black}{76.47} & 40.20 & 37.49 \\ Energy incidence [\%] & 1.30 & 1.30 & 1.30 & 1.32 & 1.11 &1.28\\ Conventional generation [MWh] & \multicolumn{3}{r}{1,395.48} & \multicolumn{3}{r}{\textcolor{black}{ 1,342.17}} \\ \hline Total consumer surplus [\$K] & \multicolumn{3}{r}{810.90} & \multicolumn{3}{r}{\textcolor{black}{815.18}} \\ Producer surplus [\$K] & \multicolumn{3}{r}{10.73} & \multicolumn{3}{r}{3.99} \\ ISO's revenue [\$K] & \multicolumn{3}{r}{1.52} & \multicolumn{3}{r}{1.3} \\ Wholesale surplus [\$K] & \multicolumn{3}{r}{823.15} & \multicolumn{3}{r}{\textcolor{black}{820.47}} \\ \hline Total social surplus [\$K]& \multicolumn{3}{r}{ 902.72} & \multicolumn{3}{r}{\textcolor{black}{901.41}} \\ \hline \end{tabular}} \label{tab:150} \end{table} For comparison purposes, Table \ref{tab0} provides the results without any requirements on volumetric or fixed charges. As proven in Proposition \ref{pr:OptimalTau}, the optimal $\tau^{b^*} = \tau^{s^*} = 0$, and the total social surplus in the cases of $R= 25$ MWh and $R = 150$ MWh are higher than their counterparts in Tables \ref{tab:25} and \ref{tab:150}, consistent with the theoretical results in Section \ref{sec:Theo}. \begin{table}[!ht] \renewcommand{\arraystretch}{0.9} \caption{Results of 25 and 150 MWh Renewable case under 0\% volumetric charge} \resizebox{\columnwidth}{!}{ \label{tab:pc} \centering \begin{tabular}{l|rrrrrr} \hline Variables\textbackslash Renewable & \multicolumn{3}{r}{25 MWh} & \multicolumn{3}{r}{150 MWh} \\\hline Volumetric charge [\$/MWh] & \multicolumn{3}{r}{0 (0)} & \multicolumn{3}{r}{0 (0)} \\ Prosumer's sale(+)/buy(-) [MWh] & \multicolumn{3}{r}{-45.97 } & \multicolumn{3}{r}{78.94} \\ Prosumer's load [MWh] & \multicolumn{3}{r}{95.97} & \multicolumn{3}{r}{96.06} \\ Backup generation [MWh] & \multicolumn{3}{r}{25.0} & \multicolumn{3}{r}{25.0} \\ Prosumer surplus [\$K] & \multicolumn{3}{r}{73.24} & \multicolumn{3}{r}{79.62} \\ Prosumer incidence [\%] & \multicolumn{3}{r}{1.36} & \multicolumn{3}{r}{1.33} \\ Prosumer fixed charge [\$/household] & \multicolumn{3}{r}{2.33} & \multicolumn{3}{r}{3.40} \\\hline Variables\textbackslash Nodes & A & B & C & A & B & C \\ \hline Conventional demand [MWh] & 383.94 & 592.70 & 501.30 & 384.25 & 593.02 & 501.31 \\ Fixed Charge [\$/household] & 1.52 & 1.31& 1.08& 1.46 & 1.25 & 1.03\\ Power price [\$/MWh] & 77.95 & 57.91 & 37.88 & 76.64 & 57.26 & 37.88 \\ Energy incidence [\%] & 1.36 & 1.36 & 1.36 & 1.33 & 1.33 &1.33\\ Conventional generation [MWh] & \multicolumn{3}{r}{1,523.94} & \multicolumn{3}{r}{1,399.64} \\ \hline Total consumer surplus [\$K] & \multicolumn{3}{r}{803.58} & \multicolumn{3}{r}{807.73} \\ Producer surplus [\$K] & \multicolumn{3}{r}{14.45} & \multicolumn{3}{r}{13.63} \\ ISO's revenue [\$K] & \multicolumn{3}{r}{1.80} & \multicolumn{3}{r}{1.74} \\ Wholesale surplus [\$K] & \multicolumn{3}{r}{819.83} & \multicolumn{3}{r}{823.11} \\ \hline Total social surplus [\$K]& \multicolumn{3}{r}{893.08} & \multicolumn{3}{r}{902.73} \\ \hline \end{tabular}} \label{tab0} \end{table} \textcolor{black}{Figure \ref{fig:2} plots the rent distribution among various entities in the market against the fraction of fixed costs assigned to the volumetric rate. The palpable difference of prosumer's surplus in Figure \ref{fig:2}(e) of R=25 MWh cf. R=150 MWh cases affects the wholesale market surplus in Figure \ref{fig:2}(b) and deserves some explanation. When R=25 MWh, the prosumers do not have enough DER generation and need to buy energy from the grid. An increase in the volumetric charge in x-axis provides an opportunity for the prosumer to ``avoid'' fixed charge via increasing self reliance or reducing consumption. As a result, its surplus steadily increases until the volumetric charge becomes 100\%, where $\tau^b=\$60$/MWh or nearly 50\% of the retail price at node A, an unbearably high level.} \textcolor{black}{On the contrary, prosumers sell energy to the grid when R=150MWh. Its surplus in this situation is affected by two counteracting forces. On the one hand, its increased contribution to fixed cost through $\tau^s$ reduces its surplus until $\tau^s$ drops to approximately -\$10/MWh; beyond which its surplus is adequately offset by a decline in fixed charge (or an increase in the volumetric charge fraction), leading its surplus to be leveled around \$80k until the volumetric charge fraction is greater than $90\%\Tau$. When the volumetric charge fraction is 100\%, the prosumer can avoid the fixed cost entirely, leading to a surge in its surplus as shown in Figure \ref{fig:2}(e).} \textcolor{black}{With regard to consumers, their surplus is affected by retailed power prices and allotted fixed costs. As alluded to earlier, forgo consumption is one way to minimize the impact of fixed cost recovery. This strategy effectively enhances consumers' surplus until the allotted fraction of volumetric charge equal to 40\%. However, beyond this level, the retail power prices become too high, leading to a drop in the surplus.} \textcolor{black}{Overall, we observe that the wholesale surplus in Figure \ref{fig:2}(b) continues to decline with the increase in the volumetric fraction under R=25MWh, forming a concave curve due to the impacts on consumer surplus in \ref{fig:2}(c). Finally, the changes in consumers surplus is ``neutralized'' by the changes of prosumers' surplus in Figure \ref{fig:2}(e) under R=150MWh, leading to the total surplus in Figure \ref{fig:2}(a).} To zoom in on energy equity, Figure \ref{fig:4} shows the energy incidence by income groups, e.g., low, medium, high and prosumers. When $R = 25$ MWh, energy equity can be maintained when the revenue levied from volumetric charge can be less than or equal to $80\% \Tau$; beyond that, the different fixed charges $\phi$ collected from different income groups can no longer be used to maintain energy equity. The similar trend is also seen for $R = 150$ MW. \vspace{-1em} \begin{figure}[htbp] \centering \includegraphics[width=3.45in]{./figs/fig1.pdf}\vspace{-1em} \caption{Distribution of surplus} \label{fig:2} \end{figure} \begin{figure}[htbp] \centering \includegraphics[width=3.45in]{./figs/fig3.pdf}\vspace{-1em} \caption{Plots of energy incidences}\vspace{-1.5em} \label{fig:4} \end{figure} \section{Conclusion \vspace{-3pt}}\label{sec:con} Recovering utilities' fixed costs has presented a significant regulatory challenge in designing electricity tariffs. While emphasis has typically been placed on economic efficiency and incentive to conserve energy, equally important are their impacts on the energy equity among different income groups. The situation is further exacerbated by the presence of prosumers who are typically among the most affluent income groups, taking advantage of the electricity tariff, adopting new technologies and optimizing their self-interests. This study examines the optimal retail tariff in the presence of prosumers. We demonstrate that a volumetric approach to recover fixed costs based on energy consumption is likely to be less efficient. Our analysis concludes that the first-best policy is to leave the wholesale market intact and rely on fixed charge to recover fixed costs. However, such a policy prescription is likely implausible: lower retail prices will likely provide a disincentive for energy conservation, against the effort to decarbonizing economy. In addition, a lumpy fixed charge on low-income households can be challenging to those families who already face harsh economic situations. Therefore, a policy that provides directly financial compensation to low-income households is economically efficient. Our analysis suggests that programs, such as CARE (California Alternate Rate of Energy), designed to mitigate low-income households' energy expenditure via lower retail rates, a volumetric approach, is unlikely to be efficient. \textcolor{black}{Because our analysis is short-run based, which does not consider the interaction between power-system operations and expansion decisions, it is subject to a number of long-run implications. In particular, while the policy is expected to improve conventional consumers' energy expenditure incidence, it may offset their economic incentive to invest in DERs, slowing the development of non-utility-scale DERs.} \textcolor{black}{Moreover, the impact on incidence can also be affected by demand elasticity. When demand is less price-responsive, consumers cannot forgo consumption in response to higher power prices, leading to higher consumption and a lower volumetric tariff.} {

0f78fde602e5246c7d8567ae8565800319c96c7c

\section{Introduction} Graphene has opened up a new era of topological materials, followed by the discovery of topological insulators, topological superconductors, and semi-metals in both two and three dimensions (See reviews~\cite{Hasan2010b, Qi2011, Armitage2018}). Since then, the interplay between topology and correlation has been the primary focus of condensed matter research. Graphene and its variants, due to its excellent electronic and mechanical properties~\cite{Neto2007, DasSarma2011}, have become wonderful platforms for hosting exotic phases of matter and also find themselves widely applicable in electric device engineering and material science. The characteristic feature of graphene is the appearance of the Dirac cone in the spectrum, closely tied to the symmetry of the underlying honeycomb lattice. The two sublattices (A and B) of the honeycomb lattice transform into each other under the simplest non-abelian point group $C_{3v}$, which contains 3-fold rotations and in-plane reflections. At the $K (K')$ point of the Brillouin zone, the wave functions of A and B sublattices form the two-dimensional ($E$) irreducible representations (irrep) of the $C_{3v}$, enforcing the Dirac cones. Once there, the Dirac cones are stable as long as the time-reversal and inversion symmetry are preserved. The on-site $p_z$ orbital of graphene transforms trivially (it belongs to the $A_1$ irrep) under the site symmetry group $C_{3v}$. It is natural to ask what happens if the on-site orbitals form the $E$ irrep of the point group. The $E$ irrep features double degeneracy and strong anisotropy and is expected to bring rich orbital physics in graphene-like Dirac materials. Such a situation arises in many distinct systems. It was initially studied in optical lattices, where the two-dimension irrep is realized by the $p_x$ and $p_y$ orbitals in the harmonic trap~\cite{Wu2007, Wu2008}. In the transition metal oxide heterostructures~\cite{Xiao2011, Ruegg2011, Yang2011a} and transition metal dichalcogenide monolayers~\cite{Qian2014}, the $d$ orbitals of the transition metal decompose based on the $C_{3v}$ symmetry and are active near the Fermi surface. In hexagonal monolayers of heavy elements, such as Germanene, Stanene, and Bismuthene, the $(p_x, p_y)$ doublets realize the orbital degrees of freedom and because of the large atomic spin-orbit coupling, contribute to the large topological gaps~\cite{Xu2013, Wu2014, Zhang2014, Reis2017}. Even in simple carbon systems, orbital physics can be realized via lattice engineering, for example, organic framework~\cite{Wang2013, Wang2013a} and graphene-kagome lattice~\cite{Chen2018}. Remarkably, recently experiments~\cite{Cao2018, Cao2018a,yankowitz2019tuning} on twisted bilayer graphene revealed Mott insulator and superconductivity phases, and it is proposed that the low-lying degrees of freedom are compatible with two orbitals on the honeycomb lattice as well~\cite{Po2018, Yuan2018, Liu2018,venderbos2018correlations,dodaro2018phases,fidrysiak2018unconventional}. Furthermore, the orbital degrees of freedom do not have to be electronic and can manifest themselves as the polarization modes of polaritons in photonic lattices~\cite{Jacqmin2014, Milicevic2017} and phonons in graphene and mechanical structures~\cite{Zhang2015, Roman2015, Stenull2016, Zhu2018}. Given all these interconnected systems and the increasing new realizations of orbital-active Dirac material, this work aims to bridge all the different systems through the symmetry principle. Despite the vastly different origins, the orbital degrees of freedom can be understood as the irreducible representations of the site symmetry of the lattice, which can lead to universal properties. We show that the symmetry alone enforces the Dirac cone at $K (K')$ point and the quadratic band touching at the $\Gamma$ point. We further discuss various gap opening mechanisms and interaction effects, which lead to the quantum spin Hall effect and quantum anomalous Hall effect. In particular, when the $E_g$ doublets realize orbital degrees of freedom, the resulting topological insulator states carry octupole order. Finally, the method we use here to study the doubly-degenerate orbitals in the honeycomb lattice can be readily generalized to the orbital degrees of freedom arising from larger point groups on lattices described by other space groups. The rest of the paper is organized as follows. In Sec.~\ref{sec:honeycomb_orb}, we discuss the symmetry of the honeycomb lattice and the orbital realization of the on-site irreducible representations, focusing on the $d$ orbitals. In Sec.~\ref{sec:band_structure}, we derive the band structure of the orbital-active honeycomb lattice systems from a simple tight-binding model. In Sec.~\ref{sec:general_symmetry}, we go beyond the simple tight-binding model and demonstrate that many interesting features of the band structure are solely protected by the lattice symmetry. In Sec.~\ref{sec:gap_opening}, we investigate various band gap opening mechanisms. In Sec.~\ref{sec:interaction}, the interplay between the band structure and the interaction effects is discussed. Sec.~\ref{sec:summary} is left for summary and outlook. \section{The honeycomb lattice and orbital symmetries} \label{sec:honeycomb_orb} \begin{figure} \includegraphics[height=0.45\columnwidth, width=0.45\columnwidth] {graphene.pdf} \includegraphics[height=0.45\columnwidth, width=0.45\columnwidth] {bilayer.pdf} \caption{(a) The standard honeycomb lattice with $C_{6v}$ point symmetry group. (b) The buckled honeycomb lattice with the symmetry group is downgraded to $D_{3d}$.} \label{fig:lattice} \end{figure} We start with reviewing the symmetry of the planar honeycomb lattice. The planar honeycomb lattice, sketched in in Fig.~\ref{fig:lattice}(a), consists of two sublattices $A$ (blue) and $B$ (red). The three nearest neighbor vectors are labeled as $\hat{e}_1\sim \hat{e}_3$. The symmetry of the lattice is described by the space group $P6mm$, a direct product of the point group $C_{6v}$ and the translation symmetry of the triangular Bravis lattice \footnote{If one also considers the mirror symmetry taking $z$ to $-z$, the point group is $D_{6h}$ and the space group is $P6/mmm$.}. The maximal point group $C_{6v}$ is realized at the centers of the hexagons. On the other hand, the point group symmetry acting on a lattice site, called site symmetry, is a subgroup of the maximal point group. The site symmetry group is important because it affects the orbital part of the wave function of the degrees of freedom living on lattice sites (such as electrons, phonons, etc.) The site symmetry of the honeycomb lattice is $C_{3v}$ generated by a 3-fold rotation axis and three vertical reflection planes (e.g., the $yz$-plane and its symmetry counterparts by rotations of $\pm 120^\circ$). In contrast, the reflection with respect to the $xz$-plane and its symmetry counterparts by rotations of $\pm 120^\circ$ interchange the $A$ and $B$ sublattices and are not included in the site symmetry. The orbital of the onsite degrees of freedom is classified by the irreducible representations of the site symmetry group. The $C_{3v}$ group has three irreducible representations (irrep), including two 1d irreps $A_{1,2}$ and a 2d irrep $E$ as explained in Appendix~\ref{appendix:group}. The irreps fully determines the symmetry structure of the onsite degrees of freedom, regardless of their microscopic origins. In this paper, we mainly focus on electron atomic orbitals. Taking $z$ axis perpendicular to the lattice plane, the $s$ and $p_z$ orbitals realize the $A_1$ irrep and lead to the remarkable electronic structure of graphene. In contrast, the $p_x$ and $p_y$ orbitals realize the two dimensional $E$ irrep. This doublet can also be organized into the complex basis $p_x \pm ip_y$ which are eigenstates of the orbital angular momentum $L_z$ with eigenvalues $\pm 1$, respectively. As to the 5-fold $d$-orbitals, the $d_{r^2-3z^2}$ falls into the $A_1$ irrep. The remaining four form two $E$ irreps: the $(d_{xz}, d_{yz})$ doublet and the $(d_{xy}, d_{x^2-y^2})$ doublet. The complex orbitals $d_{xz}\pm i d_{yz}$, and $d_{xy}\pm d_{x^2-y^2}$ carry orbital angular momentum numbers $\pm 1$ and $\mp 2$ respectively. Since the site symmetry group only has one 2d irrep, the three doublets, $(p_x, p_y)$, $(d_{xz}, d_{yz})$ and $(d_{xy}, d_{x^2-y^2})$ are equivalent as far as the symmetry is concerned. One can explicitly check that the group elements of the site symmetry $C_{3v}$ have the same matrix representation in the basis of all realizations of the $E$ irrep. We consider a closely related lattice structure sketched in Fig~\ref{fig:lattice}(b). This lattice structure, dubbed buckled honeycomb lattice, can be viewed as a bilayer of sites taken from a cubic lattice in the $(1,1,1)$ direction. The blue and red dots form a honeycomb lattice when projecting into the $(1,1,1)$ plane. Compared to the planner honeycomb lattice, the point group symmetry of the buckled lattice downgrades from $d_{6h}$ to $d_{3d}$, where the six-fold rotation becomes a rotoreflection. On the other hand, the site symmetry remains the same, described by $C_{3v}$. As a result, based on previous analysis, the realizations of the $E$ irrep in the buckled lattice must be equivalent to the $(p_x, p_y)$ doublet in the planar case. Here we focus on the $d$ orbitals and establish this equivalence. The buckled honeycomb lattice originates from the cubic lattice. Taking the $z$ along $(0,0,1)$, the 5-fold d orbitals split into a $T_{2g}$ triplet $(d_{yz}, d_{zx}, d_{xy})$ and a $E_g$ doublet $(d_{x^2-y^2}, d_{r^2-3z^2})$, which are irreps of the $O_h$ point group. The site symmetry of the buckled lattice $C_{3v}$ is a subgroup of $O_h$. The $E_g$ doublet falls into the only 2d irrep $E$ of $C_{3v}$, while the $T_{2g}$ triplet further splits into the 1d irrep $A_1$ and the 2d irrep $E$. To make the connection between the $E_g$ doublet and the orbital realization of the $E$ irrep in the planar case more explicit, we can rotate the frame of the buckled lattice so that the $z$ axis is along the 3-fold rotation axis $(1,1,1)$. Under this frame rotation, the $E_g$ doublet becomes \begin{equation}\begin{aligned}} \def\eea{\end{aligned}\end{equation} d_{x^2-y^2}&\rightarrow \frac{1}{\sqrt{3}}(d_{xy}+\sqrt{2}d_{xz}), \\ d_{r^2-3z^2}&\rightarrow \frac{1}{\sqrt{3} }(d_{x^2-y^2}+\sqrt{2} d_{yz}). \label{eq:rotated_eg} \eea Hence, the $E_g$ orbitals are a superposition of two $E$ doublets $(d_{xy}, d_{x^2-y^2})$ and $(d_{xz}, d_{yz})$ in the planar case. Therefore, as far as the site symmetry $C_{3v}$ is concerned, the $e_g$ doublet is equivalent to the $(p_x, p_y)$ doublet in the planar case with the following mapping, \begin{equation}\begin{aligned}} \def\eea{\end{aligned}\end{equation} d_{x^2-y^2} \leftrightarrow p_x, \ \ \, d_{r^2-3z^2} \leftrightarrow p_y. \eea For completeness, we present the decomposition of the five $d$-orbitals into two $E$ irreps and one $A_1$ irrep of the $C_{3v}$ group as follows, \begin{equation}\begin{aligned}} \def\eea{\end{aligned}\end{equation} \begin{cases} (d_{x^2-y^2}, d_{r^2-3z^2}) & E \\ \left (\frac{1}{\sqrt{2}}(d_{yz} -d_{zx}), \frac{1}{\sqrt{6}} (d_{yz}+d_{zx}-2d_{xy})\right ) & E \\ \frac{1}{\sqrt{3}}(d_{xy}+d_{yz}+d_{zx}) &A_1, \\ \end{cases} \label{eq:111_irrep} \eea choosing $(1,1,1)$ as the rotation axis. In addition to the $E_g$ orbitals which become an $E$-representation, the $T_{2g}$-orbitals split into one $E$ irrep and one $A_1$ irrep. In principle, the two $E$-representations from the $E_g$ and $T_{2g}$ orbitals can mix. In transition metal oxide where the buckled lattice is relevant, there is often an oxygen octahedron around each transition metal ion. The oxygen octahedron introduces a large crystal field that splits the $E_g$ and $T_{2g}$ orbitals. As a result, the mixing between the $E$ irrep of $C_{3v}$ derived from the $E_g$ orbitals and that from the $T_{2g}$ orbitals is weak. \section{Magnetic octupole moment of the $E_g$ doublet} \label{sec:octupole} Although all realizations of $E$ irrep of $C_{3v}$ are equivalent from the symmetry consideration. The $E_g$ orbitals are special physically and worth special attention. The key difference lies in the angular momentum of the complex combination of the doublets. In the case of $(p_x, p_y)$, the complex combination $p_x \pm i p_y$ takes the form $\exp(\pm i \theta)$, and thus carry angular momentum $\pm 1$ along the rotating axis. The same applies to the $(d_{xz}, d_{yz})$ doublet. In the case of $(d_{xy}, d_{x^2-y^2})$, the complex combination $d_{xy} \pm i d_{x^2-y^2}$ takes the form $\exp (\mp i 2\theta )$ and thus carries angular momentum $\mp 2$. In contrast, the angular momentum of the complex combination of the $E_g$ orbitals $d_{x^2-y^2}\pm d_{r^2-3z^2}$ vanishes. From Eq.~\eqref{eq:rotated_eg}, the complex combination of $E_g$ doublet can be viewed as the weighted superposition of the complex combinations of the $(d_{xz}, d_{yz})$ and $(d_{xy}, d_{x^2-y^2})$. The angular momentum of the two doublets cancel each other, leading to the zero angular momentum of the $E_g$ doublet. Instead of angular momentum, the complex $E_g$ orbitals carry higher rank magnetic momentum, measured by spherical tensor operators $Y_{lm}$. A list of spherical tensor operators constructed from the angular momentum operator $\vec L$ in the $d$ orbital space can be found in appendix~\ref{appendix:tensor}. Going through all the higher rank tensor operators, we find that the leading non-vanishing spherical tensor operators projecting into $E_g$ orbital is \begin{equation}\begin{aligned}} \def\eea{\end{aligned}\end{equation} P_{E_g} Y_{3,\pm 2}P_{E_g}=\mp 3\sqrt{\frac{5}{2}} i\sigma_2 \eea where $P_{E_g}$ is the projection operator. The two non-vanishing components of the rank 3 spherical tensor operators can be grouped into a single cubic harmonic tensor $\hat f_{xyz}=\frac{i}{\sqrt 2}(Y_{2,-2}-Y_{2,2})$, which is projected into $E_g$ orbital space as \begin{equation}\begin{aligned}} \def\eea{\end{aligned}\end{equation} P_{e_g}\hat f_{xyz} P_{e_g}=-3\sqrt 5 \sigma_2. \eea where $\hat f_{xyz}$ corresponds to the octupole magnetic moment. Therefore, the complex combinations of the $E_g$ orbital, instead of carrying angular momentum, carry octupole magnetic moment, which was proposed to be the \ `` hidden order " in certain strongly-correlated electronic systems \cite{Santini2000,Brink2001,Kuramoto2009,Jackeli2009,Li2016}. \section{The band structure of the orbital active honeycomb lattice} \label{sec:band_structure} In this section, we study the band structure arising from the orbital degrees of freedom on the planar and the buckled honeycomb lattice. To be concrete, we first introduce a simple nearest neighboring tight-binding model before presenting more general scenarios in the next section. \subsection{Constructing the tight-binding model} \begin{figure} \includegraphics[height=0.32\columnwidth, width=0.32\columnwidth] {pxy.pdf} \includegraphics[height=0.32\columnwidth, width=0.32\columnwidth] {dxy.pdf} \includegraphics[height=0.32\columnwidth, width=0.32 \columnwidth] {eg.pdf} \caption{The orbital configuration of the $\sigma$-bonding on the $\hat{e}_3$ bond for (a) $(p_x, p_y)$ doublet; (b) $(d_{xy}, d_{x^2-y^2})$ doublet; (c) $E_g$ doublets.} \label{fig:bonding} \end{figure} In the orbital active honeycomb lattice, the hopping between neighboring sites occurs between different components of the $E$ irrep and thus is more complicated than the orbital inactive case. There are two kinds of hopping processes allowed by symmetry on a bond. In the convention of chemistry, the one with larger hopping amplitude is called $\sigma$ bonding, and the other one is called $\pi$ bonding. The difference in the hopping amplitude arises from the anisotropy of the orbital wave function. The bonding direction of the $\sigma$ ($\pi$) bond is along the direction of the maximal (minimal) angular distribution of orbital wave functions. The $\sigma$ bond and the $\pi$ bond for all orbital realizations of the $E$ irrep are shown in Fig.~\ref{fig:bonding} for one of nearest neighboring bond $\hat{e}_3$. In the planar case, the $\sigma$ bonding orbital is $p_y$, $d_{yz}$ or $d_{x^2-y^2}$, and in the buckled case, the $\sigma$ bonding orbital is $d_{r^2-3z^2}$. The $\sigma$ bonding orbitals along other nearest neighboring bond are linear combinations of the two orbitals in the $E$ irrep, obtained from applying 3 fold rotation on $p_y$, $d_{yz}$, $d_{x^2-y^2}$ or $d_{r^2-3z^2}$. Since all the different doublets form the same irrep of $C_{3v}$ point group, they transform in the same way under rotation. To unify the notation, we use $\gamma_{x,y}$ to represent the two states in the $E$ irrep for different orbital realizations, where $\gamma_x$ stands for $p_x$, $d_{xy}$ or $d_{x^2-y^2}$, and $\gamma_y$ stands for $p_y$, $d_{x^2-y^2}$ or $d_{r^2-3z^2}$ correspondingly. We define $\gamma_1\sim \gamma_3$ to be the $\sigma$ bonding orbitals along the three nearest neighboring bonds $\hat{e}_1 \sim \hat{e}_3$, \begin{equation}\begin{aligned}} \def\eea{\end{aligned}\end{equation} \gamma_1 = \frac{\sqrt{3}}{2} \gamma_x +\frac{1}{2} \gamma_y, \ \ \gamma_2 =- \frac{\sqrt{3}}{2} \gamma_x +\frac{1}{2}\gamma_y, \ \ \gamma_3 = - \gamma_y. \eea Since the $\sigma$ bonding is much stronger than the $\pi$ bonding, we neglect the $\pi$ bonding and construct the single particle Hamiltonian of the nearest neighboring $\sigma$ bonding. Using $\gamma_1\sim\gamma_3$, the Hamiltonian can be conveniently written as \begin{equation}\begin{aligned}} \def\eea{\end{aligned}\end{equation} H_0=t_\parallel \sum_{\vec{r}\in A,j=1,2,3} \Big\{\gamma_j^\dagger (\vec{r}+\hat e_j) \gamma_j(\vec{r}) +h.c. \Big\}. \label{eq:tight-binding} \eea where the summation over $\vec{r}$ is only on the A sublattice and $\hat{e}_1\sim \hat{e}_3$ are the unit vectors pointing from A site to its three nearest neighboring B sites on the planar honeycomb lattice \begin{equation}\begin{aligned}} \def\eea{\end{aligned}\end{equation} \hat{e}_1 = \frac{\sqrt{3}}{2} \hat{e}_x +\frac{1}{2} \hat{e}_y, \ \ \hat{e}_2 =- \frac{\sqrt{3}}{2} \hat{e}_x +\frac{1}{2}\hat{e}_y, \ \ \hat{e}_3 = - \hat{e}_y. \eea We have set the nearest neighboring distance to 1. In the case of the buckled honeycomb lattice, the three nearest neighboring vectors are the same as in the planar case when the coordinates are projected onto the $(1,1,1)$ plane. The Hamiltonian has the same form for different realizations of the $E$ irrep of the site symmetry $C_{3v}$ for both the planar and buckled honeycomb lattice. This demonstrates the power and elegance of the symmetry principle. \subsection{The spectrum and wavefunctions} The Hamiltonian Eq.~\eqref{eq:tight-binding} is ready to be diagonalized in the momentum space. We define a 4-component spinor $\psi(\vec k)$ in momentum space as \begin{equation}\begin{aligned}} \def\eea{\end{aligned}\end{equation} \psi(\vec k)= (\gamma_{x, A}(\vec k),\gamma_{y,A}(\vec k), \gamma_{x,B}(\vec k), \gamma_{y,B}(\vec k))^T, \eea where $A$ and $B$ refer to the two sublattices. The annihilation operators $\gamma_{x,y}(k)$ is defined as \begin{equation}\begin{aligned}} \def\eea{\end{aligned}\end{equation} \gamma_{x, y}(\vec k)&= \frac{1}{\sqrt{N}}\sum\limits_{\vec{r}} \gamma_{x,y}(\vec{r})e^{-i \vec k\cdot \vec{r}}, \eea The crystal momentum $\vec k$ is defined in the Brillouin zone shown in Fig.~\ref{fig:spectrum}(a). In the case of the buckled honeycomb lattice, $\vec{r}$ is the projected coordinate in the $(1,1,1)$ plane. \begin{figure}[h] \includegraphics[height=0.45\columnwidth, width=0.45\columnwidth] {Brillouin.pdf} \includegraphics[height=0.45\columnwidth, width=0.45\columnwidth] {band.pdf} \caption{(a) The Brillouin zone of the honeycomb lattice. (b) The band structure of the tight binding model is described in Eq.~\eqref{eq:tight-binding} in the strong anisotropy limit. There are four bands (The spin degrees of freedom just add another copy and are omitted). The bottom band and the top band are completely flat, while the middle two bands have the same dispersion relations as in graphene.} \label{fig:spectrum} \end{figure} With this setup, the Hamiltonian takes the following block form: \begin{equation}\begin{aligned}} \def\eea{\end{aligned}\end{equation} H(k)=\begin{pmatrix} 0 &H_{AB}\\ H_{AB}^\dagger & 0 \\ \end{pmatrix}, \label{eq:H_k} \eea where \begin{equation}\begin{aligned}} \def\eea{\end{aligned}\end{equation} H_{AB}= t_\parallel \begin{pmatrix} \frac{3}{4} (e^{i \vec k\cdot \hat e_1}+e^{i\vec k\cdot \hat e_2}) & \frac{\sqrt 3}{4} (e^{i \vec k\cdot \hat e_1}+e^{i \vec k\cdot \hat e_2}) \\[0.9 em] \frac{\sqrt 3}{4} (e^{i \vec k\cdot \hat e_1}+e^{i\vec k\cdot \hat e_2}) & \frac{1}{4}(e^{i \vec k\cdot \hat e_1}+e^{i \vec k\cdot \hat e_2}+2 e^{i \vec k \cdot \hat e_3})\\[0.9em] \end{pmatrix}. \eea There are four bands in total. The middle two bands have the exact dispersion as that in graphene: \begin{equation}\begin{aligned}} \def\eea{\end{aligned}\end{equation} E_{2,3}=\mp \frac{t_\parallel}{2} \left |\sum\limits_ie^{i \vec k \cdot \hat e_i} \right | =\mp \frac{t_\parallel}{2} \sqrt{3+2 \sum \limits_{i=1}^3\cos \vec k \cdot\vec b_i}, \eea where $ \vec b_i =\frac{1}{2} \epsilon_{ijk}(\hat e_j-\hat e_k)$ are the next nearest neighboring vectors. The bands display two Dirac cones at $K$ and $K'$. In addition, Fermi surface nesting and Van Hove singularity occur at $1/4$ filling above and below the Dirac point. The wave functions associated with the middle two bands are \begin{equation}\begin{aligned}} \def\eea{\end{aligned}\end{equation} \ket{\psi(\vec k)}_{2,3}=\frac{1}{\sqrt {N_0}}\left (e^{-i\frac{\theta}{2}}\sum\limits_i \hat e_i e^{i \vec k \cdot \hat e_i },\pm e^{i\frac{\theta}{2}}\sum\limits_i \hat e_i e^{-i\vec k \cdot \hat e_i} \right ), \label{eq:psi_23} \eea where the phase $\theta=\text{arg}(\sum\limits_ie^{i \vec k\cdot \hat e_i })$ and the normalization $N_0=6-2 \sum \limits_{i=1}^3\cos \vec k \cdot \vec b_i$. On the other hand, the top band and the bottom band are perfectly flat with the energy \begin{equation}\begin{aligned}} \def\eea{\end{aligned}\end{equation} E_{1,4}=\mp\frac{3}{2}t_\parallel \eea and connect to the middle two bands at the $\Gamma$ point. The corresponding wave functions are \begin{equation}\begin{aligned}} \def\eea{\end{aligned}\end{equation} \ket{\psi(\vec k)}_{1,4}=\frac{1}{\sqrt{3N_0}}\left (\sum\limits_i \vec b_i e^{-i \vec k \cdot \hat e_i },\pm \sum\limits_i \vec b_i e^{i \vec k \cdot \hat e_i} \right ). \label{eq:psi_14} \eea The entire spectrum is plotted in Fig. \ref{fig:spectrum}. \begin{figure} \includegraphics[height=0.35\columnwidth, width=0.9\columnwidth] {local.pdf} \includegraphics[height=0.9\columnwidth, width=0.9\columnwidth] {wigner.pdf} \caption{(a) The spatial localized states in the lower flat band of the tight-binding Hamiltonian Eq.~\eqref{eq:tight-binding} for (a) $(p_x, p_y)$ doublet; (b) $(d_{xz}, d_{yz})$ doublet; (c) $E_g$ doublet. (d) The spatial localized state Wigner-crystal when the lower flat-band is $1/3$ filled.} \label{fig:flat_band} \end{figure} \subsection{The appearance of the flat-band and the localized state} The existence of the flat bands implies that one can construct a local eigenstate of the single-particle Hamiltonian. The flat band has been studied in detail in~\cite{Wu2007,Wu2008} in the context of the $p$ orbitals in the honeycomb optical lattice, and plaquette states on a hexagon are constructed as the local basis (Fig.~\ref{fig:flat_band}(a)). Here we investigate it in the orbital-active Dirac material realized by the $(d_{xy}, d_{x^2-y^2})$ and the $e_g$ doublets. The localized states can be elegantly constructed from the Bloch wave function in Eq.~\eqref{eq:psi_14}, \begin{equation}\begin{aligned}} \def\eea{\end{aligned}\end{equation} \ket{\psi_{\vec R}}_{\pm}=\frac{1}{\sqrt{N_k}} \sum \limits_k\ket{\psi(\vec k)}_{1,4} e^{-i \vec k \cdot \vec R}, \eea where $\vec R$ are the centers of the hexagons. Each hexagon hosts one localized state from each flat band. The localized states are \begin{equation}\begin{aligned}} \def\eea{\end{aligned}\end{equation}\label{eq:local} \ket{\psi_{\vec R}}_{\pm}&=\\ \sum\limits_{n=0}^5&(\pm1)^{n} \left (\cos\frac{n\pi}{3}\gamma_x^\dagger(\vec{r}_n)+\sin \frac{n\pi}{3}\gamma_y^\dagger(\vec{r}_n)\right )\ket{0}. \eea The summation is over the six vertices of the hexagon as shown in Fig.~\ref{fig:flat_band}(b) for the case of the $(d_{xy}, d_{x^2-y^2})$ doublet. The localized state has the same weight on each site but different orbital configurations, related by $\frac{\pi}{3}$ rotations. On each site, the orbital is projected into the $\pi$-bonding along the outward bond away from the hexagon. Due to the destructive interference, electrons in such localized single-particle states cannot leak out the plaquette, rendering the localized states eigenstates of the tight-binding Hamiltonian in Eq.~\eqref{eq:tight-binding} with the same energy. The above analysis can be carried over to the $E_g$ doublets on the buckled honeycomb lattice. In this case, the localized state is confined in a buckled hexagon as shown in Fig.~\ref{fig:flat_band}$(c)$. \subsection{Orbital configurations at high symmetric points} As shown in Fig.~\ref{fig:spectrum}(b), the spectrum exhibits double degeneracy at the $K(K')$ point and the $\Gamma$ point in the Brillouin zone. Now we investigate the Bloch wave function at these high symmetric points in detail. \subsubsection{$\Gamma$ point} Around the center of the Brillouin zone $\vec k=(0,0)$, the Hamiltonian Eq.~\eqref{eq:H_k}, in the unit of $t_\parallel$, can be expanded as \begin{equation}\begin{aligned}} \def\eea{\end{aligned}\end{equation} H_\Gamma (k)=&\frac{3}{2} (1-\frac{1}{4} |k|^2) \tau_1 \otimes \sigma_0 - \frac{3}{4} \left ( k_x \tau_2 \otimes \sigma_1 +k_y \tau_2 \otimes \sigma_3 \right ) \\ &-\frac{3}{16} \left ( (k_x^2-k_y^2)\tau_1 \otimes \sigma_3+ 2 k_x k_y \tau_1 \otimes \sigma_1 \right ), \label{H_Gamma} \eea where the Pauli matrices $\sigma_0 \sim \sigma_3$ ($\sigma_0$ represents the identity matrix) describe the orbital degrees of freedom $\gamma_{x,y}$ in the $E$ irrep, and $\tau_0 \sim \tau_3$ act on the space of sublattice $(A, B)$. To the leading order, the dispersion of the bands is \begin{equation}\begin{aligned}} \def\eea{\end{aligned}\end{equation} &E^\Gamma_{1,4}=\mp \frac{3}{2} t_\parallel \\ &E^\Gamma_{2,3}=\pm\frac{3}{2}t_\parallel \left ( -1+\frac{1}{4}|\Delta k|^2 \right ). \eea Therefore, the bands touch each other quadratically at both upper and lower degeneracy points. The degenerate wave functions at each touching point can be regrouped so that they only contain one of each the orbital component. At the lower touching point, the wave functions are \begin{equation}\begin{aligned}} \def\eea{\end{aligned}\end{equation} \ket{\psi(\Gamma)}^+_{x(y)}=\frac{1}{\sqrt{2}}\left ( \gamma_{x(y),A}^\dagger + \gamma_{x(y),B}^\dagger \right )\ket{0}. \label{eq:psi_Gamma_1} \eea At the upper touching point, the $B$ sublattice component acquires a minus sign, and the wave functions are \begin{equation}\begin{aligned}} \def\eea{\end{aligned}\end{equation} \ket{\psi(\Gamma)}^-_{x(y)}=\frac{1}{\sqrt{2}}\left ( \gamma_{x(y),A}^\dagger - \gamma_{x(y),B}^\dagger \right )\ket{0}. \label{eq:psi_Gamma_2} \eea \subsubsection{$K$ and $K'$ points} Around $ \vec K=(\frac{4\pi}{3\sqrt{3}},0)$, the Hamiltonian in Eq.~\eqref{eq:H_k} can be expanded as \begin{equation}\begin{aligned}} \def\eea{\end{aligned}\end{equation} H_K =&-\frac{4}{3}\Delta k_x \tau_1 \otimes \sigma_0 + \frac{4}{3}\Delta k_y \tau_2 \otimes \sigma_0 \\ &-\frac{3}{8}(2+\Delta k_x) \tau_1 \otimes \sigma_0 - \frac{3}{8}\Delta k_y \tau_2 \otimes \sigma_3 \\ &-\frac{3}{8} \Delta k_y \tau_1 \otimes \sigma_1- \frac{3}{8} (2-\Delta k_x)\tau_2 \otimes \sigma_1 \label{eq:H_K}, \eea where $\Delta \vec k =\vec k- \vec K$. The middle two bands touch each other with the dispersion, \begin{equation}\begin{aligned}} \def\eea{\end{aligned}\end{equation} E_{2,3}= \mp \frac{3}{4}t_\parallel |\Delta k|, \eea which demonstrates the Dirac cone. The doubly degenerate wave functions can be combined so that each of them only occupies one of the sublattices: \begin{equation}\begin{aligned}} \def\eea{\end{aligned}\end{equation} \ket{\psi ( \vec K)}_A=\frac{1}{\sqrt 2}\left ( \gamma^\dagger _{x,A}+i \gamma^\dagger_{y,A} \right ) \ket{0} \\ \ket{\psi (\vec K)}_B=\frac{1}{\sqrt 2}\left ( \gamma^\dagger _{x,B}-i \gamma^\dagger_{y,B} \right ) \ket{0}. \label{eq:psi_K} \eea The orbital states on the two sublattices are circular polarized and exhibit opposite chiralities. Such complex combinations of the orbitals exhibit distinct physical properties for different orbital realizations. In the case of the $(p_x,p_y)$ doublet as well as the $(d_{xz}, d_{yz})$ doublet, the circular polarized state $\ket{\gamma_1}\pm i \ket{\gamma_2}$ carries angular momentum $L_z=\pm 1$; in the case of the $(d_{xy}, d_{x^2-y^2})$ doublet, the circular polarized state carries angular momentum $\mp 2$; in the case of the $E_g$ doublet, it carries magnetic octupole moment. These complex orbital states play an important role in the topological properties of the orbital-active Dirac material and will be addressed in Sec.~\ref{sec:gap_opening}. \subsection{Response of the flat band to magnetic field} One interesting question regarding the flat band that appeared is its response to an external magnetic field. In a flat band, because the kinetic energy of the electrons is completely quenched, the usual semi-classical picture is no longer valid. Recent work~\cite{rhim2020flat} demonstrates that the response of flat bands to an external magnetic field is closely related to the quantum distance of the flat band. The quantum distance between two Bloch wavefunctions is defined as, \begin{equation}\begin{aligned}} \def\eea{\end{aligned}\end{equation} d=1-|\bra{\psi(k)}\psi(k')\rangle|^2, \eea which ranges from 0 to 1. The flat band is singular if $d$ is nonzero in the limit that $|k-k'|\rightarrow 0$. A singular point $k_0$ can be characterized by the maximal quantum distance $d_{max}$ between the wavefunctions of $k$ and $k'$ that are sufficiently close to $k_0$. In systems without orbital degrees of freedom, it is found that when $d_{max}$ is nonzero, the flat band splits into Landau levels in an energy window, and the width of the energy window is determined by $d_{max}$. In our case, the flat band touches the dispersive Dirac band at the $\Gamma$ point. The wavefunction of the flat band, expanded around the $\Gamma$ point, is \begin{equation}\begin{aligned}} \def\eea{\end{aligned}\end{equation} \ket{\psi}=\frac{1}{\sqrt{2}}\left (\sin \theta(\vec k), - \cos \theta(\vec k), \sin \theta(\vec k), -\cos\theta(\vec k) \right) \label{eq:singular} \eea where $\theta(\vec k)$ is the azimuth angle of $\vec{k}$. Therefore, the wavefunction at $\theta$ and $\theta-\pi/2$ are orthogonal to each other, rendering the maximal quantum distance $d=1$. As a result, the flat band that appeared here is singular by definition. We study the response of the flat band to an external magnetic field by including the gauge field in the hopping parameter $t\rightarrow t\exp(i\int A(\vec r)d\vec r)$. For simplicity, we choose the Landau gauge $A=B(0,x)$ and the strength of the magnetic field B is set by the flux through each hexagon $2\pi p/q$. In the presence of the magnetic field, the original four bands split into $4q$ sub-bands. The band structure is shown in Fig.~\ref{fig:magnetic_field}. While the middle two dispersive Dirac bands form the characteristic Landau levels, surprisingly, the two singular flat bands are completely inert to the magnetic field, in contrast with previous results on singular flat bands without orbital degrees of freedom. This is due to the orbital nature of the singularity in Eq.~\eqref{eq:singular}. In the presence of the magnetic field, one can still construct localized states from the Bloch wavefunction of the flat band. Instead of occupying a hexagon, the localized states now occupy a magnetic unit cell. The wavefunction is nonzero only on the boundary of the magnetic unit cell, and the orbital configuration is parallel to the tangential direction. An example of the localized states is shown in Fig.~\ref{fig:magnetic_field}(b) for the flux given by $2\pi/3$. \begin{figure} \includegraphics[height=0.49\columnwidth, width=0.49\columnwidth] {orbital_response.pdf} \includegraphics[height=0.49\columnwidth, width=0.49\columnwidth] {local_magnetic.pdf} \caption{(a) The flat band in the orbital active Dirac system is singular but does not respond to an external magnetic field due to the orbital nature of the singularity. (b) An example of the localized state in the presence of the magnetic field. Here the flux through each hexagon is $2\pi/3$. The localized state occupies three hexagons.} \label{fig:magnetic_field} \end{figure} \section{General symmetry consideration beyond strong anisotropic limit \label{sec:general_symmetry} In the last section, we demonstrate many remarkable features resulting from the tight-binding Hamiltonian Eq.~\eqref{eq:tight-binding}, including orbital enriched Dirac cone, quadratic band touching and flat bands. Since the Hamiltonian only includes nearest neighboring hopping and $\sigma$ bonding, a natural question is whether these features rely on the specific form of the Hamiltonian or are generic results from the symmetry. In this section, we address this issue by general symmetry consideration. We study the band structure of the orbital active Dirac materials around the high symmetric points in the Brillouin zone using $k\cdot p$ theory. In summary, the band flatness is not generic and can be bent by the $\pi$-bondings, for example. However, the orbital configurations at the high symmetric points and $k$ dependence of the dispersion around the $\Gamma$ and $K$ points are preserved as long as the symmetry of the system is respected. In the following, we consider the effects of the point group symmetry of the buckled honeycomb lattice $D_{3d}$. \subsection{$\Gamma$ point} At the $\Gamma$ point, the little group is the maximal point group of the lattice, $D_{3d}$. The Bloch wave function is composed of the orbital part $\ket{\gamma_{x,y}}$ and the plane wave part $\ket{\Gamma}_{A, B}$, which can be classified into the irreps of the little group separately. The orbital degrees of freedom $(\ket{\gamma_x}, \ket{\gamma_y})$ form the two dimensional $E_g$ irrep, while the planewave part $\ket{\Gamma}_A + \ket{\Gamma}_B$ and $\ket{\Gamma}_A - \ket{\Gamma}_B$ form the one dimensional $A_{1g}$ and $A_{2u}$ irreps. Therefore the composite Bloch wave function can be grouped into two two-dimensional irreps, $E_g$ and $E_u$, respectively. This indicates that the four energy levels can be grouped into two doubly-degenerate sets, where the degeneracy completely originates from the orbitals. The Bloch wave functions of the $E_g$ irrep are \begin{equation}\begin{aligned}} \def\eea{\end{aligned}\end{equation} \ket{\gamma_{x(y)}}\otimes (\ket{\Gamma}_A+\ket{\Gamma}_B), \eea and wave functions of the $E_u$ irrep are \begin{equation}\begin{aligned}} \def\eea{\end{aligned}\end{equation} \ket{\gamma_{x(y)}}\otimes (\ket{\Gamma}_A-\ket{\Gamma}_B), \eea This is consistent with Eq.~\eqref{eq:psi_Gamma_1} and Eq.~\eqref{eq:psi_Gamma_2}. Furthermore, in both degeneracy sets, the orbital part of the wave functions can be regrouped into the complex orbital states $\ket{\gamma_{\pm}}=\ket{\gamma_1} \pm i\ket{\gamma_2}$. Therefore, spin-orbit coupling is able to gap out the degeneracy. The generic dispersion around the $\Gamma$ point can be obtained from the $k \cdot p$ theory. The $k \cdot p$ Hamiltonian is invariant under the $D_{3d}$ point group symmetry. In order to write down all the symmetry allowed terms, it is convenient to first classify the operators in the orbital space $\vec \sigma$, the operators in the sublattice space $\vec \tau$ and the momentum $\vec k$ into irreps of the point group $D_{3d}$. In the sublattice space, the classification is \begin{equation}\begin{aligned}} \def\eea{\end{aligned}\end{equation} \begin{cases} \tau_0, \tau_1 & A_{1_g} \\ \tau_2, \tau_3 & A_{2_u}. \end{cases} \eea In orbital space, the classification is \begin{equation}\begin{aligned}} \def\eea{\end{aligned}\end{equation} \begin{cases} \sigma_0 , & A_{1g} \\ \sigma_2, & A_{2g} \\ (\sigma_1, \sigma_3), & E_g. \end{cases} \eea For the momentum, the classification is \begin{equation}\begin{aligned}} \def\eea{\end{aligned}\end{equation} \begin{cases} (k_x, k_y) & E_{u}\\ k_x^2+k_y^2 & A_{1g} \\ (2k_x k_y, k_x^2-k_y^2), & E_{g}. \end{cases} \eea Based on these classifications and the product table of the $D_{3d}$ point group, the most general Hamiltonian takes the following form, \begin{widetext} \begin{equation}\begin{aligned}} \def\eea{\end{aligned}\end{equation} H_\Gamma (\vec k)=&(h_1+h_2 ( k_x^2+ k_y^2))\tau_1 \otimes \sigma_0 + k_x (h_3 \tau_2 \otimes \sigma_1 + h_4 \tau_3 \otimes \sigma_1 )+ k_y (h_3 \tau_2 \otimes \sigma_3 + h_4 \tau_3 \otimes \sigma_3 )+\\ &(k_x^2- k_y^2)(h_5 \tau_0 \otimes \sigma_3 + h_6 \tau_1 \otimes \sigma_3 )+ 2 k_x k_y (h_5 \tau_0 \otimes \sigma_1 +h_6 \tau_1 \otimes \sigma_1 ), \label{H_Gamma_general} \eea \end{widetext} where $h_1\sim h_6$ are constants with the unit of energy. We will see later that at the first order of $\vec k$, the degeneracy is still preserved, so we have to include second-order terms of $\vec k$. It recovers the tight-binding Hamiltonian around the $\Gamma$ point presented in Eq.~\eqref{H_Gamma}, when the $h$s are set to \begin{equation}\begin{aligned}} \def\eea{\end{aligned}\end{equation} &h_1=-4 h_2=-2h_3 = -8 h_6=\frac{3}{2}t_\parallel \\ &h_4=h_5=0, \eea In the leading order, the dispersion is \begin{equation}\begin{aligned}} \def\eea{\end{aligned}\end{equation} &E_{1,2}&=-h_1-\left (2h_2+\frac{h_3^2+h_4^2}{h_1}\right )| k|^2\pm (h_5-h_6)| k|^2\\ &E_{3,4}&=h_1+\left (2h_2+\frac{h_3^2+h_4^2}{h_1}\right )|k|^2\pm (h_5+h_6)| k|^2. \eea At finite $k$, the degeneracy is lifted by \begin{equation}\begin{aligned}} \def\eea{\end{aligned}\end{equation} &|E^\Gamma_1 - E^\Gamma_2 |=\frac{m_-}{2}|k|^2 \\ &|E^\Gamma_3 - E^\Gamma_4 |=\frac{m_+}{2} |k|^2 . \\ \eea where the effect mass $m_\pm=4 |h_5 \pm h_6| $. Therefore, the bands touch quadratically at both degenerate points. It is known that quadratic bound touching is unstable to interaction and can lead to exotic phases such quantum Hall effect and nematicity~\cite{Sun2011}. \subsection{$K(K')$ point} At the $K$ point, the little group is $D_3$, containing the three-fold rotations perpendicular to the plane and a two-fold rotation through the plane that interchanges the two sublattices. The Bloch wave functions $\ket{\psi(K)}$, containing both the planewave part and the orbital part, can be organized into irreps of the little group. The planewave part contains two sublattice components, forming the $E$-irrep, with the $A/B$ sublattice component carrying chirality $\pm1$. The on-site orbital degrees of freedom $\gamma_x$ and $\gamma_y$ also transform as the $E$ irrep, the complex combination $\gamma_x\pm i \gamma_y$ carrying the chirality $\pm 1$. Therefore, the four composite wave functions can be decomposed into three irreps as $2\otimes2=1\oplus1\oplus 2$. There are two trivial $A_1$ representations, where the chiralities of the orbital and planewave cancel each other and an $E$ irrep where the chiralities of the orbital and planewave add up. In general, the two $A_1$ states do not have the same energy. In contrast, the two states in the $E$ irrep are degenerate from symmetry and carry opposite chirality at the Dirac point. Explicitly, the two states are: \begin{equation}\begin{aligned}} \def\eea{\end{aligned}\end{equation} &\left(\ket{\gamma_x}+ i\ket{\gamma_y}\right )\otimes \ket{K}_A\\ &\left(\ket{\gamma_x}- i\ket{\gamma_y}\right)\otimes \ket{K}_B. \label{eq:wave_k} \eea This is consistent with the wave functions of the tight-binding model at $K$ in Eq.~\eqref{eq:psi_K}. To obtain the generic dispersion around $K(K')$, we again employ the $k\cdot p$ theory. The Hamiltonian around $K$, a combination of the plane wave, orbital, and sublattice has to be invariant under the $C_{3v}$ point group. Following the same strategy, we first organize $\vec{\sigma}$, $\vec{\tau}$ and the momentum $\Delta \vec k =\vec k-\vec K$ into irreps of the little group $d_3$. In the sublattice space, we have, \begin{equation}\begin{aligned}} \def\eea{\end{aligned}\end{equation} \begin{cases} \tau_0, &A_1 \\ \tau_3, &A_2 \\ (\tau_1, -\tau_2), &E. \end{cases} \eea In orbital space, we have, \begin{equation}\begin{aligned}} \def\eea{\end{aligned}\end{equation} \begin{cases} \sigma_0 , &A_1 \\ \sigma_2, &A_2 \\ (\sigma_3, -\sigma_1), &E. \end{cases} \eea In addition, the momentum $(\Delta k_x, \Delta k_y)$ belongs to the $E$ irrep as well. Therefore, based on the product table of $D_3$ point group, the most general Hamiltonian, apart from an overall constant, reads, \begin{widetext} \begin{equation}\begin{aligned}} \def\eea{\end{aligned}\end{equation} H_K(\Delta \vec k)=&h_1 \tau_3 \otimes \sigma_2 + h_2 ( \tau_1 \otimes \sigma_3 + \tau_2 \otimes \sigma_1)+\\ &\Delta k_x \left \{+ h_3 \tau_0 \otimes \sigma_3+h_4 \tau_1 \otimes \sigma_0 + h_5 \tau_2 \otimes \sigma_2 + h_6 \tau_3 \otimes \sigma_1 +h_7 ( \tau_1 \otimes \sigma_3 - \tau_2 \otimes \sigma_1)\right \}+\\ &\Delta k_y \left \{ -h_3\tau_0 \otimes \sigma_1 - h_4 \tau_2 \otimes \sigma_0 + h_5 \tau_1 \otimes \sigma_2 + h_6 \tau_3 \otimes \sigma_3 + h_7 ( \tau_2 \otimes \sigma_3 + \tau_1 \otimes \sigma_1)\right \}. \label{eq: H_K_general} \eea \end{widetext} The expansion of the $\sigma$-bonding Hamiltonian at $K$ in Eq.~\eqref{eq:H_K} is a special case with \begin{equation}\begin{aligned}} \def\eea{\end{aligned}\end{equation} h_2=h_4=2 h_7=-\frac{3}{4} t_\parallel \\ h_1=h_3=h_5=h_6=0. \eea In the general situation, at the leading order of $\Delta \vec k$, the dispersions of the four bands read, \begin{equation}\begin{aligned}} \def\eea{\end{aligned}\end{equation} E^K_{1,4}&=-h_1\pm 2 h_2+ \mathcal{O}(|\Delta k|^2) \\ E^K_{2,3}&=h1 \pm 2 h_7 |\Delta k| +\mathcal{O}(|\Delta k|^2), \eea which is consistent with those given by the nearest-neighboring tight-binding model. The dispersion $E^K_{2,3}$ is Dirac-like as long as $h_7 \neq 0$. The situation of the $K'$ point can be obtained by performing the reflection symmetry with respect to the $y$ axis. The above analysis solely relies on the non-abelian nature of the point group and therefore is widely applicable to the orbital-active Dirac material, independent of the origin of the orbitals. \section{Gap opening mechanism} \label{sec:gap_opening} We have shown that the symmetry of the honeycomb lattice protects the degeneracy of the band structure at $K(K')$ point and $\Gamma$ point. The degenerate states form the 2-dimensional irrep of the little group at the high symmetry points. The degeneracy can be lifted by including various symmetry-breaking terms in the Hamiltonian, which introduces gaps at the Dirac point or/and the quadratic band touching point. The interplay of different symmetry-breaking terms can give rise to various topological band structures, rendering the orbital active Dirac system a flexible platform for realizing topological insulators with different edge-state properties. In this section, we discuss the gap opening mechanisms for different orbital doublets of the $E$ irrep, previously studied in different contexts~\cite{Zhang2014,Xiao2011}, in a unified manner. Based on Eq.~\eqref{eq:psi_Gamma_1} and Eq.~\eqref{eq:psi_K}, the degenerate wave functions at $\Gamma$ and $K$ can be grouped into circular polarized orbital state $\gamma_x \pm i\gamma_y$ with opposite chirality. As a result, a $\sigma_2$ term in the orbital space, which measures the chirality, is able to lift the degeneracy at both $Gamma$ and $K$ points. In addition, at $K(K')$ points, the two complex orbital states only occupy A and B sublattices, respectively. As a result, a $\tau_3$ term in the sublattice space can also gap out the Dirac points. In contrast, since the degenerate states at the $\Gamma$ point have the same weight on both sublattices, they remain degenerate after the $\tau_3$ term is added to the Hamiltonian. One can also add other terms to the $k\cdot p$ Hamiltonian to open up a gap in the spectrum. But the two terms mentioned above, $\tau_3$ and $\sigma_2$, denoted as $H_m$ and $H_\lambda$ respectively, are among the simplest and have a clear physical origin. In the real space, they have the following form, \begin{equation}\begin{aligned}} \def\eea{\end{aligned}\end{equation} H_m&=m\left\{\sum\limits_{\vec r\in A ,\sigma} \gamma^\dagger_{\sigma}(\vec r) \gamma_{\sigma}(\vec r)-\sum\limits_{\vec r\in B ,\tau} \gamma^\dagger_{\sigma}(\vec r) \gamma_{\sigma}(\vec r)\right\}\\ H_\lambda&=\lambda\left\{ \sum\limits_ {\vec r \in A, B} i \gamma_x^\dagger (\vec r) \gamma_y (\vec r ) + h.c. \right \}. \eea The term $H_m$ represents the staggering mass resulting from the imbalance between the $A$ and $B$ sublattices, which for example, occurs in TMD materials. It only depends on the particle number on each sublattice and does not rely on the particular orbital state the electrons occupy. On the other hand, $H_\lambda$ measures the chirality of the orbital and originates from spin-orbit coupling $-\lambda_0\vec{L}\cdot \vec{S}$, where $\lambda_0$ is the atomic spin-orbit coupling strength, $\vec{L}$ is the physical angular momentum operator and $\vec{S}$ is the spin operator of electrons. In free space, $\vec{L}$ acts on the Hilbert space labeled by the angular momentum $S$, $p$, $d$, etc. In the case of the planar and buckled honeycomb lattices, the spherical symmetry reduces to the site symmetry $C_{3v}$. As the result, the physical angular momentum $\vec{L}$ should be projected into the 2d irrep. The result depends on the particular orbital realizations of the irrep, even though they are equivalent under the $C_{3v}$ point group. In the following, we discuss the different orbital realizations case by case. In the case of the $(p_x, p_y)$ and $(d_{xz}, d_{yz})$ doublets, the circular polarized orbital state have angular momentum $\pm 1$. The $L_z$ operator, projecting into the two-dimensional space, becomes $\sigma_2$, while $L_x$ and $L_y$ are zero. The spin orbit coupling $-\lambda \vec{L}\cdot \vec{s}$ becomes $-\lambda \sigma_y s_z$. As a result, after the spin-orbit coupling is included in the Hamiltonian, which lifts the degeneracy at the $\Gamma$ point and the $K$ points, the complex orbital state with positive chirality, $p_x + i p_y$ or $d_{xz} + i d_{yz}$, has higher energy than its partner, as shown in Fig.~\ref{fig:gap}(a). In the case of the $(d_{xy}, d_{x^2-y^2})$ doublet, the complex orbital states carry angular momentum $\mp 2$. The $L_z$ operator in this space is $-2\sigma_y$ while the other components vanish. Therefore, the spin-orbit coupling term is $2\lambda_0 \sigma_y s_z$. Note the factor of 2 and extra minus sign compared with the previous two cases. Therefore, the complex orbital state $d_{xy} + i d_{x^2-y^2}$ has lower energy than its partner at the $\Gamma$ point and the $K$ points when the degeneracy is lifted, as shown in Fig.~\ref{fig:gap}(b). The $E_g$ doublet is special. As discussed in Sec.~\ref{sec:octupole}, the complex orbital combinations do not carry angular momentum. The angular momentum operator $L$, projecting into this space, vanishes for all components. The $\sigma_y$ term measure the octupole momentum $\hat{f}_{xyz}$ instead, which is the lowest rank of non-vanishing multipole order for $E_g$ orbitals. On the level of single-particle physics, the $\sigma_y$ term cannot be obtained directly from the spin-orbit coupling in the $E_g$ space. It can appear as a result of second-order perturbation, taking into account the virtual excitation from the $E_g$ orbitals to the $t_{2g}$ orbitals. As discussed in Eq.~\eqref{eq:111_irrep}, the $T_{2g}$ orbitals splits into one 1d irrep $A_1$ and one 2d irrep $E$ of the site symmetry group $C_{3v}$, which in general have distinct onsite energies. We denote the energy difference from the two irreps derived from the $T_{2g}$ orbitals to the $E_g$ orbitals as $\Delta_1$ and $\Delta_2$, respectively. The second order spin-orbit coupling reads, \begin{equation}\begin{aligned}} \def\eea{\end{aligned}\end{equation} &H_{\lambda}'\\ &=-\lambda_0^2P_{E_g}\left\{\frac{\vec{L}\otimes\vec S P_{A_{1g}}\vec{L}\otimes\vec S}{\Delta_1}+ \frac{\vec{L}\otimes\vec S P_{E}\vec{L}\otimes\vec S}{\Delta_2}\right\} P_{E_g}\\ &=-(\frac{\lambda_0^2}{2\Delta_1}+\frac{\lambda_0^2}{\Delta_2})\sigma_0-\frac{\lambda_0^2}{\Delta_1\Delta_2}(\Delta_1-\Delta_2) \sigma_2\otimes S_{(1,1,1)}, \label{eq:hlambda} \eea where $P_{E_g}$, $P_E$ and $P_{A_1}$ are the projection operators, and $S_{(1,1,1)}$ is the spin operator along the (1,1,1) direction. The first term is proportional to the identity operator and thus can be absorbed into the chemical potential. The second represents the effective spin-orbit coupling in the $E_g$ doublets with the spin-orbit coupling strength, \begin{equation}\begin{aligned}} \def\eea{\end{aligned}\end{equation} \lambda = \frac{\lambda_0^2}{\Delta_1\Delta_2}(\Delta_1-\Delta_2). \eea Recall that $\Delta_1-\Delta_2$ is the energy difference between the $A_1$ and $E$ irreps derived from the $T_{2g}$ orbital in Eq.~\eqref{eq:111_irrep}. Therefore, the energy splitting of the $T_{2g}$ triplet under $C_{3v}$ site symmetry is essential for nonzero $\lambda$. \begin{figure*} \includegraphics[height=0.32\textwidth, width=0.32\textwidth] {spin_orbit_1.pdf} \includegraphics[height=0.32\textwidth, width=0.32\textwidth] {spin_orbit_2.pdf} \includegraphics[height=0.32\textwidth, width=0.32\textwidth] {stagger_mass.pdf} \caption{(a) The gap opening pattern from the $H_\lambda$ term for the $(p_x, p_y)$ and $(d_{xz},d_{yz})$ doublets. (b) The gap opening pattern from the $H_\lambda$ term for the $(d_{xy}, d_{x^2-y^2})$ doublet. (c) The gap opening pattern from the stagger mass term for all cases. The complex orbitals state naturally occurs at the $K (K')$ point. } \label{fig:gap} \end{figure*} The gaps introduced by the $\sigma_y$ term in the orbital space are topological. It is straightforward to show that the four bands in Fig.~\ref{fig:gap}(a) and (b) acquire Chern numbers 1, 0, 0, -1 from the bottom to the top. As a result, edge states appear on the boundary of the material. We consider the Hamiltonian on a ribbon with finite width but infinite length, in which case $k_x$ remains a good quantum number. The spectrum is plotted in Fig.~\ref{fig:edge}(a) as a function $k_x$, showing the edge states between the four bulk bands. The orbital wave functions of the edge states are in general complex. The expectation value of the $\sigma_y$ operator in the orbital space is indicated by the color bar. When the orbital degree of freedom is the $E_g$ doublet, the edge states carry the magnetic octupole moment instead of the dipole moment, sketched in Fig.~\ref{fig:edge}(b). \begin{figure} \includegraphics[height=0.49\columnwidth, width=0.49\columnwidth] {edge_state.pdf} \includegraphics[height=0.49\columnwidth, width=0.49\columnwidth] {edge_state2.pdf} \caption{(a) The spectrum of the Hamiltonian including $H_\lambda$ term on a ribbon geometry. $\lambda$ is set to $0.2t_\parallel$. Edge states appear between the bulk bands. The orbital wave function of the edges is complex, the chirality indicated by the color bar. (b) The ribbon geometry. When the orbital realization is from the $E_g$ doublet, the edge states carry magnetic octupole momentum. } \label{fig:edge} \end{figure} The topological gaps are proportional to the coefficient $\lambda$ of the $\sigma_2$ term in the Hamiltonian. Remarkably, when the orbital degrees freedom are realized by the $(p_x, p_y)$, $(d_{xz}, d_{yz})$ or $(d_{xy}, d_{x^2-y^2})$, $\lambda$ is directly related to the atomic spin-orbit coupling strength $\lambda_0$, which could be quite large for heavy atoms. This leads to a robust topological phase, such quantum spin hall effect, at high temperatures. In contrast, in $E_g$ systems, the $\sigma_2$ term comes from the second-order perturbation of the spin-orbit coupling. Therefore, in the $E_g$ Dirac materials, the band degeneracy is much more stable than other realizations. Even though the quadratic band touching can be gaped out from dynamic spin-orbit coupling generated by interaction, the Dirac points are stable against interaction, making the $E_g$ Dirac materials an ideal 2D Dirac semi-metal. We also present the gap opening pattern from the staggering mass term $H_m$ in Fig.~\ref{fig:gap}(c), which is the same for different orbital realizations. $H_m$ term leads to a trivial band insulator by itself However, including $H_m$ in the presence of $H_\lambda$ can lead to richer topological phases with various Chern bands and edge state configurations~\cite{Zhang2014}. \section{The interaction effects} \label{sec:interaction} The interplay between the band structure and interaction can lead to many interesting phases of matter as a function of the filling fraction. In the following, we mention several most interesting ones. At filling $1/2$, the Fermi surface is at the Dirac points, which are stable against interaction. Therefore, the system remains a Dirac semi-metal for a while before it enters a Mott insulator phase. In the Mott insulator, each site is occupied by two electrons. Because of the strong intra-orbital repulsion and Hund's coupling, the two electrons prefer to stay in two different orbitals and form a triplet. As a result, the low-energy effective theory of the Mott insulator is described by a spin 1 Heisenberg model on the honeycomb lattice where the orbital degrees of freedom are inert. At filling $1/4$, the bottom two spinful bands are filled. The Fermi surface is right at the quadratic band-touching point, which is unstable against interaction~\cite{Sun2009}. When the interaction is weak, it dynamically generates the spin-orbital coupling term in both $p_x/p_y$ and $E_g$ orbital-active Dirac material, which gives rise to the quantum spin Hall effect. As the interaction strength grows, the quantum spin Hall phase gives way to a nematic phase that breaks the 3-fold rotation symmetry. At filling $1/8$, the Fermi surface is within the bottom two spinful bands. In the strong anisotropy limit, the two bands become flat, which enhances the interaction effect. Due to the Coulomb interaction, the system favors a flat-band ferromagnetic state. Therefore, effectively, one of the spinful flat bands is filled, and the Fermi surface is at the quadratic band touching point again. The resulting weak interaction phase exhibits the anomalous quantum Hall effect. As the filling becomes even lower due to the band flatness, the systems start to Wigner-crystallize. In particular, when the flat band is $1/3$ filled, the localized states close-pack the lattice. If only on-site interactions are considered, such a close-packing many-body state is the \textit{exact} many-body ground state with massive spin degeneracy from each localized single-particle state. This close-packing state breaks the original lattice translation symmetry with an enlarged $\sqrt{3}\times \sqrt{3}$ unit cell, sketched in Fig.~\ref{fig:flat_band}(b). When long-range interactions are considered, the Wigner crystal appears at even lower fillings. When the spin-orbit coupling is included, the flat band becomes nearly flat and acquires Chern number $\pm 1$. In this case, the Chern fractional insulator~\cite{Sun2011} also becomes a ground state candidate and competes with the Wigner crystal phase. \section{Discussion and summary} \label{sec:summary} We have studied the orbital-active Dirac materials in a unified manner. The various orbital realizations can be understood as the irreps of the point group symmetry $C_{3v}$. Both belonging to the two-dimensional $E$ irrep of $C_{3v}$, the $p_x/p_y$ doublet and the $E_g$ doublet can map to each other, and the Dirac materials based on these two sets of different doublet has the same property. Based on the symmetry, we show the band structure features Dirac cones at $K(K')$ points and a quadratic band touching $\Gamma$. The interplay between the interesting band structure and interaction leads to a rich phase diagram of the orbital-active Dirac materials. The Dirac material is not only limited to the electronic systems but also applies to systems of phonons and polaritons~\cite{Jacqmin2014, Milicevic2017, Zhang2015, Roman2015, Stenull2016, Zhu2018}, where their polarization modes realize the orbital degrees of freedom. The symmetry argument in Sec.~\ref{sec:general_symmetry} also enforces chiral valley phonons in materials with a honeycomb structure, such as Boron Nitride and TMD. Thus the interplay between the chirality of electrons wave function and the chirality of the phonons opens a new door for valleytronics. \section{Acknowledgment} C. W. is supported by the National Natural Science Foundation of China through Grant No. 11729402, No. 11729402 and No. 122234016.

439af6c38bb549275fa2f9d7bfdb2f06e7aa44cb

\section{Introduction} The linear $M_{\rm bh}$--$M_{\rm *,sph}$ relation \citep{1988ApJ...324..701D, 1989IAUS..134..217D}, see also \citet{1992ASPC...31..417Y}, sometimes referred to as the Magorrian relation \citep{1998AJ....115.2285M}, has repeatably been heralded as a critical ingredient to understanding the coevolution of galaxies and their central massive black holes. Black hole feedback is said to regulate the gas and thereby control the star formation \citep[e.g.,][and references therein]{1998A&A...331L...1S, 2007MNRAS.380..877S, 2008ApJ...676...33D, 2010MNRAS.402.1536S, 2013MNRAS.432.3401G} and thus establish/explain the $M_{\rm bh}$--$M_{\rm *,sph}$ relation.\footnote{\citet{2005SSRv..116..523F}, \citet{2006ccha.book.....L} and \cite{2016ASSL..418..263G} review the discovery of black holes and their scaling relations.} Despite early evidence for a non-linear $M_{\rm bh}$--$M_{\rm *,sph}$ relation \citep[e.g.,][]{1998ApJ...505L..83L, 2001ApJ...553..677L, 1999ApJ...519L..39W, 2000MNRAS.317..488S}, there has been a tendency to cling to the simplicity of the original trend. However, along with increases in sample size and improvements in galaxy decomposition --- which have led to both a better understanding of galaxies and a better measurement of their spheroidal component\footnote{We use the term `spheroid' to denote bulges and (pure) elliptical galaxies.} ---, has come an ever-refined insight into the $M_{\rm bh}$--$M_{\rm *,sph}$ diagram through the detection of (galaxy morphology)-dependent substructure and departures from the near-linear relation. Clues that something was amiss with the notion of a near-linear $M_{\rm bh}$--$M_{\rm *,sph}$ relation were presented in \citet{2012ApJ...746..113G}, which reported a steeper near-quadratic relation for spheroids with a \citet{1963BAAA....6...41S} light profile\footnote{A review of the \citet{1963BAAA....6...41S} model can be found in \citet{2005PASA...22..118G}.} and a near-linear relation for spheroids with a core-S\'ersic light profile\footnote{\cite{1966ApJ...143.1002K} and \citet{1972IAUS...44...87K} noted that such galaxies have shallow inner light profiles notably flatter than expected from their outer $R^{1/4}$-like light profile.} \citep{2003AJ....125.2951G}. This work built on a key tip-off in the final paragraph of \citet{2007MNRAS.379..711G}\footnote{\citet{2012ApJ...746..113G} noted that the final exponent in the second last sentence of the Appendix of \citet{2007MNRAS.379..711G} should have read 1/0.5 rather than 0.5 to give $M_{\rm bh} \propto L^{1/0.5}$.} and was later expressed as a (cool gas)-rich versus (cool gas)-poor galaxy sequence in \citet{2013ApJ...764..151G} and \citet{2013ApJ...768...76S}. \citet{2015ApJ...798...54G} revealed that the near-quadratic relation also appeared to encompass active galactic nuclei (AGNs) with virial masses\footnote{Virial masses were derived using a virial factor $f=2.8$ \citep{2011MNRAS.412.2211G}.} as low as $2\times10^5$ M$_{\odot}$. With improved data, \citet{2016ApJ...817...21S} found that the distributions in the $M_{\rm bh}$--$M_{\rm *,sph}$ diagram were better described by a `blue sequence' for late-type galaxies (LTGs) --- which are all S\'ersic galaxies --- and a `red sequence' for early-type galaxies (ETGs), which can be S\'ersic galaxies or core-S\'ersic galaxies. This red versus blue sequence was later emphasised by others, including \citet{2016ApJ...831..134V} and \citet{2020ApJ...898...83D}. Doubling the sample size of spiral galaxies used by \citet{2016ApJ...817...21S}, \citet{2019ApJ...873...85D} could better constrain the $M_{\rm bh}$--$M_{\rm *,sph}$ relation for the LTGs, finding a slope of 2.17$\pm$0.32 to $2.44^{+0.35}_{-0.31}$ depending on the regression analysis used. Doubling the sample size of ETGs used by \citet{2016ApJ...817...21S}, \citet{2019ApJ...876..155S} measured a slope of 1.27$\pm$0.07 for the ETGs but crucially explained why this was misleading. \citet{2019ApJ...876..155S}, and \citet{2016ApJS..222...10S}, knew which ETGs were (pure) elliptical galaxies and which were lenticular or ellicular\footnote{Ellicular galaxy is the name given by \citet{2016ApJ...831..132G} to the ES galaxy type introduced by \citet{1966ApJ...146...28L}. They have intermediate-scale discs which do not dominate the light at large radii, in contrast to the familiar lenticular (S0) galaxies whose large-scale discs do \citep{2019MNRAS.487.4995G}.} galaxies. Separating the ETGs into those with and without discs, \citet{2019ApJ...876..155S} revealed that they followed separate $M_{\rm bh}$--$M_{\rm *,sph}$ relations with similar slopes ($\approx$1.9$\pm$0.2, based on $M_*/L_{3.6} = 0.6$) but offset by an order of magnitude in $M_{\rm bh}$. Therefore, as \citet{2019ApJ...876..155S} explained, published slopes for the near-linear $M_{\rm bh}$--$M_{\rm *,sph}$ relation, i.e., the `red-sequence', are dependent on the sample's arbitrary number of ETGs with and without discs. See, for example, \citet[][with a slope of 0.93$\pm$0.10]{2007MNRAS.379..711G}, \citet[][with a slope of 1.16$\pm$0.08]{2013ARA&A..51..511K}, \citet[][with a slope of 0.846$\pm$0.064]{2016ApJ...818...47S}, and \citet[][with a slope of 1.04$\pm$0.10]{2016ApJ...817...21S}. The slope is not a measure of physical importance --- as has been thought and reported for over a quarter of a century regarding galaxy/black hole coevolution --- but rather a reflection of the sample selection. This revelation has been shown to impact black hole correlations involving not just the spheroid's stellar mass but also the spheroid's size \citep{2020ApJ...903...97S} and the spheroid's range of density measures \citep{2022ApJ...927...67S}. This new knowledge is important because it rewrites our understanding of the interplay between spheroids and their central massive black holes. This realisation was refined by performing multicomponent decompositions of the galaxy light, with recourse to kinematic information and accounting for distinct physical entities such as bars, rings, bulges, and discs detected in the images and the Fourier harmonic analysis of the isophotes \citep{2015ApJ...810..120C}. Here, with updated data, we offer a likely explanation for the offset between the relations followed by elliptical and ellicular/lenticular galaxies. We also raise some of the ensuing implications. In particular, we more clearly elucidate the origin and `red herring' nature of the near-linear $M_{\rm bh}$--$M_{\rm *,sph}$ relation in regard to understanding the (limited caretaker) role for AGN feedback in elliptical galaxies. We previously used a simple conversion of starlight-to-mass in our (galaxy morphology)-dependent scaling diagrams: specifically, $M_*/L_{*,3.6\, \mu m} = 0.6$ \citep{2014ApJ...788..144M}.\footnote{This was based on a \citet{2003PASP..115..763C} `initial mass function'.}\footnote{In practice, while $M_*/L_{obs,3.6}=0.6$ was used for the ETGs \citep{2019ApJ...876..155S}, a lower value of $M_*/L_{\rm obs,3.6}=0.453$ was applied to the LTGs \citep{2019ApJ...873...85D} because the observed luminosity at 3.6~$\mu$m includes both starlight and the glow of warm dust. This reduced ratio encapsulated the mean ratio $L_{*,3.6}/L_{\rm obs.3.6} \approx 0.75$ for LTGs \citep{2015ApJS..219....5Q}.} Such an approach meshes well with the notion that many compact `red nuggets' at redshifts $z \sim 2.5 \pm1$ (both massive and not so massive) have become the bulges of some of today's lenticular and spiral galaxies \citep{2015ApJ...804...32G, 2016MNRAS.457.1916D, 2017ApJ...840...68G, 2022MNRAS.514.3410H}. Such an origin for these bulges would make them old, as \citet{1996AJ....111.2238P} and \citet{2009MNRAS.395...28M} reported, and therefore require a high mass-to-light ratio. However, not every bulge needs to be old. Here we explore colour-dependent $M_*/L_{obs,3.6}$ ratios for a sample of $\sim$100 galaxies pre-observed with the Spitzer Space Telescope and close enough to resolve their bulges ($R_{\rm e} \gtrsim 2\arcsec$). That is, we allow for departures from the assumption that all the bulges have the same $M_*/L \equiv \Upsilon_*$ ratio. Here, we use a $B-V$ colour-dependent mass-to-light ratio prescription to derive the stellar masses. Appendix~\ref{Apdx1} offers an alternative optical-NIR prescription for the $\Upsilon_*$ ratio based on the $V-$[3.6] colour. It provides an analysis less sensitive to star formation (given that star formation may be more reflective of the disc than the spheroid). In Section~\ref{Sec_DaA}, we summarise the salient features of our galaxy sample and describe the prescription for deriving their colour-dependent $M_*/L_{obs,3.6}$ ratios. We have also updated a few black hole masses, some spheroid luminosities, and many galaxy distances, slightly impacting the black hole masses and absolute magnitudes. We provide a data table of final values with sufficient information to trace the origin of the data readily. In Section~\ref{Sec_Results}, we present the $M_{\rm bh}$--$M_{\rm *,sph}$ and $M_{\rm bh}$--$M_{\rm *,gal}$ diagrams and relations as a function of galaxy morphology (E, ES/S0, and S). Section~\ref{Sec_b2e} presents the size-mass relation for our sample of spheroids and uses this to reveal how dry mergers, and the transition from bulges to E galaxies, naturally produce the offset $M_{\rm bh}$--$M_{\rm *,sph}$ and $M_{\rm bh}$--$R_{\rm e,sph}$ relations for E galaxies relative to the bulges in ES/S0 galaxies and also the offset between ES/S0 and E galaxies in the $M_{\rm bh}$--$M_{\rm *,gal}$ diagram. Section~\ref{Sec_Disc_2} explains the apparent overmassive and underermassive black holes (in bulges) relative to the original near-linear relation, with Section~\ref{Sec_Disc_3} presenting the location of relic `red nuggets' at the top of the bulge sequence. The stripped S0 galaxy M32 --- the prototype for the `compact elliptical' galaxy class --- is discussed in Section~\ref{Sec_Disc_4}. Section~\ref{Sec_Disc_5} identifies and discusses what may be the primary bivariate black hole relation in the $M_{\rm bh}$--$M_{\rm *,sph}$ diagram. Section~\ref{Sec_Disc_6} then discusses the galaxy stellar mass function and the (moot) role of AGN feedback in shaping it instead of potentially just maintaining it. Finally, several other implications are briefly mentioned in Section~\ref{Sec_imp}. It is important to note that the authors have been mindful of using the strict interpretation of morphological terms in this paper. An elliptical (E) galaxy has no substantial disc other than perhaps a small nuclear disc, whereas ellicular (ES) and lenticular (S0) galaxies have an intermediate-scale and a large-scale disc, respectively. The expression `early-type galaxy' (ETG) is used to generically refer to the E, ES, and S0 galaxies without a spiral pattern, while the expression `late-type galaxy' (LTG) refers to spiral (S) and irregular (Irr) galaxies. This notation is confined to high-surface brightness galaxies that define the galaxy classification grid seen in \cite{2019MNRAS.487.4995G} and built on the Aitken-Jeans-Lundmark-Hubble galaxy sequence discussed there. The term `spheroid' refers to both an elliptical galaxy and the bulge of a disc galaxy, while the term bulge refers to the bulges of S, ES, and S0 galaxies but not E galaxies. The only (mild) confusion\footnote{\citet{1966ApJ...143..192Z} pointed out that these elliptical-like galaxies are notably more compact than the more commonly known `elliptical galaxies' - many pf which turned out to be lenticular galaxies with large-scale discs.} to this nomenclature is that we will sometimes refer to relic `red nuggets' --- unevolved spheroidal-shaped galaxies from $z\sim 2.5\pm1$ which have not acquired a large-scale disc of stars by today --- as belonging to the bulge sequence. Why we do this will become apparent as one reads on. \section{Data and Analysis}\label{Sec_DaA} \subsection{The sample}\label{Sec_Sample} \citet{2019ApJ...873...85D} and \citet{2019ApJ...876..155S} provide galaxy decompositions for LTGs and ETGs with directly measured black hole masses obtained from the literature. The galaxy decomposition process involved the extraction of a nested set of one-dimensional profiles, including the surface brightness profile, the ellipticity profile, the position angle profile, and an array of profiles quantifying the amplitude of Fourier Harmonic terms used to describe the isophotal deviations from perfect ellipses \citep{2015ApJ...810..120C}. These one-dimensional profiles enable accurate two-dimensional reconstruction of the galaxy without stochastic irregularities due to, for example, star formation or undigested neighbours. Such irregularities remain in the `residual image', obtained by subtracting the smooth reconstruction from the original image, where they can more readily be studied without the (often overwhelming) glow of the host galaxy \citep[e.g.,][]{2021ApJ...923..146G}. The surface brightness profile of the geometric-mean axis\footnote{The geometric-mean axis, also know as the `equivalent axis' $R_{\rm eq}$, is given by the geometric-mean of the major (a) and minor (b) axis. These $R_{\rm eq}=\sqrt{ab}$ radii are {\it equivalent} to a circularised version of a galaxy's quasi-elliptical isophotes.\label{footReq}} is then recreated by optimally fitting a suite of galaxy components. One of the advantages of this approach is that it is not limited to models in which galaxy components may have fixed ellipticity and position angles, as with directly fitting the two-dimensional image. For instance, a single-component triaxial bulge with a radially-varying ellipticity and position angle profile might get broken into two or more components when attempting to model it in two dimensions. The bulk of the sample was previously imaged with the Spitzer Space Telescope at a wavelength of 3.6~$\mu$m. The galaxies were `disassembled' to reveal their components and better establish the luminosity of their spheroidal component. Their samples were supplemented by using optical and near-IR $K_s$-band images when the Spitzer data were either unavailable or when better spatial resolution was required to probe the bulge component. To keep things simple, and minimise the introduction of possible biases, here we avoid potential offsets arising from the use of a range of filters and thus adopted stellar mass-to-light ratios. We do this by solely using the galaxy sample whose structural composition was studied at 3.6~$\mu$m. This sample consists of 73 ETGs\footnote{This sample of 73 ETGs is comprised of 40 from \citet{2016ApJS..222...10S}, of which three (NGC: 821; 1399; and 3377) are remodelled in \cite{Graham:Sahu:22}, plus 33 from \citet{2019ApJ...876..155S}, of which two (NGC~2787 and NGC~5419) are remodelled in \cite{Graham:Sahu:22}.} plus 31 LTGs\footnote{\citet[][their Table~3]{2019ApJ...873...85D} contains 28 galaxies with Spitzer data, including NGC~4395 and NGC 6926 which are bulgeless, and including NGC~224 taken from \citet{2016ApJS..222...10S}. Two of these 28 (NGC~1320 and NGC~4699) are remodelled in \cite{Graham:Sahu:22}. A further three galaxies (NGC~2273, NGC~4945 and UGC~3789) from \citet{2016ApJS..222...10S} are included, taking the tally to 31.}, coming from the larger sample of 84 ETGs \citep{2019ApJ...876..155S} and 43 LTGs \citep{2019ApJ...873...85D}. The smaller fraction of LTGs with useful Spitzer data is a consequence of the need to resolve the bulge component of the galaxy. As such, more LTGs than ETGs had previously required Hubble Space Telescope (HST) data. In passing, it is noted that the (peanut shell)-shaped structures associated with buckled bars \citep{1981A&A....96..164C, 2005MNRAS.358.1477A, 2016MNRAS.459.1276C} --- sometimes referred to as `pseudobulges' --- were either modelled as a `barlens' \citep[e.g.,][their Figure~3]{2011MNRAS.418.1452L, 2019ApJ...876..155S} or effectively folded into the Ferrers bar component during the galaxy decomposition, which can be seen for every galaxy in \citet{2016ApJS..222...10S}, \citet{2019ApJ...873...85D} and \citet{2019ApJ...876..155S}. As noted above, we revisited the decomposition of seven of these (73+31=) 104 galaxies in \citet{Graham:Sahu:22}, and we use the results here. The distances, shown in Table~\ref{Table-data}, are regarded as luminosity distances. As such, the small correction for cosmological surface brightness dimming, 2.5$\log(1+z)^4$ is (implicitly) applied when we convert the 3.6~$\mu$m apparent magnitudes, $m$ --- given in the above four papers --- into absolute magnitudes, $\mathfrak{M}$, using the expression $m-\mathfrak{M}=25+5\log\,D_L$, where $D_L$ is the luminosity distance in Mpc. No Galactic extinction correction has been applied to the 3.6~$\mu$m data because any excess emission from dust in the Milky Way glowing at 3.6~$\mu$m would have effectively been removed during the sky-subtraction procedure \citep[see Section~2.2.1 in ][]{2019ApJ...876..155S}. Finally, no K-correction or evolutionary corrections were applied given the small redshifts involved, typically $z \approx$ 0.01--0.02. The spheroid and galaxy absolute magnitudes were expressed in units of solar luminosity using $\mathfrak{M}_{\odot,3.6} = 6.02$ (AB mag), equal to 3.26 (Vega mag). These were then converted into stellar masses using the prescription described in the following subsection. These masses appear in Table~\ref{Table-data}, along with the references to where one can see each galaxy's decomposition. These references are also the source for the sizes of the spheroids, quantified using the effective half-light radius, $R_{\rm e,sph}$, measured along the `equivalent axis', $R_{\rm eq}$ (see footnote~\ref{footReq}). The masses for the black holes, updated according to the new galaxy distances, are also provided in Table~\ref{Table-data}. Unless an update is indicated, the nearly 100 references for these black hole masses can be traced through \citet{2020ApJ...903...97S}. Following the exclusion of mergers by \citet{2013ARA&A..51..511K}, we exclude from the upcoming Bayesian linear regressions, but not the plots, one LTG plus four ETGs previously identified by others as somewhat unrelaxed mergers (NGC~2960 plus NGC~1194, NGC~1316, NGC~5018 and NGC~5128). We additionally exclude the stripped galaxy NGC~4342 \citep{2014MNRAS.439.2420B} and the dwarf galaxy NGC~404 \citep[Mirach's Ghost: ][]{2017ApJ...836..237N}\footnote{NGC~404 lies within seven arcmin of the second magnitude star Mirach.}. NGC~404 is the only galaxy in our sample with $M_{\rm bh} < 10^6$ M$_\odot$, thereby making it vulnerable to potentially biasing the analyses due to the weight it may have in torquing the regression lines. The ETG with the suspiciously\footnote{NGC~1275 resides 13 degrees from the Galactic plane and has $\sim$0.6 mag of Galactic extinction in the $B$-band.} blue colour of $\sim$0.6 in Figure~\ref{Fig_colour_IP13} is NGC~1275, although we left this galaxy in the sample as its inclusion/exclusion had no appreciable impact. We also included the rather blue LTG NGC~4303 but needed to exclude NGC~5055 due to its uncertain black hole mass \citep{2004A&A...420..147B, 2021MNRAS.500.1933S} and Circinus, an unrelaxed S galaxy known to be undergoing considerable starbursts in addition to hosting an AGN. The $B-V=0.174$ (Vega) colour of Circinus is less than 0.5 and well outside of the applicability range of the $M/L$ equations we are about to introduce. The nine galaxies excluded from the linear regression are marked with a dagger symbol ($\dagger$) in Table~\ref{Table-data}, which includes all 104 galaxies initially considered here. \begin{figure*} \begin{center} \includegraphics[trim=0.0cm 0cm 0.0cm 0cm, height=0.35\textwidth, angle=0]{B_V_vs_Sph_Mag_94} \includegraphics[trim=0.0cm 0cm 0.0cm 0cm, height=0.35\textwidth, angle=0]{BV_color_ML_IP13} \caption{Left-hand panel: (Galactic extinction)-corrected $B-V$ galaxy colour (Vega mag) versus the 3.6~$\mu$m absolute magnitude $\mathfrak{M}_{3.6}$ (AB mag) of the spheroid. Both NGC~1275 and Circinus are close to the Galactic plane. Right-hand panel: The thick curve (with the triangles) shows the same $B-V$ colour versus the logarithm of the stellar mass-to-light ratio at 3.6~$\mu$m obtained from Equation~\ref{Eq_MonL_IP13}, adapted from the dusty galaxy model of \citet[][their Table~6]{2013MNRAS.430.2715I}. For reference, we also show with a thin curve their simple galaxy model (their Table~3, our Equation~\ref{Eq_MonL_simple}). Both models have been converted here to a \citet{2002Sci...295...82K} IMF, see Section~\ref{Sec_IMF} for details. (Circinus falls below the blue-end cut-off of $B-V=0.5$ for the applicable range of these mass-to-light ratio equations.) } \label{Fig_colour_IP13} \end{center} \end{figure*} \begin{table*} \centering \caption{Black hole, spheroid and galaxy masses, and spheroid sizes}\label{Table-data} \begin{tabular}{llllclccccl} \hline Galaxy & Type & Dist.\ (Mpc) & kpc/$\arcsec$ & $R_{\rm e,sph,equiv}$ (kpc) & $\log(M_{\rm bh}/M_\odot)$ & $B-V$ & $V-$[3.6] & $\log(M_{\rm *,sph}/M_\odot)$ & $\log(M_{\rm *,gal}/M_\odot)$ & Ref. \\ (1) & (2) & (3) & (4) & (5) & (6) & (7) & (8) & (9) & (10) & (11) \\ \hline Circinus$^{\dagger}$ & S & 4.21$\pm$0.76 (T08) & 0.020 & 0.46 & 6.25$\pm$0.11 & 0.174 & 2.41 & 9.46$\pm$0.29 & 9.97$\pm$0.22 & DGC19\\ IC 1459 & E & 28.1$\pm$3.6 (T01$^{\prime}$) & 0.134 & 7.68 & 9.38$\pm$0.20 & 0.966 & 3.84 & 11.69$\pm$0.17 & 11.69$\pm$0.17 & SG16\\ IC 2560 & S & 32.9$\pm$5.0 (NED$^i$) & 0.157 & 0.62 & 6.52$\pm$0.11 & [0.70] & [3.17] & 9.66$\pm$0.23 & 10.69$\pm$0.18 & DGC19\\ IC 4296$^*$ & E & 46.9$\pm$3.7 (M00$^{\prime}$) & 0.222 & 9.12 & 9.10$\pm$0.09 & 0.954 & 3.36 & 11.72$\pm$0.14 & 11.74$\pm$0.14 & SGD19\\ NGC 0224 & S & 0.75$\pm$0.03 (R12) & 0.004 & 0.69 & 8.15$\pm$0.16 & 0.865 & 3.58 & 10.23$\pm$0.15 & 11.01$\pm$0.13 & SG16\\ NGC 0253 & S & 3.47$\pm$0.11 (R11) & 0.017 & 0.47 & 7.00$\pm$0.30 & [0.70] & [3.19] & 9.76$\pm$0.25 & 10.71$\pm$0.13 & DGC19\\ NGC 0404$^{\dagger}$ & ES/S0 & 3.06$\pm$0.37 (K02) & 0.015 & 0.058 & 5.74$\pm$0.1 [4f] & 0.889 & 2.77 & 8.03$\pm$0.50 & 9.19$\pm$0.16 & SGD19\\ NGC 0524 & ES/S0 & 27.7$\pm$1.1 (J21) & 0.132 & 1.10 & 9.00$\pm$0.10 & 0.977 & 3.13 & 10.88$\pm$0.15 & 11.38$\pm$0.13 & SGD19\\ NGC 0821 & E & 23.2$\pm$1.8 (T01$^{\prime}$) & 0.111 & 3.45 & 7.59$\pm$0.17 & 0.893 & 2.90 & 10.84$\pm$0.15 & 10.90$\pm$0.14 & GS22\\ NGC 1023 & ES/S0 & 11.0$\pm$0.8 (T01$^{\prime}$) & 0.053 & 0.39 & 7.62$\pm$0.05 & 0.946 & 3.15 & 10.33$\pm$0.16 & 10.89$\pm$0.14 & SG16\\ NGC 1097 & S & 24.8$\pm$0.6 (R14) & 0.119 & 1.35 & 8.38$\pm$0.04 & 0.726 & 3.52 & 10.84$\pm$0.25 & 11.42$\pm$0.13 & DGC19\\ NGC 1194$^{\dagger}$ & ES/S0 & 54.1$\pm$3.8 (NED) & 0.256 & 0.91 & 7.82$\pm$0.04 & 0.893 & 3.61 & 10.78$\pm$0.16 & 11.01$\pm$0.14 & SGD19\\ NGC 1275 & E & 69.0$\pm$9.0 (NED$^i$) & 0.324 & 17.4 & 8.88$\pm$0.21 & 0.616 & 4.04 & 11.56$\pm$0.18 & 11.60$\pm$0.17 & SGD19\\ NGC 1300 & S & 20.7$\pm$1.5 (NED) & 0.100 & 0.74 & 7.86$\pm$0.14 & 0.654 & 2.86 & 9.68$\pm$0.20 & 10.55$\pm$0.14 & DGC19\\ NGC 1316$^{\dagger}$ & ES/S0 & 17.8$\pm$5.9 (NED$^i$) & 0.086 & 1.37 & 8.16$\pm$0.29 & 0.871 & 3.60 & 11.05$\pm$0.35 & 11.69$\pm$0.31 & SG16\\ NGC 1320 & S & 36.8$\pm$2.6 (NED) & 0.175 & 0.20 & 6.77$\pm$0.22 & 0.828 & 3.50 & 10.13$\pm$0.16 & 10.70$\pm$0.14 & GS22 \\ NGC 1332 & ES/S0 & 22.0$\pm$1.8 (T01$^{\prime}$) & 0.105 & 1.89 & 9.15$\pm$0.07 & 0.932 & 3.41 & 11.15$\pm$0.15 & 11.17$\pm$0.14 & SG16\\ NGC 1374 & ES/S0 & 19.0$\pm$1.1 (T01$^{\prime}$) & 0.091 & 1.07 & 8.76$\pm$0.05 & 0.908 & 3.19 & 10.30$\pm$0.16 & 10.59$\pm$0.13 & SGD19\\ NGC 1398 & S & 24.8$\pm$4.5 (T08) & 0.119 & 1.24 & 8.03$\pm$0.11 & 0.888 & 3.36 & 10.76$\pm$0.29 & 11.44$\pm$0.20 & DGC19\\ NGC 1399 & E & 19.2$\pm$1.4 (T01$^{\prime}$) & 0.092 & 5.71 & 8.67$\pm$0.06 & 0.949 & 3.74 & 11.46$\pm$0.16 & 11.46$\pm$0.16 & GS22\\ NGC 1407 & E & 27.8$\pm$3.3 (T01$^{\prime}$) & 0.133 & 6.29 & 9.65$\pm$0.08 & 0.969 & 3.31 & 11.60$\pm$0.17 & 11.66$\pm$0.16 & SGD19\\ NGC 1600 & E & 71.7$\pm$2.7 (J21) & 0.336 & 16.7 & 10.25$\pm$0.04 & 0.972 & 3.56 & 12.06$\pm$0.13 & 12.06$\pm$0.13 & SGD19\\ NGC 2273 & S & 30.3$\pm$4.0 (NED$^i$) & 0.145 & 0.28 & 6.95$\pm$0.06 & 0.828 & 3.41 & 10.35$\pm$0.22 & 10.83$\pm$0.17 & SG16\\ NGC 2549 & ES/S0 & 12.2$\pm$1.6 (T01$^{\prime}$) & 0.059 & 0.18 & 7.15$\pm$0.60 & 0.913 & 3.19 & 9.67$\pm$0.19 & 10.21$\pm$0.17 & SG16\\ NGC 2778 & ES/S0 & 22.1$\pm$3.1 (T01$^{\prime}$) & 0.106 & 0.23 & 7.18$\pm$0.35 & 0.911 & 3.06 & 9.49$\pm$0.23 & 10.15$\pm$0.17 & SG16\\ NGC 2787 & ES/S0 & 7.2$\pm$1.2 (T01$^{\prime}$) & 0.035 & 0.14 & 7.59$\pm$0.09 & 0.942 & 3.34 & 9.37$\pm$0.24 & 10.10$\pm$0.19 & GS22\\ NGC 2960$^{\dagger}$ & S & 73.0$\pm$5.1 (NED) & 0.342 & 0.75 & 7.07$\pm$0.05 & [0.70] & [3.34] & 10.44$\pm$0.16 & 10.86$\pm$0.14 & DGC19\\ NGC 2974 & S & 21.5$\pm$2.5 (T08) & 0.103 & 0.67 & 8.23$\pm$0.07 & 0.952 & 3.45 & 10.48$\pm$0.22 & 10.98$\pm$0.16 & DGC19\\ NGC 3031 & S & 3.48$\pm$0.13 (K12) & 0.017 & 0.73 & 7.83$\pm$0.09 & 0.879 & 3.21 & 10.34$\pm$0.25 & 10.83$\pm$0.13 & DGC19\\ NGC 3079 & S & 16.5$\pm$2.9 (T08) & 0.079 & 0.34 & 6.38$\pm$0.12 & 0.670 & 4.03 & 9.88$\pm$0.29 & 10.64$\pm$0.20 & DGC19\\ NGC 3091 & E & 58.6$\pm$10.9 (T07) & 0.276 & 14.1 & 9.62$\pm$0.08 & 0.962 & 3.74 & 11.86$\pm$0.20 & 11.86$\pm$0.20 & SG16\\ NGC 3115 & ES/S0 & 9.3$\pm$0.4 (T01$^{\prime}$) & 0.045 & 1.55 & 8.94$\pm$0.25 & 0.929 & 3.30 & 10.87$\pm$0.14 & 10.95$\pm$0.13 & SG16\\ NGC 3227 & S & 25.7$\pm$3.2 (K15) & 0.123 & 1.03 & 7.97$\pm$0.14 & 0.800 & 3.19 & 10.31$\pm$0.22 & 11.07$\pm$0.16 & DGC19\\ NGC 3245 & ES/S0 & 20.1$\pm$1.9 (T01$^{\prime}$) & 0.097 & 0.23 & 8.30$\pm$0.12 & 0.888 & 3.06 & 10.12$\pm$0.17 & 10.70$\pm$0.15 & SG16\\ NGC 3368 & S & 10.8$\pm$1.5 (NED$^i$) & 0.052 & 0.25 & 6.89$\pm$0.11 & 0.838 & 3.33 & 9.95$\pm$0.17 & 10.83$\pm$0.17 & DGC19\\ NGC 3377 & E & 10.8$\pm$0.4 (T01$^{\prime}$) & 0.052 & 2.30 & 7.89$\pm$0.03 & 0.830 & 3.29 & 10.30$\pm$0.14 & 10.36$\pm$0.13 & GS22\\ NGC 3379 & E & 10.9$\pm$1.6 (K15) & 0.053 & 2.70 & 8.62$\pm$0.13 & 0.938 & 3.44 & 10.97$\pm$0.20 & 10.97$\pm$0.20 & SG16\\ NGC 3384 & ES/S0 & 11.2$\pm$0.7 (T01$^{\prime}$) & 0.054 & 0.30 & 7.23$\pm$0.05 & 0.907 & 3.33 & 10.14$\pm$0.16 & 10.67$\pm$0.14 & SG16\\ NGC 3414 & E & 24.3$\pm$3.7 (T01$^{\prime}$) & 0.117 & 2.98 & 8.38$\pm$0.09 & 0.948 & 3.39 & 10.95$\pm$0.19 & 10.98$\pm$0.18 & SG16\\ NGC 3489 & ES/S0 & 11.6$\pm$0.8 (T01$^{\prime}$) & 0.056 & 0.095 & 6.76$\pm$0.07 & 0.816 & 2.98 & 9.53$\pm$0.20 & 10.30$\pm$0.14 & SG16\\ NGC 3585 & E & 19.3$\pm$1.6 (T01$^{\prime}$) & 0.093 & 8.03 & 8.49$\pm$0.13 & 0.914 & 3.84 & 11.38$\pm$0.15 & 11.39$\pm$0.14 & SG16\\ NGC 3607 & E & 25.0$\pm$3.2 (K15) & 0.120 & 7.86 & 8.16$\pm$0.18 & 0.911 & 3.54 & 11.43$\pm$0.17 & 11.46$\pm$0.17 & SG16\\ NGC 3608 & E & 22.1$\pm$1.4 (T01$^{\prime}$) & 0.106 & 4.60 & 8.30$\pm$0.17 & 0.922 & 3.45 & 10.98$\pm$0.14 & 10.98$\pm$0.14 & SG16\\ NGC 3627 & S & 10.4$\pm$1.8 (K15) & 0.050 & 0.20 & 6.94$\pm$0.09 & 0.701 & 3.24 & 9.72$\pm$0.21 & 10.76$\pm$0.20 & DGC19\\ NGC 3665 & ES/S0 & 34.7$\pm$2.4 (T07) & 0.166 & 2.12 & 8.76$\pm$0.10 & 0.933 & 3.54 & 11.14$\pm$0.16 & 11.39$\pm$0.14 & SGD19\\ NGC 3842 & E & 87.5$\pm$4.1 (J21) & 0.407 & 30.0 & 9.94$\pm$0.13 & 0.941 & 3.84 & 11.91$\pm$0.14 & 11.93$\pm$0.13 & SG16\\ NGC 3923 & E & 22.1$\pm$2.9 (T01$^{\prime}$) & 0.106 & 8.35 & 9.47$\pm$0.13 & 0.929 & 3.74 & 11.55$\pm$0.17 & 11.55$\pm$0.17 & SGD19\\ NGC 3998 & ES/S0 & 13.6$\pm$1.2 (T01$^{\prime}$) & 0.065 & 0.31 & 8.33$\pm$0.43[4a] & 0.936 & 3.47 & 10.12$\pm$0.26 & 10.61$\pm$0.15 & SG16\\ NGC 4026 & ES/S0 & 13.2$\pm$1.7 (T01$^{\prime}$) & 0.063 & 0.15 & 8.26$\pm$0.12 & 0.901 & 3.27 & 10.19$\pm$0.22 & 10.44$\pm$0.17 & SGD19\\ NGC 4151 & S & 19.0$\pm$2.5 (H14) & 0.094 & 0.56 & 7.69$\pm$0.37 & 0.706 & 3.44 & 10.28$\pm$0.22 & 10.62$\pm$0.17 & DGC19\\ NGC 4258 & S & 7.6$\pm$0.17 (H13) & 0.037 & 0.98 & 7.60$\pm$0.01 & 0.676 & 3.37 & 10.02$\pm$0.19 & 10.70$\pm$0.13 & DGC19\\ NGC 4261 & E & 30.4$\pm$2.7 (T01$^{\prime}$) & 0.145 & 6.86 & 9.20$\pm$0.09[4b] & 0.975 & 3.69 & 11.52$\pm$0.16 & 11.54$\pm$0.15 & SG16\\ NGC 4291 & E & 25.2$\pm$3.7 (T01$^{\prime}$) & 0.121 & 1.86 & 8.51$\pm$0.37 & 0.927 & 3.38 & 10.80$\pm$0.18 & 10.80$\pm$0.18 & SG16\\ NGC 4303 & S & 19.3$\pm$0.6 (R14) & 0.093 & 0.20 & 6.78$\pm$0.17 & 0.510 & 2.90 & 9.60$\pm$0.25 & 10.66$\pm$0.13 & DGC19\\ NGC 4339 & ES/S0 & 15.8$\pm$1.3 (T01$^{\prime}$) & 0.076 & 0.49 & 7.62$\pm$0.33 & 0.887 & 2.96 & 9.73$\pm$0.21 & 10.22$\pm$0.14 & SGD19\\ NGC 4342$^{\dagger}$ & ES/S0 & 22.8$\pm$0.8 (B09$^{\prime}$) & 0.110 & 0.52 & 8.65$\pm$0.18 & 0.932 & 3.52 & 10.04$\pm$0.15 & 10.36$\pm$0.15 & SGD19\\ NGC 4350 & ES/S0 & 17.0$\pm$0.5 (B15) & 0.082 & 1.59 & 8.87$\pm$0.41 & 0.926 & 3.43 & 10.39$\pm$0.25 & 10.66$\pm$0.13 & SGD19\\ NGC 4371 & ES/S0 & 16.4$\pm$3.4 (T07) & 0.079 & 0.70 & 6.84$\pm$0.12 & 0.948 & 3.34 & 9.99$\pm$0.30 & 10.70$\pm$0.22 & SGD19\\ NGC 4374 & E & 17.7$\pm$0.9 (T01$^{\prime}$) & 0.085 & 11.0 & 8.95$\pm$0.05 & 0.944 & 3.75 & 11.61$\pm$0.15 & 11.61$\pm$0.15 & SG16\\ NGC 4388 & S & 18.0$\pm$3.6 (S14) & 0.087 & 1.24 & 6.90$\pm$0.10 & 0.711 & 3.34 & 10.07$\pm$0.30 & 10.45$\pm$0.21 & DGC19\\ NGC 4395 & S & 4.56$\pm$0.17 (S18) & 0.022 & ... & 5.62$\pm$0.17 & 0.445 & 2.96 & ... & 9.15$\pm$0.13 & DGC19 \\ NGC 4429 & ES/S0 & 16.6$\pm$0.8 (C08) & 0.080 & 0.90 & 8.18$\pm$0.08 & 0.950 & 3.40 & 10.60$\pm$0.20 & 11.04$\pm$0.13 & SGD19\\ \hline \end{tabular} \end{table*} \setcounter{table}{0} \begin{table*} \centering \caption{Continued} \begin{tabular}{llllclccccl} \hline Galaxy & Type & Dist.\ (Mpc) & kpc/$\arcsec$ & $R_{\rm e,sph,equiv}$ (kpc) & $\log(M_{\rm bh}/M_\odot)$ & $B-V$ & $V-$[3.6] & $\log(M_{\rm *,sph}/M_\odot)$ & $\log(M_{\rm *,gal}/M_\odot)$ & Ref. \\ (1) & (2) & (3) & (4) & (5) & (6) & (7) & (8) & (9) & (10) & (11)\\ \hline NGC 4434 & ES/S0 & 22.3$\pm$0.7 (B09$^{\prime}$) & 0.107 & 0.57 & 7.85$\pm$0.17 & 0.861 & 3.07 & 9.95$\pm$0.20 & 10.22$\pm$0.13 & SGD19\\ NGC 4459 & ES/S0 & 15.5$\pm$1.6 (T01$^{\prime}$) & 0.075 & 0.98 & 7.82$\pm$0.10 & 0.910 & 3.28 & 10.56$\pm$0.21 & 10.77$\pm$0.15 & SG16\\ NGC 4472 & E & 15.7$\pm$0.7 (T01$^{\prime}$) & 0.075 & 10.2 & 9.36$\pm$0.04 & 0.940 & 3.72 & 11.75$\pm$0.13 & 11.75$\pm$0.13 & SG16\\ NGC 4473 & E & 15.1$\pm$0.9 (T01$^{\prime}$) & 0.073 & 2.69 & 8.07$\pm$0.36 & 0.935 & 3.30 & 10.75$\pm$0.13 & 10.83$\pm$0.13 & SG16\\ NGC 4486 & E & 16.8$\pm$0.8 (EHT) & 0.081 & 7.06 & 9.81$\pm$0.06[4c] & 0.940 & 3.60 & 11.67$\pm$0.15 & 11.67$\pm$0.15 & SG16\\ NGC 4501 & S & 17.0$\pm$0.5 (B15) & 0.082 & 1.67 & 7.31$\pm$0.08 & 0.696 & 3.53 & 10.46$\pm$0.25 & 11.02$\pm$0.13 & DGC19\\ NGC 4526 & ES/S0 & 16.3$\pm$1.5 (T01$^{\prime}$) & 0.078 & 1.16 & 8.65$\pm$0.04 & 0.940 & 3.38 & 10.79$\pm$0.26 & 11.13$\pm$0.15 & SGD19\\ NGC 4552 & E & 14.8$\pm$1.0 (T01$^{\prime}$) & 0.071 & 5.08 & 8.67$\pm$0.05 & 0.944 & 3.44 & 11.01$\pm$0.16 & 11.07$\pm$0.16 & SGD19\\ NGC 4564 & ES/S0 & 14.4$\pm$1.1 (T01$^{\prime}$) & 0.069 & 0.41 & 7.77$\pm$0.06 & 0.901 & 3.20 & 10.08$\pm$0.16 & 10.35$\pm$0.14 & SG16\\ NGC 4578 & ES/S0 & 16.2$\pm$0.5 (B09$^{\prime}$) & 0.078 & 0.49 & 7.28$\pm$0.35 & 0.842 & 3.25 & 9.79$\pm$0.15 & 10.24$\pm$0.13 & SGD19\\ NGC 4594 & S & 9.55$\pm$0.44 (Mc16) & 0.046 & 1.90 & 8.81$\pm$0.03 & 0.935 & 3.11 & 11.04$\pm$0.25 & 11.26$\pm$0.13 & DGC19\\ NGC 4596 & ES/S0 & 17.0$\pm$0.5 (B15) & 0.082 & 0.74 & 7.90$\pm$0.20 & 0.921 & 3.37 & 10.28$\pm$0.20 & 10.86$\pm$0.13 & SG16\\ NGC 4621 & E & 17.6$\pm$1.6 (T01$^{\prime}$) & 0.084 & 7.64 & 8.59$\pm$0.06 & 0.912 & 3.56 & 11.24$\pm$0.16 & 11.28$\pm$0.15 & SG16\\ NGC 4649 & E & 16.2$\pm$1.1 (T01$^{\prime}$) & 0.078 & 6.29 & 9.66$\pm$0.10 & 0.946 & 3.58 & 11.56$\pm$0.14 & 11.56$\pm$0.14 & SGD19\\ NGC 4697 & E & 11.3$\pm$0.7 (T01$^{\prime}$) & 0.054 & 12.2 & 8.26$\pm$0.04 & 0.884 & 3.81 & 11.07$\pm$0.15 & 11.12$\pm$0.14 & SG16\\ NGC 4699 & S & 20.4$\pm$3.8 (K13) & 0.098 & 0.23 & 8.27$\pm$0.09 & 0.860 & 3.24 & 10.30$\pm$0.25 & 11.27$\pm$0.20 & GS22 \\ NGC 4736 & S & 5.0$\pm$0.4 (T01$^{\prime}$) & 0.024 & 0.23 & 6.83$\pm$0.11 & 0.735 & 3.48 & 10.03$\pm$0.21 & 10.51$\pm$0.14 & DGC19\\ NGC 4742 & ES/S0 & 14.9$\pm$1.1 (T01$^{\prime}$) & 0.072 & 0.25 & 7.13$\pm$0.18 & 0.774 & 2.91 & 9.78$\pm$0.16 & 10.07$\pm$0.14 & SGD19\\ NGC 4762 & ES/S0 & 17.0$\pm$0.5 (B15) & 0.082 & 0.18 & 7.24$\pm$0.14 & 0.841 & 3.35 & 9.74$\pm$0.15 & 10.83$\pm$0.13 & SGD19\\ NGC 4826 & S & 7.2$\pm$0.7 (T01$^{\prime}$) & 0.035 & 0.42 & 6.18$\pm$0.12 & 0.803 & 3.31 & 9.88$\pm$0.21 & 10.74$\pm$0.15 & DGC19\\ NGC 4889 & E & 96.3$\pm$6.7 (NED) & 0.446 & 27.1 & 10.3$\pm$0.44 & 1.031 & 3.93 & 12.26$\pm$0.14 & 12.26$\pm$0.14 & SG16\\ NGC 4945 & S & 3.56$\pm$0.20 (Mo16) & 0.017 & 0.16 & 6.13$\pm$0.30 & [0.70] & [3.09] & 9.29$\pm$0.20 & 10.42$\pm$0.13 & SG16\\ NGC 5018$^{\dagger}$ & ES/S0 & 38.4$\pm$2.7 (NED) & 0.183 & 1.13 & 8.00$\pm$0.08 & 0.836 & 3.10 & 10.93$\pm$0.16 & 11.31$\pm$0.14 & SGD19\\ NGC 5055$^{\dagger}$ & S & 8.87$\pm$0.39 (M17) & 0.043 & 1.87 & 8.94$\pm$0.10 & 0.704 & 3.41 & 10.49$\pm$0.25 & 10.81$\pm$0.13 & DGC19\\ NGC 5077 & E & 39.8$\pm$7.4 (T07) & 0.189 & 4.35 & 8.85$\pm$0.23 & 0.987 & 3.60 & 11.41$\pm$0.20 & 11.41$\pm$0.20 & SG16\\ NGC 5128$^{\dagger}$ & ES/S0 & 3.76$\pm$0.05 (K07) & 0.018 & 1.09 & 7.65$\pm$0.12 & 0.899 & 3.60 & 10.71$\pm$0.25 & 11.14$\pm$0.12 & SG16\\ NGC 5252 & ES/S0 & 104.0$\pm$7.3 (NED) & 0.480 & 0.71 & 9.03$\pm$0.40 & [0.90] & [3.42] & 10.97$\pm$0.27 & 11.50$\pm$0.15 & SGD19\\ NGC 5419 & E & 57.0$\pm$4.0 (NED) & 0.269 & 10.8 & 9.87$\pm$0.14 & 0.986 & 3.42 & 11.87$\pm$0.16 & 11.87$\pm$0.14 & GS22\\ NGC 5576 & E & 24.5$\pm$1.6 (T01$^{\prime}$) & 0.118 & 5.82 & 8.19$\pm$0.10 & 0.863 & 3.35 & 10.90$\pm$0.16 & 10.90$\pm$0.16 & SG16\\ NGC 5813 & ES/S0 & 31.0$\pm$2.6 (T01$^{\prime}$) & 0.148 & 2.10 & 8.83$\pm$0.06 & 0.940 & 3.17 & 10.96$\pm$0.16 & 11.34$\pm$0.14 & SGD19\\ NGC 5845 & ES/S0 & 25.0$\pm$2.4 (T01$^{\prime}$) & 0.120 & 0.63 & 8.41$\pm$0.22 & 0.972 & 3.38 & 10.26$\pm$0.21 & 10.46$\pm$0.15 & SGD19\\ NGC 5846 & E & 24.0$\pm$2.2 (T01$^{\prime}$) & 0.115 & 9.59 & 9.04$\pm$0.06 & 0.961 & 3.79 & 11.55$\pm$0.15 & 11.55$\pm$0.15 & SG16\\ NGC 6251 & E & 104.6$\pm$7.3 (NED) & 0.483 & 14.5 & 8.77$\pm$0.16 & [0.90] & [3.62] & 11.87$\pm$0.15 & 11.87$\pm$0.15 & SG16\\ NGC 6861 & ES/S0 & 27.0$\pm$4.0 (T01$^{\prime}$) & 0.129 & 2.60 & 9.30$\pm$0.08 & 0.962 & 3.61 & 11.07$\pm$0.19 & 11.15$\pm$0.18 & SGD19\\ NGC 6926 & S & 85.6$\pm$6.0 (NED) & 0.399 & ... & 7.68$\pm$0.50 [4d] & 0.586 & 3.02 & ... & 11.13$\pm$0.14 & DGC19 \\ NGC 7052 & E & 61.9$\pm$2.6 (J21) & 0.291 & 5.83 & 9.35$\pm$0.05[4e] & [0.90] & [3.53] & 11.46$\pm$0.13 & 11.46$\pm$0.13 & SGD19\\ NGC 7332 & ES/S0 & 22.2$\pm$2.0 (T01$^{\prime}$) & 0.106 & 0.26 & 7.06$\pm$0.20 & 0.877 & 3.40 & 10.17$\pm$0.17 & 10.79$\pm$0.15 & SGD19\\ NGC 7457 & ES/S0 & 12.7$\pm$1.2 (T01$^{\prime}$) & 0.061 & 0.40 & 6.96$\pm$0.30 & 0.843 & 3.07 & 9.34$\pm$0.17 & 10.12$\pm$0.15 & SGD19\\ NGC 7582 & S & 22.2$\pm$4.0 (N11) & 0.106 & 0.48 & 7.72$\pm$0.12 & 0.737 & 3.60 & 10.28$\pm$0.29 & 10.90$\pm$0.20 & DGC19\\ NGC 7619 & E & 46.6$\pm$1.7 (J21) & 0.221 & 12.8 & 9.36$\pm$0.09 & 0.968 & 3.69 & 11.69$\pm$0.14 & 11.71$\pm$0.13 & SG16\\ NGC 7768 & E & 108.2$\pm$7.6 (NED) & 0.499 & 21.0 & 9.09$\pm$0.15 & 0.906 & 3.83 & 11.90$\pm$0.16 & 11.90$\pm$0.16 & SG16\\ UGC 3789 & S & 50.7$\pm$5.2 (R13) & 0.240 & 0.58 & 7.07$\pm$0.05 & [0.70] & [3.27] & 10.11$\pm$0.26 & 10.68$\pm$0.15 & SG16\\ \hline \end{tabular} Column~1: Galaxy name ($^{\dagger}$ Excluded from the Bayesian linear regression. $^*$ IC~4296 = Abell 3565-BCG). Column~2: Broad galaxy type from \citet[][their Table~A1]{2020ApJ...903...97S}. Column~3: Adopted `luminosity distance' in Mpc: T01$^{\prime}$ = \citet{2001ApJ...546..681T} with a 0.083 mag reduction to their distance moduli (see Section~\ref{Sec_IMF}); T08 = \citet{2008ApJ...676..184T, 2009AJ....138..323T}; Y12 = \citet{2012PASJ...64..103Y}; M00$^{\prime}$ = \citet{2000A&A...361...68M} with a 0.083 mag reduction to their distance moduli; R12 = \citet{2012ApJ...745..156R}; R11 = \citet{2011ApJS..195...18R}; K02 = \citet{2002A&A...389..812K}; R14 = \citet{2014AJ....148..107R}; R13 = \citet{2013ApJ...767..154R}; NED$^i$ = median redshift-independent distance from NED (and the standard deviation associated with the {\it mean} redshift-independent distance); NED = (Virgo + GA + Shapley)-corrected Hubble flow distance based on $H_0=73$ km s$^{-1}$ Mpc$^{-1}$; T07 = \citet{2007A&A...465...71T} mean value via NED; K12 = \citet{2012ApJ...747...15K}; K15 = \citet{2015AstBu..70....1K}; H14 = \citet{2014Natur.515..528H}; H13 = \citet{2013ApJ...775...13H}; B09$^{\prime}$ = \citet{2009ApJ...694..556B} with a 0.023 mag reduction to their distance moduli (see Section~\ref{Sec_IMF}), and using the distance for NGC~4365 in the case of NGC~4342 \citep{2014MNRAS.439.2420B}; S14 = \citet{2014MNRAS.444..527S}; S18 = \citet{2018ApJS..235...23S}; C08 = \citet{2008ApJ...683...78C}; EHT = \citet{2019ApJ...875L...1E}; Mc16 = \citet{2016AJ....152..144M}; Mo16 \ \citet{2016MNRAS.457.1419M}; B15 = \citep{2015A&A...579A.102B}; K13 = \citet{2013MNRAS.429.2677K}; J21 = \citet{2021ApJS..255...21J}; M17 = \citet{2017AJ....154...51M}; K07 = \citet{2007AJ....133..504K}; N11 = \citet{2011A&A...532A.104N}. Column~4: Scale size conversion based on the `angular-size distance' (not shown) and using $H_0=73$ km s$^{-1}$ Mpc$^{-1}$, $\Omega_m=0.3$ and $\Omega_{\Lambda}=0.7$. Column~5: Effective half-light radius of the spheroid, derived from the geometric-mean axis, equivalent to a circularised version of the quasi-elliptical isophotes. Column~6: Black hole masses from the compilation in \citet[][and the reference-chain therein]{2020ApJ...903...97S}, after adjusting to the distances in Column~3. Exceptions are: 4a = \citet{2018MNRAS.473.2930D}; 4b = \citet{2021ApJ...908...19B}; 4c = \citet{2019ApJ...875L...6E}; 4d = \citet[][upper limit]{2018ApJ...854..124Z}; 4e = \citet{2021MNRAS.503.5984S}; and 4f = \citet{2020MNRAS.496.4061D}. Columns~7 and 8: (Galactic extinction)-corrected $B-V$ and $V-$[3.6] colours, obtained from NED. The former are used in Equation~\ref{Eq_MonL_IP13} and the latter in the Appendix equation~\ref{Eq-Schom-21}, to derive the 3.6~$\mu$m stellar mass-to-light ratios for calculating the stellar-masses. Values in [brackets] are assumed rather than measured. Column~9: Spheroid stellar mass based on the Spitzer apparent magnitude reported by either: SG16 \citep{2016ApJ...817...21S}; DGC19 \citep{2019ApJ...873...85D}; SGD19 \citep{2019ApJ...887...10S}; or GS22 \citep{Graham:Sahu:22}. The revised distances, colour-dependent $M_*/L$ ratios, and updated magnitudes for seven systems presented in SG22, results in the updates, shown here, for the stellar masses reported in \citet{2020ApJ...903...97S}. Column~10: Updated galaxy stellar mass. Column~11: Reference displaying the multicomponent decomposition used to obtain both the spheroid magnitude and size, and the galaxy magnitude. \end{table*} \subsection{Stellar mass-to-light ratios}\label{Sec_IP13} As illustrated by, for example, \citet[][their Figure~1]{2014AJ....148...77M}, \citet[][their Section~7]{2017ApJS..233...13Z}, and \citet[][their Figure~4]{2019ApJ...876..155S}, the colour-dependent mass-to-light ratio prescriptions from different papers do not agree with each other. Even after correcting for the different assumptions in the stellar population models, the equations from different papers do not agree \citep[][their Figure~6]{2014AJ....148...77M}. \citet[][their Figure~10]{2019MNRAS.483.1496S} present half a dozen ($B-V$)-dependent relations for the mass-to-light ratio. The relation from \citet{2013MNRAS.430.2715I} sits in the middle and is therefore adopted here as a middle ground. In the Appendix, we additionally show the result of adopting the latest relation from \citet{2022arXiv220202290S}. We have taken the $B_T$ and $V_T$ total galaxy magnitudes from \citet{1991rc3..book.....D}, as listed in the NASA/IPAC Extragalactic Database (NED)\footnote{\url{http://nedwww.ipac.caltech.edu}}, and then corrected these for Galactic extinction based on the near-infrared maps of \citet{2011ApJ...737..103S}, again, as provided by NED. These Johnson-Cousins $B$ and $V$ magnitudes are Vega magnitudes and benefit from having been (i) derived from wide field-of-view imaging from which the sky-background was readily available, and (ii) taken in both the Northern and Southern hemisphere, thereby capturing most of our sample. For three ETGs (NGC~6251, NGC~5252, and NGC~7052), and three LTGs (IC~2560, NGC~253, and NGC~2960), either the $B$- or the $V$-band magnitude was not available. For these three ETGs we assigned a $B-V$ colour of 0.9, and for these three LTGs we assigned a $B-V$ colour of 0.7. In passing, we recognise that spheroid colours, rather than galaxy colours, would be advantageous. However, this would encompass considerable additional work, requiring multicomponent decomposition of two optical bands. Moreover, while both LTGs and ETGs can have colour gradients --- i.e., varying colour with radius, which can be due to the bulge-to-disc transition --- our sample is dominated by early-type spirals (Sa--Sb) and early-type galaxies (E-S0) for which the bulge and disc colours within individual galaxies may not be too dissimilar \citep{1996AJ....111.2238P}. The (Galactic extinction)-corrected $B_T-V_T$ galaxy colour, hereafter $B-V$, is shown in Figure~\ref{Fig_colour_IP13} and provided in Table~\ref{Table-data}. It was used to obtain the stellar mass-to-light ratio at 3.6~$\mu$m as follows. We started with the ($B-V$)-dependent expression for the $K$-band $M_*/L_K$ ratio taken from Table~6 of \citet{2013MNRAS.430.2715I}. It is based on realistic dusty models, designed for ``samples that include a range of morphologies, intrinsic colours and random inclinations''. It is such that \begin{equation} \label{EqIP13_tab6} \log(M_*/L_K) = 0.866 (B-V) - 0.926, \end{equation} which is reportedly based on the \citet{1998ASPC..134..483K} `initial mass function' (IMF)\footnote{The IMF is the histogram of stellar birth masses.}, and good for $0.5 < B-V < 1.1$. \citet{2013MNRAS.430.2715I} report a $\pm$0.1~dex (25\%) uncertainty on these $M_*/L_K$ ratios. As they noted, the combination of dust attenuation (dimming the optical magnitudes) and reddening (of the $B-V$ colour) somewhat cancel to provide $M_*/L_K$ ratios that are consistent with those derived from their simpler (dust free) galaxy model. This partial nulling behaviour was noted by \citet{2003ApJS..149..289B} and can be seen in \citet[][their Fig.~13]{2007MNRAS.379.1022D}, when assuming the dust models of \citet{2000A&A...362..138P}. Here, we convert Eq.~\ref{EqIP13_tab6} into a 3.6~$\mu$m equation using the following relation taken from the start of Section 5.6 in \citet[][see their Fig.~7]{2019MNRAS.483.1496S}: \begin{equation} m_K - m_{3.6} = 0.54 - 0.42(B-V). \nonumber \end{equation} Taking 2.5 times the logarithm of $(M_*/L_K)L_K = (M_*/L_{3.6})L_{3.6}$ and substituting in this $m_K - m_{3.6}$ colour term, one obtains \begin{equation} \log(M_*/L_{3.6}) = 1.034(B-V) - 1.142. \label{Eq-ML36} \end{equation} While the use of individual $m_K - m_{3.6}$ colours rather than the above mean ($B-V$)-dependent relation might seem preferable, in practice it can become problematic due to the different method used to determine the total $K$-band and Spitzer magnitudes \citep[e.g.,][their Fig.~2]{2013ApJ...768...76S}. As noted above, \citet{2013MNRAS.430.2715I} initially derived a colour-dependent $M_*/L$ relation for a less complicated galaxy model based on composite stellar populations. This may be more applicable for the ETGs, and is such that $\log(M_*/L_K) = 1.055 (B-V) - 1.066$, for $0.2 < B-V < 1.0$. Morphing this in the same manner as done to Equation~\ref{EqIP13_tab6} gives the relation \begin{equation} \log(M_*/L_{3.6}) = 1.223(B-V) - 1.282. \label{Eq_ippy1} \end{equation} In the following subsection, we adjust these expressions (equations~\ref{Eq-ML36} and \ref{Eq_ippy1}) to align them with the \citet{2002Sci...295...82K} IMF. \pagebreak \subsubsection{Consideration of the IMF}\label{Sec_IMF} The stellar mass-to-light ratios from the above, and all, stellar population models are dependent upon the assumed IMF \citep[see ][]{2013pss5.book..115K}. To convert from the \citet{1998ASPC..134..483K} IMF\footnote{The concluding section in \citet{1998ASPC..134..483K} notes an IMF slope $\alpha_1=1.3$, and $\alpha_2=2.2$ from their previous sections, along with $\alpha_3=2.7$ for representing the IMF of the Galactic field.} (spanning 0.1--100 M$_\odot$) --- which was used by \citet{2013MNRAS.430.2715I} and is inherent in the previous equations --- to an alternative IMF, the logarithm of the $M_*/L$ ratio needs to be adjusted by a (near) constant factor \citep[e.g.,][]{2014ARA&A..52..415M}. This is shown by \citet[][their Figure~12]{2006MNRAS.372.1149F}, where it can be seen that switching to the Salpeter IMF (spanning 0.1--100 M$_\odot$)\footnote{The mass limits within which the IMF used by \citet{2013MNRAS.430.2715I} was integrated are stated in \citet{2004MNRAS.347..691P}.} requires adding 0.225 dex to the above equations. Conversion to other assumed IMFs can be done following the offsets provided by, for example, \citet[][their p.306]{2003ApJS..149..289B} or \citet[][their Table~2]{2010MNRAS.404.2087B}. In this paper, we adopt the \citet{2002Sci...295...82K} IMF, detailed further in \citet{2013pss5.book..115K}, and have therefore added 0.225 dex and subtracted 0.30 dex from Eq.~\ref{Eq-ML36} to give \begin{equation} \log(M_*/L_{3.6}) = 1.034(B-V) - 1.067. \label{Eq_MonL_IP13} \end{equation} Following \citet{2013MNRAS.430.2715I}, we assign a 25 percent uncertainty to these $M_*/L_{3.6}$ ratios. In the same way, Equation~\ref{Eq_ippy1} becomes \begin{equation} \log(M_*/L_{3.6}) = 1.223(B-V) - 1.207. \label{Eq_MonL_simple} \end{equation} We have included this additional relation (Equation~\ref{Eq_MonL_simple}), shown in the right-hand panel of Figure~\ref{Fig_colour_IP13}, simply to demonstrate that it yields similar $M_*/L_{3.6} \equiv \Upsilon_{*,3.6}$ ratios to those from Equation~\ref{Eq_MonL_IP13}. We proceed using Equation~\ref{Eq_MonL_IP13} to derive the spheroid and galaxy stellar masses for all. In Table~\ref{Table-data}, we list the spheroid and galaxy stellar-masses, and the black hole masses taken from the compilation in \citet{2020ApJ...903...97S}, unless indicated otherwise. The luminosity distances are also tabulated here. Distances from \citet{2001ApJ...546..681T} have been reduced by $\sim$4 percent due to a 0.083 mag reduction in their distance moduli. This arose from a 0.06 mag reduction after a recalibration by \citep[][their Section 4.6]{2002MNRAS.330..443B}\footnote{We opt not to use the attempted refinement offered by Equation~A1 and Figure~7 in \citet{2010ApJ...724..657B}.} plus a 0.023 mag reduction due to a reduced distance modulus for the Large Magellanic Cloud \citep{2019Natur.567..200P} involved in the initial calibration. The black hole masses depend linearly on the angular distance to the host galaxies, and these masses have been updated here to reflect this. Following \citet[][their Eq.~9]{2019ApJ...876..155S}, the quoted uncertainties on the stellar masses include three uncertainties added in quadrature. These relate to the distance (see Table~\ref{Table-data}), the $M_*/L$ ratio (a 25 percent uncertainty is suggested by \citet{2013MNRAS.430.2715I}), and the apparent magnitude. Here, we use a 0.15 mag uncertainty for the galaxy magnitude, and thus also for the spheroidal component of pure elliptical galaxies. This primarily captures uncertainty in the extrapolation of the light profile to large radii \citep[][their Figure~1]{2005PASA...22..118G} and this value also falls in the middle of the $-$0.11 to +0.18 range reported by \citet[][their Section~4.2.4]{2016ApJS..222...10S}. For those galaxies with two or more components, we assign uncertainties reflecting the complexity of the decomposition and thus the accuracy of the spheroid magnitude. These uncertainties were at elevated levels in \citet{2016ApJS..222...10S} and, in turn, \citet{2019ApJ...876..155S} because they were based on the published range of spheroid magnitudes from decompositions that, in retrospect, were clearly in error due to, for example, missed discs or bars. Having narrowed in on a better suite of components for each galaxy, the typical uncertainty on the spheroid magnitude is reduced. We adopt the following grading schema for the uncertainties on the magnitudes: Grade 0 (0.15 mag: single-component galaxy); Grade 1 (0.2 mag: ES galaxies and those with only minor inner components); Grade 1.5 (0.25 mag: typically a clean bulge+disc fit, or if several arcseconds of inner data were excluded, or if intracluster light (ICL) is present); Grade 2 (0.40 mag: usually a bar+bulge+disc fit); Grade 3 (0.55 mag: typically many components present). The forthcoming regressions were, however, tested and found to be stable (at the 1$\sigma$ level) to a broad range of uncertainties. In Appendix~\ref{Apdx1}, we repeat the forthcoming analysis of Section~\ref{Sec_Results} using an alternative optical-infrared colour-dependent prescription for the mass-to-light ratio, taken from \citet{2022arXiv220202290S}. This additional analysis supports one of our primary conclusions: that violent, disc-destroying, mergers of (red) bulge$+$disc galaxies \citep[e.g.,][]{2006ApJ...636L..81N} produce an offset population of elliptical galaxies in the $M_{\rm bh}$--$M_{\rm *,sph}$ diagram. Neither the initial (bulge) nor the secondary (elliptical galaxy) relations have a near-linear slope in this diagram. \begin{figure} \begin{center} \includegraphics[trim=0.0cm 0cm 0.0cm 0cm, width=1.0\columnwidth, angle=0]{second} \caption{Black hole mass versus spheroid stellar mass (obtained via Equation~\ref{Eq_MonL_IP13}) for different galaxy morphologies (E, ES/S0 and S). As noted in the inset legends, each panel sampled the galaxies differently. The darker shading reveals the 1$\sigma$ uncertainty on each relations' slope and intercept --- as determined by `confband.py' from SciPy \citep{2020SciPy-NMeth} ---, while the lighter shading shows the root mean square (rms) scatter. All quantities are shown in Table~\ref{Table-IP13}. The middle panel reveals that the single relation for ETGs, shown in the upper panel, overlooks a key division between ETGs with and without discs. Similarly, the single relation for disc galaxies, shown in the lower panel, overlooks the division between disc galaxies with and without a spiral pattern and thus the varying abundance of cold gas and star formation. While one may use Table~\ref{Table-data} to identify every galaxy shown here, for some galaxies mentioned in the text, we have added small labels which can be seen by zooming in. These are small to avoid overly detracting from the underlying patterns. } \label{Fig_M_Msph_IP13} \end{center} \end{figure} \begin{figure} \begin{center} \includegraphics[trim=0.0cm 0cm 0.0cm 0cm, width=\columnwidth, angle=0]{Mbh_Msph_ratio_vs_Msph_E_ESS0_S-mod} \caption{$M_{\rm bh}/M_{\rm *,sph}$ mass ratio versus $M_{\rm *,sph}$. We have mapped the lines, and their 1$\sigma$ uncertainty, from the middle panel of Figure~\ref{Fig_M_Msph_IP13} For a given spheroid mass, the mean $M_{\rm bh}/M_{\rm *,sph}$ ratio is different by roughly an order of magnitude for ETGs with and without discs. As shown by the arrows, equal-mass mergers of S0 galaxies (illustrated here with two different bulge-to-total stellar mass ratios, $B/T=1/4$ and $1/2$) can shift systems towards the lower right. For example, an E galaxy built from an equal-mass merger of two S0 galaxies with $B/T=1/2$ will enact both a downward shift of $\sim$0.3 dex ($=\log2$) and a rightward shift of $\sim$0.6 dex ($=\log4$). Given the slope of $0.53\pm0.15$ for the distribution of ES/S0 galaxies in this diagram, equal-mass mergers between S0 galaxies with $B/T=1/2$ (or $=1/4$) will create a new relation with a vertical offset of $\sim$0.6 (or $\sim$0.8) dex. On the other hand, equal-mass mergers of E galaxies with $B/T=1$ will create a remnant shifted to the right by only $\sim$0.3 dex, as shown by the three example arrows.} \label{Fig_msph_rat_IP13} \end{center} \end{figure} \begin{figure} \begin{center} \includegraphics[trim=0.0cm 0cm 0.0cm 0cm, width=\columnwidth, angle=0]{Mbh_BT_ratio} \caption{ Black hole mass versus the logarithm of the spheroid-to-total stellar mass ratio for S, ES/S0 and E galaxies (for which the ratio tends to 1). The symbols and shading have the same meaning as in Figure~\ref{Fig_M_Msph_IP13}. To help understand the transition from the $M_{\rm bh}$--$M_{\rm *,sph}$ diagram (Figure~\ref{Fig_M_Msph_IP13}) to the $M_{\rm bh}$--$M_{\rm *,gal}$ diagram (Figure~\ref{Fig_MMgal_IP13}), we have shown how the `spheroid-to-total' ratio relates to the black hole mass. A linear regression of $M_{\rm *,sph}/M_{\rm *,gal}$ and $M_{\rm bh}$ has been performed here (see Table~\ref{Table-IP13}). Note: Some E galaxies have $B/T$ ratios smaller than some ES galaxies due to additional undigested components and/or nuclear discs. } \label{Fig_BonT_IP13} \end{center} \end{figure} \section{Results}\label{Sec_Results} \subsection{The $M_{\rm bh}$--$M_{\rm *,spheroid}$ and $M_{\rm bh}$--$M_{\rm *,galaxy}$ diagrams}\label{Sec_Diagrams} In Figure~\ref{Fig_M_Msph_IP13}, we show the $M_{\rm bh}$--$M_{\rm *,sph}$ diagram, using Equation~\ref{Eq_MonL_IP13}, for the three morphological types: E, ES/S0 and S. In the upper panel we combine the E, ES and S0 galaxies, which represent the ETGs. In the middle panel, there is no grouping of the different galaxy types, while in the lower panel, we combine the ES, S0 and S galaxies, representing the disc galaxies. The first point we make is that the slope of the $M_{\rm bh}$--$M_{\rm *,sph}$ relation for bulges in either S galaxies or ES/S0 galaxies, and the slope of the $M_{\rm bh}$--$M_{\rm *,sph}$ relation for elliptical galaxies, is neither equal to 1 nor close to it. This is quantified in Table~\ref{Table-IP13} and described further in Section~\ref{Sec-Relations}. The different relations for the bulges and elliptical galaxies can also be seen in the $(M_{\rm bh}/M_{\rm *,sph})$--$M_{\rm *,sph}$ diagram (Figure~\ref{Fig_msph_rat_IP13}). For a given spheroid stellar-mass, Figure~\ref{Fig_msph_rat_IP13} reveals different $(M_{\rm bh}/M_{\rm *,sph})$ ratios for elliptical galaxies and the bulges of disc galaxies. This different ratio has received little attention in the literature and has never been explained. The arrows in this diagram trace the expected movement due to simple, equal-mass, dry mergers of galaxies with some illustrative bulge-to-total stellar mass ratios, $B/T$. The merger of two E galaxies produces a shift to the right, while a merger of two identical S0 galaxies with a typical $B/T=0.25$ \citep[e.g.,][]{2005MNRAS.362.1319L} produces a considerable shift to the lower right. Considering the mean regression lines, the elliptical galaxies appear to be built, on average, by just one major merger. One can also appreciate how brightest cluster galaxies (BCGs), which are typically E galaxies, occupy the right-hand side of the distribution in this diagram. Figure~\ref{Fig_BonT_IP13} shows, for different morphological types, the trend between black hole mass and $B/T$, or more precisely, the spheroid-to-total stellar mass ratio. $B/T$ is not some near-constant value for all S0 galaxies; a range of ratios is known \citep[e.g., ][]{2008MNRAS.388.1708G}. Aside from the exclusions mentioned in Section~\ref{Sec_Sample}, here we exclude the ES/S0 galaxy NGC~4762 given the excessive weight its small $B/T$ ratio has on our sample's regression.\footnote{We do not wish to imply that S0 galaxies with low $B/T$ ratios are in error, only that this $B/T$ data point for NGC~4762 interferes with the current mapping between Figures~\ref{Fig_M_Msph_IP13} and \ref{Fig_MMgal_IP13}.} Figure~\ref{Fig_BonT_IP13} reveals that the S0 galaxies with the lower $B/T$ ratios have smaller black hole masses, as is observed among the S galaxies. These trends aid our understanding of the transition from the $M_{\rm bh}$--$M_{\rm *,sph}$ diagram (Figure~\ref{Fig_M_Msph_IP13}) to the $M_{\rm bh}$--$M_{\rm *,gal}$ diagram (Figure~\ref{Fig_MMgal_IP13}). For a given black hole mass, the smaller $B/T$ ratios in the LTGs imply that there will be a steeper $M_{\rm bh}$--$M_{\rm *,gal}$ relation for LTGs than for ES/S0 galaxies. That is, given the greater disc-to-bulge flux ratios and smaller spheroid masses when moving from Sa to Sc galaxies, the spiral galaxies transition to a steeper $M_{\rm bh}$--$M_{\rm *,gal}$ relation than the early-type disc galaxies (ES/S0). This is seen in Figure~\ref{Fig_MMgal_IP13}, in which the $M_{\rm bh}$--$M_{\rm *,gal}$ relation for the E galaxies is basically\footnote{Nuclear discs and additional, possibly undigested, galaxy components result in a slight difference.} the same as the $M_{\rm bh}$--$M_{\rm *,sph}$ relation for the E galaxies. The $M_{\rm bh}$--$M_{\rm *,gal}$ relation for the E galaxies is, however, offset from the $M_{\rm bh}$--$M_{\rm *,gal}$ relation for the ES/S0 galaxies. The darker shading in this diagram reveals that the relations are not consistent with each other at the 1$\sigma$ level. This reveals that the E and ES/S0 galaxies are not offset from each other in the $M_{\rm bh}$--$M_{\rm *,sph}$ diagram due to the exclusion of the non-spheroid stellar mass. \begin{figure} \begin{center} \includegraphics[trim=0.0cm 0cm 0.0cm 0cm, width=0.99\columnwidth, angle=0]{fifth-fig} \caption{Similar to Figure~\ref{Fig_M_Msph_IP13} but now showing the {\em galaxy's} stellar mass, obtained via Equation~\ref{Eq_MonL_IP13}, rather than the {\em spheroid's} stellar mass. The faint dashed lines are the $M_{\rm bh}$--$M_{\rm *,sph}$ relations from Figure~\ref{Fig_M_Msph_IP13}. The bulgeless galaxy NGC~4395, with a rather blue $B-V$ colour of 0.445, and thus low $M_*/L_{3.6}$ ratio, can be seen in the lower left next to NGC~404. } \label{Fig_MMgal_IP13} \end{center} \end{figure} \begin{figure} \begin{center} \includegraphics[trim=0.0cm 0cm 0.0cm 0cm, width=0.98\columnwidth, angle=0]{Mbh_Mgal_ratio_vs_Mgal-mod} \caption{Similar to Figure~\ref{Fig_msph_rat_IP13} but now showing the $M_{\rm bh}/M_{\rm *,gal}$ mass ratio versus $M_{\rm *,gal}$. The lines from the middle panel of Figure~\ref{Fig_MMgal_IP13} have been mapped here. On average, the low-mass E galaxies have roughly the same $M_{\rm bh}/M_{\rm *,gal}$ ratio as the S0 galaxies which merged to create them, but twice as much stellar mass, in accord with a likely dry equal-mass merger. The high-mass E galaxies, including some brightest cluster galaxies, have four times as much mass as the S0s with a similar $M_{\rm bh}/M_{\rm *,gal}$ ratio, reflective of two such mergers or equally one equal-mass E galaxy merger. } \label{Fig_mgal_rat_IP13} \end{center} \end{figure} In Figure~\ref{Fig_mgal_rat_IP13}, we present the $(M_{\rm bh}/M_{\rm *,gal})$--$M_{\rm *,gal}$ diagram. As noted in \citet{2019ApJ...876..155S}, one can {\em approximate} the ETGs (E/ES/S0) with a single relation in this diagram if, for example, one is pursuing rough predictions for black hole masses in other ETGs. However, there is more detail to it than this, and this detail is one of the keys to understanding the black hole mass scaling diagrams. As with Figure~\ref{Fig_M_Msph_IP13}, one can again see that the addition of the non-spheroid stellar mass, primarily from the disc and bar, does not align the lenticular and elliptical galaxies. This reveals that the offset in the $M_{\rm bh}$--$M_{\rm *,sph}$ diagram, between the bulge component of ES/S0 galaxies and E galaxies, is not an artifact of separating/reducing the light in some ETGs (those with discs) but not others (those without) when we were plotting the spheroid stellar mass. The arrows in Figure~\ref{Fig_mgal_rat_IP13} reveal that the distribution of elliptical galaxies is readily explained if they are built from the dry merger of lenticular galaxies, which is widely thought to be the case, and also the merger of elliptical galaxies.\footnote{While we have used simple equal-mass mergers to illustrate S0-to-E, and E-to-E, transformations, there are other options, such as several minor mergers.} However, what is not well-recognised is the ensuing offset between the E and ES/S0 galaxies in various black hole mass scaling diagrams populated with real data. Here, we have: \begin{itemize} \item built on \citet{2019ApJ...876..155S} which established that there is not a single (fundamental) $M_{\rm bh}$--$M_{\rm *,sph}$ relation for ETGs; \item revealed that mergers have not built a near-linear $M_{\rm bh}$--$M_{\rm *,sph}$ relation, due to the folding-in of the disc/bar stellar mass, reducing the $M_{\rm bh}$--$M_{\rm *,sph}$ ratio; and \item established that a single $M_{\rm bh}$--$M_{\rm *,gal}$ relation is not a fundamental relation for ETGs, with merger-built E galaxies offset from the S0 galaxies. \end{itemize} We will return to these points with additional supportive evidence in Section~\ref{Sec_b2e}. \subsection{The $M_{\rm bh}$--$M_{\rm *,sph}$ and $M_{\rm bh}$--$M_{\rm *,gal}$ relations and ratios}\label{Sec-Relations} We have used hierarchical Bayesian model fitting through the state-of-the-art platform for statistical modelling known as {\sc Stan} \citep{2017JSS....76....1C, Rstan:2016}\footnote{\url{https://mc-stan.org/}}. The statistical model used for our linear regression considered uncertainties in both variables and is aimed at obtaining a symmetric relation between the two variables. A bivariate normal density was used to represent the distribution of latent (`true') $\log M_*$ and $\log M_{\rm bh}$ values that might occur in our sample. As noted in \citet{2019ApJ...873...85D}, ``this is conceptually equivalent to the generative framework sketched by \citet{2010arXiv1008.4686H}, in which the observed data points are imagined to be drawn from a distribution centred around a `line of best fit', except that here we allow Bayesian `shrinkage' by estimating the underlying distribution along the line rather than keeping this as an improper uniform prior.'' Details of the statistical model framework are described in \citet[][their Appendix~A]{2019ApJ...873...85D}. The best-fitting relations are shown in Table~\ref{Table-IP13}, along with the slope, $A$, and intercept, $B$, at the normalisation point. The normalisation point of, for example, the $M_{\rm bh}$--$M_{\rm *,sph}$ relation is $\upsilon(5\times10^{10}\,M_\odot)$. If using equation~\ref{Eq_MonL_IP13} to convert light into stellar mass and thus derive a value of $M_{\rm *,sph}$ for use in the $M_{\rm bh}$--$M_{\rm *,sph}$ equation (to predict $M_{\rm bh}$), one has that $\upsilon=1$. If, however, a different light-to-mass ratio prescription is preferred and used to derive one's estimate of $M_{\rm *,sph}$, then one needs to apply the appropriate value of $\upsilon$, as illustrated in \citet[][see their Figure~4 and Equations~6 to 8]{2019ApJ...876..155S} for colour-dependent light-to-mass ratio prescriptions in different passbands. While we could drop this $\upsilon$ term from our equations, as typically done before \citet[][see their Equations~10 and 11]{2019ApJ...873...85D}, its inclusion serves to remind readers that a specific prescription for $\Upsilon_*$ has been used to derive the equation and that they need to apply a conversion if using an alternate prescription. The root mean square scatter, $\Delta_{\rm rms}$, in the $\log\,M_{\rm bh}$ direction is also tabulated for reference, although it is noted that this is not the quantity that is minimised with a Bayesian regression. We have also applied three additional linear regression codes to our data, and found consistent results with our primary Bayesian analysis. For example, the Bisector regression from the Bivariate Correlated Errors and Intrinsic Scatter ({\sc BCES}) routine \citep{1996ApJ...470..706A} gave the following slopes in the $M_{\rm bh}$--$M_{\rm *,sph}$ diagram for the E, ES/S0, and S galaxies: 1.62$\pm$0.17; 1.49$\pm$0.13; and 2.19$\pm$0.33. Using a symmetrical treatment\footnote{While the modified-{\sc FITEXY} routine performs a non-symmetrical regression of a sample of ($X$, $Y$) data pairs, a symmetrical treatment can be obtained by running the regression twice, the second time with the $Y$ and $X$ variables swapped around with each other. The {\em bisector} of the resulting two regression lines provides an expression which effectively treats the data equally \cite[e.g.,][their Section~3.1.1]{2006ApJ...637...96N,2007ApJ...655...77G}.} of the modified-{\sc FITEXY} routine from \citet{2002ApJ...574..740T} yielded slopes equal to 1.65$\pm$0.12, 1.53$\pm$0.11 and 2.20$\pm$0.26, respectively. Finally, the Bayesian {\sc linmix} code from \citet{2007ApJ...665.1489K} yielded: 1.61$\pm$0.14 (E); 1.52$\pm$0.13 (ES/S0); and 2.14$\pm$0.34 (S). From Table~\ref{Table-IP13}, we have that $M_{\rm bh} \propto M_{\rm *,sph}^{2.25\pm0.39}$ for the bulges of the spiral galaxies. This has 1$\sigma$ uncertainties which overlap with those from the steeper relation reported by \citet{2019ApJ...873...85D}, in which $M_{\rm bh} \propto M_{\rm *,sph}^{2.44\pm0.33}$ for a larger sample of 40 spiral galaxies observed with a range of filters and $M_*/L$ ratios. It appears that the bulges of spiral galaxies define a steeper $M_{\rm bh}$--$M_{\rm *,sph}$ relation than the bulges of S0 galaxies. Three evolutionary pathways for the spiral galaxy bulges are offered in Section~\ref{Sec_Disc_1}. In fair agreement with the relations found here for the ETGs using the hierarchical Bayesian model fitting (see Table~\ref{Table-IP13}), \citet{2019ApJ...876..155S} report a slope of 1.86$\pm$0.20 for 36 ES/S0 galaxies and 1.90$\pm$0.20 for 40 E galaxies. As seen in \citet{2019ApJ...876..155S}, the present sample of E galaxies trace a relation which is roughly parallel to that defined by the bulges of S0 galaxies. Using multicomponent decompositions, \citet{2016ApJ...817...21S} reported a median $M_{\rm bh}/M_{\rm *,sph}$ value of $\sim$0.68 percent for 45 ETGs, which they thought followed a near-linear $M_{\rm bh}$--$M_{\rm *,sph}$ relation. This result was based on the use of an $M_*/L_{3.6}$ ratio of $\sim$0.60 and a \citet{2003PASP..115..763C} IMF. \citet{2010MNRAS.404.2087B} suggest a reduction of just 0.05 dex to the logarithm of $\Upsilon_*$ ($\equiv M_*/L$) to convert from the \citet{2003PASP..115..763C} IMF to the \citet{2002Sci...295...82K} IMF. Therefore, this $M_{\rm bh}/M_{\rm *,sph}$ mass ratio of 0.68 percent increases to 0.76 percent for the \citet{2002Sci...295...82K} IMF. This is comparable to the median $M_{\rm bh}/M_{\rm *,sph}$ ratio for core-S\'ersic galaxies reported in \citet{2013ApJ...764..151G}, which was obtained using $M_*/L_{K}=0.8$ from \citet{2003ApJS..149..289B} and based on a diet-Salpeter IMF. Their reported $K$-band mass-to-light ratio of 0.49 drops by 0.15 dex, or to $M_*/L_{K}=0.57$, when switching to the \citet{2002Sci...295...82K} IMF. Consequently, their $M_{\rm bh}/M_{\rm *,sph}$ ratio of 0.49 percent increases to 0.69 percent once calibrated against the \citet{2002Sci...295...82K} IMF, and is thus in good agreement with the above value of 0.76 percent.\footnote{We thank Peter Behroozi for pointing out this issue in 2017, surrounding clarification of the adopted IMF before comparing $M_{\rm bh}/M_{\rm *,sph}$.} However, as \citet{2019ApJ...876..155S} uncovered, and as can be seen in Figure~\ref{Fig_msph_rat_IP13}, this near-constant $M_{\rm bh}/M_{\rm *,sph}$ mass ratio of $\sim$0.7 percent for ETGs is misleading. Individual ratios, at a fixed $M_{\rm *,sph}$, differ by an order of magnitude depending on whether the system is an S0 or an E galaxy. Furthermore, the ratio can vary by an order of magnitude within either of these two galaxy types. Turning to the {\em galaxy} masses, Figure~\ref{Fig_mgal_rat_IP13} illustrates that while the S galaxies tend to have lower $M_{\rm bh}/M_{\rm *,gal}$ ratios than the ETGs in our sample, due in part to the greater disc-to-bulge ratios in S galaxies, there is more to it than that. On average, for a given $M_{\rm bh}/M_{\rm *,gal}$ ratio, the E galaxies have higher masses than the S0 galaxies, which is expected if S0 galaxies merge to form E galaxies. This observation also expresses itself as a lower $M_{\rm bh}/M_{\rm *,gal}$ ratio in E galaxies than S0 galaxies at a given galaxy mass, modulo the scarcity of low-mass E and high-mass S0 galaxies --- another signature of the dry merger phenomena. In passing we note that it almost goes without saying that applying consistent $\Upsilon_*$ ratios between different studies is vital for avoiding artificial mismatches such as that reported in \citet{2016MNRAS.460.3119S}. Realised some years ago, and detailed in \citep{SahuGrahamHon22}, this mismatch led us to introduce the mass-to-light conversion term, $\upsilon$, in \citet{2019ApJ...873...85D}.\footnote{The lower-case upsilon symbol was introduced to facilitate changes to the mass-to-light ratio, $\Upsilon$, in a similar manner to how $h$ can enact changes to the adopted Hubble-Lema\^itre constant H$_0$.} This was developed further in \citet{2019ApJ...876..155S} and explains the $\upsilon$ term included in Table~\ref{Table-IP13}. \begin{table} \centering \caption{Black hole mass scaling relations}\label{Table-IP13} \begin{tabular}{lccc} \hline Galaxy type & Slope (A) & Intercept (B) & $\Delta_{\rm rms}$ \\ \hline \multicolumn{4}{c}{$\log(M_{\rm bh}/M_\odot) = A\log[M_{\rm *,sph}/\upsilon(5\times10^{10}\,M_\odot)] +B$} \\ E (35) & 1.64$\pm$0.17 & 7.79$\pm$0.17 & 0.38 \\ ES/S0 (32) & 1.53$\pm$0.15 & 8.67$\pm$0.15 & 0.44 \\ S (26) & 2.25$\pm$0.39 & 8.66$\pm$0.28 & 0.57 \\ ES/S0 \& S (58) & 1.84$\pm$0.16 & 8.63$\pm$0.14 & 0.55 \\ E \& ES/S0 (67) & 1.16$\pm$0.07 & 8.30$\pm$0.11 & 0.43 \\ \hline \multicolumn{4}{c}{$\log(M_{\rm bh}/M_\odot) = A\log[M_{\rm *,gal}/\upsilon\,10^{11}\,M_\odot] +B$} \\ E (35) & 1.69$\pm$0.17 & 8.22$\pm$0.15 & 0.38 \\ ES/S0 (32) & 1.93$\pm$0.28 & 8.57$\pm$0.17 & 0.61 \\ S (26) & 3.23$\pm$0.57 & 7.91$\pm$0.18 & 0.60 \\ ES/S0 \& S (58) & 2.38$\pm$0.27 & 8.28$\pm$0.13 & 0.80 \\ E \& ES/S0 (67) & 1.59$\pm$0.11 & 8.37$\pm$0.11 & 0.49 \\ \hline \multicolumn{4}{c}{$\log(M_{\rm bh}/M_\odot) = A\{\log[M_{\rm *,sph}/M_{\rm *,gal}] -(-0.5) \} + B$} \\ ES/S0 (31) & 4.00$\pm$0.58 & 7.71$\pm$0.16 & 0.64 \\ S (26) & 3.23$\pm$0.63 & 7.99$\pm$0.21 & 0.78 \\ \hline \multicolumn{4}{c}{$\log(M_{\rm sph}/ \upsilon M_\odot) = A\log[R_{\rm e,sph,eq}/\rm kpc] +B$} \\ All (93) & 1.14$\pm$0.04 & 10.48$\pm$0.08 & 0.29 \\ \hline \multicolumn{4}{c}{$\log(M_{\rm bh}/M_\odot) = A\log[R_{\rm e,sph,eq}/\rm kpc] +B$} \\ E (35) & 2.38$\pm$0.33 & 6.88$\pm$0.32 & 0.58 \\ ES/S0 (32) & 1.98$\pm$0.24 & 8.52$\pm$0.16 & 0.54 \\ S (26) & 2.40$\pm$0.40 & 8.02$\pm$0.20 & 0.65 \\ \hline \end{tabular} The slope and intercept of the relations shown in Figures~\ref{Fig_M_Msph_IP13}, \ref{Fig_BonT_IP13} \ref{Fig_MMgal_IP13}, \ref{Fig_R_Msph_IP13} and \ref{Fig_evolve_IP13} have been obtained using a Bayesian analysis that treats the data symmetrically. The root mean square (rms) scatter reported here, $\Delta_{\rm rms}$, is the vertical scatter about each relation. The spheroid and galaxy stellar masses have been obtained using the $M_*/L$ prescription given in Equation~\ref{Eq_MonL_IP13}. The $\upsilon$ term is mentioned towards the end of Section~\ref{Sec-Relations} is equal to 1 if one uses stellar masses consistent with those obtained via Equation~\ref{Eq_MonL_IP13}. \end{table} Here, we find $M_{\rm bh}/M_{\rm *,sph} = 0.0018$ (0.18 percent) for elliptical galaxies with $M_{\rm *,sph}=10^{11}$ M$_\odot$, and 1.7 percent for bulges of the same stellar mass. The order of magnitude difference between these morphological types can be seen in Figure~\ref{Fig_msph_rat_IP13}. This difference appears to widen slightly when moving to lower spheroid masses. Furthermore, one can see how dry mergers of S0 galaxies, building E galaxies, can explain this observation. \subsection{From bulges to elliptical galaxies}\label{Sec_b2e} Much of the accretion-driven growth of black holes is known to occur in regular star-forming disc galaxies \citep[e.g.,][]{2009ApJ...691..705G, 2011ApJ...726...57C}. That is, the AGNs tend to reside in normal, often isolated, spiral galaxies \citep[e.g.,][]{2005ApJ...627L..97G, 2012ApJ...744..148K}. AGNs are not particularly prevalent during or after major mergers. Given that elliptical galaxies have little to no star formation, it is apparent that the gaseous processes driving growth in the $M_{\rm bh}$--$M_{\rm *,sph}$ diagram occur in bulges. As noted by multiple studies, the black hole accretion rate relative to the star formation rate is such that it is not expected to establish a linear $M_{\rm bh}$--$M_{\rm *,gal}$ relation but instead a steeper relation \citep[e.g.,][]{2012ApJ...755..146S, 2013ApJ...765L..33L, 2014A&A...566A..53D, 2019ApJ...885L..36D}. This bodes well for the steeper-than-linear trend seen for spiral galaxies in the $M_{\rm bh}$--$M_{\rm *,gal}$ diagram. Should some of the star-formation be occurring in bulges, then the higher (black hole accretion rates)-to-(star-formation rates) in spiral galaxies may mesh well with the trend seen in the $M_{\rm bh}$--$M_{\rm *,sph}$ diagram. By considering the sizes of the spheroids, we can build a more informed scenario for what we are witnessing in the $M_{\rm bh}$--$M_{\rm *,sph}$ and $M_{\rm bh}$--$M_{\rm *,gal}$ diagrams. We will see how dry mergers can account for the steeper-than-linear $M_{\rm bh}$--$M_{\rm *,sph}$ relation observed for the E galaxies. \begin{figure} \begin{center} \includegraphics[trim=0.0cm 0cm 0.0cm 0cm, width=\columnwidth, angle=0]{IP13_Msph_Re-mod} \caption{Geometric-mean axis, aka `equivalent (circular) axis', effective half-light radii, $R_{\rm e,sph,eq}$, of the spheroid versus the stellar mass of the spheroid, $M_{\rm *,sph}$. Modification of \citet[][their Figure~9]{2020ApJ...903...97S} using the stellar mass-to-light prescriptions in Equation~\ref{Eq_MonL_IP13} and only those galaxies with 3.6~$\mu$m data. The arrows show the apparent movement caused by creating an E galaxy from an equal-mass dry merger between two S0 galaxies with bulge-to-total ratios equal to one-quarter and one-half and between two elliptical galaxies with $B/T=1$. The length of the horizontal arrows are based on maintaining the observed relation $M_{\rm *,sph} \propto R_{\rm e,sph,eq}^{1.14\pm0.04}$. The E galaxies with the larger radii at $\log(M_{\rm *,sph}/M_\odot) \approx$ 10.25 and 11.05 are NGC~3377 and NGC~4697, respectively.} \label{Fig_R_Msph_IP13} \end{center} \end{figure} \begin{figure*} \begin{center} \includegraphics[trim=0.0cm 0cm 0.0cm 0cm, width=1.0\textwidth, angle=0]{Mbh-Re-Msph-IP13-mod} \caption{ Left-hand panel: $M_{\rm bh}$ versus $R_{\rm e,sph,eq}$. Adaption of Figure~13 from \citet{2020ApJ...903...97S}, using only those galaxies with 3.6~$\mu$m data and updated vales reported in Table~\ref{Table-data}. Right-hand panel: Evolution in the $M_{\rm bh}$--$M_{\rm *,sph}$ diagram arising from dry equal-mass mergers of galaxies with bulge-to-total ($B/T$) ratios of 1.0, 0.5 and 0.25 to produce an E galaxy. The E galaxy sequence is seen to follow a steeper-than-linear relation rather than the linear relation expected from E+E galaxy mergers \citep{2007ApJ...671.1098P}. The somewhat discrepant (low $M_{\rm *,sph}$ or high $M_{\rm bh}$) spiral and elliptical galaxy with $\log(M_{\rm bh}/M_\odot) \approx 7.9$ dex are NGC~1300 and NGC~3377, respectively.} \label{Fig_evolve_IP13} \end{center} \end{figure*} In Figure~\ref{Fig_R_Msph_IP13}, we show the effective half-light size of the spheroids, $R_{\rm e,sph}$, versus their stellar mass, $M_{\rm *,sph}$. These radii are given in Table~\ref{Table-data}, along with the reference showing the modelled light profile from which the radii were measured. We used the geometric-mean axis, aka the `equivalent axis', $r=\sqrt{ab}$, along which the size of the radii are equivalent to a circularised version of the quasi-elliptical isophotes. There is no discontinuity in the $R_{\rm e,sph}$--$M_{\rm *,sph}$ diagram between the different types of spheroids. This continuity holds whether the spheroids coexist with a disc that either does or does not contain a spiral pattern, or whether they exist on their own with no disc, i.e., are an elliptical galaxy. The $R_{\rm e,sph}$--$M_{\rm *,sph}$, or equally $M_{\rm *,sph}$--$R_{\rm e,sph}$, relation is seen in Figure~\ref{Fig_R_Msph_IP13} to have a slope close to unity. Curiously, there is little evidence for any broad curvature in the distribution of $R_{\rm e,sph}$ and $M_{\rm *,sph}$. This differs from what is seen in the $M_{\rm *,gal}$--$R_{\rm e,gal}$ relation for ETGs \citep[][their Figure~1b]{2006AJ....132.2711G} due to the presence and then dominance of discs as one moves to lower masses. Not surprisingly, this near-linear slope matches that seen at the bright end of the $M_{\rm *,gal}$--$R_{\rm e,gal}$ relation for ETGs, which is dominated by E galaxies \citep[e.g.,][their Figure~9]{2008MNRAS.388.1708G, 2018MNRAS.477.5327K, 2019ApJ...886...80D, 2020ApJ...903...97S}, i.e., systems without discs. The simulations from \citet{2009ApJ...703.1531N}, involving elliptical galaxies undergoing minor and major dry merger events, build a near-linear $M_{\rm *,gal}$--$R_{\rm e,gal}$ relation \citep{2009ApJ...706L..86N}. This relation will be explored further in \citet{HGS2022} with a sample twice that used here and having multicomponent decompositions and a consistent set of $\Upsilon_*$ ratios. In the left-hand panel of Figure~\ref{Fig_evolve_IP13}, we see the black hole masses versus the half-light radii of the host spheroids, as measured from the geometric-mean axis. Given the strong relation between the sizes and the masses of the {\em spheroids} seen in Figure~\ref{Fig_R_Msph_IP13}, it is not too surprising that the structure in the $M_{\rm bh}$--$R_{\rm e,sph}$ diagram (Figure~\ref{Fig_evolve_IP13}) shows a similarity to that seen in the $M_{\rm bh}$--$M_{\rm *,sph}$ diagram (Figure~\ref{Fig_M_Msph_IP13}). Using the $M_{\rm *,sph}$--$R_{\rm e,sph}$ diagram (Figure~\ref{Fig_R_Msph_IP13}), one can map the expected shift in the $M_{\rm bh}$--$M_{\rm *,sph}$ diagram for equal-mass mergers of S0 galaxies that produce an E galaxy. This is shown in the right-hand panel of Figure~\ref{Fig_evolve_IP13}, and can be understood in terms of the galaxies effectively folding in their disc stars to make the new E galaxy and thereby lowering the $M_{\rm bh}/M_{\rm *,sph}$ ratio, as seen in Figure~\ref{Fig_msph_rat_IP13}. This scenario also readily explains the offset between bulges and E galaxies seen in Figure~\ref{Fig_evolve_IP13}. For example, a merger of two equal-mass S0 galaxies with B/T=0.25 \citep[e.g., ][]{2005MNRAS.362.1319L, 2008MNRAS.388.1708G} will double $M_{\rm bh}$ and increase the spheroid (now elliptical galaxy) mass 8-fold once the disc light is incorporated. In the $M_{\rm bh}$--$M_{\rm *,sph}$ diagram, this merger moves a high stellar mass S0 up by 0.3 dex and across by 0.9 dex, placing it on the sequence of elliptical galaxies. From the relation $M_{\rm *,sph} \propto R_{\rm e,sph,eq}^{1.14\pm0.04}$ (based on the Spitzer sample used here), we have that a 0.9 dex increase in $M_{\rm *,sph}$ is associated with a 0.79 dex increase in $R_{\rm e,sph,eq}$. Such an increase from a major merger event is plotted in Figures~\ref{Fig_R_Msph_IP13} and \ref{Fig_evolve_IP13}. Figure~\ref{Fig_evolve_IP13} provides a previously unstated measure-of-sorts of the average number of major mergers the E galaxies in our sample have experienced. There is evidence here, and in Figure~\ref{Fig_mgal_rat_IP13}, that BCGs have experienced a greater number of mergers, and this will be explored in more detail in a subsequent paper. \section{Discussion}\label{Sec_Disc_1} \subsection{Overmassive and undermassive black holes}\label{Sec_Disc_2} Early observational bias favouring the detection of systems with big black holes led to samples dominated by elliptical galaxies and lenticular galaxies with massive bulges. This sample selection produced an apparent near-linear $M_{\rm bh}$--$M_{\rm *,sph}$ relation, which steepened as lower-mass black holes were gradually included. Figure~\ref{Fig_schematic} reveals how this near-linear `red sequence' in the $M_{\rm bh}$--$M_{\rm *,sph}$ diagram arises by sampling both massive bulges and elliptical galaxies. For many such elliptical galaxies, their spheroid mass may be dominated by the disc masses of their progenitor galaxies. This explains the approximately order of magnitude lower $M_{\rm bh}/M_{\rm *,sph}$ ratio in elliptical galaxies when compared to bulges of the same `spheroid' mass (Figure~\ref{Fig_msph_rat_IP13}). As noted earlier, this is not because the galaxies' disc masses are excluded from the $M_{\rm bh}$--$M_{\rm *,sph}$ diagram; the $M_{\rm bh}/M_{\rm *,gal}$ ratio is not equal among ES/S0 and E galaxies at a given galaxy stellar-mass (Figure~\ref{Fig_mgal_rat_IP13}). The notion of a `red sequence' representing the underlying fundamental connection between black holes and `classical bulges', i.e., bulges built by mergers, introduces problems that disappear when considering the $M_{\rm bh}$--$M_{\rm *,sph}$ diagram in terms of a bulge sequence and an offset merger-built population of elliptical galaxies. Most obvious is that the E galaxies do not follow the (near-linear) red sequence but define a steeper non-linear relation (see Table~\ref{Table-IP13}). In addition, the low- and high-mass bulges appear as outliers from the near-linear `red sequence', invoking a misleading perception as to the need for separate formation physics. It had led to the notion that massive bulges and relic galaxies are a disconnected population with overmassive black holes relative to galaxies on the near-linear relation (see Figure~\ref{Fig_schematic}). They are, however, not overmassive relative to the bulge sequence. Furthermore, while some black holes in BCGs appear overmassive relative to the `red sequence', they are not overmassive relative to the elliptical galaxy $M_{\rm bh}$--$M_{\rm *,sph}$ sequence. By appreciating the role of mergers, we can understand how the morphology-dependent relationships arose in the $M_{\rm bh}$--$M_{\rm *,sph}$ and $M_{\rm bh}$--$M_{\rm *,gal}$ diagrams. The near-linear red-sequence also resulted in claims that low-mass bulges were yet another disconnected population with undermassive black holes relative to galaxies on the red-sequence. However, they are not undermassive relative to the bulge sequence. We again note that while our sample of bulges does contain members which reside below the `red sequence', they are not the (peanut shell)-shaped structures associated with unstable bars, nor are they nuclear or inner discs which we model with separate components. \begin{figure} \begin{center} \includegraphics[trim=0.0cm 0cm 0.0cm 0cm, width=1.0\columnwidth, angle=0]{BH_schematic_1} \caption{Schematic of the $M_{\rm bh}$--$M_{\rm *,sph}$ diagram for ETGs. The steep band on the left shows a relation for bulges (in S0 and ES galaxies). The steep band on the right shows the relation for E galaxies. The shallower relation, shown in grey, tracks an apparent `red sequence' obtained by sampling some massive bulges and some elliptical galaxies. Past claims for apparently overmassive and undermassive black holes, relative to this near-linear `red sequence', can be understood in terms of the host bulge belonging to the (steeper) bulge sequence. } \label{Fig_schematic} \end{center} \end{figure} \subsubsection{Relic red nuggets at the top of the bulge sequence}\label{Sec_Disc_3} The gaseous processes that gave rise to some bulges may have occurred long ago. Indeed, many local bulges could be the descendants of the `red nuggets' observed at $z \approx 2.5 \pm 1$ \citep{2005ApJ...626..680D, 2011ApJ...739L..44D} and potentially now cloaked in a large-scale disc \citep[][and references therein]{2015ApJ...804...32G, 2016MNRAS.457.1916D, 2022MNRAS.514.3410H}. If any of the high-$z$ red nuggets did not acquire a disc by today---which may be likely if they started life in a proto-(galaxy cluster), given (i) the propensity for ram-pressure stripping of cold gas by hot gas, and (ii) the reduction in galaxy mergers due to high fly-by speeds---, then they will remain a compact massive galaxy today \citep[e.g.,][]{1966ApJ...143..192Z, 1968cgcg.bookR....Z, 1971cscg.book.....Z, 2010MNRAS.408L..21S, 2010ApJ...712..226V, 2013ApJ...777..125P, 2014ApJ...780L..20T}. Such local `compact galaxies' are also referred to as `relic galaxies' if their stars are old \citep[e.g.,][]{2017MNRAS.467.1929F}. In separating spheroid types, the `relic galaxies' are seen not to be associated with the merger-built E galaxies in the $M_{\rm bh}$--$M_{\rm *,sph}$ diagram but rather sit on the bulge sequence, with the most massive red relics located at the top. Consequently, Figure~\ref{Fig_schematic} reveals how massive relic galaxies appear to have overmassive black holes relative to the near-linear `red sequence' but are consistent with the bulge sequence for ETGs. NGC~1332 is not an elliptical galaxy but a relic ES galaxy which has not acquired/built a large-scale disc. The dominant spheroidal component in NGC~1332 has $\log(M_{\rm *,sph}/M_{\odot}) = 11.15\pm0.15$ dex and a geometric mean radius $R_{\rm e,sph} \approx 1.9$~kpc. It has the second highest black hole mass of the ES+S0 galaxies in our sample, and can be seen to reside at the top of the bulge sequence in Figure~\ref{Fig_evolve_IP13}. We have also labelled the ES galaxy NGC~6861 in Figure~\ref{Fig_evolve_IP13}, which has $\log(M_{\rm *,sph}/M_{\odot}) = 11.07\pm0.19$ and $R_{\rm e,sph} \approx 2.6$~kpc. If a high-$z$ `red nugget' acquired a disc over time, then today the `red nugget' would be the compact massive spheroid of a disc galaxy. NGC~5252, for example, likely has such a relic bulge \citep{2019ApJ...876..155S}; and also a relic quasar \citep{2005A&A...431..465C}. Given the old ages of discs in massive lenticular galaxies, the bulk of their stars formed long ago, no doubt acquired through direct accretion and mergers but also possibly via star-formation in their gas discs at cosmic noon \citep[e.g., ][]{2020MNRAS.497.3273F}. \subsubsection{Merger-built Brightest Cluster Galaxies} The creation of BCGs via (multiple) mergers produces the largest elliptical galaxies, found at the centres of galaxy groups and clusters. The steeper-than-linear $M_{\rm bh} \propto M_{\rm *,sph}^{1.72}$ relation seen in \citet[][their Figure~3]{2018ApJ...852..131B} for BCGs is explained here as a combination of mergers folding in the disc mass and a steep origin relation for the bulges of the pre-merged progenitor galaxies in the $M_{\rm bh}$--$M_{\rm *,sph}$ diagram. That is, the elliptical galaxies, which include the BCGs\footnote{Our galaxy stellar-masses are derived by excluding the surrounding ICL light, either because it was fit as a component during the galaxy decomposition and excluded here, or because the images and light profiles were not deep enough for the ICL to have been an issue.}, should not be thought of as a departure from a near-linear $M_{\rm bh}$--$M_{\rm *,sph}$ relation. Instead, they represent a shift to a somewhat parallel relation to that defined by the bulges of ETGs. Of course, when E$+$E dry mergers build new E galaxies, the evolutionary path in the $M_{\rm bh}$--$M_{\rm *,sph}$ diagram will be along a vector with a slope of 1. One might imagine that in the Universe's distant future, one would start to see a linear relation for BCGs emerge from the top end of the current relation for the E galaxies. This is, however, something which we will leave for the simulators. \subsection{The stripped S0 galaxy M32}\label{Sec_Disc_4} In Figure~\ref{Fig_R_Msph_IP13}, M32 appears to the left of the $M_{\rm bh}$--$M_{\rm *,sph}$ relation defined by bulges in ETGs. However, it resides within the 1$\sigma$ scatter about this relation. The slight preference to the left may reflect that the bulge, along with the disc, in M32 has been eroded by its massive neighbour, M31. This process can reduce the bulge mass and inflate its half-light radius. `Compact elliptical' (cE) galaxies like M32, which have lost the gravitational tug-of-war to retain ownership of their stars \citep{Roche:1850, 1972ApJ...178..623T}, stand out in the galaxy colour-magnitude diagram due to their low luminosity for their colour \citep[e.g.,][their Figure~11]{2019MNRAS.484..794G}. For M32, the $V-I$ (Vega) colour of $\sim$1.2--1.4 mag \citep{1998AJ....116.2263L} implies $M_*/L_I \approx 3 M_\odot/L_\odot$ \citep{2022arXiv220202290S}, or $\sim$2.4 after converting to a \citet{2002Sci...295...82K} IMF. Coupled with 0.09 mag of Galactic extinction, the absolute magnitude for the spheroidal component of M32, $M_I = -17.0$ mag \citep[Vega: ][]{2009MNRAS.397.2148G}, corresponds to $M_* \approx 0.8\times10^9\,M_\odot$. Performing a multicomponent fit to M32's major-axis light profile, \citet{2009MNRAS.397.2148G} measured an effective half light-radius of 26$\arcsec$.3 for the bulge component. For an ellipticity of 0.3 at this radius, this translates to an equivalent-axis $R_{\rm e,sph,eq}=22\arcsec$. Using a scale of 4~pc per 1$\arcsec$, this angular size is equal to a physical size of 88~pc, as shown in Figure~\ref{Fig_evolve_IP13}. We add the dwarf E galaxy Pox~52 (93 Mpc distant), with $M_{\rm bh} = (3.2\pm1)\times10^5$ M$_\odot$ and $M_{\rm *,sph}=1.2\times10^9$ M$_\odot$. We use $R_{\rm e,eq}=436$~pc \citep{2008ApJ...686..892T}, based on a minor-to-major axis ratio $b/a = 0.79$ \citep{2004ApJ...607...90B} and $R_{\rm e,maj}=490$~pc \citep{2008ApJ...686..892T}. Pox~52 follows both the $M_{\rm bh}$--$M_{\rm *,sph}$ and $M_{\rm bh}$--$R_{\rm e,sph}$ relations well. There is another spheroid in our sample, albeit not used in our regression analyses, with a smaller mass and size than that of M32. The dwarf S0 galaxy NGC~404 can be seen in Figure~\ref{Fig_evolve_IP13} to follow the S0 galaxy sequence in the $M_{\rm bh}$--$R_{\rm e,sph}$ diagram but reside to the left of the S0 galaxy sequence in the $M_{\rm bh}$--$M_{\rm *,sph}$ diagram. This LINER galaxy has a bright ($r \approx 2\arcsec$) nuclear spiral pattern \citep[compare CG~611: ][]{2017ApJ...840...68G} and is encircled by a much larger H\,{\footnotesize I} gas disc with knotty, irregular tendrils of UV hotspots and H\,{\footnotesize II} regions \citep{2010ApJ...716...71W, 2013ApJ...772L..23B}. However, this galaxy is excluded from the fitting process because its location at the lower extremum of our data might excessively torque the fitted relation. This becomes problematic using such a datum if its measurements are in error or if the scaling relation does not extend linearly to such low black hole masses. \citet{2017ApJ...836..237N} reported a 3$\sigma$ upper limit to the black hole mass of $1.5\times10^4$ M$_\odot$, which was recently revised to $M_{\rm bh} = 5.5^{+4.1}_{-3.8}\times10^5$ M$_\odot$ \citep{2020MNRAS.496.4061D}. In passing, we note how this discrepancy highlights the affect of systematic errors not captured by the small formal/random errors typically reported for most black hole mass measurements. We also attach a 0.5 dex uncertainty to our spheroid mass, which may be three times less massive than our adopted value from \citet{2019ApJ...876..155S} if this galaxy has an anti-truncated disc \citep{Graham:Sahu:22}, resulting in a steeper inner-disc component at the expense of the bulge. LEDA~87000 is a galaxy that likely harbours a central intermediate-mass black hole \citep{2015ApJ...809L..14B}. Although \citet{2016ApJ...818..172G} observed it to follow the near-quadratic $M_{\rm bh}$--$M_{\rm *,sph}$ relation followed by LTGs, inspection of subsequent Hubble Space Telescope images reveals that the previously poorly-resolved `barge' component\footnote{`Barge' is an amalgamation of Bar$+$Bulge.} --- as seen in ground-based images --- was all bar and no bulge \citep{2017ApJ...850..196B}. This represents something of a growing trend in which the closer one looks, the more `bulges' --- when simply defined as the excess of light above the inward extrapolation of an outer exponential disc --- retreat by giving up ground to bars or other features \citep[e.g.,][]{2003ApJ...582L..79B, 2003ApJ...597..929E, 2005MNRAS.362.1319L, 2022MNRAS.514.3410H}. \subsection{The primary relation} \label{Sec_Disc_5} The larger, merger-built elliptical galaxies are seen to define a secondary, or at least subsequent, relation in the $M_{\rm bh}$--$M_{\rm *,sph}$ diagram. In terms of Darwinian evolution on a galaxy scale, they can be thought of as coming into existence via punctuated equilibrium rather than gradualism. Major dry mergers between S0 galaxies, in which the S0 galaxies effectively fold in all their disc stellar mass to create an elliptical galaxy, are accompanied by a substantial oversized jump in the stellar mass (relative to the jump in the black hole mass) and a large jump in the half-light size of the new spheroid, i.e., the elliptical galaxy. Such evolution explains the two prominent relations observed in the $M_{\rm bh}$--$M_{\rm *,sph}$ diagram and the $M_{\rm bh}$--$R_{\rm e,sph,eq}$ diagram for ETGs (Figure~\ref{Fig_evolve_IP13}). Broadly speaking, some bulges may have arisen from a kind of rapid monolithic collapse, or at least result from an early-formation process that created the observed high-$z$ `red nuggets', while most elliptical galaxies likely formed from a binary merger or hierarchical merging (in the case of the BCGs) over the age of the Universe. As such, a meaningful cosmological probe into the evolution of the galaxy/black hole scaling relations needs to be mindful of the galaxy morphology. For example, a sample of elliptical galaxies at $z=1$ can not be directly compared with a sample of bulges at $z=0$; to do so would be comparing apples and oranges. To summarise, the notion of a single near-linear $M_{\rm bh}$--$M_{\rm *,sph}$ relation is inadequate and seems to offer misdirection in understanding galaxies and black holes. The averaging of black hole and galaxy masses through mergers has established neither the expected nor an observed near-linear $M_{\rm bh}$--$M_{\rm *,sph}$ relation. While we have presented the most accurate $M_{\rm bh}$--$M_{\rm *,sph}$ diagram to date, and interpreted the broad brush stroke near--parallel relations shown in \citet{2019ApJ...876..155S}, there is further information to be gleaned from this diagram. Mergers, both wet and dry, which do not fold in all of the disc's stellar mass will be addressed in \citet{Graham:Sahu:22}, where we develop something of a phylogenetic tree diagram within the bivariate space of $M_{\rm bh}$ versus $M_{\rm *,sph}$. One should expect to observe morphology-dependent substructure in other black hole scaling diagrams. For example, as previously noted, the broad red/blue sequence for ETGs/LTGs in the $M_{\rm bh}$--$M_{\rm *,gal}$ diagram has been observed in the $M_{\rm bh}$--colour diagram \citep{2020ApJ...898...83D}. This broad division may also appear in the $M_{\rm bh}$--metallicity, $Z$, diagram \citep{2003ApJ...596...72W, 2008MNRAS.390..814K}. Depending how the number of globular clusters, $N_{\rm GC}$, traces a galaxy's stellar mass \citep[][and references therein]{2014A&A...565L...6M}, one may also expect the ETGs and LTGs to follow different trends in the $M_{\rm bh}$--$N_{\rm GC}$ diagram \citep{2009MNRAS.392L...1S, 2010ApJ...720..516B}. The number of red and blue globular clusters around each galaxy may yield yet further subdivisions \citep[see][]{1998AJ....116.2841K, 2002A&A...395..761K}, as may their kinematics \citep[e.g.,][]{2012MNRAS.426L..51S, 2013MNRAS.433..235P}. With our new understanding of the relevance of galaxy morphology and galaxy formation history, the role of mergers, and refined insight into what may be considered the primary relations versus their modified/evolved form, one is also better placed to tackle the question of whether or not a two-dimensional plane within a three-dimensional space may provide an improved description over bivariate linear relations. For example, does a third axis, in addition to $M_{\rm bh}$ and $M_{\rm sph}$, uncover a distribution on a more fundamental plane? Our analysis, considering additional parameters obtained from physically-motivated multicomponent decomposition, such as the spheroid S\'ersic index and stellar density \citep[e.g.,][]{2007ApJ...655...77G, 2016ApJ...818...47S, 2022ApJ...927...67S}, along with spheroid size, mass, and velocity dispersion, will be presented in a forthcoming paper. Here, we restrict ourselves to briefly explaining why the combination of $M_{\rm bh}$, $M_{\rm *,sph}$, and $R_{\rm e,sph,eq}$ (or equally\footnote{The term $\langle I \rangle _{\rm e}$ is the mean intensity within $R_{\rm e}$. Given that $M_{\rm *,sph} \propto R_{\rm e,sph,eq}^2 \langle I \rangle _{\rm e}$ (by definition), modulo the (stellar mass)-to-light ratio, and given $M_{\rm *,sph} \propto R_{\rm e,sph,eq}^{1.14}$ (Figure~\ref{Fig_R_Msph_IP13}), we have that $\langle I \rangle _{\rm e} \propto R_{\rm e,sph,eq}^{-0.86}$ and thus $M_{\rm *,sph} \propto \langle I \rangle _{\rm e}^{-1.33}$.} $\langle I \rangle _{\rm e}$) may {\em not} produce a useful plane about which the scatter in the $\log(M_{\rm bh})$ direction is less than that seen about the $M_{\rm bh}$--$M_{\rm *,sph}$ relation. For the following thought experiment, we can consider two parallel relations in the $M_{\rm bh}$--$M_{\rm *,sph}$ diagram, one for S0 galaxy bulges and the other for an offset population of merger-built E galaxies. We can use the knowledge that the (logarithm of the) half-light spheroid radius scales with the (logarithm of the) spheroid stellar mass (Figure~\ref{Fig_R_Msph_IP13}). One way to think of the problem is that we wish to introduce an $R_{\rm e,sph,eq}$ term to effectively shift the E galaxies to the left in the $M_{\rm bh}$--$M_{\rm *,sph}$ diagram, to make them overlap with the bulges and thereby reduce the scatter seen in this diagram (see Figure~\ref{Fig_evolve_IP13}). However, we need to bear in mind that this procedure will also shift the bulges to the left, given that we are assuming no knowledge of morphology and just using the parameters $M_{\rm bh}$, $M_{\rm *,sph}$, and $R_{\rm e,sph,eq}$. It turns out that to achieve overlap of the elliptical and bulge samples, the necessary subtraction of a $\log\,R_{\rm e}$ term from the $\log\,M_{\rm *,sph}$ values in the $M_{\rm bh}$--$M_{\rm *,sph}$ diagram results in a near-vertical distribution of points with a near-infinite slope. Remember, $R_{\rm e,sph,eq}$ scales almost linearly with $M_{\rm *,sph}$. This shall be shown in a forthcoming paper but we felt it was of sufficient interest to provide some initial insight here. \subsection{Is there a role for AGN feedback in shaping the turnover and (high mass)-end of the galaxy mass function?}\label{Sec_Disc_6} The observational results herein represent a considerable departure from the connection galaxies are often claimed or thought to have with their central black hole. More accurate spheroid masses --- particularly from a greater awareness that many ETGs are S0s rather than Es --- have revealed how the coevolution of {\em bulges} and supermassive black holes have built a super-linear\footnote{We use the term `super-linear' to denote a power-law with a slope steeper than 1 but not as high as 2.} or near-quadratic\footnote{We use the term `near-quadratic' to describe a power-law with a slope close to a value of 2.} $M_{\rm bh}$--$M_{\rm *,sph}$ relation. \citet{2012ApJ...746..113G} and \citet{2013ApJ...764..151G} highlighted this steeper slope and discussed how dry mergers might be producing an offshoot of core-S\'ersic galaxies, creating (what was thought to be) a near-linear slope at high black hole masses in the $M_{\rm bh}$--$M_{\rm *,sph}$ diagram. However, this idea did not account for the incoming disc mass during some mergers, or for the more recent observation that merger-built elliptical galaxies (with and without depleted cores) do not follow a near-linear $M_{\rm bh}$--$M_{\rm *,sph}$ relation. Obviously, AGN feedback has thus also not produced a near-linear $M_{\rm bh}$--$M_{\rm *,sph}$ relation for the elliptical galaxies. Moreover, the location of the elliptical galaxies in the $M_{\rm bh}$--$M_{\rm *,sph}$ diagram would appear to not be due to AGN feedback but rather major mergers in which the angular momenta of the progenitor galaxy's discs have largely cancelled. This observation is apparent from the $M_{\rm bh}$--$M_{\rm *,sph}$ diagram, the $M_{\rm bh}/M_{\rm *,sph}$ ratios (Figure~\ref{Fig_msph_rat_IP13}), the $M_{\rm bh}/M_{\rm *,gal}$ ratios (Figure~\ref{Fig_mgal_rat_IP13}) and the $M_{\rm bh}$--$R_{\rm e,sph}$ diagram (Figure~\ref{Fig_evolve_IP13}). While `quasar mode' AGN feedback (discussed in Section~\ref{Sec_imp}) might contribute to a link between black hole mass and {\em bulge} mass for some lower mass systems, it is not yet well established how much it may regulate the gas and star formation in the discs of galaxies \citep[e.g.,][]{2014MNRAS.441.1615G, 2018ApJ...869..113D}. Given that most of the stellar mass in disc galaxies resides in their discs, with $B/T < 0.5$ for most S0 and S galaxies \citep{2008MNRAS.388.1708G}, the role of AGN in shaping the {\em galaxy} stellar-mass function appears limited. Given that mergers, rather than AGN feedback, have likely built the elliptical galaxies which dominate the high-mass end of the galaxy mass function \citep[e.g.][]{2022MNRAS.513..439D}, the scope for AGN feedback driving and shaping coevolution in high-mass galaxies appears quenched \citep[e.g.,][]{2003ApJ...599...38B, 2013MNRAS.436.1750R, 2022MNRAS.511..506C}. \begin{figure} \begin{center} \includegraphics[trim=0.0cm 0cm 0.0cm 0cm, width=1.0\columnwidth, angle=0]{MF_schematic} \caption{Schematic of the galaxy stellar-mass function. Here we speculate, with reasoning in Section~\ref{Sec_Disc_6}, as to the nature of the galaxy stellar-mass function. While AGN feedback might regulate bulge growth and help to establish the $M_{\rm bh}$--$M_{\rm *,sph}$ relation in some disc galaxies, it may be the hot gas halo which curtails star formation in discs and keeps it truncated in elliptical galaxies. Mergers, rather than AGN feedback, appear to have established the elliptical galaxy $M_{\rm bh}$--$M_{\rm *,sph}$ relation (Figures~\ref{Fig_evolve_IP13} and \ref{Fig_schematic}). } \label{Fig_schematic_MF} \end{center} \end{figure} Of course, a galaxy does not need to blow out its gas --- via, say, supernovae or an AGN --- in order to cause a cessation of star formation. A galaxy could instead prevent the cooling of gas which might form stars \citep{2003ApJ...599...38B}. Creating a hot gas halo in/around massive pressure-supported spheroids may have this effect \citep[e.g.,][see our Figure~\ref{Fig_schematic_MF}]{2020MNRAS.491.1311M}. While star formation and stellar winds might not generate the escape speeds required to clear gas from a massive galaxy (and its dark matter halo), they contribute a hot gas source, as does gas `shock-heating' during a galaxy collision \citep[e.g., ][]{2019ApJ...878..161J}. X-ray sputtering from hot gas \citep{2021A&A...649A..18G} also breaks up dust clouds and thereby removes the shielding from ionising radiation that dust may have provided potential stellar nurseries. Furthermore, these winds can keep the AGN `pilot light' on by supplying low-level fuelling \citep[e.g.,][]{1991ApJ...376..380C, 2006ApJ...640..143S} for the AGN. We term such an energy source a `Benson burner'.\footnote{This is a play on words combining Bunsen burner - used for heating and sterilisation - and a reference to the idea sparked by \citep{2003ApJ...599...38B}.} Should hot gas halos efficiently suppress star formation, then rather than ejecting gas which might form stars, it is about acquiring and retaining (hot) gas. The system needs to be capable of maintaining, and thus also massive enough to retain, a hot gas halo rather than have it evaporate or collapse into a disk where it may cool and form stars. A {\em hot 'n dry} (hot gas and dry merger) combination may help explain the upper-end of the galaxy mass function where star formation has dwindled or ceased. Unlike energetic but directional AGN jets \citep[which can both suppress and trigger star formation: ][]{2013ApJ...772..112S, 2014AN....335..531G, 2015ApJ...799...82C}, a hot gas halo can permeate the entire galaxy, including the disc. The relation between black hole mass and both X-ray gas temperature and luminosity \citep{2018ApJ...852..131B, 2019ApJ...884..169G, 2019MNRAS.488L.134L} may add credibility to this picture. Low levels of omni-directional particle outflows and electromagnetic radiation from the central `Benson burner' would also help counter cooling \citep{2006MNRAS.368L..67B, 2006MNRAS.370..645B, 2006MNRAS.365...11C, 2017MNRAS.465...32B} --- seen as X-ray radiation coming from the hot gas halo \citep[e.g.,][]{1976ApJ...207..460S, 1984MNRAS.208..185N}. This would help hold star-formation at bay, at least in a closed-box model with no substantial infall of cold gas \citep[e.g.,][]{2009Natur.457..451D}. When cooler gas is available, sporadic feeding and associated percolation events may produce bubbles and cavities observed at various wavelengths \citep{1984Natur.310..568S, 1985PASJ...37..359T, 2003ApJ...582..246B, 2004ApJ...607..800B}. However, this so-called `radio mode' AGN feedback would only maintain the $M_{\rm bh}$--$M_{\rm *,sph}$ relation, which we have argued is established by other means \citep{2016MNRAS.463.3948D, 2019SSRv..215....5W}. For the first time, we have used the black hole scaling relations to confirm that AGN primarily have a caretaker role among elliptical galaxies, and we have revealed how mergers rule the roost and dictate the $M_{\rm bh}/M_{\rm *,sph}$ ratio and presumably also the $M_{\rm *,sph}/M_{\rm dark matter}$ ratio \citep{2016MNRAS.463.3948D, 2021MNRAS.507.4274M}. This result is tied to the offset trend seen for elliptical galaxies in the $M_{\rm bh}$--$M_{\rm *,sph}$, $M_{\rm bh}$--$M_{\rm *,gal}$ and $M_{\rm bh}$--$R_{\rm e,sph}$ diagrams. It is not due to spheroids with partially depleted cores, which some E galaxies have but others do not, and which some S0 galaxy bulges possess. Such spheroids, whose central `phase space' is depleted of stars, tend to occupy the $M_{\rm bh}$--$M_{\rm *,sph}$ `red sequence', which is a `red herring' due to the partial picture it provided. In particular, it missed the wet and damp mergers, and thus the steep $M_{\rm bh}$--$M_{\rm *,sph}$ relation for the ensemble of elliptical galaxies. We will pursue this further by addressing merger-built lenticular galaxies with depleted cores, such as NGC~5813, and major wet mergers, for example NGC~5128, in \citet{Graham:Sahu:22}. \subsection{Some further thoughts}\label{Sec_imp} It is evident that the coevolution of bulges and their central black holes have not produced a simple near-linear $M_{\rm bh}$--$M_{\rm *,sph}$ relation. The steep $M_{\rm bh}$--$M_{\rm *,sph}$ relation for bulges has implications for countless simulations, semi-analytic works, theories, and papers that may have calibrated themselves to a near-linear $M_{\rm bh}$--$M_{\rm *,sph}$ relation. For example, as shown by \citet[][their Figure~6]{2018ApJ...852..131B}, while the Horizon-AGN simulation \citep{2014MNRAS.444.1453D} produces an $M_{\rm bh}$--$M_{\rm *,sph}$ `red sequence' with a slope around 1.1 to 1.2, it does not have the scatter to accommodate the steeper relations defined by either the bulges or the elliptical galaxies. That is, it appears to have not captured the key merger-induced jump from bulges to elliptical galaxies. While some studies are ahead of the pack, producing steeper relations \citep[e.g.,][]{2006MNRAS.373.1173F, 2012MNRAS.420.2662D, 2012MNRAS.423.2397K, 2017MNRAS.472L.109A, 2018MNRAS.479.4056W, 2019ApJ...885L..36D, 2020MNRAS.494.2747M, 2022MNRAS.513.3768I, 2022MNRAS.511.5756T}, it has been hard for the notion of a steep $M_{\rm bh}$--$M_{\rm *,sph}$ relation to get oxygen given the significant paradigm shift that it implies. It is, therefore, perhaps worth reiterating an element from \citet[][their section~4.3]{2013ApJ...764..151G}, which introduced a related revision to the `quasar mode' (aka cold-gas mode) of black hole growth used in some semi-analytic models \citep{2004ApJ...600..580G, 2005MNRAS.361..776S}. The steep $M_{\rm bh}$--$M_{\rm dyn,sph}$ relation detected by \citet{2012ApJ...746..113G}, which supplanted the single linear relation from \citet{2004ApJ...604L..89H}, challenged the past assumption of accretion-induced black hole growth that is linearly proportional to the inflowing mass of cold gas. \citet[][their Equation~8]{2006MNRAS.365...11C} and others have popularised this black hole feeding scenario to model how AGN outflows account for what was thought to be a linear $M_{\rm bh}$--$M_{\rm *,sph}$ relation. \citet{2013ApJ...764..151G} presented a revised prescription for the increase in black hole mass, $\delta M_{\rm bh}$, associated with wet mergers, such that \begin{equation} \delta M_{\rm bh} \propto \left(\frac{M_{\rm min}}{M_{\rm maj}}\right) \left[ \frac{M^X_{\rm cold}}{1+(280\, {\rm km\, s}^{-1})/V_{\rm virial}}\right]. \label{EqQ} \end{equation} The exponent $X$ represents the logarithmic slope of the $M_{\rm bh}$--$M_{\rm *,sph}$ relation for bulges, and they specified $X=2$. $M_{\rm min}$ and $M_{\rm maj}$ are the total baryonic masses from the minor and major galaxies involved in the accretion/merger event, and $M_{\rm cold}$ is their combined cold gas mass. The velocity $V_{\rm virial}$ is the merged system's circular or `virial' velocity, normalised at 280 km s$^{-1}$ \citep{2000MNRAS.311..576K}. This modified equation may prove helpful for exploring and understanding galaxy/(black hole) evolution through semi-analytic approaches, although it does not encompass the cessation of star-formation due to hot X-ray halos in massive systems, or the pivotal role of dry mergers in shaping the distribution seen in the $M_{\rm bh}$--$M_{\rm *,sph}$ diagram. There are also significant ramifications for predictions of gravitational waves from space-based interferometers \citep[e.g.,][]{2005LRR.....8....8M, 2013CQGra..30x4009S, 2020MNRAS.492..256K, 2022arXiv220505099S} and pulsar timing arrays monitored with ground-based radio telescopes \citep[e.g., ][]{2010CQGra..27h4013H, 2013Sci...342..334S, 2019MNRAS.488..401C}. For example, the steep $M_{\rm bh}$--$M_{\rm *,sph}$ relation for bulges should be considered, if not favoured over the near-linear `red-sequence' when assigning BH masses to galaxies in works attempting to estimate the background signal from binary black hole mergers. One caveat here is that the mergers involving a BCG may involve systems on both the bulge and the elliptical sequence. As noted in \citet{2019MNRAS.484..794G}, predictions for black hole masses will be too high if using the original near-linear $M_{\rm bh}$--$M_{\rm *,sph}$ relation at low spheroid masses. This over-prediction can result in over-looking potential populations of intermediate-mass black holes ($10^2 < M_{\rm bh}/M_{\odot} < 10^5$). Furthermore, application of the steeper relation has already been shown to result in an order of magnitude reduction to the expected detection rate of extreme mass ratio inspiral (EMRI) events from compact stellar-mass objects around massive black holes \citep[e.g.,][]{2012A&A...542A.102M}. That work can be further refined based on the updated relations herein, providing better expectations for what the European Laser Interferometer Space Antenna ({\it LISA}) \citep{1997CQGra..14.1399D} and TianQin \citep{2016CQGra..33c5010L} can hope to achieve based on their current design plans. As noted above, the pursuit of long-wavelength gravitational waves, from the coalescence of binary supermassive black holes \citep[e.g.,][]{2003ApJ...582L..15K, 2006ApJ...646...49R, 2019ApJ...884...36L, 2022ApJ...926L..35O}, is an endeavour underway via monitoring pulsar arrival times using radio telescopes. These studies will benefit from an improved knowledge of the varying $M_{\rm bh}/M_{\rm *,sph}$, and $M_{\rm bh}/M_{\rm *,gal}$, ratios in pre-merged galaxies. This can enable revised predictions for, and possibly aid in the tentative confirmation of, a long-wavelength gravitational wave background \citep{2020ApJ...905L..34A, 2021MNRAS.508.4970C}. Related to the EMRI events are the nuclear stars clusters that coexist with \citep[e.g.,][]{2008AJ....135..747G, 2008ApJ...678..116S, 2009MNRAS.397.2148G} and feed \citep[e.g.,][]{2002RvMA...15...27K, 2002ApJ...576..753L} the central black hole in galaxies. The revised/steeper $M_{\rm bh}$--$M_{\rm *,sph}$ relations, coupled with the $M_{\rm nsc}$--$M_{\rm *,sph}$ relations \citep{2003ApJ...582L..79B, 2003AJ....125.2936G}, led to the discovery of the $M_{\rm bh}$--$M_{\rm nsc}$ relation \citep{2016IAUS..312..269G, 2020MNRAS.492.3263G}. This should be useful for modelling not only gravitational radiation events but also the expected frequency of tidal disruption events \citep[TDEs:][]{2001astro.ph..6422K, 2004ApJ...600..149W, 2015JHEAp...7..148K, 2016MNRAS.455..859S, 2017MNRAS.465.3840C, 2020MNRAS.498..507T}, which have been observed in data dating back to 1990. There are currently around 100 such known events.\footnote{\url{https://tde.space/}} If a non-rotating Schwarzschild-Droste \citep{1916AbhKP1916..189S, 1917KNAB...19..197D} black hole is more massive than $\sim$10$^8$ M$_\odot$, and thus the gravitational gradient at, and beyond, the `event horizon' is not strong enough to pull a star apart, there will be no TDE \citep{1999MNRAS.309..447M}. The star will cross the event horizon and disappear without displaying its hot interior. As we have seen, most of the systems with $M_{\rm bh} \lesssim 10^8$ M$_\odot$ follow the near-quadratic $M_{\rm bh}$--$M_{\rm *,sph}$ relation for bulges, suggesting the need to use this steeper relation rather than the near-linear `red-sequence', which pertains to (some) systems with $M_{\rm bh} \gtrsim 10^8$ M$_\odot$. One should, however, be mindful that the spin-reduced size of the event horizon in a rotating \citet{1963PhRvL..11..237K} black hole can result in a star's tidal disruption radius being greater than the event horizon for black hole masses up $\sim$7$\times 10^8$ M$_\odot$ for maximally spinning black holes \citep{1992MNRAS.259..209B, 2012PhRvD..85b4037K}. The morphology-dependent black hole scaling relations also demand a re-examination of the virial $f$-factors used to convert AGN virial masses into black hole masses \citep[e.g.,][]{2009ApJ...694L.166B, 2018ApJ...864..146B}. Failing to account for the different morphologies and formation history of the spheroids hosting the AGN or inactive black holes will produce erroneous results. As noted in \citet{2012ApJ...746..113G}, there is a wealth of additional and immediate implications and insight from the steeper-than-linear $M_{\rm bh}$--$M_{\rm *,sph}$ relations. These include black hole mass predictions in other galaxies, constructing the black hole mass function, and deriving the black hole mass density based on reliable bulge and elliptical galaxy mass functions. In passing, it is noted that some care with the Hubble-Lema\^itre constant, or little $h$, is required for such calculations, as noted in \citet{2007MNRAS.380L..15G} and \citet{2013PASA...30...52C}. \section*{Acknowledgements} This research was supported under the Australian Research Council's funding scheme DP17012923. Part of this research was conducted within the Australian Research Council's Centre of Excellence for Gravitational Wave Discovery (OzGrav), through project number CE170100004. This research has made use of the NASA/IPAC Extragalactic Database (NED) and the NASA/IPAC Infrared Science Archive. We used the {\sc Rstan} package available at \url{https://mc-stan.org/}. We also used python packages numpy \citep{harris2020array}, matplotlib \citep{Hunter:2007} and SciPy \citep{2020SciPy-NMeth}. \section{Data Availability} The data underlying this article are available in the NASA/IPAC Infrared Science Archive. \bibliographystyle{mnras}

08c29fbb811ba167109d819fe85dc1afbe445758

\section{Introduction} \label{sec:intro} In online teaching, it is important for educators to record written contents to make the lectures engaging. Two popular techniques for recording are: \textit{(A)} utilizing an overhead camera which is straightforward, and \textit{(B)} using a graphics tablet or tab~\cite{Tab2021}, i.e. \textit{Intuos}\footnote{\texttt{www.wacom.com/en-us/products/pen-tablets}}. These tabs are becoming popular, but many teacher find it difficult as they are expensive~\cite{linda_2018}, and many lack abilities to use such instruments. Besides, most general-purpose graphics tabs have no built-in screen; hence, users must maintain \textit{hand-eye synchronization}~\cite{santos_2020}, which can be displeasing. Excessive use of these gadgets might lead to illness, like \textit{musculoskeletal}. ~\cite{Xu2020}. Tablets with integrated touch displays are even more expensive. In this paper, we present an alternative to graphics tab which is affordable yet providing its essential features. Our goal is to combine the old-fashioned pen-paper technique with the modern computer-based recording technology. Our solution avoids huge budget, and does not necessitate hand-eye synchronization. It needs only a \textbf{laptop}, a \textbf{pen}, and \textbf{paper}; and we name it ``\textit{{Do-It-Yourself Graphics Tab}}'', or ``\textit{{DIY Graphics Tab}}''. Fig.~\ref{fig:teaser} presents an overview \textit{DIY Graphics Tab}. To configure, user places a paper in front of a laptop and tilts its lid so that the webcam can capture the area containing the paper $\bigl[$\protect\includegraphics[scale=.19]{crc1.png}{$^{\large (a, b)}\bigr]$}. We found that, a tilting angle around $45$ degrees \textit{w.r.t} the base is adequate. Our system then processes the frame, and only contents are rendered on screen---as if it were taken from a ``bird's eye" view $\bigl[$\protect\includegraphics[scale=.19]{crc1.png}{$^{\large (d)}\bigr]$}. Our system takes care of the obstacles, such as---occlusion by hand, random movement of paper, background, lighting conditions, and perspective distortion due to the angular view.\\ Few works were done on capturing lectures and analyzing the contents, i.e., \cite{Davila2021, UralaKota2019, kota2018automated, davila2017whiteboard, lee2017robust, Yeh2014, Wienecke2003}, but they focused on whiteboard. Research that inclines mostly with our domain is \textit{WebcamPaperPen}~\cite{webcamPP2014}. In terms of ease of use and simplicity in configuration, our work outperforms \textit{WPP} (Table~\ref{tab:compare}). \begin{table}[h] \centering \begin{tabular}{lll} \textit{\textbf{\small Negative features}} & \textit{\textbf{\small WPP}} & \textit{\textbf{\small DIYGT}} \\ \hline \vspace{-1mm} \begin{tabular}[c]{@{}l@{}}\small Requires manual detection? \end{tabular} & {\cmark} & \begin{tabular}[c]{@{}l@{}}{\xmark}\end{tabular} \\ \vspace{-1mm} \begin{tabular}[c]{@{}l@{}}\small Requires extra equipment?\end{tabular} & \begin{tabular}[c]{@{}l@{}}{\cmark}\end{tabular} & \xmark \\ \vspace{-1mm} \begin{tabular}[c]{@{}l@{}}\small Requires steadiness?\end{tabular} & \cmark & \xmark \\ \vspace{-1mm} \begin{tabular}[c]{@{}l@{}}\small Requires specific type of pen?\end{tabular} & \begin{tabular}[c]{@{}l@{}}{\cmark}\end{tabular} & \xmark \\ \vspace{-1mm} \small Restriction on handedness? & \begin{tabular}[c]{@{}l@{}}{\cmark}\end{tabular} & \xmark \\ \vspace{-1mm} \small Requires eye-hand sync? & \begin{tabular}[c]{@{}l@{}}{\cmark}\end{tabular} & \begin{tabular}[c]{@{}l@{}} {\xmark}\end{tabular} \\ \begin{tabular}[c]{@{}l@{}}\small Requires prior experience?\end{tabular} & \cmark & \xmark \\ \hline \end{tabular} \caption{A comparison between \textit{WPP}~\cite{webcamPP2014} and ours.} \vspace{-11mm} \label{tab:compare} \end{table} \section{Implementation} Fig.~\ref{fig:teaser}\protect\includegraphics[scale=.19]{crc2.png} shows our pipeline. After a frame is captured $\bigl[$\protect\includegraphics[scale=.19]{crc2.png}{$^{\large (a)}\bigr]$}, it goes through several stages, which are explained below. $\bigl[$\protect\includegraphics[scale=.19]{crc2.png}{$^{\large (b, c)}\bigr]$}\textbf{:} To extract paper region, we developed a dataset of images of papers in different positions and lighting, and occlusions. Coordinates of a paper corners are stored as a convex hull of a quadrilateral mask. We trained \textit{Mask-RCNN}~\cite{he2017mask} model using our dataset and performed segmentation. $\bigl[$\protect\includegraphics[scale=.19]{crc2.png}{$^{\large (d, e)}\bigr]$}\textbf{:} Segmentation provided a mask covering paper. We extracted the largest contour of mask to detect corners. $\bigl[$\protect\includegraphics[scale=.19]{crc2.png}{$^{\large (f)}\bigr]$}\textbf{:} The \textit{$4$-point perspective transformation} \cite{rosebrock_2014, articleIPM, DBLP:journals/corr/abs-1812-00913} is applied on the corners to unwarp the paper region. $\bigl[$\protect\includegraphics[scale=.19]{crc2.png}{$^{\large (g)}\bigr]$}\textbf{:} We applied adaptive thresholding~\cite{nina2011recursive} followed by morphological operations and connected component analysis~\cite{dougherty2003hands} to avoid palm, fingers and noises.\\ To make the system robust, user needs to specify his/her handedness. For a left-handed person our system simply flips the frame in segmentation phase along the axis. Fig.~\ref{fig:res} shows more result. Materials and demos are available \href{https://imruljubair.github.io/project/project-page.html#tab}{{\color{blue}here}}. \section{Evaluation} We performed a survey on $29$ educators from university level, inviting them to use \textit{DIY Graphics Tab} for recording their lectures. We divide them into $3$ groups depending on their technical backgrounds on computer, which are---\textit{ {\textbf{\color{red}fTB}:}} teachers with no or very few technical background. \textit{\textbf{\color{blue}TB$-$GT}:} Teachers with technical background with no previous experience of using graphics tablet. \textit{\textbf{\color{gray}TB$+$GT}:} Teachers with technical background and also has previous experience of using graphics tablet. We asked participants to utilize our method and grade it on a scale of \textcircled{\small 1} to \textcircled{\small 5} (\textit{very poor} to \textit{excellent})~\cite{likert1932technique, mcleod_1970}. We found $44.4\%$ of \textit{\textbf{\color{gray}TB$+$GT}} gave \textcircled{\small 4}. Our method received \textcircled{\small 5} from a majority ($57.1\%$) of \textit{\textbf{\color{blue}TB$-$GT}}. Almost $77\%$ of our target users, \textit{\textbf{\color{red}fTB}}, gave highest score. The average rating is $4.44$. The lid must be slanted in our system which rise concerns as the screen becomes non-visible to instructors. But we are motivated from overhead camera configuration---where the user primarily concentrates on desk. However, we conducted a poll to assess if the tilting may be a problem. We observe that around $72\%$ of teachers are comfortable with it. They think the non-visible screen is not a major concern as long as they can save the costs of graphic tabs. We also collected a questionnaire-based evaluation from testers to determine the cost-effectiveness of our \textit{DIY Graphics Tab}. To maintain our evaluation on the same ground, we mostly use the questions from \textit{WebcamPaperPen}~\cite{webcamPP2014, mastersthesis}. The user responses are listed below. We observed a majority of users believed that using a graphical tab is highly expensive and found our technology to be cost efficient. \begin{itemize} \item \small{It replaces graphics tab perfectly for lecture recording ($31.03\%$)}. \item \small{The actual graphics tab is better but it is too expensive, I would rather use it ($41.38\%$).} \item \small{After a certain period I would get tired of it and would buy a proper graphics tab to improve my content ($20.69\%$).} \item \small{It does not fit my lecture recording ($6.9\%$).} \end{itemize} \begin{figure}[t] \vspace{-4mm} \includegraphics[width=0.4\textwidth]{Presentation2.jpg} \caption{Result of \textit{DIY Graphics Tab}.} \label{fig:res} \vspace{-4mm} \end{figure} \section{Conclusion} In this paper, we introduced \textit{DIY Graphics Tab} as a substitute for graphic tab using pen--paper and webcam. We used \textit{Mask-RCNN} to obtain the region of paper image and applied perspective transformation to achieve top-down view. Our system has shortcomings that demand futher research, e.g., flickering, haziness. etc. We plan to enrich dataset and use advanced segmentation techniques, e.g., \textit{YOLACT}~\cite{bolya2019yolact, yolactplus}. We also like to convert our system into a real-time application. \newpage \newpage \bibliographystyle{ACM-Reference-Format}

533ba6b595df7cdb1adb4e6ba25abd4c05956571

\section{Introduction} \label{S1} Aluminium alloys are widely used in industry for their various advantages: they are light weighted metallic alloys, have a high corrosion resistance and good mechanical properties \cite{Mondolfo1979,Develay1992}. Since the early work of Guinier \cite{Guinier1938} and Preston \cite{Preston1938}, it has been demonstrated that the increase of hardness in age hardened Al-Cu alloys is due to the formation of nanoscaled zones that have been first detected by small-angle X-ray scattering (SAXS) known as the Guinier-Preston (GP) zones \cite{Sigli2018}. When such alloy is quenched, then aged at room temperature (\textit{i.e.}{}: naturally aged), the formation of these GP zones is activated by the excess vacancies after quenching, which are slowly annihilated near residual dislocations and grain boundaries. When an Al-Cu alloy is heat treated, the precipitation sequence is \cite{Starink2004,Son2005,Rodriguez2018}: Super Saturated Solid Solution $\rightarrow$ Cu clustering $\rightarrow$ GP zones $\rightarrow$ $\theta''$ $\rightarrow$ $\theta'$ $\rightarrow$ $\theta$. The formation of GP zones is obtained quickly at ambient temperature, thus the copper clustering sequence is usually neglected to describe the precipitation sequence of this alloy. GP zones in Al-Cu alloys are circular nanoscaled disks parallel to \{100\} planes, isolated in one layer \cite{Silcock1954,Bourgeois2011}. The precipitation in Al-Cu alloys has been extensively studied in the literature through hardness measurements, transmission electron microscopy (TEM), differential scanning calorimetry (DSC) and strength models have been developed to correlate the increase of hardening to the precipitation state \cite{Myhr2001,Rodriguez2018}. Atomistic calculations have been used to predict the evolution during ageing of the precipitation \cite{Hutchinson2010,Sigli2018,Deschamps2021} in addition to determine the formation energy of particles, their geometries and interaction with dislocations \cite{Zander2008,Singh2010,Staab2018,Miyoshi2019}. These parameters can be then implemented in classical nucleation and growth theories or/and clusters dynamics \cite{Clouet2009c,Stegmuller2019}. Publications about naturally aged alloys and the role of excess vacancies are numerous, but less information is reported on GP zone nucleation and growth during the first step of natural ageing. Besides, aluminium alloys are among materials bearing the highest potential to contain, and to facilitate the transport of hydrogen fuel due to their various advantages and they are already used in fuel-cell-based, electric vehicles \cite{Horikawa2019}. More generally, since hydrogen is becoming a key component of the energy transition worldwide, it can either be used as an energy carrier or directly as a fuel in vehicles, including automobiles and planes. However, due to its small size and high mobility, hydrogen influences the mechanical properties of metals and alloys leading to premature failures of engineering structures. This phenomenon is called hydrogen embrittlement (HE) and several models have been proposed in the literature to describe the underlying physical mechanisms (see for details \cite{Lynch2012,Kirchheim2014,Robertson2015}). For all models, HE involves the energy reduction of one process in the presence of hydrogen to activate a mechanism (\textit{e.g.}{}: grain boundary segregation \cite{Hondros1996,Oudriss2012}, Cottrell atmosphere of dislocation \cite{Gu2018,Hachet2020b}, shielding effect promoting slip band localisation \cite{Beachem1972,Birnbaum1994,Delafosse2001}, enhancement of vacancy formation \cite{McLellan1997,Harada2005,Fukai2006}, and so on...). These models can describe accurately HE of pure metals, but can fail describing HE in more complex materials (\textit{e.g.}{} in alloys) where mechanical properties are dependent on the distribution of strengthening precipitates and their interactions with defects, in particular dislocations. Aluminium alloys are also not immune to hydrogen ingress and when hydrogen is introduced as a solute, it easily diffuses and segregates to crystalline defects \cite{Scully2012,Zhao2022}. \textit{Ab initio}{} calculations have shown that hydrogen atoms strongly interact with vacancies in aluminium \cite{Wolverton2004,Nazarov2014,Connetable2018}. Previous studies focusing on the hydrogen/vacancy interactions in metals have shown that hydrogen decreases the formation energy of vacancy clusters containing hydrogen \cite{Fukai2006,Nazarov2014} and the vacancy migration energy \cite{Du2020}. Hydrogen may also delay the clustering of solutes and the coarsening of precipitates in some Al alloys \cite{Zhao2022,Hachet2022b}. Therefore, it is critical to understand how these interactions impact the kinetic and thermodynamic of precipitates in aluminium alloys (which may evolve even at room temperature) to predict microstructural evolutions and eventually to reduce the damaging effect of hydrogen. In this study, the influence of hydrogen on the early stage of GP zone formation during natural ageing in an Al-5Cu alloy is investigated. The following section is focused on experimental data that highlight the influence of hydrogen on GP zone nucleation and growth during the first step of natural ageing. Then, \textit{ab initio}{} calculations are presented to first demonstrate the impact of copper on the interaction between hydrogen and vacancy. The second part of the \textit{ab initio}{} calculations is focused on the effect of hydrogen on the diffusion of vacancy and copper in FCC aluminium. \section{Experimental evidence of the hydrogen influence on the GP zone formation and hardening kinetics} \label{S2} \subsection{Experimental details} \label{S21} The investigated material is provided by Goodfellow$^{\circledR}$ with the following composition (wt.\%): 5.3\%Cu-0.7\%Fe-0.4\%Si-0.3\%Pb, Al balance (standard AA2011, called Al-5Cu further). Disc shaped samples (with a diameter of 6\,mm and a thickness of 1\,mm) are solutionised at 810\,K during 1\,h, water quenched, then naturally aged at room temperature either in air or 5\,h in a 0.1\,M NaOH solution. Before the introduction in the solution, samples are quickly (few minutes) mechanically grinded using SiC foil paper with a particle size of 8\,µm to remove the oxide layer grown during the solution heat treatment. Aqueous solution containing NaOH is aggressive towards aluminium and its oxide, it prevents the formation of a passive layer and leads to H incorporation in the alloy \cite{Birnbaum1997,Scully2012}. After 5\,h in NaOH, the samples are further aged in air at room temperature and the hardness evolution is compared to the alloy directly aged in air. The increasing hardness resulting from the GP zone nucleation and growth is firstly measured by micro-hardness measurements, using a Future tech FM7 device at room temperature. The micro-hardness values presented in this study are the average of at least 6 indents obtained with a micro Vickers diamond indenter using a load of 500\,g and a dwell time of 10\,s. High angle annular dark field scanning TEM (HAADF-STEM) images are recorded with collection angles ranging from 50 to 180\,mrad using a JEOL ARM 200 microscope, operated at 200\,kV. Thin foil specimens are prepared with a twin-jet electro-polisher (TENUPOL 5 from Struers$^{\circledR}$) using a mixture of $30\%\mathrm{HNO_3}-70\%\mathrm{CH_3OH}$ (\%vol) at 243\,K. Final thinning is carried out by low-energy ion milling conducted with a GATAN$^{\circledR}$ Precision Ion Polishing System. \subsection{Hardness kinetic variations of the naturally aged Al-5Cu alloy due to hydrogen} \label{S22} After water quenching, the hardness of the alloy is 91$\pm$1\,HV, and it increases progressively to reach a maximum of 117$\pm$2\,HV after $\sim50\,h$ (see fig. \ref{FigHVAgeTime_AlCu}). When the alloy is aged 5\,h in NaOH solution to introduce hydrogen, the microhardness is significantly lower than the alloy directly aged in air: after 5\,h, the hardness of the alloy naturally aged in air is 105$\pm$2\,HV while it is only 95$\pm$2\,HV when NaOH treatment is carried out. However, after several additional hours (50 to 200 hours) at room temperature in air, the micro-hardness further increases and catches up the hardness of the material without hydrogen. This suggest that hydrogen atoms quickly desorb from the alloy and do not significantly affect the final microstructure (fig. \ref{FigHVAgeTime_AlCu}). To confirm these measurements, HAADF-STEM observations are carried out further. \begin{figure}[bth!] \centering \includegraphics[width=0.99\linewidth]{FigHVAgeTime_AlCu.eps} \caption{Vickers micro-hardness evolution as a function of the natural ageing time when the alloy is diractly aged (DA) in air and after 5h in NaOH.} \label{FigHVAgeTime_AlCu} \end{figure} Naturally aged materials are observed by HAADF-STEM in (001) zone axis to clearly exhibit GP zones. They are observed after being naturally aged 1 and 9 days. When the alloy is directly aged 1 day in air, small GP zones appear, as shown in fig. \ref{FigSTEMNA}.a and they become significantly larger after 9 days (fig. \ref{FigSTEMNA}.c). When it is aged 5\,h in NaOH, GP zones are not visible after 1 day (fig \ref{FigSTEMNA}.b), but become visible after 9 days (fig \ref{FigSTEMNA}.d), with an average diameter similar to those observed in the alloy directly aged 1 day in air (fig \ref{FigSTEMNA}.a). These observations are consistent with the delayed hardening (fig. \ref{FigHVAgeTime_AlCu}). However, the hardnesses of the alloy aged in both conditions are relatively similar after 9 days (fig. \ref{FigHVAgeTime_AlCu}), suggesting that hydrogen does not affect the final microstructure, which obviously is not the case when fig \ref{FigSTEMNA}.c and \ref{FigSTEMNA}.d are compared. \begin{figure*}[bth!] \centering \includegraphics[width=\linewidth]{FigSTEM_NA.eps} \caption{HAADF-STEM micrographs of the precipitation evolution of the Al-5Cu alloy directly aged in air (a and c) and 5\,h in NaOH followed by an ageing in air (b and d). The total natural ageing time is 1 day (a and b) or 9 days (c and d). All samples are oriented in the (001) zone axis and their corresponding FFTs are set in the bottom right corner.} \label{FigSTEMNA} \end{figure*} GP zone diameters $d$ are directly measured on STEM-HAADF images and their distributions are plotted in fig. \ref{FigdGPZ}. There is no distribution for the alloy stored in 5\,h in NaOH, then aged 1 day in air because GP zones could not be observed (fig. \ref{FigSTEMNA}.b). Longer natural ageing gives larger GP zones with a mean diameter varying form 1.4\,nm$\pm$0.5\,nm after 1 day to 2.6\,nm$\pm$0.8\,nm after 9 days. Besides, the diameter distribution becomes significantly broader during natural ageing. When the alloy is stored 5\,h in NaoH prior to 9 days in air, the mean diameter is 1.7\,nm$\pm$0.5\,nm with a narrow distribution similar to that of the alloy aged during 1 day directly in air. \begin{figure}[bth!] \centering \includegraphics[width=\linewidth]{FigGPZ_Distri.eps} \caption{Distribution of GP zone diameters measured on HRSTEM images in alloy directly aged in air during 1 day and 9 days (DA) and 5\,h in contact of hydrogen, then aged in air (5\,h NaOH). The measurements were carried out on of 200\,GP zones for each distribution.} \label{FigdGPZ} \end{figure} Due to the limited foil thickness, some GP zones are partially imaged and the true diameter $d_t$ of GP zones can be determined from the apparent diameter $d$ using \cite{Nie2008,Rodriguez2018}: \begin{equation} d = \left(\frac{\pi d_t/4 + \delta}{\delta + d_t} \right)d_t, \label{eqdapp} \end{equation} with $\delta$ the thin foil thickness. In this work, thin foil thickness $\delta$ was measured by electron energy loss spectroscopy. Assuming an electron mean free path in aluminium of 120\,nm \cite{Bardal2000}, the thickness of thin foils was systematically between 40\,nm and 100\,nm. The density of GP zones may be also estimated, following \cite{Nie2008,Rodriguez2018}: \begin{equation} n_P = \frac{N_x + N_y + N_z}{S (d_t + \delta)}, \label{eqnP} \end{equation} with $N_x$, $N_y$ $N_z$, the number of GP zones counted in the x, y and z axes, parallel to the $[100]$, $[010]$ and $[001]$ directions, respectively and $S$ the surface of observation ($S=5.10^{-3} {\rm \mu m}$). Since only GP zones perpendicular to (001) Al planes are visible on HRSTEM images (fig. \ref{FigSTEMNA}), the number of precipitates in the z-direction ($N_z$) is obtained from $N_x$ and $N_y$ using: \begin{equation} N_z = 0.5 \times (N_x + N_y)\left(\frac{d_t+\delta}{\sqrt{S}}\right). \label{eqNz} \end{equation} Assuming that the thickness of all GP zones is half the aluminium lattice parameter ($a_{Al} = 0.404\,{\rm nm}$), their volume fraction writes as: \begin{equation} f_V = n_P \frac{\pi d_t^2 a_{Al}}{8}. \label{eqfV} \end{equation} \begin{table*}[h!] \centering \caption{Quantitative measurements of the GP zone and cluster densities $n_P$, apparent diameter $d$ and volume fraction $f_V$ in the Al-5Cu alloy directly aged in air or after 5h in NaOH.} \label{tblSTEMNA} \begin{tabular}{cccccc} \hline \ &\ & \multicolumn{2}{c}{1 day} & \multicolumn{2}{c}{9 days} \\ \ & \ & Air & NaOH & Air & NaOH \\ \hdashline GP & $<d>$ (nm) & 1.4$\pm$0.5 & \ & 2.6$\pm$0.8 & 1.7$\pm$0.5 \\ zones & $n_P$ (10$^{-3}$ nm$^{-3}$) & 2.4$\pm$0.1 & \ & 3.6$\pm$0.1 & 1.6$\pm$0.1 \\ \ & $f_V$ (\%) & 0.1 & \ & 0.4 & 0.1 \\ \hdashline Cu & $<d>$ (nm) & 1.4$\pm$0.5 & 1.3$\pm$0.4 & \ & 1.1$\pm$0.4 \\ clusters & $n_P$ (10$^{-3}$ nm$^{-3}$) & 0.8$\pm$0.1 & 1.8$\pm$0.1 & \ & 1.1$\pm$0.1 \\ \ & $f_V$ (\%) & 0.06 & 0.12 & \ & 0.04 \\ \hline \end{tabular} \end{table*} Fast Fourier transforms (FFTs) of each HAADF images are also displayed in fig \ref{FigSTEMNA}. Spots in FFTs at the \{110\} positions are clearly exhibited for the alloy directly aged during 1 day, but they disappear after 9 days when GP zones are well developed. These spots also appear for the alloy aged 5\,h in NaOH, followed by 1 day in air with a stronger intensity as compared to the directly aged alloy. These spots do not completely disappear after 9 days and thus seem to be linked to features appearing before GP zones. \{110\} reflections are forbidden reflections on diffraction pattern and they appear only on FFTs of HRSTEM images. Similar reflections on FFT patterns have been reported in naturally aged Al-Zn-Mg alloys. Authors have attributed these features to either $\rm{Al_3Zn}$ dispersoids or artefacts caused by the TEM sample preparations \cite{Lervik2021}. Since we do not have such dispersoids and since similar preparation conditions were applied for all thin foils, these spots can only be the result of real features revealed by images. Since they disappear when GP zones are well developed, they might results from Cu clusters that form prior to GP zones \cite{Starink2004,Son2005,Rodriguez2018}. To study these clusters, a mask is applied on the FFTs of each image having strong signals at the \{110\} positions (fig. \ref{FigCuClDistri}.a), isolating these spots. Then, a reconstructed image showing areas responsible of the strong \{110\} signals is obtained using an inverse FFT function (fig. \ref{FigCuClDistri}.b) Assuming that these areas correspond to Cu clusters, their sizes are measured and the obtained distributions are displayed on fig \ref{FigCuClDistri}.c. Contrary to GP zones, these nanosized particles are not detectable without this image filtering and disappear when the alloy is directly aged in air during 9 days (no signal is observed at the \{110\} positions of the FFTs of fig \ref{FigSTEMNA}.c). The precipitate density is directly obtained by counting the observed clusters. The cluster size is estimated from the average of the longest and shortest distances measured on images. The obtained distributions are similar for all states but the average diameter is significantly smaller for samples stored in 5\,h NaOH \begin{figure}[bth!] \centering \includegraphics[width=\linewidth]{FigCuClDistri.eps} \caption{Determination and distribution of the copper clusters' size. a. HAADF-STEM image of Al-5Cu aged 5\,h in NaOH and 1 day in air with the masks applied on the corresponding FFT. b. Inverse FFT image showing regions with strong \{110\} signal on FFT. c. Distribution of the cluster diameters measured on filtered images corresponding to \ref{FigSTEMNA}.a, \ref{FigSTEMNA}.b and \ref{FigSTEMNA}.c. The measurement were carried out 200 clusters for each distribution.} \label{FigCuClDistri} \end{figure} The determined GP zone and Cu cluster densities $n_P$, apparent diameters $d$ and volume fraction $f_V$ are summarised in table \ref{tblSTEMNA}. These experiments clearly indicate a significant change of microstructure due to hydrogen incorporation: there is a delay of GP zone formation, observable using microhardness and HAADF-STEM. Complementary information is detailed in \ref{AppDSC} where DSC measurements have been performed on this alloy differently aged. Hydrogen strongly interacts with excess vacancies \cite{Fukai2006,Connetable2018,Du2020,Hachet2022b}, and thus seems to delay the formation of GP zones probably by reducing the diffusion of copper or/and by increasing the energy barrier of GP zones nucleation. \textit{Ab initio}{} calculations are then carried out to get a deeper understanding of fundamental mechanisms leading to the delayed formation of GP zones when hydrogen is introduced on the alloy. \section{Influence of copper atoms on the hydrogen-vacancy interactions in aluminium} \label{S3} \textit{Ab initio}{} calculations are performed to quantify the variation of copper diffusion when the atom is close to a hydrogen vacancy complex. The first step is to estimate the interaction enthalpy between vacancy and hydrogen $H^{inter}_{Vac-H}$ in aluminium and in a diluted Al-Cu alloy. The aim is to determine if this complex is more stable in the vicinity of copper atom before determining the influence of H on the diffusion variation of copper. \subsection{Computational details} \label{S31} The following \textit{ab initio}{} calculations used density functional theory (DFT) in the Quantum Espresso code \cite{Hohenberg1964,Kohn1965,Giannozzi2009}. Pseudopotentials built with the projected augmented wave method \cite{Blochl1994,Kresse1996} are used with a kinetic energy cutoff of 600\,eV. The exchange-correlation is described with the generalised gradient approximation with the Perdew-Burke-Ernzerhof functional \cite{Perdew1996}. All calculations are performed at constant pressure with a 0.2\,eV Methfessel-Paxton broadening \cite{Methfessel1989}. All simulations cells are 4$\times$4$\times$4 repetition of the primitive cell and the Brillouin zone was sampled using a 8$\times$8$\times$8 Monkhorst and Pack \cite{Monkhorst1976}. Atomic positions are relaxed until all ionic forces are inferior to 10\,meV/\AA{} for static and for climbing nudged elastic band (C-NEB) calculations. With these parameters and using the method described in previous study \cite{Hachet2018}, the obtained lattice parameters and elastic constants are deduced and given in table \ref{tblaCijAlCu}. The experimental values are also given and showing that these parameters describe accurately both metals at atom scales \cite{Kittel2004}. \begin{table}[h!] \centering \caption{Lattice parameters and elastic constants of pure Al and Cu.} \label{tblaCijAlCu} \begin{tabular}{ccccc} \hline \ & $a$ (nm) & $C_{11}$ (GPa) & $C_{12}$ (GPa) & $C_{44}$ (GPa) \\ \hline Al & 0.40394 & 116 & 61 & 33 \\ \cite{Kittel2004} & 0.405 \ & 114 & 62 & 28 \\ \hline Cu & 0.36304 & 177 & 124 & 80 \\ \cite{Kittel2004} & 0.361 & 176 & 125 & 82 \\ \hline \end{tabular} \end{table} \subsection{Interaction between hydrogen and vacancy in the vicinity of copper} \label{S32} The first calculations aim at determining the interaction enthalpy of hydrogen vacancy complex $H_{Vac-H}^{inter}$ with and without Cu atom in its vicinity. This enthalpy determined at zero pressure ($p=0$) is defined by: \begin{equation} H^{inter}_{Vac-H} = H^{sp}_{Vac-H}(p) - H^{sp}_{Vac}(p) - H^{sp}_{H}(p) + H^{sp}_{Bulk}(p), \label{eqHinter} \end{equation} with $H^{sp}_{Vac-H}$ and $H^{sp}_{Vac}$, the enthalpies of the supercell containing the hydrogen-vacancy complex and the hydrogen free vacancy, respectively. The enthalpies $H^{sp}_{H}$ and $H^{sp}_{Bulk}$ are from the vacancy free systems with and without hydrogen in solid solution, respectively. Since the tetrahedral interstitial sites are the most stable sites for hydrogen near vacancy in pure Al \cite{Wolverton2004,Nazarov2014,Connetable2018}, we introduced hydrogen in these sites for pure Al and in a diluted Al-Cu system. For the latter, the different interstitial sites ($T_i$ sites) are not equivalent when the copper atom is in the vicinity of the vacancy as shown in fig. \ref{FigdCuH}. However, due to the crystal symmetry, the distance between Cu and H when H is in $T_1$ is identical to the distance between Cu and H when H is in $T_2$. These equivalent positions are called $P_1$ (with $d_{Cu-H}^{P_1}$=0.04\,nm). The distance between Cu and H is also identical when H is in $T_3$, $T_4$, $T_5$ and $T_6$, these equivalent positions are called $P_2$ ($d_{Cu-H}^{P_2}$=0.08\,nm). Finally, the position $P_3$ stands for H in sites $T_7$ and $T_8$ ($d_{Cu-H}^{P_3}$=0.11\,nm). Therefore, $H^{inter}_{Vac-H}$ has to be calculated for only 3 positions in a diluted Al-Cu system. \begin{figure}[bth!] \centering \includegraphics[width=0.99\linewidth]{FigdCuH.eps} \caption{Illustration of the hydrogen possible positions near vacancy in a diluted Al-Cu. The vacancy in the Al matrix (grey) containing copper (blue) is represented with the grey square. The different possible positions for hydrogen atoms are represented with various shades of red.} \label{FigdCuH} \end{figure} The interaction enthalpies between hydrogen and vacancy in pure aluminium and in a diluted Al-Cu are displayed in fig. \ref{FigEinterVacH}. The interaction enthalpy is always attractive when hydrogen is close to the vacancy, both in pure Al and in diluted Al-Cu. In pure aluminium, $H^{inter}_{Vac-H} = -0.34$\,eV, similar to the interaction energy reported in previous work using \textit{ab initio}{} calculations ($E^{inter}_{Vac-H} = -0.335$\,eV \cite{Wolverton2004} and $E^{inter}_{Vac-H} = -0.33$\,eV \cite{Connetable2018}) but lower than experimental data ($-0.45\pm0.07$\,eV \cite{Linderoth1988}). In diluted Al-Cu, $H^{inter}_{Vac-H}$ is also negative, meaning that vacancies attract hydrogen, but the attraction is stronger when the vacancy is not bound to a Cu atom. Small variations are also observed depending on the position of H in the lattice: if H is inserted in $P_1$ (\textit{i.e.}{}: sites $T_1$ and $T_2$ in fig. \ref{FigdCuH}), the attraction is the weakest with $H^{inter}_{Vac-H}(P_1) = -0.29$\,eV. If H is inserted in $P_3$ (\textit{i.e.}{}: sites $T_7$ and $T_8$ in fig. \ref{FigdCuH}), $H^{inter}_{Vac-H}(P_3) = -0.31$\,eV and the attraction is the strongest when H is in $P_2$ position (\textit{i.e.}{}: sites from $T_3$ to $T_6$ in fig. \ref{FigdCuH}), with $H^{inter}_{Vac-H}(P_2) = -0.33$\,eV. These interaction enthalpy variations show that copper atoms have an influence on the interaction between hydrogen atoms and vacancies in FCC aluminium matrix, but hydrogen still segregates near vacancies even with a copper atom in its vicinity and the most stable sites for H corresponds to $P_2$ and $P_3$ positions. \begin{figure}[bth!] \centering \includegraphics[width=0.99\linewidth]{FigEinterVacH.eps} \caption{Interaction enthalpy between vacancy and hydrogen in pure Al and in diluted Al-Cu system. For diluted Al-Cu, energies are displayed as a function of the distance between Cu and H atoms.} \label{FigEinterVacH} \end{figure} \section{Atomic scale modelling of hydrogen consequences on the diffusion of copper in aluminium} \label{S4} Further, the diffusion coefficient of copper in aluminium is evaluated using the five jump frequency model from LeClaire \cite{LeClaire1978}, which has been used in previous studies to describe the diffusion of vacancies in aluminium matrix \cite{Yang2021} or substitutionnal solute (including copper) in different matrix, including aluminium \cite{Mantina2009a,Mantina2009b,Naghavi2017}. The first step to determine $D_{Cu}$ is to calculate the self diffusion coefficient for Al, $D_{Al}$ and the impact of hydrogen on it. Therefore, the influence of H on the self-diffusion coefficient of $D_{Al}$ is firstly investigated then the consequence of H of $D_{Cu}$ is studied. \subsection{Hydrogen impact on self diffusion in pure aluminium} \label{S41} Since the interaction between vacancy and hydrogen is attractive in diluted Al-Cu, the self-diffusion coefficient for Al ($D_{Al}$) and the diffusion of copper ($D_{Cu}$) near hydrogen vacancy complex is determined, further. The self diffusion coefficient $D_{Al}$ can be defined as \cite{LeClaire1978}: \begin{equation} D_{Al}(T) = f_0\omega_0(T)a(T)^2C_{0}^{j}(T), \label{eqDAl} \end{equation} with $f_0$ a correlation factor, which is constant and equal to 0.7815 for FCC crystals, $a(T)$ the lattice parameter of aluminium determined at finite temperature following the method described in \cite{Metsue2016,Hachet2018} by considering the atomic vibrations of aluminium atoms (see \ref{AppLatParamDet} for more details on its determination at finite temperature). The concentration $C_{0}^{j}(T)$ (with $j = Vac$ or $Vac-H$) is the equilibrium concentration of vacancies or hydrogen-vacancy complexes, respectively and defined as \cite{Metsue2014a,Naghavi2017}: \begin{equation} C^{Vac}_0(T) = \exp\left(-\frac{H_{Vac}^f - TS_{Vac}^f}{k_BT} \right), \label{eqCVac} \end{equation} where $k_B$ and $T$ are the Boltzmann constant and the temperature of the system, respectively. $H^f_{Vac}$ and $S^f_{Vac}$ are the vacancy formation enthalpy and entropy written in the hydrogen free system as: \begin{align} H_{Vac}^f &= H^{sp}_{Vac} - \frac{N-1}{N} H^{sp}_{Bulk}, \label{1}\\ S_{Vac}^f &= S^{sp}_{Vac} - \frac{N-1}{N} S^{sp}_{Bulk}, \label{2} \end{align} with $H^{sp}_{Vac}$, $S^{sp}_{Vac}$, $H^{sp}_{Bulk}$ and $S^{sp}_{Bulk}$ the enthalpy and entropy of the supercell of a perfect crystal with and without a vacancy, respectively. The obtained formation enthalpy $H_{Vac}^f$ is equal to 0.64\,eV, in agreement with the literature \cite{Mantina2008,Vo2017,Connetable2018}. In previous work, Naghavi \textit{et al}{} have shown how the self diffusion coefficient of Cobalt is impacted by modifying $S^f_{Vac}$ \cite{Naghavi2017}. Hence, we have explicitly calculated $S^{sp}_{Vac}$ and $S^{sp}_{Bulk}$ with the PHON program \cite{Alfe2009}, which calculates force constant matrices and phonon frequencies in both crystals. For hydrogen vacancy complexes, their equilibrium concentration $C^{Vac-H}_0(T)$ is: \begin{equation} C^{Vac-H}_0(T) = \exp\left(-\frac{H_{Vac-H}^f - TS_{Vac-H}^f}{k_BT} \right), \label{eqCVac-H} \end{equation} assuming that $S^f_{Vac-H}=S^f_{Vac}$ and with $H^f_{Vac-H}$ defined as \cite{Metsue2016,Connetable2018}: \begin{equation} H_{Vac-H}^f = H^{sp}_{Vac-H} - \frac{N-1}{N} H^{sp}_{Bulk} - 0.5\times H_{H_2}, \label{eqHfVac_HH} \end{equation} with $H_{H_2}$ the enthalpy of the hydrogen molecule at zero pressure $p = 0$ MPa as a reference. It leads to $H^f_{Vac-H}=1.05$\,eV, close to data reported in the literature (1.32\,eV \cite{Xie2016} and 1.02\,eV \cite{Connetable2018}). The jump frequency $\omega_0(T)$ is defined as \cite{LeClaire1978}: \begin{equation} \omega_{0} = \nu_0\exp\left(-\frac{H_{0}^m}{k_BT} \right), \label{eqw0} \end{equation} with $\nu_0$ the attempt frequency equal to $\sqrt(3/5)k_B/h \theta_D$ ($h$ and $\theta_D$ are the Planck constant and Debye temperature of aluminium, respectively) \cite{Satta1998,Kittel2004}. $H_{0}^m$ is the migration enthalpy of vacancies or hydrogen vacancy complexes. Since all calculations are performed at zero pressure, it is assumed that the migration enthalpy $H_{0}^m$ is equivalent to the migration energy $E^m_{0}$ \cite{Carling2003,Naghavi2017}, which is obtained from the transition states of the C-NEB calculations of vacancy (with or without H) diffusing from one site to another. Without H, the vacancy migration enthalpy presented in fig. \ref{Figw0} is found $H_{Vac}^m=E_{Vac}^m=0.59$\,eV, in agreement with the literature \cite{Mantina2008}. Assuming Al atoms exchange a similar way with the vacancy in pure Al and in a diluted Al-Cu, hydrogen can either be still associated to the vacancy (when H is in $P_1$) or dissociated from the vacancy (when H is in $P_3$). The migration energy of a hydrogen-vacancy complex (with H in $P_1$ and $P_3$) is also obtained using C-NEB calculations and results are also presented in fig. \ref{Figw0}. When H is still associated to the vacancy during the Al atom displacement, an increase of the migration energy is observed ($E_{Vac-H}^m(P_1)=1.08$\,eV) while $E_{Vac-H}^m(P_3)=0.61$\,eV, close to $E_{Vac}^m$. However, the final state is less stable when H is in $P_3$, the energy difference corresponding to the hydrogen-vacancy binding energy. \begin{figure}[bth!] \centering \includegraphics[width=\linewidth]{Figw0.eps} \caption{Energy variation of Al atom migrating in pure Al near vacancy and hydrogen-vacancy complex.} \label{Figw0} \end{figure} The self-diffusion coefficient calculated as a function of the temperature is displayed in fig. \ref{FigDAl}. The coefficient $D_{Al}$ obtained from our \textit{ab initio}{} calculations is close to the values reported in the literature in a range of 360\,K to 933\,K \cite{Mantina2008,Lundy1962,Volin1968,Messer1975}. In presence of H, $D_{Al}$ is always reduced but the diffusion coefficient is less impacted when H is in $P_3$ than when it is in $P_1$ (due to the increase of migration energy $E_{Vac-H}^m(P_1) = 1.08$\,eV). \begin{figure}[bth!] \centering \includegraphics[width=\linewidth]{FigDAl.eps} \caption{Calculated self-diffusion coefficient for Al, including previous work of Mantina \textit{et al}{} \cite{Mantina2008}, Lundy and Murdock \cite{Lundy1962}, Messer \textit{et al}{} \cite{Messer1975}, Volin and Ballufi \cite{Volin1968}}. \label{FigDAl} \end{figure} The diffusion of impurities or vacancies usually follows an Arrhenius type temperature dependence \cite{LeClaire1978}: \begin{equation} D_{Al}(T) = D_0 \exp{\left(- \frac{Q}{k_B T}\right)}, \label{eqDAlArr} \end{equation} with $D_0$ a pre-exponential factor and $Q$ the activation energy. From \textit{ab initio}{} calculations (fig. \ref{FigDAl}), curves are fitted using eq. (\ref{eqDAlArr}) to quantify the change of $D_0$ and $Q$ due to hydrogen incorporation. The results are given in table \ref{tblD0QDAl}. Without hydrogen, calculated data are similar to those reported in the literature but an important increase of the activation energy is obtained when hydrogen-vacancy complexes exist. \begin{table}[h!] \centering \caption{Pre-exponential factor $D_0$ and activation energy $Q$ of self-diffusion coefficient for Al with and without H. A comparison with data from the literature is also provided \cite{Mantina2008,Lundy1962,Volin1968,Messer1975}. * are for experimental data and ** are for \textit{ab initio}{} calculations} \label{tblD0QDAl} \begin{tabular}{cccc} \hline \ & $D_0$ ($10^{-6}$ m$^2$.s$^{-1}$) & $Q$ (eV) & $T$ (K) \\ \hline * \cite{Lundy1962} & 171 & 1.48 & 720-920 \\ * \cite{Volin1968} & 17.6 & 1.31 & 360-480 \\ * \cite{Messer1975} & 13.7 & 1.28 & 550-750 \\ ** \cite{Mantina2008} & 7.75 & 1.26 & \ \\ Without H & 2.41 & 1.23 & \ \\ With H in $P_1$ & 2.41 & 2.13 & \ \\ With H in $P_3$ & 2.41 & 1.66 & \ \\ \hline \end{tabular} \end{table} Although the interaction enthalpy between a vacancy and a H atom is negative, these calculations show that in pure Al, hydrogen delays the diffusion of vacancy by increasing the vacancy migration enthalpy (when H is in $P_1$ only) and vacancy formation enthalpy (when H is in $P_1$ and $P_3$). Thus, hydrogen should also impact the diffusion of copper, as studied in the next section. \subsection{Copper diffusion in the vicinity of hydrogen-vacancy complexes in Al matrix} \label{S42} The influence of hydrogen on the diffusion coefficient of copper ($D_{Cu}$) is investigated in this section using the five jump frequency model developed by LeClaire \cite{LeClaire1978}. The model includes several assumptions: (i) the solute diffusion is controlled by a vacancy mechanism, (ii) the interaction between solute and vacancy is only limited to monovacancy being in the nearest neighbour of the solute and (iii) Cu atoms do not interact with other Cu atoms (diluted system). It has been demonstrated that such model accurately describes the diffusion of solute in aluminium \cite{Mantina2009a,Mantina2009b} or in cobalt \cite{Naghavi2017}. This approach considers that solute atoms diffuse predominantly with a vacancy mechanisms through the five jump frequencies $\omega_i, i = 0,4$ defined by \cite{LeClaire1978}: \begin{equation} \omega_i(T) = \nu_i\exp{-\frac{H_{i}^m}{k_BT}}. \label{eqwi} \end{equation} $\omega_1$ corresponds to the jump frequency for Al atom-vacancy jumps between a pair of sites that are both nearest neighbours of Cu atom. $\omega_2$ is the jump frequency for Cu atom-vacancy exchange. $\omega_3$ is the jump frequency of Al, which dissociates the vacancy and Cu. $\omega_4$ is the opposite of $\omega_3$ (\textit{i.e.}{}: jump frequency of Al atom, which binds Cu and the vacancy). The fifth jump is $\omega_0$, the self-diffusion of Al (see eq. (\ref{eqw0}) in section \ref{S32}). The attempt frequency $\nu_i$ is equal $\sqrt(3/5)k_B/h \theta_D$ (like in the previous section), with $\theta_D$ the Debye temperature of aluminium or copper depending on the atom that is exchanged with the vacancy. \begin{figure}[bth!] \centering \includegraphics[width=0.67\linewidth]{FigCuVac.eps} \caption{Illustration of 4 atom jumps based on the LeClaire model (the fifth jump, $\omega_0$ correspond to the jump frequency of an Al atom without copper).} \label{FigCuVac} \end{figure} NEB calculations are performed to determine the migration energies of the different jumps without H and with H in $P_1$ and $P_3$ positions. Like in section \ref{S41}, these positions are chosen over $P_2$ in order to have H either still associated with the vacancy before and after the jump or dissociated from it . Besides, all calculations are performed at zero pressure, therefore it is also assumed that the migration enthalpies $H_{i}^m$ are equivalent to the migration energies $E^m_{i}$ \cite{Carling2003,Naghavi2017}. Results are plotted in fig. \ref{Figwi}. When H is still associated to the vacancy, the migration energy is always larger than when the atom is exchanged with a hydrogen free vacancy. When H is the furthest from the mobile atom, H is dissociated from the vacancy and the final state has an higher energy than its initial state corresponding to the interaction energy between H and vacancy. Consequently, when H is in $P_1$, $E^m_4$ is lower than $E^m_3$, while it is symmetrical when the vacancy is not linked to a hydrogen atom. \begin{figure}[bth!] \centering \includegraphics[width=\linewidth]{Figwi.eps} \caption{Energy variation of atoms corresponding to the different jump frequencies for vacancy and hydrogen vacancy complexes. a. Energy migration of Al that does not dissociate copper with the vacancy. b. Energy migration of Cu in Al matrix. c. Energy migration of Al that dissociates (orange) and associates (dark-magenta) the vacancy with Cu atom.} \label{Figwi} \end{figure} Then, the diffusion coefficient of copper $D_{Cu}$ is obtained using \cite{LeClaire1978}: \begin{equation} D_{Cu}(T) = f_2\omega_2a(T)^2C_2^{j}(T), \label{eqDCu} \end{equation} with $f_2$ a correlation factor, similar to $f_0$ in eq. (\ref{eqDAl}), which is written \cite{Manning1964}: \begin{equation} f_2 = \frac{2\omega_1+\omega_3 F(\alpha)}{2\omega_2+2\omega_1+\omega_3 F(\alpha)}, \label{eqf2} \end{equation} where $\alpha=\omega_4/\omega_0$ and \begin{multline} F(\alpha) = 7 - \left(\frac{10\alpha^4+180.5\alpha^3+927\alpha^2+1341\alpha}{2\alpha^4+40.2\alpha^3+254\alpha^2+597\alpha+436}\right). \label{eqF2} \end{multline} The concentration $C^{j}_2(T)$ is similar to $C^{j}_0(T)$ of eq. (\ref{eqDAl}) and is written as: \begin{equation} C^{j}_{2}(T) = \exp\left(-\frac{H_{j}^f - TS_{i}^f + \Delta G_b}{k_BT} \right), \label{eqC2} \end{equation} with the $\Delta G_b$, corresponding to the binding energy between copper and vacancy. This energy is linked to the concentration of copper atoms near vacancies and can be determined using the jump frequencies ratio within the assumptions detailed in the beginning of this section \cite{LeClaire1978}: \begin{equation} \frac{\omega_3}{\omega_4} = \exp{\left(-\frac{\Delta G_b}{k_B T}\right)}. \label{eqDGb} \end{equation} Identical to section \ref{S41}, the concentration and migration energy of hydrogen-vacancy complexes are calculated for systems having hydrogen in $P_1$ and $P_3$ and the resulting coefficient diffusion of copper is plotted in fig. \ref{FigDCu}. \begin{figure}[bth!] \centering \includegraphics[width=\linewidth]{FigDCu.eps} \caption{Calculated diffusion coefficient of Cu as a function of T with and without H compared to previous work of Mantina \textit{et al}{} \cite{Mantina2009a}, Du \textit{et al}{} \cite{Du2003}, Peterson and Rothman \cite{Peterson1970}, Anand \textit{et al}{} \cite{Anand1965} and Murphy \cite{Murphy1961}}. \label{FigDCu} \end{figure} While a slightly higher diffusion coefficient is obtained at low temperature compared to literature data, relatively consistent values are obtained at high temperatures ($T<500\,K$). Moreover, the diffusion coefficient of copper is systematically lower in presence of hydrogen. Assuming that hydrogen atoms stay in its interstitial site, the effect is more important when H is far from the copper atom (in $P_3$) than when it is close to Cu (in $P_1$). This result is obtained because when H is far from Cu, the energy barrier to associate the hydrogen vacancy complex ($E^m_4$) or to dissociate it with the Cu atom ($E^m_3$) is large. Besides, when Cu atom is exchanged with the vacancy, it dissociates H-vacancy complex (which is less stable than the initial state). These effects have a more important impact on the diffusion coefficient than having hydrogen in $P_1$, which increases the migration energy of copper $E^m_2$. Assuming that the diffusion of copper also follows an Arrhenius type temperature dependence (eq. \ref{eqDAlArr}), $D_0$ and $Q$ can also be estimated for Cu in pure Al, by fitting curves of fig. \ref{FigDCu}. The fitted results are displayed in table \ref{tblD0QDCu}. Without hydrogen, the calculated data are significantly lower than those reported in the literature. An important increase of the activation energy is also obtained when hydrogen is linked to the vacancies, but the variation of the activation energy is less pronounced for $D_{Cu}$ than for $D_{Al}$. \begin{table}[h!] \centering \caption{Pre-exponential factor $D_0$ and activation energy $Q$ of the diffusion coefficient of Cu in FCC Al with and without H. A comparison with data from the literature is also provided \cite{Mantina2009a,Du2003,Peterson1970,Anand1965,Murphy1961}. * are for experimental data and ** are for \textit{ab initio}{} calculations.} \label{tblD0QDCu} \begin{tabular}{cccc} \hline \ & $D_0$ ($10^{-6}$ m$^2$.s$^{-1}$) & $Q$ (eV) & $T$ (K) \\ \hline * \cite{Murphy1961} & 29 & 1.35 & 777-908 \\ * \cite{Anand1965} & 15 & 1.31 & 623-903 \\ * \cite{Peterson1970} & 65 & 1.40 & 594-928 \\ * \cite{Du2003} & 44 & 1.39 & \ \\ ** \cite{Mantina2009a} & 4.4 & 1.25 & \ \\ Without H & 1.9 & 1.16 & \ \\ With H in $P_1$ & 1.7 & 1.48 & \ \\ With H in $P_3$ & 1.7 & 1.59 & \ \\ \hline \end{tabular} \end{table} \section{Discussion} \label{S5} The experiments of section \ref{S2} show a delay of the formation and growth of GP zones when hydrogen is introduced in solid solution. In the following section, a classical strength model is applied to correlate the microstructure evolution observed through HAADF-STEM with the hardness measurements. Then, results from experiments are compared with the diffusion coefficients obtained in section \ref{S3} and discussed further. \subsection{Relationship between structure evolution and hardness} \label{S51} Classical strengthening model is applied to establish a relationship between the hardness measurements and the HAADF-STEM observations (model developed in previous work for Al alloys \cite{Myhr2001,Bardel2014,Rodriguez2018,Bellon2020}). It is assumed that the microhardness is linked to the yield stress $\sigma_Y$ through a Tabor factor (T = 2.8) \cite{Hutchings2009} and $\sigma_Y$ is linked to the stress needed to induce dislocations or to make them mobile (\textit{i.e.}{}: the critical resolved shear stress) $\tau_Y$ through a Taylor factor (M = 3.1) \cite{Bardel2014,Rodriguez2018}: \begin{equation} \tau_Y = \frac{T}{M}HV. \label{eqHVty} \end{equation} The stress $\tau_Y$ can be determined from microstructural features and additive contributions; which may be considered as a first approximation as \cite{Myhr2001,Bardel2014,Rodriguez2018}: \begin{equation} \tau_Y = \tau_0 + \tau_d + \tau_{gb} + \tau_{ss} + \tau_P \label{eqty} \end{equation} where $\tau_0$ is the friction stress, $\tau_d$ the forest hardening due to dislocations, $\tau_{gb}$ the stress contribution of the grain boundaries, $\tau_{ss}$ the stress contribution Cu of solid solution and $\tau_P$ the stress contribution of particles. The stress contributions $\tau_0$, $\tau_d$ and $\tau_{gb}$ are assumed constant for all ageing times and they are determined from hardness measurements of the alloy naturally aged 10\,minutes, assuming that all the copper atoms are in solid solution. The stress contribution $\tau_{ss}$ may be written \cite{Rodriguez2018,Bellon2020}: \begin{equation} \tau_{ss} = HX_{Cu}^{n}, \label{eqtss} \end{equation} with $X_{Cu}$ the mass fraction of Cu in solid solution in the Al matrix. $H$ and $n$ two constants equal to 7.2\,MPa and 1, respectively \cite{Zhu2008,Bellon2020}. Then, the concentration of copper in solid $X_{Cu}$ is updated for each state by knowing the volume fraction of precipitates (Cu clusters and GP zones) inducing a variation of $\tau_{ss}$. When the alloy is naturally aged, Cu atoms agglomerate as clusters then quickly form GP zones. According to the HAADF-STEM images, the kinetic is slower when the alloy is in contact of hydrogen prior to ageing in air. Both GP zones and Cu clusters affect hardness due to interactions with dislocations. Since they are nanoscaled with a diameter $d_t$ inferior to $d_C = 10 \rm{nm}$, a critical diameter below which dislocations shear these particles (and when the particles are larger than $d_C$, they are by-passed by dislocations) \cite{Myhr2001}, the shear stress $\tau_P$ for a moving dislocation writes as \cite{Myhr2001,Rodriguez2018}: \begin{equation} \tau_{P} = \frac{1}{b}\sqrt{\frac{3 f_V}{2 \pi}}\left(0.72 \mu b^2 \right) \left(\frac{2}{d_C}\right)^{1.5}\left(\frac{d_t}{2}\right)^{0.5}, \label{eqtp} \end{equation} with $b$ and $\mu$, the Burgers vector and shear modulus equal to 0.286 nm and 27 GPa \cite{Myhr2001}, respectively. $d_t$ and $f_V$ are experimental data (table \ref{tblSTEMNA}) and they are used to determine $\tau_P$ for each condition. It is important to note that eq. (\ref{eqtp}) was established for spherical particles, which is clearly not the case for GP zones. However, previous studies have shown that this equation gives also consistent results for GP zones \cite{Zander2008,Rodriguez2018}. Fig. \ref{FigTauY} shows the variations of $\tau_Y$ estimated from hardness measurements and eq. (\ref{eqHVty}). These estimates are compared to values calculated from model, eqs. (\ref{eqty}-\ref{eqtp}), with only the contribution of GP zones and with the contribution of GP zones and Cu clusters. The shear stresses obtained from the strength model exhibit the same trend than values estimated from hardness measurements: (i) $\tau_Y$ increases with the ageing time and (ii) $\tau_Y$ is always larger when the alloy is aged in air than when it is stored in NaOH prior to ageing in air. However, the best match is obtained when both clusters and GP zones are considered. This result confirms the delay of GP zone nucleation and growth due to hydrogen incorporation. \begin{figure}[bh!] \centering \includegraphics[width=\linewidth]{FigTauY.eps} \caption{Variation of $\tau_Y$ during natural ageing of AlCu, obtained from the hardness measurements and from the strength model considering GP zones only and including particles appearing as copper clusters.} \label{FigTauY} \end{figure} \subsection{Comparison between experiments and calculations} \label{S52} The microhardness measurements show a delay of the hardening kinetic when the alloy is stored in 5\,h in NaOH. The HAADF-STEM observations confirm this delay and even after being naturally aged 9 days, microstructures are still different when the alloy is directly aged in air or stored 5\,h in NaOH. This difference can be due to hydrogen reducing the diffusion of copper or/and an increase of the energy barrier to form GP zones, locking copper agglomerates as clusters. Since hydrogen strongly interacts with the excess vacancies \cite{Fukai2006,Connetable2018,Du2020,Hachet2022b}, our numerical study focused on the effect of hydrogen on the diffusion of the copper in aluminium through \textit{ab initio}{} calculations. According to fig. \ref{FigDCu}, the ratio of diffusion coefficient $D^{With H}_{Cu}/D^{Without H}_{Cu}$ is $\sim10^{-6}$ at 300\,K. Defining the effective diffusion length of copper as $\lambda_{Cu} = \sqrt{6 D_{Cu} t}$, then $\lambda^{With H}_{Cu}/\lambda^{Without H}_{Cu} \sim 10^{-3}$. The HAADF-STEM images of fig. \ref{FigSTEMNA} show that the structure is similar in the alloy directly aged 1\,h in air and in the alloy stored 5\,h in NaOH and aged 9 days in air. Assuming that this difference is only due to the impact of hydrogen on the diffusion coefficient, then $\lambda^{With H}_{Cu}(t=9\,d) = \lambda^{Without H}_{Cu}(t=1\,d)$ and $D^{With H}_{Cu}/D^{Without H}_{Cu} \sim 10^{-1}$. This is consistent but much lower than results from \textit{ab initio}{} calculations. However, during the first hours in NaOH, hydrogen diffuses in Al-5Cu and some time is required to create H-vacancy complexes, thus part of them annihilate before the creation of a hydrogen-vacancy complex. Besides, during the ageing in air after the 5\,h in NaOH, hydrogen may desorb from the alloy, which probably reduces the quantity of hydrogen-vacancy complexes. These two features may explain the difference \d{of diffusion coefficient} between experimental observations and calculations. Nevertheless, the \textit{ab initio}{} calculations are focused only on single H-vacancy complexes while previous work has shown that depending on the hydrogen content, hydrogen-vacancy complexes with several hydrogen atoms may be more stable than single H-vacancy complexes \cite{Gunaydin2008}. However, these H atoms would be placed in additional tetrahedral interstitial sites of the vacancy, decreasing even more the diffusion coefficient and may lead to a non-Arrhenius behavior of the diffusion coefficient (as observed for $D_{Al}^H$ in ref. \cite{Gunaydin2008}). \textit{Ab initio}{} calculations revealed that hydrogen and vacancy are attractive, even with copper in the vicinity of the vacancy (and H is more stable far from Cu atoms, in positions $P_2$ and $P_3$). The previous calculations for pure Al (sec. \ref{S41}) show that hydrogen affects the mobility of vacancy and thus should also impact the diffusion of copper as confirmed in section \ref{S42}. However, it is important to note that these \textit{ab initio}{} calculations have some limitations. It has been established that hydrogen enhances the formation of vacancies in aluminium \cite{Fukai2006,Xie2016} even though an increase of the formation enthalpy of hydrogen-vacancy complex is obtained (which is also in agreement with the literature \cite{Xie2016,Connetable2018}). This is because \textit{ab initio}{} calculations focus on monovacancies bound exclusively to one hydrogen atom. According to previous \textit{ab initio}{} calculations, up to 10\,H atoms can be incorporated in one vacancy (12\,H atoms for \cite{Lu2005} and 13\,H for \cite{Nazarov2014}). While increasing the number of H-trapped atoms in one vacancy may increases the formation enthalpy of the hydrogen-vacancy complex \cite{Connetable2018}, several hydrogen atoms included in one vacancy can form $H_2$ molecules. This might reduce the formation energy of these complexes or lead to divacancies, which have a stronger attraction with H \cite{Nazarov2014}. In the situation where several H are trapped in vacancies or divacancies, the complex may reduce its formation energy considerably and would not impact the migration energy of the vacancy (\textit{e.g.}{}: when H atoms are far of the moving atom). Theoretically, even an increase of the self-diffusion due to H can not be excluded. Such increase of self-diffusion has been reported for nickel where the formation energies of hydrogen-vacancy follows the superabundant vacancy model with up to 6\,H atoms inserted \cite{Metsue2016}. However, the increase of the migration energy of these complexes due to H incorporation is smaller \cite{Wang2015}. As a consequence, when one hydrogen atom is incorporated in one vacancy, hydrogen delays its diffusion, but the opposite may occur when more hydrogen atoms are incorporated in one vacancy, and an increase of the vacancy mobility can even be obtained \cite{Du2020}. Similar mechanisms cannot be excluded in the diluted Al-Cu system, but experimental data show that Cu diffusion is significantly reduced, thus based on our \textit{ab initio}{} calculations, single H-vacancy complexes are probably dominant. \section{Conclusions} \label{S6} This present work studied the consequences of hydrogen on the GP zone formation and growth in an Al-5Cu alloy. When it is naturally aged in air, excess vacancies diffuses and are annihilated in residual dislocations and grain boundaries, allowing copper to diffuse leading to the nucleation of GP zones, which that harden the alloys. When the alloy is in contact with hydrogen during the beginning of the natural ageing, a delay of the hardening kinetics is noted with a change of the microstructure observed through HAADF-STEM. According to the \textit{ab initio}{} calculations performed in this study, hydrogen-vacancy complexes are stable in the vicinity of Cu in a diluted Al-Cu system. Therefore, these complexes \d{(containing one or more H atoms)} affect the diffusion of Cu in Al matrix leading to a delayed nucleation and growth of GP zones. Modelling at atomic scale the self-diffusion for Al and the diffusion of Cu in Al in the presence of hydrogen-vacancy complexes highlights two phenomenons : (i) hydrogen in the path of the atom that exchange with the vacancy will increase the energy barrier and (ii) hydrogen far from this atom will be dissociated from the vacancy and it will be less stable after the atom jump. As a consequence, the diffusion coefficient of copper is always smaller in presence of hydrogen close to or trapped in the vacancy. In addition, even if hydrogen always reduces the mobility of Cu in diluted Al-Cu, it was demonstrated that the position of H in the lattice of a diluted Al-Cu alloy influences the diffusion coefficient. Therefore, statistical study using techniques such as kinetic Monte-Carlo simulations would be performed in the near future to determine the evolution of the diffusion coefficient of copper as a function of the hydrogen concentration. This statistical study would also help determining the number of H atoms insides a vacancy for a given temperature and H concentration, which is an unknown parameter in the present work. Finally, these results are obtained without studying the consequences of hydrogen on the precipitate/matrix interface, which will also be the object of a future study. \section*{Acknowledgments} The authors thanks Dr. D. Embury, Dr. A. Saiter-Fourcin, Dr. F. Vurpillot and G. Da Costa for fruitful discussions. The authors acknowledge the financial support of the French Agence Nationale de la Recherche (ANR), through the program “Investissements d’Avenir” (ANR-10-LABX-09-01), LabEx EMC3, and the région NORMANDIE. This work was partially supported by the CRNS Federation IRMA - FR 3095. The calculations are performed using the computing ressources of CRIANN (Normandy, France) whithin the framework of the project No. 2021016.

80fed43bb08b102bf9c878c5fea1830cd13b3028

\section{Introduction} Ultra-reliable and low-latency communication (URLLC) is one of the generic applications required to be covered in the fifth-generation (5G) \cite{durisi2016toward,bennis2018ultrareliable,zhang2021ris}. As a result, it has been attracted significant interests since it enables several innovative usages, especially in industrial production, such as remote heavy industrial machines operation and factory automation \cite{simsek20165g,liu2018tractable}. However, compared with conventional communication systems, the achievable rate under URLLC is quite different since short blocklength is adopted to shorten the latency such that the classical Shannon-sense capacity no longer holds. Specifically, the URLLC rate is a complicated function of the transmission power, the precoding vector, the bandwidth, the transmission time, and the decoding error probability \cite{polyanskiy2010channel}. Indeed, guaranteeing URLLC represents unique challenges to resource allocation design due to the non-convexity introduced by the finite blocklength. In the literature, much attention has been devoted to designing effective resource allocation algorithms that support URLLC \cite{she2018joint,nasir2020resource,he2021beamforming}. However, the systems considered in these works are all cellular networks and their performance is known to be limited by severe inter-cell interference. Cell-free massive multiple-input multiple-output (MIMO) architecture is a new promising solution to overcome the issue discussed above \cite{zhang2021local,zhang2021improving,zheng2022cell,9737367}. It reaps the advantages of massive MIMO and network MIMO, since massive distributed access points (APs) facilitate coherent signal transmission to serve all the users without any cell boundaries \cite{bjornson2019making,ngo2017cell,zhang2020prospective}. However, current literature focuses on the resource allocation in cell-free massive MIMO systems for URLLC is still limited. For example, in \cite{nasir2021cell}, the authors applied the path-following algorithm (PFA) for optimizing the power allocation with a special class of conjugate beamforming to maximize the users' minimum URLLC rate and the energy efficiency. However, an adaptive and optimized precoding design at the APs is generally more effective that the fixed one. Besides, in \cite{lancho2022cell}, the upper bounds of the uplink and downlink decoding error probabilities (DEPs) were derived by using the saddlepoint method to support URLLC. While the closed-form expression of DEP can characterize the performance, it is generally intractable for the the design of cooperatively efficient resource allocation. As such, there is an emerging need for designing the precoding with the performance metric of the URLLC rate. Motivated by the above discussion, the PFA-based precoding design for maximizing the users' minimum URLLC rate is studied in this correspondence. First, a PFA-based centralized precoding design is proposed which generates a sequence of feasible points and converges to a locally optimal solution of the design optimization problem. Second, we propose a decentralized PFA-based precoding design by dividing the APs into several non-overlapping cooperative clusters in which the APs only share the data and instantaneous channel state information (CSI) in each cluster to design the precoding vectors to reduce the computational complexity. Simulation results show that compared with the centralized precoding, the decentralized PFA precoding can achieve 80\% of the 95\%-likely URLLC rate and 89\% of the average URLLC rate with only 12\% of the computational complexity of the counterpart. We also investigate the impact of the precoding schemes, the length of transmission duration, and the size of the AP cluster on the URLLC rate via extensive simulations. \section{System Model} We consider a cell-free massive MIMO system, which consists of $L$ APs and $K$ single-antenna users that are distributed arbitrarily over a large area. We assume that each AP is equipped with $N$ antennas. Moreover, all the APs are connected with each other and a central processing unit (CPU) via dedicated fronthaul links with sufficient capacity. All APs serve all users on the same time-frequency resource through time division duplex (TDD) operation \cite{9743355}. The channel coefficient between AP $l$ and user $k$, ${{\bf{h}}_{kl}} \in {{\mathbb{C}}^{N \times 1}}$, is assumed to follow a correlated Rayleigh fading distribution. We adopt a classic block fading model for modeling the channels such that ${{\bf{h}}_{kl}}$ remains constant in $t$ channel uses of the time-frequency blocks and experience an independent realization in every block. Note that the channel coefficients can be acquired at the APs by existing channel estimation algorithms \cite{bjornson2017massive} and this is beyond the scope of this work as we aim to optimize the precoding for URLLC. Therefore, we assume that perfect CSI is available at the APs. In the downlink payload data transmission phase, the received signal at user $k$ can be expressed as ${y_k} =\sum\limits_{l = 1}^L {\bf{h}}_{kl}^H{{\bf{w}}_{kl}}{s_k} + \sum\limits_{l = 1}^L {{\bf{h}}_{kl}^H} \sum\limits_{i \ne k}^K {{\bf{w}}_{il}}{s_i} + {n_k}$, where ${s_i} \sim {{\cal N}_{\mathbb{C}}}\left( {0,1} \right)$ at AP $l$, ${{\bf{w}}_{il}} \in {{\mathbb{C}}^{N \times 1}}$ is the precoding vector for user $i$ at AP $l$, and ${n_k} \sim {{\cal N}_{\mathbb{C}}}\left( {0,{\sigma ^2}} \right)$ represents the thermal noise at user $k$. Then, the corresponding effective signal-to-interference-plus-noise ratio (SINR) is given as \begin{equation} {\varphi _k} = \frac{{{{\left| {{\bf{ h}}_k^H{{\bf{w}}_k}} \right|}^2}}}{{\sum\limits_{i \ne k}^K {{{\left| {{\bf{ h}}_k^H{{\bf{w}}_i}} \right|}^2}} + {\sigma ^2}}}, \end{equation} where ${{\bf{h}}_k} = {\left[ {{\bf{h}}_{k1}^H, \cdots ,{\bf{h}}_{kL}^H} \right]^H} \in {{\mathbb{C}}^{LN \times 1}}$ and ${{\bf{w}}_i} = {\left[ {{\bf{w}}_{i1}^H, \cdots ,{\bf{w}}_{iL}^H} \right]^H} \in {{\mathbb{C}}^{LN \times 1}}$. By treating the inter-user interference ${\bf{h}}_{kl}^H\sum\limits_{i \ne k}^K {{{\bf{w}}_{il}}{s_i}}$ as Gaussian noise, where $p_{il}^{{\rm{dl}}} \buildrel \Delta \over = {\left\| {{{\bf{w}}_{il}}} \right\|^2}$ is the power allocated to user $i$ at AP $l$, the achievable rate in nats/sec/Hz for user $k$ for the case of sufficiently long blocklength is given by the Shannon rate function ${{\tilde R}_k}= \ln \left( {1 + {\varphi _k}} \right)$, and the achievable URLLC rate in nats/sec/Hz for user $k$ can be approximated as \cite[eq. (30)]{nasir2020resource} \begin{equation}\label{URLLC Rate} {R_k} = \ln \left( {1 + {\varphi _k}} \right) - \sqrt {\frac{1}{{tB}} \times {V_k}} \times {Q^{ - 1}}\left( {\epsilon} \right), \end{equation} where $t$ is the transmission duration, $B$ is the communication bandwidth, ${V_k}$ is the channel dispersion \cite{nasir2020resource} which can be expressed as ${V_k} = 1 - \frac{1}{{{{\left( {1 + {\varphi _k}} \right)}^2}}}$, ${Q^{ - 1}}\left( \cdot \right)$ is the inverse of the Gaussian Q-function, i.e., $Q\left( x \right) = \int_x^\infty {\frac{1}{{\sqrt {2\pi } }}\exp \left( { - {t^2}/2} \right)} dt$, and ${\epsilon}$ is the decoding error probability. Note that (\ref{URLLC Rate}) is the normal approximation when the channel ${{\bf{h}}_k}$ is assumed to be quasi-static and deterministic over the transmission duration $t$. The subtrahend in (\ref{URLLC Rate}) captures the rate penalty due to the finite block length, $tB$. \section{Max-min Rate Based Precoding Design}\label{Design} \subsection{Centralized Precoding Design} In the centralized precoding design, the optimization of the precoding vectors takes place at the CPU, where the estimate of the global instantaneous CSI ${{{\bf{h}}}_{kl}},\forall k \in \left\{ {1, \cdots ,K} \right\},\forall l \in \left\{ {1, \cdots ,L} \right\}$, available. The centralized max-min URLLC rate optimization problem can be expressed as \begin{align} &\mathop {\max }\limits_{\bf{w}} \mathop {\min }\limits_{k = 1, \cdots ,K} \left\{ {{R_k}\left( {\bf{w}} \right)} \right\}\label{P1}\tag{3a}\\ &{\rm{s.}}{\rm{t.}}\;\;\;\;\;\;\sum\limits_{k = 1}^K {{{\left\| {{{\bf{w}}_{kl}}} \right\|}^2}} \le {p_{\max }},\forall l,\label{3b}\tag{3b} \end{align} where ${\bf{w}} = \left\{ {{{\bf{w}}_{kl}}:k = 1, \cdots ,K,l = 1, \cdots ,L} \right\}$ and ${p_{\max }}$ is the maximum power at each AP. The problem (\ref{P1}) is non-convex due to the URLLC rate function ${R_k}\left( {\bf{w}} \right)$. With the help of \cite{nasir2020resource}, we apply the PFA to develop a concave lower bound for ${R_k}\left( {\bf{w}} \right)$. Without loss of generality, the URLLC rate expression for user $k$ can be rewritten as ${R_k}\left( {\bf{w}} \right) = {f_k}\left( {\bf{w}} \right) - a{g_k}\left( {\bf{w}} \right)$, where $a = {Q^{ - 1}}\left( {\epsilon } \right)/\sqrt {t{ B}}$, ${f_k}\left( {\bf{w}} \right) = \ln \left( {1 + {\varphi _k}\left( {\bf{w}} \right)} \right)$, and ${g_k}\left( {\bf{w}} \right) = \sqrt {1 - 1/{{\left( {1 + {\varphi _k}\left( {\bf{w}} \right)} \right)}^2}}$. Now, we aim to establish a convex lower bound for ${f_k}\left( {\bf{w}} \right)$ and a concave upper bound for ${g_k}\left( {\bf{w}} \right)$. Let ${{\bf{w}}^{\left( n \right)}}$ be a feasible point for (\ref{P1}) that is computed from the $\left( {n - 1} \right)$th iteration of the iterative PFA. \subsubsection{Lower bounding for ${f_k}\left( {\bf{w}} \right)$} According to \cite{nasir2020resource}, the following inequality holds for all ${\bf{x}} \in {{\mathbb{C}}^{{M_1}}},{\bf{y}} \in {{\mathbb{C}}^{{M_2}}}$ and ${\bf{\bar x}} \in {{\mathbb{C}}^{{M_1}}},{\bf{\bar y}} \in {{\mathbb{C}}^{{M_2}}}$ \begin{align}\label{I1}\tag{4} \ln\! \left( \!\!{1\!\! +\! \frac{{{{\left\| {\bf{x}} \right\|}^2}}}{{{{\left\| {\bf{y}} \right\|}^2} \!\!+\! {\sigma ^2}}}}\!\! \right) \!\!\ge\!\! a\!\! -\! \frac{{{{\left\| {{\bf{\bar x}}} \right\|}^2}}}{{2{\cal R}\!\!\left\{ {{{{\bf{\bar x}}}^H}{\bf{x}}} \right\} \!\!- \! {{\left\| {{\bf{\bar x}}} \right\|}^2}}}\!-\! b{\left\| {\bf{x}} \right\|^2} \!\!-\! c{\left\| {\bf{y}} \right\|^2}. \end{align} Applying the inequality in (\ref{I1}) for $x = {\bf{h}}_k^H{{\bf{w}}_k}$, $y = {{\cal L}_k}\left( {\bf{w}} \right)$, $\bar x = {\bf{h}}_k^H{\bf{w}}_k^{\left( n \right)}$, $\bar y = {{\cal L}_k}\left( {{{\bf{w}}^{\left( n \right)}}} \right)$, where ${{\cal L}_k}\left( {\bf{w}} \right)$ arranges ${\bf{h}}_k^H{{\bf{w}}_i},i \ne k$ into a vector of dimension $K-1$, we obtain \begin{align}\label{f_n} {f_k}\left( {\bf{w}} \right) &\ge \bar a_k^{\left( n \right)} - \frac{{{{\left| {{\bf{h}}_k^H{\bf{w}}_k^{\left( n \right)}} \right|}^2}}}{{2{\cal R}\left\{ {{{\left( {{\bf{w}}_k^{\left( n \right)}} \right)}^H}{{{\bf{h}}}_k}{\bf{h}}_k^H{{\bf{w}}_k}} \right\} - {{\left| {{\bf{h}}_k^H{\bf{w}}_k^{\left( n \right)}} \right|}^2}}}\notag\\ &- \!\bar b_k^{\left( n \right)}{\left| {{\bf{h}}_k^H{{\bf{w}}_k}} \right|^2} \!-\! \bar c_k^{\left( n \right)}\sum\limits_{i \ne k} {{{\left| {{\bf{h}}_k^H{{\bf{w}}_i}} \right|}^2}} \!\buildrel \Delta \over = \!f_k^{\left( n \right)}\left( {\bf{w}} \right)\tag{5}, \end{align} with the constraint of \begin{equation}\label{trust region}\tag{6} 2{\cal R}\left\{ {{{\left( {{\bf{w}}_k^{\left( n \right)}} \right)}^H}{{{\bf{h}}}_k}{\bf{h}}_k^H{{\bf{w}}_k}} \right\} - {\left| {{\bf{h}}_k^H{\bf{w}}_k^{\left( n \right)}} \right|^2} > 0, \end{equation} where $\bar a_k^{\left( n \right)} \!=\! {f_k}\left( {{{\bf{w}}^{\left( n \right)}}} \right) \!+\! 2 \!-\! \frac{{{{\left| {{\bf{h}}_k^H{\bf{w}}_k^{\left( n \right)}} \right|}^2}}}{{\beta _k^{\left( n \right)}}}\frac{{\sigma ^2}}{{\alpha _k^{\left( n \right)}}}$, $0 \!< \!\bar b_k^{\left( n \right)} \!=\! \frac{{\bar a_k^{\left( n \right)}}}{{\beta _k^{\left( n \right)}{{\left| {{\bf{h}}_k^H{\bf{w}}_k^{\left( n \right)}} \right|}^2}}}$, $0 \!<\! \bar c_k^{\left( n \right)} \!= \!\frac{{{{\left| {{\bf{h}}_k^H{\bf{w}}_k^{\left( n \right)}} \right|}^2}}}{{\beta _k^{\left( n \right)}\alpha _k^{\left( n \right)}}}$, $\alpha _k^{\left( n \right)} \!\buildrel \Delta \over =\! \sum\limits_{i \ne k} \!{{{\left| {{\bf{h}}_k^H{\bf{w}}_i^{\left( n \right)}} \right|}^2}} \!+\! {{\sigma ^2}}$, and $\beta _k^{\left( n \right)} \!\buildrel \Delta \over = \! \sum\limits_{i = 1}^K {{{\left| {{\bf{h}}_k^H{\bf{w}}_i^{\left( n \right)}} \right|}^2}} \!+\! {{\sigma ^2}}$. According to \cite{nasir2020resource}, the function $f_k^{\left( n \right)}\left( {\bf{w}} \right)$ is concave over the trust region (\ref{trust region}) and achieves the same value as ${f_k}\left( {\bf{w}} \right)$ at ${{{\bf{w}}^{\left( n \right)}}}$, $f_k^{\left( n \right)}\left( {{{\bf{w}}^{\left( n \right)}}} \right) = {f_k}\left( {{{\bf{w}}^{\left( n \right)}}} \right)$. \subsubsection{Upper bounding for ${g_k}\left( {\bf{w}} \right)$} Since the function $f\left( x \right) = \sqrt x$ is concave on $x > 0$, the following inequality for all $x > 0$ and $\bar x > 0$ holds true \begin{align}\label{I2}\tag{7} \sqrt x \!=\! f\left( x \right)\le f\left( {\bar x} \right) \!+\! {\left. {\frac{{\partial \!f\!\left( x \right)}}{{\partial x}}} \right|_{x = \bar x}}\left( {x \!-\! \bar x} \right)\!=\! \frac{{\sqrt {\bar x} }}{2} \!+ \!\frac{x}{{2\sqrt {\bar x} }}, \end{align} where $\frac{{\partial f\left( x \right)}}{{\partial x}}$ refers to the partial derivative of the function $f\left( x \right)\le f\left( {\bar x} \right)$ with respect to $x$. Applying the inequality in (\ref{I2}) for $x = 1 - 1/{\left( {1 + {\varphi _k}\left( {\bf{w}} \right)} \right)^2}$ and $\bar x = 1 - 1/{\left( {1 + {\varphi _k}\left( {{{\bf{w}}^{\left( n \right)}}} \right)} \right)^2}$ and using \begin{align}\label{I3} &{{{{\left( {\sum\limits_{i \ne k} {{{\left| {{\bf{h}}_k^H{{\bf{w}}_i}} \right|}^2}} + {{\sigma ^2}}} \right)}^2}}}/{{{{\left( {\sum\limits_{i = 1}^K {{{\left| {{\bf{h}}_k^H{{\bf{w}}_i}} \right|}^2}} + {{\sigma ^2}}} \right)}^2}}}\notag\\ &\ge\!\!\! \frac{{4\alpha _k^{\left( n \right)}}}{{{{\left( {\beta _k^{\left( n \right)}} \right)}^2}}}\!\!\left( \!{\sum\limits_{i \ne k}\! {\left(\! {2{\cal R}\!\left\{ \!{{{\left( \! {{\bf{h}}_k^H{\bf{w}}_i^{\left( n \right)}} \!\right)}^*}{\bf{h}}_k^H{{\bf{w}}_i}} \right\} \!\!-\! {{\left| {{\bf{h}}_k^H{\bf{w}}_i^{\left( n \right)}} \!\right|}^2}} \!\right)} } { + {{\sigma ^2}}} \!\!\right)\notag\\ &-\!\! \frac{{2{{\left( \!{\alpha _k^{\left( n \right)}} \!\right)}^2}}}{{{{\left(\! {\beta _k^{\left( n \right)}} \!\right)}^3}}}\!\!\left( \!{\sum\limits_{i = 1}^K \!{{{\left| {{\bf{h}}_k^H{{\bf{w}}_i}} \right|}^2}} \!\! +\! {{\sigma ^2}}} \!\!\right)\!-\! \frac{{{{\left( \!{\sum\limits_{i \ne k}\! {{{\left| {{\bf{h}}_k^H{{\bf{w}}_i}} \right|}^2}} \!\!+\! {{\sigma ^2}}} \!\!\right)}^2}}}{{{{\left( {\beta _k^{\left( n \right)}} \right)}^2}}}\tag{8}, \end{align} with the constraints of \begin{align} &\sum\limits_{i = 1}^K {{{\left| {{\bf{h}}_k^H{{\bf{w}}_i}} \right|}^2}} + {{\sigma ^2}} \le 2\beta _k^{\left( n \right)},\label{cons_g1}\tag{9}\\ &\frac{1}{{{{\left( {\beta _k^{\left( n \right)}} \right)}^2}}}\left( {\sum\limits_{i = 1}^K {{{\left| {{\bf{h}}_k^H{{\bf{w}}_i}} \right|}^2}} + {{\sigma ^2}}} \right)\le \!\! \frac{2}{{\alpha _k^{\left( n \right)}}}\notag\\ &\;{\times}\!\!\left( \! {\sum\limits_{i \ne k}\! {\left( {2{\cal R}\!\left\{ {{{\left( {{\bf{h}}_k^H{\bf{w}}_i^{\left( n \right)}} \! \right)}^*}{\bf{h}}_k^H{{\bf{w}}_i}} \!\right\}} \right.} } {\left. { \!\!-\! {{\left| {{\bf{h}}_k^H{\bf{w}}_i^{\left( n \right)}} \right|}^2}} \right) \!\!+\! \!{{\sigma ^2}}} \right),\label{cons_g2}\tag{10} \end{align} we have \begin{align}\label{g_n} {g_k}\left( {\bf{w}} \right) &\le d_k^{\left( n \right)} - \frac{{4\alpha _k^{\left( n \right)}e_k^{\left( n \right)}}}{{{{\left( {\beta _k^{\left( n \right)}} \right)}^2}}}\left( {\sum\limits_{i \ne k} {\left( {2{\cal R}\left\{ {{{\left( {{\bf{h}}_k^H{\bf{w}}_i^{\left( n \right)}} \right)}^*}{\bf{h}}_k^H{{\bf{w}}_i}} \right\}} \right.} } \right. \notag\\ &\left.{\left. { - \!{{\left| {{\bf{h}}_k^H{\bf{w}}_i^{\left( n \right)}} \right|}^2}} \right) \!\!+\! {{\sigma ^2}}} \!\right)\!\! + \!\! \frac{{2{{\left( \!{\alpha _k^{\left( n \right)}} \!\right)}^2}e_k^{\left( n \right)}}}{{{{\left( {\beta _k^{\left( n \right)}} \right)}^3}}}\!\!\left( \!{\sum\limits_{i = 1}^K \!{{{\left| {{\bf{h}}_k^H{{\bf{w}}_i}} \right|}^2}} \!\! +\!\! {{\sigma ^2}}} \!\! \right)\notag\\ &+\frac{{{{\left( {\sum\limits_{i \ne k} {{{\left| {{\bf{h}}_k^H{{\bf{w}}_i}} \right|}^2}} + {{\sigma ^2}}} \right)}^2}e_k^{\left( n \right)}}}{{{{\left( {\beta _k^{\left( n \right)}} \right)}^2}}}\buildrel \Delta \over = g_k^{\left( n \right)}\left( {\bf{w}} \right)\tag{11}, \end{align} where $0 \!< \!d_k^{\left( n \right)} \!= \!\frac{{\sqrt {1 \!-\! 1/{{\left( {1 \!+\! {\varphi _k}\left( {{{\bf{w}}^{\left( n \right)}}} \right)} \right)}^2}} }}{2} \!+\! \frac{1}{{2\sqrt {1 \!-\! 1/{{\left( {1 \!+\! {\varphi _k}\left( {{{\bf{w}}^{\left( n \right)}}} \right)} \right)}^2}} }}$, and $0 < e_k^{\left( n \right)} = \frac{1}{{2\sqrt {1 - 1/{{\left( {1 + {\varphi _k}\left( {{{\bf{w}}^{\left( n \right)}}} \right)} \right)}^2}} }}$. The function $g_k^{\left( n \right)}\left( {\bf{w}} \right)$ is convex and achieves the same value as ${g_k}\left( {\bf{w}} \right)$ at ${{\bf{w}}^{\left( n \right)}}$, $g_k^{\left( n \right)}\left( {{{\bf{w}}^{\left( n \right)}}} \right) = {g_k}\left( {{{\bf{w}}^{\left( n \right)}}} \right)$. \subsubsection{Concave Lower bound for ${R_k}\left( {\bf{w}} \right)$} By applying (\ref{f_n}) and (\ref{g_n}), we have ${R_k}\left( {\bf{w}} \right) \ge f_k^{\left( n \right)}\left( {\bf{w}} \right) - ag_k^{\left( n \right)}\left( {\bf{w}} \right) \buildrel \Delta \over = R_k^{\left( n \right)}\left( {\bf{w}} \right)$, under the trust region constrained by (\ref{trust region}), (\ref{cons_g1}), and (\ref{cons_g2}). The function $R_k^{\left( n \right)}\left( {\bf{w}} \right)$ is concave and matches with the function ${R_k}\left( {\bf{w}} \right)$ at ${{\bf{w}}^{\left( n \right)}}$: \begin{equation}\label{R_n}\tag{12} {R_k}\left( {{{\bf{w}}^{\left( n \right)}}} \right) = R_k^{\left( n \right)}\left( {{{\bf{w}}^{\left( n \right)}}} \right). \end{equation} At the $n$th iteration, we solve the following convex problem with the computational complexity ${\cal O}\left( {{{\left( {LNK} \right)}^3}\left( {2K + 1} \right)} \right)$ to generate the next feasible point ${\bf{w}}^{\left( {n+1} \right)}$: \begin{equation}\label{P2}\tag{13} \mathop {\max }\limits_{\bf{w}} \mathop {\min }\limits_{k = 1, \cdots ,K} \left\{ {R_k^{\left( n \right)}\left( {\bf{w}} \right)} \right\} \;\;\;{\rm{s.}}{\rm{t.}}\;\;\;\text{(\ref{3b}),\;(\ref{trust region}),\;(\ref{cons_g1}),\;(\ref{cons_g2})}. \end{equation} According to (\ref{f_n}) and (\ref{g_n}), we can conclude that $\mathop {\min }\limits_{k = 1, \cdots ,K} {R_k}\left( {{{\bf{w}}^{\left( {n + 1} \right)}}} \right) \ge \mathop {\min }\limits_{k = 1, \cdots ,K} {R_k}\left( {{{\bf{w}}^{\left( n \right)}}} \right),\;\forall n$, which guarantees the monotonicity in convergence. According to \cite{nasir2020resource,nasir2021cell,xing2020matrix1,xing2020matrix2}, it is important to have a proper initial point ${{\bf{w}}^{\left( 0 \right)}}$ with the positive URLLC rate. Thus, we start from any random point ${{\bf{w}}^{\left( 0 \right)}}$ satisfying the convex power constraint $\sum\limits_{k = 1}^K {{{\left| {{{\bf{w}}_{kl}}} \right|}^2}} \le K,\forall l$ and (\ref{trust region}), and then iterate \begin{equation}\label{Shannon}\tag{14} \mathop {\max }\limits_{\bf{w}} \mathop {\min }\limits_{k = 1, \cdots ,K} f_k^{\left( n \right)}\left( {\bf{w}} \right) \;\;\;{\rm{s.}}{\rm{t.}}\;\;\;\text{(\ref{3b})}, \end{equation} The solution obtained by these iterations can be adopted as the feasible initial point ${{\bf{w}}^{\left( 0 \right)}}$. Finally, Algorithm 1 provides the pseudo-code for the applied path-following procedure. \begin{algorithm}[t] \caption{Path-Following Algorithm for Solving Problem (\ref{P1})} \begin{algorithmic}[1] \State \textbf{Initialization}: Iterate the convex problem (\ref{Shannon}) until the convergence to obtain an initial point ${{\bf{w}}^{\left( 0 \right)}}$. Set $n=0$. \State Using (\ref{f_n}) to obtain a concave lower bound for ${f_k}\left( {\bf{w}} \right)$ with constraint (\ref{trust region}). \State Using (\ref{g_n}) to obtain a convex upper bound for ${g_k}\left( {\bf{w}} \right)$ with constraints (\ref{cons_g1}) and (\ref{cons_g2}). \State Using (\ref{R_n}) to obtain a concave lower bound for ${R_k}\left( {\bf{w}} \right)$ under the trust region constrained by (\ref{trust region}), (\ref{cons_g1}), and (\ref{cons_g2}). \State \textbf{Repeat until (\ref{P1}) converges} : Solve the convex problem (\ref{P2}) to generate ${\bf{w}}^{\left( {n+1} \right)}$. \end{algorithmic} \end{algorithm} \subsection{Decentralized Precoding Design} The previously proposed centralized precoding design requires all the APs to upload the instantaneous CSI to the CPU, which put a significant burden on the fronthaul signaling. Besides, the computational complexity of the centralized precoding design can be exceedingly high for a huge number of antennas. As such, there is a desire for designing the precoding in a decentralized manner which only requires local instantaneous CSI at the APs. In practice, the APs can be divided into several non-overlapping cooperation clusters in which the APs in the same cluster shares both the data and the instantaneous CSI to design the precoding vectors. The APs in different clusters only have the knowledge of the statistical CSI, such as the mean and the variance. Note that although APs are divided into clusters, each user is served by all the APs instead of the APs in the cluster which the user resides in. Assume each cluster contains $M$ APs, therefore, there are $L/M$ clusters in the network. As stated before, each AP can obtain the instantaneous CSI of the APs in the same cluster and the statistical CSI of the APs in different clusters. Therefore, the virtual SINR of user $k$ in cluster ${\cal{L}}$ for designing the precoding vector can be expressed as \begin{equation}\label{VSINR-1}\tag{15} \varphi _{k{\cal L}}^{\rm{V}}\left( {{{\bf{w}}_{k{\cal L}}}} \right)\! \! = \!\!\frac{{{{\left| {\sum\limits_{l \in {\cal L}} {{\bf{h}}_{kl}^H{{\bf{w}}_{kl}}} + \sum\limits_{\bar l \notin {\cal L}} {{\mathbb{E}}\left\{ {{\bf{h}}_{k\bar l}^H} \right\}{{\bf{w}}_{k\bar l}}} } \right|}^2}}}{{\sum\limits_{i \ne k}^K {{{\left| {\sum\limits_{l \in {\cal L}} {{\bf{h}}_{kl}^H{{\bf{w}}_{il}}} + \sum\limits_{\bar l \notin {\cal L}} {{\mathbb{E}}\left\{ {{\bf{h}}_{k\bar l}^H} \right\}{{\bf{w}}_{i\bar l}}} } \right|}^2}} \!+ \!{\sigma ^2}}}. \end{equation} Since we consider Rayleigh fading channels, we have ${\mathbb{E}}\left\{ {{\bf{h}}_{k\bar l}^H}\right\} = {\bf{0}}$. Therefore, (\ref{VSINR-1}) can be written as \begin{equation}\label{VSINR-2}\tag{16} \varphi _{k{\cal L}}^{\rm{V}}\left( {{{\bf{w}}_{k{\cal L}}}} \right) = \frac{{{{\left| {\sum\limits_{l \in {\cal L}} {{\bf{h}}_{kl}^H{{\bf{w}}_{kl}}} } \right|}^2}}}{{\sum\limits_{i \ne k}^K {{{\left| {\sum\limits_{l \in {\cal L}} {{\bf{h}}_{kl}^H{{\bf{w}}_{il}}} } \right|}^2}} + {{\sigma ^2}}}}. \end{equation} The decentralized max-min URLLC rate optimization problem can be expressed as \begin{align}\label{P_distributed} &\mathop {\max }\limits_{{\bf{w}}_{\cal L}^{\rm{V}}} \mathop {\min }\limits_{k = 1, \cdots ,K} R_{k{\cal L}}^{\rm{V}}\left( {{\bf{w}}_{\cal L}^{\rm{V}}} \right)\notag\\ &\;{\rm{s.}}{\rm{t.}}\;\;\;\;\;\sum\limits_{k = 1}^K {{{\left| {{{\bf{w}}_{k{\cal L}}}} \right|}^2}} \le {p_{\max}},\forall l \in {\cal L},\tag{17} \end{align} where ${\bf{w}}_{\cal L}^{\rm{V}}$ represents the precoding vectors designed for all the users by APs in cluster ${\cal{L}}$ according to (\ref{VSINR-2}), and $R_{k{\cal L}}^{\rm{V}}\left( {{\bf{w}}_{k{\cal L}}^{\rm{V}}} \right) = \ln \!\left( {1 + \varphi _{k{\cal L}}^{\rm{V}}\left( {{\bf{w}}_{k{\cal L}}^{\rm{V}}} \right)} \right) - \sqrt {\frac{1}{{tB}} \times V_{k{\cal L}}^{\rm{V}}} \times{Q^{ - 1}}\left( \epsilon \right)$, $V_{k{\cal L}}^{\rm{V}} = 1 - \frac{1}{{{{\left( {1 + \varphi _{k{\cal L}}^{\rm{V}}\left( {{\bf{w}}_{k{\cal L}}^{\rm{V}}} \right)} \right)}^2}}}$. The problem (\ref{P_distributed}) can be solved in a similar approach as the one for (\ref{P1}). When the problem (\ref{P_distributed}) has been solved for all the clusters, we can obtain the precoding vector for user $k$ by \begin{equation}\label{w}\tag{18} {{\bf{w}}_k} = {\left[ {{{\left( {{\bf{w}}_{k1}^{\rm{V}}} \right)}^H}, \cdots ,{{\left( {{\bf{w}}_{k\left( {L/M} \right)}^{\rm{V}}} \right)}^H}} \right]^H}. \end{equation} Then, the URLLC rate of user $k$ can be obtained by computing (\ref{URLLC Rate}) using the precoding vector obtained from (\ref{w}). The computational complexity for each iteration in decentralized precoding design is ${\cal O}\left( {{{\left( {\left( {\frac{L}{M}} \right)NK} \right)}^3}\left( {2K + 1} \right)} \right)$. Compared with the centralized precoding, the computational complexity decreased by $M^3$. \section{Numerical Results} In this section, we evaluate the performance of the proposed PFA precoding design for the centralized and the decentralized fashion and investigate the impact of the precoding schemes, the length of transmission duration $t$, the number of antennas equipped at each AP $N$, and the size of the AP cluster $M$ on the URLLC rate. We first describe our adopted simulation parameters. We adopt the similar parameters setting as in \cite{ngo2017cell} as the basis to establish our simulation system model. $L$ APs and $K$ users are deployed in a rectangular area of $96\times48$ $\text{m}^{2}$. In particular, the APs are deployed on a rectangle grid. The area is wrapped around at the edges to avoid the boundary effects \cite{ngo2017cell}. The horizontal spacing between APs are $24$ m, and the vertical spacing is $12$ m. The $K$ users are deployed randomly. We adopt a similar propagation model as in \cite{bjornson2019making}. Besides, we set $L = 16$, $\tau_p = 3$, and $\epsilon = 10^{-5}$. Note that in all the figures, the achievable rates are calculated in bits/s/Hz. \begin{figure}[t!] \centering \includegraphics[width=3in]{centralized_MMSE_vs_PFA.eps} \caption{CDF of the achievable rate achieved by the centralized PFA precoding and the duality-based MMSE precoding with $t = 0.05$ ms, $B = 1$ MHz, $K = 6$, and $N = 4$.} \label{fig_centralized_MMSE_vs_PFA} \end{figure} Fig. \ref{fig_centralized_MMSE_vs_PFA} shows the cumulative distribution functions (CDFs) of the achievable rate per user achieved by the proposed PFA centralized precoding and the duality-based MMSE precoding with $t = 0.05$ ms, $B = 1$ MHz, $K = 6$, and $N = 4$ which is given by \begin{equation}\label{MMSE}\tag{19} {{\bf{w}}_k} = \frac{{{{\bf{v}}_k}}}{\left\| {{{\bf{v}}_{kl}}} \right\|},\;\;\; {{\bf{v}}_k} = p{\left( {\sum\limits_{i = 1}^K p {{{\bf{h}}}_i}{\bf{h}}_i^H + {\sigma ^2}{{\bf{I}}_{LN}}} \right)^{ - 1}}{{{\bf{h}}}_k}, \end{equation} where $p$ is the transmit power intend for each user at each AP. It can be observed that the proposed PFA centralized precoding scheme performs very well. The achievable rate per user distribution with the proposed PFA centralized precoding almost uniformly outperforms the duality-based MMSE precoding, and the former is more steeper. Specifically, applying the PFA centralized precoding leads to 32\% improvement in terms of average URLLC rate and 65\% improvement in terms of 95\%-likely URLLC rate. Note that the duality-based MMSE precoding in (\ref{MMSE}) is only a heuristic solution utilizing the uplink-downlink duality and cannot effectively minimize the MSE ${\mathbb{E}}\left\{ {\left. {{{\left| {{y_k} - {s_k}} \right|}^2}} \right|{{{\bf{h}}}_{kl}}} \right\}$. Moreover, compared with the PFA centralized precoding, the duality-based MMSE precoding has a lower computational complexity since it only requires $\frac{{{N^2}{L^2}K + NLK}}{2} + \frac{{{N^3}{L^3} - NL}}{3} + {N^2}{L^2}$ complex-valued multiplications. Besides, as expected, the performance of Shannon rate serves as a performance upper bound of the URLLC rate at the expense of infinitely long code length. \begin{figure}[t!] \centering \includegraphics[width=3in]{T_revise.eps} \caption{Optimized 95\%-likely achievable rate versus the transmission time $t$ with $N = 4$ and $B = 1$ MHz.} \label{fig_T} \end{figure} Fig. \ref{fig_T} plots the optimized 95\%-likely achievable rate by Algorithm 1 versus the transmission time $t$ with $N = 4$ and $B = 1$ MHz . As expected, the URLLC rate increases along with the transmission time $t$ according to the expression of the URLLC rate. Note that the Shannon rate is fixed since it is computed assuming a sufficient long blocklength, e.g., $t \to \infty$. Besides, when the number of user increases from 6 to 15, we can observe that the achievable rate decreases since there are more users competing for limited resources that reduces the flexibility of the resource allocation for effective beamforming. The performance gap between the Shannon rate and URLLC rate is also reduced with the increasing number of users as the performance of these two scheme is limited by the user with the weakest channel gain. \begin{figure}[t!] \centering \includegraphics[width=3in]{URLLC_revise.eps} \caption{CDF of the URLLC rate achieved by the PFA precoding in the centralized and decentralized way with $t = 0.05$ ms and $N = 4$.} \label{fig_URLLC} \end{figure} Fig. \ref{fig_URLLC} shows the performance of the PFA precoding in the centralized and decentralized fashion in terms of the URLLC rate. The curve ``C-PFA'' represents the URLLC rate computed using the centralized PFA precoding design. Also, the curve ``D-4-cluster'', ``D-2-cluster'', and ``D-16-cluster'' stand for the performance of the decentralized PFA precoding design with 4 APs, 8 APs, and 1 AP in each cluster, respectively. The first observation from Fig. \ref{fig_URLLC} is that compared with the centralized PFA precoding, the 95\%-likely URLLC rate with the decentralized PFA precoding is generally lower. This is because when the decentralized PFA precoding is adopted, only the instantaneous CSI within the cluster and the statistical CSI outside the cluster are used for optimization in each cluster. As there is a mismatch between the statistical CSI and the instantaneous CSI, the optimization for the decentralized setting is less effective for the utilization of the system resources. Besides, the performance of the 2-cluster decentralized PFA precoding outperforms the centralized PFA precoding for the strong users. The reason is that the performance of the centralized PFA precoding is always limited by the worst-case users, since substantial resources are allocated to equalize all the SINRs, while the decentralized PFA precoding benefits from being more scalable. Compared with the 2-cluster decentralized PFA precoding, when adopting the 4-cluster or 16-cluster decentralized PFA precoding, the mismatch between the statistical CSI and the instantaneous CSI is pronounced, so the performance is the worse. Specifically, compared with the centralized precoding, the 95\%-likely URLLC rate is reduced from 16.73 bits/s/Hz to 13.25 bits/s/Hz with the 2-cluster decentralized PFA precoding and to 8.95 bits/s/Hz with the 4-cluster decentralized PFA precoding. Moreover, when the fully distributed 16-cluster decentralized PFA precoding is adopted, the 95\%-likely URLLC rate is only 0.17 bits/s/Hz. However, since the computational complexity is also reduced, the performance loss of adopting the 2-cluster decentralized PFA precoding instead of the centralized precoding is tolerable. In particular, the 2-cluster decentralized PFA precoding achieves 80\% of the 95\%-likely URLLC rate, 89\% of the average URLLC rate, and 12\% of the computational complexity of the centralized precoding. The second observation is that the CDF of users' URLLC rate is not as steep as the counterpart when the decentralized PFA precoding design is adopted. The reason is that the optimization target of each cluster contains virtual SINR rather than the actual SINR, leading to under utilisation of system resources. \section{Conclusion} In this correspondence, we considered the precoding design in the cell-free massive MIMO system for URLLC in the centralized and decentralized fashion. PFA was designed for maximizing the users' minimum URLLC rate and its performance was evaluated with different settings of the transmission time, the number of antennas per AP, and the size of the AP cluster. Simulation results showed that the centralized PFA precoding design can effectively improve the performance of 95\%-likely achievable rate and the decentralized PFA precoding with a reasonable setting can approach the performance of the former but with low computational complexity. In the future, we will jointly optimize the precoding vector, the cluster formation, and the number of APs in each cluster in a distributed fashion for URLLC. \bibliographystyle{IEEEtran}

d86c75b5c9133792bc51ad99ffe176e87340f080

\section{Introduction} \label{intro} In the broader area of computational finance, the mere establishment of the existence of solution to a problem is not sufficient towards achieving the tangible financial goals, for the problem that has been posed. Accordingly, as is the case for many applications, we seek a solution that (in practice) happens to be an approximation to the actual solution being sought. To this end, we begin by observing that for problems in quantitative finance, one can arrive at either an analytical or (possibly) a semi-analytical solution, in only a handful of cases. Therefore, for the most part, one needs to devise efficient methods to arrive at the desired and appropriate solution to the posed problem which, in turn, necessitates the resorting to computational techniques. At the heart of this paper, lies a specific computational technique, widely used in the finance industry, namely, the Monte Carlo simulation approach. Accordingly, we begin our presentation with a brief narrative about this approach, with the main focus of the discussion being steered towards the recent research and development in the area of Multilevel Monet Carlo (MLMC). The interested readers may refer to \cite{giles2008multilevel,giles2008improved,giles2013antithetic,giles2013multilevel,lemaire2017multilevel,belomestny2013multilevel,heinrich2001multilevel} for greater clarity on the approach of MLMC and the key developments with respect to algorithms, as well as the applications in financial engineering problems. In this article we primarily focus on the importance sampling approach developed and studied in \cite{kebaier2018coupling,alaya2016improved} and also on how MLMC has led to the development of algorithms for efficient risk estimation in the field of financial risk management, discussed by the authors in \cite{giles2019multilevel}. However, we give a brief overview of Monte Carlo and MLMC before directing our discussion towards the aforesaid specific topics. Monte Carlo methods have become one of the driving computing tools in the finance industry. The necessity of simulating high-dimensional stochastic models, which in turn may be attributed to the linear development in the complexity corresponding to the size of the problem itself, is one of the primary reasons that this approach is becoming the critical computational strategy in the industry. The main objective of this method, in case of computational finance is to reach the necessary degree of accuracy, which is coupled with a high computational cost. More specifically, we intend to approximate $\mathbf{E}[Y]$ where, $Y = G(X)$ is functional of the random variable $X$. The traditional Monte Carlo approach requires a computational complexity of an order of $O(\epsilon^{-3})$ to attain the root mean square (RMS) error of $O(\epsilon)$ in a biased context \cite{giles2008multilevel}. This limitation led to the introduction of the multilevel framework in \cite{giles2008multilevel} to address this issue and achieve $O(\epsilon^{-2})$ computational complexity in the biased framework. The idea behind the multilevel architecture is to employ independent standard Monte Carlo on various resolution levels and use the differences as the control variate for the Monte Carlo simulation at its most granular level, which in mathematical terms is given by, \begin{equation} \label{rev:eq1} \mathbf{E}[Y_{L}]=\mathbf{E}[Y_{1}]+\sum\limits_{l=2}^{L}\mathbf{E}[Y_{l}-Y_{l-1}],~\text{where}~ Y_{l}=G(X_{l}). \end{equation} Using the standard Monte Carlo as the estimator to approximate the expectation on the right hand side of \eqref{rev:eq1}, we obtain, \begin{equation} \label{rev:eq2} \widehat{Y}_{L}=\frac{1}{N_{1}}\sum\limits_{k=1}^{N_{1}}Y_{1}^{k}+\sum\limits_{l=2}^{L}\frac{1}{N_{l}}\sum\limits_{k=1}^{N_{l}}\left(Y_{l}^{k}-Y_{l-1}^{k}\right), \end{equation} and therefore, $\mathbf{E}[Y_{L}] \approx \widehat{Y}_{L}$. Here $X_{l}$ is the approximation of the random variable $X$ on level $l$ and this approximation is contingent on the application under consideration. For example, if the underlying stochastic process is driven by a stochastic differential equation (SDE), then $X_{l}$ is the approximation of $X$, with some time discretization parameter $h_{l}$. With all the preludes being presented in the preceding discussion, we are now in a position to examine the following the seminal result due to Giles \cite{giles2008multilevel}. \begin{theorem} Let $G$ denote a functional of the random variable X, and let $Y_{l} = G(X_l)$ denote the corresponding level $l$ numerical approximation. If there exist independent estimators $Z_{l}$, based on $N_{l}$ Monte Carlo samples, and positive constants $\alpha,\beta,\gamma, c_{1},c_{2},c_{3},c_{4}$ such that $\displaystyle{\alpha \geq \frac{1}{2}\min\left(\alpha,\beta\right)}$ and \begin{enumerate} \item $\displaystyle{\left|\mathbf{E}\left[Y_{l}-Y_{l-1}\right]\right| \leq c_{1}2^{-\alpha l}}$. \item $\displaystyle{\mathbf{E}\left[Z_{l}\right]= \begin{cases} \mathbf{E}\left[Y_{1}\right],~l=0, \\ \mathbf{E}\left[Y_{l}-Y_{l-1}\right],~l>0. \end{cases}}$. \item $\displaystyle{\mathbf{V}[Z_{l}] \leq c_{2}N_{l}^{-1}2^{-\beta l}}$. \item $\displaystyle{C_{l} \leq c_{3}N_{l}2^{\gamma l}}$, where $C_{l}$ is the computational complexity of $Z_{l}$, \end{enumerate} then there exists a positive constant $c_{4}$ such that for any $\epsilon < e^{-1}$, there are values $L$ and $N_{l}$ for which the multilevel estimator $\displaystyle{Z=\sum\limits_{l=1}^{L} Z_{l}}$,has a MSE with bound, \[MSE\equiv \mathbf{E}\left[\left(Z-\mathbf{E}[Y]\right)^{2}\right] < \epsilon^{2},\] with a computational complexity $C$, having the bound, \[\mathbf{E}[C] \leq \begin{cases} c_{4}\epsilon^{-2},~\beta>\gamma, \\ c_{4}\epsilon^{-2}\left(\log\epsilon\right)^{2},~\beta=\gamma, \\ c_{4}\epsilon^{-2-\frac{(\gamma-\beta)}{\alpha}},~0<\beta <\gamma. \end{cases}\] \end{theorem} It is quite evident from the above theorem that the computational complexity is driven by the strong convergence of the estimator \textit{i.e.,} $\mathbf{V}[Z_{l}]$. Therefore, one of the main challenges while developing a MLMC based estimator is to study the order of strong convergence of the underlying approximation. With this brief introduction of MLMC, we now direct our discussion towards its recent developments, pertaining to algorithm and financial applications. \section{Importance Sampling Multilevel Algorithm} \label{importance_sampling} Since the advent of MLMC in literature, one of the directions of its progression has been through various attempts to combine this algorithm, with the already existing variance reduction techniques. For instance, Giles, in \cite{giles2013antithetic, giles2014antithetic} studied and analyzed the combination of antithetic variates and MLMC in order to bypass the Levy area simulation, encountered while using Milstein discretization scheme, in order to simulate higher dimensional SDEs. However, in our discussion we primarily focus on the combination of importance sampling algorithm and Multilevel estimator. The idea of incorporating importance sampling with multilevel estimators is derived from the seminal paper by Arouna \cite{arouna2004adaptative}. Arouna's idea relied upon the parametric change of measure and using a search algorithm to approximate the optimal change of the measure parameter, in order to minimize the variance of the standard Monte Carlo estimator. Before we discuss the research undertaken in the area of multilevel pertaining to importance sampling algorithm, we give a brief overview of the parametric importance sampling approach. Consider a general problem of estimating $\mathbf{E}[G(X)]$, where $X$ is a $d$-dimensional random variable. If $f(x)$ is the multivariate density function, then, \[\mathbf{E}[G(X)]=\int G(x)f(x)dx=\int G(x+\theta)f(x+\theta)dx=\int h(\theta, x)f(x)dx,\] where, $\displaystyle{h(\theta,x)=\frac{G(x+\theta)f(x+\theta)}{f(x)}}$. This implies that, $\mathbf{E}[G(X)]=\mathbf{E}[h(\theta,X)]$. Therefore, we need to determine the optimal value of $\theta$ such that $\text{Var}[h(\theta,X)]$ is minimum. Mathematically this is represented as, \begin{equation} \label{rev:eq3} \theta^{*}=\arg\min_{\theta \in \mathbf{R}^{d}}\text{Var}[h(\theta,X)]. \end{equation} In order to solve the above problem, one can resort to the usage of the Robbins-Monro algorithm that deals with a sequence of random variable $\left(\theta_{i}\right)_{i\in\mathbb{N}}$, which approximate $\theta^{*}$ accurately. However, the convergence of this algorithm requires certain restrictive conditions, which are known as the non explosion condition (given in \cite{alaya2015importance}), \[\mathbf{E}[h^{2}\left(\theta, X\right)] \leq C\left(1+|\theta|^{2}\right)~\text{for all}~\theta \in \mathbf{R}^{d}.\] In order to deal with this restrictive condition, the authors in \cite{chen1987convergence,chen1986stochastic} introduced a truncation based procedure which was furthered in \cite{andrieu2005stability,lelong2008almost}. An unconstrained procedure to approximate $\theta^{*}$, by using the regularity of the density function in an extensive manner, was introduced in \cite{lemaire2010unconstrained} along with the proof of convergence of the algorithm. Beside the stochastic approximation algorithm, one can also use deterministic algorithm such as sample average approximation, which, though being computationally expensive, provides for a better approximation to $\theta^{*}$. In problems dealing with the pricing of the options, for the most part the underlying stochastic process $(X_t)_{0\leq t \leq T}$ with $T>0$ being a finite time horizon, is governed by some SDEs. The general form of these SDEs is given as follows: \begin{equation} \label{rev:eq4} dX_{t}=b(X_{t})dt+\sum\limits_{j=1}^{q} \sigma_{j}(X_{t})dW_{t}^{j},~X_{0}=x \in \mathbb{R}^{d}, \end{equation} where, $\displaystyle{W:=\begin{pmatrix}W_{1}&W_{2}&\dots&W_{q}\end{pmatrix}}$ is a $q$-dimensional Brownian motion on a filtered probability space $\left(\Omega,(\mathcal{F}_{t}\right)_{0 \leq t \leq T},\mathbb{P})$, with $b:\mathbb{R}^{d} \rightarrow \mathbb{R}^{d}$ and $\sigma_{j}:\mathbb{R}^{d} \rightarrow \mathbb{R}^{d}$ being the functions satisfying the following condition: \begin{equation} \label{rev:assumption_1} \forall x,y \in \mathbb{R}^{d},~\lvert b(x)-b(y)\rvert+\sum\limits_{j=1}^{q}\lvert \sigma_{j}(x)-\sigma_{j}(y)\rvert < K_{b,\sigma}\lvert x-y \rvert,~\text{where}~ K_{b,\sigma}>0. \tag{$\text{C}_1$} \end{equation} Assumption \eqref{rev:assumption_1} ensures the existence and the uniqueness of solution to \eqref{rev:eq4}. For the most part, constructing an analytical or semi-analytical solution to \eqref{rev:eq4} is not possible, and therefore we need to rely on discretization schemes such as Euler or Milstein in order to simulate the SDEs. For detailed discussion on these discretization schemes, the interested readers may refer to \cite{kloeden1992stochastic}. Further, following the idea of \cite{arouna2004adaptative}, we consider a family of stochastic process $\left(X_{t}(\theta)\right)_{0 \leq t \leq T}$, with $\theta \in \mathbb{R}^{d}$, being governed by the following SDE: \begin{equation} \label{rev:eq5} dX_{t}(\theta) = (b(X_{t}(\theta))+\sigma(X_{t}(\theta))\theta)dt + \sum\limits_{j=1}^{q}\sigma_{j}(X_{t}(\theta))dW_{t}^{j},~\sigma(x)=\begin{pmatrix}\sigma_{1}(x)&\dots&\sigma_{q}(x)\end{pmatrix}. \end{equation} As a consequence of the Girsanov's Theorem, we know that there exists a risk-neutral probability measure $\mathbf{P}_{\theta}$, which is equivalent to $\mathbf{P}$ such that, \begin{equation} \label{rev:eq6} \frac{d\mathbf{P}_{\theta}}{{d\mathbf{P}|}_{\mathcal{F}_t}}=\exp{\left(-\innerproduct{\theta}{ W_{t}}- \frac{1}{2}\lvert\theta\rvert^{2}t\right)}, \end{equation} under which the process $\displaystyle{\left(\theta t+W_{t}\right)_{0\leq t \leq T}}$ is a Brownian motion. Therefore, \begin{equation} \label{rev:eq7} \mathbf{E}_{\mathbf{P}}\left[G(X_{T})\right] = \mathbf{E}_{\mathbf{P}_{\theta}}\left[G(X_{T}(\theta))\right]= \mathbf{E}_{\mathbf{P}}\left[G(X_{T}(\theta))e^{-\innerproduct{\theta}{ W_{T}}- \frac{1}{2}\lvert\theta\rvert^{2}T}\right] \end{equation} Therefore, following the discussion above we have, \[\mathbf{E}\left[G(X_{T})\right]=\mathbf{E}\left[h(\theta, X_{T})\right].\] here, $h(\theta, X_{T}) = G(X_T(\theta))e^{-\innerproduct{\theta}{ W_{T}}- \frac{1}{2}\lvert\theta\rvert^{2}T}$. Now the idea of importance sampling Monte Carlo method is to estimate $\mathbf{E}\left(G(X_{T})\right)$, where $\theta$ is given by, \begin{equation} \label{rev:eq8} \theta^{*}=\arg\min_{\theta \in \mathbb{R}^{d}} \text{Var}~\left(G(X_{T}(\theta))e^{-\innerproduct{\theta}{ W_{T}}- \frac{1}{2}\lvert\theta\rvert^{2}T}\right). \end{equation} In the context of Multilevel estimator, we present two approaches studied in \cite{kebaier2018coupling,alaya2016improved}, adapting the ideas studied by authors in \cite{arouna2004adaptative,alaya2015importance} and extending it to multilevel scenarios. Under the parametric change of measure, the general multilevel estimator is given defines as, \begin{equation} \label{rev:eq9} \mathbf{E}[Y_{L}]=\mathbf{E}[Y_{1}^{\theta_1}]+\sum\limits_{l=2}^{L}\mathbf{E}[Y_{l}^{\theta_l}-Y_{l-1}^{\theta_l}],~\text{where}~Y_{l}^{\theta}= G(X_{l}^{\theta})e^{-\innerproduct{\theta_l}{ W_{T}^l}-\frac{1}{2}\lvert\theta_l\rvert^{2}T}. \end{equation} Under the framework of multilevel estimator, the parametric importance sampling estimator looks like, \begin{equation} \label{rev:eq10} \displaystyle\widehat{Y}_L^{\theta}=\frac{1}{N_{1}}\sum\limits_{k=1}^{N_{1}}Y_{1}^{k,\theta_1}+\sum\limits_{l=2}^{L}\frac{1}{N_{l}} \sum\limits_{k=1}^{N_{l}}(Y_{l}^{k,\theta_l}-Y_{l-1}^{k,\theta_{l}}). \end{equation} Considering the variance of the above estimator, we have \cite{kebaier2018coupling}, \begin{equation} \label{rev:eq11} \displaystyle\text{Var}[\widehat{Y}_L^{\theta}]=\frac{1}{N_{1}}\mathbf{v}_{1}(\theta_1)+\sum\limits_{l=2}^{L}\frac{1}{N_{l}} \sum\limits_{k=1}^{N_{l}}\frac{(M-1)T}{M^l}\mathbf{v}_{l}(\theta_{l}), \end{equation} where, \[\mathbf{v}_{1}(\theta_{1})=\text{Var}[Y_{1}^{\theta_{1}}]~\text{and}~\mathbf{v}_{l}(\theta_{l})=\text{Var}[Y_{l}^{\theta_{1}}-Y_{l-1}^{\theta_{1}}].\] Therefore, as discussed, in order to solve the problem of minimizing the overall variance of the estimator described above, we intend to minimize the variance at each level of resolution, \textit{i.e.,} we aim at approximating $\theta_{l}^{*}$ for $l=1,\dots,L$, such that, \begin{equation} \label{rev:eq12} \theta_{1}^{*}=\arg\min_{\theta \in \mathbf{R}^{d}}v_{1}(\theta_{1})~\text{and}~\theta_{l}^{*}=\arg\min_{\theta \in \mathbf{R}^{d}}v_{l}(\theta_{l}). \end{equation} Further, pertinent to the discussion carried out in \cite{alaya2015importance} and another application of the Girsanov's Theorem, the above problem can be reformulated as, \begin{eqnarray} \label{rev:eq13} \theta_{1}^{*}&=&\arg\min_{\theta_{1}\in \mathbf{R}^{d}}\mathbf{E}\left[G(X_{1})^{2}e^{-\innerproduct{\theta_1}{W_{T}^{1}}+\frac{1}{2}\lvert\theta_{1}\rvert^{2}T}\right] \nonumber\\ \theta_{l}^{*}&=&\arg\min_{\theta \in \mathbf{R}^{d}}\mathbf{E}\left[\frac{M^{l}}{(M-1)T}\left(G(X_{l})-G(X_{l-1})\right)^{2} e^{-\innerproduct{\theta_l}{W_{T}^{l}}+\frac{1}{2}\lvert\theta_{l}\rvert^{2}T}\right]. \end{eqnarray} We present below the two algorithm namely, the sample average approximation and stochastic approximation, in order to approximate the $\theta_{l}$'s as the solution to \eqref{rev:eq13}. \subsection{Sample Average Approximation} The sample average approximation deals with approximating the above expectations using $\widetilde{N}_l$ sample paths. More specifically, \begin{equation} \label{rev:eq14} \mathbf{E}\left[G(X_{1})^{2}e^{-\innerproduct{\theta_1}{ W_{T}^{1}}+\frac{1}{2}\lvert\theta_{1}\rvert^{2}T}\right] \approx \frac{1}{\widetilde{N}_{1}}\sum\limits_{j=1}^{\widetilde{N}_{1}}G(X_{1}^{k})^{2} e^{-\innerproduct{\theta_{1}}{W_{T}^{1,k}}+\frac{1}{2}\lvert\theta_{1} \rvert^{2}T} \equiv \mathcal{V}_{1}, \end{equation} and, \begin{eqnarray} \label{rev:eq15} && \mathbf{E}\left[\left(G(X_{l})- G(X_{l-1})\right)^{2}e^{-\innerproduct{\theta_l}{W_{T}^{l}}+ \frac{1}{2}\lvert\theta_{l}\rvert^{2}T}\right] \nonumber\\ & \approx & \frac{1}{\widetilde{N}_{l}}\sum\limits_{j=1}^{\widetilde{N}_{l}}\frac{M^{l}}{(M-1)T}\left(G(X_{l}^{k})-G(X_{l-1}^{k})\right)^{2}e^{-\innerproduct{\theta_{l}}{ W_{T}^{l,k}}+\frac{1}{2}\lvert\theta_{l}\rvert^{2}T} \equiv \mathcal{V}_{l}. \end{eqnarray} Having approximated the expectation in the minimization problem, the authors used the standard Newton-Raphson algorithm on the functions $\mathcal{V}_1$ and $\mathcal{V}_l$ in order to approximate $\theta_{l}^{*}$ for $l=1,\dots,L$. In \cite{alaya2015importance} it is proved that if the functional $G(X)$ satisfies the non-degeneracy conditions \textit{i.e.,} $\mathbf{P}((G(X^{1}_{T})\neq 0) >0 $ and $\mathbf{P}\left(((G(X^{l}_{T})-G(X^{l-1}_{T}))\neq 0\right) >0 $ and further assuming they have finite second moment, then by Lemma 2.1 in \cite{alaya2015importance}, $\mathcal{V}_{1}$ and $\mathcal{V}_{l}$ are infinitely continuously differentiable. Moreover, both $\mathcal{V}_{1}$ and $\mathcal{V}_{l}$ are both strongly convex, thus implying the existence of the unique minimum $\theta_{1}^{*}$ and $\theta_{l}^{*}$ as the solution to equation \eqref{rev:eq13}. \subsection{Adaptive Stochastic Approximation} Under the stochastic approximation, studied in \cite{alaya2016improved} the aim of determining the optimal change of parameter $\theta_{l}^{*}$ for $l=1,\dots,L$ is carried out using the Robbins-Monro algorithm. Here, we briefly describe the algorithm. Consider a compact convex set $\Theta \subset \mathbb{R}^{q}$ such that $ 0 \in \text{int}(\Theta)$. Then the recursive algorithm with projection is defined as follows, \begin{equation} \label{rev:eq16} \theta_{l}^{n+1}=\textbf{Proj}_{\Theta}\left[\theta_{l}^{n}-\gamma_{n+1}H_{l}(\theta_{l}^{n},Y_{l},W_{T}^{l})\right], \end{equation} where, $\text{Proj}_{\Theta}(\theta)=\min_{\theta \in \Theta}\abs{\theta-\theta_{0}}$. The sequence $(\gamma_n)_{n \geq 1}$ is a decreasing sequence of positive real numbers satisfying, \begin{equation} \label{rev:eq17} \sum\limits_{n = 1}^{\infty} \gamma_{n}=\infty~\text{and}~\sum\limits_{i=1}^{\infty} \gamma_{n}^{2} < \infty. \end{equation} Also, \begin{equation} \label{rev:eq18} H_{l}(\theta_{l}^{n},Y_{l},W_{T}^{l})=\left\{\begin{array}{ll} \left(\theta_{1}T-W_{T}^{1}\right)\left(G(X_{1})^{2}e^{-\innerproduct{\theta_1}{W_{T}^1}+\frac{1}{2}\lvert\theta_{1}\rvert^{2}T)}\right),&l=1, \\ \left(\theta_{l}T-W_{T}^{l}\right)\left[\frac{M^{l}}{(M-1)T} \left(G(X_{l})-G(X_{l-1})\right)^2e^{-\innerproduct{\theta_l}{W_{T}^{l}}+\frac{1}{2}\lvert\theta_{l}\rvert^{2}T}\right],& l=2,\dots,L. \end{array} \right\} \end{equation} The algorithm described above is the constrained version of the Robbins-Monro algorithm. The inclusion of the projection operator in the recursive algorithm is to satisfy the non-explosion condition described above. Similar to the discussion carried out in the previous section, if the non-degeneracy conditions are satisfied \textit{i.e.,} $\mathbf{P}\left(G(X^{1}_{T})\neq 0\right) >0 $ and $\mathbf{P}\left(\left(G(X^{l}_{T})-G(X^{l-1}_{T})\right)\neq 0\right) >0 $, further assuming the finite second moment of $G(X_{1})$ and $G(X_{l})- G(X_{l-1})$, we can conclude the convergence of the $\theta_{l}^{*}$, constructed recursively using equation \eqref{rev:eq16}, for various level of resolutions. The term adaptive is used in the sense that, the estimation of the optimal importance sampling parameter and the multilevel Monte Carlo run simultaneously. The multilevel estimator in this case is given as follows, \begin{equation} \label{rev:eq19} \widehat{Y}_{L}^{\theta}=\frac{1}{N_{1}}\sum\limits_{k=1}^{N_{1}}Y_{1}^{k,\theta_{1}^{k-1}}+\sum\limits_{l=2}^{L}\frac{1}{N_{l}}\sum\limits_{k=1}^{N_{l}}\left(Y_{l}^{k,\theta_{l}^{k-1}}-Y_{l-1}^{k,\theta_l^{k-1}}\right). \end{equation} However, for the purpose of the practical implementation, one needs to stop the approximations procedure after finite number of iterations. Having approximated the $\theta_{l}^{*}$ for $l=1,\dots,L$, we use the multilevel algorithm described by equation \eqref{rev:eq16} to estimate our expectation. It is quite evident from the way the algorithms have been described that the importance sampling algorithm combined with a multilevel estimator is more computationally complex than the standard multilevel algorithm. However, the variance reduction achieved by these combinations compensates for the high computational complexity. That is, the hybrid algorithm achieves the desired RMS error much faster than the MLMC estimator. The studies carried out in \cite{kebaier2018coupling,alaya2015importance} demonstrate the accuracy of the hybrid importance sampling multilevel algorithm over standard multilevel algorithm, through a series of numerical examples, where the underlying SDEs are multi-dimensional. The slight drawback of the sample average approximation method, though more stable than the adaptive stochastic algorithm, is the slow convergence rate to the optimal value. As for the stochastic approximation, the algorithm is sensitive to the learning parameter $\gamma_n$ and therefore is unstable. It may be pointed out that the study performed above only deals with the Euler Multilevel Monte Carlo, restricted to the use of Euler discretization to simulate the underlying SDEs. More recently, a study carried out by authors in \cite{sinha2022multilevel} generalize this approach, undertaking higher order discretization schemes such as Milstein to simulate the underlying SDEs. The interested reader can refer to the references mentioned therein to get a more rigorous understanding of this hybrid algorithm. \section{MLMC and Efficient Risk Estimation.} \label{multi_risk} Risk measurement and consequent management is one of the essential components of financial engineering. The computation of the former (risk measures) for a financial portfolio is both challenging and computationally intensive, which may be ascribed to computations involving nested expectation, which entails multiple evaluations of the loss to the portfolio, for distinct risk scenarios. Further, the cost of computing loss of portfolio entailing thousands of derivatives becomes progressively expensive with an increase in the size of the portfolio \cite{giles2019sub}. Value-at-Risk (VaR), Conditional VaR (CVaR), and the likelihood of a large loss are the necessary risk metrics used to estimate the risk of a financial portfolio. At the core of these estimation, is the necessity of evaluating the nested expectation, given by, \begin{equation} \label{rev:eq20} \eta=\mathbf{E}\left[H\left(\mathbf{E}[X\lvert Y]\right)\right ] \end{equation} where, $H$ is the Heaviside function. More specifically, suppose we need to compute the probability of the expected loss being greater than $L_{\eta} \in \mathbf{R}$, \textit{i.e.,} we are interested in the following computation: \begin{equation} \label{rev:eq21} \eta=\mathbf{E}\left[H(\mathbf{E}[\Delta\lvert R_{\tau}]-L_{\eta})\right], \end{equation} where $\mathbf{E}[\Delta \lvert R_{\tau}]$ is the expected loss in a risk-neutral world, with $R_{\tau}$ being a possible risk scenario at some short risk (time) horizon $\tau$. Also, $\Delta$ is the average loss of many losses incurred from different financial derivatives, depending upon similar underlying assets \cite{giles2019sub}, that is, \begin{equation} \label{rev:eq22} \Delta=\frac{1}{K}\sum\limits_{i=1}^{K} \Delta_{i}, \end{equation} where $K$ is the total number of derivatives and $\Delta_i$ is the loss from the $i$-th derivative. The average is considered to ensure the boundedness of $\Delta$, when the portfolio size of $K$ increases. A standard and straight forward way to estimate the nested expectation \eqref{rev:eq20} is the usage of Monte Carlo method. This involves, simulating $M$ independent scenarios of the risk parameter $R_{\tau}$, and for each risk scenario, $N$ total loss samples, which are independent. This method was explored in \cite{gordy2010nested} and an extended analysis was carried out in \cite{giorgi2017limit}. The total computational cost to perform the above simulation is $O(\max(K\epsilon^{-2},\epsilon^{-3}))$ in order to achieve the root-mean-squared (RMS) error of $\epsilon$ \cite{giles2019sub}. In order to handle this issue we present the ideas studied in \cite{giles2019multilevel} under the realm of MLMC. \subsection{Adaptive Sampling Multilevel estimator} \label{adap_multi} As mentioned in the previous section, the cost of the standard Monte Carlo to achieve the root-mean-squared error of $\epsilon$ is $O(\epsilon^{-3})$. To improve the computational complexity, the authors in \cite{broadie2011efficient} developed an efficient through the adaptation of the sample size required in the inner sampler of Monte Carlo, to the particular outer sampler random variable. Under certain conditions, the authors were able to achieve the $O(\epsilon^{-5/2})$ computational complexity to achieve the RMS of $\epsilon$. Giles in \cite{giles2019multilevel} extended this approach to the multilevel framework and was able to achieve $O\left(\epsilon^{-2}\lvert \log \epsilon \rvert^{2}\right)$ computational cost for a RMS error tolerance $\epsilon$. Before presenting the work initiated by Giles, we put forth a brief review of the studies carried out in \cite{gordy2010nested} and \cite{broadie2011efficient}. The authors in \cite{gordy2010nested}, estimated the inner expectation of the equation \eqref{rev:eq20}, \textit{i.e.,} $\mathbf{E}[X\lvert Y = y]$, for a given $y$, using the unbiased Monte Carlo estimator, with $N$ sample paths, as given by, \begin{equation} \label{rev:eq23} \widehat{Z}_{N}(y)=\frac{1}{N}\sum\limits_{n=1}^{N} x_{n}(y), \end{equation} where, $\{x_{n}(y)\}_{n}$ are the mutually independent samples from the random variable $X$, conditioned on $Y=y$. Again using the Monte Carlo for the outer expectation, we have, \begin{equation} \label{rev:eq24} \eta \approx \frac{1}{M}\sum\limits_{m=1}^{M} H\left(\widehat{Z}_{N}(y_{m})\right), \end{equation} where $\{y_{m}\}_{m}$ are the mutually independent samples from the random variable $Y$. Further, they proved that if the two random variables $\mathbf{E}[X\lvert Y]$ and $\widehat{Z}_N$ have the joint density $d_{N}(y,z)$ and assuming that for $i=0,1,2$, $\displaystyle{\frac{\partial}{\partial y_{i}}d_{N}(y,z)}$ exists, plus there exists a non-negative function $d_{i,N}$, such that, \[\left\lvert \frac{\partial}{\partial y_{i}}d_{N}(y,z) \right \rvert \leq d_{i,N},~\text{for all}~ N,y,z,~\text{and}~ \sup_{N} \int\limits_{-\infty}^{\infty} \lvert z \rvert^{q}d_{i,N}(z)dz < \infty \tag{$\text{C}_2$} \label{rev:assumption_2},\] for all $0\leq q \leq 4$, then the RMS error of the estimator \eqref{rev:eq24} is $O\left(M^{-1/2}+N^{-1}\right)$. Therefore, in order to achieve the RMS error of $O(\epsilon)$ we need $M=O(\epsilon^{-2})$ and $N=O(\epsilon^{-1})$, leading to the total computational complexity of $O(\epsilon^{-3})$. Authors in \cite{broadie2011efficient} developed an adaptive sampling technique to deal with high computational complexity previously discussed. Their approach was based on the likelihood that an additional sample will result in a negative estimate of $\widehat{Z}_{N+1}$ having estimated that $\widehat{Z}_{N} > 0$ for given $Y$. More specifically they showed that, \[\mathbf{P}\left[\widehat{Z}_{N+1}\leq 0 \lvert \widehat{Z}_{N}\right] \leq \frac{\sigma^{2}}{\left(N\widehat{Z}_{N}(Y)+\mu\right)^{2}}\approx \frac{\sigma^{2}}{N^{2}\mu^{2}},\] where $\mu = \mathbf{E}[X\lvert Y]$ and $\sigma^2 = \text{Var}[X\lvert Y]$. Therefore, if $\displaystyle{N \geq \frac{\epsilon^{-1/2}\sigma}{\abs{\mu}}}$, then the probability that $H\left(\widehat{Z}_{N}(Y)\right)= H\left(\widehat{Z}_{N+1}(Y)\right) \approx H\left(\mathbf{E}[X \lvert Y]\right)$ is equal to $1-\epsilon$. Based on these observations, the authors in \cite{broadie2011efficient} introduced two algorithms, the first being based on the minimization of the total number of samples for all inner Monte Carlo samplers with respect to given tolerance $\epsilon$, and the second being iterative, estimating $\abs{\mu}$ and $\sigma$ after every iteration, for given value of $Y$, using $N$ samples further adding more inner samples till $\displaystyle{\frac{N\mu}{\sigma}}$ exceeds some error margin threshold. Under these two algorithms it was observed that the overall computational complexity is $O(\epsilon^{-5/2})$ \cite{giles2019multilevel}. The authors in \cite{giles2019multilevel} introduced the above algorithms in the realm of multilevel simulation, wherein they used multilevel estimator in order to achieve an approximation to the outer expectation, while making use of the sample size in the inner expectation as the discretization parameter. More specifically, \begin{equation} \label{rev:eq25} \widetilde{\eta} \coloneqq \sum\limits_{l=0}^{L}\frac{1}{M_{l}}\sum\limits_{m=1}^{M_{l}} H\left( \widehat{Z}_{N_{l}}^{f,l,m}(y^{l,m})\right)-H\left(\widehat{Z}_{N_{l-1}}^{c,l,m}(y^{l,m})\right), \end{equation} where, \begin{equation} \label{rev:eq26} \widehat{Z}_{N_{l}}^{f,l,m}(y)=\frac{1}{N_l}\sum\limits_{n=1}^{N_{l}} x^{f,l,m,n}(y), \end{equation} with $\{x^{.,l,m,n}(y)\}$ being the i.i.d samples of the random variable $X$, given $Y=y$. Also, $H\left(\widehat{Z}_{-1}^{c,0,\dots}(y)\right)\equiv 0$. Now under the assumptions \ref{rev:assumption_2}, it can be proved that \cite{giles2019multilevel}, \[\babs{\mathbf{E}\left[H\left(\widehat{Z}_{N_{l}}(Y)\right)-H\left(\mathbf{E}[X\lvert Y]\right)\right]}=O\left(N_{l}^{-1}\right).\] Further, under the assumption that there exists constants $\delta_{0}$ and $\rho_{0}$ such that, $\displaystyle{\rho(\delta)\leq \rho_{0}}$, for all $\delta \in [0,\delta_0]$ where $\delta$ is the random variable with density $\rho$, the authors in \cite{giles2019multilevel} proved that, if $X$ and $Y$ are the two random variables, satisfying the stated assumption, then, \begin{equation} \label{rev:eq27} \text{Var}\left[H\left(\widehat{Z}_{N}(Y)\right)-H\left(\mathbf{E}[X\lvert Y]\right)\right]=O(N^{-1/2}). \end{equation} The above result determines the strong convergence property necessary to analyze the full potential of the MLMC estimator, in this scenario. However, if $\displaystyle{N_{l}=N_{0}2^{l}}$, then with standard MLMC complexity analysis it is easy to determine that the computational complexity required to achieve RMS error of $\epsilon$, we need $O\left(\epsilon^{-5/2}\right)$ computational complexity. To cater to this high computational demand, even in the framework of MLMC, the authors undertook the adaptive approach developed in \cite{broadie2011efficient} and extended it to the framework of MLMC. Giles extended the studies carried out by authors in \cite{broadie2011efficient} to multilevel paradigm with an aim to reduce the overall computational cost to $O\left(\epsilon^{-2}\abs{\log \epsilon)}^2\right)$. In addition to the assumptions stated above, it is further assumed that, \begin{equation} \label{assumption_3} \sup_{y} \mathbf{E}[\sigma^{-q}\abs{X-\mathbf{E}[X\lvert Y]}^{q}\lvert Y=y] < \infty, \quad 2 < q < \infty. \tag{$C_3$} \end{equation} Thus, under the above stated assumptions, it was proved in Lemma 2.5 (for the perfect adaptive sampling) and Theorem 2.7 of \cite{giles2019multilevel}, that if the maximum number of sample paths is restricted to, \begin{equation} \label{rev:eq28} N =\bigg \lceil\max\left(O\left(\epsilon^{-1}\right),C^{2}\frac{\sigma^2}{\abs{\mu}^2}\right)\bigg \rceil, \end{equation} then the further number of sample path of various level of resolutions are given by, \begin{equation} \label{rev:eq29} N_{l}=\bigg \lceil N_{0}4^{l}\max\left(2^{-l},\min\left(1,\left(C^{-1} N_{0}^{1/2}2^{l}\frac{\abs{\mu}}{\sigma}\right)^{-r}\right)\right)\bigg \rceil, \end{equation} with $C$ being some confidentiality constant and $\displaystyle{1 < r < 2-\frac{2}{q}}$ for the perfect adaptive sampling and \\ $\displaystyle{1< r < 2-\frac{\sqrt{4q+1}-1}{q}}$ when the values of $\abs{\mu}$ and $\sigma$ is approximated. Therefore, \begin{equation} \label{rev:eq30} \text{Var}\left[H\left(\widehat{Z}_{N}(Y)\right)-H\left(\mathbf{E}[X\lvert Y]\right)\right]=O\left(2^{-l}\right), \end{equation} thereby leading to the overall computational complexity of the desired order. In a detailed discussion carried out in Section 4 of \cite{giles2019multilevel}, it was proved (pertaining to the calculation of VaR and CVaR) that in order to achieve the overall computational cost of $O(\epsilon)$ RMS error, the required computational complexity is $O\left(\max(\epsilon^{-2}\abs{\log \epsilon)}, K \epsilon^{-2})\right)$ for the estimation of VaR and CVaR, respectively. The numerical test on a model problem undertaken shows the efficacy of the algorithm constructed. Readers are directed to the referred paper for detailed discussion on the proofs of the above stated results. It may be noted that the computational complexity increases with an increase in the portfolio size, $K$. A random sub-sampling approach, extending it to a multilevel framework, thereby addressing the dependency on the portfolio size, to achieve the desired RMS error was recently introduced in \cite{giles2019sub}. \section{Conclusion} In this paper, we gave a brief overview of the recent trends in the paradigm of the multilevel algorithm concerning the importance sampling, in the case of option pricing and an adaptive sampling approach while determining the VaR and CVaR for large portfolios. The algorithms discussed serves as the improvement in the computational efficiency of the standard multilevel estimators, each having its merits and shortcomings. As discussed in Section \ref{importance_sampling}, the importance sampling algorithm combined with multilevel estimators significantly decreases variance at various resolution levels. However, the decrease in variance comes at the cost of increased computational complexity in either case and an increase in the sensitivity to approximate the optimal parameter. As for developing the MLMC based algorithm for efficient risk estimation discussed in Section \ref{multi_risk}, the adaptive sampling approach introduced in this paradigm leads to a significant improvement in the overall computational complexity to achieve the desired root mean squared error. However, the dependence of computational complexity on the size of the portfolio is a subtle shortcoming of the discussed algorithm. Overall, the presented ideas have substantially contributed to the research and development of the multilevel algorithm for various applications encountered in financial engineering problems. However, the scope to enrich the standard algorithm with non-standard variance reduction techniques is still an exciting path for future research.\\ \\ \bibliographystyle{elsarticle-num}

d85bd8fd0b4b47d18ff271dcd308bab643060dee

\section{Introduction} \label{sec:introduction} Recent work on GPT-2~\citep{gpt2} and GPT-3~\citep{gpt3} have shown that large language models possess few-shot learning capabilities and zero-shot instruction following capabilities, despite only being pre-trained with a self-supervised causal language modeling objective (which is to predict the next token). An arbitrary task can be converted into a natural language task specification, often called a \textit{prompt}. Prompting a task in this way makes its format similar to the language modeling objective used to pre-train large language models. In the zero-shot setting, this prompt contains just the task with instructions, whereas in the few-shot setting, the prompt contains both the task and several example demonstrations. When a language model is tasked to generate text to complete this prompt, it can perform the task in the process. The broader paradigm of reframing all tasks as text generation is known as \textit{prompt-based learning}. In the few-shot setting, the learning that occurs from examples provided in a given prompt (the context) is known as \textit{in-context learning} \citep{promptingsurvey}. In the zero-shot setting, models perform \textit{instruction following} \citep{instructgpt}, with their performance guided through natural language instructions provided in the prompt. Emergent prompt-based learning capabilities have mainly been demonstrated for unidirectional language models. Bidirectional language models have stronger learned representations \citep{bert,xlm,t5}; however, they have not been able to broadly demonstrate the same few-shot in-context learning capabilities or zero-shot instruction following capabilities due to the incompatibility bidirectional denoising pre-training objectives have with the prompting paradigm. Instead, they typically require fine-tuning. Bidirectional models are not able to generate long, fluent completions to prompts since they are usually only trained to output single tokens or short spans of text to in-fill masked tokens during pre-training. We discuss this more in-depth in Section \ref{sec:directionality}. Today, language model architects are faced with a difficult choice between unidirectional or bidirectional models. The authors of GPT-3 lay out this design dilemma in \citet{gpt3}: \begin{displayquote} \small{``GPT-3 has several structural and algorithmic limitations ... as a result our experiments do not include any bidirectional architectures or other training objectives such as denoising ... our design decision comes at the cost of potentially worse performance on tasks which empirically benefit from bidirectionality ... making a bidirectional model at the scale of GPT-3, and/or trying to make bidirectional models work with few- or zero-shot learning, is a promising direction for future research, and could help achieve the `best of both worlds'.''} \end{displayquote} \fewshotfigure In this paper, we directly address this dilemma. We contribute a new technique, \textsc{Sap} (\textbf{S}equential \textbf{A}utoregressive \textbf{P}rompting), that enables bidirectional language models to take advantage of prompting and allows them to perform at the level of unidirectional models in few- or zero-shot learning without fine-tuning. \textsc{Sap} iteratively prompts bidirectional models, concatenating previous generations back into the prompt, to produce longer generations from models that were only pre-trained to output short, mask-infill spans. We acknowledge efficiency concerns in Section \ref{sec:conclusion} and we discuss the importance and impact of \textsc{Sap} and its results to the field regardless of those concerns. Using the machine translation task as an in-depth case study, we empirically demonstrate mT5~\citep{xue-etal-2021-mt5}, a bidirectional language model, used with \textsc{Sap} outperforms its unidirectional counterparts, GPT-3 and XGLM~\citep{gpt3,xglm} in both the few-shot and zero-shot settings, while utilizing approximately 50\% fewer parameters. We then examine \textsc{Sap}'s effectiveness on other tasks such as question answering and summarization, demonstrating that bidirectional models can be prompted for tasks beyond machine translation. Our work hints at the possibility of more efficient and performant few-shot learners through pre-trained language models that incorporate bidirectionality. We discuss this impact and outline future research directions to this end in Section \ref{sec:conclusion}. In summary, our key contributions are: \begin{enumerate} \item{We introduce \textsc{Sap}, a technique that enables bidirectional language models to work with few-shot and zero-shot prompt-based learning at a level that exceeds unidirectional models. Our results demonstrate in-context learning and instruction following are emergent properties of a broader class of language models, rather than only unidirectional models, addressing a long-standing challenge in language model design and use.} \item{We perform an in-depth study of the effectiveness of a bidirectional language model, mT5, with \textsc{Sap} on the machine translation task. Evaluating over 14 language pairs, despite using approximately 50\% fewer parameters than GPT-3 and XGLM, we find \textsc{Sap} with mT5 has improved average few-shot and zero-shot performance over all language pairs, and especially has improved performance on individual low-resource language pairs. } \item{We propose a range of improvements---filtering, prompt ensembling, and English-centric bootstrapping---to the unsupervised machine translation procedure outlined by \citet{gpt3unsupervised} to better adapt the bootstrapping process for unsupervised low-resource machine translation.} \item{We assess \textsc{Sap}'s performance on the tasks of question answering and summarization, and we find the technique enables few-shot in-context learning and zero-shot instruction following capabilities of bidirectional models in tasks beyond machine translation.} \end{enumerate} \section{Related Work} \subsection{Unidirectional and Bidirectional Language Models} \label{sec:directionality} Transformer-based language models \citep{attention} can be broadly categorized into bidirectional and unidirectional models. Bidirectional models are models that use a denoising pre-training objective (such as masked language modeling), allowing them to utilize \textit{bidirectional} context when learning language representations. Unidirectional language models are models with a causal---or a left-to-right---language modeling objective (such as next token prediction), restricting them to be \textit{unidirectional} when learning representations \citep{promptingsurvey}. The T5 family of models, such as T5 v1.1 and mT5, and BART-style models \citep{bart} are bidirectional, while GPT-style models, such as GPT-2, GPT-3, and XGLM are unidirectional. Usually, but not always, bidirectional models are paired with an encoder-decoder architecture, while unidirectional models are paired with a decoder-only architecture \citep{bert, t5,xue-etal-2021-mt5,gpt2,gpt3,xglm,bigsciencearchobjective}. BERT-style models are an example of an exception. BERT-style models are bidirectional, but they cannot be easily utilized for prompting and text generation since they are encoder-only \citep{bertspeak}. Of the available bidirectional models, T5 models are the only models with a long enough sequence length (unlimited with their relative position embeddings) to support many in-context prompt examples and with a large enough number of parameters to be effective zero-shot and few-shot performers \citep{gpt2,gpt3,scaling}. See Appendix \ref{sec:survey} for a survey of popular open source language models. Aside from sequence length and model size, BART is not purely trained on the span denoising objective \textsc{Sap} exploits, but is also trained on many other corruption objectives like ``Sentence Permutation.'' For this reason, we utilize the T5 models for experiments and leave the exploration of the generalization of \textsc{Sap} to other models, that could become available later, as future work. \citet{bert} and \citet{t5} have both shown that after transfer learning, bidirectional denoising pre-training objectives such as BERT's masked language modeling and T5's random span corruption outperform causal language modeling on downstream tasks. \citet{gpt3} concedes this to be a potential source of weakness for the GPT-3 model on certain tasks where bidirectionality is important. Despite the advantages of denoising objectives, prompting and in-context learning capabilities have not been broadly demonstrated for bidirectional language models like T5, disqualifying them when few-shot in-context learning and zero-shot instruction following is desired. \citet{lester-etal-2021-power} explains this may be because: \begin{displayquote} \small{``...a T5 model pre-trained exclusively on span corruption, such as T5.1.1, has never seen truly natural input text (free of sentinel tokens), nor has it ever been asked to predict truly natural targets''} \end{displayquote} In other words: when pre-trained on their denoising objectives, language models like T5 that utilize bidirectionality are only conditioned to output a single token or short spans of tokens (the in-fill of the mask) rather than full and complete sentences; this inhibits their ability to generate arbitrary-length natural responses to a variety of prompts. Despite the stronger learned representations of bidirectional models, their shortcomings in prompt-based learning motivate \citet{gpt3} and \citet{xglm} to explicitly choose unidirectional models over bidirectional models for GPT-3 and XGLM. \subsection{Prompting Bidirectional Language Models} \label{sec:priortechniques} Unlike prior approaches to incorporate prompt-based learning capabilities into bidirectional models, our technique, \textsc{Sap}, neither requires fine-tuning, weight updates, nor supervised instruction-tuning datasets. It demonstrates that bidirectional language models develop \textit{innate} few-shot learning capabilities with in-context learning and zero-shot instruction following capabilities. \paragraph{Cloze-style prompts} \citet{cloze} and \citet{cloze2} find that bidirectional models such as RoBERTa and ALBERT \citep{roberta,albert} can be ``prompted'' with cloze-style phrases. They propose a few-shot training paradigm called \textsc{Pet} where the model's predicted mask in-fill, called a ``verbalizer,'' is used to label fine-tuning examples for the model. These verbalizers are only a single word or a few words, e.g. ``yes'', ``no'', ``amazing'', ``worse''. \citet{electrazero} follow a similar technique, but with the ELECTRA model \citep{electra}. These works primarily demonstrate zero-shot effectiveness on classification tasks such as sentiment analysis, rather than more challenging generation tasks such as machine translation or question answering. Furthermore, they still require fine-tuning for effective few-shot learning, a major limitation that does not achieve the prompt-based in-context learning or instruction following abilities of unidirectional models such as GPT-3. \paragraph{LM-adaptation} \citet{lester-etal-2021-power} finds some success with prompting the T5 v1.1 models after continued pre-training on the unidirectional prefix-LM objective described in \citet{t5}. The resulting model, T5 v1.1 LM-adapted (T5+LM), is described as a late-stage adaptation to a unidirectional objective. Adaptation requires performing weight updates, and given that representations learned by the original denoising objective have been shown to be superior \citep{t5}, we hypothesize that such an adaptation could degrade the quality of the learned representations. \paragraph{Prompt-tuning} \citet{lester-etal-2021-power} and \citet{prefixtuning} find by fine-tuning only a portion of the parameters in an otherwise frozen pre-trained bidirectional language model, a ``soft prompt'' can be discovered through backpropagation. Soft prompts are prompts discovered in the embedding space of the model and are not grounded in natural language. As a form of parameter-efficient fine-tuning \citep{parameff}, this approach requires training the prompt embeddings and benefits from initialization from LM-adaptation, both of which require performing weight updates. The nature of soft prompts lacking grounding in natural language makes their use and flexibility limited, a stark difference from the instruction following capabilities of unidirectional models \citep{promptingsurvey}. \paragraph{Instruction-tuning} Language models can be fine-tuned on a supervised dataset consisting of natural language prompts and their respective target completions \citep{flan,t0,instructgpt,metalicl}. This ``instruction-tuning'' technique allows these models to improve performance on instruction following and therefore exhibit few-shot and zero-shot capabilities through prompting. The T0 model in particular is an instruction-tuned version of the T5+LM model~\citep{lester-etal-2021-power}, augmenting it with prompting capabilities. While instruction-tuning likely bolsters the instruction following performance of a model, we hypothesize that by instruction-tuning, the T0 model is to some degree surfacing the innate prompting ability that the bidirectional model already has. We provide evidence towards this hypothesis by demonstrating that bidirectional models can be prompted without instruction-tuning. \subsection{Unsupervised Machine Translation through Prompting} GPT-2~\citep{gpt2} and GPT-3~\citep{gpt3} have shown it is possible to perform few-shot machine translation and unsupervised zero-shot machine translation with large language models using prompting and in-context learning. The XGLM model~\citep{xglm} trains a similar architecture to GPT-3 on a diverse multilingual corpus, resulting in improvements on few-shot, low-resource machine translation. \citet{gpt3unsupervised} introduce bootstrapping and self-amplification techniques to further improve unsupervised zero-shot performance on machine translation. \firstwordablationtable \section{Few-shot Machine Translation} \label{sec:few-shot} To motivate our method for enabling few-shot in-context learning in bidirectional language models, we first focus on applying $\text{mT5}_\text{3.7B}$ (mT5-XL) \citep{xue-etal-2021-mt5} to the machine translation task as an in-depth case study since this task benefits greatly from bidirectionality \citep{xlm,xglm}. We largely follow the procedure of \citet{xglm}, except with mT5 and \textsc{Sap}. mT5 is a massively multilingual bidirectional model trained on random span corruption, a variant of masked language modeling. We demonstrate that with \textsc{Sap}, mT5 can perform few-shot machine translation using prompting and in-context examples with no fine-tuning. We first formulate a prompt format that utilizes its random span masking scheme to complete the translation task, such as: \\[.5em]\centerline{\fbox{\begin{minipage}{15.5em} \small{ Translate Spanish to English.\\ Spanish: El clima es soleado.\textcolor{gray}{</s>}\\ English: The weather is sunny.\textcolor{gray}{</s>}\\ Spanish: Mi perro es un cachorro.\textcolor{gray}{</s>}\\ English: My dog is a puppy.\textcolor{gray}{</s>}\\ Spanish: Los árboles son importantes.\textcolor{gray}{</s>}\\ English: \textcolor{red}{<X>} } \end{minipage}}} \subsection{Sequential Autoregressive Prompting (\textsc{Sap}) Technique} By requiring mT5 to in-fill \textcolor{red}{<X>}\footnote{We use the first sentinel token from the mT5 vocabulary as our mask token.}, we are effectively asking it to translate the Spanish sentence. However, due to the limitations of the denoising pre-training objective on prompting (described in Section \ref{sec:directionality}), we observe mT5 often outputs a partial translation of the beginning of the source sentence, rather than the full translation. To overcome this, we prompt mT5 $T$ times until the model generates a stop token \textcolor{gray}{</s>}\footnote{We repurpose the 100th sentinel token from the mT5 vocabulary as our stop token.}, resulting in a longer translation. At each time step of iteration, we keep the first word generated (using the space character as delimiter) and concatenate it into the last line of the prompt to use in the next time step. This iterative prompting enables us to extract longer generations. Formally, we denote the generation at each time step $t$ as $G_t$. We denote the first word generated at each time step as $F_t$, where $F_t = \texttt{SPLIT}(G_t, \texttt{" "})\texttt{[0]}$. We update the prompt at each time step $P_t$ to include the cumulative generation from all previous time steps concatenated in the last line of the prompt. The prompt used at each time step $P_t$ is as follows: \\[.5em]\centerline{\fbox{\begin{minipage}{15.5em} \small{ Translate Spanish to English.\\ Spanish: El clima es soleado.\textcolor{gray}{</s>}\\ English: The weather is sunny.\textcolor{gray}{</s>}\\ Spanish: Mi perro es un cachorro.\textcolor{gray}{</s>}\\ English: My dog is a puppy.\textcolor{gray}{</s>}\\ Spanish: Los árboles son importantes.\textcolor{gray}{</s>}\\ English: \texttt{CONCAT}($F_0$, \ldots, $F_{t-1})$ \textcolor{red}{<X>} } \end{minipage}}}\\[.5em] In Table \ref{table:firstword-ablation}, we also consider sequential prompting---concatenating the entire generation $G_t$ instead of just the first word of the generation $F_t$---but find that it produces significantly inferior results as low-quality tokens are generated after the first word. By conditioning the model to generate the next word in the translation based on previous words generated, this technique resembles autoregression. mT5 is already autoregressive, but it is autoregressive only at the decoder level. Adding previously generated words back into the prompt allows them to pass through the encoder layers as well. For this reason, we call this technique \textsc{Sap} (\textbf{S}equential \textbf{A}utoregressive \textbf{P}rompting). To provide a signal to stop generation, we add our stop token at the end of each example in the prompt. We stop prompting after the model generates a stop token.\footnote{We also implement a basic post-processing step to strip any generated text after a repeated sequence of three or more tokens following settings available in common decoding implementations \citep{transformers}.} The overall process is graphically depicted, with stop tokens omitted, in Figure \ref{fig:fewshot}. \subsection{Results} \label{sec:fewshot-results} Following \citet{xglm}, we evaluate our technique on 14 languages from the FLORES-101 dataset \citep{flores101} that span high-resource and low-resource languages\footnote{HR: English (en), German (de), French (fr), Catalan (ca), Finish (fi), Russian (ru), Bulgarian (bg), Chinese (zh); LR: Korean (ko), Arabic (ar), Swahili (sw), Hindi (hi), Malayalam (my), Tamil (ta)}. We evaluate SentencePiece BLEU (spBLEU) \citep{flores101} in every direction, leading to an evaluation over 182 language pairs in total. Abbreviated results can be found in Table \ref{table:fewshot-flores-results-abbrev}, and the matrix of full results can be found in Appendix \ref{sec:flores-results}. Examples generations can be found in Appendix \ref{sec:examplegenerations}. On an average spBLEU score over all 182 pairs, our model matches the performance of the unidirectional XGLM and GPT-3 models---with approximately 50\% fewer parameters and 16x fewer examples. Notably, our technique has significant improved performance on language pairs with at least one low-resource language, while trailing only slightly on high-resource pairs. \bootstrapfigure \section{Unsupervised Zero-shot Machine Translation} \label{sec:zero-shot} To extend our in-depth case study on the machine translation task, we now perform fully unsupervised zero-shot machine translation with \textsc{Sap} and mT5 following the procedure of \citet{gpt3unsupervised}, which uses a self-amplification technique to boost performance. A comparison of zero-shot performance without self-amplification can be found in Appendix \ref{sec:selfamp-ablation}. We ultimately will replace the examples in the few-shot prompt with synthetic parallel examples. These synthetic parallel examples are bootstrapped in a completely unsupervised fashion using a zero-shot translation prompt with no examples. The zero-shot prompt format looks like: \\[.5em]\centerline{\fbox{\begin{minipage}{15.5em} \small{ Translate Spanish to English.\\ Spanish: Los árboles son importantes.\textcolor{gray}{</s>}\\ English: \textcolor{red}{<X>} } \end{minipage}}}\\[.5em] We adapt the bootstrap process of \citet{gpt3unsupervised} to retrieve these synthetic parallel examples. The process, as depicted in Figure \ref{fig:bootstrap}, consists of three steps: \begin{itemize}[leftmargin=*] \item[] \textbf{Step 1 (sampling)}: Generate synthetic parallel examples using a zero-shot translation prompt (with no examples) to translate sentences from a monolingual source language corpus. \item[] \textbf{Step 2 (filtering)}: Filter out low-quality synthetic examples to keep only high-quality synthetic examples using an unsupervised scoring technique (discussed in Section \ref{sec:mt5score}). \item[] \textbf{Step 3 (self-amplification)}: Translate any source language sentence desired using these synthetic parallel examples in the few-shot prompt. \end{itemize} We iteratively run multiple rounds of this bootstrap by repeating step 2 and step 3 to form a better few-shot prompt. The few-shot prompt after self-amplification is used to translate more source language sentences. These are then filtered using the scoring technique used in step 2 and so on. In our experiments, we run four bootstrapping rounds and sample 100 source language sentences from the training dataset in each round. Note that none of the target language parallel sentences from the training dataset are used in this zero-shot setting; following \citet{gpt3unsupervised}, only the source language sentences are used. \subsection{Filtering Down to High-quality Translations} \label{sec:mt5score} The filtering step of the bootstrap requires an unsupervised scoring method for assessing the quality of translations. We first use \texttt{{langdetect}}\footnote{\url{https://pypi.org/project/langdetect/}}, a language identifier, as a simple rule-based filter to ensure the generated text is in the desired target language. We then score the remaining generated translations against their corresponding original sentence in the source language. For this unsupervised multilingual similarity metric, we utilize the BERTScore~\citep{bertscore} algorithm with $\text{mT5}_{\text{300M}}$ (mT5-small)\footnote{The BERTScore Python library~\citep{bertscore} directly supports using mT5 instead of BERT.}, dubbing it ``mT5Score''. We ablate the use of mT5Score as a filter in Appendix \ref{sec:mt5score-ablation}. We take the top two synthetic parallel examples with the highest mT5Score in the filtering step and use those as synthetic few-shot examples in the prompt in the self-amplification step. \subsection{Translating with an Ensemble of Prompts} Because the two examples used in the prompt can greatly affect the quality of the generated translations, some prompts containing low-quality synthetic examples may cause poor translations for certain sentences. To combat this and reduce variation in performance, we keep the top $N$ synthetic examples instead of two synthetic examples. We use these to form $\frac{N}{2}$ different few-shot prompts with two synthetic parallel examples each. Each sentence in the test set is then translated with these $\frac{N}{2}$ different prompts to produce $\frac{N}{2}$ translations. The best translation of the $\frac{N}{2}$ translations is chosen in a fully unsupervised manner with mT5Score, as done in the filtering step of the bootstrap. We find this ensembling technique helps make unsupervised zero-shot performance competitive with few-shot performance. Experiments varying the number of prompts in the ensemble can be found in Appendix \ref{sec:promptensemble-ablation}. Unless otherwise stated, we use a 4 prompt ensemble in this paper: $\frac{N}{2} = 4$. In sum, we sample and zero-shot translate 100 sentences from a monolingual corpus, keep the top eight synthetic parallel examples scored by mT5Score, and use them to form four few-shot prompts, each of which has two synthetic examples. \fewshotfloresresultstableabbrev \subsection{English-centric Bootstrapping} \label{sec:englishcentric} While \citet{gpt3unsupervised} only performed a bootstrap on English-French and French-English pairs, we perform bootstrapping on some language pairs which may contain at least one low-resource language or non-English language. It has been found that multilingual language models perform best in English, due to imbalance of languages in the pre-training corpus where English has the highest amount of data \citep{xglm}. Therefore, when running the bootstrap on various language pairs, we modify the bootstrap to favor generating English, or pivot through English when neither the source nor target language is English. Ablation experiments can be found in Appendix \ref{sec:englishcentric-ablation}. We outline examples of our modified English-centric bootstrapping process for various language pairs in Appendix \ref{sec:englishcentric-examples}. \subsection{Results} We report results with the same method used for our few-shot evaluation. Abbreviated results can be found in Table \ref{table:fewshot-flores-results-abbrev} and the matrix of full results can be found in Appendix \ref{sec:flores-results}. In this unsupervised setting, we find our zero-shot results exceed our 2-shot results; furthermore, they significantly exceed the performance of the XGLM and GPT-3 results reported in \citet{xglm} on an average spBLEU score over all 182 pairs (+1.0 spBLEU). Again, we note strong performance on language pairs that contain one or more low-resource languages. Intuitively, we can explain the zero-shot performance surpassing the few-shot performance through our use of prompt ensembling in the zero-shot setting. As prompt ensembling utilizes four prompts with two synthetic parallel examples each, it essentially uses eight synthetic examples, instead of just two real examples in the few-shot setting. Our synthetic examples are nearly as high-quality as real examples (similar to the findings of \citet{gpt3unsupervised}) as demonstrated by Appendix \ref{sec:promptensemble-ablation}. Prompt ensembling not only reduces performance variation if low-quality synthetic examples are selected during the bootstrap, but it also boosts performance beyond the few-shot setting as demonstrated by Table \ref{table:firstword-ablation} and the Appendix \ref{sec:promptensemble-ablation} experiments (Russian-English 26.9 $\rightarrow$ 27.9 spBLEU). In Appendix \ref{sec:zeroshot-results}, we also evaluate on WMT14~\citep{wmt14} to compare with the results reported in \citet{gpt3unsupervised} using $\text{GPT-3}_\text{175B}$. \section {Other Language Generation Tasks} \label{sec:other-tasks} We next demonstrate that bidirectional models have a generalized ability, beyond machine translation, to be prompted for arbitrary tasks. We evaluate their performance on question answering and summarization language generation tasks. Example generations can be found in Appendix \ref{sec:examplegenerations}. \subsection{Question Answering} We compare the zero-shot question answering performance of mT5 against XGLM on the XQuAD dataset \citep{xquad}, a multilingual question answering dataset, in Table \ref{table:xquad}. We find mT5 with \textsc{Sap} outperforms XGLM significantly (+1.7 EM, +12.3 F1). In Table \ref{table:squad}, we also compare against T5+LM~\citep{lester-etal-2021-power}. As T5+LM is English-only, we compare using the English-only SQuAD v1.1 dataset \citep{squad}. We still utilize the multilingual mT5 with \textsc{Sap} due to observations that the English-only T5 v1.1 model does not perform as well as mT5 in prompt-based learning\footnote{We discuss this observation in more detail in Appendix \ref{sec:t5v11observation}.}. \textsc{Sap} achieves +6.7 EM and +5.6 F1 over T5+LM. \textsc{Sap}, as an iterative technique, is useful for producing long generations from a bidirectional model for tasks such as machine translation. We find, however, it still has utility on tasks like question answering where answer generations are shorter spans of text. We ablate utilizing \textsc{Sap} with mT5 against the simple approach of prompting mT5 once and using the mask in-fill generated on SQuAD v1.1. In the few-shot (16-shot) setting, we find that utilizing \textsc{Sap} still markedly improves performance (+12.5 EM, +5.5 F1) even on short-form generation tasks like question answering. \xquadtable \squadandsummarizationtable \subsection{Summarization} We next perform summarization on the CNN/Daily Mail dataset \citep{cnndailymail,cnndailymail2,cnndailymail3} as another long-form text generation task. We compare mT5 with T5+LM and ablate the usage of \textsc{Sap} once again in Table \ref{table:summarization}. In the few-shot setting, we find a significant lead against T5+LM (+7.1 ROUGE-L). Of that +7.1 ROUGE-L boost, the ablation of our usage of \textsc{Sap} finds the technique itself is responsible for a large component of the boost (+5.3). \section {Conclusion and Future Directions} \label{sec:conclusion} We demonstrate \textsc{Sap} with the bidirectional mT5 model enables few-shot and zero-shot machine translation and zero-shot multilingual question answering, outperforming unidirectional models despite using far fewer parameters and examples. Our results suggest that the bidirectional representations learned by models such as mT5 contribute to this improved performance. Still, we concede that our results do not conclusively prove bidirectionality explains the difference in performance. Beyond bidirectionality and pre-training objectives, mT5, XGLM, and GPT-3 further differ in architecture, pre-training corpus, and hyperparameters. A complete ablation experiment would be computationally expensive, and we leave this as future work. The main limitation of \textsc{Sap} lies in its computational efficiency, discussed further in Appendix \ref{sec:limitations} along with potential mitigations. Importantly, these results demonstrate bidirectional models possess few-shot in-context learning and zero-shot instruction following capabilities innately, without the post-hoc modifications required by prior work. Our results finally contribute strong evidence towards the strength and efficiency of bidirectional pre-training objectives and motivate further research into bidirectional architectures, pre-training objectives, and language models designed and optimized for prompting and few-shot learning. We hypothesize these future bidirectional training schemes could yield an approach that overcomes the efficiency limitations of \textsc{Sap}, while maintaining the performance and parameter size reduction benefits. Concurrent recent work that compares or mixes unidirectional and bidirectional pre-training objectives \citep{bigsciencearchobjective,unifying,alexatm} already provide some early evidence towards this hypothesis.

9622cbc0ea01299260b90241d4dcb52540d386cb

\section{Introduction} \label{section:intro} In recent years, deep neural networks (DNNs) have grown in size from 3,246 trainable parameters in 1989 (LeNet \cite{LetNet}) to tens of millions of parameters (AlexNet \cite{AlexNet}, ResNet \cite{he2015resnet}). This has led to a massive increase in networks' ability to capture complex patterns from input data. However, this has also led to the risk of overfitting training data. As an extreme case, Zhang et al. \cite{Overfitting} showed that any sufficiently large network can memorize the labels for instances in a dataset, even if they are randomly assigned. This effect is more pronounced when there is insufficient training data for a network or the labels associated with the data are noisy. The latter case is common when training data and their annotations are automatically crawled from the Internet. Conceptually, an instance of training data could have two types of features: \begin{itemize} \item \emph{Class-wise Discriminative Features}: These are features that are useful for determining the class that each sample belongs to. For example, a picture of a cat may be identified as such by the presence of whiskers. \item \emph{Identity-wise Discriminative Features}: These are features that are useful for determining precisely which sample is which. For example, the length and number of whiskers would help to identify a specific cat. \end{itemize} A deep neural network can use both features to perform empirical risk minimization. However, the second type of feature is less likely to generalise well to unseen data. Relying on it could cause overfitting. Motivated by the above insight, this paper proposes a method to encourage the learning of Class-wise Features whilst discouraging the learning of Identity-wise Features. Intuitively, this leads to classifiers only using those features that are both discriminative across classes and common within each class. We do this by assigning each sample a unique ID and training a network that can classify samples whilst failing to determine the individual identity of each sample. This proposed process is performed adversarially in a manner akin to the DANN method \cite{ganin2015domainadversarial, ganin2015unsupervised} in domain adaptation. DANN maintains a domain classifier that can identify the domain of the input data while adversarially learning features that can reduce the discriminative power of the domain classifier. In our case, we essentially treat each individual sample as its own domain. In order to perform adversarial training, it was necessary to implement a gradient reversal layer (GRL) as described by Ganin et al. \cite{ganin2015domainadversarial}. However, proper adversarial training requires a balance between the regular backpropagation of the classification task and the reversed backpropagation of the sample identification task. This requires careful fine-tuning. While it is possible to discover a training schedule that maintains this balance, it is tricky and subject to error. We therefore also propose a method called Dynamic Gradient Reversal (DGR), which requires no tuning. Experimentation shows that DGR can be used anywhere a GRL is used. We further explore two use cases for ASIF. The first one is to improve the generalisation performance of a deep neural network when training with a small amount of data. The second one is to increase the resilience against inaccurately labelled training data. In the second case, we show that the proposed method can be directly applied as a regularization approach or can be used to identify the incorrect class labels. In short, the main contributions of this paper are: \begin{itemize} \item A technique to reduce the tendency for large networks to overfit to specific samples in the training data. \item A training method that maintains or improves accuracy while reducing the number of class-variant features to the absolute minimum. \item A gradient reversal algorithm that can be used anywhere a DANN-style Gradient Reversal Layer is used. \end{itemize} The source code behind these experiments is available at \url{https://github.com/avichapman/identity-feature-suppression}. The rest of the paper is organised as follows. In Sect. \ref{RelatedWorks}, we provide further background in the problem area. We then describe the methodology behind our approach in Sect. \ref{Method}. In Sect. \ref{Experiments}, we describe the experiments performed to explore the characteristics of ASIF. Finally, in Sect. \ref{Conclusions} we summarise our results and suggest fruitful directions for future work. \section{Related Works} \label{RelatedWorks} \subsection{Domain Adaptation} Domain Adaptation is a family of techniques for learning a task on data from one domain and applying that task to another domain \cite{Wilson2020Survey}. Pan and Yang \cite{PanSurvey2010} define a 'domain' as consisting of a feature space and a marginal probability distribution of the features across the population. Domain Adaptation works on the assumption that the feature space stays the same. That is, the things about a data sample that can be measured never change. For example, in any given country it is possible to measure the probability of a pregnancy resulting in fraternal twins. However, the marginal probability distribution between domains can vary widely. Applied to the aforementioned example, this would mean that the probability of fraternal twins in one country may be very different from that same measurement in another country. Several recent works \cite{Dredze2009MultidomainLB, joshi-etal-2012-multi, hassan2018unsupervised} have attempted to take advantage of data across all domains to learn to predict within a single domain. Sebag et al. \cite{schoenauersebag2019multidomain} attempted to use adversarial learning to learn the distributions of the various domains in the training set. Wilson, Doppa and Cook \cite{Wilson2020} proposed a method called Convolutional deep Domain Adaptation model for Time Series data (CoDATS), which took advantage of adversarial learning to encourage a feature extractor network to learn features that were domain invariant. The CoDATS network consisted of three parts: a feature extractor, a classifier and a domain predictor. The classifier was trained against labelled data for a source domain. At the same time, the domain predictor was trained to predict whether a given sample belonged to the source domain or a target domain. They used a gradient reversal layer (DGR) to create an adversarial relationship between the two tasks. The DGR worked by passing a feature vector through untouched when feeding forward and reversing the gradient when backpropagating. Using this technique, Wilson et al. were able to demonstrate superior classification on the target domain, despite the network having never seen labels for the target domain. However, this technique is limited in that it requires each domain to have a label, which doesn't work in blurry edge cases or when the target domain is unknown at training time. \subsection{Learning from Noisy Labels} \label{section:RelatedWorksNoisyLabels} In real life datasets, perfect label accuracy is unrealistic. Human annotators make errors due to fatigue and other considerations. Moreover, labels are often obtained through means such as Amazon's Mechanical Turk or as pseudo-labels generated via semi-supervised means. Referred to as \emph{noisy labels}, these inaccurate labels can have a deleterious effect on training accuracy. Zhang et al. \cite{DBLP:journals/corr/ZhangBHRV16} demonstrated that a sufficiently complex DNN could learn a dataset with an arbitrarily high level of label noise. This overfitting behaviour negatively affects performance when evaluated on test data. Fr\'enay and Verleysen \cite{FrenayLabelNoiseSurvey} divided label noise into two types: \emph{Instance-independent} and \emph{Instance-dependent}. \emph{Instance-independent} label noise is characterised by the lack of a probabilistic relationship between a label being wrong and the underlying features of a given sample. In the literature, this is often further sub-divided into Symmetric and Asymmetric noise. \begin{itemize} \item Symmetric noisy labels are modelled by randomly changing label values from their true class to some other class with a certain probability. \item Asymmetric label noise is modelled by randomly changing the labels for samples of certain classes to a similar class with a given probability. An example of this would be changing 'cars' to 'trucks' in the CIFAR10 dataset. This is slightly more realistic. \end{itemize} \emph{Instance-dependent} label noise, on the other hand, varies with the particular characteristics of each sample. This mirrors reality, in that more ambiguous samples are more likely to be misclassified. Chen et al. \cite{Chen2021InstanceDependentNoise} described a method for producing realistic instance-dependent noise. Their method involved training a DNN on a dataset for a certain number of epochs and recording the associated loss of each sample. The samples with the highest averaged loss were deemed to be the most counter-intuitive samples and therefore the most likely to be mislabelled in real life. While this does not work as an online means of determining misclassification likelihood, it is more than sufficient when the training data is known beforehand. DNNs tend to learn general classification rules first before proceeding to memorise individual samples in a dataset. This means that early in training, a sample's classification loss can be used as an indication of whether the sample's label is incorrect. A large loss may indicate a label is wrong. This is often referred to as the \emph{small-loss trick}. Han et al. \cite{Han2018CoTeaching}, Jiang et al. \cite{Jiang2018Mentornet} and Yu et al. \cite{Yu2019Disagreement} all utilise the small-loss trick to detect noisy labels. A popular method for combating overfitting to noisy labels is to change the rate at which DNNs learn instance-specific cases by modifying the loss function. This means that samples with an abnormally high loss (and thus likely to be wrongly labelled) are ignored or have reduced affect on the training outcome. Zhang et al. \cite{GCE} proposed the Generalised Cross Entropy loss, which bridged the gap between Cross Entropy loss and the MAE/unhinged loss. Menon et al. \cite{PHUber} went further and proposed a partially Huberised Cross Entropy loss, which utilized gradient clipping to arrive at a more robust training solution. Both of these methods are included for comparison in this paper. \section{Method} \label{Method} This section describes the proposed ASIF training method in further detail. \subsection{Notations and Definitions} The following notation will be used in this paper: \begin{itemize} \item $N$: The total number of samples being trained on. \item $C$: The total number of classes in the dataset. \item $N_c$: The total number of samples of a given class. \item $B$: The total number of samples in each mini-batch. \item $\eta$: The amount of label noise, 0-1. \item $\mathscr{L}_{\text{id}_c}$: The identification task loss associated with class $c$. \item $\mathscr{L}_\text{cls}$: The classification task loss. \item $\Psi$: The Feature Extractor. \item $F_\Psi$: The dimension of the feature vector output from $\Psi$. \item $h_I$: The Identifier module. \item $h_C$: The Classifier module. \end{itemize} \subsection{Method Description} As discussed in Section \ref{section:intro}, ASIF attempts to overcome the problem of networks overfitting to specific instances in the training data. To that end, it attempts to select \emph{Class-wise Features} while suppressing \emph{Identity-wise Features}. Recall that the former are features useful in classifying a sample into one of several classes, while the latter features are useful in determining the identity of a specific sample. ASIF attempts to perform both tasks: \emph{Classification} and \emph{Identification}. \emph{Classification} divides the dataset into $C$ classes and attempts to determine the class that each sample belongs to. Similarly, \emph{Identification} divides the dataset into $N$ 'identities' (one for each sample) and attempts to ascertain the identity of each and every sample. The module contains global parameters as well as parameters that are tuned for each class to perform identification amongst the individual samples within that class. A gradient reversal layer exists to encourage the network to fail in its \emph{Identification} task. \subsection{Network Structure} \begin{figure*} \centering \includegraphics[width=\textwidth]{asif_model_horizontal.png} \caption{Architecture of our proposed ASIF network. \emph{(a)} A shared Feature Extractor $\Psi$ extracts features for use by all downstream tasks. A Classifier module $h_C$ performs the \emph{Classification} task, while an Identifier module $h_I$ performs the \emph{Identification} task. \emph{(b)} The Identifier module contains shared parameters that are trained on all samples, as well as dedicated parameters for each class of samples producing $C$ outputs, one for each class. There is a Dynamic Gradient Reversal (DGR) layer between the shared and per-class parameters.} \label{fig:asif_model} \end{figure*} As Figure \ref{fig:asif_model} indicates, the ASIF network consists of several components: \subsubsection{Feature Extractor} The Feature Extractor ($\Psi$) can be any off-the-shelf network. For the purposes of the experiments we've done as part of this paper, we used ResNet18 \cite{he2015resnet}. For ease of base-lining, we used the ResNet18 implementation provided as part of Nishi et al.'s \cite{Nishi2021Augmentation} work. The output of the feature extractor is a feature vector $\Psi(x)$, where $\Psi(x) \in \mathbf{R}^{F_\Psi}$. This is passed to a linear layer $h_C$, the 'Classifier' in Figure \ref{fig:asif_model}(a), to extract logits for use with the \emph{Classification} task. The logits are combined with the classification label using a Cross Entropy loss ($\mathscr{L}_\text{cls}$). Note that the loss applied here can be varied. Investigation of other losses for the Classification task is left to future work. \subsubsection{Identification Module} The Identification Module $h_I$ attempts to identify individual samples. It has $C$ outputs, one for each class, where $h_{I_c}(\Psi(x)) \in \mathbf{R}^{N_c}$. As shown in Figure \ref{fig:asif_model}(b), the module is divided into three parts: a 'public' part whose parameters are trained on all samples regardless of label, a DGR, and $C$ sets of private parameters that are only trained on samples with the matching label. Each set of private parameters $1..C$ outputs one of the outputs $h_{I_c}(\Psi(x))$. By sharing as many parameters as possible between the classes, we allow for a network that more easily scales to a larger number of classes. The 'public' section of the Identifier consists of two linear layers with a Batch Normalization, ReLU and dropout in between. The 'private' section has a Batch Normalization, ReLU, dropout and final linear layer for each class. All hyperparameters, such as dropout levels, were manually optimised using cross-validation. Each mini-batch output from the Identifier is filtered down to just the batch members from its class and a Cross-Entropy Loss ($\mathscr{L}_\text{id}$) is applied. The losses are then averaged in proportion to each class' share of the mini batch and combined with the classification loss. This structure allows the ASIF network to optimise: \begin{equation} \label{eqn:asif_loss} \min_{\Psi, h_C}\max_{\{{h_I}_c\}}\mathbb{E}[h_C(\Psi(x))] \\ + \lambda_\text{id}\mathbb{E}[\frac{1}{C}\sum\limits_{c=1}^{C} {h_I}_c(\Psi(x))], \end{equation} where ${h_I}_c$ denotes the Identifier output for class $c$. It also includes a coefficient $\lambda_\text{id}$ for the Identifier losses. The values used in the experiments can be found in the appendix. \subsubsection{Dynamic Gradient Reversal Layers} Similar to DANN \cite{ganin2015domainadversarial}, we reverse the gradient during backpropagation from the \emph{Identification} task. To review, DANN contains a Gradient Reversal Layer (GRL) which, on feeding forward has the value $R(x) = x$. However, when backpropagating it has the value $\frac{\partial R}{\partial x} = -\lambda I$, where $I$ is the identity matrix and $\lambda$ is a hyperparameter. The value of $\lambda$ is set according to a schedule tuned to ensure a proper balance is maintained between the competing tasks. Unfortunately, choosing an incorrect value for $\lambda$ leads to the \emph{Identification} task becoming confidently wrong. To overcome this limitation, this paper proposes a Dynamic Gradient Reversal (DGR) scheme. Unlike DANN, our method does not require a tune-able hyper-parameter. Moreover, our experiment shows that DGR alleviates overfitting better when compared with DANN-like gradient reversal layers, as shown in Figure \ref{fig:dgr_vs_dann}. \begin{figure} \centering \includegraphics[width=3.2in]{dgr_vs_dann.png} \caption{Classification loss, training and test accuracy versus training epoch when training with Symmetrical 80\% noise and CIFAR10. The use of Dynamic Gradient Reversal (DGR) leads to reduced overfitting, as indicated by the classification loss not dropping as fast. Also note that the DGR training and test accuracies remain in lockstep, while the DANN accuracies diverge during later training.} \label{fig:dgr_vs_dann} \end{figure} \begin{algorithm} \caption{Dynamic Gradient Reversal}\label{alg:dgr} \begin{algorithmic} \State $\hat{\mathscr{L}_\text{id}} \gets log(N_c)$ \State $\lambda \gets 1.0$ \While{Still Training} \State Perform \emph{Identification Task} \State Backpropagate using $\lambda$ \State Record $\mathscr{L}_\text{id}$ \State $\lambda \gets \frac{\mathscr{L}_\text{id} - \hat{\mathscr{L}_\text{id}}}{\hat{\mathscr{L}_\text{id}}}$ \EndWhile \end{algorithmic} \end{algorithm} Formally, Dynamic Gradient Reversal (DGR) is designed to maintain the \emph{Identification Task} in a state of maximum uncertainty. This is done by establishing a baseline ideal loss ($\hat{\mathscr{L}_\text{id}}$) corresponding to maximum entropy in the identity logits as shown in Equation \ref{eqn:max_uncertainty}: \begin{equation} \label{eqn:max_uncertainty} \hat{\mathscr{L}_\text{id}} = -\sum_{I=1}^{N_c} \frac{1}{N_c} log(\frac{1}{N_c}) = log(N_c). \end{equation} Then we automatically choose $\lambda$ by comparing the current $\mathscr{L}_\text{id}$ and $\hat{\mathscr{L}_\text{id}}$. The detailed process of calculating $\lambda$ is described in Algorithm \ref{alg:dgr}. The algorithm has the same computational cost as DANN, since the tuning of the $\lambda$ value is done using the loss, which is computed anyway. \subsection{Noisy Label Detection} One of the challenges of obtaining real-life data is the difficulty of creating accurate labels. It is typical to use a large human workforce and/or some degree of automation (for example, scraping the Internet) to acquire the labels. Inaccurate (or 'noisy') labels often become associated with the training data. This can lead to reduced accuracy in trained networks. To detect noisy labels, the \emph{small-loss trick} \cite{Han2018CoTeaching, Jiang2018Mentornet, Yu2019Disagreement} can be applied, which is described above in Section \ref{section:RelatedWorksNoisyLabels}. DNNs learn general cases first and high losses indicate tricky cases or bad labels. Given ASIF's regularising effect during training, the feature extractor will take much longer to overfit to the edge cases. We thus theorise that the small-loss trick applied to the Classification Task's loss $\mathscr{L}_\text{cls}$, would be far more robust when trained with ASIF than without. \section{Experiments} \label{Experiments} In this section, we evaluate the performance of ASIF using reduced training sets and noisy labels. We also delve into the characteristics of ASIF. Experiments were run using both the CIFAR10 and Fashion-MNIST \cite{FashionMNIST} datasets. All training techniques are judged based on their macro F1 scores when classifying on the test set. The CIFAR10 dataset contains 50,000 small 32x32 sample images, split evenly across ten classes. It contains a further 10,000 images for evaluation - 1000 per class. The Fashion-MNIST dataset contains 60,000 small 28x28 grayscale images of fashion products, split evenly across ten classes. It contains a further 10,000 images for evaluation. To serve as a point of comparison, all experimental configurations were also run with several other methods. In these non-ASIF cases, the Identifier module was removed from the network. Three different losses were applied to the output logits from the Classifier: \begin{itemize} \item CE: Cross-Entropy. This is the same as used in the ASIF experiments. \item GCE: Generalised Cross-Entropy \cite{GCE} \item PHuber: partially Huberised Cross-Entropy \cite{PHUber} \end{itemize} \subsection{Reduced Training Sets} \label{ReducedTrainingSets} To investigate ASIF's ability to resist overfitting with small training sets, we designed an experiment which trained the network on the CIFAR10 dataset with various numbers of samples per class. The inputs to all techniques were the same, with standard regularisation applied, but no data augmentation. Runs were conducted with $N$ = [10k, 20k, 30k, 40k, all]. Each configuration was run three times and their macro F1 scores across the test set averaged. \begin{table}[] \centering \begin{tabular}{||c c c c c||} \hline $N$ & CE & GCE & PHuber & ASIF \\ [0.ex] \hline\hline 10k & 69.1 \textpm \, 1.2 & 64.8 \textpm \, 2.4 & 73.2 \textpm \, 0.6 & \textbf{74.2 \textpm \, 0.2} \\ \hline 20k & 74.0 \textpm \, 0.5 & 72.6 \textpm \, 1.1 & 79.4 \textpm \, 0.0 & \textbf{80.2 \textpm \, 0.1} \\ \hline 30k & 78.7 \textpm \, 0.4 & 75.6 \textpm \, 0.7 & 82.1 \textpm \, 0.1 & \textbf{83.6 \textpm \, 0.3} \\ \hline 40k & 81.5 \textpm \, 0.7 & 80.6 \textpm \, 0.1 & 83.4 \textpm \, 0.1 & \textbf{85.5 \textpm \, 0.3} \\ \hline 50k & 84.3 \textpm \, 0.6 & 83.0 \textpm \, 0.4 & 83.9 \textpm \, 0.4 & \textbf{86.8 \textpm \, 0.4} \\ [1ex] \hline \end{tabular} \caption{Macro F1 Scores when training on reduced training sets on CIFAR10.} \label{tab:small_set_cifar10_results} \end{table} We discovered that ASIF has a marked advantage over its competitors in this domain. Table \ref{tab:small_set_cifar10_results} summarises the results for CIFAR10. Fashion-MNIST results can be found in Table \ref{tab:small_set_fashion_mnist_results} in the appendix. \subsection{Training with Label Noise} To test ASIF in the presence of label noise, we intentionally modified the labels in the datasets. Two types of label noise were investigated. The first was Symmetrical instance-invariant noise, as described by Zhang et al. \cite{DBLP:journals/corr/ZhangBHRV16} and by Zhang and Sabuncu \cite{DBLP:journals/corr/abs-1805-07836}. To apply this sort of noise, a percentage of the dataset corresponding to the desired noise level $\eta$ were selected for label modification. Symmetrical instance-invariant noise is not realistic, since labelling errors are more likely to happen with ambiguous samples than with obviously distinct ones. We used the technique described by Chen et al. \cite{Chen2021InstanceDependentNoise} to produce realistic instance-dependent noise. Their process produced a list of samples and associated average losses. We ranked the samples by descending loss and selected the top $N \times \eta$ samples. Those samples then had their labels randomly swapped. Having trained ASIF against six values of $\eta$ and two different noise techniques, we have been able to show superior performance in the very high noise domain around 70\%-90\% noise. Figure \ref{fig:inst_noise_vs_accuracy_improvement} clearly illustrates ASIF's accuracy advantage in that area with CIFAR10. \begin{figure} \centering \includegraphics[width=3.2in]{sym_noise_vs_accuracy_improvement.png} \caption{CIFAR10 results. ASIF confers a clear accuracy improvement over the baseline training method (Cross Entropy) with noisy labels, especially in the high-noise end of the range. GCE and PHuber training results are included for comparison.} \label{fig:inst_noise_vs_accuracy_improvement} \end{figure} In the easier scenario of Symmetric Instance-Invariant noise, PHuber is still competitive with ASIF. However, even in those cases, ASIF beats CE and GCE by a wide margin. Please see Tables \ref{tab:noise_cifar10_inst} and \ref{tab:noise_cifar10_sym} for CIFAR10 full results. ASIF's corresponding performance with Fashion-MNIST can be found in Tables \ref{tab:noise_fashion_mnist_sym} and \ref{tab:noise_fashion_mnist_inst} the appendix. \begin{table}[] \centering \begin{tabular}{||c c c c c||} \hline $\eta$ & CE & GCE & PHuber & ASIF \\ [0.ex] \hline\hline 0 & 84.3 \textpm \, 0.6 & 83.0 \textpm \, 0.4 & 83.9 \textpm \, 0.4 & \textbf{86.8 \textpm \, 0.4} \\ \hline 0.2 & 79.5 \textpm \, 0.5 & 79.1 \textpm \, 0.1 & 80.5 \textpm \, 0.3 & \textbf{81.5 \textpm \, 0.0} \\ \hline 0.4 & 70.4 \textpm \, 0.3 & 70.5 \textpm \, 0.1 & \textbf{74.2 \textpm \, 0.4} & 73.1 \textpm \, 0.6 \\ \hline 0.6 & 59.2 \textpm \, 0.7 & 58.4 \textpm \, 0.3 & \textbf{65.7 \textpm \, 0.1} & 65.6 \textpm \, 1.9 \\ \hline 0.7 & 52.8 \textpm \, 1.0 & 50.9 \textpm \, 0.6 & 44.1 \textpm \, 0.6 & \textbf{61.6 \textpm \, 0.9} \\ \hline 0.8 & 45.3 \textpm \, 1.4 & 42.4 \textpm \, 0.5 & 34.0 \textpm \, 0.2 & \textbf{53.5 \textpm \, 1.0} \\ \hline 0.9 & 38.2 \textpm \, 0.5 & 31.5 \textpm \, 0.9 & 27.4 \textpm \, 2.3 & \textbf{41.8 \textpm \, 1.3} \\ [1ex] \hline \end{tabular} \caption{Macro F1 Scores with Instance-Dependent Noisy Labels on CIFAR10.} \label{tab:noise_cifar10_inst} \end{table} \begin{table}[] \centering \begin{tabular}{||c c c c c||} \hline $\eta$ & CE & GCE & PHuber & ASIF \\ [0.ex] \hline\hline 0 & 84.3 \textpm \, 0.6 & 83.0 \textpm \, 0.4 & 83.9 \textpm \, 0.4 & \textbf{86.8 \textpm \, 0.4} \\ \hline 0.2 & 65.5 \textpm \, 0.5 & 75.1 \textpm \, 0.4 & \textbf{81.8 \textpm \, 0.1} & 77.6 \textpm \, 0.7 \\ \hline 0.4 & 49.0 \textpm \, 2.3 & 60.3 \textpm \, 1.1 & \textbf{78.6 \textpm \, 0.8} & 72.0 \textpm \, 0.4 \\ \hline 0.6 & 41.0 \textpm \, 1.1 & 48.3 \textpm \, 1.3 & \textbf{73.5 \textpm \, 0.3} & 63.3 \textpm \, 1.9 \\ \hline 0.7 & 29.7 \textpm \, 1.1 & 40.8 \textpm \, 0.5 & \textbf{67.3 \textpm \, 1.3} & 55.6 \textpm \, 0.5 \\ \hline 0.8 & 21.4 \textpm \, 0.7 & 29.3 \textpm \, 0.9 & \textbf{53.6 \textpm \, 4.4} & 43.1 \textpm \, 2.4 \\ \hline 0.9 & 5.0 \textpm \, 2.9 & 16.6 \textpm \, 0.5 & 17.7 \textpm \, 3.1 & \textbf{27.3 \textpm \, 2.6} \\ [1ex] \hline \end{tabular} \caption{Macro F1 Scores with Symmetric Instance-Invariant Noisy Labels on CIFAR10.} \label{tab:noise_cifar10_sym} \end{table} \subsection{Detecting Incorrect Labels} In addition to showing that ASIF confers an advantage when training in the face of noisy labels, we can also show that ASIF can help to detect which samples have bad labels. To do so, periodically through the training we recorded the associated $\mathscr{L}_\text{cls}$ values for each sample and then ranked them by descending value. The top $N \times \eta$ samples were selected as 'probably false'. These samples' labels were then compared with the true labels to determine the dirty label picking balanced accuracy. Our results indicate that utilising ASIF to pick out bad labels can confer as much as a 10\% advantage over the baseline CE loss. This is a significant improvement over both GCE and PHuber. In the easier scenario of Symmetric Instance-Invariant noise, ASIF shows a detection advantage of as much as 13\% over CE. Please see Tables \ref{tab:label_picking_cifar10_inst} and \ref{tab:label_picking_cifar10_sym} for full results. \begin{table}[] \centering \begin{tabular}{||c c c c c||} \hline $\eta$ & CE & GCE & PHuber & ASIF \\ [0.ex] \hline\hline 0.2 & 64.2 \textpm \, 0.8 & 65.0 \textpm \, 1.0 & 67.6 \textpm \, 1.1 & \textbf{70.0 \textpm \, 1.6} \\ \hline 0.4 & 73.4 \textpm \, 0.5 & 70.5 \textpm \, 1.4 & 72.6 \textpm \, 5.0 & \textbf{78.0 \textpm \, 1.6} \\ \hline 0.6 & 73.2 \textpm \, 1.6 & 67.7 \textpm \, 0.5 & 74.5 \textpm \, 0.6 & \textbf{79.7 \textpm \, 3.3} \\ \hline 0.7 & 69.7 \textpm \, 1.6 & 66.2 \textpm \, 1.1 & 64.5 \textpm \, 3.1 & \textbf{79.6 \textpm \, 2.0} \\ \hline 0.8 & 65.9 \textpm \, 0.7 & 60.7 \textpm \, 0.7 & 62.9 \textpm \, 3.6 & \textbf{71.2 \textpm \, 1.2} \\ \hline 0.9 & 58.1 \textpm \, 2.0 & 58.3 \textpm \, 0.9 & 61.1 \textpm \, 0.6 & \textbf{61.9 \textpm \, 3.0} \\ [1ex] \hline \end{tabular} \caption{F1 Score of Instance-Dependent Noisy Label Detection on CIFAR10} \label{tab:label_picking_cifar10_inst} \end{table} \begin{table}[] \centering \begin{tabular}{||c c c c c||} \hline $\eta$ & CE & GCE & PHuber & ASIF \\ [0.ex] \hline\hline 0.2 & 79.1 \textpm \, 3.1 & \textbf{86.7 \textpm \, 0.6} & 82.0 \textpm \, 2.0 & 85.6 \textpm \, 0.7 \\ \hline 0.4 & 75.9 \textpm \, 0.2 & 83.7 \textpm \, 0.3 & 84.6 \textpm \, 0.5 & \textbf{86.2 \textpm \, 1.3} \\ \hline 0.6 & 73.3 \textpm \, 0.9 & 77.6 \textpm \, 1.1 & 83.1 \textpm \, 1.4 & \textbf{84.1 \textpm \, 0.4} \\ \hline 0.7 & 67.1 \textpm \, 1.6 & 73.2 \textpm \, 1.1 & \textbf{79.7 \textpm \, 3.2} & 79.5 \textpm \, 1.3 \\ \hline 0.8 & 60.7 \textpm \, 0.4 & 65.7 \textpm \, 0.7 & 71.7 \textpm \, 0.6 & \textbf{73.2 \textpm \, 2.0} \\ \hline 0.9 & 52.2 \textpm \, 2.2 & 54.8 \textpm \, 0.9 & 59.4 \textpm \, 2.0 & \textbf{62.9 \textpm \, 0.9} \\ [1ex] \hline \end{tabular} \caption{F1 Score of Symmetric Instance-Invariant Noisy Label Detection on CIFAR10} \label{tab:label_picking_cifar10_sym} \end{table} \subsection{Analysis of the Features Learned With ASIF} \label{FeatureAnalysis} We conducted an analysis to gain a better understanding of the effect ASIF has on the features learned. First, we address the degree to which ASIF inhibits memorisation of features. The theory behind ASIF is that we penalise the learning of features that can be used to identify specific instances of data, while allowing the learning of features that are necessary to differentiate between classes of data. We tested this by training a single layer classifier on the feature vectors obtained from the frozen pre-trained feature extractor $\Psi(x)$ until performance stopped improving. The goal was correctly identifying each individual sample. If ASIF performed as theorised, the best loss obtained would be worse for ASIF than it would be for the baseline. The results, shown in Table \ref{tab:feature_vector_ids_losses}, confirm this supposition. \begin{table}[] \centering \begin{tabular}{||c c c||} \hline Dataset & CE & ASIF \\ [0.ex] \hline\hline CIFAR10 & 0.005 \textpm \, 0.008 & 0.782 \textpm \, 0.159 \\ \hline Fashion-MNIST & 1.059 \textpm \, 0.332 & 8.307 \textpm \, 0.202 \\ [1ex] \hline \end{tabular} \caption{Loss Obtained identifying feature vectors} \label{tab:feature_vector_ids_losses} \end{table} Second, we turn our attention to feature dis-entanglement. There is a tension behind the two goals of class discrimination and instance discrimination, since features can be used toward both ends. We hypothesize that the result of this struggle would be that only the fewest, most useful features would be selected for classification. This would act to dis-entangle class-wise and identity-wise features, with class-variant features confined to a small subspace and the rest of the features showing no significant statistical differences between classes. We set out to investigate the distribution of the features selected by ASIF. To do so, we performed an experiment. We trained a fresh linear classifier $h$ on the feature vectors obtained from the frozen pre-trained feature extractor $\Psi(x)$. We then used the absolute value of the linear weights $h_\Theta$ to ascertain which feature dimensions were least important to the classification and removed those from the vectors, before retraining another fresh classifier on the resulting vectors. We performed this process repeatedly, reducing the vector size with each iteration until the resulting vectors had only five dimensions. With each iteration, we recorded the best classification accuracy obtained. Figure \ref{fig:clustering_accuracy_vs_features} shows the results obtained from this experiment. In all cases, the majority of the features played no important role in classification. However, as the least important features continued to be removed, the Cross Entropy baseline became the first to show a reduced accuracy when there were around 55 feature dimensions remaining. \begin{figure} \centering \includegraphics[width=3.2in]{clustering_experiment_accuracy.png} \caption{The accuracy obtained when training a single-layer classifier on the output of feature extractors trained with Cross Entropy Loss (Red) and ASIF (Blue) using between 1 and 100 of the most important features.} \label{fig:clustering_accuracy_vs_features} \end{figure} ASIF, on the other hand, does not seriously lose accuracy until we are down to the ten most important feature dimensions. This feature compression is desirable as it has been linked to generalization \cite{tishby2015deep, shwartzziv2017opening}. \section{Conclusions and Future Work} \label{Conclusions} We have introduced ASIF, which differentiates between Class-wise and Identity-wise Discriminative Features, promoting the former whilst suppressing the latter. Through experimentation, we have shown that ASIF has a regularising effect that can reduce overfitting to individual samples and increase robustness against label noise. We have shown that ASIF can result in accuracies that are better than standard training results, especially in the high-noise domain. Moreover, we have also proposed the DGR method which allows for a gradient reversal layer without tune-able hyper-parameters. We have shown that the use of ASIF results in the selection of far fewer class-variant features. This feature of ASIF-trained feature extractors has great potential for use in Data Privacy and Anonymization applications, or anywhere else that autoencoders are useful. Moreover, when training on the full dataset, the selected features are likely to be domain invariant and lead to much better generalisation. Extending the Identification task to the unsupervised realm may encourage more domain invariant feature selection and seems like a fruitful direction for future research. ASIF has limitations when it comes to scaling up. Because its Identifier module has a linear layer that specifies individual instances in a fixed dataset, the size of that dataset has a practical upper limit. Because each class has its own set of weights, this memory constraint also places a limit on the number of classes trained. Moreover, while suppressing identity features has been shown to be helpful in the settings laid out in this paper, they are crucial in certain other tasks - such as face recognition. This limits ASIF's applicability in some areas. Despite these limitations, differentiating between Class-wise and Identity-wise Features is nonetheless a novel approach. \section*{Acknowledgment} This work was supported with supercomputing resources provided by the Phoenix HPC service at the University of Adelaide. \section{CIFAR10 ASIF Configurations} \label{appendix:cifar10_asif_config} Table \ref{tab:cifar10_asif_experiment_configs} shows the configurations used to produce the CIFAR10 ASIF results. \begin{table}[] \centering \begin{tabular}{||c c c c c||} \hline Noise & $\eta$ & N & LR & $\lambda_\text{if}$ \\ [0.ex] \hline\hline Instance & 0.2 & 50k & 0.0001 & 0.01 \\ \hline Instance & 0.4 & 50k & 0.0001 & 0.1 \\ \hline Instance & 0.6 & 50k & 0.0001 & 0.1 \\ \hline Instance & 0.7 & 50k & 0.0001 & 0.1 \\ \hline Instance & 0.8 & 50k & 0.0001 & 0.1 \\ \hline Instance & 0.9 & 50k & 0.0001 & 0.1 \\ \hline Symmetric & 0.2 & 50k & 0.0001 & 10.0 \\ \hline Symmetric & 0.4 & 50k & 0.0001 & 100.0 \\ \hline Symmetric & 0.6 & 50k & 0.0001 & 1.0 \\ \hline Symmetric & 0.7 & 50k & 0.0001 & 1.0 \\ \hline Symmetric & 0.8 & 50k & 0.001 & 1.0 \\ \hline Symmetric & 0.9 & 50k & 0.001 & 10.0 \\ \hline None & 0 & 10k & 0.001 & 1.0 \\ \hline None & 0 & 20k & 0.0001 & 1.0 \\ \hline None & 0 & 30k & 0.0001 & 1.0 \\ \hline None & 0 & 40k & 0.0001 & 1.0 \\ \hline None & 0 & 50K & 0.001 & 1.0 \\ [1ex] \hline \end{tabular} \caption{CIFAR10 ASIF Experimental Configurations.} \label{tab:cifar10_asif_experiment_configs} \end{table} \section{Fashion-MNIST ASIF Configurations} \label{appendix:cifar10_fashion_mnist_config} Table \ref{tab:fashion_mnist_asif_experiment_configs} shows the configurations used to produce the Fashion-MNIST ASIF results. \begin{table}[] \centering \begin{tabular}{||c c c c c||} \hline Noise & $\eta$ & N & LR & $\lambda_\text{if}$ \\ [0.ex] \hline\hline Instance & 0.2 & 60k & 0.001 & 0.001 \\ \hline Instance & 0.4 & 60k & 0.0001 & 0.1 \\ \hline Instance & 0.6 & 60k & 0.001 & 0.001 \\ \hline Instance & 0.7 & 60k & 0.001 & 0.001 \\ \hline Instance & 0.8 & 60k & 0.001 & 1.0 \\ \hline Instance & 0.9 & 60k & 0.001 & 0.001 \\ \hline Symmetric & 0.2 & 60k & 0.0001 & 0.001 \\ \hline Symmetric & 0.4 & 60k & 0.001 & 0.001 \\ \hline Symmetric & 0.6 & 60k & 0.001 & 0.001 \\ \hline Symmetric & 0.7 & 60k & 0.001 & 1.0 \\ \hline Symmetric & 0.8 & 60k & 0.0001 & 10.0 \\ \hline Symmetric & 0.9 & 60k & 0.0001 & 0.1 \\ \hline None & 0 & 10k & 0.001 & 0.001 \\ \hline None & 0 & 20k & 0.001 & 0.001 \\ \hline None & 0 & 30k & 0.001 & 0.001 \\ \hline None & 0 & 40k & 0.001 & 0.001 \\ \hline None & 0 & 60K & 0.001 & 0.01 \\ [1ex] \hline \end{tabular} \caption{Fashion-MNIST ASIF Experimental Configurations.} \label{tab:fashion_mnist_asif_experiment_configs} \end{table} \section{Fashion-MNIST Detailed Results} \label{appendix:fashion_mnist_detailed_results} Experiments run on CIFAR10 were also performed on the Fashion-MNIST dataset to test repeatability of the results. Unlike in the case of CIFAR10, tests were only run on Cross Entropy and ASIF, not on GCE or PHuber. Results are shown here: \begin{itemize} \item Reduced Datasets: Table \ref{tab:small_set_fashion_mnist_results} \item Symmetrical Instance-Invariant Noise: Table \ref{tab:noise_fashion_mnist_sym} \item Instance-Dependent Noise: Table \ref{tab:noise_fashion_mnist_inst} \end{itemize} \begin{table}[] \centering \begin{tabular}{||c c c||} \hline $N$ & CE & ASIF \\ [0.ex] \hline\hline 10k & 89.1 \textpm \, 0.2 & \textbf{89.6 \textpm \, 0.1} \\ \hline 20k & 90.5 \textpm \, 0.1 & \textbf{91.2 \textpm \, 0.3} \\ \hline 30k & 91.4 \textpm \, 0.1 & \textbf{91.9 \textpm \, 0.1} \\ \hline 40k & 92.0 \textpm \, 0.1 & \textbf{92.5 \textpm \, 0.3} \\ \hline 60k & 93.0 \textpm \, 0.0 & \textbf{93.1 \textpm \, 0.2} \\ [1ex] \hline \end{tabular} \caption{Macro F1 scores when training on reduced training sets on Fashion-MNIST.} \label{tab:small_set_fashion_mnist_results} \end{table} \begin{table}[] \centering \begin{tabular}{||c c c||} \hline $\eta$ & CE & ASIF \\ [0.ex] \hline\hline 0 & 93.0 \textpm \, 0.0 & \textbf{93.1 \textpm \, 0.2} \\ \hline 0.2 & 88.1 \textpm \, 0.0 & \textbf{90.3 \textpm \, 0.4} \\ \hline 0.4 & 83.7 \textpm \, 0.1 & \textbf{88.8 \textpm \, 1.1} \\ \hline 0.6 & 74.9 \textpm \, 1.1 & \textbf{86.9 \textpm \, 0.8} \\ \hline 0.7 & 66.7 \textpm \, 2.2 & \textbf{85.3 \textpm \, 0.8} \\ \hline 0.8 & 57.3 \textpm \, 4.8 & \textbf{81.4 \textpm \, 2.3} \\ \hline 0.9 & 19.4 \textpm \, 3.5 & \textbf{73.2 \textpm \, 0.4} \\ [1ex] \hline \end{tabular} \caption{Macro F1 Scores when training on Fashion-MNIST with Symmetric Instance-Invariant Noisy Labels.} \label{tab:noise_fashion_mnist_sym} \end{table} \begin{table}[] \centering \begin{tabular}{||c c c||} \hline $\eta$ & CE & ASIF \\ [0.ex] \hline\hline 0 & 93.0 \textpm \, 0.0 & \textbf{93.1 \textpm \, 0.2} \\ \hline 0.2 & \textbf{85.6 \textpm \, 0.2} & 82.5 \textpm \, 4.3 \\ \hline 0.4 & \textbf{74.7 \textpm \, 0.6} & 72.4 \textpm \, 1.9 \\ \hline 0.6 & 53.4 \textpm \, 0.6 & \textbf{58.3 \textpm \, 1.5} \\ \hline 0.7 & 43.1 \textpm \, 1.0 & \textbf{44.8 \textpm \, 1.6} \\ \hline 0.8 & 30.5 \textpm \, 0.6 & \textbf{34.3 \textpm \, 0.8} \\ \hline 0.9 & 21.4 \textpm \, 1.6 & \textbf{24.7 \textpm \, 2.4} \\ [1ex] \hline \end{tabular} \caption{Macro F1 Scores when training on Fashion-MNIST with Instance-Dependent Noisy Labels.} \label{tab:noise_fashion_mnist_inst} \end{table} \end{document}

f90a2030da5956afe08f646b575dd5e0e074ace9

\section{Introduction} \label{s:Introduction} Emulsions are suspensions of immiscible liquids (such as oil and water) and play a central role in a wide range of industrial processes such as food processing~\citep{walker2015development, zhang2015influence,guzey2006formation}, pharmaceutical processes~\citep{spernath2006microemulsions} or oil production~\citep{dicharry2006stability,angardi2021critical}. Moreover, current research is examining the application of fuel--water emulsions for more efficient and environmentally friendly power generation. Examples include gasoline--water direct injection (GWDI) for future gasoline engines~\citep{hoppe2017evaluation,heinrich2017gasoline} or fuel--water emulsions for small gas turbines~\citep{chmielewski2020combustion} and diesel engines~\citep{welscher2021comprehensive}. In particular for power generation applications, a better understanding of the stability of emulsions and the timescale of the segregation process is of central importance. To this end, we numerically study the segregation of emulsions in decaying turbulence under gravity. The formation of an emulsion requires energy input in form of kinetic energy, to deform and break up droplets. For a non-iso-density emulsion, the mixing of the lighter and the heavier phase requires additional energy input. In order to keep an emulsion stable, a continuous supply of energy is then required. Without further energy input, emulsions are unstable due to the natural tendency to minimize the potential and surface energy. To minimize the net potential energy, the heavier phase sinks, which is governed by the gravitational acceleration $g$ and the density difference between the two phases. To minimize the surface energy, droplets coalesce, thus reducing the interface area. A higher surface tension $\sigma$ leads to a higher variation of surface energy and thus increases the tendency for coalescence. Furthermore, it should be noted that interface minimization can be prevented by the presence of surfactants~\citep{kilpatrick2012water,goodarzi2019comprehensive}, such as the naturally occurring surfactants asphaltene and resins. Both rising and coalescence lead to segregation of emulsions, which can be quantified by the height of the lighter phase, i.e., the position of its center of mass in the direction of the gravitational acceleration, and the interface area, respectively. However, the two processes of rising and coalescence mutually interact with each other since larger droplets rise more easily and acceleration in one direction promotes coalescence. Experimental studies on the segregation process of emulsions mainly focus on chemical engineering aspects and monitor the height of the coalescing interface. There are several studies in the literature related to modeling of gravity assisted oil-water emulsion separation in oil production processes~\citep{dalingaros1987prediction, jeelani1998effect,jeelani1985prediction,henschke2002determination, jeelani2005creaming, lobo1993dispersion,aleem2020experimental}. The proposed models are designed for gravity settlers to separate the water during the oil production process and deliver correlations for the temporal evolution of zone heights. A review of these models is provided by \citet{frising2006liquid}. Most of these models require various input and modeling parameters and are very sophisticated. Further, limited optical access makes experimental studies of emulsion segregation processes challenging and requires advanced measurement techniques~\citep{koegl2020characterization}. For this reason, information on the interface area in segregating emulsions is hard to access. In this work, we want to complement these experimental studies with a numerical investigation for a generic configuration. Several numerical studies of emulsions and emulsification processes have been reported in the past, mostly focusing on droplet size distributions. First numerical emulsion studies employed Lattice-Boltzmann (LB) methods, such as e.g. \citet{Perlekar:2012ip,Skartlien:2013ec} and \citet{Mukherjee:2019ka}. More recently, \citet{CrialesiEsposito:2021ui} and \citet{Begemann:2022CJ} utilized direct numerical simulations (DNS) combined with the volume of fluid method (VOF) for these investigations. Moreover, several numerical studies focused specifically on the break-up of droplets such as e.g. \citet{Komrakova:2019kz} or \citet{Shao:2018gg}. The reverse process to emulsification, namely coalescence and rising of the lighter phase, was studied in the following papers: \citet{Dodd:2016eo} investigated droplet coalescence and droplet turbulence interaction in decaying turbulence and found that the energy release due to coalescence processes has an impact on the decay of the turbulent kinetic energy. \new{An important and central aspect of multiphase configurations is the effect of the buoyancy force due to a density difference between the phases and gravity. Previous numerical studies on the effect of the buoyancy force considered, for example, bubble-laden downflow configurations~\citep{brauer2021turbulent,Trautner:2021de,hasslberger2020direct} or rising bubbles~\citep{meller2022sub,hasslberger2018flow}. In addition, \citet{Saeedipour:2021gg} and \citet{estivalezes2022phase} have recently performed simulations of the phase inversion test case, where the lighter phase is initialized at the bottom of a box and rises due to gravitational acceleration. Despite the central importance of the buoyancy force on the segregation of non-iso-density emulsions, we are not aware of any previous numerical simulation studies on this. The effect of the buoyancy force on emulsion stability and emulsion segregation is of particular importance for various applications, especially with respect to emulsions in power generation (see above). With the present work, we aim to complement experimental studies~\citep{al2009experimental, sazonov2019untersuchung} with numerical investigations of emulsion stability and segregation. CFD simulations can overcome some of the limitations and challenges in experimentally characterizing emulsions~\citep{mcclements2007critical,koegl2020characterization} and provide new and more detailed insights.} The present work builds upon our recent paper on emulsification and emulsions~\citep{Begemann:2022CJ}. \new{Using the enhanced linear forcing approach proposed in our recent paper, we can generate a statistically stationary emulsion with a prescribed turbulent kinetic energy, and therewith obtain well-defined initial conditions for studying the segregation. In the present work, we} study the segregation of emulsions resembling oil--in--water \new{liquid--liquid} emulsions in terms of density ratio. In our study, we vary \new{the buoyancy force (by varying the gravitational acceleration $g$)} and the surface tension coefficient $\sigma$, the latter resulting in different droplet size distributions of the emulsions. Hence, we focus on the parameters affecting the segregation progress (minimization of potential and surface energy). For our studies, we use DNS with the finite volume approach and the VOF method. The emulsions are generated by a linear forcing of turbulence augmented with a PID controller~\citep{Begemann:2022CJ}. We then switch off the forcing and activate the gravitational acceleration and let the emulsions segregate in decaying turbulence under gravity. The paper is structured as follows. In \cref{s:Method}, we describe the computational method. \Cref{s:Configs} presents the considered configurations and the numerical setup. The results are presented in \cref{s:Results}, which firstly studies the segregation process in detail, then focuses on the energy releases to elucidate the dominant mechanisms and finally studies the timescale of the segregation. \Cref{s:Conclusions} summarizes the findings and draws conclusions. \section{Computational method} \label{s:Method} The simulations are conducted with the open source code PARIS (PArallel, Robust, Interface Simulator)~\citep{Aniszewski:2021cpc}. PARIS has been specifically designed for simulations of multiphase flows and is often used for studies of atomization processes, see e.g.~\citep{klein2017comparison,hasslberger2019flow,salvador2018analysis,ling2017spray} as well as other multiphase flow configurations~\citep{hasslberger2018flow}. The solver uses the single fluid formulation \citep{ProsperettiTryggvason:2007CompMeth} of the incompressible Navier-Stokes equations. The continuity and momentum equation are given as \begin{equation} \label{eq:continuity} \frac{ \partial u_{ i } }{ \partial x_{ i } } = 0\,, \end{equation} \begin{equation} \label{eq:momentum} \begin{split} & \rho \left( \frac{ \partial u_{ i } }{ \partial t } + \frac{ \partial u_{ i } u_{ j } }{ \partial x_{ j } } \right) = \\ & - \frac{ \partial p }{ \partial x_{ i } } + \frac{ \partial }{ \partial x_{ j } } \left[ \mu \left( \frac{ \partial u_{ i } }{ \partial x_{ j } } + \frac{ \partial u_{ j } }{ \partial x_{ i } } \right) \right] + \sigma n_{ i } \kappa \delta_{ s } + \rho g_{ i } \end{split} \end{equation} with the density $\rho$, the dynamic viscosity $\mu$, the $i^{th}$ velocity component $u_i$, the pressure $p$ and the gravitational acceleration $g_i$. In each cell, the density and viscosity values are linearly interpolated using the local volume fraction $\alpha$ of the dispersed phase, which is tracked with the geometrical VOF method~\citep{HirtNichols:1981jcp}: \begin{equation} \label{eq:densityViscosity} \rho=\alpha \rho_d + \left( 1-\alpha \right) \rho_c,\quad \mu=\alpha \mu_d + \left( 1-\alpha \right) \mu_c. \end{equation} The subscripts $d$ and $c$ denote the dispersed and the carrier phase, respectively. The Continuous-Surface-Force (CSF) approach \citep{BrackbillKotheZemach:1992jcp} determines the surface tension force from the surface tension coefficient $\sigma$, the interface normal $n_i=\frac{\partial \alpha}{\partial x_i}/| \nabla \alpha |$, the interface indicator function $\delta_S=|\nabla \alpha|$ and the interface curvature $\kappa=\frac{\partial n_i}{\partial x_i}$. The latter is accurately computed using a state-of-the-art height function approach~\citep{Popinet:2009jcp}. Details on its implementation in PARIS can be found in \citet{Aniszewski:2021cpc}. The advection of the VOF marker function is performed using a geometrical interface reconstruction algorithm. The respective transport equation is given as \begin{equation} \label{eq:VOFtransport} \frac{\partial \alpha}{\partial t}+u_i \frac{\partial \alpha}{\partial x_i}=0,\,\,\,\alpha= \begin{cases} 1, & \text{if $\pmb{x}$ is in dispersed phase.}\\ 0, & \text{if $\pmb{x}$ is in carrier phase.} \end{cases} \end{equation} A red-black Gauss-Seidel solver with overrelaxation is employed to solve the Poisson equation for pressure in the framework of the projection method. The simulation is advanced in time using a second-order predictor-corrector method. For the spatial discretization, the finite-volume approach is realized using a cubic grid. The velocity components are stored on a staggered grid, while the pressure and the VOF marker function, as well as the local densities and the viscosities resulting from the latter, are computed at the cell centers. The third-order Quadratic Upstream Interpolation for Convective Kinematics (QUICK) scheme~\citep{leonard1979stable} has been chosen to discretize the convective term of the momentum equation, while its viscous term is treated using central differences \newmk{and therewith second-order accurate.} \begin{figure*}[!htb] \centering \subfigure[]{\includegraphics[width=0.23\linewidth]{Fig1_setup_a}} \subfigure[]{\includegraphics[width=0.23\linewidth]{Fig1_setup_b_}} \hspace{0.2cm} \subfigure[]{\includegraphics[width=0.22\linewidth]{Fig1_setup_c}} \caption{ Simulation setup. (a)~Turbulent emulsion at statistically stationary state, (b)~forcing is turned off, gravitational acceleration $g$ is activated and slip walls are prescribed in the direction of the gravitational acceleration (orange walls), (c)~segregation under $g$. } \label{fig:sim_setup} \end{figure*}% To generate a turbulent emulsion in homogeneous isotropic turbulence (HIT), we employ the linear Lundgren forcing~\citep{Lundgren:2003uf} extended by a PID controller, see \citet{Begemann:2022CJ}. This extension provides a constant turbulent kinetic energy and accelerates the emulsification process. \begin{table}[!tb] \center \caption{Constant emulsion parameters.} \label{tab:Setup} \begin{tabular}{ccccccccc} \hline \multicolumn{1}{c}{$\phi$} & \multicolumn{1}{c}{$\rho_c$} & \multicolumn{1}{c}{$\rho_d$} & \multicolumn{1}{c}{$\nu_c=\nu_d $} & \multicolumn{1}{c}{$k$ } & \multicolumn{1}{c}{$\varepsilon$ } & \multicolumn{1}{c}{$Re_\lambda$}& \multicolumn{1}{c}{L} & \multicolumn{1}{c}{N} \\ - & $\si{kg/m^3}$&$\si{kg/m^3}$& $\si{m^2/s}$&$\si{m^2/s^2}$ &$\si{m^2/s^3}$& -&$\si{m}$& - \\ \hline 1/8 & 1 & 0.9 & 0.001 & 0.5 & 0.153 &104& $2\pi$ & 384\\%& 0.1 \hline \end{tabular} \end{table} \begin{table}[!tb] \center \caption{Considered cases. The baseline values \newmk{(BL)} are $\sigma_\mathrm{BL} = \num{2e-2}\,\si{N/m}$, $d_{H\mathrm{BL}} = 0.1468\,\si{m}$ and $g_\mathrm{BL}=4.59\,\si{m/s^2}$. Note that the correlation between $d_{H}$ and $\sigma$ for constant $\rho_c$ and $\varepsilon$ reads $d_{H} = \sigma^{3/5}$, see also \cref{eq:Hinze}. The last three columns contain the dimensionless segregation number $Seg$, the dimensionless energy release ratio $\Psi$ and the ratio $Seg/\Psi$, see \cref{ss:Mechanisms} for details.} \label{tab:CasesSeg} \begin{tabular}{lrrrrrrr} \hline \multicolumn{1}{l}{Case} & \multicolumn{1}{c}{\newnew{$We_l$}} & \multicolumn{1}{c}{$\sigma/\sigma_\mathrm{BL}$} & \multicolumn{1}{c}{$d_H/d_{H\mathrm{BL}}$} & \multicolumn{1}{c}{$g/g_\mathrm{BL}$}& \multicolumn{1}{r}{$\;\;\;Seg\;$} & \multicolumn{1}{c}{$\;\;\;\Psi\;$}& \multicolumn{1}{c}{$Seg/\Psi$}\\ \hline \textit{Baseline} & \newnew{21} & 1.0 & 1.00 & 1.0 & 10.6 & 2.5 & 4.2\\ \textit{Low}~$g$ & \newnew{21} & 1.0 & 1.00 & 0.5 & 5.3 & 1.3 & 4.2\\ \textit{High}~$g$ & \newnew{21} & 1.0 & 1.00 & 2.0 & 21.2 & 5.0 & 4.2\\ \textit{Low}~$\sigma$ & \newnew{70} & 0.3 & 0.49 & 1.0 & 18.3 & 4.6 & 4.0\\ \textit{High}~$\sigma$ & \newnew{4} & 5.0 & 2.63 & 1.0 & 5.6 & 1.8 & 3.2\\ \hline \end{tabular} \end{table} \section{Considered configurations} \label{s:Configs} In this paper, we study the segregation of emulsions under \new{different buoyancy forces obtained by varying gravitational acceleration $g$}. Further, we also consider emulsions with different droplet size distributions, which are obtained by a variation of the surface tension coefficient $\sigma$. Emulsions feature polydisperse droplet size distributions. A reference value for droplet size distributions in emulsions is given by the Hinze scale $d_H$~\citep{hinze1955fundamentals}, which is expected to be the most stable maximum droplet diameter in emulsions for HIT. $d_H$ is determined by \begin{equation} d_H= \left(We_{d,crit}/2 \right)^{3/5} \left( \rho_c/\sigma \right)^{-3/5}\varepsilon^{-2/5} \,, \label{eq:Hinze} \end{equation} where $We_{d,crit}$ denotes the critical \newnew{droplet} Weber number, for which we assume $We_{d,crit}= 1.17$ following \citet{hinze1955fundamentals} and recent numerical studies~\citep{CrialesiEsposito:2021ui, Mukherjee:2019ka}. Moreover, $\rho_c$ is the density of the carrier fluid, $\sigma$ the surface tension coefficient and $\varepsilon$ the dissipation rate associated with the turbulence intensity. \newnew{For emulsions, a Weber number $We_l$ using a characteristic length scale can be defined as \begin{equation} We_{l} = \frac{\rho_c {u^\prime}^2 l}{\sigma}, \label{eq:Wel} \end{equation taking into account the effect of surface tension. In experimental studies (e.g.~\citep{Perlekar:2012ip}), for example, the diameter of the stirrer is used as length scale. For the HIT considered here, the integral length scale of the turbulent flow field is employed. Therefore, emulsions at stationary state can be characterized by $We_{l}$ as demonstrated and discussed in \citet{Begemann:2022CJ}. Moreover, using $l = ({u^{\prime}}^2)^{3/2}/\varepsilon$, the correlation \begin{equation} d_H/l \propto We_{l}^{-3/5} \, \label{eq:dHWel} \end{equation}% is obtained. } The emulsions are generated in HIT with a constant turbulent kinetic energy $k$. For isotropic turbulence, $k$ is given by $k = (3/2) {u^{\prime}}^2$, where $u^\prime$ denotes the velocity fluctuation. In case of linear forcing, the dissipation rate $\varepsilon$ is determined by $k$ and the integral length scale $l$, which is $20\%$ of the domain length~\citep{Rosales:2005kp,Carroll:2013jt}, using the correlation $l = ({u^{\prime}}^2)^{3/2}/\varepsilon$. Additionally, for HIT, a characteristic dimensionless number is the Taylor Reynolds number $Re_\lambda = \lambda u^\prime/\nu$ formed with the Taylor micro-scale $\lambda = \sqrt{15\nu/\varepsilon}u^\prime$. We simulate five different configurations. \newmk{The parameters common for all configurations} are summarized in \cref{tab:Setup}. The volume fraction of the dispersed phase is $\phi = V_d/(V_d + V_c) = 1/8 $ for all cases. The carrier and dispersed fluid have a density of $\rho_c=1\,\si{kg/m^3}$ and $\rho_d=0.9\,\si{kg/m^3}$, respectively, thus making the dispersed fluid the lighter one. The kinematic viscosities are both set to $\nu_d=\nu_c=0.001\,\si{m^2/s}$. We here study emulsions generated at a turbulence intensity of $k=0.5\,\si{m^2/s^2}$ in a cubic domain with length $L = 2\,\pi$ and thus a Taylor Reynolds number of $Re_\lambda = 104$ and a dissipation rate of $\varepsilon=0.153\,\si{m^2/s^3}$. We discretize the domain with $N=384$ cells in each direction ($\approx \num{57e6}$ cells in total). This grid resolution has been chosen to fulfill the criterion $K_{max}\eta\geq 1.5$, see e.g.~\citet{Pope:2001turb}, where $K_{max}$ is the maximum wavenumber $K_{max}=N \pi /L$ and $\eta$ is the Kolmogorov scale $\eta=(\nu^3/\varepsilon)^{1/4}$, given by the kinematic viscosity $\nu$ and the dissipation rate $\varepsilon$. A grid study of this configuration can be found in our recent paper~\citep{Begemann:2022CJ}. For the simulation of the segregation process, first, turbulent emulsions in HIT are generated as described in detail in \citet{Begemann:2022CJ}. \new{We consider a cubic box with periodic boundary conditions.} \new{In order to generate turbulent emulsions,} we \new{first} perform single-phase simulations to obtain a fully developed single-phase HIT. Then we initialize the dispersed phase as spherical droplets, which break up in the linearly forced HIT and an emulsion is obtained as visualized in \cref{fig:sim_setup}~(a). At a statistically stationary state, we turn the forcing off and prescribe slip walls in the direction of the gravitational force, see \cref{fig:sim_setup}~(b). A thin layer (four cells) of the carrier fluid is additionally initialized at the bottom to detach dispersed structures there. Then, we let the emulsions segregate (\cref{fig:sim_setup}~(c)). The time $t$ is measured from the time instant when the forcing is turned off. \Cref{tab:CasesSeg} lists the considered configurations for the segregation. \newnew{The configurations are adopted from our previous investigation~\citep{Begemann:2022CJ} of emulsification processes and emulsions at statistically stationary state. The emulsions before segregation can be described by the dimensionless Weber number $We_l$.} Starting from a baseline case (\textit{Baseline}), the gravitational acceleration $g$ for the segregation process is varied. More precisely, $g$ is halved in case \textit{Low}~$g$ and doubled in case \textit{High}~$g$ with respect to the baseline case. Furthermore, also the surface tension coefficient $\sigma$ is varied for the cases \textit{Low}~$\sigma$ and \textit{High}~$\sigma$, resulting in different droplet size distributions at the statistically stationary state before the segregation. Note that for constant $\rho_c$ and $\varepsilon$, which is the case here, the correlation between $d_{H}$ and $\sigma$ reads $d_{H} = \sigma^{3/5}$, see also \cref{eq:Hinze}. For the variation of $\sigma$, the entire emulsification process has been simulated to obtain the respective emulsions at statistically stationary state. \newnew{The employed value for the surface tension $\sigma_\mathrm{BL} = \num{2e-2}\,\si{N/m}$ approximates that of realistic fluids~\citep{koegl2020characterization}. The gravitational acceleration $g_\mathrm{BL}=4.59\,\si{m/s^2}$ has been chosen to obtain a Bond number $Bo=\Delta \rho g R^2/\sigma$ of Hinze droplets comparable to realistic examples of liquid--liquid emulsions (estimated values $\Delta \rho = 100\,\si{kg/m^3}$, $g=9.81\,\si{m/s^2}$, $R=\num{1.6e-3}\,\si{m}$, $\sigma = \num{2e-2}\,\si{N/m}$). Additionally, it is worth noting that for gravity-driven rising/falling of a dispersed phase, the ratio of the density difference to the density of the disperse phase $\Delta \rho /\rho_d$ is decisive, and the ratio in our simulations corresponds to that of realistic liquid--liquid emulsions.} To monitor the segregation progress, we track the interface area $A$ and the center of mass of each phase $h_i$. The interface area of the dispersed phase is numerically approximated with the volume integral of the gradient of the VOF marker function $|\nabla\alpha|$. As reference for the interface area, the theoretically completely segregated state with $A_{\infty} =L^2$ is used. \begin{figure*}[] \centering \includegraphics[width=0.75\linewidth]{Fig2_reduced} \caption{Visualization of the segregation process for varying $g$. Rows correspond to different $g$ values with (a) \textit{Low}~$g$, (b) \textit{Baseline}, (c) \textit{High}~$g$, and columns to different time instants (i, ii, iii, iv) ($t=\{ 0,\,5,\,10,\,20\,\si{s}\}$). The images show the iso-surface of the volume fraction corresponding to $\alpha = 0.5$.} \label{fig:SegG} \end{figure*}% \begin{figure*}[] \centering \includegraphics[width=0.75\linewidth]{Fig3_reduced} \caption{Visualization of the segregation process for varying $\sigma$. Rows correspond to different $\sigma$ values with (a) \newnew{\textit{Low}~$\sigma$}, (b) \textit{Baseline}, (c) \newnew{\textit{High}~$\sigma$}, and columns to different time instants (i, ii, iii, iv) ($t=\{ 0,\,5,\,10,\,20\,\si{s}\}$). The images show the iso-surface of the volume fraction corresponding to $\alpha = 0.5$.} \label{fig:SegS} \end{figure*} \section{Results} \label{s:Results} We first study the segregation process for varying $g$ and $\sigma$ based on the obtained simulation results (\cref{ss:Results}). In \cref{ss:Mechanisms,ss:ERR}, we analyze the segregation process from the perspective of the energy releases driving it. At the end of this section (\cref{ss:u}), we derive correlations to estimate the characteristic velocity and timescale of the segregation process. \subsection{Effect of $g$ and $\sigma$} \label{ss:Results} \Cref{fig:SegG} visualizes the segregation for different gravitational accelerations $g$, while \cref{fig:SegS} shows the same process for emulsions with different surface tension coefficients $\sigma$. In \cref{fig:SegG} (variation of $g$), a difference in the segregation progress can already be seen at the second time step visualized (see \cref{fig:SegG}~(ii)). At a higher $g$, a larger fraction of the lighter phase has risen and fewer dispersed structures are visible in the lower part of the box. As time progresses, the faster segregation at higher $g$ becomes more evident. Finally, at the last time step shown (see \cref{fig:SegG}~(iv)), a clear difference for different $g$ can be seen. At the highest $g$ (see \cref{fig:SegG}~(c, iv)) only a few very small structures are visible in the lower part. Further, it should be noted that in all configurations (a--d), some structures of the heavier phase are enclosed at the upper boarder, resulting in the interfaces visible there. \Cref{fig:SegS} illustrates the segregation for different $\sigma$. Here, the distribution of the dispersed phase in the emulsion differs significantly. The higher the surface tension coefficient, the larger are the structures of dispersed fluid. For the case \textit{Low}~$\sigma$ (see \cref{fig:SegS}~(a)), there are many small structures with a smaller buoyancy force, which is proportional to $\Delta \rho V$, higher drag forces as well as more interactions in between dispersed structures. For these reasons, the segregation progress at lower $\sigma$ is significantly slower than for the cases with a higher $\sigma$. At the highest $\sigma$ (see \cref{fig:SegS}~(c)), comparably large structures are present and they experience a higher buoyancy force than smaller structures. Additionally, the high $\sigma$ promotes coalescence even more. For the \textit{High}~$\sigma$ case, a nearly complete segregation is reached at the last time step visualized (see \cref{fig:SegS}~(c,iv)). \begin{figure*}[] \centering \subfigure{\includegraphics[width=0.99\linewidth]{Fig4_PP_v2}} \caption{Visualization of the segregation process. \new{(a) Front view on the emulsion, (b) recorded evolution of the center of mass of the dispersed phase, (c) time series of the orange section highlighted in (a) at a frame rate of 2 frames/s.}} \label{fig:ts_seg} \end{figure*} \new{Moreover, the time series in \cref{fig:SegG,fig:SegS} also illustrate the different droplet shapes during the segregation process. As expected, small droplets have a quasi-spherical shape due to the dominance of the surface tension forces, while larger droplets are rather ellipsoidal. \Cref{fig:SegS} shows the effect of the surface tension coefficient on the droplet shapes at comparable size. Comparing the last time steps of the \textit{Low}~$\sigma$ and the \textit{Baseline} case (\cref{fig:SegS}~(a-b,iv)), it can be seen that droplets of comparable size are more ellipsoidal for the \textit{Low}~$\sigma$ case. Regarding the droplet shape, there is an interplay between the surface tension force, which aims at a spherical shape, gravity, which affects the buoyancy force and also the hydrodynamic pressure inside the droplet, and the turbulent flow field. The well-known Grace-Diagram~\citep{grace1976shapes} allows for estimations of drop and bubble shapes as a function of the E\"{o}tv\"{o}s number, also known as Bond number, and the bubble Reynolds number. The E\"{o}tv\"{o}s/Bond number and its significance will be discussed in more detail in \cref{ss:Mechanisms}. } The segregation process under gravity can be characterized by the height of each phase and the interface area. In process engineering, mostly the height of the lighter phase or a coalescence or creaming interface is used to quantify the segregation, see e.g.~\citet{aleem2020experimental}. These quantities are optically easily accessible and are therefore commonly used. Note that the height refers to the position in the direction of gravitational acceleration. \newmk{In \cref{fig:ts_seg}, we have tried to adopt the experimental procedure for characterizing segregation to our simulation results and have specially post-processed them for this purpose.} Further, for a more accurate evaluation, we have monitored the height of the center of mass of each phase during our simulations. As can be seen in \cref{fig:ts_seg}~(b,c), the monitored height of the center of mass matches the optical impression. In a bounded domain, the changes of height are directly coupled by the correlation $\Delta h_c = -\phi/(1-\phi)\Delta h_d$ (see also \cref{eq:h}), which can be seen in the visualization in \cref{fig:ts_seg}. In the following we only present the height of the center of mass of the dispersed phase, where the change is more evident. \FloatBarrier \pagebreak \begin{figure*}[!tb] \subfigure[]{\includegraphics[trim={2cm 10cm 13cm 1cm},clip,width=0.49\linewidth]{Fig5_PP}} \subfigure[]{\includegraphics[trim={14cm 10cm 1cm 1cm},clip,width=0.49\linewidth]{Fig5_PP}} \subfigure[]{\includegraphics[trim={2cm 1cm 13cm 10cm},clip,width=0.49\linewidth]{Fig5_PP}} \subfigure[]{\includegraphics[trim={14cm 1cm 1cm 10cm},clip,width=0.49\linewidth]{Fig5_PP}} \caption{Segregation progress measured by the height of the center of mass of the dispersed phase (a,b) and the interface area (c,d). The effect of different gravitational forces is shown in (a,c), whereas the effect of different surface tension coefficients is illustrated in (b,d).} \label{fig:HeightsComp} \end{figure*} \Cref{fig:HeightsComp} visualizes the segregation progress measured by the height (a,b) and the interface area (c,d). The left column~(a,c) depicts the data for different gravity accelerations and the right column~(b,d) illustrates the progress for different surface tension coefficients. As discussed above, a stronger gravitational force promotes segregation and leads to a faster change in the heights of the center of mass. The surface tension coefficient also alters the segregation progress, since a higher surface tension coefficient accelerates the segregation measured of the height of the lighter phase. As can be seen in \cref{fig:SegS}, larger droplets are present for higher surface tension coefficients and the coalescence process is faster, which enhances the rise of the lighter phase. Additionally, the interface area can also be considered to characterize the segregation progress. For the present configuration, the recorded data is shown in \cref{fig:HeightsComp}~(c,d). It has to be noted that due to the upper and lower bounds in our configuration, structures must coalesce at a certain point in time, and thus the final coalescence process is clearly \newmk{dominated} by the gravitational acceleration. \Cref{fig:HeightsComp}~(c) reveals that after about $t=7\,\si{s}$ the segregation measured by the interface area is predominantly governed by the gravitational acceleration. Before that ($t<7\,\si{s}$), a smaller $g$ leads to a somewhat faster decay of the interface area, see \cref{fig:HeightsComp}~(c). We explain this by the fact that at a lower $g$ the structures remain at the same height a little longer and thus have more time to coalesce. For the evolution of the relative interface area $A/A_{em}$ at varying $\sigma$ (\cref{fig:HeightsComp}~(d)), no clear trend can be observed. It should be noted that the interface area of the emulsion $A_{em}$ is significantly larger for smaller surface tension coefficients, which biases the representation. The time derivative $\partial A/\partial t$, or more precisely that of the surface tension energy $\sigma \partial A/\partial t$, is more suitable for a comparison and will be analyzed in \cref{ss:ERR}. Further, the strongly fluctuating evolution at the \textit{High}~$\sigma$ case is related to the smaller number of droplets present in this configuration. Overall, we conclude that for the present configuration, a clear correlation for the evolution of the interface can be expected only when there is no gravitational effect. The comparison of the evolution of the height and the interface area demonstrates that for the configurations studied, the height is more representative of the segregation progress and is therefore considered in the following. \subsection{Driving mechanisms and dimensionless segregation number} \label{ss:Mechanisms} Without energy input, emulsions are unstable and segregate over time as visualized in the previous section (\cref{ss:Results}). The segregation is driven by the energy release due to the minimization of the net potential energy and that of the surface energy. The change of potential energy of a two phase flow composed of a carrier and dispersed phase is given by \begin{equation} \Delta E_{pot} = (\rho_c V_c \Delta h_c +\rho_d V_d \Delta h_d) g\,. \label{eq:EPot} \end{equation} In a bounded domain, the change of heights is restricted to $V_c \Delta h_c + V_d\Delta h_d = 0$. Using this and $V_d= \phi V$ and $V_c= (1-\phi) V$, the \newmk{relation} \begin{equation} \Delta h_c = -\frac{\phi}{1-\phi}\Delta h_d \label{eq:h} \end{equation} is obtained. Thus, \cref{eq:EPot} can be simplified to \begin{equation} \Delta E_{pot} = -\Delta \rho g V_d \Delta h_d \label{eq:EPot2} \end{equation} with $\Delta \rho=\rho_c-\rho_d$, which is in the considered configuration positive since $\rho_c > \rho_d$. Consequently, the rise of the lighter phase (here dispersed phase) releases energy ($\Delta E_{pot}<0$). The change of surface energy is given by \begin{equation} \Delta E_{\sigma} = \sigma \Delta A. \label{eq:ES} \end{equation} Breakup leads to an increase of the interface area ($\Delta A>0$) and requires energy input, while coalescence leads to a reduction of the interface area ($\Delta A<0$), thus releasing energy ($\Delta E_{\sigma}<0$). To identify the dominant mechanism promoting the segregation, we propose a non-dimensional energy release ratio $\Psi$ of these two driving mechanisms \begin{equation} \Psi = \frac{\Delta E_{pot}}{\Delta E_{\sigma}}=\frac{ -\Delta \rho g V_d \Delta h_d} {\sigma \Delta A}. \label{eq:ERR1} \end{equation} The correlation above depends on several case specific quantities as $V_d$, $\Delta h_d$ and $\Delta A$, where especially the latter is often a-priori unknown. In the following we aim to derive a dimensionless correlation, which depends only on the fluid properties $\Delta \rho$ and $\sigma$, the gravitational acceleration $g$ and two characteristic length scales - $d$ for a representative droplet size and $H$ for a representative length scale of the segregation progress. To this end, we substitute the change of the interface area $\Delta A= A_\infty - A_{em}$, where $A_\infty$ is the interface area at fully segregated state and $A_{em}$ is the interface area of the emulsion at stationary state, by $\Delta A\approx -A_{em}$, which is valid for $A_\infty << A_{em}$. Further, $A_{em}$ can be expressed with the Sauter mean diameter $d_{32}=6 V_d /A_{em}$, for which constant relations to the Hinze scale have been reported $d_{32} \propto d_H$ \citep{pacek1998sauter,Yi:2020dn,boxall2012droplet}. This allows for the approximation \begin{equation} \Delta A \approx -6 V_d/d_{32}. \label{eq:DeltaA} \end{equation} Additionally substituting $\Delta h_d$ with a representative length scale $H$ gives \begin{equation} \Psi \approx \frac{1}{6}\frac{\Delta \rho g d_{32} H} {\sigma }. \label{eq:ERR2} \end{equation} Using this correlation, a dimensionless segregation number $Seg$ can be defined as \begin{equation} Seg =\frac{\Delta \rho g d H} {\sigma }, \label{eq:ERR2} \end{equation} where the representative length scale for the segregation process $H$ is the height the lighter (here dispersed) phase rises on average. In a bounded domain with length $L$, $H$ can be calculated using $H = (1 - 0.5\phi ) L - 0.5 L = 0.5(1-\phi) L$, which here is $H =7/8\pi$. $d$ stands for a representative diameter of the emulsion, which can be approximated with $d_{32}$ or $d_H$. It is worth noting that the dimensionless segregation number $Seg$ closely resembles the dimensionless Bond number $Bo$, also known as E\"otv\"os number, with \begin{equation} Bo=\frac{\Delta \rho g R^2} {\sigma }, \label{eq:Bo} \end{equation} for rising/falling bubbles or droplets with radius $R$. However, $Bo$ has a different physical relevance since it describes the ratio of body forces to surface forces and characterizes the bubble/droplet shape and the tendency for a breakup. \new{$R$ in \cref{eq:Bo} can be substituted by a different characteristic length scale of the bubble/droplet.} \newnew{Because of their different physical meanings, $Seg$ and $Bo$ use different length scales. The length scale for $Bo$ is associated with the bubble/droplet, while for $Seg$ the product of a length scale associated with the bubble/droplet (denoted here by $d$) and one associated with the segregation process (denoted here by $H$) is used. Note that a Bond number $Bo$ with a characteristic length scale of $\sqrt{d H}$, which is physically difficult to motivate, leads to the same expression as $Seg$.} \Cref{tab:CasesSeg} contains the segregation number $Seg$ determined with $d_H$ together with the dimensionless energy release ratio $\Psi$ evaluated using the changes between $t = 0\,\si{s}$ and $t = 25\,\si{s}$. A higher $Seg$ (or higher $\Psi$) indicates that the release of potential energy dominates for segregation, while a lower number implies a more important role of the surface tension energy release. It should be noted that the segregation number $Seg$ (or $\Psi$) only indicates the ratio of the two energy releases and does not provide any information about the timescale of the segregation process, which is analyzed in \cref{ss:u}. The ratio $Seg/\Psi$ is also included in \cref{tab:CasesSeg}. For a variation of $g$, the ratio is constant confirming the validity of the above made approximations. When $\sigma$ is varied, the ratio is approximately the same, but the values scatter. We conjecture that this is due to the approximation used for $\Delta A$ (\cref{eq:DeltaA}). \begin{figure*}[!tb] \subfigure[]{\includegraphics[trim={2cm 0cm 17cm 0cm},clip,width=0.49\linewidth]{Fig6_PP_v2}} \subfigure[]{\includegraphics[trim={17cm 0cm 2cm 0cm},clip,width=0.49\linewidth]{Fig6_PP_v2}} \caption{Energy release rates due to the reduction of the net potential height $\dot{E}_{pot}$ and the reduction of the interface area $\dot{E}_{\sigma}$ for varying $g$ in (a) and varying $\sigma$ in (b). } \label{fig:ERates} \end{figure*} \begin{figure*}[!tb] \subfigure[]{\includegraphics[trim={2cm 0cm 15cm 0.5cm},clip,width=0.49\linewidth]{Fig7_PP_v2}} \subfigure[]{\includegraphics[trim={15cm 0cm 2cm 0.5cm},clip,width=0.49\linewidth]{Fig7_PP_v2}} \caption{Decay of the normalized kinetic energy in different configurations. (a) for varying $g$, (b) for varying $\sigma$.} \label{fig:Ekin} \end{figure*} \subsection{Energy release rates and decay of turbulent kinetic energy} \label{ss:ERR} For the actual segregation, not only the amount of energy release is crucial, but also the time over which it is released. To this end, we propose to evaluate energy release rates. Using \cref{eq:EPot2,eq:ES}, the energy release rates are \begin{equation} \dot{E}_{pot} = \Delta \rho V_d \dot{h}_d g \quad \dot{E}_{\sigma} = \sigma \dot{A}. \label{eq:Rates} \end{equation} As discussed above, during the segregation progress both quantities are negative and therewith drive this process. \Cref{fig:ERates} visualizes the energy release rates. For the variation of the gravitational acceleration (\cref{fig:ERates}~(a)), the energy release for the potential energy $\dot{E}_{pot}$ clearly increases with increasing $g$, while the energy release due to the reduction of the surface energy $\dot{E}_{\sigma}$ is similar \newmk{for the three $g$-variations, see \cref{fig:ERates}~(a).} The difference between $\dot{E}_{pot}$ at different $g$ is not only caused by the different $g$ values, but also the resulting different $\dot{h}$, amplifying the differences in the energy release. The variation of the surface tension coefficient (see \cref{fig:ERates}~(b)) does not reveal such a clear trend. As expected, the energy release from the surface tension term $\dot{E}_{\sigma}$ increases with increasing $\sigma$, however, the value of $\sigma$ also affects the release of the potential energy $\dot{E}_{pot}$. At a smaller $\sigma$, the dispersed phase remains in a more disturbed state, i.e., smaller droplets, for a longer time. This leads to a greater number of droplet interactions and mutual hindering effects, limiting the rise of the lighter phase and thus the release of potential energy. Conversely, a higher $\sigma$ means fewer droplets as well as faster coalescence and, consequently, less restriction on the release of potential energy. Hence, the release of potential energy appears to scale proportional to the droplet size in the emulsions, see also \cref{fig:ERates}~(b). Further, the release of potential and surface energy affects the decay of the turbulent kinetic energy, which is depicted in \cref{fig:Ekin}. \Cref{fig:Ekin}~(a) shows that a higher $g$ results in higher energy release due to the decreasing potential energy, which is transformed in kinetic energy and retards the decay of the latter. The variation of $\sigma$ also affects the decay of turbulent kinetic energy (\cref{fig:Ekin}~(b)). However, due to the complex interplay of droplet size distributions and the release of gravitational energy, no clear trend is observed. For a detailed study of the effect of varying $\sigma$ on the decay of turbulent kinetic energy in decaying turbulence without gravitational force, we refer the reader to \citet{Dodd:2016eo}. \FloatBarrier \subsection{Time scale of the segregation process} \label{ss:u} For practical applications, the duration of the segregation process and the effects of parameter changes on this duration are of particular interest. Therefore, we attempt to derive a characteristic time scale for segregation progress. For this purpose, we consider the time evolution of the height of the lighter phase and its time derivative which represents an average rising velocity. \begin{figure*} \includegraphics[width=0.99\linewidth]{Fig8_all} \caption{Temporal derivative of the center of mass of the dispersed phase $\dot{h}_d$ (a,b), normalized temporal derivative $\dot{h}_d/\dot{h}_d^*$ (c,d) and temporal evolution of \newmk{$h/L$} scaled by the proposed timescale (e,f). The left column (a,c,e) shows these quantities for different gravitational accelerations $g$ and the right column (b,d,f) illustrates them for different surface tension coefficients $\sigma$. \newline\newline\newline\newline\newline\newline\newline\newline} \label{fig:Velocities} \end{figure*} \Cref{fig:Velocities}~(a,b) illustrates the temporal derivative of the height of the center of mass of the dispersed phase $\dot{h}_d$. As discussed in the previous subsection, in case of a variation of $g$, $\dot{h}_d$ clearly increases for an increasing $g$, see \cref{fig:Velocities}~(a). The droplet size distribution (associated with $\sigma$) alters the release rate of potential energy and thus $\dot{h}_d$. A more dispersed emulsion, characterized by a smaller $d_H$, has a higher hindering effect and limits the release of the potential energy, whereas a less dispersed emulsion, characterized by a higher $d_H$, allows for a higher energy release of potential energy. Consequently, $\dot{h}_d$ increases with increasing $\sigma$ (increasing $d_H$), see \cref{fig:Velocities}~(b). The average rising velocity of a single droplet due to gravitational acceleration, neglecting friction forces, is given by \begin{equation} U_\mathrm{g} = \frac{H}{t}= \sqrt{\frac{1}{2}\frac{\Delta \rho } {\rho_d}g H}, \label{eq:Ug_0} \end{equation} where $H$ stands for the height, which the dispersed phase has risen. The detailed derivation of this \newmk{relation} is provided in the appendix. This velocity, of course, significantly overestimates the average rising velocity of the considered configuration as friction forces and droplet--droplet interactions are neglected. However, the gravity-based velocity (\cref{eq:Ug_0}), together with the observations described above, motivate the formulation \begin{equation} \dot{h}_d \propto U_g \xi(d)\,, \label{eq:v} \end{equation} where $\dot{h}_d$ is assumed to be proportional to a gravity based velocity $U_g$ and a factor $\xi$ depending on the droplet size distribution of the emulsion. The proposed \newmk{relation} is a simplification for the configurations considered here. For other configurations, additional effects of other parameters, such as viscosity, would also have to be incorporated. We approximate $d$ with the Hinze scale $d_H$ and propose \begin{equation} \xi = (d_H/d_{ref})^\gamma\, . \label{eq:v2} \end{equation}% For the considered configuration, we have fitted $\gamma = 0.5$ and for simplicity taken $d_{ref} = d_{H\mathrm{BL}}$. This results in the correlation \begin{equation} \dot{h}_d^* = c_p U_g \xi( d) = c_p \sqrt{\frac{1}{2}\frac{\Delta \rho } {\rho_d}g H} \sqrt{\frac{d_H}{ d_{ref}}}, \label{eq:v3} \end{equation} where $c_p$ denotes a \newmk{proportionality} factor depending on the choice of $d_{ref}$. For $d_{ref} = d_{H\mathrm{BL}}$, $c_p$ is fitted to $c_p = 0.25$. The measured velocities $\dot{h}_d$ normalized by $\dot{h}_d^*$ (\cref{eq:v3}) are plotted in \cref{fig:Velocities}~(c,d). \newmk{In all cases, the normalized velocities have their maximum at about 1, which confirms the validity of the approximations made.} Further, the derived correlation allows for an estimation of a timescale for the segregation. Using \cref{eq:v,eq:v2} the following is obtained: \begin{equation} \tau_\mathrm{Seg} \propto \frac{H}{U_{g} (d_H/d_{ref})^{\gamma}} \,. \label{eq:t_g2} \end{equation} Using additionally the definition of $U_\mathrm{g}$ (\cref{eq:Ug_0}) results in \begin{equation} \tau_\mathrm{Seg} \propto \sqrt{\frac{\rho_d H}{\Delta \rho g}}\left(\frac{d_{ref}}{d_H}\right)^{\gamma}\,. \label{eq:t_g3} \end{equation} Thus, for a variation of only $g$, the timescale is proportional to $\tau \propto 1/\sqrt{g}$ and for a variation of only the droplet size distribution ($\sigma$), the timescale is proportional to $\tau \propto 1/{{d_H}^{\gamma}}$. For the considered configuration, the timescale of the segregation can be explicitly calculated as \begin{equation} \tau_\mathrm{Seg} = \frac{H}{c_p U_{g} (d_H/d_{ref})^{\gamma}}. \label{eq:t_g4} \end{equation} Using $\gamma = 0.5$, $c_p = 0.25$, and \cref{eq:Ug} gives \begin{equation} \tau_\mathrm{Seg} = 2^{5/2} \sqrt{\frac{\rho_d H}{\Delta \rho g}} \sqrt{\frac{d_{ref}}{d_H}}, \label{eq:t_g5} \end{equation} where $H = 7/8\pi$. \Cref{fig:Velocities}~(e,f) shows the height of the dispersed phase $h$ plotted over the time normalized by the characteristic timescale. It can be seen that for both variations, the segregation measured by height is completed at $\tau_\mathrm{Seg}$. Moreover, it is noteworthy that the scaled temporal evolution of the heights almost coincides to one line when $g$ is varied, see \Cref{fig:Velocities}~(e). \section{Conclusions} \label{s:Conclusions} In this work, we \new{have} numerically \new{studied} the segregation of turbulent emulsions under different gravitational accelerations $g$ and with different droplet size distributions obtained by altering the surface tension coefficient $\sigma$. To this end, we first generated turbulent emulsions in a linearly forced HIT and then turned off the forcing and activated the gravitational acceleration. \new{This approach enabled us to study the segregation process using well-defined initial conditions.} \new{To our knowledge, this work represents the first numerical investigation of the gravity-driven segregation process. With it, we extend previous numerical studies focusing on emulsification or emulsions at the statistically stationary state. Moreover, the time-resolved, three-dimensional visualization of the segregation progress obtained by our DNS studies supplements existing experimental studies on segregation.} We have approached this topic from the thermodynamic perspective of energy releases, adding an important complementary perspective to this physical process. Segregation can be quantified by the height of each phase and the interface area. In the present study, we have primarily analyzed the temporal evolution of the height of the dispersed phase. Moreover, we have addressed the energy release of the two central processes, namely the rise of the lighter phase (release of potential energy) and coalescence (release of surface energy). Based on our observations, we have defined a dimensionless segregation number $Seg$ that characterizes the ratio of potential energy release to surface energy release, allowing for an identification of the dominating process. In addition, we evaluated and compared the energy release rates. Our simulation results show that a smaller droplet size, i.e., smaller $\sigma$, hinders and limits the release of potential energy. Finally, we have derived a correlation to estimate the average rising velocity of the lighter phase, which also allows for the derivation of a characteristic timescale. We found that the average rising velocity is a fraction of a gravity based velocity and depends on the size of the droplets in the emulsion. Scaling of the velocities and the time with the empirically derived correlation showed good agreement. The presented work can be considered as a first important step towards the numerical assessment of emulsion segregation. Subject of current investigations is the evaluation of droplet size distributions during the segregation process. Therefore, in order to obtain statistically reliable data, a multitude of identical segregation processes have to be simulated simultaneously. Furthermore, in future studies we plan to consider configurations with varying density differences between the dispersed and carrier phase to assess the effect of the density difference on the segregation and its time scale. \section*{Declaration of Competing Interest} The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. \section*{Acknowledgment} This project received funding by \textit{dtec.bw} - Digitalization and Technology Research Center of the Bundeswehr - under the project \textit{MORE}, which is gratefully acknowledged. Further, the authors thank the Gauss Centre for Supercomputing e.V. (www.gauss-centre.eu) for funding this project by providing computing time on the GCS Supercomputer SuperMUC-NG at Leibniz Supercomputing Centre (www.lrz.de). \section*{Data Availability} The data that support the findings of this study are available from the corresponding author upon reasonable request. \FloatBarrier \section*{Appendix} \subsection*{Gravity based velocity derived from the force balance on a droplet} \label{ss:ug} A gravity-based characteristic velocity $U_\mathrm{g}$ can be derived based on the force balance on a droplet with mass $m$ ($m=\rho_d V_{droplet}$). Neglecting friction forces and other losses the force balance reads \begin{equation} m_d \ddot{x}= F_b-F_g, \end{equation} where $F_b$ denotes the buoyancy force with $F_b=\rho_c V_{droplet} g$ and $F_g$ denotes the gravitation force with $F_g=\rho_d V_{droplet} g$. This leads to the following acceleration: \begin{equation} \ddot{x}= \frac{\Delta \rho}{\rho_d} g. \label{eq:xDotDot} \end{equation} Integrating twice in time and using the initial conditions $\dot{x}=0,\,x=0$ gives \begin{equation} x = \frac{\Delta \rho}{\rho_d} g \frac{1}{2}t^2. \label{eq:x} \end{equation} The distance over a time $t$ is set to $x=H$ and $t$ can be expressed as \begin{equation} t = \sqrt{2 \frac{\rho_d}{\Delta \rho}\frac{H}{g}}. \label{eq:t} \end{equation} Thus, the average velocity resulting from gravitation over a distance $H$ can be determined with \begin{equation} U_\mathrm{g} = \frac{H}{t}= \sqrt{\frac{1}{2}\frac{\Delta \rho } {\rho_d}g H}. \label{eq:Ug} \end{equation} This \newmk{relation} can alternatively be derived by evaluating the velocity based on the kinetic energy equivalent of the release of potential energy, as shown below. \subsection*{Gravity based velocity derived from the energy release of the potential energy} The gravity based characteristic velocity (\cref{eq:Ug}) can also be derived by evaluating the velocity based on the kinetic energy equivalent of the release of potential energy. For consistency, we also consider the kinetic energy in the carrier phase. This gives the following balance \begin{equation} \frac{1}{2} \rho_d V_d U_{g,max}^2+\frac{1}{2} \rho_c V_c U_{c,g,max}^2 = - \Delta E_{pot}, \label{eq:Ebil1} \end{equation} where $U_{g,max}$ refers to the maximum velocity in the dispersed phase to be consistent with the nomenclature used so far and $U_{c,g,max}$ to that in the carrier phase. Using \cref{eq:h}, we can recast the left part to \begin{equation} \frac{1}{2}\rho_d V_d U_{g,max}^2(1+\frac{\rho_c}{\rho_d}\frac{\phi}{1-\phi}) = - \Delta E_{pot}. \label{eq:Ebil2} \end{equation} We simplify this expression with $\beta = \frac{\rho_c}{\rho_d}\frac{\phi}{1-\phi}$. Note that for small void fractions and density ratios close to 1 this expression vanishes ($\beta\to 0$). Inserting the expression for the release of potential energy from \Cref{eq:EPot2} yields \begin{equation} \frac{1}{2}\rho_d V_d U_{g,max}^2(1+\beta) = \Delta \rho V_d g \Delta h_d. \label{eq:Ebil3} \end{equation} Based on this correlation, we can derive a maximum velocity associated with the energy release of the potential energy that reads \begin{equation} U_{g,max} = \sqrt{\frac{2}{1+\beta} \frac{\Delta \rho }{\rho_d} g \Delta h_d}. \label{eq:Udmaxg} \end{equation} Assuming that the initial velocity is zero and the acceleration is linear, the average velocity can be approximated by $U_\mathrm{g} \approx \frac{1}{2} U_{g,max}$. Further, we substitute $\Delta h_d$ with $H$ and obtain \begin{equation} U_{g} = \sqrt{\frac{1}{2(1+\beta)} \frac{\Delta \rho }{\rho_d} g H}. \label{eq:Ug2} \end{equation} For $\beta=0$ this expression is identical to \cref{eq:Ug}. In the considered configurations we have $\beta=0.1587$ which corresponds to a prefactor of $\sqrt{1/(1+\beta)}=0.9289$ for $U_{g}$ and has been neglected in the evaluations shown in the paper for the sake of simplicity. \FloatBarrier \section*{References}

de02922af2b2a55c3c94e93d4a9da10353a6454b

\section{Preliminary} \section{Introduction} \IEEEPARstart{R}{ecent} advances within the Smart Grid (SG) paradigm are geared towards the incorporation of several Internet of Things (IoT) based devices and advanced computing technologies to ensure reliability, flexibility and efficiency of critical power systems \cite{Lamnatou_Chemisana_Cristofari_2022}. With the prevalence of Artificial Intelligence (AI), the enormous amount of highly granular power-related data generated by such intelligent devices enable energy service providers to improve load forecasts, maximize financial gains, devise effective demand management and other grid operation strategies, etc \cite{Sakhnini_Karimipour_Dehghantanha_Parizi_Srivastava_2021}. Besides, consumers can experience better quality of service through personalization of the power system applications and tools \cite{9478223}. However, present data analytics solutions for SGs primarily emphasize on centralized and decentralized approaches that require the direct sharing of data and/or learning models to dedicated central servers \cite{9381850} . In such cases, the sharing of fine-grained load consumption profiles collected from individual smart meters to central data servers imposes several privacy concerns to energy data owners \cite{husnoo2021false, reda2021taxonomy}. For instance, several studies \cite{10.1145/1878431.1878446, 10.1145/2528282.2528295} have highlighted that simple analysis of load consumption patterns recorded by smart meters can reveal household occupancy rates, the presence of people within a house, and sleep/wake-up time of residents, without any prior knowledge. Indeed, higher resolution of smart meter data leads to higher granularity in information and allows third parties to infer more sensitive information about households. In such a scenario, Federated Learning (FL) emerges as a viable privacy-preserving distributed computing alternative which transfers computation to energy data owners and allows the training of a global model through collaboration of devices without requiring the migration of data to a central repository for model training \cite{9084352}. Typically, edge devices in an energy system network iteratively train a local model and update the resulting parameters to a central aggregator which accumulates and processes the parameters and then sends back the updated parameters to the edge devices. The communication rounds continue until successful convergence of the model. In spite of the privacy preservation benefits due to the omission of raw data sharing requirements, FL is also efficient in terms of communication resource usage and has higher scalability \cite{9084352}. Recently, FL has gained much attention from researchers to explore its potential benefits within several smart grid domains, namely short-term load forecasting \cite{husnoo2022fedrep, 9148937}, energy theft detection \cite{9531953}, to name a few. Nevertheless, despite its promising privacy-preserving potentials, recent literature has revealed that FL may fail to provide sufficient privacy guarantees in certain circumstances. For example, researchers have discovered that they are able to reconstruct the original raw data from the sharing of gradients of the model during iterations \cite{iDLG}. Furthermore, due to the distributed nature of FL, it is vulnerable to Byzantine faults/attacks whereby the client nodes behave arbitrarily which may be a result of adversarial manipulations or software/hardware faults \cite{FLTrust}. Therefore, it is imperative to design FL mechanisms that are fault-tolerant to such behaviours, provide good generalisation performance and are communication efficient. Consequently, we investigate this research gap in the field of smart grids by contributing to the following: \begin{enumerate}[leftmargin=*] \item Inspired by the idea of gradient quantization, we develop a state-of-the-art privacy-preserving federated learning-based framework that leverages the SIGNSGD algorithm to improve the robustness of FL strategies for residential short-term load forecasting against Byzantine attacks. \item Specifically, in this paper, we highlighted three key data integrity attacks against short term load forecasting FL models. We design the data integrity threat models and their counter measures. \item We further extend the proposed framework towards a privacy-preserving SIGNSGD-based FL approach whereby the clients locally perturb their trained parameters by adding noise prior to uploading to the server for aggregation to prevent parameter information leakage and ensure privacy preservation more effectively. \item We conduct comprehensive case studies and extensive empirical evaluations to verify the effectiveness of our proposed scheme using a real Australian energy consumption dataset obtained from Ausgrid Network. \end{enumerate} \noindent The rest of this paper is structured as follows. Section \ref{sect:prelim} provides some background information in relation to our conceptual framework. Section \ref{sect:probdef} covers the problem definition section where we discuss some popular adversarial byzantine threat models on FL. In Section \ref{propmethod}, we describe our proposed FL architecture followed by Section \ref{Results} which focusses on the evaluation and comparison of our proposed framework under several scenarios. Finally, Section \ref{Conclusion} concludes this manuscript and provides some potential future directions for research. \section{Related Works} Typically, byzantine threats on federated learning scenarios consist of updating arbitrary model parameters from the clients to the server in the aim of impacting the convergence of the model \cite{BarossoFedThreatSurvey}. More specifically, byzantine attacks are typically untargeted threats during which adversarial clients either train their local models on corrupted datasets or fabricate random model updates. Inherently, byzantine threats are usually less stealthy and can be detected through close analysis of the global model performance \cite{9220780}. To address byzantine resiliency in FL, a number of works have been proposed in recent literature. Throughout this section, we briefly summarize the main studies undertaken in regards to byzantine resiliency in FL. A common approach to byzantine-resiliency in FL is to employ aggregation operators which are based on statistically robust estimators. For instance, the authors in \cite{FLTrust, 10.1145/3154503, 9029245} leveraged the use of Byzantine-robust aggregation rules by comparing the local updates of clients and filtering out statistical outliers prior to global model updates. Furthermore, Blanchard et al. \cite{10.5555/3294771.3294783} proposed a computationally expensive \textit{Krum} algorithm which performs gradient update selection and has the least sum of distances from the nearest gradient updates during each iteration. In addition, \cite{pmlr} introduced \textit{Bulyan} as an extension of Krum to recursively find subset of nodes using Krum and eventually perform an element-wise pruned mean on the updates to exclude the high magnitude values. Similarly, a handful of other byzantine-robust aggregation operators \cite{distChen, RSALi, pmlr_v80_yin18a, 9153949,PillutlaAggregation, GonzalezByzantine, ShuhaoResidual} have been proposed in existing literature to mitigate the vulnerability of FL to byzantine attacks. Another interesting study in \cite{9669031} utilized a mixed-strategy game-theoretic approach between the server and the clients whereby each client can either update good or corrupted model parameters while the server can either choose to accept or discard them. By employing the Nash Equilibrium property, the clients' updates were selected based on their probability of providing the correct updates. In addition to the design of byzantine-robust operators, several other defence strategies have been employed through anomaly detection \cite{ShiqiDefending, 9054676, 8975792}, pre-processing methods \cite{https://doi.org/10.48550/arxiv.2004.04986}, etc. However, while much work has been carried out to mitigate the threats of FL, little to no work has been carried out on secure, privacy-preserving and fault-tolerant FL frameworks for residential short-term electrical load forecasting to the best of our knowledge. \section{Preliminary} \label{sect:prelim} Throughout this section, we will discuss some preliminary and related background knowledge on FL and Differential Privacy (DP). Furthermore, within this section, we shall discuss a conventional FL set-up for short term load forecasting which will be used as a baseline during the evaluation of our proposed scheme. \subsection{Federated Learning} For the past couple of decades, Artificial Intelligence (AI) has transformed every walk of life and proven its benefits within several fields. However, one of the biggest real-world challenge faced by AI is the design of high-performing models due to natural data fragmentation coupled with security and privacy enforcement. Eventually, to alleviate this issue, \cite{Fedpap} introduced a fundamentally novel learning technique known as \textit{Federated Learning} which enables the collaborative decentralised training of machine learning models without the physical migration of raw data as depicted in Fig. \ref{fig:fedillus}. \begin{figure}[!h] \centering \includegraphics[width=7cm]{Federated_Learning_illustration.png} \caption{An illustration of the steps involved in FL.} \label{fig:fedillus} \end{figure} Suppose we have $N$ clients and each client $C_i$ holds a local training dataset $D_i$ where $i \in {1,2,...,N}$. An active $C_i$, participating in in the local training, aims to collaboratively learn the weights $w_i$ of the shared global model such that a certain empirical loss $L_i$ is minimized. Therefore, we can formulate the optimization problem solved by multiple data owners as $w^* = \underset{w_i}{\mathrm{arg\ min}} \displaystyle \sum_{i=1}^N L_{i} (w_{i})$. Specifically, each communication round proceeds as shown in Fig. \ref{fig:fedillus} through the following steps: (1) The central server sends a unanimous global model $w$ to the active FL clients. (2) Each client trains the local model by using its own local dataset $D_i$ in order to solve the optimization problem $\underset{w_i}{\mathrm{min}}\ L_i(D_i, w_i)$. (3) Each client updates its local model parameters to the central server. (4) The server computes the global model update by aggregating the parameters received from the local models such that. (5) Lastly, the server sends back the updated parameters to the local models. This iterative process is continued until convergence of the global model. Furthermore, there are two baseline approaches to train models in a FL set-up namely Federated Averaging (Fed-Avg) and Federated Stochastic Gradient Descent (Fed-SGD). Generally, Fed-SGD \cite{pmlr-v119-malinovskiy20a} averages the locally computed gradient at every step of the learning phase while Fed-Avg \cite{9488877} averages local model updates when all the clients have completed training their models. However, as mentioned before, regardless of the approach used, FL is prone to several privacy and security threats, which have been discussed as following. \subsection{Differential Privacy} \label{sect:diffpriv} Due to the several drawbacks of data anonymization techniques such as loss of data utility, risks of re-identification, etc., Differential Privacy (DP) emerged as a formal framework that enables the quantification of the preservation of individual privacy within a statistical database during the release of useful aggregate information \cite{9594795}. Therefore, we formally define some related concepts in relation to DP as in the following: \noindent \textbf{Definition 1}: A randomized algorithmic mechanism $M: X \longrightarrow R$ with domain $X$ and range $R$ satisfies ($\epsilon$, $\delta$)-differential privacy if for all measurable sets $S \subseteq R$ and if for any two adjacent inputs $D$, $D' \in X$, the following holds: $Pr[M(D) \in S] \leq exp(\epsilon) \times Pr[M(D') \in S] + \delta$. Here $Pr$ denotes probability \cite{9594795}. \noindent \textbf{Definition 2}: The privacy loss $L$ of a randomized algorithmic mechanism $M: X \longrightarrow R$ for any result $v \in R$ and for any two data samples $D$, $D' \in X$ is expressed as: $L(v, D, D') = log \dfrac{Pr[M(D) = v]}{Pr[M(D') = v]}$ \cite{9594795}. One of the most popular noise addition mechanisms for DP is the Gaussian Mechanism. A given noise distribution $n \sim N(0, \sigma^2)$ preserves ($\epsilon$,$\delta$)-DP where $N$ is a Gaussian distribution with 0 mean and variance $\sigma$, such that the noise scale is $\sigma \geq c\Delta s/\epsilon$ and the constant $c \geq \sqrt{2ln(1.25/\delta)}$ for $\epsilon \in (0,1)$ where $\Delta s$ is the sensitivity of the real-valued function. However, it is important to note that choosing the right amount of noise is a significant challenge that still lingers within research. \subsection{Federated Load Forecasting with Fed-SGD (Benchmark)} During Fed-SGD, a distributed stochastic gradient descent algorithm is applied within a federated environment to jointly train the global model. As shown in Algorithm \ref{FedSGDalgo}, during each communication round, each client $k$ computes the gradient $g_k$ by initially optimizing the loss of the local model using their local dataset $D_{k}$. The local gradients $g_k$ is then sent to the control centre whereby they are aggregated and the new gradient updates are pushed back to the local models. Eventually, the whole process is continued until convergence. \begin{algorithm} \textbf{Input}: learning rate $\eta$, each client $k$, local data $D_{k}$. Control centre initializes and distributes unanimous model $m_0$ and encrypted parameter initialization $||\hat{m}_0||$ to all clients $N$. \For{each communication round $T_{cl} = 1,2,..., t$}{ \For{each client $k \in N$}{ Compute gradient $g_k$ by training model on local dataset $D_k$. Send to Control Centre. \textbf{end} } Control Centre aggregates the local gradient updates as $g$. Control centre pushes updated gradients back to the local models. \textbf{end} } \caption{Short-term Load Forecasting with Fed-SGD.} \label{FedSGDalgo} \end{algorithm} \section{Problem Definition \& Adversarial Models} \label{sect:probdef} Federated learning enables promising privacy-preserving data analytics for smart grids by pushing model training to devices, thus requiring no direct data sharing \cite{9084352}. Nonetheless, recent literature has revealed its failure to sufficiently guarantee privacy preservation due to update leakage \cite{bhowmick2019protection}, deep leakage\cite{geng2022general}, byzantine attacks \cite{247652}, etc. Throughout this paper, we aim to address byzantine threats in relation to federated learning for electrical load forecasting. Before we present our proposed defense strategy, in this section, we consider three types of byzantine threat models on federated load forecasting as in the following: \begin{enumerate}[leftmargin=*] \item \textbf{Threat Model 1} \textit{(Local Data Poisoning)}: We consider the scenario where a subset of the total number clients $k$ to be malicious or are controlled by a malicious attacker. Malicious clients may be injected to the federated learning framework through the addition of adversarially-controlled smart metering devices. The goal of the adversary is to manipulate the learnt parameters such that the global model $M$ has high indiscriminate errors, thus implying that the attack objective is: $Attack(D_{k} \cup D_{k}', m_{t}^k) = max \displaystyle \sum_{i=1}1[f(x_{i}'; m_{t}^k ) = t_{i}'] $, where $m_{t}^k$ represents the updated model. Each malicious client is able to stealthily alter their local training sample, $D_k$ but is unable to access and manipulate the data of other participants or the model learning process. Let $D_k = \{(x_i, t_i)|i = 1,...,n\}$ denote the pristine local training dataset with $n$ samples where $x_i$ is the time instance and $y_i$ is the corresponding electrical load consumed. Each malicious client $k$ modifies their dataset $D_k$ such that the trigger $v$ is inserted into $x_i$ whereby $x_{i}' = x_i + v, t_i$. The sign $+$ denotes the addition of the poison trigger $v$ to $x$ such that the poisoned dataset $D_{k}' = \{(x_{i}', t_{i}'|i = 1,...,n\}$. The poisoned dataset $D_{k}'$ is then used for model training. The adversary's goal is to ensure the degradation of forecasts of the auxiliary data by the global model. \item \textbf{Threat Model 2} \textit{(Model Leakage \& Poisoning)}: In this scenario, we assume that the adversary can arbitrarily manipulate the local models sent from the clients to the central aggregator for illicit purposes but cannot observe the training data of other honest clients. Similarly, during this type of threat, the ultimate adversarial goal is to manipulate the learnt global model such that it has a high error rate indiscriminately for testing examples. Such attacks directly negatively impact the usability of the model and will eventually lead to denial-of-service attacks. \item \textbf{Threat Model 3} \textit{(Colluding attack)}: Lastly, we consider the cross-device scenario whereby multiple malicious clients are present during the federated training iteration. The adversaries intentionally collude with each other during a single iteration by sending the same update. i.e., each of the attackers send the same learnt update during some of the training iterations such that the goal of this threat model relies upon the manipulation of the learnt global model to induce high error rates. \end{enumerate} \section{Proposed Method} \label{propmethod} Within this section, we propose a new FL framework to circumvent the aforementioned byzantine threats on FL for short-term load forecasting. The key idea lies in sharing just the sign of the gradients to preserve privacy. We present the our developed solution as in the following: \subsection{System Model Overview} \begin{figure} \centering \includegraphics[width=8cm]{DP-Fed.png} \caption{An illustration of proposed approach.} \label{fig:proposedapp} \end{figure} As previously discussed, the objective of this study is to design a robust and privacy-preserving FL framework for residential short-term load forecasting. As shown in Fig \ref{fig:proposedapp}, our proposed method consists of three components as discussed. \begin{enumerate}[leftmargin=*] \item \textit{Electrical Appliances}: Whenever someone within a household uses one of the electrical appliances, the load consumption is collected by the smart meter \item \textit{Smart Meter}: Each customer has a smart meter that is connected a Home Area Network. Each smart meter collects energy load consumption profiles. The data collected is locally stored on the HAN of the consumer such that local models can be trained using their own dataset. \item \textit{Control Centre}: The control centre is responsible for broadcasting a learning model and default model parameters, aggregation of parameters after training and finally broadcasting the updated model parameters. \end{enumerate} \subsection{Algorithm Design} Within a conventional federated learning setting with $N$ clients, at round $t$, a selected client $k \in N$ performs local gradient descent iterations $T_{gd}$ using a common broadcasted local model $m_{t-1}$ on its local training sample $D_{k}$ such that a new updated model $m_{t}^k$ is obtained. Each client $k$ then sends its updated parameters $\Delta m_{t}^k = m_{t}^k - m_{t-1}^k$ to the central orchestrator which in turn aggregates model updates from all $N$ clients $\forall k \in N$ such that $m_{t} = m_{t-1} + \sum_{k \in N} \dfrac{|D_{k}|}{\sum_{j} |D_{j}|} \Delta m_{t}^k$. The model training continues until convergence and is subsequently terminated after a set number of rounds $T_{cl}$. \begin{algorithm} \textbf{Input}: learning rate $\eta$, each client $k$ local data $D_{k}$. Control centre distributes unanimous model $m_{0}$ and encrypted parameter initialization $||\hat{m_{0}}||$ to all clients $N$. \For{each communication round, $T_{cl} =1,..., t$}{ \For{each client $k$}{ Compute the gradient $g_{k}= m_{t}^k - m_{t-1}^k$ by training on local dataset $D_k$. Obtain sign vector $sign(\Delta m_{t}^k)$ from $g_k.$ Perturb $sign(\Delta m_{t}^k)$ with a random Gaussian noise $\zeta_{k}$ such that $\sum_{k \in N} sign(\Delta m_{t}^k) + \zeta_{k}$ satisfies differential privacy. Encrypt $sign(\Delta m_{t}^k) + \zeta_{k}$ into $E_{k}[sign(\Delta m_{t}^k) + \zeta_{k}]$ and send to control centre. \textbf{end} } Control Centre aggregates encrypted updates $\sum_{k} E_{N_{k}} (sign(\Delta m_{t}^k) + \zeta_{k})$. Control Centre pushes $sign(g_N)$ to all clients, $N$. \textbf{end} } \caption{Proposed Framework} \label{proposedalgo} \end{algorithm} However, in the context of smart grids, conventional federated learning settings pose several privacy risks as earlier discussed. Therefore, we propose a novel privacy-preserving federated learning framework for electrical load forecasting through model weight quantization as in \cite{jin2021stochasticsign}. Specifically, as shown in Algorithm \ref{proposedalgo} and Figure \ref{fig:proposedapp}, a selected client $k$ initially computes the gradient update $g_{k} = m_{t}^k - m_{t-1}^k$ from which it obtains the sign vector $sign(\Delta m_{t}^k) = sign(m_{t}^k - m_{t-1}^k)$ where $sign(\Delta m_{t}^k)$: $\mathbb{R}^n \longrightarrow {-1,1}^n$. A random Gaussian noise $\zeta_{k}$ is then added to perturb $sign(\Delta m_{t}^k)$ such that $\sum_{k \in N} sign(\Delta m_{t}^k) + \zeta_{k}$ satisfies differential privacy. Furthermore, to prevent an adversary from learning $sign(\Delta m_{t}^k) + \zeta_{k}$ accurately in circumstances where $N$ is large, each client $k$ updates the encrypted results $E_{k}[sign(\Delta m_{t}^k) + \zeta_{k}]$ to the central aggregator. The orchestrator in turn sums all the encrypted model updates from $N$ such that $\sum_{k} E_{N_{k}} (sign(\Delta m_{t}^k) + \zeta_{k})$. This aggregation follows the selection of the median of all $N$ clients signs at every position of the update vector. The model training continues until convergence and is subsequently terminated after a set number of rounds $T_{cl}$. \subsection{Convergence Analysis} In the following, we will present a formal analysis of the SIGNSGN approach through the use of refined assumptions derived from conventional SGD assumptions. \noindent\textbf{Assumption 1} \textit{(Lower Bound)}: Given an objective/loss function $f$, at any point $x$, $f(x) \geq f^(x^*)$, where $f^(x^*)$ represents the objective value and $x^*$ represents the global minima of f(x). This standard assumption is indeed necessary to ensure the convergence to a stationary point. \noindent\textbf{Assumption 2} \textit{(Smoothness)}: Given an objective/loss function $f$, the gradient of $f$ (derivative of the function with respect to $x$) when evaluated on any coordinate $(x, y)$ can be represented as $g(x)$. Then, for $\forall x, y$ and for some non-negative constant $L_{i}$, we require that $|f(y) - [f(x) + g(x)^T (y-x)]| \leq \frac{1}{2} \sum_{i}L_{i}(y_{i} - x_{i})^2$. This assumption is an extension of the Lipschitz Continuity condition which is essential to guarantee that the loss $l$ of $f$ is smooth and convergence of gradient descent algorithms. \noindent\textbf{Assumption 3} \textit{(Variance Bound)}: Upon receiving the query $ x \in \mathbb{R}^n$, the stochastic gradient oracle results in an independent, unbiased estimate $\hat{g}$ that has bounded variance per coordinate $\mathbb{E}[\hat{g}(x)] = g(x)$, $\mathbb{E}[(\hat{g}(x)_{i} - g(x)_{i})^2 \leq \sigma{i}^2 $ where $\sigma{i}^2$ is the uniform variance bound. The classical convergence analysis of SGD is carried out under the assumption that the norm of the stochastic gradient is uniformly bounded. While this might hold for some loss functions, bounded variance may be violated where $f$ is strongly convex as $x \longrightarrow \infty$. However, this assumption is necessary to grasp the fundamental properties of stochastic optimisation algorithms. \noindent\textbf{Assumption 4} \textit{(Gradient Noise)}: At any given point $x$, each component of the stochastic gradient vector, $\hat{g}(x)$, must have a unimodal distribution that is also symmetric about the mean. This assumption ensures that the addition of extra noise for the purpose of differential privacy does not skew the distribution and decrease utility. Under these assumptions, we have the following result: \noindent \textbf{Theorem 1} \textit{(Non-convex convergence rate of SIGNSGD)}: Run algorithm 1 for $K$ iterations under Assumptions 1 to 3. Set the learning rate as $\delta_k = \dfrac{1}{\sqrt{||L||_1 K}}$ where $n_k = K$. Let $N$ be the cumulative number of stochastic gradient calls up to step $K$, i.e. $N = O(K^2)$. Then we have $\mathbb{E}[\dfrac{1}{K} \displaystyle \sum_{k = 0}^{K-1}||g_k||_1 ]^2 \leq \dfrac{1}{\sqrt{N}}[\sqrt{||L||_1 } (f_0 - f_* \dfrac{1}{2}) + 2||\sigma||_1]$. \section{Simulation \& Results} \label{Results} In this section, we provide the results of the experimental evaluations of our proposed approach. We first introduce the dataset used and the settings shared by all experiments. Next, the performance of the proposed approach is presented and compared throughout different scenarios. Lastly, we discuss the overall results. \subsection{Experimental Setup} This research was conducted using \textit{Solar Home Electricity Data} from Eastern Australia's largest electricity distributor, Ausgrid. The dataset composes of half-hourly electricity consumption data of 300 de-identified customers which is measured using gross meters during the period starting 1\textsuperscript{st} July 2012 to 30\textsuperscript{th} June 2013. We initially filter the data based on General Consumption (GC) category. It is then converted to the suitable time-series format. It is then split into test (30\%) and train (70\%) subsets. Every experiment carried out have the following general configurations. There is a set number of clients (10 clients) each holds a local subset of the data and there is a server which helps to coordinate the FL scenario. The model performance is evaluated using three metrics: \textit{Mean Squared Error (MSE)}, \textit{Root Mean Squared Error (RMSE)} and lastly, \textit{Mean Absolute Percentage Error (MAPE)}. \subsection{Comparison with Baseline (No Attack)} Throughout this section, we present the experimental results to compare the performance of the proposed approach against the conventional Fed-SGD approach. As shown in Fig. \ref{fig:trainloss}(a), it can be seen that the Fed-SGD reaches convergence after the 47\textsuperscript{th} communication round while the proposed approach converges after the 40\textsuperscript{th} communication round. \begin{figure}[!h] \centering \subfloat[\centering Convergence of Federated LSTM-CNN model ]{{\includegraphics[width=4cm]{FedSGDvsProposed.png} }}% \qquad \subfloat[\centering MAPE (\%) per client ]{{\includegraphics[width=3.9cm]{ComparisonMAPEprop.png} }}% \caption{Comparison between Fed-SGD and proposed approach.}% \label{fig:trainloss} \end{figure} \begin{table}[!h] \caption{Evaluation of Fed-SGD with several models \label{CompaGedSGD}} \begin{tabular}{|l|l|l|l|l|l|} \hline \textbf{Metric} & \textbf{RNN} & \textbf{GRU} & \textbf{LSTM} & \textbf{CNN} & \textbf{LSTM-CNN} \\ \hline \textbf{MSE} & 0.2657 & 0.1973 & 0.1634 & 0.2567 & 0.1583 \\ \hline \textbf{RMSE} & 0.5346 & 0.4042 & 0.3463 & 0.5243 & 0.3008 \\ \hline \textbf{MAPE (\%)} & 16.4 & 10.9 & 11.0 & 12.8 & 9.7 \\ \hline \end{tabular} \end{table} \begin{table}[!h] \caption{Evaluation of proposed method with several models \label{CompaSignSGDModel}} \begin{tabular}{|l|l|l|l|l|l|} \hline \multicolumn{1}{|l|}{\textbf{Metric}} & \multicolumn{1}{l|}{\textbf{RNN}} & \multicolumn{1}{l|}{\textbf{GRU}} & \multicolumn{1}{l|}{\textbf{LSTM}} & \multicolumn{1}{l|}{\textbf{CNN}} & \multicolumn{1}{l|}{\textbf{LSTM-CNN}} \\ \hline \multicolumn{1}{|l|}{\textbf{MSE}} & \multicolumn{1}{l|}{0.2662} & \multicolumn{1}{l|}{0.1864} & \multicolumn{1}{l|}{0.1803} & \multicolumn{1}{l|}{0.2456} & \multicolumn{1}{l|}{0.1437} \\ \hline \multicolumn{1}{|l|}{\textbf{RMSE}} & \multicolumn{1}{l|}{0.5432} & \multicolumn{1}{l|}{0.4127} & \multicolumn{1}{l|}{0.3890} & \multicolumn{1}{l|}{0.5329} & \multicolumn{1}{l|}{0.3243} \\ \hline \multicolumn{1}{|l|}{\textbf{MAPE (\%)}} & \multicolumn{1}{l|}{15.9} & \multicolumn{1}{l|}{11.1} & \multicolumn{1}{l|}{10.8} & \multicolumn{1}{l|}{13.6} & \multicolumn{1}{l|}{9.7} \\ \hline \end{tabular} \end{table} \begin{table*}[!ht] \centering \caption{Evaluation of proposed FL framework against Threat Model 1 \& 2 \label{EvaluationFLThreat}} \begin{tabular}{cc|cc|cc|} \cline{3-6} \multicolumn{2}{l|}{\textbf{}} & \multicolumn{2}{c|}{\textbf{Fed-SGD}} & \multicolumn{2}{c|}{\textbf{Proposed Solution}} \\ \hline \multicolumn{1}{|c|}{\textbf{\begin{tabular}[c]{@{}c@{}}\% of Compromised\\ Clients\end{tabular}}} & \textbf{Metric} & \multicolumn{1}{c|}{\textbf{Threat Model 1}} & \textbf{Threat Model 2} & \multicolumn{1}{l|}{\textbf{Threat Model 1}} & \multicolumn{1}{l|}{\textbf{Threat Model 2}} \\ \hline \multicolumn{1}{|c|}{\multirow{3}{*}{\textbf{10}}} & \textbf{MSE} & \multicolumn{1}{c|}{0.2910} & 0.3134 & \multicolumn{1}{c|}{0.1621} & 0.1532 \\ \cline{2-6} \multicolumn{1}{|c|}{} & \textbf{RMSE} & \multicolumn{1}{c|}{0.4732} & 0.5490 & \multicolumn{1}{c|}{0.3251} & 0.3029 \\ \cline{2-6} \multicolumn{1}{|c|}{} & \textbf{MAPE (\%)} & \multicolumn{1}{c|}{18.2} & 20.1 & \multicolumn{1}{c|}{10.1} & 9.9 \\ \hline \multicolumn{1}{|c|}{\multirow{3}{*}{\textbf{20}}} & \textbf{MSE} & \multicolumn{1}{c|}{0.4180} & 0.4519 & \multicolumn{1}{c|}{0.1835} & 0.1642 \\ \cline{2-6} \multicolumn{1}{|c|}{} & \textbf{RMSE} & \multicolumn{1}{c|}{0.7893} & 0.9201 & \multicolumn{1}{c|}{0.3502} & 0.3129 \\ \cline{2-6} \multicolumn{1}{|c|}{} & \textbf{MAPE (\%)} & \multicolumn{1}{c|}{25.7} & 27.1 & \multicolumn{1}{c|}{12.2} & 10.8 \\ \hline \multicolumn{1}{|c|}{\multirow{3}{*}{\textbf{30}}} & \textbf{MSE} & \multicolumn{1}{c|}{0.7319} & 0.8192 & \multicolumn{1}{c|}{0.2678} & 0.2134 \\ \cline{2-6} \multicolumn{1}{|c|}{} & \textbf{RMSE} & \multicolumn{1}{c|}{1.2398} & 1.4576 & \multicolumn{1}{c|}{0.4249} & 0.3965 \\ \cline{2-6} \multicolumn{1}{|c|}{} & \textbf{MAPE (\%)} & \multicolumn{1}{c|}{38.9} & 42.2 & \multicolumn{1}{c|}{17.3} & 14.1 \\ \hline \end{tabular} \end{table*} \begin{table}[] \caption{Evaluation of proposed FL framework against Threat Model 3\label{EvaluationFLThreat3}} \begin{tabular}{|c|c|c|c|} \hline \multicolumn{1}{|l|}{\textbf{\% of Comp. Clients}} & \multicolumn{1}{l|}{{ \textbf{Metric}}} & \multicolumn{1}{l|}{{\textbf{Fed-SGD}}} & \multicolumn{1}{l|}{{\textbf{Proposed Solution}}} \\ \hline & \textbf{MSE} & 0.3103 & 0.1732 \\ \cline{2-4} & \textbf{RMSE} & 0.5321 & 0.3324 \\ \cline{2-4} \multirow{-3}{*}{\textbf{20}} & \textbf{MAPE (\%)} & 19.3 & 11.2 \\ \hline & \textbf{MSE} & 0.5231 & 0.2034 \\ \cline{2-4} & \textbf{RMSE} & 0.8743 & 0.3958 \\ \cline{2-4} \multirow{-3}{*}{\textbf{30}} & \textbf{MAPE (\%)} & 34.0 & 14.0 \\ \hline & \textbf{MSE} & 0.7793 & 0.2901 \\ \cline{2-4} & \textbf{RMSE} & 1.2343 & 0.4302 \\ \cline{2-4} \multirow{-3}{*}{\textbf{40}} & \textbf{MAPE (\%)} & 39.5 & 16.4 \\ \hline \end{tabular} \end{table} \begin{table}[] \centering \caption{Evaluation of proposed FL framework under different Privacy Budgets \label{threat3}} \begin{tabular}{|c|c|c|c|} \hline \textbf{$\epsilon$-Budget} & \textbf{Metric} & \textbf{Fed-SGD} & \multicolumn{1}{l|}{\textbf{Proposed Solution}} \\ \hline \multirow{3}{*}{\textbf{0.01}} & \textbf{MSE} & 0.1583 & 0.1437 \\ \cline{2-4} & \textbf{RMSE} & 0.3 & 0.3243 \\ \cline{2-4} & \textbf{MAPE (\%)} & 9.7 & 9.7 \\ \hline \multirow{3}{*}{\textbf{0.1}} & \textbf{MSE} & 0.4320 & 0.1645 \\ \cline{2-4} & \textbf{RMSE} & 0.8173 & 0.3192 \\ \cline{2-4} & \textbf{MAPE (\%)} & 26.4 & 10.5 \\ \hline \end{tabular} \end{table} \begin{figure*}[!h] \centering \subfloat[\centering Impact of Threat Model 1 ]{{\includegraphics[width=5cm]{Attack1FedSGD.png} }}% \subfloat[\centering Impact of Threat Model 2 ]{{\includegraphics[width=5cm]{ImpactT2FedSGD.png} }}% \subfloat[\centering Impact of Threat Model 3 ]{{\includegraphics[width=5cm]{ImpactofEPonFedSGD.png} }}% \caption{Impact of Attacks on Fed-SGD}% \label{fig:impactFedSGD} \end{figure*} \begin{figure*}[!h] \centering \subfloat[\centering Impact of Threat Model 1 ]{{\includegraphics[width=5cm]{Threat1Mitig.png} }}% \subfloat[\centering Impact of Threat Model 2 ]{{\includegraphics[width=5cm]{Threat2Mitig.png} }}% \subfloat[\centering Impact of Threat Model 3 ]{{\includegraphics[width=5cm]{Threat3Mitig.png} }}% \caption{Mitigating threat models using our proposed method}% \label{fig:impactSIGNSGD} \end{figure*} As our proposed solution converges faster that the traditional Fed-SGD one, we can conclude that the proposed approach provides a fast algorithmic convergence. Furthermore, we use the three aforementioned evaluation metrics to compare and contrast the performance of the proposed solution against Fed-SGD with several models as presented in Table \ref{CompaGedSGD} and Table \ref{CompaSignSGDModel}. The experimental results reveal that the the proposed framework reaches similar performance as compared to the Fed-SGD approach. Similarly, in Fig. \ref{fig:trainloss}(b), the MAPE per active household within the FL set ups are contrasted which shows that our proposed approach reaches relatively similar performance as compared to the Fed-SGD. More specifically, after the comparison, we can deduce that our proposed framework reaches good generalization performance for short-term load forecasting within acceptable error ranges. Moreover, after comparing the proposed framework based on models as presented in Table \ref{CompaSignSGDModel}, it can be deduced that LSTM-CNN model shows the best overall forecasting performance with an average MAPE of 9.7\% in both the conventional Fed-SGD and the proposed FL framework. \subsection{Impact on attacks on proposed framework} In this section, we evaluate the robustness of our proposed FL framework against the adversarial threat models as described in Section \ref{sect:probdef}. To discuss the impact of Byzantine Attacks on the standard Fed-SGD and our proposed approach, we further divide the results into the two following sections: \subsubsection{Impact of attacks on Fed-SGD} After evaluating the impact of the three byzantine threat models as in Section \ref{sect:probdef}, we present the results within this section. In Table \ref{EvaluationFLThreat}, we evaluated the performance of Fed-SGD under Threat Model 1 \& 2 respectively. For both threat models, there is a direct relationship between the percentage of compromised active FL clients and the mean error of the FedSGD FL model, that is, once the percentage of compromised clients increases, the mean error of the FL model decreases. Specifically, for threat models 1 \& 2, at 10\% of compromised clients, the MAPE of the FL model is 18.2\% and 20.1\% respectively, thereby following an upward trend such that at 30\% of compromised, the MAPE of the FL model reaches around 38.9\% and 42.2\% respectively. It is worth noting that once a third of the clients are compromised/malicious, there is almost around an average of 40\% difference between the actual value and the forecasted value. On the other hand, Table \ref{EvaluationFLThreat3} investigates the impact of the colluding attack on the Fed-SGD setup. Similarly, the the number of compromised clients is directly proportional to the mean error of the FL model. As the number of colluding adversaries increases, the mean error of the FL model also increases. Furthermore, based on Fig. \ref{fig:impactFedSGD}(a) and Fig. \ref{fig:impactFedSGD}(b), we can note that as the percentage of compromised clients increase, the FL model loss starts to diverge after a certain number of communication rounds due to threat models 1 \& 2. Similarly, based on Fig. \ref{fig:impactFedSGD}(c), as the number of colluding adversaries increases, the FL model loss starts to diverge after a certain number of communication rounds. \subsubsection{Impact of proposed FL framework on attacks} Within the previous section, we discussed the impact of attacks on the standard Fed-SGD setup. However, throughout this one, we will discuss the impact of our proposed solution on mitigating the threat models presented in Section \ref{sect:probdef}. As presented in Table \ref{EvaluationFLThreat}, when our proposed solution is under attack by threat models 1 \& 2, at 10\% of compromised clients, the mean error of the FL model stayed relatively similar to the MAPE of the model prior to any attacks. Gradually, with increasing percentage of compromised clients, it can be seen that there is a slight increase in the MAPE. Specifically, from 20\% to 30\% of compromised clients, the MAPE is 5.1\% and 3.3\% for threat models 1 \& 2 respectively. However, the small increase in the mean error of the FL model is still within acceptable ranges. Similarly, based on Table \ref{EvaluationFLThreat3}, it is evident that there is a very slight increase (within acceptable error ranges) in the MAPE value as the number of compromised clients increases. Furthermore, based on Fig. \ref{fig:impactSIGNSGD}, we notice that the under all percentages of compromised clients, our proposed model is optimized such that it converges after a certain number of communication rounds/iterations. Therefore, we can eventually conclude that our proposed approach effectively mitigates byzantine attacks. \subsection{Results Discussion} With increasing concerns and regulation enforcement in regards to security and privacy within the smart grid paradigm, it is crucial to develop privacy-preserving and robust short term load forecasting solutions. FL, whilst still being in its infant stage, requires further improvements under different circumstances. Therefore, throughout this study, we investigate Byzantine attacks in relation to federated short term load forecasting. Furthermore, we propose and design a robust defense solution to mitigate those threats. From Table \ref{CompaGedSGD} and \ref{CompaSignSGDModel} above, it can be seen that our proposed approach reaches comparable forecasting performance as FedSGD when there are no attacks. Similarly, when compared to several other time-series forecasting models, our proposed approach matches that of Fed-SGD. More specifically, we achieved the best overall performance of our proposed approach using the LSTM-CNN model with a MAPE of 9.7\% for both FL setups. Therefore, we selected LSTM-CNN as the principal model to evaluate our proposed approach under the three threat models as discussed in Section \ref{sect:probdef}. Based on the experimental results presented in Tables \ref{EvaluationFLThreat} and \ref{EvaluationFLThreat3} as well as the Figs. \ref{fig:impactFedSGD} and \ref{fig:impactSIGNSGD}, under the conventional Fed-SGD approach, we notice an overall degradation in the performance of the model with increasing intensity of attacks. For instance, an increase in the percentage of compromised clients results an upward shift in the mean error of the model. On the flip side, we notice that our proposed approach can withstand such attacks with minimal impact on the mean error of the FL model. This leads us to conclude that it is indeed a resilient and privacy-preserving FL set-up for residential short-term load forecasting. \section{Conclusion} \label{Conclusion} The rapid adoption of FL within the smart grid ecosystem has spiked the interest of researchers to address its security and privacy issues. Byzantine attack mitigation plays a crucial role in securing and enhancing the robustness of FL for short-term load forecasting. Therefore, throughout this manuscript, we propose a state-of-the-art FL-based approach that leverages the notions of gradient quantization and differential privacy to overcome this challenge. Furthermore, we empirically demonstrate that our proposed solution effectively mitigate popular byzantine threats and provides relatively similar performance as compared to standard FL setups. Finally, the next steps in this research are to: (1) design and evaluate our proposed FL framework against stronger byzantine attacks, and, (2) take into consideration the existence of distributed energy resources to improve the grid model. \bibliographystyle{IEEEtran}

f7a170e0d0294d6773f8ac4a453864b9b89cd75a

\section{Introduction} \label{sec_introduction} The Epoch of Reionisation (EoR) marks the second major phase transition in the Universe. With the emergence of the first galaxies, ultraviolet (UV) radiation gradually ionises the neutral hydrogen (\HI) in the intergalactic medium (IGM) until the Universe is reionised by $z\simeq5.3$ \citep{Keating2020, Zhu2021, Bosman2021}. However, as only the brighter galaxies during the EoR are observed to date, key questions detailing the reionisation process remain outstanding: Did the few bright and more massive or the numerous faint and low-mass galaxies contribute more to reionisation? Feedback processes, such as heating by supernovae (SN) and photoionisation, suppress star formation in low-mass galaxies \citep{Gnedin2014, ocvirk2016, Ocvirk2018, Hutter2021a}, and reduce the contribution of very low-mass galaxies to reionisation. An even more critical quantity that regulates the ionising radiation (with energies $E>13.6$~eV) escaping from galaxies and driving the reionisation of the IGM is the fraction of ionising photons $f_\mathrm{esc}$ that escape from galaxies into the IGM \citep{Hutter2021b}. While the presence of \HI in the IGM during the EoR impedes direct measurements of $f_\mathrm{esc}$, different theoretical models and simulations have investigated the physical processes determining and dependencies of $f_\mathrm{esc}$ \citep[e.g.][]{Ferrara2013, Wise2014, Kimm2014, Kimm2019}. Cosmological radiation hydrodynamical simulations suggest that $f_\mathrm{esc}$ decreases towards deeper gravitational potential \citep[e.g.][]{Wise2014, Kimm2014, Kimm2017, Kimm2019, Xu2016, anderson2017, lewis2020}. High-resolution simulations of the ISM indicate that $f_\mathrm{esc}$ is dominated by the escape from star-forming clouds. The ionising radiation of massive stars and their explosions as SN ionise, heat and destroy the star-forming clouds clearing the way for the ionising radiation to escape \citep{Howard2018, Kim2019, He2020, Kimm2021}. The complex dependency of $f_\mathrm{esc}$ on the underlying gravitational potential, the gas distribution and stellar populations in the ISM leaves marks not only in the radiation emitted by galaxies but also in the ionisation topology, the time and spatial distribution of the ionised regions around galaxies. Current and forthcoming observations of galaxies and the ionisation state of the IGM have the potential to constrain galactic properties, such as $f_\mathrm{esc}$, and the reionisation process. On the one hand, detecting the 21cm signal from \HI in the IGM with forthcoming large radio interferometers (e.g. Square Kilometre Array) will measure the ionisation topology, which provides constraints on the dependence of $f_\mathrm{esc}$ on galaxy mass \citep{kim2013b, Seiler2019, Hutter2020}. On the other hand, being extremely sensitive to the attenuation by \HI in the IGM, the observable Lyman-$\alpha$ (Ly$\alpha$) radiation at $1216$\AA~ from high-redshift galaxies has gained popularity in probing reionisation for the following reason: A $z\gtrsim6$ galaxy only exhibits detectable Ly$\alpha$ emission when: (i) it is surrounded by an ionised region that is large enough to allow a sufficient fraction of its emerging Ly$\alpha$ line to traverse the IGM, or (ii) it is gas-rich enough (corresponding to a high \HI column density) such that the red part of the Ly$\alpha$ line emerging from the galaxy is redshifted out of absorption, or (iii) it has strong outflows that also redshift the emerging Ly$\alpha$ line out of absorption, or it is a combination of all three. The first criterion suggest that more massive galaxies able to retain more gas might be the most likely to show observable Ly$\alpha$ emission during the EoR: their higher rates of forming stars emitting ionising photons lead to an increased production of Ly$\alpha$ radiation in the ISM and the growth of large ionised regions around them. The latter is accelerated by their ionised regions merging earlier with those of the surrounding lower mass objects attracted by their deeper gravitational potentials \citep{Chardin2012, Furlanetto2016, Chen2019}. As reionisation progresses and the ionised regions grow, increasingly lower mass galaxies become visible as Ly$\alpha$ emitters (LAEs), which leads not only to a higher fraction of galaxies showing Ly$\alpha$ emission but also to a reduced clustering of LAEs \citep{mcquinn2007, jensen2013, Hutter2015, Sobacchi2015}. This picture is increasingly supported by observations of $z>6$ LAEs. Not only the fraction of Lyman Break Galaxies (LBGs) showing Ly$\alpha$ emission rises from $z\simeq8$ to $z\simeq6$ \citep{Schenker2014, Pentericci2014, Pentericci2018, Fuller2020}, but also the majority of Ly$\alpha$ emission at $z\gtrsim6.5$ is detected in galaxies with a bright UV continuum \citep{Cuby2003, Willott2013, Oesch2015, Sobral2015, Zitrin2015, Roberts-Borsani2016, Matthee2017, Songaila2018, Hashimoto2018, Taylor2020, Endsley2022, Endsley_Stark2022}. Moreover, the close proximity of UV-bright LAEs suggests that LAEs are located in over-dense regions \citep{Vanzella2011, Castellano2016, Castellano2018, Jung2020, Tilvi2020, Hu2021, Endsley_Stark2022} that exhibit the first and largest ionised regions during the EoR. This hypothesis is also in line with the observed double-peaked Ly$\alpha$ profiles in $z\gtrsim6.5$ galaxies \citep{Songaila2018, Hu2016, Matthee2018, Meyer2021}, indicating that the ionised regions surrounding them are so large that even the part bluewards the Ly$\alpha$ resonance redshifts out of resonance. Current theoretical predictions of the large-scale LAE distribution confirm this picture, suggesting that the LAEs we see during the EoR are more massive galaxies naturally located in over-dense regions \citep[c.f.][]{Dayal2011, jensen2013, Hutter2014, Mesinger2015, Weinberger2018, Qin2022}. Yet, all these LAE models effectively assume a constant $f_\mathrm{esc}$ value across the entire galaxy population at a given redshift. This assumption remains highly uncertain as $f_\mathrm{esc}$ is very sensitive to the ISM and the circumgalactic medium (CGM) of galaxies that again depend on the underlying gravitational potential of a galaxy. However, it is essential, since $f_\mathrm{esc}$ defines the critical processes that shape the Ly$\alpha$ luminosities observed from galaxies. An $f_\mathrm{esc}$ varying with galactic properties and the underlying gravitational potential might alter the galaxy population seen as LAEs for the following reasons: Firstly, within a galaxy, most Ly$\alpha$ radiation is produced by recombining hydrogen atoms and scales with the number of \HI ionising photons absorbed within the galaxy ($\propto 1-f_\mathrm{esc}$). Secondly, a fraction of these Ly$\alpha$ photons undergoes only a few scattering events when they escape through the same low-density tunnels that facilitate the escape of \HI ionising photons. In contrast, the other fraction that traverses optically thick clouds upon its escape is scattered and absorbed by hydrogen and dust, respectively \citep[see, e.g.][]{Verhamme2015, Dijkstra2016, Kimm2019, Kakiichi2021}. These different escape mechanisms result not only in $f_\mathrm{esc}$ posing a lower limit to the fraction of Ly$\alpha$ photons escaping from a galaxy but also determining the Ly$\alpha$ line profile that emerges from a galaxy. Detailed low-redshift galaxy observations increasingly supported the $f_\mathrm{esc}$-sensitivity of these Ly$\alpha$ properties \citep{Verhamme2017, Jaskot2019, Gazagnes2020}. Thirdly, $f_\mathrm{esc}$ shapes the IGM ionisation topology by determining the number of ionising photons available to ionise the IGM surrounding a galaxy. While a higher $f_\mathrm{esc}$ value enlarges the ionised region surrounding a galaxy and enhances the transmission of Ly$\alpha$ radiation through the IGM \citep{Dayal2011, Hutter2014}, the corresponding Ly$\alpha$ line emerging from a galaxy will be more peaked around the Ly$\alpha$ resonance and raise the absorption by \HI in the IGM. Given this complex $f_\mathrm{esc}$-dependency of the observed Ly$\alpha$ luminosity, it remains unclear whether different dependencies of $f_\mathrm{esc}$ with galaxy properties (e.g. increasing or decreasing with rising halo mass) would (i) identify the same galaxies as LAEs (exceeding a threshold Ly$\alpha$ luminosity) and/or (ii) lead to different spatial large-scale distribution of the LAEs' Ly$\alpha$ luminosities. In other words, which of these $f_\mathrm{esc}$-dependent Ly$\alpha$ processes dominates the observed Ly$\alpha$ luminosities? For example, is the $f_\mathrm{esc}$-dependency of the intrinsic Ly$\alpha$ luminosity dominant, and we yield a weaker clustering of LAEs when $f_\mathrm{esc}$ value decreases with rising halo mass? Or do they compensate each other once we reproduce the observed Ly$\alpha$ luminosity functions (Ly$\alpha$ LFs)? To address these questions, we use our {\sc astraeus} framework that models galaxy evolution and reionisation self-consistently \citep{Hutter2021a, Ucci2022}, and simulate different reionisation scenarios that gauge the physically plausible range of $f_\mathrm{esc}$ dependencies, i.e. $f_\mathrm{esc}$ decreasing and increasing with rising halo mass. Moreover, we parameterise results from numerical Ly$\alpha$ radiative transfer (RT) simulations of clumpy media \citep{Gronke2017} and build an analytic model for the fraction of Ly$\alpha$ photons escaping and the corresponding Ly$\alpha$ line profile emerging from high-redshift galaxies. Importantly, we explore three different Ly$\alpha$ line profile models, including (i) a Gaussian profile around the Ly$\alpha$ resonance where the Ly$\alpha$ escape fraction is directly related to the dust attenuation of the UV continuum \citep[used in previous LAE models outlined in][]{Dayal2011, Hutter2014}, (ii) a Ly$\alpha$ line profile emerging from a shell of outflowing dusty gas clumps, which we model by using the different Ly$\alpha$ escape regimes identified in \citet{Gronke2017}, and (iii) a Ly$\alpha$ line profile emerging from a shell of outflowing gas clumps with a fraction $f_\mathrm{esc}$ of the solid angle interspersed by gas-free tunnels. The latter two give rise to various combinations of a central peak around the Ly$\alpha$ resonance (Ly$\alpha$ photons hardly scatter in an optically thin medium) and two peaks in the red and blue wings (Ly$\alpha$ photons are scattered in an optically thick medium). By deriving the observed Ly$\alpha$ luminosities of all simulated galaxies for all combinations of reionisation scenarios and Ly$\alpha$ line models, we address the following questions: Which $f_\mathrm{esc}$-dependent Ly$\alpha$ process, i.e. intrinsic production, escape or transmission through the IGM of Ly$\alpha$ radiation, dominates the observed Ly$\alpha$ luminosity? Can the observed Ly$\alpha$ luminosities of galaxies inform us on their emerging Ly$\alpha$ line profile? Given the ionisation topology depends sensitively on the assumed dependency of $f_\mathrm{esc}$ with halo mass, are the same or different galaxies identified as LAEs and do they differ in the spatial distribution of their Ly$\alpha$ luminosities? This paper is organised as follows. In Section \ref{sec_model} we briefly describe the {\sc astraeus} model, its implementation of dust and the different reionisation simulations. In Section \ref{sec_modelling_LAEs} we introduce the different Ly$\alpha$ line profile models and their corresponding attenuation by dust. We then (Section \ref{sec_number_and_properties_LAEs}) discuss how the Ly$\alpha$ line profiles depend on halo mass in our different reionisation scenarios, how free model parameters, such as the ISM clumpiness or size of the dust gas clumps, need to be adjusted to fit the observed Ly$\alpha$ LFs, and how the galaxy properties determining the observed Ly$\alpha$ luminosities depend on the halo mass of a galaxy. In Section \ref{sec_spatial_distribution_LAEs} we identify the location of LAEs in the large-scale density and ionisation structure and assess whether the spatial distribution of LAEs differs for different $f_\mathrm{esc}$-dependencies on halo mass/ionisation topologies. Finally, we briefly discuss which Lyman Break galaxies are preferentially identified as LAEs (Section \ref{sec_LAE_LBG_relation}) and conclude in Section \ref{sec_conclusions}. In this paper we assume a $\Lambda$CDM Universe with cosmological parameter values of $\Omega_\Lambda=0.69$, $\Omega_m=0.31$, $\Omega_b=0.048$, $H_0=100h=67.8$km~s$^{-1}$Mpc$^{-1}$, $n_s=0.96$ and $\sigma_8=0.83$, and a Salpeter initial mass function \citep[IMF;][]{salpeter1955} between $0.1\msun$ to $100\msun$. \section{The model and simulations} \label{sec_model} In this paper, we use the {\sc astraeus} framework. This framework couples a semi-analytic galaxy evolution model (an enhanced version of {\sc delphi}; \citet{Dayal2014}) with a semi-numerical reionisation scheme ({\sc cifog}; \citet{Hutter2018a}) and runs the resulting model on the outputs of a dark matter (DM) only N-body simulation. In this Section, we provide a brief description of the physical processes implemented in {\sc astraeus} \citep[for more details, see][]{Hutter2021a} and introduce the different reionisation simulations. \subsection{N-body simulation} \label{subsec_Nbody_simulation} As part of the Multidark simulation project, the underlying DM N-body simulation ({\sc very small multidark planck; vsmdpl}) has been run with the {\sc gadget-2 tree+pm} code \citep{springel2005}. In a box with a side length of $160h^{-1}$Mpc, it follows the trajectories of $3840^3$ DM particles. Each DM particle has a mass of $6\times10^6 h^{-1}\msun$. For a total of $150$ snapshots ranging from $z=25$ to $z=0$, the phase space {\sc rockstar} halo finder \citep{behroozi2013_rs} has been used to identify all halos and subhalos down to $20$ particles or a minimum halo mass of $1.24 \times 10^8h^{-1}\msun$. To obtain the local horizontal merger trees (sorted on a redshift-by-redshift basis within a tree) for galaxies at $z=4.5$ that {\sc astraeus} requires as input, we have used the pipeline internal {\sc cutnresort} scheme to cut and resort the vertical merger trees (sorted on a tree-branch by tree-branch basis within a tree) generated by {\sc consistent trees} \citep{behroozi2013_trees}. For the first $74$ snapshots that range from $z=25$ to $z=4.5$, we have generated the DM density fields by mapping the DM particles onto $2048^3$ grids and re-sampling these to $512^3$ grids used as input for the {\sc astraeus} pipeline. \subsection{Galaxy evolution} \label{subsec_galaxy_evolution} {\sc astraeus} tracks key processes of early galaxy formation and reionisation. At each time step and for each galaxy, it tracks the amount of gas that is accreted, the gas and stellar mass merging, star formation and associated feedback from SNII and metal enrichment, as well as the large-scale reionisation process and its associated feedback on the gas content of early galaxies. \subsubsection{Gas and stars} \label{subsubsec_gas_and_stars} In the beginning, when a galaxy starts forming stars in a halo with mass $M_h$, it has a gas mass of $M_g^i(z)=f_g (\Omega_b/\Omega_m) M_h(z)$, with $f_g$ being the gas fraction not evaporated by reionisation, i.e. $f_g=1$ and $f_g<1$ as the galaxy forms in a neutral and ionised region, respectively. In subsequent time steps a galaxy gains gas from its progenitors ($M_g^\mathrm{mer}(z)$) and smooth accretion ($M_g^\mathrm{acc}$), while its total gas mass never exceeds the limit given by reionisation feedback: $M_g^i = \min\left( M_g^\mathrm{mer}(z) + M_g^\mathrm{acc}(z), f_g (\Omega_m/\Omega_b) M_h \right)$ with $M_g^\mathrm{acc}=M_h(z) - \sum_{p=1}^\mathrm{N_p} M_{h,p}(z + \Delta z)$ and $M_g^\mathrm{mer}(z)=\sum_{p=1}^\mathrm{N_p} M_{h,p} (z + \Delta z)$ where $N_p$ is the galaxy's number of progenitors and $M_{h,p}$ the halo mass of each progenitor. At each time step, a fraction of the merged and accreted (initial) gas mass is transformed into stellar mass, $M_\star^\mathrm{new}(z)=(f_\star^\mathrm{eff}/\Delta t) M_g^i(z)$.\footnote{We note that this definition has been altered compared to the first version of {\sc astraeus} in \citep{Hutter2021a}.} Here $f_\star^\mathrm{eff}$ represents the fraction of gas that forms stars over a time span $\Delta t$ and is limited by the minimum amount of stars that need to form to eject all gas from the galaxy, $f_\star^\mathrm{ej}$, and an upper limit, $f_\star$. $f_\star^\mathrm{eff}$ depends on the gravitational potential: more massive galaxies form stars at the constant rate $f_\star$, while low-mass galaxies form stars at the limited rate $f_\star^\mathrm{ej}$ due to SN and radiative feedback. While we account for radiative feedback from reionisation by modifying the initial gas mass reservoir with the factor $f_g$, $f_\star^\mathrm{eff}$ incorporates the suppression of star formation in low-mass halos as gas is heated and ejected by SNII explosions. Our model incorporates a delayed SN feedback scheme, i.e. at each time step the effective star formation efficiency accounts for the SNII energy released from stars formed in the current and previous time steps, following the mass-dependent stellar lifetimes \citep{padovani1993}. In contrast to \citet{Hutter2021a}, we have updated our model and do not assume stars to form in bursts to calculate the number of SNII exploding within a time step but $M_\star^\mathrm{new}(z)$ to form at a constant star formation over the entire time step (see Appendix \ref{app_delayed_non-bursty_SNscheme} for a detailed calculation). The star formation efficiency in the SN feedback-limited regime is given by $f_\star^\mathrm{ej}(z) = \frac{v_c^2}{v_c^2 + f_w E_{51} \nu_z} \left[ 1 - \frac{f_w E_{51} \sum_j \nu_j M_{\star,j}^\mathrm{new}(z_j)}{M_\mathrm{g}^i(z)~ v_c^2} \right]$, with $v_c$ being the rotational velocity of the halo, $E_{51}$ the energy released by a SNII, $f_w$ the fraction of SNII energy injected into the winds driving gas outflows, $M_{\star,j}^\mathrm{new}(z_j)$ the stellar mass formed during previous time steps $j$, and $\nu_j$ the fraction of stellar mass formed in previous time step $j$ that explodes in the current time step given the assumed IMF. {\sc astraeus} incorporates multiple models for radiative feedback from reionisation, ranging from a weak and time-delayed ({\it Weak Heating}) to a strong instantaneous feedback ({\it Jeans mass}). In this work, we use the intermediate and time-delayed {\it Photoionisation} model, where the characteristic mass defining the gas fraction not evaporated by reionisation grows on a dynamical timescale to the respective Jeans mass \citep[for a detailed description see][]{Hutter2021a}. We list the {\sc astraeus} model parameters and their assumed values in Table \ref{table_model_params}. $f_\star$ and $f_w$ have been adjusted to reproduce the observed UV LFs, stellar mass functions, global star formation rate density, and global stellar mass density at $z=10-5$. \begin{table} \centering \caption{{\sc astraeus} model parameters and chosen values in this work.} \label{tab:example_table} \begin{tabular*}{\columnwidth}{ccc} \hline\hline Parameter & Value or reference & Description\\ \hline $f_\star$ & $0.025$ & Maximum star-formation efficiency \\ $f_w$ & $0.2$ & SN coupling efficiency \\ - & Photoionization & Radiative feedback model \\ IMF & \citet{salpeter1955} & For stellar evolution, enrichment, SED \\ SED & \textsc{Starburst99} & ionizing SED model \\ \hline \end{tabular*} \label{table_model_params} \end{table} \subsubsection{Metals and dust} \label{subsubsec_metals_and_dust} The current {\sc astraeus} model also incorporates the metal enrichment by stellar winds, SNII and SNIa explosions \citep[for a detailed description see][]{Ucci2022}. At each time step, we assume that gas smoothly accreted has the average metallicity of the gas in the IGM, $Z_\mathrm{IGM}$. Metals are produced through stellar winds, SNII and SNIa explosions. The amount of newly forming metals depends on the number of massive stars exploding as SN in the current time step according to \citet{padovani1993}, \citet{yates2013} and \citet{maoz2012}. For the corresponding stellar metal yields, {\sc astraeus} uses the latest yield tables from \citet{Kobayashi2020b}. We assume that gas and metals are perfectly mixed. Thus, the metals ejected from the galaxy are proportional to the ejected gas mass and the metallicity of the gas in the galaxy. This ejected metal mass contributes to $Z_\mathrm{IGM}$. In this work, we have extended the {\sc astraeus} model \citep{Hutter2021a, Ucci2022} to follow the formation, growth, destruction, astration and destruction of dust in each galaxy \citep[c.f.][for details]{Dayal2022}. We note that we consider dust to be part of our metal reservoir (i.e. $M_\mathrm{dust}\leq M_\mathrm{m}$). At each time step, {\sc astraeus} computes the evolution of the dust mass $M_\mathrm{dust}$ in a galaxy by solving the following differential equation \begin{eqnarray} \frac{\mathrm{d}M_\mathrm{dust}}{\mathrm{d}t} &=& \dot{M}_\mathrm{dust}^\mathrm{prod} + \dot{M}_\mathrm{dust}^\mathrm{grow} - \dot{M}_\mathrm{dust}^\mathrm{dest} - \dot{M}_\mathrm{dust}^\mathrm{astr} - \dot{M}_\mathrm{dust}^\mathrm{ej}. \label{eq_dust} \end{eqnarray} The first term on the right hand side (RHS) of Eqn. \ref{eq_dust} denotes the production of dust in SNII and AGB stars through condensation of metals in stellar ejecta \begin{eqnarray} \dot{M}_\mathrm{dust}^\mathrm{prod} &=& y_\mathrm{SNII} \gamma_\mathrm{SNII} + \dot{M}_\mathrm{dust}^\mathrm{AGB}, \end{eqnarray} with $y_\mathrm{SNII}=0.45\msun$ being the dust mass formed per SNII, \begin{eqnarray} \gamma_\mathrm{SN}(t) &=& \int_{8\msun}^{40\msun} \mathrm{SFR}(t - \tau_m) \phi(m) \mathrm{d}m \end{eqnarray} the number of SNII events, \begin{eqnarray} \dot{M}_\mathrm{dust}^\mathrm{AGB}(t) &=& \int_{0.85\msun}^{50\msun} y_\mathrm{AGB}(m) \mathrm{SFR}(t - \tau_m) \phi(m) \mathrm{d}m \end{eqnarray} the contribution from AGB stars and $y_\mathrm{AGB}$ the dust yields from AGB stars. In agreement with \citep{Ucci2022}, we adopt the latest yield tables from \citet{Kobayashi2020b} for $y_\mathrm{AGB}$. The second term on the RHS of Eqn. \ref{eq_dust} describes the dust grain growth through the accretion of heavy elements in dense molecular clouds in the ISM, \begin{eqnarray} \dot{M}_\mathrm{dust}^\mathrm{grow} &=& \left( Z' - \frac{M_\mathrm{dust}}{M_\mathrm{g}^i} \right) f_\mathrm{cold~gas} \frac{M_\mathrm{dust}}{\tau_\mathrm{gg,0} Z_\odot} \end{eqnarray} where $Z'$ is the metallicity after accretion and star formation, $M_\mathrm{dust}$ is the dust mass, $f_\mathrm{cold~gas}$ the fraction of cold and molecular gas, and $\tau_\mathrm{gg}=\tau_\mathrm{0,gg} / Z$ the accretion timescale adopted from \citet{Asano2013} \citep[see also][]{Triani2020}. We assume $f_\mathrm{cold~gas}=0.5$ and $\tau_\mathrm{gg,0}=30$Myrs. The third term in Eqn. \ref{eq_dust} describes the destruction of dust by SN blastwaves, for which we adopt the analytic description outlined in \citet{mckee1989} \begin{eqnarray} \dot{M}_\mathrm{dust}^\mathrm{dest} &=& \left( 1 - f_\mathrm{cold~gas} \right) \frac{M_\mathrm{dust}}{M_\mathrm{g}^i}\ \gamma_\mathrm{SN} \epsilon\ M_\mathrm{SN, bw}, \end{eqnarray} with $\epsilon$ being the effifiency of dust destruction in a SN-shocked ISM and $M_\mathrm{SN, bw}$ the mass accelerated to $100$~km~s$^{-1}$ by the SN blast wave. In line with \citet{mckee1989} and \citet{lisenfeld_ferrara1998} we adopt $\epsilon=0.03$ and $M_\mathrm{SN, bw}=6.8\times10^3\msun$. Finally, Eqn. \ref{eq_dust} accounts also for the destruction of dust by astration as new stars form from the metal-enriched gas, \begin{eqnarray} \dot{M}_\mathrm{dust}^\mathrm{astr} &=& Z^\mathrm{i}\ \frac{M_\star^\mathrm{new}}{\Delta t}, \end{eqnarray} and the ejection of metals through winds powered by the energy injected by SN, \begin{eqnarray} \dot{M}_\mathrm{dust}^\mathrm{ej} &=& Z'\ \frac{M_\mathrm{g}^\mathrm{ej}}{\Delta t}. \end{eqnarray} The parameter values ($y_\mathrm{SNII}$, $\tau_\mathrm{gg,0}$, $\epsilon$, $M_\mathrm{SN,bw}$) quoted reasonably reproduce the observed UV LFs when the UV is attenuated by dust as follows \citep[UV LFs data points include][]{Atek2015, Atek2018, Bouwens2015, Bouwens2016, Bouwens2017, Bowler2014, Bowler2015, Calvi2016, Castellano2010a, Castellano2010b, Finkelstein2015, Ishigaki2018, Livermore2017, McLeod2015, McLeod2016, McLure2009, McLure2013, Oesch2013, Oesch2018, Ouchi2009, Schenker2013, Schmidt2014, Tilvi2013, vanderBurg2010, Willott2013, Zheng2012}: From the dust mass, $M_d$, we obtain the total optical depth to UV continuum photons as \citep[see e.g.][]{Dayal2011} \begin{eqnarray} \tau_\mathrm{UV,c} &=& \frac{3 \Sigma_d}{4as}, \end{eqnarray} with $\Sigma = M_d / (\pi r_d^2)$ being the dust surface mass density, $r_d$ the dust distribution radius, and $a=0.03~\mu$m and $s=2.25$g~cm$^{-3}$ the radius and material density of graphite/carbonaceous grains \citep{todini-ferrara2001}. Since we assume that dust and gas are perfectly mixed, we equate the dust distribution radius, $r_d$, with the radius of the gas, $r_g=4.5 \lambda r_\mathrm{vir} \left[ (1+z)/6 \right]^{1.8}$. Here $\lambda$ is the spin parameter of the simulated halo, $r_\mathrm{vir}$ the virial radius, and the third factor accounts for the redshift evolution of the compactness of galaxies and ensures that the observed UV LFs at $z=5-10$ are well reproduced. For a slab-like geometry, the escape fraction of UV continuum photons of a galaxy is then given by \begin{eqnarray} f_\mathrm{esc}^\mathrm{c} &=& \frac{1-\exp(-\tau_\mathrm{UV,c})}{\tau_\mathrm{UV,c}}, \end{eqnarray} and its observed UV luminosity by \begin{eqnarray} L_\mathrm{c}^\mathrm{obs} &=& f_\mathrm{esc}^\mathrm{c} L_\mathrm{c}, \end{eqnarray} with the intrinsic UV luminosity, $L_\mathrm{c}$, being computed as outlined in Section 2.2.4 in \citet{Hutter2021a}. \subsection{Reionisation} \label{subsec_reionisation} At each time step {\sc astraeus} follows the time and spatial evolution of the ionised regions in the IGM. For this purpose, it derives the number of ionising photons produced in each galaxy, $\dot{Q}$, by convolving the galaxy's star formation rate history with the spectra of a metal-poor ($Z=0.05$Z$_\odot$) stellar population. Spectra have been obtained from the stellar population synthesis model {\sc starburst99} \citep{Leitherer1999}. Again we assume that stars form continuously over a time step. Then the number of ionising photons that contribute to the ionisation of the IGM is then given by \begin{eqnarray} \dot{N}_\mathrm{ion} &=& f_\mathrm{esc} \dot{Q}, \end{eqnarray} where $f_\mathrm{esc}$ is the fraction of ionising photons that escape from the galaxy into the IGM. From the resulting ionising emissivity and gas density distributions {\sc astraeus} derives the spatial distribution of the ionised regions by comparing the cumulative number of ionising photons with the number of absorption events \citep[see {\sc cifog},][, for details]{Hutter2018a}. Within ionised regions, it also derives the photoionisation rate and residual \HI fraction in each grid cell. The ionisation and photoionisation fields obtained allow us then to determine on the fly whether the environment of a galaxy has been reionised and account for the corresponding radiative feedback by computing the gas mass the galaxy can hold on to ($f_g M_g^i$). \subsection{Simulations} \label{subsec_simulations} \begin{figure} \centering \includegraphics[width=0.5\textwidth]{fesc_halomass_relations.png} \caption{The ionising escape fraction $f_\mathrm{esc}$ for the three models, decreasing (solid orange line), being constant (dash dotted magenta line) and increasing (dotted blue line) with halo mass $M_h$.} \label{fig_fesc} \end{figure} \begin{figure} \centering \includegraphics[width=0.5\textwidth]{hist_ion_all_models.png} \caption{Ratio of the mass- and volume-averaged neutral hydrogen fraction (top panel) and volume averaged neutral hydrogen fraction (bottom panel) as a function of redshift. In each panel, we show results for our three $f_\mathrm{esc}$ models: decreasing (solid orange line), being constant (dash dotted magenta line) and increasing (dotted blue line) with halo mass $M_h$. In the lower panel, grey points indicate observational constraints from: GRB optical afterglow spectrum analyses \citep[light triangles;][]{Totani2006, Totani2014}, quasar sightlines \citep[Medium squares;][]{Fan2006}, Lyman-$\alpha$ LFs \citep[dark circles]{Konno2018}, \citep[dark squares;][]{Kashikawa2011}, \citep[dark diamonds][]{Ouchi2010}, \citep[dark pentagons][]{Ota2010} and \citep[dark triangles][]{Malhotra2004}, Lyman-$\alpha$ emitter clustering \citep[dark plus signs;][]{Ouchi2010} and the Lyman-$\alpha$ emitting galaxy fraction \citep[dark crosses;][]{Pentericci2011, Schenker2012, Ono2012, Treu2012, Caruana2012, Caruana2014, Pentericci2014}.} \label{fig_hist_ion} \end{figure} In the following we consider three different reionisation scenarios that explore the physically plausible space of the ionising escape fraction $f_\mathrm{esc}$ (c.f. Fig. \ref{fig_fesc}): \begin{enumerate} \item {\sc mhdec}: $f_\mathrm{esc}$ decreases with rising halo mass of a galaxy (red solid line) \begin{eqnarray} f_\mathrm{esc} &=& f_\mathrm{esc,low} \left( \frac{f_\mathrm{esc,high}}{f_\mathrm{esc,low}} \right)^{\frac{\log_{10} (M_h / M_{h,\mathrm{low}}) }{\log_{10} (M_{h,\mathrm{high}} / M_{h,\mathrm{low}} )}} \label{eq_fesc} \end{eqnarray} with $f_\mathrm{esc,low}=0.55$, $f_\mathrm{esc,high}=0.05$, $M_{h,\mathrm{low}}=2\times10^8h^{-1}\msun$ and $M_{h,\mathrm{high}}=10^{10}h^{-1}\msun$. \item {\sc mhconst}: $f_\mathrm{esc}=0.16$ for each galaxy (magenta dash-dotted line). \item {\sc mhinc}: $f_\mathrm{esc}$ increases with rising halo mass of a galaxy (blue dotted line) following Eqn. \ref{eq_fesc} with $f_\mathrm{esc,low}=0.08$, $f_\mathrm{esc,high}=0.4$, $M_{h,\mathrm{low}}=10^9h^{-1}\msun$ and $M_{h,\mathrm{high}}=10^{11}h^{-1}\msun$. \end{enumerate} These three $f_\mathrm{esc}$ prescriptions have been adjusted to reproduce the electron optical depth measured by Planck \citep{planck2018} and fit the observational constraints from LAEs, quasar absorption spectra and gamma ray bursts (as depicted in the lower panel of Fig. \ref{fig_hist_ion}). In addition, for {\sc mhinc} the maximum $f_\mathrm{esc}$ value of more massive galaxies is also limited by the observed Ly$\alpha$ LFs. Despite having very similar electron optical depths, these three $f_\mathrm{esc}$ prescriptions lead to different ionisation histories and topologies (see Fig. \ref{fig_hist_ion} and \ref{fig_XHImaps_with_LAEs}). As $f_\mathrm{esc}$ decreases with rising halo mass, reionisation is dominated by the low-mass galaxies ($M_h\lesssim10^{10}\msun$), leading to on average smaller ionised regions and lower photoionisation rates. Since these low-mass galaxies appear earlier, reionisation begins earlier (see solid red line in Fig. \ref{fig_hist_ion}); however, as shown in \citet{Hutter2021a} for the {\it Photoionisation} model their overall star formation rate decreases around $z\simeq7$, resulting in the Universe being reionised at a later time and exhibiting a higher average residual \HI fraction in ionised regions. In contrast, as $f_\mathrm{esc}$ increases with rising halo mass, more massive galaxies ($M_h\gtrsim10^{10}\msun$) drive reionisation. On average, ionised regions are larger and more clustered around more massive galaxies, and photoionisation rates within these ionised regions are higher. Reionisation begins later with the appearance of more massive galaxies and ends earlier as the abundance of these massive galaxies increases. \section{Modelling Ly$\alpha$ emitters} \label{sec_modelling_LAEs} In this Section, we introduce the different models for the emergent Ly$\alpha$ line profiles (Section \ref{subsec_Lya_line_models}) and fractions of Ly$\alpha$ radiation escaping from a galaxy (Section \ref{subsec_dust_attenuation}), describe the attenuation of Ly$\alpha$ radiation by \HI in the IGM, and the derivation of the observed Ly$\alpha$ luminosity of a galaxy (Section \ref{subsec_IGM_attenuation}). We summarise the combinations of emerging Ly$\alpha$ line profile and dust attenuation models investigated in this paper in Section \ref{subsubsec_emerging_Lya_profile_models}. \subsection{Emerging Ly$\alpha$ line profiles} \label{subsec_Lya_line_models} We investigate three Ly$\alpha$ line profiles $J(x)$: (1) a thermally Doppler-broadened Gaussian centred at the Ly$\alpha$ resonance; (2) a single, double or triple-peaked profile that depends on the clumpiness and \HI column density of the gas in a galaxy; (3) a single, double or triple-peaked profile that depends both on the ionising escape fraction $f_\mathrm{esc}$ and the clumpiness and \HI column density of the gas in a galaxy. While the first model represents a simple assumption used in previous works \citep[e.g.][]{Dayal2011, Hutter2014}, the latter two models are inspired by observations and detailed Ly$\alpha$ radiative transfer simulations \citep[e.g.][]{Dijkstra2016, Gronke2017}. The Ly$\alpha$ line emerging from a galaxy is given by the intrinsic Ly$\alpha$ luminosity, $L_\alpha^\mathrm{intr}= \frac{2}{3}Q(1-f_\mathrm{esc})~h\nu_\alpha$, the escape fraction Ly$\alpha$ photons from the galaxy, $f_\mathrm{esc}^\mathrm{Ly\alpha}$, and the line profile $J(x)$. \begin{eqnarray} L_\alpha^\mathrm{ISM}(\nu) &=& L_\alpha^\mathrm{intr} f_\mathrm{esc}^\mathrm{Ly\alpha} J(x) \end{eqnarray} In the remainder of this Section, we detail our different models for the Ly$\alpha$ line profiles and escape fractions. \subsubsection{Central Gaussian} \label{subsubsec_Lya_line_model_gaussian} This model assumes that the emission sites of Ly$\alpha$ radiation, the hydrogen atoms within a galaxy, move at velocities that reflect the galaxy's rotation. The corresponding Doppler-broadened Ly$\alpha$ line profile is then given by \begin{eqnarray} J_\mathrm{centre}(x) &=& \frac{1}{\sqrt{\pi}} \frac{\sigma_\mathrm{th}}{\sigma_r} \exp \left[ -x^2 \frac{\sigma_\mathrm{th}^2}{\sigma_r^2} \right], \label{eq_Jcentre} \end{eqnarray} where we have expressed the frequency deviation from the Ly$\alpha$ resonance $\nu_\alpha$ in terms of the thermal line broadening $\sigma_\mathrm{th}=(v_\mathrm{th}/c)\nu_\alpha$ with $v_\mathrm{th}=\sqrt{2k_B T/m_\mathrm{H}}$, yielding $x=\frac{\nu - \nu_\alpha}{\sigma_\mathrm{th}}$. $k_B$ is the Boltzmann constant, $m_\mathrm{H}$ the mass of a hydrogen atom and $T$ the temperature of the \HI gas. Since the $\sigma_\mathrm{th}$-dependence of $x$ cancels any dependency of $J_\mathrm{centre}(\nu)$ on $\sigma_\mathrm{th}$, the assumed gas temperature has no effect on the emerging Ly$\alpha$ line profile (we use $T=10^4$~K in Fig. \ref{fig_profiles}). $\sigma_r\simeq(v_r/c)\nu_\alpha$ describes the Doppler broadening of the line due to the rotation of the galaxy. The rotation velocity of the galaxy $v_r$ is closely linked to the halo rotational velocity $v_c= (3\pi G H_0)^{1/3} \Omega_m^{1/6} (1+z)^{1/2} M_h^{1/3}$, ranging between $v_r=v_c$ and $v_r=2v_c$ \citep{mo1998, cole2000}. We assume $v_r=1.5 v_c$. \subsubsection{Single, double or triple-peaked in a clumpy/homogeneous medium:} \label{subsubsec_Lya_line_model_clumpy} This model describes the Ly$\alpha$ line profile emerging from a clumpy medium. It implements the regimes and characteristic escape frequencies identified in \citet{Gronke2017}. We consider a slab with a thickness of $2B$ and a total optical depth of $2\tau_0$. The source is located at the slab's midplane and injects photons with a frequency $x_i$ close to the Ly$\alpha$ resonance $x=0$. If the slab medium is homogeneous, \citet{neufeld1990} derived the emergent Ly$\alpha$ profile to be \begin{eqnarray} J_\mathrm{slab}(T,\tau_0, x, x_i) &=& 4 \pi \frac{\sqrt{6}}{24} \frac{x^2}{a(T)~ \tau_0} \frac{1}{\cosh\left( \sqrt{\frac{\pi^4}{54}} \frac{|x^3 - x_i^3|}{a(T)~\tau_0}\right)}, \label{eq_Jslab} \end{eqnarray} with $a(T)=\frac{A_\alpha}{4\pi \sigma_\mathrm{th}(T)}$ and $\int_{-\infty}^{\infty} J_\mathrm{slab}(x,x_i)\ \mathrm{d}x = 1$. $A_\alpha$ is the Einstein for the spontaneous emission of Ly$\alpha$ photons. For $x_i=0$, the emerging Ly$\alpha$ spectrum peaks at $x_p= \pm \left(k a \tau_0/\sqrt{\pi} \right)^{1/3}$. However, in our model, we assume the gas in a galaxy not to be static but outflowing at a constant velocity $v$. This translates effectively to a Doppler shift of the injection frequency from $x_i=0$ to $x_i=\frac{v}{v_\mathrm{th}}$. In the following we will revisit the regimes for Ly$\alpha$ escape in a clumpy medium that have been identified in \citet{Gronke2017} and extend them to an "outflowing" slab (or injection frequency $x_i\neq0$). The clumpy medium is characterised by the total optical depth of the clumps and the average number of clumps each Ly$\alpha$ photon escaping the slab scatters with. For a slab consisting of clumps with each having an optical depth $\tau_\mathrm{0,cl}$ at the line centre, Ly$\alpha$ photons escaping the slab will encounter on average $f_c$ clumps and have a total optical depth of $\tau_0=\frac{4}{3} f_c \tau_\mathrm{0,cl}$\footnote{The factor $4/3$ arises from the mean path length through a sphere.} at the line centre. The emerging Ly$\alpha$ line profile depends sensitively on the total and clump optical depth at line centre, $\tau_0$ and $\tau_{0,cl}$, respectively, and the number of clumps the Ly$\alpha$ photons scatter with. \citet{Gronke2017} identified the following regimes: \begin{itemize} \item {\it Free-streaming regime:} The clumpy medium is optically thin ($\tau_\mathrm{0}<1$), and Ly$\alpha$ photons can stream through. The emerging line profile peaks around $x=0$. \item {\it Porous regime:} The clumps are optically thick to Ly$\alpha$ photons ($\tau_\mathrm{0,cl}>1$), but only a fraction $1-\exp(-\tau_{0,cl})$ of the Ly$\alpha$ photons scatter with a clump. The emerging line profile is again peaked around $x=0$. \item {\it Random walk regime:} The clumps are optically thick to Ly$\alpha$ ($\tau_\mathrm{0,cl}>1$), and each Ly$\alpha$ photon encounters $N_\mathrm{sct,rw}\propto f_c^2$ scattering events \citep{Hansen_Oh2006}. However, the number of scattering events is too low for the Ly$\alpha$ photons to scatter in frequency space far enough into the wings to escape through excursion. Hence, the emerging line profile peaks also around $x=0$. \item {\it Homogeneous regime:} The clumps are optically thin ($\tau_\mathrm{0,cl}\leq1$) and Ly$\alpha$ photons scatter $\sim\tau_0$ times ($N_\mathrm{sct,exc}\propto f_c$) and escape via excursion: they follow a random walk in space {\it and} frequency and escape as they are scattered into the wings where the clumps become optically thin. The emerging line profile is a double-peaked with the two peaks being located at $x_p \simeq \left(k a \tau_0\right/\sqrt{\pi})^{1/3}$ for an injection frequency $x=0$ \citep{Adams1975}, $x_{p,1} \simeq \pm \left[ \left(\frac{k a \tau_0}{\sqrt{\pi}} \right)^{k} + x_i^{3k} \right]^{1/3k}$ and $x_{p,2} \simeq \pm \left(\frac{k a \tau_0}{\sqrt{\pi}}\right)^{1/3} \left[ 1 - \frac{k}{2\pi} \frac{x_i^3}{x_i^3 + k a \tau_0/\sqrt{\pi}} \right]$ for an injection frequency $x=\pm x_i$ and $x_i\geq0$.\footnote{The expressions for $x_{p,1}$ and $x_{p,2}$ have been identified as sufficient numerical fits to different $\tau_0$ and $x_i$ values.} \end{itemize} \citet{Gronke2017} derived the boundary criteria between these regimes for a static clumpy medium. We advance these criteria to a clumpy slab moving with a constant velocity $v$ from the source (mimicking outflows) or an injection frequency $x \neq 0$. To derive the critical number of clumps, $4/3 f_c$, required for each regime, we first consider the time and distances covered that it takes a Ly$\alpha$ to transverse the slab. {\it Excursion:} As Ly$\alpha$ photons transverse or escape the slab, they scatter with \HI many times. This alters their direction and frequency $x$, and they essentially perform a random walk. However, as the Ly$\alpha$ cross section is higher close to the line centre, most scatterings will occur close to the line centre and remain spatially close. Only as the Ly$\alpha$ photons are scattered into the wings of the Ly$\alpha$ absorption profile their mean free paths become larger, allowing them to escape the slab \citep{Adams1975}. The series of these so-called wing scatterings that allow Ly$\alpha$ photons to escape are referred to as excursion. We can estimate the mean displacement and time spent in such an excursion event: a Ly$\alpha$ photon with frequency $x$ will scatter on average $N_\mathrm{sct,exc}\sim x^2$ times before it has traversed a slab of thickness $B$. Its average mean free path is then $\lambda_\mathrm{mfp,exc}(x) = B \sigma_0/ (k\tau_0\sigma_\mathrm{HI}(x)) = B / (k \tau_0 H(a,x))$ using the wing approximation of the Ly$\alpha$ cross section and $k$ being a geometrical factor determined in \citet{Adams1975} and amounting to $\sqrt{3}$. This and the random walk nature of the Ly$\alpha$ escape imply an average displacement of \begin{eqnarray} d_\mathrm{exc} &=& \sqrt{N_\mathrm{sct,exc}} \lambda_\mathrm{mfp,exc}(x) = \frac{\sqrt{N_\mathrm{sct,exc}}B}{k\tau_0 H_v(a,x)} = \frac{x B}{k\tau_0 H_v(a,x)} \nonumber \\ \end{eqnarray} and time spent in the excursion of \begin{eqnarray} t_\mathrm{exc} &=& N_\mathrm{sct,exc} \frac{\lambda_\mathrm{mfp,exc}(x)}{c} = \frac{N_\mathrm{sct,exc} B}{c k\tau_0 H_v(a,x)} = \frac{x^2 B}{c k\tau_0 H_v(a,x)}, \nonumber \\ \end{eqnarray} with $H_v(a,x)=\frac{a}{\sqrt{\pi}(x)^2}$ being the effective line absorption profile in the wings. {\it Random Walk:} As the clumps become optically thick, the Ly$\alpha$ photons do not escape the slab via excursion anymore but by random walking: the number of scattering events is smaller than required for excursion and scale with the square of the number of clumps, $N_\mathrm{sct,rw}\propto f_c^2$ \citep{Hansen_Oh2006}. With the mean free path given by the average clump separation $\lambda_\mathrm{mfp,rw}=kB/f_c$, the average displacement and time are then \begin{eqnarray} d_\mathrm{rw} &=& \sqrt{N_\mathrm{sct,rw}} \lambda_\mathrm{mfp,rw} = \frac{\sqrt{N_\mathrm{sct,rw}} k B}{f_c} = k B \end{eqnarray} and \begin{eqnarray} t_\mathrm{rw} &=& \sqrt{N_\mathrm{sct,rw}} \frac{\lambda_\mathrm{mfp,rw}}{c} = \frac{\sqrt{N_\mathrm{sct,rw}} k B}{c f_c} = \frac{k B f_c}{c}. \end{eqnarray} \begin{enumerate} \item {\it Division between random-walk and homogeneous regime in optically thick medium:} For a given total optical depth at the line centre, $\tau_0$, we can derive the critical number of clumps along a line of sight that marks the transition from the random (clumps are optically thick) to the homogeneous regime (clumps become optically thick at the excursion frequency). We estimate this transition to arise when both regimes contribute equally to the flux of escaping Ly$\alpha$ photons. \begin{eqnarray} \frac{F_\mathrm{rw}}{F_\mathrm{exc}} &=& \frac{t_\mathrm{exc}}{t_\mathrm{rw}} = \frac{x^2}{k^2 \tau_0 H(x) f_c} = \frac{\sqrt{\pi} x^4}{k^2 a \tau_0 f_c} = 1 \label{eq_FrwDivFexc} \end{eqnarray} With $\tau_0=4/3 f_c \tau_\mathrm{0,cl}$, the critical number of clumps for Ly$\alpha$ photons escaping at frequency $x$ yields then as \begin{eqnarray} f_{c,\mathrm{crit}} &=& \frac{\sqrt{3}\pi^{1/4}}{2k} \frac{x^2}{\sqrt{a \tau_\mathrm{0,cl}}}. \end{eqnarray} As long as the wings remain optically thick, the majority of Ly$\alpha$ photons (with injection frequency $x_i$) will escape at \begin{eqnarray} x_\mathrm{esc} &\simeq& \begin{cases} \left( \frac{k a \tau_0}{\sqrt{\pi}} \right)^{1/3} & \mathrm{for}\ x_i \leq \left(\frac{k a \tau_0}{\sqrt{\pi}} \right)^{1/3}\\ x_i & \mathrm{otherwise},\\ \end{cases} \label{eq_xesc} \end{eqnarray} leading to \begin{eqnarray} f_{c\mathrm{,crit}} &=& \begin{cases} \frac{2}{\sqrt{3}k\pi^{1/4}} \sqrt{a \tau_\mathrm{0,cl}} & \mathrm{for}\ x_i^2 \leq \frac{4 a \tau_\mathrm{0,cl}}{3 \sqrt{\pi}}\\ \frac{\sqrt{3}\pi^{1/4}}{2k} \frac{x_i^2}{\sqrt{a \tau_\mathrm{0,cl}}} & \mathrm{otherwise}.\\ \end{cases} \label{eq_fccrit} \end{eqnarray} This $f_{c\mathrm{,crit}}$ value marks the transition from the random walk to the excursion regime. Its value depends strongly on the typical escape frequency of Ly$\alpha$ photons, $x_\mathrm{esc}$, which again determines the dependency of $f_{c\mathrm{,crit}}$ on the clump optical depth $\tau_\mathrm{0,cl}$ and injection frequency $x_i$. We can understand these dependencies as follows: For injection frequencies close to $x=0$, $x_i^2 \leq \frac{4 a \tau_\mathrm{0,cl}}{3 \sqrt{\pi}}$, optically thicker clumps require a higher escaping frequency $x_\mathrm{esc}$ and thus a higher total optical depth. This can only be achieved by interacting with more clumps (higher $f_{c,\mathrm{crit}}$). However, for $x_i^2 > \frac{4 a \tau_\mathrm{0,cl}}{3 \sqrt{\pi}}$, raising the injection frequency causes the clumps to be optically thinner at the escape frequency. Consequently, Ly$\alpha$ photons need to scatter more often (higher $f_{c,\mathrm{crit}}$) to reach the escape frequency. However, this rise in the number of scattering events is "normalised" to the escape frequency of the static solution ($x_i=0$), i.e. the larger the optical depth is, the smaller is the difference between $x_\mathrm{esc}=x_i$ and $x_\mathrm{esc}=\left(k a \tau_0/\sqrt{\pi} \right)^{1/3}$, and hence the smaller is the number of "extra" scattering events required. Because the transition described by $f_{c,\mathrm{crit}}$ is not sharp, we model the Ly$\alpha$ line profile emerging from the moving slab by superposing the Ly$\alpha$ radiation escaping in the homogeneous (using $J_\mathrm{slab}$ in Eqn. \ref{eq_Jslab}) and random walk regimes (using $J_\mathrm{centre}$ in Eqn. \ref{eq_Jcentre} and assuming $\sigma_r = \sigma_\mathrm{th}$). \begin{eqnarray} J_\mathrm{rh}(\tau_0, x, x_i) &=& (1 - f_\mathrm{rw}) J_\mathrm{slab}(T, \tau_0, x, x_i) + f_\mathrm{rw} J_\mathrm{centre}(T, x) \label{eq_Jrh} \nonumber \\ \end{eqnarray} We derive the corresponding ratio $f_\mathrm{fw}$ by assuming that the Ly$\alpha$ flux escapes predominantly where the Ly$\alpha$ profiles peak, \begin{eqnarray} f_\mathrm{rw} &=& \frac{F_\mathrm{rw} / F_\mathrm{exc}}{F_\mathrm{rw} / F_\mathrm{exc} + \frac{J_\mathrm{centre}(0)}{2(J_\mathrm{slab}(\tau_0,x_{p,1},x_i) + J_\mathrm{slab}(\tau_0,x_{p,2},x_i))}}. \label{eq_frh} \end{eqnarray} We note that this description reproduces the Ly$\alpha$ line profiles for resting clumps, fixed $\tau$ values, and varying $f_c$ values in \citet{Gronke2017}. \item {\it Division between porous and homogeneous regime in optically thin medium:} As the medium becomes optically thinner, Ly$\alpha$ photons that scatter into the wings can escape the slab before completing their excursion, i.e. $x_\mathrm{esc}=x_\mathrm{max}=\max(x_\star, x_i)$. Here the transition from an optically thin medium to an optically thick medium is described by $k a \tau_0=\sqrt{\pi} x_\mathrm{max}^3$,\footnote{While the wings become optically thin for $k\tau(x_\mathrm{max})\geq1$ corresponding to $k a \tau_0=\sqrt{\pi} x_\mathrm{max}^2$, we adapt this criterion to be consistent with a continuous function for $x_\mathrm{esc}$.} with $x_\star$ being the frequency where the Ly$\alpha$ absorption profile transitions from the Gaussian core to the Lorentzian wings. While the slab is optically thin at $x_\mathrm{max}$, depending on whether the clumps are optically thin or thick at $x=0$, the escape of Ly$\alpha$ photons is described by the homogeneous and porous regime, respectively. Again we estimate the transition to arise when both regimes contribute equally to the flux of escaping Ly$\alpha$ photons. We note that if clumps are optically thin at line centre ($\tau_\mathrm{0,cl}<1$), not every clump encounter leads to a scattering event; this reduces the time to escape $t_\mathrm{exc}$ by a factor $1-e^{-\tau_\mathrm{0,cl}}$. \begin{eqnarray} \frac{F_\mathrm{por}}{F_\mathrm{hom}} &=& \frac{t_\mathrm{exc}}{t_\mathrm{rw}~(1-e^{-\tau_\mathrm{0,cl}})} = \frac{\sqrt{\pi} x_\mathrm{max}^4}{k^2 a \tau_0 f_c (1-e^{-\tau_\mathrm{0,cl}})} = 1 \end{eqnarray} We yield the critical number of clumps that mark the transition from the porous to the homogeneous regime as \begin{eqnarray} f_{c\mathrm{,crit}} &=& \frac{x_\mathrm{max}}{k (1-e^{-\tau_\mathrm{0,cl}})}. \end{eqnarray} The emerging Ly$\alpha$ line profile accounts again for Ly$\alpha$ photons escaping in homogeneous ($J_\mathrm{slab}$, see Eqn. \ref{eq_Jslab}) and porous regime ($J_\mathrm{centre}$, see Eqn. \ref{eq_Jcentre}). \begin{eqnarray} J_\mathrm{ph}(\tau_0, x, x_i) &=& (1 - f_\mathrm{por}) J_\mathrm{slab}(T, \tau_0, x, x_\mathrm{max}) + f_\mathrm{por} J_\mathrm{centre}(T, x) \label{eq_Jph} \nonumber \\ \end{eqnarray} However, to ensure that $J_\mathrm{slab}$ peaks at $x_\mathrm{max}$, we use $J_\mathrm{slab}(T, \tau_0, x, x_i)$ as written in Eqn. \ref{eq_Jslab} for $|x_i|>x_\star$ and replace $a\tau_0$ by $\sqrt{\pi}x_\star^3$ in Eqn. \ref{eq_Jslab} otherwise. The ratio between the two different escape regimes is then again given by assuming that most Ly$\alpha$ photons escape at the peak frequencies, \begin{eqnarray} f_\mathrm{por} &=& \frac{F_\mathrm{por} / F_\mathrm{hom}}{F_\mathrm{por} / F_\mathrm{hom} + \frac{J_\mathrm{centre}(0)}{2(J_\mathrm{slab}(\tau_0,x_{p,1},x_\mathrm{max}) + J_\mathrm{slab}(\tau_0,x_{p,2},x_\mathrm{max}))}}. \nonumber \\ \label{eq_fph} \end{eqnarray} \end{enumerate} To derive the Ly$\alpha$ line profile emerging from a simulated galaxy, we obtain the injection frequency $x_i$ and optical depth at the Ly$\alpha$ line centre $\tau_0$ from the galaxy's initial gas mass $M_g^i$ and SN energy $E_\mathrm{SN}$ as follows: \paragraph*{Outflow velocity:} \label{par_outflow_velocity} The injection frequency $x_i$ is the velocity $v$ of the outflowing gas in terms of the thermal velocity $v_\mathrm{th}$. We derive the outflow velocity $v$ of the gas from the SN energy injected into the gas residing in the galaxy with a total gas mass $M_g$ as \begin{eqnarray} v &=& \left( \frac{2 E_\mathrm{SN}}{M_g} \right)^{1/2} = \left( \frac{2 M_g^\mathrm{ej} v_c^2}{M_g^\mathrm{i}} \right)^{1/2} \simeq v_c \left( \frac{2 f_\star^\mathrm{eff}}{f_\star^\mathrm{ej}} \right)^{1/2}\\ v_c &=& \sqrt{\frac{G M_\mathrm{vir}}{r_\mathrm{vir}}}, \end{eqnarray} with $f_\star^\mathrm{eff}$ being the effective star formation efficiency, and $f_\star^\mathrm{ej}$ the star formation efficiency required to eject all gas as defined for the delayed SN feedback scheme in \citet{Hutter2021a}. This results in outflow velocities of $60$km/s and $143$km/s for $10^9\msun$ and $10^{11}\msun$ halos, respectively. We note that $v$ is linked to the escape velocity $v_e$ of the halo as $\frac{v}{v_e} = \frac{v}{\sqrt{2}v_c} = \sqrt{\frac{M_g^\mathrm{i}}{M_g^\mathrm{ej}}}$. \paragraph*{Optical depth and number of clumps encountered:} The optical depth at the Ly$\alpha$ line centre yields as \begin{eqnarray} \tau_0 &=& \frac{4}{3} f_c \tau_\mathrm{0,cl} = N_\mathrm{HI} \sigma_\mathrm{HI}. \end{eqnarray} $\tau_\mathrm{0,cl}$ is a free parameter in our model and reflects the optical depth of a cloud in the ISM. Thus we can estimate it from the median mass ($M_\mathrm{cl}$) and size ($r_\mathrm{cl}$) of molecular clouds ($M_\mathrm{cl}\simeq 10^5\msun$ and $r_\mathrm{cl}=20$~pc) as \begin{eqnarray} \tau_\mathrm{0,cl} &=& \sigma_\mathrm{HI} \frac{M_\mathrm{cl}}{r_\mathrm{cl}}. \end{eqnarray} \paragraph*{\HI column density:} \label{par_HI_column_density} We derive the neutral hydrogen column density $N_\mathrm{HI}$ from the initial gas mass, $M_g^\mathrm{i}$ as \begin{eqnarray} N_\mathrm{HI} &=& \xi \frac{3 M_\mathrm{HI}}{4\pi r_g^2 m_\mathrm{H}} = \xi \frac{3 X_c (1-Y) M_g^\mathrm{i}}{4\pi\ (4.5\lambda r_\mathrm{vir})^2 m_\mathrm{H}} = \xi \frac{3 f_m M_\mathrm{vir}}{81\pi \lambda^2 r_\mathrm{vir}^2 m_\mathrm{H}} \nonumber \\ &=& \xi \left( \frac{9\pi^2 H_0^2 \Omega_m}{G} \right)^{2/3} (1+z)^2 M_\mathrm{vir}^{1/3} \frac{3 f_m}{81\pi \lambda^2 m_\mathrm{H}}. \end{eqnarray} Here $r_g$ describes the gas radius, for which we assume $r_g= 4.5 \lambda r_\mathrm{rvir}$. $X_c$ and $Y$ are the cold gas and helium mass fractions, respectively. Gas accretion and SN feedback processes determine the relation between the initial gas mass and the halo mass, $f_m$, which ranges typically between $\sim10^{-3}$ for low-mass galaxies to $\sim 10^{-1}$ for more massive galaxies. $\xi$ is a geometrical correction factor that depends on $\tau_0$ and the dust optical depth at the Ly$\alpha$ resonance $\tau_d$. Its maximum values is $0.35$ and we describe its derivation and dependencies in Appendix \ref{app_correction_factor}. For the cosmological parameters assumed in this paper, we yield \begin{eqnarray} N_\mathrm{HI} = 6.5\times 10^{17} \mathrm{cm}^2\ (1+z)^2 \frac{\xi f_m}{\lambda^2} \left( \frac{M_\mathrm{vir}}{10^8 M_\odot} \right)^{1/3}. \end{eqnarray} \subsubsection{Ionising escape fraction dependent in a clumpy/homogeneous medium} \label{subsubsec_Lya_line_model_porous} For this Ly$\alpha$ line profile model, we assume a model similar to the so-called picket fence model \citep{Heckman2011}. Here a fraction $f_\mathrm{esc}$ of the ionising radiation escapes through low-density channels, while the other fraction of ionising photons is absorbed by the dense shell. Correspondingly, the Ly$\alpha$ photons escaping through the channels scatter only a few times, while those escaping through the shell encounter many scattering events. The former gives rise to a single-peaked Ly$\alpha$ line centred around $x=0$, while the latter creates a broader double-peaked Ly$\alpha$ line (assuming a homogeneous slab model with peaks at $x_p$). In detail, we assume that Ly$\alpha$ photons escaping through the low-density channels encounter a total optical depth $\tau_\mathrm{channel}$, leading to the fraction of ionising photons escaping through these channels being given by \begin{eqnarray} f_\mathrm{esc}^\mathrm{channel} &=& \exp\left( - \tau_\mathrm{channel}^\mathrm{LyC} \right) = \exp\left ( - \tau_\mathrm{channel} \frac{\sigma_\mathrm{HI}^\mathrm{LyC}}{\sigma_\mathrm{HI}} \right). \end{eqnarray} Here $\sigma_\mathrm{HI}$ and $\sigma_\mathrm{HI}^\mathrm{LyC}$ are the absorption cross sections of neutral hydrogen for Ly$\alpha$ and ionising photons, respectively. As $f_\mathrm{esc}^\mathrm{channel}$ is very likely to be lower than unity, the fraction of the solid angle covered by channels \begin{eqnarray} f_\mathrm{channel} &=& \frac{f_\mathrm{esc}}{f_\mathrm{esc}^\mathrm{channel}} \end{eqnarray} will exceed $f_\mathrm{esc}$. From the Ly$\alpha$ channel optical depth, $\tau_\mathrm{channel}$, and the gas mass divided into gas located in channels and shells, we derive the optical depth of Ly$\alpha$ photons traversing the shell as \begin{eqnarray} \tau_\mathrm{shell} &=& \frac{\tau_0 - f_\mathrm{channel} \tau_\mathrm{channel}}{1 - f_\mathrm{channel}}, \label{eq_tau_shell} \end{eqnarray} with the total optical depth $\tau_0$ being derived as described in Section \ref{par_HI_column_density}. We assume the gas in the shell to be outflowing while the gas in the channels is at rest.\footnote{To compute the outflow velocity $x_i$, we use the entire gas mass instead of the gas sitting in the shell. Thus, if gas resides in channels, we underestimate $x_i$ slightly.} For a given clump optical depth, $\tau_\mathrm{0,cl}$, we determine the Ly$\alpha$ escape regime for the dense shell and channels, and derive the corresponding fractions of Ly$\alpha$ radiation that escapes without significant scattering, $f_\mathrm{shell}$ and $f_\mathrm{channel}$ respectively, using Eqns \ref{eq_frh} and \ref{eq_fph}. The Ly$\alpha$ line emerging from the galaxy contains Ly$\alpha$ photons that traverse the dense shell \begin{eqnarray} J_\mathrm{shell}(T, \tau_\mathrm{shell}, x, x_i) &=& f_\mathrm{shell} J_\mathrm{centre}(T, x) \\ && + \left( 1-f_\mathrm{shell} \right) J_\mathrm{slab}(T, \tau_\mathrm{shell}, x, x_i) \nonumber \end{eqnarray} and escape through the channels \begin{eqnarray} J_\mathrm{channel}(T, \tau_\mathrm{channel}, x, x_i) &=& f_\mathrm{channel} J_\mathrm{centre}(T, x) \\ && + \left( 1-f_\mathrm{channel} \right) J_\mathrm{slab}(T, \tau_\mathrm{channel}, x, x_i). \nonumber \end{eqnarray} It is given by \begin{eqnarray} J(\tau, x, x_i) &=& f_\mathrm{channel}\ J_\mathrm{channel}(T, \tau_\mathrm{channel}, x, x_i) \\ && +\ \left(1 - f_\mathrm{channel}\right)\ J_\mathrm{shell}(T, \tau_\mathrm{shell}, x, x_i), \nonumber \end{eqnarray} with the outflow velocity $x_i$ being derived as outlined in Section \ref{par_outflow_velocity}. However, in the following, we consider the extreme case of $\tau_\mathrm{channel}=0$ (resulting in $f_\mathrm{channel}=1$). \subsection{Dust attenuation} \label{subsec_dust_attenuation} We employ two different dust models. The first one links the Ly$\alpha$ escape fraction to the escape fraction of UV continuum photons, $f_\mathrm{esc}^\mathrm{c}$. The second one is more complex. It assumes a clumpy medium where the attenuation of Ly$\alpha$ by dust follows different relations in the regimes identified in \citet{Gronke2017}. Both models assume a slab-like geometry and we describe their details in the following. \subsubsection{Simple attenuation model} \label{subsubsec_dust_simple_model} In this model, we assume that (i) dust and gas are perfectly mixed, (ii) the dust distribution is slab-like, and (iii) the dust attenuation of Ly$\alpha$ photons is proportional to the dust attenuation of UV continuum photons. The escape fraction of Ly$\alpha$ photons, $f_\mathrm{esc}^\mathrm{Ly\alpha}$, is then directly related to the escape of UV continuum photons, $f_\mathrm{esc}^\mathrm{c}$, derived in Section \ref{subsubsec_metals_and_dust}. \begin{eqnarray} f_\mathrm{esc}^\mathrm{Ly\alpha} &=& p\ f_\mathrm{esc}^\mathrm{c} \end{eqnarray} We use $p$ as a free parameter to obtain the observed Ly$\alpha$ luminosity functions at $z=6.6-7.3$. \subsubsection{Refined attenuation model} \label{subsubsec_dust_refined_model} This model assumes that dust and gas are perfectly mixed and distributed in clumps. The dust attenuation of Ly$\alpha$ photons depends on the total optical depth of the dust, $\tau_\mathrm{d,total}$, the optical depth of a clump, $\tau_\mathrm{d,cl}$, and the number of clumps, $f_c$, encountered along the sightline from the midplane to the surface of the slab. We derive its value by estimating the dust absorption cross section. Following \citet{Galliano2022} and assuming the radius and density of graphite/carbonaceous grains (see Section \ref{subsubsec_metals_and_dust}), we assume $\kappa_\mathrm{abs} \simeq \frac{Q_\mathrm{abs}}{a s} \simeq 2\times10^5$~cm$^{2}/$g with $Q_\mathrm{abs}\simeq1$ being the absorption efficiency. \footnote{We note that this is in rough agreement with the dust extinction cross sections of the Small and Large Magellanic clouds $\kappa_\mathrm{ext} = \sigma_\mathrm{d} / m_\mathrm{H} = \sigma_\mathrm{d,ref} \frac{M_z}{M_\mathrm{d}\ Z_\mathrm{ref} m_\mathrm{H}} \simeq 4\times 10^5$~cm$^{2}/$g, with the extinction efficiency $Q_\mathrm{ext}=Q_\mathrm{abs}+Q_\mathrm{sca}$ being given by the similar sized absorption ($Q_\mathrm{abs}$) and scattering efficiencies ($Q_\mathrm{sca}$) at Ly$\alpha$, a dust-to-metal mass ratio $M_\mathrm{d}/M_Z\simeq0.25$, $\sigma_\mathrm{ref}\simeq4\times10^{-22}$cm$^2$ and $Z_\mathrm{ref}\simeq0.0025$ for SMC and $\sigma_\mathrm{ref}\simeq7\times10^{-22}$cm$^2$ and $Z_\mathrm{ref}\simeq0.005$ for LMC \citep[for further explanations see][]{Laursen2010}.} \begin{eqnarray} \tau_\mathrm{d,total} &=& \frac{4}{3} f_c \tau_\mathrm{d,cl} = \xi\ \frac{3}{4\pi} \frac{M_\mathrm{d}}{r_\mathrm{d}^2} \kappa = \frac{M_\mathrm{d}}{M_\mathrm{HI}} \frac{\kappa_\mathrm{abs} m_\mathrm{H}}{\sigma_\mathrm{HI}} \tau_0 \end{eqnarray} The resulting estimates for $\tau_\mathrm{d,total}$ and $\tau_\mathrm{d,cl}$ allow us to compute the Ly$\alpha$ escape fractions in the different escape regimes as follows. \paragraph{Free-streaming regime:} \label{par_dust_freestreaming} In an optically thin slab ($\tau_0<1$), the Ly$\alpha$ photons stream through $\sim f_c$ clumps. On their way, they are attenuated by the dust in clumps and hence, the total dust optical depth determines the Ly$\alpha$ escape fraction, $\tau_\mathrm{d,total}$, as \begin{eqnarray} f_\mathrm{esc}^\mathrm{Ly\alpha, fs} &=& \exp\left(-\tau_\mathrm{d, total}\right) = \exp\left(-\frac{4}{3} f_c \tau_\mathrm{d,cl}\right) \end{eqnarray} We note that in this regime, the number of clumps along the sightline $f_c$ and clump optical depth $\tau_\mathrm{0,cl}$ are degenerate. \paragraph{Random walk regime:} \label{par_dust_randomwalk} In the random walk regime, both the slab and individual clumps are optically thick ($\tau_\mathrm{0,cl}\geq1$). As a result, Ly$\alpha$ photons escape by mostly being scattered by the clumps, and their escape fraction is determined by the number of clumps encountered along their random walk, $N_\mathrm{cl}(f_c)$, and the absorption probability per clump interaction $\epsilon$. According to \citet{Hansen_Oh2006}, it is then given by \begin{eqnarray} f_\mathrm{esc}^\mathrm{Ly\alpha, rw} &=& f_\mathrm{HO06} = \frac{1}{\cosh(\sqrt{2 N_\mathrm{cl}(f_c)\ \epsilon})} \end{eqnarray} We assume $N_\mathrm{cl}(f_c)\simeq \frac{3}{2}f_c^2 + 2 f_c$ as found in \citet{Gronke2017}. The scaling of $N_c$ with $f_c$ also agrees with the findings in \citet{Hansen_Oh2006} and prefactors vary slightly due to different geometries of the scattering surface. However, since $\epsilon$ is sensitive to how deep the photons permeate the clump, it depends non-trivially on the clump optical depth and movement. For simplicity, we assume $\epsilon=1-\exp(-\tau_\mathrm{d,cl})$, which has been shown to be in agreement with numerical simulations for small $\tau_\mathrm{d,cl}$ values \citep{Gronke2017}. We note that in this regime the Ly$\alpha$ escape fraction depends solely on the number of clumps encountered $f_c$, and hence $f_c$ and $\tau_\mathrm{d,cl}$ are not degenerate as in the free-streaming or homogeneous regime. \paragraph{Homogeneous regime:} \label{par_dust_homogeneous} In the homogeneous regime, the slab is optically thick ($\tau_0\geq1$), while the individual clumps are optically thin ($\tau_\mathrm{0,cl}<0$). During their initial random walk, the Ly$\alpha$ photons scatter with $N_\mathrm{cl}(f_{c,\mathrm{crit}})$ clumps before they diffuse into the wings and escape by free-stream through $f_c$ clumps. The resulting Ly$\alpha$ escape fraction, \begin{eqnarray} f_\mathrm{esc}^\mathrm{Ly\alpha, hom} &=& f_\mathrm{HO06}(f_{c,\mathrm{crit}}) \exp(-\tau_\mathrm{d, total}), \end{eqnarray} depends on $f_{c,\mathrm{crit}}$ and $\tau_\mathrm{d,total}$, with $f_{c,\mathrm{crit}}$ being determined by $\tau_0$ and $\tau_\mathrm{0,cl}$. \begin{eqnarray} f_{c,\mathrm{crit}} = \begin{cases} \frac{2}{\sqrt{3}k \pi^{1/4}} \sqrt{a \tau_\mathrm{0,cl}} & \mathrm{for}\ ka\tau_0 \geq \sqrt{\pi} x_\mathrm{max}^3\ \mathrm{and}\ x_i^2\leq \frac{4 a \tau_\mathrm{0,cl}}{3\sqrt{\pi}} \\ \frac{\sqrt{3} \pi^{1/4}}{2k} \frac{x_i^2}{\sqrt{a \tau_\mathrm{0,cl}}} & \mathrm{for}\ ka\tau_0 \geq \sqrt{\pi} x_\mathrm{max}^3\ \mathrm{and}\ x_i^2 > \frac{4 a \tau_\mathrm{0,cl}}{3\sqrt{\pi}} \\ \frac{x_\mathrm{max}}{k \left(1-e^{-\tau_\mathrm{0,cl}}\right)} & \mathrm{for}\ ka\tau_0 < \sqrt{\pi} x_\mathrm{max}^3 \end{cases} \end{eqnarray} \paragraph{Porous regime:} \label{par_dust_porous} In the porous regime, the individual clumps are optically thick ($\tau_\mathrm{0,cl}\geq1$), but only a fraction $1-e^{-f_c}$ of the Ly$\alpha$ photons will encounter a clump along their sightlines. The other fraction of Ly$\alpha$ photons does not interact with any clumps and is thus not attenuated by dust as they escape the slab.\footnote{We note that our expression is here a lower limit of $f_\mathrm{esc}^\mathrm{Ly\alpha, por}$ as we assume the Ly$\alpha$ radiation interacting with clumps to experience attenuation as if they streamed through the clump. It might be more appropriate to consider these Ly$\alpha$ photons to be absorbed as in the random walk regime, $f_\mathrm{esc}^\mathrm{Ly\alpha, por} = e^{-f_c} + \left[1 - e^{-f_c}\right] \frac{1}{\cosh(\sqrt{2 N_\mathrm{cl}(f_c)\ \epsilon})}$, however in practise galaxies in the porous regime have not much, if any, dust.} \begin{eqnarray} f_\mathrm{esc}^\mathrm{Ly\alpha, por} &=& e^{-f_c} + \left[1 - e^{-f_c}\right] \exp\left(-\frac{4}{3} f_c \tau_\mathrm{d,cl}\right) \end{eqnarray} \subsubsection{Emerging Ly$\alpha$ line profile models} \label{subsubsec_emerging_Lya_profile_models} We briefly summarize the combinations of Ly$\alpha$ line and dust attenuation models that we will investigate in this paper. \paragraph*{Gaussian:} The Ly$\alpha$ line profile emerging from a galaxy is given by the central Gaussian Ly$\alpha$ line profile (Section \ref{subsubsec_Lya_line_model_gaussian}). To account for the attenuation by dust, we apply the Ly$\alpha$ escape fraction, $f_\mathrm{esc}^\mathrm{Ly\alpha}$, derived in our simple dust model (Section \ref{subsubsec_dust_simple_model}) to all frequencies $x$. \paragraph*{Clumpy:} This model assumes an outflowing shell of dusty gas clumps, whereas gas and dust are perfectly mixed. It combines the Ly$\alpha$ line model described in Section \ref{subsubsec_Lya_line_model_clumpy} with the refined dust model depicted in Section \ref{subsubsec_dust_refined_model}. The gas in the galaxies is assumed to have a temperature of $T=10^4$~K.\footnote{We have chosen $T=10^4$~K for simiplicity. If we were to assume the virial temperature ($T_\mathrm{vir}$), the double-peak line profile would narrow as $T_\mathrm{vir}$ increases.} In contrast to the {\it Gaussian} model, we dust-attenuate the Ly$\alpha$ line of each escape regime (homogeneous, random, porous) by its corresponding escape fraction $f_\mathrm{esc}^\mathrm{Ly\alpha}$. The emerging Ly$\alpha$ line profile is then the superposition of the line profiles of all relevant escape regimes, \begin{equation} L_\alpha^\mathrm{ISM}(x) = f_\mathrm{esc, slab}^\mathrm{Ly\alpha} (1-f) J_\mathrm{slab}(x) + f_\mathrm{esc, centre}^\mathrm{Ly\alpha} f J_\mathrm{centre}(x), \end{equation} with $f_\mathrm{esc, slab}^\mathrm{Ly\alpha}=f_\mathrm{esc, hom}^\mathrm{Ly\alpha}$, and $(f, f_\mathrm{esc, centre}^\mathrm{Ly\alpha})$ given by $(1, f_\mathrm{esc, fs}^\mathrm{Ly\alpha})$, $(f_\mathrm{rw}, f_\mathrm{esc, rw}^\mathrm{Ly\alpha})$ or $(f_\mathrm{por}, f_\mathrm{esc, por}^\mathrm{Ly\alpha})$ depending on the total and clump optical depths $\tau_0$ and $\tau_\mathrm{0,cl}$. \paragraph*{Porous:} This model is very similar to the {\it Clumpy} model. However, it considers the outflowing shell of clumps to be pierced with gas and dust-free channels through which a fraction $f_\mathrm{esc}$ of the Ly$\alpha$ photons escape without scattering. It combines the Ly$\alpha$ line model described in Section \ref{subsubsec_Lya_line_model_porous} and assuming $\tau_\mathrm{channel}^\mathrm{LyC}=0$ with the refined dust model depicted in Section \ref{subsubsec_dust_refined_model}. Again we assume the gas in the galaxy to be heated to a temperature of $T=10^4$~K, and the Ly$\alpha$ line of each escape regime (homogeneous, random, porous) to be dust-attenuated by its corresponding escape fraction $f_\mathrm{esc}^\mathrm{Ly\alpha}$. The emerging Ly$\alpha$ line profile is again a superposition of the Ly$\alpha$ photons escaping through the channels and the clumpy shell, \begin{eqnarray} L_\alpha^\mathrm{ISM}(x) &=& f_\mathrm{esc}~ J_\mathrm{channel}(x)\ +\ (1- f_\mathrm{esc})~ J_\mathrm{shell}(x) \\ J_\mathrm{channel}(x) &=& J_\mathrm{centre}(x)\\ J_\mathrm{shell}(x) &=& f_\mathrm{esc, slab}^\mathrm{Ly\alpha}~ (1-f)~ J_\mathrm{slab}(x) \ +\ f_\mathrm{esc, centre}^\mathrm{Ly\alpha}~ f~ J_\mathrm{centre}(x) \nonumber \\ \end{eqnarray} with $f_\mathrm{esc, slab}^\mathrm{Ly\alpha}=f_\mathrm{esc, hom}^\mathrm{Ly\alpha}$, and $(f, f_\mathrm{esc, centre}^\mathrm{Ly\alpha})$ given by $(1, f_\mathrm{esc, fs}^\mathrm{Ly\alpha})$, $(f_\mathrm{rw}, f_\mathrm{esc, rw}^\mathrm{Ly\alpha})$ or $(f_\mathrm{por}, f_\mathrm{esc, por}^\mathrm{Ly\alpha})$ depending on the total and clump optical depths $\tau_0$ and $\tau_\mathrm{0,cl}$. We note that $\tau_0$ exceeds the $\tau_0$ value in the {\it Clumpy} model when $f_\mathrm{esc}>0$ (see Eqn. \ref{eq_tau_shell}), as the same amount of gas and dust is distributed over a smaller solid angle. \subsection{IGM attenuation} \label{subsec_IGM_attenuation} The Ly$\alpha$ radiation escaping from a galaxy is attenuated by the \HI it encounters along the line of sight from the location of emission ($r_\mathrm{em}$) to the location of absorption ($r_\mathrm{obs}$). Expressing the frequency $\nu$ of a photon in terms of its rest-frame velocity $x = v/b = (\nu_\alpha/\nu - 1) c/b$ relative to the Ly$\alpha$ line centre, the transmitted fraction of radiation at frequency $x$ is given by \begin{eqnarray} T_{\alpha,x}(x) &=& \exp \left[ -\tau_\alpha(x) \right] \\ \tau_\alpha(x) &=& \int_{r_\mathrm{em}}^{r_\mathrm{obs}} \sigma_0~ \phi(x + x_\mathrm{p}(r))\ n_\mathrm{HI}(r)\ \mathrm{d}r. \label{eq_IGMtransmission_tau_alpha} \end{eqnarray} Here $\tau_\alpha$ describes the optical depth to Ly$\alpha$, while $n_\mathrm{HI}(r)$ and $v_\mathrm{p}(r)= b x_\mathrm{p}(r)$ the \HI density and peculiar velocity (in the rest-frame of the emitted Ly$\alpha$ radiation) at a physical distance $r$ from the emitter, respectively. $\sigma_0$ is the absorption cross section, described in the cgs system as \begin{eqnarray} \sigma_0 &=& \frac{\pi e^2 f}{m_e c^2} = \frac{3 \lambda_\alpha^2 A_{21}}{8\pi}, \end{eqnarray} where $f=0.4162$ is the oscillator strength, $e$ the electron charge, $m_e$ the electron mass, $\lambda_\alpha=1216\AA$ the wavelength of a photon at the Ly$\alpha$ line centre, and $A_{21}=6.265\times10^8$s$^{-1}$ the Einstein coefficient for spontaneous emission of Ly$\alpha$ photons. $\phi(x)$ depicts the Ly$\alpha$ profile for absorption and is given by a Voigt profile consisting of a Gaussian core \begin{eqnarray} \phi_\mathrm{Gauss}(x) &=& \frac{\lambda_\alpha}{\sqrt{\pi} b} ~\exp\left( - x^2 \right) \\ b &=& \sqrt{\frac{2 k_B T_\mathrm{IGM}}{m_H}}, \end{eqnarray} and Lorentzian damping wings \begin{eqnarray} \phi_\mathrm{Lorentz}(x) &=& \frac{A_{21} \lambda_\alpha^2}{4\pi^2 (x/b)^2 + A_{21}^2 \lambda_\alpha^2}. \end{eqnarray} Here $b$ is the Doppler parameter, $T_\mathrm{IGM}$ the temperature of the IGM, $k_B$ the Boltzmann constant, and $m_\mathrm{H}$ the mass of a hydrogen atom. While pressure line broadening is unimportant in regions of low \HI density and the profile can be approximated by the Gaussian core, the absorption in the Lorentzian damping wings is non-negligible in regions of high \HI density. In practise, we mimic the Voigt profile by assuming the Gaussian core profile $\phi(x)=\phi_\mathrm{Gauss}(x)$ for $|x|<x_\star$ and the Lorentzian wing profile $\phi(x)=\phi_\mathrm{Lorentz}(x)$ otherwise. Fitting to numerical results yields the transition frequency as \begin{eqnarray} x_\star &=& 0.54 \log_{10}(b) \end{eqnarray} for temperatures between $T=0.01$K and $10^8$K. Our calculations of $T_\alpha$ include the Hubble flow and peculiar velocities $v_\mathrm{p}$: outflows (inflows) of gas from a galaxy that correspond to positive (negative) $v_\mathrm{p}$ values will redshift (blueshift) the Ly$\alpha$ photons and lead to an increase (decrease) in $T_\alpha$. For each galaxy in a simulation snapshot we derive $T_\alpha$ along all directions along the major axes (i.e. along and against the x, y and z axes). By stepping through the simulation box that is divided into $512$ cells on the side (and each cell having a size of $461$ckpc), we derive the $n_\mathrm{HI}(r)$ and $v_\mathrm{p}(r)$ profiles from the {\sc astraeus} ionisation and {\sc vsmdpl} density and velocity grids. For any galaxy, we start the profiles at the galaxy position $r_\mathrm{em}=0$ and end them once the highest frequency $x_\mathrm{max} = v_\mathrm{max}/b=40$ tracked in our Ly$\alpha$ line profiles has redshifted out of absorption at $r\simeq v_\mathrm{max}/ [H_0 \Omega_m^{1/2}(1+z)^{1/2}] \simeq 13.6/(1+z)^{1/2}$cMpc. We assume $T_\mathrm{IGM}=10^4$K in ionised and $T_\mathrm{IGM}=10^2$K in neutral regions. Since the Ly$\alpha$ line redshifts out of resonance very quickly (the light travel time for distance $r$ at $z=7$ is less than $2$~Myrs, shorter than the simulation time steps), a single simulation snapshot suffices for computing the $T_\alpha$ values of the galaxies in that snapshot. We also assume periodic boundary conditions when computing $T_\alpha$. Finally, we derive the observed, i.e. dust and IGM attenuated, Ly$\alpha$ luminosity and line profile along each major axes (resulting in 6 lines of sight) as \begin{eqnarray} L_{\alpha,x}(x) &=& L_\alpha^\mathrm{ISM}(x)\ T_{\alpha,x}(x) = L_\alpha^\mathrm{intr}\ f_\mathrm{esc}^\mathrm{Ly\alpha}\ J(x), \end{eqnarray} where $f_\mathrm{esc}^\mathrm{Ly\alpha}$ and $J(x)$ are the respective Ly$\alpha$ escape fraction and line profile for a one of the models as outlined in Section \ref{subsubsec_emerging_Lya_profile_models}. The total observed Ly$\alpha$ luminosity $L_\alpha$ and total fraction of Ly$\alpha$ radiation transmitted through the IGM are yielded when integrating the respective quantity over the frequency $x$. \begin{eqnarray} L_\alpha &=& \int_{-\infty}^{\infty} L_{\alpha,x}(x)\ \mathrm{d}x \\ T_\alpha &=& \int_{-\infty}^{\infty} T_{\alpha,x}(x)\ \mathrm{d}x \end{eqnarray} In the following, we use all lines of sight as independent probes when line-of-sight-sensitive Ly$\alpha$ quantities are analysed. We derive the observed Ly$\alpha$ luminosities ($L_\alpha$) for all galaxies at $z=20$, $15$, $12$, $10$, $9$, $8$, $7.3$, $7$ and $6.6$ for any combination of emerging Ly$\alpha$ line model ({\it Gaussian}, {\it Clumpy}, {\it Porous}) and reionisation scenario ({\sc mhdec}, {\sc mhconst}, {\sc mhinc}). Free model parameters ($p$ for the {\it Gaussian} model, $\tau_\mathrm{0,cl}$ for the {\it Clumpy} and {\it Porous} models) have been chosen to visually best-fit the observed Ly$\alpha$ LFs at $z\simeq6.7$, $7.0$ and $7.3$ (see Tab. \ref{tab_Lya_line_and_ionisation_topology_models}). For simplicity and better comparison we assume in all models the gas in galaxies to have the temperature of photo-ionised gas, $T=10^4$~K. Moreover, we note that since the {\sc mhconst} scenario represents an intermediate case and provides no further insights, we limit our discussion to the {\sc mhdec} and {\sc mhinc} scenarios in the remainder of this paper. \begin{table} \centering \begin{tabular}{c|c|c|c|c} Parameter & Scenario & {\it Gaussian} & {\it Clumpy} & {\it Porous} \\ \hline $\tau_\mathrm{0,cl}$ & {\sc mhdec} & - & $8.5\times10^5$ & $1.7\times10^6$ \\ $\tau_\mathrm{0,cl}$ & {\sc mhconst} & - & $8\times10^5$ & $1.6\times10^6$ \\ $\tau_\mathrm{0,cl}$ & {\sc mhinc} & - & $5\times10^5$ & $1.5\times10^6$ \\ $p$ & {\sc mhdec} & 0.8 & - & - \\ $p$ & {\sc mhconst} & 0.9 & - & - \\ $p$ & {\sc mhinc} & 1.1 & - & - \\ $T$ & all & $10^4$~K & $10^4$~K & $10^4$~K \\ \end{tabular} \caption{Parameters for our three different Ly$\alpha$ line profile models} \label{tab_Lya_line_and_ionisation_topology_models} \end{table} \section{Numbers and properties of Ly$\alpha$ emitting galaxies} \label{sec_number_and_properties_LAEs} In this Section, we aim to identify which physical process -- the intrinsic Ly$\alpha$ production ($L_\alpha^\mathrm{intr}$), the absorption by dust within the galaxies ($f_\mathrm{esc}^\mathrm{Ly\alpha}$), or the scattering by \HI in the IGM ($T_\alpha$) -- dominates the observed Ly$\alpha$ emission. To this end, we analyse (i) how the IGM attenuation profile $T_\alpha(x)$ depends on galaxy mass and the $f_\mathrm{esc}$-sensitive ionisation topology, (ii) how the Ly$\alpha$ line profiles emerging from a galaxy depend on the density and velocity distributions of gas and dust within a galaxy and $f_\mathrm{esc}$, and how much it affects the fraction of Ly$\alpha$ radiation that is transmitted through the IGM, and (iii) to which degree the $f_\mathrm{esc}$ dependency of $L_\alpha^\mathrm{intr}$, $f_\mathrm{esc}^\mathrm{Ly\alpha}$, and $T_\alpha$ leave characteristic imprints in the Ly$\alpha$ luminosity functions and the population emitting visible Ly$\alpha$ emission. \begin{figure*} \centering \includegraphics[width=0.99\textwidth]{hist1Dprofile_Lyaline_Freq_Mvir_zall_allModels.png} \caption{Intrinsic (top) and observed (bottom) Ly$\alpha$ line profile and IGM transmission (centre) at $z=8.0$, $7.3$, $7.0$, $6.6$ for a homogeneous static gas shell (left), a clumpy outflowing gas shell (centre), and a clumpy outflowing gas shell with holes through which Ly$\alpha$ radiation escapes without scattering. Solid (dashed dotted) lines show results for the reionisation scenario where $f_\mathrm{esc}$ decreases (increases) with halo mass $M_h$.} \label{fig_profiles} \end{figure*} \subsection{The transmission through the IGM} \label{subsec_discussion_IGM_transmission} We start by discussing the frequency-dependent IGM transmission $T_{\alpha,x}$ shown in the top row of Fig. \ref{fig_profiles}. These profiles depend solely on the underlying ionisation topology and density distribution of the IGM. From the different panels depicting the average $T_{\alpha,x}$ in different halo mass bins of width $\Delta\log_{10}M_h=0.125$, we see that all $T_{\alpha,x}$ profiles follow a common trend: $T_{\alpha,x}$ decreases towards higher frequencies with an stronger decline around the Ly$\alpha$ resonance ($x=0$). Photons bluewards the Ly$\alpha$ resonance redshift into the Ly$\alpha$ resonance as they propagate through the IGM and have the largest likelihood to be absorbed by the \HI present. Photons redwards the Ly$\alpha$ resonance are also redshifted, but their likelihood of being absorbed by \HI decreases significantly as their energy drops. In each panel in the top row of Fig. \ref{fig_profiles} we show $T_{\alpha,x}$ for the two reionisation scenarios {\sc mhdec} (yellow/orange/brown lines) and {\sc mhinc} (blue lines) and redshifts $z=8.0$, $7.3$, $7.0$, $6.6$ (bright to dark lines as redshift decreases). In general, i.e. for both reionisation scenarios and all halo masses, $T_{\alpha,x}$ increases as the ionised regions grow around galaxies and the IGM is increasingly ionised (bright to dark lines): firstly, a larger ionised region shifts not only the point of strongest Ly$\alpha$ absorption to higher frequencies $x$ but also reduces the absorption in the damping wings of the Ly$\alpha$ absorption profile. Secondly, lower \HI fractions in ionised regions diminish the number of \HI atoms absorbing Ly$\alpha$ photons. These two mechanisms shape $T_{\alpha,x}$ redwards and bluewards the Ly$\alpha$ resonance. As Ly$\alpha$ photons travel through the IGM and redshift, photons emitted at frequencies $x\gtrsim0$ see the Gaussian core of the Ly$\alpha$ absorption profile $\phi(x)$ and are absorbed by \HI abundances as low as $\chi_\mathrm{HI}\gtrsim10^{-4}$; thus they are sensitive to the residual \HI fraction in ionised regions. Correspondingly, we see in Fig. \ref{fig_profiles} that $T_{\alpha,x}$ increases for $x\gtrsim0$ with decreasing redshift as the photoionisation rate around galaxies increases and lowers the residual \HI fraction in ionised regions. However, photons emitted at frequencies $x\lesssim 0$ are absorbed by the damping wings of the Ly$\alpha$ absorption profile $\phi(x)$. Since the Ly$\alpha$ absorption cross section is lower in the damping wings, the abundance of \HI needs to be significantly higher for Ly$\alpha$ photons to be absorbed; thus, as the sizes of ionised regions decrease, photons emitted at these frequencies are increasingly absorbed by the neutral regions located beyond the ionised regions around the emission sites. For this reason, we find $T_{\alpha,x}$ for $x\lesssim0$ to increase as the sizes of the ionised regions around galaxies rise with increasing halo mass and decreasing redshift. The rising sizes of ionised regions also become manifest in the shift of the frequency at which $T_{\alpha,x}$ has a value of $0.5$ to higher frequencies. Its dependence on the size of the ionised regions around galaxies makes $T_{\alpha,x}$ a tracer of the ionisation topology: our two extreme reionisation scenarios where $f_\mathrm{esc}$ either increases ({\sc mhinc}, blue dotted lines) or decreases ({\sc mhdec}, yellow to brown solid lines) with rising halo mass $M_h$ exhibit very different ionisation topologies (see Fig. \ref{fig_XHImaps_with_LAEs}). These differences are imprinted in $T_{\alpha,x}$ as follows. Firstly, since in the {\sc mhinc} scenario the higher $f_\mathrm{esc}$ values of more massive galaxies ($M_h\gtrsim10^{10}\msun$) raise the photoionisation rate within ionised regions (leading to lower $\chi_\mathrm{HI}$ values, also seen in Fig. \ref{fig_hist_ion} at $z\lesssim6$), the corresponding $T_{\alpha,x}$ values are higher bluewards the Ly$\alpha$ resonance than in the {\sc mhdec} scenario. Moreover, in the {\sc mhinc} scenario, reionisation proceeds faster, leading to the Universe being more ionised at $z<7$, and the bias of the ionising emissivity towards more massive galaxies grows with time, raising the photoionisation rate in the ionised regions. Both effects contribute to the relative increase in $T_{\alpha,x}$ from {\sc mhdec} to {\sc mhinc} to rise towards lower redshifts bluewards the Ly$\alpha$ resonance. Secondly, as the size of the ionised regions around galaxies is imprinted in $T_{\alpha,x}$ redwards the Ly$\alpha$ resonance, {\sc mhinc} shows lower (higher) $T_{\alpha,x}$ values at $z\gtrsim7$ ($z\lesssim7$) than the {\sc mhdec} scenario for galaxies with $M_h<10^{11}\msun$: At $z\gtrsim7$, ionised regions become increasingly smaller towards lower mass halos ($M_h\lesssim10^{9.5}\msun$) and higher redshifts as the corresponding $f_\mathrm{esc}$ values and global ionisation fraction decrease. However, at $z\lesssim7$, this trend reverses as the ionised regions become large enough for the red wing of the Ly$\alpha$ to be redshifted out of the absorption resonance of the Gaussian core. Towards more massive halos and higher global ionisation fractions, $T_{\alpha,x}$ becomes sensitive to the residual \HI fraction in ionised regions (c.f. $T_{\alpha,x}$ in {\sc mhinc} (light blue dotted line) exceeds $T_{\alpha,x}$ in {\sc mhdec} (yellow solid line) at $z=6.6$). It is interesting to note that the respective $T_{\alpha,x}$ values are very similar in both reionisation scenarios, despite the $f_\mathrm{esc}$ values of more massive halos ($M_h>10^{10}\msun$) differing by about one order of magnitude or more. \begin{figure*} \centering \includegraphics[width=0.99\textwidth]{Lya_LFs_allModels.png} \caption{Observed Ly$\alpha$ luminosity functions at $z=20$, $15$, $12$, $10$, $9$, $8$, $7.3$, $7$, $6.6$ for a homogeneous static gas shell (left), a clumpy outflowing gas shell (centre), and a clumpy outflowing gas shell with holes through which Ly$\alpha$ radiation escapes without scattering. Solid (dashed dotted) lines show results for the reionisation scenario where $f_\mathrm{esc}$ decreases (increases) with halo mass $M_h$.} \label{fig_LyaLF} \end{figure*} \subsection{The Ly$\alpha$ line profiles and luminosity functions} The Ly$\alpha$ line profile emerging from a galaxy represents a quantity that (i) is shaped by the density and velocity distribution of gas and dust within the galaxy and (ii) affects which fraction of the Ly$\alpha$ radiation escaping from a galaxy is transmitted through the IGM. In this Section, for our three models of the emerging Ly$\alpha$ line profiles, we discuss the following: (i) How do the assumed gas and dust distributions affect the attenuation of Ly$\alpha$ by dust in a galaxy and the emerging Ly$\alpha$ line profile? (ii) How does the Ly$\alpha$ line profile affect the Ly$\alpha$ transmission through the IGM? And, since the luminosity function of the intrinsic Ly$\alpha$ luminosity ($L_\alpha^\mathrm{intr}$) will be steeper for the scenario where $f_\mathrm{esc}$ increases ({\sc mhinc}) than when it decreases ({\sc mhdec}) with rising halo mass, (iii) which characteristics are required for the Ly$\alpha$ line profiles of the simulated galaxy population to reproduce the observed Ly$\alpha$ luminosity functions (Ly$\alpha$ LFs)? \subsubsection{The Gaussian model} The {\it Gaussian} line model centers the Ly$\alpha$ line at the Ly$\alpha$ resonance. The second row in Fig. \ref{fig_profiles} shows that its width increases as the rotational velocity of a galaxy increases with rising halo mass. Both the increase in the line width and the size of the ionised region surrounding the galaxy lead to higher IGM transmission values of Ly$\alpha$ radiation as galaxies become more massive (c.f. third row in Fig. \ref{fig_profiles}). At the same time, the fraction of Ly$\alpha$ photons that escape from the galaxies drops as the abundance of dust increases. We use the ratio between the Ly$\alpha$ and UV continuum escape fractions to adjust the Ly$\alpha$ luminosities emerging from the galaxies and fit the observed Ly$\alpha$ LFs in each of our reionisation scenarios. In the {\sc mhinc} scenario the more massive galaxies -- that dominate the observed Ly$\alpha$ LF -- have higher $f_\mathrm{esc}$ values than in the {\sc mhdec} scenario; to compensate the corresponding lower $L_\alpha^\mathrm{intr}$ values (and steeper slope of the intrinsic Ly$\alpha$ LF), we need a higher $f_\mathrm{esc}^\mathrm{Ly\alpha}/f_\mathrm{esc}^\mathrm{UV}$ ratio ($1.1$) than in the {\sc mhdec} scenario ($0.8$). Despite this compensation, the slopes of the observed Ly$\alpha$ LFs at $z\lesssim8$ (c.f. left panel in Fig. \ref{fig_LyaLF}) is still steeper for the {\sc mhinc} than for the {\sc mhdec} scenario. \subsubsection{The Clumpy model} In the {\it Clumpy} model, the clumpiness of the gas in the outflowing shell and the attenuation by dust molecules in these clumps determine the shape of the Ly$\alpha$ line profile. We note that in the following clumpiness describes the number of clumps in the dusty gas shell, i.e. a higher clumpiness corresponds to fewer clumps and thus a higher ratio between the clump ($\tau_\mathrm{0,cl}$) and total line optical depth ($\tau_0$). We find the following characteristic trends for the Ly$\alpha$ line profile: Firstly, the clumpier the gas in the shell is, the more Ly$\alpha$ radiation escapes around the Ly$\alpha$ resonance (profile showing a central peak), and the fewer Ly$\alpha$ photons escape through excursion or via the wings (double peak profile). Secondly, when assuming the same clump size for all galaxies -- as we do in this paper -- the gas clumpiness decreases as galaxies become more massive and contain more gas. Thus, from low-mass to more massive galaxies, we find the Ly$\alpha$ line profile to shift from a central peak to a double-peak profile (see the fourth row in Fig. \ref{fig_profiles} from left to right), reflecting the transition from the porous or random regime to the homogeneous regime (see Section \ref{subsubsec_Lya_line_model_clumpy}). This transition also goes in hand with an increased transmission through the IGM, which we can see when comparing the Ly$\alpha$ profiles emerging from galaxies (fourth row) with those after having traversed the IGM (fifth row in Fig. \ref{fig_profiles}). The Ly$\alpha$ luminosity at $x=0$ decreases by $\sim0.4-0.8$ orders of magnitude for all halo masses (from $10^{41.2}$erg~s$^{-1}$ to $10^{40.8}$erg~s$^{-1}$ for $M_h\simeq10^{11}\msun$ and from $10^{39.6}$erg~s$^{-1}$ to $10^{39.0}$erg~s$^{-1}$ for $M_h\simeq10^{9}\msun$ for e.g. {\sc mhdec} model), while the peak Ly$\alpha$ luminosity of the red wing decreases only about $\sim 0-0.3$ orders of magnitude at all halo masses. While the blue wing is similarly or more attenuated than the central peak in the IGM, the total fraction of Ly$\alpha$ radiation transmitted through the IGM for a fully-double peaked profile exceeds that of profiles with a central peak component. Furthermore, as the galaxies' gravitational potentials flatten with decreasing redshift, $\tau_0$ decreases and leads to (i) a narrower double-peak profile (following the dependence of the peak position on $\tau_0^{1/3}$) and (ii) a stronger central peak (the gas becomes clumpier as the ratio $\tau_\mathrm{0,cl}/\tau_0$ increases). A change in the clumpiness of the outflowing gas and dust shell (or clump optical depth $\tau_\mathrm{0,cl}$ and $\tau_\mathrm{d,cl}$) goes not only in hand with a change in the Ly$\alpha$ profile affecting $T_\alpha$ but also an altered attenuation of the escaping Ly$\alpha$ radiation by dust. Thus, adjusting the clump optical depth allows us to enhance and reduce the Ly$\alpha$ luminosities and reproduce the observed Ly$\alpha$ LFs: As we increase the size of the clumps, i.e. increase $\tau_\mathrm{0,cl}$, Ly$\alpha$ photons will scatter with fewer clumps, leading to (i) a higher fraction $f_\mathrm{esc}^\mathrm{Ly\alpha}$ escaping, and (ii) a higher fraction escaping at the Ly$\alpha$ resonance, which again leads to stronger attenuation by \HI in the IGM. However, effectively there exists a limit in enhancing Ly$\alpha$ emission from galaxies by decreasing $\tau_\mathrm{0,cl}$. Once the emerging Ly$\alpha$ profile is fully double-peaked, the attenuation by \HI in the IGM can not be decreased any further (by changing the injected Ly$\alpha$ line profile), and a further enhancement of observable Ly$\alpha$ emission could be solely due to the attenuation by dust. But the latter is not possible, as in the homogeneous escape regime and for sufficiently low $\tau_\mathrm{0,cl}$ values such that the escape frequency $x_\mathrm{esc}$ follows the injection frequency $x_i$ (c.f. Eqns \ref{eq_xesc} and \ref{eq_fccrit}), the attenuation by dust increases with decreasing clump size (see Section \ref{subsubsec_dust_refined_model}). With the observed Ly$\alpha$ LFs being dominated by the more massive galaxies ($M_h\gtrsim10^{10}\msun$, as we will discuss in the next Section), we find that the Ly$\alpha$ luminosities do not increase further for $\tau_\mathrm{0,cl}\lesssim10^5$. Fortunately, all our reionisation scenarios can fit the observed Ly$\alpha$ LFs for $\tau_\mathrm{0,cl}>10^5$ (see centre panel in Fig. \ref{fig_LyaLF}): as the intrinsic Ly$\alpha$ LFs is lower at the bright end in the {\sc mhinc} scenario, a lower $\tau_\mathrm{0,cl}$ value ($5\times10^5$) is required than for the {\sc mhdec} scenario ($8.5\times10^5$). Nevertheless, the slopes of the resulting observed Ly$\alpha$ LFs at $z\lesssim8$ keep the trends of the intrinsic Ly$\alpha$ LFs, with the bright ends of the Ly$\alpha$ LFs being steeper in the {\sc mhinc} than in the {\sc mhdec} scenario. \subsubsection{The Porous model} The {\it Porous} model represents a refinement of the {\it Clumpy} model. It adds gas-free channels through which Ly$\alpha$ and ionising photons escape freely. This explains why, to first order, we find the trends in the last two rows of Fig. \ref{fig_profiles} to be similar to those in the fourth and fifth rows. A lower clumpiness of gas and dust in the outflowing shell induces a stronger prevalence of the double-peak component in the Ly$\alpha$ line profile emerging from a galaxy, enhancing the IGM transmission $T_\alpha$ and absorption by dust within the galaxy, and causing the corresponding Ly$\alpha$ LFs to shift to lower values. On the other hand, it differs from the {\it Clumpy} model substantially, as $f_\mathrm{esc}$ determines the minimum fraction of Ly$\alpha$ radiation that escapes at the Ly$\alpha$ resonance and contributes to the central peak in our modelling. Hence, as long as $\tau_\mathrm{0,cl}$ remains above the $\tau_\mathrm{0,cl}$ value that leads to the same fraction of Ly$\alpha$ escaping in the central peak than given by $f_\mathrm{esc}$ (referred to as $\tau_\mathrm{0,cl}^\mathrm{crit}$ in the following), the {\it Porous} model inherits the trend of the {\it Clumpy} model. As $\tau_\mathrm{0,cl}$ drops below $\tau_\mathrm{0,cl}^\mathrm{crit}$, a further decrease in $\tau_\mathrm{0,cl}$ has no effect on the Ly$\alpha$ line profile emerging from a galaxy, and the observed Ly$\alpha$ LFs remains "fixed". The resulting upper limit of $f_\mathrm{esc}^\mathrm{Ly\alpha}$ is essential, as together with $L_\alpha^\mathrm{intr}$ it provides an upper limit to $f_\mathrm{esc}$ values that fit the observed Ly$\alpha$ LFs. We find this upper limit to be about $f_\mathrm{esc}\sim0.5$ in our {\sc astraeus} model. Due to their opposing dependencies of $f_\mathrm{esc}$ with halo mass, the Ly$\alpha$ profiles in the {\it Porous} model show the largest differences between the {\sc mhdec} and {\sc mhinc} scenarios among our three Ly$\alpha$ line profile models. While the double-peak component is more prominent in the most massive galaxies ($M_h\simeq10^{11}\msun$) in the {\sc mhdec} scenario, the central peak is stronger in the {\sc mhinc} scenario. To fit the observed Ly$\alpha$ LFs, we find that we require for both reionisation scenarios a more clumpy gas and dust distribution than in the {\it Clumpy} model, i.e. a (higher) $\tau_\mathrm{0,cl}$ value of $1.5-1.7\times10^6$. These increased $\tau_\mathrm{0,cl}$ values enhance the corresponding $f_\mathrm{c,crit}$ values and thus the dust attenuation in the homogeneous regime giving rise to the double-peak components and counteract the increased escape close to the Ly$\alpha$ resonance. This model-integrated correlation between $f_\mathrm{esc}^\mathrm{Ly\alpha}$ and $f_\mathrm{esc}$ counteracts the trend of flattening (steepening) the slope of the intrinsic Ly$\alpha$ LFs due to $f_\mathrm{esc}$ decreasing (increasing) with rising halo mass: If $f_\mathrm{esc}$ is low (high), more (less) Ly$\alpha$ radiation is subject to dust attenuation. This model feature explains why the observed Ly$\alpha$ LFs of the {\sc mhdec} ({\sc mhinc}) simulation are steeper (shallower) than in the {\it Clumpy} model. \begin{figure*} \centering \includegraphics[width=0.99\textwidth]{2Dhistogram_numDens_medianLyaProperties-Mvir_zall.png} \caption{Median of indicated galactic properties (lines) and their $\sim1.3\sigma$ uncertainties (shaded regions) as a function of halo mass $M_h$ at $z=8.0$, $7.3$, $7.0$, $6.6$ for a homogeneous static gas shell. Solid (dashed dotted) lines show results for the reionisation scenario where $f_\mathrm{esc}$ decreases (increases) with halo mass $M_h$.} \label{fig_histograms} \end{figure*} As the dust composition and absorption cross sections remain uncertain during the EoR, we note that a lower (higher) dust absorption cross section $\kappa_\mathrm{abs}$ could still reproduce the observed Ly$\alpha$ LFs in our {\it Clumpy} and {\it Porous} models by raising (decreasing) the clump optical depth $\tau_\mathrm{0,cl}$. However, this would go along with an enhanced (reduced) double-peak and reduced (enhanced) central-peak component in the average Ly$\alpha$ line profile emerging from galaxies. Finally, we briefly comment on how our emerging and IGM-attenuated Ly$\alpha$ profiles compare to those obtained from radiative hydrodynamical simulations of clouds and small cosmological volumes ($\sim10^3$cMpc$^3$). While the {\it Clumpy} and {\it Porous} reproduce the double- and triple-peak profiles and their dependence on $N_\mathrm{HI}$ and $f_\mathrm{esc}$ found in cloud simulations \citep{Kakiichi2021, Kimm2019, Kimm2021} by construction, their Ly$\alpha$ line profiles differ from those obtained from the {\sc sphinx} simulation \citep{Garel2021}. In {\sc sphinx} the median angle-averaged Ly$\alpha$ line profile has been found to be less double-peaked towards brighter galaxies, with the blue peak being seemingly increasingly suppressed. This is the opposite trend of our findings. The discrepancy lies in the differently assumed or simulated ISM structures: While our LAE models assume an idealised scenario of same-sized outflowing dusty gas clumps, the {\sc sphinx} simulation follows the formation of star-forming clouds within galaxies. With rising galaxy mass, we expect the simulated {\sc sphinx} galaxies to contain a higher number of star-forming clouds with various velocity and size distributions. A single or very few star-forming clouds -- as found in low-mass galaxies -- will give rise to a double-peaked Ly$\alpha$ line profile. Adding the profiles of multiple/many star-forming clouds at different velocities will give rise to increasingly more complex Ly$\alpha$ line profiles as galaxies become more massive. Adjusting our current Ly$\alpha$ line profile models to the complex structure of the ISM will be the subject of future work. \subsection{The dependence of Ly$\alpha$ properties on halo mass} In this Section, we provide a more detailed discussion of how the intrinsic Ly$\alpha$ luminosity ($L_\alpha^\mathrm{intr}$), the Ly$\alpha$ escape fraction, the Ly$\alpha$ transmission through the IGM, the observed Ly$\alpha$ luminosity, and Ly$\alpha$ equivalent width depend on halo mass and evolve with redshift for the different reionisation scenarios. To this end, we show these quantities as a function of halo mass for both reionisation scenarios ({\sc mhdec}: yellow/orange/brown lines; {\sc mhinc}: blue lines) and redshifts $z\simeq8$, $7.3$, $7$ and $6.6$ in Fig. \ref{fig_histograms} and list the corresponding average \HI fractions in Table \ref{table_XHI_reionisation_scenarios}. Solid and dot-dashed lines in Fig. \ref{fig_histograms} depict the median value for galaxies in the given halo mass bin, and shaded regions indicate the range spanned by $68\%$ of the values. For line-of-sight-dependent Ly$\alpha$ properties ($T_\alpha$, $L_\alpha$, EW$_\alpha$), we include all $6$ lines of sight. \begin{table} \centering \begin{tabular}{|c|c|c|c|} $z$ & $\langle\chi_\mathrm{HI}\rangle^\mathrm{MHINC}$ & $\langle\chi_\mathrm{HI}\rangle^\mathrm{CONST}$ & $\langle\chi_\mathrm{HI}\rangle^\mathrm{MHDEC}$\\ \hline 8.0 & 0.84 & 0.80 & 0.71 \\ 7.3 & 0.69 & 0.66 & 0.59 \\ 7.0 & 0.52 & 0.52 & 0.49 \\ 6.6 & 0.23 & 0.29 & 0.34 \\ \end{tabular} \caption{The evolution of the global \HI fractions of the IGM for our reionisation scenarios.} \label{table_XHI_reionisation_scenarios} \end{table} \paragraph*{Intrinsic Ly$\alpha$ luminosity $L_\alpha^\mathrm{intr}$:} As the most recent star formation dominates the production of ionising photons within galaxies, we find $L_\alpha^\mathrm{intr}$ to follow the SFR-$M_h$ relation \citep[for a detailed discussion, see][]{Hutter2021a}. While the range of SFR values is broad for low-mass halos ($M_h\lesssim10^{9.5}\msun$) where SN feedback drives stochastic star formation, the SFR-$M_h$ relation becomes tighter towards more massive galaxies as SN feedback ejects an increasingly lower fraction of gas from the galaxy. Being mainly produced by recombining hydrogen atoms within a galaxy, the Ly$\alpha$ radiation produced within the galaxy correlates with the escape fraction of ionising photons as $1-f_\mathrm{esc}$. As we can see from the first row in Fig. \ref{fig_histograms}, this dependency on $f_\mathrm{esc}$ leads to higher (lower) Ly$\alpha$ luminosities for more massive galaxies, lower (higher) Ly$\alpha$ luminosities for low-mass galaxies, and thus a shallower (steeper) LFs in the {\sc mhdec} ({\sc mhinc}) scenario. \paragraph*{Ly$\alpha$ escape fraction $f_\mathrm{esc}^\mathrm{Ly\alpha}$:} As the dust content in galaxies increases with their mass, we find $f_\mathrm{esc}^\mathrm{Ly\alpha}$ to decrease with rising halo mass at all redshifts and for all Ly$\alpha$ line models. However, the different assumed distributions of dust and their resulting attenuation of Ly$\alpha$ radiation lead to differences in the details of this global trend: Firstly, the {\it Gaussian} model shows a steeper decline in $f_\mathrm{esc}^\mathrm{Ly\alpha}$ for galaxies with $M_h\gtrsim10^{10.5}\msun$ than the {\it Clumpy} and {\it Porous} models. Secondly, $f_\mathrm{esc}^\mathrm{Ly\alpha}$ is always higher in the {\sc mhinc} than in the {\sc mhdec} scenario. This is necessary to reproduce the observed Ly$\alpha$ LFs by compensating the lower intrinsic Ly$\alpha$ luminosities with a more clumpy gas-dust distribution in the {\sc mhinc} scenario. In case of the {\it Clumpy} model, it also highlights how a decrease in the clump optical depth by a factor $\sim 2$ can increase $f_\mathrm{esc}^\mathrm{Ly\alpha}$ by reducing the fraction of Ly$\alpha$ photons escaping in the homogeneous regime (i.e. a decrease in $f_\mathrm{c,crit}$ and $\tau_\mathrm{0,cl}$ leads to a reduced number of clumps encounters $N_\mathrm{cl}$ and the clump albedo $\epsilon$). Thirdly, for the {\sc mhdec} ({\sc mhinc}) scenario, the $f_\mathrm{esc}^\mathrm{Ly\alpha}$ values show higher (lower) values in the {\it Porous} model than in the {\it Clumpy} model for $M_h\lesssim10^{10}\msun$. The reason for this difference is as follows. In both scenarios the higher $\tau_\mathrm{0,cl}$ values in the {\it Porous} model increase the dust attenuation of Ly$\alpha$ escaping in the homogeneous regime. But only a fraction $1-f_\mathrm{esc}$ of the Ly$\alpha$ photons is subject to dust attenuation. This unattenuated escape of Ly$\alpha$ radiation imprints the mass-dependency of $f_\mathrm{esc}$ in the $f_\mathrm{esc}^\mathrm{Ly\alpha}$ values. However, for galaxies with $M_h\gtrsim10^{10}\msun$, this imprint ($f_\mathrm{esc}^\mathrm{Ly\alpha}$ enhancement in {\it Porous} model) is only visible in the {\sc mhinc} scenario where $f_\mathrm{esc}$ values are sufficiently large ($>0.1$); in the {\sc mhdec} scenario $f_\mathrm{esc}$ values are too small. \paragraph*{Ly$\alpha$ IGM transmission $T_\alpha$:} As outlined in Section \ref{subsec_discussion_IGM_transmission}, the surrounding ionised region (in particular its size and residual \HI fraction) and the Ly$\alpha$ line profile emerging from a galaxy determine how much of a galaxy's escaping Ly$\alpha$ radiation is transmitted through the IGM. For more massive galaxies with $M_h\gtrsim10^{10}\msun$, $T_\alpha$ is mainly shaped by the Ly$\alpha$ profile. This is because the ionised regions surrounding them are sufficiently large -- due to their enhanced ionising emissivity and their clustered neighbourhood -- for the Ly$\alpha$ radiation to redshift out of absorption. Hence, at these high halo masses, any trends in $T_\alpha$ reflect the ratio between the Ly$\alpha$ radiation escaping around the Ly$\alpha$ resonance and escaping through the wings: the more Ly$\alpha$ escapes in the central peak, the lower is the $T_\alpha$ value. Indeed, as can be seen in Fig. \ref{fig_histograms}, the {\it Gaussian} model concentrating the emerging Ly$\alpha$ radiation around the Ly$\alpha$ resonance shows the lowest median $T_\alpha$ values at $M_h\gtrsim10^{10}\msun$ among all Ly$\alpha$ line profile models. In the {\it Clumpy} model, where the fraction of Ly$\alpha$ escaping through the wings increases with rising halo mass, we find the median $T_\alpha$ value to increase accordingly. This effect is more evident for the {\sc mhinc} scenario as it transitions from a Ly$\alpha$ line profile with a dominating central peak at $M_h\simeq10^{11}\msun$ to one with a prevailing double peak component. The {\it Porous} model also confirms that $T_\alpha$ is highly sensitive to the Ly$\alpha$ line profile. In the {\sc mhinc} scenario, the double peak component is weaker and increases less with halo mass, leading to slightly lower $T_\alpha$ values than in the {\it Clumpy} model for $M_h\simeq10^{10-11}\msun$ and $T_\alpha$ hardly changing with halo mass. In the {\sc mhdec} scenario, we see the same effect but to a lower degree. However, for less massive galaxies ($M_h\lesssim10^{10}\msun$), $T_\alpha$ is more sensitive to the properties of their surrounding ionised regions. Since the ionised regions around less massive galaxies can differ significantly depending on their environment and phase in their stochastic star formation cycle (see \citet{Hutter2021b} and \citet{Legrand2022b} for environment dependence), their $T_\alpha$ values span across an extensive range from as low as effectively zero to as high as $\simeq70\%$. Nevertheless, the median $T_\alpha$ value shows a definite trend. It increases with rising halos mass for all models and at all stages of reionisation. With increasing halo mass, galaxies are surrounded by larger ionised regions as they form more stars emitting ionising photons and are more likely to be located in clustered regions that are reionised earlier. The larger the surrounding ionised regions are, the higher the transmission of Ly$\alpha$ radiation through the IGM. We can see this relationship when comparing the median $T_\alpha$ values of the {\sc mhdec} and {\sc mhinc} simulations. In the {\sc mhdec} scenario low-mass galaxies are surrounded by larger ionised regions at $z\gtrsim7$ than in the {\sc mhinc}, causing their corresponding $T_\alpha$ values to be raised (c.f. orange/brown solid lines vs dark blue/blue lines in the third row of Fig. \ref{fig_histograms}). At $z\lesssim7$, however, reionisation progresses faster and the photoionisation rate in clustered ionised regions yields a lower residual \HI fraction in the {\sc mhinc} simulation, both leading to a higher median $T_\alpha$ value for the {\sc mhinc} than {\sc mhdec} scenario at $z\simeq6.6$. Finally, we briefly discuss how the Ly$\alpha$ line profile emerging from a galaxy affects $T_\alpha$ for less massive galaxies. From Fig. \ref{fig_histograms} we see that the $T_\alpha$ values differ between our three different Ly$\alpha$ line profile models: While at all stages of reionisation the $T_\alpha$ values for $M_h\lesssim10^{10}\msun$ are very similar in the {\it Porous} and {\it Clumpy} model, the {\it Porous} model shows lower $T_\alpha$ values for $M_h\gtrsim10^{10}\msun$ at $z\lesssim7$ than the {\it Clumpy} model in the {\sc mhinc} scenario. This drop goes in hand with the increased central peak component in these more massive galaxies (c.f. Fig. \ref{fig_profiles} and the previous Section). The median $T_\alpha$ values of the {\it Gaussian} model always lie below those of the {\it Clumpy} and {\it Porous} models; a larger fraction of Ly$\alpha$ radiation escapes closer to the Ly$\alpha$ resonance and is thus subject to stronger attenuation by the IGM. \paragraph*{Variance of the IGM transmission along different lines of sight:} To investigate how strongly the transmission of Ly$\alpha$ radiation through the IGM depends on the direction, we show the standard deviation of $T_\alpha$ values over the $6$ lines of sight aligning with the major axes in relation to the corresponding mean value, $\sigma_{T_\alpha} / \langle T_\alpha \rangle = \sqrt{\langle T_\alpha^2\rangle - \langle T_\alpha\rangle^2}/ \langle T_\alpha \rangle$, in the fourth row of Fig. \ref{fig_histograms}. At all redshifts and for all models, $\sigma_{T_\alpha} / \langle T_\alpha \rangle$ decreases with rising halo mass and decreasing redshift for the following reason. As galaxies grow in mass, they produce more ionising photons that can ionise larger regions around them and are also more likely to be located in more strongly clustered ionised regions, both enhancing and homogenising Ly$\alpha$ transmission through the IGM along different lines of sight. However, we note that parts of the decrease of $\sigma_{T_\alpha} / \langle T_\alpha \rangle$ with decreasing redshift is also due to $\langle T_\alpha \rangle$ rising. Since it is hard to disentangle these two effects, we will focus on relative differences between the different reionisation scenarios and Ly$\alpha$ line profile models. Firstly, the more the emerging Ly$\alpha$ line profile is concentrated around the Ly$\alpha$ resonance, the more sensitive is $T_\alpha$ to the varying \HI abundance around a galaxy, and the larger is the variance across different lines of sight (c.f. the higher $\sigma_{T_\alpha} / \langle T_\alpha \rangle$ values in the {\it Gaussian} compared to the other two models, and in the {\it Porous} compared to the {\it Clumpy} model for $M_h\gtrsim10^{10.5}\msun$ when central peak component dominates). Secondly, we focus on Ly$\alpha$ line profiles more sensitive to the environmental \HI abundance of a galaxy ({\it Gaussian} model). When accounting for the $\langle T_\alpha \rangle$ values to be lower in the {\sc mhinc} than in the {\sc mhdec} scenario at $z\simeq7$ (see median $T_\alpha$ values in the third row of Fig. \ref{fig_histograms}), we can deduce that the variance of $T_\alpha$ across different lines of sight is higher in the {\sc mhdec} than in the {\sc mhinc} scenario. Indeed in the {\sc mhinc} scenario, the shape of ionised regions is closer to spheres and less filamentary, which results in more ``homogeneous" $T_\alpha$ values. \paragraph*{Observed Ly$\alpha$ luminosity $L_\alpha$:} For any model and reionisation scenario, the trend of $L_\alpha$ with rising halo mass depends on the respective trends of $L_\alpha^\mathrm{intr}$, $f_\mathrm{esc}^\mathrm{Ly\alpha}$, and $T_\alpha$. Being surrounded by smaller ionised regions, the low $T_\alpha$ values of less massive galaxies ($M_h\lesssim10^{10}\msun$) strongly suppress and shape their emerging Ly$\alpha$ radiation. In contrast, the $T_\alpha$ values of more massive galaxies ($M_h\gtrsim10^{10}\msun$) show only weak trends with halo mass and similar values throughout reionisation. For this reason, the trends of their $L_\alpha$ values with halo mass are predominantly shaped by the corresponding trends of $L_\alpha^\mathrm{intr}$ and $f_\mathrm{esc}^\mathrm{Ly\alpha}$. Though, for model parameters that reproduce the observed Ly$\alpha$ LFs, a relative increase (decrease) of $L_\alpha^\mathrm{intr}$ towards higher halo masses, such as in the {\sc mhinc} ({\sc mhdec}) scenario, is compensated by an $f_\mathrm{esc}^\mathrm{Ly\alpha}$ that decreases more (less) strongly with halo mass. Nevertheless, the resulting relation between $L_\alpha$ and halo mass does not significantly change. It shows that only more massive galaxies where SN and radiative feedback do not considerably suppress star formation exhibit observable Ly$\alpha$ emission of $L_\alpha\gtrsim10^{41}$erg~s$^{-1}$. \paragraph*{Observed Ly$\alpha$ equivalent width EW$_\alpha$:} We compute the Ly$\alpha$ equivalent width EW$_\alpha$ from $L_\alpha$ and the observed UV continuum luminosity at $1500$\AA~ ($L_c$). The trend of the median EW$_\alpha$ with halo mass follows that of $L_\alpha$, with median EW$_\alpha$ values ranging from $\sim20$\AA~ for galaxies in $M_h\simeq10^{10}\msun$ halos to $\sim50$\AA~ for galaxies in $M_h\simeq10^{11.3}\msun$ halos. More massive galaxies with a strongly attenuated UV continuum -- the fraction of these galaxies increases towards higher halo masses due to the higher abundance of dust -- and high $L_\alpha$ values show EW$_\alpha$ values up to $\sim300$\AA~ (and very few up to $\sim1000$\AA) in the {\it Clumpy} and {\it Porous} models. However, these high EW$_\alpha$ values are not present in the {\it Gaussian} model for the following reason: in this model, the escape of Ly$\alpha$ and UV continuum radiation differs just by a constant factor, while the dust attenuation of Ly$\alpha$ and UV continuum photons within a galaxy are not only linked via the dust mass in the {\it Clumpy} and {\it Porous} models. \paragraph*{} In summary, we find that only more massive galaxies ($M_h\gtrsim10^{10}\msun$) where star formation is not substantially suppressed by SN and radiative feedback from reionisation show significant Ly$\alpha$ emission of $L_\alpha\gtrsim10^{41}$erg~s$^{-1}$. This limitation of observable Ly$\alpha$ emission to more massive galaxies allows the $f_\mathrm{esc}$-dependency of the intrinsic Ly$\alpha$ luminosity to be compensated by a weaker or stronger attenuation of Ly$\alpha$ by dust within a galaxy. If less massive galaxies were visible in Ly$\alpha$, they would break this degeneracy as they would not contain enough dust to attenuate the Ly$\alpha$ radiation in all scenarios sufficiently. \section{The spatial distribution of Ly$\alpha$ emitting galaxies} \label{sec_spatial_distribution_LAEs} \begin{figure*} \centering \includegraphics[width=0.99\textwidth]{fescMHDECfescMHINC_LyaProfileFESC_tauClump1.7e6_LyaLum1.e42_dustGRONKE_clumpCrossSec2.e5_clumpAlbedo-1._XHI_z256.png} \caption{Neutral hydrogen fraction fields at $z=8.0$ (left), $z=7.0$ (centre), and $z=6.6$ (right) for the {\sc mhdec} (top) and {\sc mhinc} models (bottom). We show a $1.6h^{-1}$cMpc-thick (5 cells) slice through the centre of the simulation box. The blue color scale depicts the volume-averaged value of the neutral fraction in each cell. Red stars show the location of LAEs, with their sizes and colour scale encoding the observed Ly$\alpha$ luminosity along the $z$-direction.} \label{fig_XHImaps_with_LAEs} \end{figure*} \begin{figure*} \centering \includegraphics[width=0.99\textwidth]{FESC_2Dhistogram_numDens_Lya_obs_XHI-DENS.png} \caption{2D probability distribution in $\chi_\mathrm{HI}$ and overdensity for all simulation cells (grey) and galaxies with $L_\alpha\geq10^{42}$erg~s$^{-1}$ (green), $L_\alpha\geq10^{42.5}$erg~s$^{-1}$ (blue), and $L_\alpha\geq10^{43}$erg~s$^{-1}$ (red) in the {\it Porous} model. The top (bottom) row shows results for the reionisation scenario where $f_\mathrm{esc}$ decreases (increases) with halo mass $M_h$.} \label{fig_Lya_obs_XHI-DENS} \end{figure*} In this Section, we analyse where galaxies with observable Ly$\alpha$ emission are located in the large-scale structure and how their environment and Ly$\alpha$ luminosity distributions differ in our reionisation scenarios ({\sc mhdec} and {\sc mhinc}). For this purpose, we discuss the environment of Ly$\alpha$ emitting galaxies in terms of their large-scale spatial distribution (Fig. \ref{fig_XHImaps_with_LAEs}), their surrounding over-density ($1+\delta$) and \HI fraction ($\chi_\mathrm{HI}$) (Fig. \ref{fig_Lya_obs_XHI-DENS}), and their 3D autocorrelation functions (Fig. \ref{fig_3Dcorrfuncs_FESC}). As we yield very similar results for our three Ly$\alpha$ lines profile, we use the {\it Porous} model as a representative case. \subsection{The environment} Before detailing the location of Ly$\alpha$ emitting galaxies in the large-scale matter distribution, we briefly discuss the ionisation structure of the IGM using Fig. \ref{fig_XHImaps_with_LAEs} and \ref{fig_Lya_obs_XHI-DENS}. Fig. \ref{fig_XHImaps_with_LAEs} shows the ionisation fields at $z=8$, $7$ and $6.7$ for the {\sc mhdec} (top) and {\sc mhinc} scenarios (bottom). As can be seen in this Figure, if $f_\mathrm{esc}$ decreases with halo mass ({\sc mhdec} scenario), reionisation is not only more extended but also ionised regions are on average smaller, follow more the large-scale density distribution and thus have less bubble-like shapes than if $f_\mathrm{esc}$ increases with halo mass ({\sc mhinc} scenario). The grey contours in Fig. \ref{fig_Lya_obs_XHI-DENS}, showing the two-dimensional probability density distribution of the \HI fraction ($\chi_\mathrm{HI}$) and over-density of the IGM at $z=8$, $7.3$, $7$ and $6.7$ (derived from all cells of the $512^3$ ionisation and density grids output by {\sc astraeus}), complement the picture. These contours indicate that not only an increasing fraction of the volume becomes ionised as reionisation progresses (from right to left) but also the $\chi_\mathrm{HI}$ values in ionised regions decrease (e.g. from $\chi_\mathrm{HI}\simeq10^{-4}$ ($10^{-4.3}$) in average dense regions with $\log_{10}(1+\delta)+1$ at $z=8$ to $\chi_\mathrm{HI}\simeq10^{-4.7}$ ($10^{-5.2}$) at $z=6.7$ for the {\sc mhdec} ({\sc mhinc}) scenario). The latter is because as galaxies grow in mass with decreasing redshift, their emission of ionising photons increases, leading to a rise of the photoionisation rates within ionised regions and thus lower $\chi_\mathrm{HI}$ values. Moreover, at the same time, as the photoionisation rate within ionised regions becomes increasingly homogeneous, the enhanced number of recombinations in denser regions \citep[for the detailed modelling description see][]{Hutter2018a} leads to a positive correlation between the \HI fraction and density in ionised regions. However, the exact value of the photoionisation rate within ionised regions and its spatial distribution depend strongly on the ionising emissivities escaping from the galaxies into the IGM. If less clustered low-mass galaxies drive reionisation -- as in the {\sc mhdec} scenario (top row in Fig. \ref{fig_Lya_obs_XHI-DENS}) --, the resulting photoionisation rate is more homogeneous and lower than if the more strongly clustered massive galaxies are the main drivers of reionisation (c.f. {\sc mhinc} scenario in the bottom row of Fig. \ref{fig_Lya_obs_XHI-DENS}). The difference in the photoionisation rate's magnitude explains the shift of the $\chi_\mathrm{HI}$ values by an order of magnitude to lower values in under-dense to moderately over-dense regions ($\log_{10}(1+\delta)\lesssim1.2$) when going from the {\sc mhdec} to the {\sc mhinc} scenario. In contrast, the more inhomogeneous distribution of the photoionisation rate's values enhances this drop in $\chi_\mathrm{HI}$ in over-dense regions where the most massive galaxies are located. As we can see from the red stars in Fig. \ref{fig_XHImaps_with_LAEs} and coloured contours in Fig. \ref{fig_Lya_obs_XHI-DENS}, Ly$\alpha$ emitting galaxies always lie in ionised regions. Although these galaxies trace the ionisation topology, their populations hardly differ for our two opposing reionisation scenarios. This absence of a significant difference is due to their massive nature \citep[see also e.g.][]{Kusakabe2018}: hence, all Ly$\alpha$ emitting galaxies lie in over-dense regions, with the ones brighter in Ly$\alpha$ located in denser regions (c.f. green to blue to red contours). The latter trend is mainly because more massive galaxies, which exhibit higher star formation rates and produce more ionising and Ly$\alpha$ radiation, are located in denser regions. \begin{figure*} \centering \includegraphics[width=0.99\textwidth]{LyaProfileFESC_3Dcorrfuncs.png} \caption{Top panels: 3D correlation function of galaxies that exceed an observed Ly$\alpha$ luminosity of $L_\alpha>10^{42}$erg~s$^{-1}$ (left), $L_\alpha>10^{42.5}$erg~s$^{-1}$ (centre) and $L_\alpha>10^{43}$erg~s$^{-1}$ (right) at $z=10$, $9$, $8$, $7.3$, $7$, $6.6$ for the {\it Porous} model. Solid (dashed dotted) lines show results for the reionisation scenario where $f_\mathrm{esc}$ decreases (increases) with halo mass $M_h$ and assumes $\tau_\mathrm{0,cl}=4\times10^5$ ($2\times10^5$). The grey to black lines indicate the corresponding LBG ($M_\mathrm{UV}<-17$) 3D correlation functions from $z=10$ to $6.6$. Bottom panels: Ratio between the 3D LAE correlation functions of the {\sc mhinc} and the {\sc mhdec} scenario at fixed redshifts.} \label{fig_3Dcorrfuncs_FESC} \end{figure*} \subsection{The clustering} In this Section, we address the question whether the Ly$\alpha$ luminosity-dependent distribution of LAEs could differ for reionisation scenarios with opposing trends of $f_\mathrm{esc}$ with halo mass. For this purpose, we analyse the 3D autocorrelation function for LAE samples with different minimum Ly$\alpha$ luminosities (Fig. \ref{fig_3Dcorrfuncs_FESC}). We define a galaxy to be an LAE if it has an observed Ly$\alpha$ luminosity of $L_\alpha\geq10^{42}$erg~s$^{-1}$. Before we discuss the differences between our opposing $f_\mathrm{esc}$ descriptions, we give a brief overview of the global trends and their origins. Firstly, as predicted by hierarchical structure formation, all autocorrelation functions in Fig. \ref{fig_3Dcorrfuncs_FESC} decrease from small to large scales, implying stronger clustering of galaxies on small scales than on large scales. Secondly, the dropping amplitude of the LAE autocorrelation functions with decreasing redshift (from ochre to blue lines) reflects the growth and increasing ionisation of ionised regions. Thirdly, since the $L_\alpha$ value of a galaxy is strongly correlated to its halo mass in our galaxy evolution model, selecting galaxies with increasingly brighter Ly$\alpha$ luminosities (left to right in Fig. \ref{fig_3Dcorrfuncs_FESC}) corresponds to selecting more massive galaxies. The latter explains the increasing amplitude and stronger clustering. Comparing the correlation functions of the $L_\alpha$ selected galaxies with those of LBGs (galaxies with $M_\mathrm{UV}\geq-17$) shows that the Ly$\alpha$ selected galaxies are more massive than our LBGs (solid grey lines). It also shows that the decrease in the clustering of LAEs is partially due to galaxies of a given mass becoming a less biased tracer of the underlying density field as the density of the Universe drops with decreasing redshift. Comparing the autocorrelation functions of our two opposing $f_\mathrm{esc}$ descriptions, we find that the {\sc mhinc} scenario (dotted lines) has higher autocorrelation amplitudes than the {\sc mhdec} scenario (solid lines) throughout reionisation and for all minimum Ly$\alpha$ luminosities studied. This difference decreases towards larger scales. The reason for these higher amplitudes is twofold: On the one hand, the {\sc mhinc} scenario has a lower global average ionisation fraction at $z\gtrsim7$ than the {\sc mhdec} scenario (see Fig. \ref{fig_hist_ion}). Its ionised regions are located around more biased tracers of matter, i.e. more massive galaxies, leading to a stronger clustering. While the scenarios' difference in $\langle\chi_\mathrm{HI}\rangle$ reaches its maximum with $\sim0.13$ around $z\simeq8$, the difference in the autocorrelation amplitudes rises even towards higher redshifts. This is because, with increasing redshift, galaxies of the same mass become more biased tracers of the underlying matter distribution. Thus, the same difference in $\langle \chi_\mathrm{HI}\rangle$ at higher $\langle \chi_\mathrm{HI}\rangle$ values leads to a larger difference in the clustering of LAEs, since the Ly$\alpha$ luminosity of a galaxy correlates strongly with its halo mass. We note that selecting LAEs with a higher minimum Ly$\alpha$ luminosity also corresponds to selecting more biased tracers and yields higher correlation amplitudes (c.f. different panels in Fig. \ref{fig_3Dcorrfuncs_FESC}). On the other hand, during the early stages of reionisation, ionised regions grow preferentially around the most biased tracers of the underlying matter field (most massive galaxies) in the {\sc mhinc} scenario. Thus, we would expect that, at the same $\langle\chi_\mathrm{HI}\rangle$ value, LAEs in this scenario are more clustered than LAEs in the {\sc mhdec} scenario where the $f_\mathrm{esc}$ decreasing with rising halo mass counteracts the biased growth of ionised regions. Indeed, at $z\lesssim7$, the correlation amplitude in the {\sc mhinc} scenario is higher or similar than in the {\sc mhdec} scenario, although the Universe is similarly or more ionised in the former, respectively. This difference becomes more apparent as we consider higher minimum Ly$\alpha$ luminosities of $L_\alpha>10^{42.5}$erg~s$^{-1}$. It is driven by the higher photoionisation rates in the ionised regions around massive galaxies. We conclude that, since LAEs coincide with the most massive galaxies located in dense and ionised regions, their clustering is primarily a tracer of the global ionisation state of the IGM. While the exact ionisation topology at fixed $\langle\chi_\mathrm{HI}\rangle$ values has only a secondary effect on the clustering of LAEs during the second half of reionisation, the spatial distribution of LAEs provides a relatively robust tool to map the detailed ionisation history at early times. \section{The relation of LAEs to LBGs} \label{sec_LAE_LBG_relation} \begin{figure*} \centering \includegraphics[width=0.99\textwidth]{2Dhistogram_numDens_median_LAE-LBG-relation-Lc_obs_zall.png} \caption{The two top rows depict the fraction of LBGs showing Ly$\alpha$ emissions with $L_\alpha\geq10^{42}$erg~s$^{-1}$ and $\mathrm{EW_\alpha}$ exceeding the value marked in the panels for the different Ly$\alpha$ line profile models as marked. The bottom row shows the corresponding medians of the $\mathrm{EW_\alpha}$ values (lines) and their $\sim1.3\sigma$ uncertainties (shaded regions). Solid (dashed dotted) lines show results for the reionisation scenario where $f_\mathrm{esc}$ decreases (increases) with halo mass $M_h$.} \label{fig_LAE_LBG} \end{figure*} In this Section, we address the question of what defines whether an LBG shows Ly$\alpha$ emission and why the fraction of LBGs with observable Ly$\alpha$ emission changes as the observed UV continuum luminosity (at $1500$\AA) or the minimum Ly$\alpha$ equivalent rise. For this purpose, we show both the fraction of LBGs with a Ly$\alpha$ equivalent width of at least EW$_\alpha\geq25$\AA~ (top row) and EW$_\alpha\geq50$\AA~ (central row) and the median EW$_\alpha$ value (bottom row) as a function of the UV continuum luminosity in Fig. \ref{fig_LAE_LBG}. For our three different Ly$\alpha$ line profile models, we find the median EW$_\alpha$ to exhibit similar values of $\sim5-30$\AA~ at all redshifts shown. Furthermore, the EW$_\alpha$ values range to lower values as galaxies become UV fainter. As galaxies become less massive, this spread in EW$_\alpha$ values reflects the increasingly broader range of star formation rate values to lower values, which traces back to the larger variety of mass assembly histories that increasingly include progenitors with particularly SN feedback-suppressed star formation. The $M_\mathrm{UV}$-dependency of the fraction of LBGs with Ly$\alpha$ emission ($f_\mathrm{LAE}$) also reflects this shift towards lower EW$_\alpha$ values (c.f. top and central row of Fig. \ref{fig_LAE_LBG}): firstly, $f_\mathrm{LAE}$ decreases towards lower UV luminosities, and secondly, this decrease is stronger for lower than higher EW$_\alpha$ cuts. These trends imply that UV bright galaxies are more likely to show higher EW$_\alpha$ values for all our Ly$\alpha$ line profile models and reionisation scenarios. For example, while only $<10\%$ of galaxies with $M_\mathrm{UV}\simeq-18$ exceed EW$_\alpha>25$\AA, $>40\%$ of galaxies with $M_\mathrm{UV}\gtrsim-20$ exceed EW$_\alpha>25$\AA~ and $>5\%$ even EW$_\alpha>50$\AA. Moreover, both the slight rise of EW$_\alpha$ and $f_\mathrm{LAE}$ values with decreasing redshift and their variation among our different reionisation scenarios can be attributed to the increasing fraction of Ly$\alpha$ radiation that is transmitted through the IGM as the Universe becomes more ionised (see $T_\alpha$ in Fig. \ref{fig_histograms}). For example, at a given redshift $z>7$ ($z<7$), the EW$_\alpha$ and $f_\mathrm{LAE}$ values are higher (lower) in the {\sc mhdec} than in the {\sc mhinc} scenario, which is due to a more (less) ionised IGM. Similarly, the lower EW$_\alpha$ values reached in the {\it Gaussian} model for UV fainter galaxies are due to the stronger absorption of Ly$\alpha$ radiation by \HI in the IGM. Finally, we note that since in the {\it Clumpy} and {\it Porous} models the attenuation of the UV continuum and Ly$\alpha$ by dust do not necessarily correlate with each other (as e.g. parts of Ly$\alpha$ can escape via random walk), a few galaxies that are attenuated strongly in the UV but less in Ly$\alpha$ show high EW$_\alpha$ values of $\sim1000$\AA. Thus, the main driver of these high EW$_\alpha$ values is the dust attenuation of the UV continuum assumed in our models. Comparing our fraction of LBGs showing Ly$\alpha$ emission with those obtained in observations \citep[e.g.][]{Schenker2012, Schenker2014, Caruana2014, Pentericci2014, Pentericci2018, Mason2019}, we find that (i) the observed trend of $f_\mathrm{LAE}$ decreasing towards higher UV luminosity agrees roughly with our results for EW$_\alpha>50$\AA~ but not for EW$_\alpha>25$\AA, and (ii) our $f_\mathrm{LAE}$ values are higher than those inferred from observations (again more so for EW$_\alpha>25$\AA~ than EW$_\alpha>50$\AA). These discrepancies hint either at our model predicting too high Ly$\alpha$ or too low UV luminosities (particularly for more massive galaxies) despite reproducing the observed Ly$\alpha$ and UV LFs, or observations missing bright LAEs. Interestingly, we find that the fraction of LBGs with high EW$_\alpha$ values of $f_\mathrm{LAE}(\mathrm{EW}_\alpha>100$\AA$)\simeq1-12\%$ and $f_\mathrm{LAE}(\mathrm{EW}_\alpha>240$\AA$)\simeq1-2\%$ in the {\it Clumpy} and {\it Porous} models are in rough agreement with the results from deep MUSE observations at $z=3-6$ that consider only LAEs with detected UV continuum \citep{Kerutt2022}. A higher abundance of high EW$_\alpha$ values has been found in various high-redshift LAE observations \citep[e.g.][]{Shibuya2018, Malhotra_Rhoads2002, shimasaku2006}. Nevertheless, our $f_\mathrm{LAE}$ values agree roughly with the results from radiative hydrodynamical simulations post-processed with Ly$\alpha$ radiative transfer, such as {\sc sphinx} \citep[c.f. Fig. B1 in][]{Garel2021}. \section{Conclusions} \label{sec_conclusions} We apply our new framework for LAEs to different reionisation scenarios, and analyse how the escape fraction of \HI ionising photons, $f_\mathrm{esc}$, and its dependence on halo mass affect the luminosity-dependent number and spatial distributions of LAEs. Besides $f_\mathrm{esc}$ affecting the IGM ionisation topology and the strength of the Ly$\alpha$ line produced in the ISM, its sensitivity to the density and velocity structure of ISM gas and dust has been found to correlate with the Ly$\alpha$ line profile emerging from a galaxy and the fraction of Ly$\alpha$ radiation escaping into the IGM. Notably, the emerging Ly$\alpha$ line profile reflects the attenuation by dust in the ISM and can also change the fraction of Ly$\alpha$ radiation that traverses the IGM unattenuated by \HI. For this reason, we build an analytical model for Ly$\alpha$ line profiles that emerge from a Ly$\alpha$ source surrounded by an outflowing shell of dusty gas clumps interspersed with low-density channels. Our model reproduces the numerical radiative transfer results of a shell with dust gas clumps of different sizes as presented in \citet{Gronke2017}. By coupling this model to {\sc astraeus}, a semi-numerical model coupling galaxy evolution and reionisation self-consistently, we derive the Ly$\alpha$ line profiles emerging from the simulated galaxy population and explore the resulting large-scale distribution of LAEs for different dependencies of $f_\mathrm{esc}$ on halo mass (decreasing, constant, increasing) and Ly$\alpha$ line profiles (Gaussian profile, outflowing shell of dusty clumps interspersed with low-density channels or not). For this parameter space, we analyse the resultant ionisation topologies, the dependencies of Ly$\alpha$ line profiles and Ly$\alpha$ properties on halo mass, and the location of galaxies with observable Ly$\alpha$ emission in the large-scale structure. Our main results are the following: \begin{enumerate} \item For an outflowing shell consisting of clumps of the same size, the Ly$\alpha$ line profile emerging from a galaxy develops from a central peak at the Ly$\alpha$ resonance to a double peak profile as it becomes more massive. Adding low-density channels results in either an enhancement or weakening of this trend as $f_\mathrm{esc}$ decreases or increases with rising halo mass. \item In all reionisation scenarios and Ly$\alpha$ line profile models, LAEs are more massive galaxies with $M_h\gtrsim10^{10}\msun$. These galaxies exhibit continuous star formation and are biased tracers of the underlying mass density distribution. Both allow efficient transmission of the Ly$\alpha$ line through the IGM by facilitating the build-up of ionised regions around them. In contrast, less massive galaxies are surrounded by smaller ionised regions, which results in their Ly$\alpha$ radiation being significantly attenuated by \HI in the IGM. \item As LAEs are more massive galaxies and the most biased tracers of the underlying mass density distribution, they are located in the densest and most highly ionised regions. This finding holds for any inside-out reionisation scenario where dense regions containing massive galaxies are ionised before under-dense voids \citep[see also][]{Hutter2014, Hutter2017}. In such scenarios, the spatial distribution of LAEs is primarily sensitive to the global ionisation fraction and only in second-order to the ionisation topology or the trend of $f_\mathrm{esc}$ with halo mass. \item As the observable Ly$\alpha$ LFs are composed of the Ly$\alpha$ emission from more massive galaxies, a decrease in their intrinsic Ly$\alpha$ luminosities (Ly$\alpha$ produced in the ISM) due to higher $f_\mathrm{esc}$ values can be compensated by reducing the attenuation by dust in the ISM \citep[echoing the degeneracy found in][]{Hutter2014}. However, if $f_\mathrm{esc}$ exceeds $\sim0.5$ for the most massive galaxies ($M_h\gtrsim10^{11}\msun$), their intrinsic Ly$\alpha$ luminosity is too low to reproduce the observed Ly$\alpha$ LFs \citep[see also][]{Hutter2014}. \end{enumerate} All combinations of our reionisation scenarios and Ly$\alpha$ line profile models result in Ly$\alpha$ and UV luminosities in reasonable agreement with observational constraints. However, although two of the three Ly$\alpha$ line profile models investigated use parameterisations of numerical Ly$\alpha$ radiative transfer simulation results, they represent idealised scenarios where the gas in each galaxy is distributed in clumps of the same mass and moves outwards at the same velocity. In reality, the density and velocity distributions of gas and dust in the ISM are more complex: Firstly, the dusty gas clumps will have different masses, with a distribution close to that of a scale-free one at the massive end. Such a mass distribution would result in more massive galaxies having larger clumps than less massive galaxies, which again would lead to a homogenisation of their Ly$\alpha$ line profiles where more massive (less massive) galaxies have an enhanced (weakened) central peak component and a weakened (enhanced) double-peak component. This change in the Ly$\alpha$ line profiles would result in the Ly$\alpha$ radiation being less (more) attenuated by dust in the ISM and traversing the IGM more (less) efficiently. Secondly, the low-density channels might not be gas-free or fully ionised, giving rise to a narrower double-peak profile instead of the central peak profile used in this work. Additionally, the gas may exhibit a turbulent velocity structure that could broaden the double-peak component. Both partially neutral low-density channels and an inhomogeneous velocity structure are likely to enhance the transmission of Ly$\alpha$ through the IGM. Thirdly, the attenuation of Ly$\alpha$ radiation by dust in the ISM depends on the distribution of dust in clumps. While our model assumes that gas and dust are perfectly mixed (resulting in an absorption probability per clump of $\epsilon\propto 1-\exp(-\tau_\mathrm{d,cl})$), a scenario where dust condensates in the centre surrounded by a shell of hydrogen gas would lower the absorption probability per clump and enhance the escape fraction of Ly$\alpha$ photons from a galaxy. While our Ly$\alpha$ line profile models are limited by the simplified structure assumed for the ISM, they cover the extreme cases of a central peak profile and a double-peak profile for more massive galaxies. Since these more massive galaxies drive the observed Ly$\alpha$ LFs, models accounting for a more detailed ISM structure are likely to lie in between and not change our key findings. In particular, the insensitivity of the spatial distribution of LAEs to any dependence of $f_\mathrm{esc}$ with halo mass suggests that LAEs alone can not help to constrain any gradual dependence of $f_\mathrm{esc}$ with galactic properties. Any dependency introduced in the intrinsic Ly$\alpha$ luminosity we can compensate by the opposed trend of the Ly$\alpha$ escape fraction, achieved by changing the ISM gas and dust distribution. This insensitivity to $f_\mathrm{esc}$ dependencies makes LAEs relatively robust tracers of the underlying density field that we can use to pin down the ionisation topology. Constraining $f_\mathrm{esc}$ during the EoR will require a combination of ionisation topology measurements through the \HI 21cm signal and measurements of other emission lines. \section*{Acknowledgements} We thank Max Gronke and Peter Laursen for useful discussions. AH, GY, LL, PD and SG acknowledge support from the European Research Council's starting grant ERC StG-717001 (``DELPHI"). AH, MT, PD also acknowledge support from the NWO grant 016.VIDI.189.162 (``ODIN") and the European Commission's and University of Groningen's CO-FUND Rosalind Franklin program. PD thanks the Institute for Advanced Study (IAS) Princeton, where a part of this work was carried out, for their generous hospitality and support through the Bershadsky Fund. GY acknowledges financial support from MICIU/FEDER under project grant PGC2018-094975-C21. We thank Peter Behroozi for creating and providing the {\sc rockstar} merger trees of the {\sc vsmdpl} and {\sc esmdpl} simulations. The authors wish to thank V. Springel for allowing us to use the L-Gadget2 code to run the different Multidark simulation boxes, including the {\sc vsmdpl} and {\sc esmdpl} used in this work. The {\sc vsmdpl} and {\sc esmdpl} simulations have been performed at LRZ Munich within the project pr87yi. The authors gratefully acknowledge the Gauss Centre for Supercomputing e.V. (www.gauss-centre.eu) for funding this project by providing computing time on the GCS Supercomputer SUPERMUC-NG at Leibniz Supercomputing Centre (www.lrz.de). The CosmoSim database (\url{www.cosmosim.org}) provides access to the simulations and the Rockstar data. The database is a service by the Leibniz Institute for Astrophysics Potsdam (AIP). The Cosmic Dawn Center (DAWN) is funded by the Danish National Research Foundation under grant No. 140. This research made use of \texttt{matplotlib}, a Python library for publication quality graphics \citep{hunter2007}; and the Python library \texttt{numpy} \citep{numpy}. \section*{Data Availability} The source code of the semi-numerical galaxy evolution and reionisation model within the {\sc astraeus} framework and the employed analysis scripts are available on GitHub (\url{https://github.com/annehutter/astraeus}). The underlying N-body DM simulation, the {\sc astraeus} simulations and derived data in this research will be shared on reasonable request to the corresponding author. \bibliographystyle{mnras}

817c40acb6ae629307ae69a4b2a307daa30935a1

\section{Introduction} Developing machines equipped with mathematical reasoning capabilities is one of the long-standing goals of artificial intelligence. Solving math word problems (MWPs) is a well-defined task to diagnose the ability of intelligent systems to perform numerical reasoning and problem-solving as humans. A surge of datasets has been proposed to facilitate the research in this domain \citep{upadhyay2017annotating,amini2019mathqa,miao2020diverse,cobbe2021training}. However, most existing MWP datasets focus on textual math word problems only. Tables, widely distributed in different documents such as invoices, health records, and financial reports, contain rich structured information different from unstructured text. Solving math word problems in such a tabular context is much more challenging than existing MWP benchmarks since the system needs to make cell selections and align heterogeneous information before performing further numerical reasoning. \begin{figure}[t!] \centering \vspace{-5mm} \includegraphics[width=1.0\textwidth]{figures/fig_dataset.pdf} \caption{Two examples from the \textsc{TabMWP}\xspace dataset. The example above is a \textit{free-text} problem with a numerical answer; the example below is a \textit{multi-choice} problem with a textual answer.} \vspace{-5mm} \label{fig:dataset} \end{figure} To fill this gap, we propose \text{Tab}ular \text{M}ath \text{W}ord \text{P}roblems (\textsc{TabMWP}\xspace{}), a new large-scale dataset that contains 38,431 math word problems with tabular context, taken from grade-level math curricula. There are two question types: \textit{free-text} questions in which the answer is an integer or decimal number, and \textit{multi-choice} questions where the answer is a text span chosen from option candidates. Different from existing MWP datasets, each problem in \textsc{TabMWP}\xspace{} is accompanied by a tabular context, which is represented in three formats: an image, a semi-structured text, and a structured table. Each problem is also annotated with a detailed solution that reveals the multi-step reasoning steps to ensure full explainability. To solve problems in \textsc{TabMWP}\xspace, a system requires multi-hop mathematical reasoning over heterogeneous information by looking up table cells given textual clues and conducting multi-step operations to predict the final answer. Take the problem above in Figure \ref{fig:dataset} as an example. To answer the question ``\textit{how much will she spend (if Tracy buys three kinds of beads)}?'', we first need to look up the corresponding three rows in the given table, calculate the individual cost for each kind of bead, and finally sum three costs up to get the answer of 31.44. Inspired the success of the large pre-trained language model GPT-3 \citep{chen2020big} in solving math word problems \citep{wei2022chain,wang2022self}, we first build a strong baseline using few-shot GPT-3 on \textsc{TabMWP}\xspace{}. A few in-context examples are randomly selected from the training set, along with the test example, and are constructed as a prompt for GPT-3 to predict the answer. However, recent studies have shown that this type of few-shot learning can be highly unstable across different selections of in-context examples \citep{zhao2021calibrate, liu2022makes, lu2022fantastically}. It could be worse on \textsc{TabMWP}\xspace{} since its problems are distributed across multiple question types and diverse table layouts. \cite{liu2022makes} try to address this issue by retrieving semantically similar examples. However, this method might not work well on \textsc{TabMWP}\xspace{} because it is not capable of measuring the similarity of structured information, such as the number of cells in tables. To alleviate this challenge, we further propose a novel approach that can learn to select in-context examples from a small amount of training data via policy gradient for prompt learning, termed \textsc{PromptPG}\xspace. As illustrated in Figure \ref{fig:model}, an agent learns to find optimal in-context examples from a candidate pool, with the goal of maximizing the prediction rewards on given training examples when interacting with the GPT-3 environment. A policy network defines the strategy of how to select the in-context examples given the current training example. The policy network is built on top of the language model BERT \citep{devlin2018bert} with fixed parameters, followed by a one-layer linear neural network with learnable parameters. The learnable parameters are updated following the policy gradient strategy \citep{sutton1998introduction}. Unlike random selection \citep{wei2022chain,wang2022self}, brute-force search, or retrieval-based selection \citep{liu2022makes}, \textsc{PromptPG}\xspace learns to construct the prompt dynamically given the candidate pool when interacting with the GPT-3 API. We implement two state-of-the-art methods as baselines, i.e., UnifiedQA \citep{khashabi2020unifiedqa} on general question answering and TAPEX \citep{liu2022tapex} on tabular question answering. Both are implemented in pre-trained and fine-tuned settings. Experimental results show that our model \textsc{PromptPG}\xspace can achieve an overall accuracy of 68.23\% on \textsc{TabMWP}\xspace, which greatly surpasses previous methods by a large margin of up to 5.31\%. Further analysis demonstrates that \textsc{PromptPG}\xspace can select better in-context examples compared with a wide range of existing selection strategies and reduce the prediction variance significantly compared to random selection. The main contributions of our work are as follows: (a) We present a new large-scale dataset, \textsc{TabMWP}\xspace, the first dataset for math word problems with tabular context; (b) We propose a novel approach, \textsc{PromptPG}\xspace, which learns the prompt dynamically via policy gradient to select in-context examples for few-shot GPT-3. To the best of our knowledge, it is the first work that applies reinforcement learning to select in-context examples for the few-shot GPT-3 model; (c) Experimental results show that \textsc{PromptPG}\xspace achieves an improvement of up to 5.31\% on \textsc{TabMWP}\xspace over existing methods, with reduced selection instability compared to random selection. \begin{figure}[t!] \centering \vspace{-5mm} \includegraphics[width=0.9\textwidth]{figures/fig1_model.pdf} \caption{Our proposed \textsc{PromptPG}\xspace is able to learn to select performing in-context examples via policy gradient when interacting with the GPT-3 API without any manually designed heuristics.} \vspace{-3mm} \label{fig:model} \end{figure} \section{The \textsc{TabMWP}\xspace Dataset} \subsection{Task Formulation} A tabular math word problem $p$ is represented as a pair ($t$, $q$), where $t$ is a table context and $q$ is a question. The table $t$ could be represented in a visual format as an image, semi-structured text, or a structured database. In this work, we focus on the semi-structured format as the table context for simplicity. The table $t$ features complicated layouts and formats: it contains multiple rows and columns, and each cell can be a string of text, a string of a number, or a mix of them. Depending on the question and answer types, the question $q$ may be accompanied by multiple-choice options $c=\{c_1, c_2, \dots, c_n\}$ or a unit $u$. Given a semi-structured tabular context $t$ and an unstructured question text $q$, the task is to generate the answer $a$, which is either numerical only text for a \textit{free-text} question, or a text span from given options for a \textit{multiple-choice} question. \subsection{Dataset Construction} \textbf{Data collection.} We construct \textsc{TabMWP}\xspace{} based on openly available content and more details are provided in Appendix \ref{appx:dataset}. Only math word problems that are accompanied by a tabular context and a detailed solution are collected. We develop a script to extract the tabular context, the question, options that apply, the correct answer, and the solution for each problem. These elements can be precisely identified using HTML tags. For each table, we take a screenshot and store its raw text. \textbf{Data preprocessing.} To make \textsc{TabMWP}\xspace{} compatible with various baselines, we represent the tabular context as three formats: an image, \textit{semi-structured} text, and a \textit{structured} spreadsheet. The semi-structured format is created by converting the raw table text into a flattened token sequence, with each row separated by a newline character `\texttt{$\backslash$n}' and each column separated by `\texttt{$\mid$}'. The semi-structured text is further transformed to the structured format, which can be easily retrieved and executed by SQL-based methods \citep{liu2022tapex} using packages like \texttt{pandas}. For clarity, the table title is separated from the raw table. Examples of three formats are shown in Appendix \ref{appx:dataset}. For better quantitative evaluation, we formalize the \textsc{TabMWP}\xspace{} problems as two question types: (a) \textit{free-text} questions, where the answer is numerical text only and the unit text is separately extracted; and (b) \textit{multi-choice} questions, the answer of which is the text span from choice options. The order of choice options is shuffled to alleviate distribution bias. Redundant information in solutions is removed, and some solutions are manually rewritten to be more human-readable. Finally, problems with the same table, question, and answer text are regarded as redundant and thus removed. We further conduct quality control to ensure data quality, which is discussed in Appendix \ref{appx:dataset}. \subsection{Dataset Statistics} \begin{wraptable}{r}{0.42\textwidth} \vspace{-3mm} \centering \renewcommand\tabcolsep{1.5pt} \small \begin{tabular}{lr} \toprule \textbf{Statistic} & \textbf{Number} \\ \midrule Total questions & 38,431 \\ ~~* \textit{free-text} questions & 28,719 \\ ~~* \textit{multi-choice} questions & 9,712 \\ \midrule \# of different questions & 28,876 \\ \# of different answers & 6,153 \\ \# of different solutions & 35,442 \\ \midrule \# of different tables & 37,644 \\ \# of tables with a title & 23,259 \\ \midrule \# of table cells (Average/Max) & 12.9 / 54 \\ \# of table rows (Average/Max) & 5.9 / 11 \\ \# of table columns (Average/Max) & 2.2 / 6 \\ \midrule Question length (Average/Max) & 22.1 / 92 \\ Answer length (Average/Max) & 1.1 / 27 \\ Solution length (Average/Max) & 49.5 / 350 \\ \bottomrule \end{tabular} \captionof{table}{Key statistics for \textsc{TabMWP}\xspace{}.} \vspace{-5mm} \label{tab:statistics} \end{wraptable} \textbf{Key statistics.} The \textsc{TabMWP}\xspace dataset contains 38,431 tabular math word problems, which are partitioned with 6:2:2 into the training, development, and test splits, corresponding to 23,059, 7,686, and 7,686 problems. Their main statistics are shown in Table \ref{tab:statistics}. 74.7\% of the questions in \textsc{TabMWP}\xspace belong to \textit{free-text} questions, while 25.3\% are \textit{multi-choice} questions. There are 28,876 different questions, 6,153 different answers, and 35,442 different solutions, indicating that \textsc{TabMWP}\xspace has a rich diversity in the problem distribution. The questions have an average of 22.1 words in length and solutions of 49.5, showing that they have lexical richness. One distinct characteristic of \textsc{TabMWP}\xspace is that each problem is accompanied by a tabular context, without which the problem would be unsolvable. There are 37,644 different tables in total, and 60.5\% of the tables have a title. The table has an average of 5.9 rows and 2.2 columns, which results in an average of 12.9 cells and a maximum of 54 cells. These statistics suggest that tables in \textsc{TabMWP}\xspace distribute diversely across semantics and layouts. \textbf{Comparison to existing datasets.} As shown in Table \ref{tab:datasets}, \textsc{TabMWP}\xspace differs from related datasets in various aspects: (1) \textsc{TabMWP}\xspace is the first dataset to study math word problems over tabular context on open domains and is the largest in terms of data size; (2) Problems in \textsc{TabMWP}\xspace are annotated with the tabular context, unlike previous MWP datasets in the first segment; (3) Different from Table QA datasets like FinQA, TAT-QA, and MultiHiertt, a lack of either mathematical reasoning or the tabular context renders the problems in \textsc{TabMWP}\xspace unanswerable; (4) There are two question types in \textsc{TabMWP}\xspace, and the answer could be a text span, an integer number, or a decimal number; (5) Each problem is annotated with natural language solutions to reveal multi-hop reasoning steps. \begin{figure}[th!] \centering \renewcommand\tabcolsep{4.0pt} \resizebox{1.0\linewidth}{!}{ \begin{tabular}{p{2.8cm}rrcccccccccc} \toprule \multirow{3}{*}{Dataset} & \multirow{3}{*}{Size} & \multirow{3}{*}{\#Table} & \multirow{2}{*}{Need} & \multirow{2}{*}{Need} & \multicolumn{2}{c}{Table Type} & \multicolumn{2}{c}{Question Type} & \multicolumn{3}{c}{Answer Type} & \multirow{2}{*}{Solution} \\ \cmidrule(lr){6-7} \cmidrule(lr){8-9} \cmidrule(lr){10-12} & & & Math? & Table? & Domain & Format & Free-text & MC & Text & Integer & Decimal & Type \\ \midrule Dolphin18K \citeyearpar{huang2016well} & 831 & \textcolor{red!50!black}{\ding{55}}~~~ & \textcolor{green!50!black}{\ding{51}} & \textcolor{red!50!black}{\ding{55}} & \textcolor{red!50!black}{\ding{55}} & \textcolor{red!50!black}{\ding{55}} & \textcolor{green!50!black}{\ding{51}} & \textcolor{red!50!black}{\ding{55}} & \textcolor{red!50!black}{\ding{55}} & \textcolor{green!50!black}{\ding{51}} & \textcolor{green!50!black}{\ding{51}} & formula \\ DRAW-1K \citeyearpar{upadhyay2017annotating} & 1,000 & \textcolor{red!50!black}{\ding{55}}~~~ & \textcolor{green!50!black}{\ding{51}} & \textcolor{red!50!black}{\ding{55}} & \textcolor{red!50!black}{\ding{55}} & \textcolor{red!50!black}{\ding{55}} & \textcolor{green!50!black}{\ding{51}} & \textcolor{red!50!black}{\ding{55}} & \textcolor{red!50!black}{\ding{55}} & \textcolor{green!50!black}{\ding{51}} & \textcolor{green!50!black}{\ding{51}} & formula \\ Math23K \citeyearpar{wang2017deep} & 23,162 & \textcolor{red!50!black}{\ding{55}}~~~ & \textcolor{green!50!black}{\ding{51}} & \textcolor{red!50!black}{\ding{55}} & \textcolor{red!50!black}{\ding{55}} & \textcolor{red!50!black}{\ding{55}} & \textcolor{green!50!black}{\ding{51}} & \textcolor{red!50!black}{\ding{55}} & \textcolor{red!50!black}{\ding{55}} & \textcolor{green!50!black}{\ding{51}} & \textcolor{green!50!black}{\ding{51}} & formula \\ MathQA \citeyearpar{amini2019mathqa} & 37,297 & \textcolor{red!50!black}{\ding{55}}~~~ & \textcolor{green!50!black}{\ding{51}} & \textcolor{red!50!black}{\ding{55}} & \textcolor{red!50!black}{\ding{55}} & \textcolor{red!50!black}{\ding{55}} & \textcolor{red!50!black}{\ding{55}} & \textcolor{green!50!black}{\ding{51}} & \textcolor{red!50!black}{\ding{55}} & \textcolor{green!50!black}{\ding{51}} & \textcolor{green!50!black}{\ding{51}} & formula \\ ASDiv \citeyearpar{miao2020diverse} & 2,305 & \textcolor{red!50!black}{\ding{55}}~~~ & \textcolor{green!50!black}{\ding{51}} & \textcolor{red!50!black}{\ding{55}} & \textcolor{red!50!black}{\ding{55}} & \textcolor{red!50!black}{\ding{55}} & \textcolor{green!50!black}{\ding{51}} & \textcolor{red!50!black}{\ding{55}} & \textcolor{green!50!black}{\ding{51}} & \textcolor{green!50!black}{\ding{51}} & \textcolor{green!50!black}{\ding{51}} & formula \\ SVAMP \citeyearpar{patel2021nlp} & 1,000 & \textcolor{red!50!black}{\ding{55}}~~~ & \textcolor{green!50!black}{\ding{51}} & \textcolor{red!50!black}{\ding{55}} & \textcolor{red!50!black}{\ding{55}} & \textcolor{red!50!black}{\ding{55}} & \textcolor{green!50!black}{\ding{51}} & \textcolor{red!50!black}{\ding{55}} & \textcolor{red!50!black}{\ding{55}} & \textcolor{green!50!black}{\ding{51}} & \textcolor{red!50!black}{\ding{55}} & formula \\ GSM8K \citeyearpar{cobbe2021training} & 8,792 & \textcolor{red!50!black}{\ding{55}}~~~ & \textcolor{green!50!black}{\ding{51}} & \textcolor{red!50!black}{\ding{55}} & \textcolor{red!50!black}{\ding{55}} & \textcolor{red!50!black}{\ding{55}} & \textcolor{green!50!black}{\ding{51}} & \textcolor{red!50!black}{\ding{55}} & \textcolor{red!50!black}{\ding{55}} & \textcolor{green!50!black}{\ding{51}} & \textcolor{red!50!black}{\ding{55}} & text \\ IconQA \citeyearpar{lu2021iconqa} & \underline{107,439} & \textcolor{red!50!black}{\ding{55}}~~~ & \textcolor{green!50!black}{\ding{51}} & \textcolor{red!50!black}{\ding{55}} & \textcolor{red!50!black}{\ding{55}} & \textcolor{red!50!black}{\ding{55}} & \textcolor{green!50!black}{\ding{51}} & \textcolor{green!50!black}{\ding{51}} & \textcolor{green!50!black}{\ding{51}} & \textcolor{green!50!black}{\ding{51}} & \textcolor{red!50!black}{\ding{55}} & \textcolor{red!50!black}{\ding{55}} \\ \midrule FinQA \citeyearpar{chen2021finqa} & 8,281 & 2,766 & \textcolor{green!50!black}{\ding{51}} & 76.6\% & finance & text & \textcolor{green!50!black}{\ding{51}} & \textcolor{red!50!black}{\ding{55}} & \textcolor{red!50!black}{\ding{55}} & \textcolor{green!50!black}{\ding{51}} & \textcolor{green!50!black}{\ding{51}} & program \\ TAT-QA \citeyearpar{zhu2021tat} & 16,552 & 2,747 & 50.0\% & \textcolor{green!50!black}{\ding{51}} & finance & text & \textcolor{green!50!black}{\ding{51}} & \textcolor{red!50!black}{\ding{55}} & \textcolor{red!50!black}{\ding{55}} & \textcolor{green!50!black}{\ding{51}} & \textcolor{green!50!black}{\ding{51}} & \textcolor{red!50!black}{\ding{55}} \\ MultiHiertt \citeyearpar{zhao2022multihiertt} & 10,440 & 9,843 & \textcolor{green!50!black}{\ding{51}} & 89.8\% & finance & text & \textcolor{green!50!black}{\ding{51}} & \textcolor{red!50!black}{\ding{55}} & \textcolor{red!50!black}{\ding{55}} & \textcolor{green!50!black}{\ding{51}} & \textcolor{green!50!black}{\ding{51}} & \textcolor{red!50!black}{\ding{55}} \\ \midrule \textbf{\textsc{TabMWP}\xspace{} (ours)} & \textbf{38,431} & \textbf{37,644} & \textcolor{green!50!black}{\ding{51}} & \textcolor{green!50!black}{\ding{51}} & \textbf{open} & \textbf{text*} & \textcolor{green!50!black}{\ding{51}} & \textcolor{green!50!black}{\ding{51}} & \textcolor{green!50!black}{\ding{51}} & \textcolor{green!50!black}{\ding{51}} & \textcolor{green!50!black}{\ding{51}} & \textbf{text}\\ \bottomrule \end{tabular} } \captionof{table}{A comparison of MWP and Table QA datasets that require numerical reasoning. \textit{text*}: each table in \textsc{TabMWP}\xspace{} is accompanied by an image format.} \vspace{-2mm} \label{tab:datasets} \end{figure} \section{Methods} \subsection{Few-shot GPT-3 for \textsc{TabMWP}\xspace{}} Provided with a few in-context examples of math word problems as the context, GPT-3 can generate the answer for a test problem, and show impressive performance across different MWP datasets \citep{wei2022chain,wang2022self}. Inspired by its success, we first build a strong baseline using few-shot GPT-3 on our \textsc{TabMWP}\xspace{} dataset. Specifically, a few training examples, along with the test example $p_i$, are provided to GPT-3 for the answer prediction. Each training example consists of a table context $t$, a question $q$, options $c$ that apply, and an answer $a$. To make the few-shot GPT-3 model workable on \textsc{TabMWP}\xspace{}, we utilize the semi-structured format as the tabular context. Following \cite{wei2022chain}, a solution $s$ can be augmented in front of the answer $a$ to reveal the multi-step reasoning process, which is able to boost the prediction performance. \subsection{Dynamic Prompting via Policy Gradient} The in-context examples can be randomly \citep{wei2022chain, wang2022self} or retrieval-based selected \citep{liu2022makes} from the training set. Recent research, however, has shown that few-shot GPT-3 can be highly unstable across different selections of in-context examples and permutations of those examples \citep{zhao2021calibrate, liu2022makes, lu2022fantastically}. This instability may be more severe on \textsc{TabMWP}\xspace{}, where examples are more distinct because they include both unstructured questions of various types and semi-structured tables in various layouts. To alleviate this issue, we aim to propose a novel approach that can learn to select performing in-context examples using a policy gradient strategy, without brute-force searching or manually designed heuristics. Formally, given a \textsc{TabMWP}\xspace{} problem $p_i$, we want the agent to find $K$ in-context examples $e_i = \{e_i^1, e_i^2,...,e_i^K\}$ from a candidate pool $E_{\text{cand}}$, and generate the answer $\hat{a}_i$, maximizing a reward $r_i = R(\hat{a}_i | p_i)$. The in-context examples are selected according to a policy \begin{equation} e_i^k \sim \pi_{\theta}(e_i|p_i), ~e_i^k \in E_{\text{cand}}, e_i^k ~\text{are independent for} ~k=\{1,2,...,K\}, \end{equation} where $\theta$ are the policy's parameters. The answer is generated through: $\hat{a}_i = \text{GPT-3}(e_i, p_i)$ using the selected examples and the given problem as the input prompt. The reward is then computed by evaluating the generated answer $\hat{a}_i$ with respect to the ground truth answer $a_i$: \begin{equation} r_i = R(\hat{a}_i | p_i) = \textsc{Eval}(\hat{a}_i, a_i), ~r_i \in \{-1, 1\}. \end{equation} The function $\textsc{Eval}()$ returns a reward of $1$ if the generated answer aligned with the label and $-1$ otherwise. Our goal is to maximize the expected reward of the generated answer under the policy $\mathbb{E}_{e_i \sim \pi_{\theta}(e_i|p_i)}[R(\text{GPT-3}(e_i,p_i))]$. We optimize the reward with respect to the parameters of the policy network using the Policy Gradient method~\citep{sutton1998introduction}. The expected reward cannot be computed in closed form, so we compute an unbiased estimation with Monte Carlo Sampling, \begin{equation} \mathbb{E}_{e_i \sim \pi_{\theta}(e_i|p_i)}\left[R(\text{GPT-3}(e_i,p_i))\right] \approx \frac{1}{N}\sum_{i=1}^{N}R(\text{GPT-3}(e_i,p_i)), ~e_i \sim \pi_{\theta}(e_i|p_i), \end{equation} where $N$ is the size of each batch yielded from our training problem set $P_{\text{train}}$. In this work, we experiment using the REINFORCE policy gradient algorithm~\citep{williams1992simple}: \begin{equation} \fontsize{9.5pt}{\baselineskip}\selectfont \begin{aligned} \nabla \mathbb{E}_{e_i \sim \pi_{\theta}(e_i|p_i)}\left[R(\text{GPT-3}(e_i,p_i))\right] &= \mathbb{E}_{e_i \sim \pi_{\theta}(e_i|p_i)} \nabla_{\theta}\log(\pi_{\theta}(e_i|p_i))R(\text{GPT-3}(e_i,p_i)) \\ &\approx \frac{1}{N}\sum_{i=1}^N\nabla_{\theta}\log(\pi_{\theta}(e_i|p_i))R(\text{GPT-3}(e_i,p_i)), ~e_i \sim \pi_{\theta}(e_i|p_i). \end{aligned} \end{equation} Intuitively, if the predicted answer is correct, we update the policy so that the probability of selecting the same prompts gets higher. Otherwise, we update the policy to reduce the probability of selecting such less matched examples. The learning process is summarized in Algorithm \ref{alg:policy} in the appendix. To get the contextualized representation of the given problem and candidate examples, we use the BERT~\citep{devlin2018bert} \texttt{[CLS]} token representation as the problem encoding. We add a small linear layer on top of the BERT final pooling layer. That allows our model to learn both the semantic similarity that the pre-trained BERT model provides and the hidden logical similarity shared among the math problems. During training, the parameters of BERT are fixed and only the appended linear layer is updated, i.e., $\theta$ is composed of the learnable parameters $\mathbf{W}$ and $\mathbf{b}$: \begin{equation} \begin{aligned} \mathbf{h}(e_i) &= \mathbf{W}(\textsc{BERT}(e_i)) + \mathbf{b}, \\ \mathbf{h}(p_i) &= \mathbf{W}(\textsc{BERT}(p_i)) + \mathbf{b}, \\ \pi_{\theta}(e_i|p_i) &= \frac{\exp{[\mathbf{h}(e_i) \cdot \mathbf{h}(p_i)}]}{\sum_{e_i' \in E_{\text{cand}}} \exp{[\mathbf{h}(e_i') \cdot \mathbf{h}(p_i)}]}. \end{aligned} \end{equation} \section{Experiments} \subsection{Experimental Settings} \textbf{Baselines.} We first develop two large language models, UnifiedQA \citep{khashabi2020unifiedqa} and TAPEX \citep{liu2022tapex}, in both pre-trained and fine-tuned settings, as strong baselines on \textsc{TabMWP}\xspace. Different model sizes are included to examine the performance across different model capacities. We further implement the zero-shot GPT-3 model, the few-shot GPT-3 model, and their chain-of-thought (CoT) reasoning variants \citep{wei2022chain}. We also study the heuristic guess baseline and human performance to analyze the lower and upper bounds on \textsc{TabMWP}\xspace, respectively. \textbf{Evaluation metric.} The answer part is extracted from the GPT-3 generation using manually designed regular regressions. To evaluate the baselines and our method, we utilize the accuracy metric to determine if the generated answer is correct given the ground truth answer. For \textit{free-text} problems where the answer is set as a number, we normalize the prediction and the label to decimal numbers with two-digit precision and check if their values are equivalent. For \textit{multi-choice} problems, we choose the most similar one from options to the generated answer following \cite{khashabi2020unifiedqa}. \textbf{Implementation details.} Fine-tuned UnifiedQA and TAPEX baselines are trained on the train split and evaluated on the test split. Few-shot GPT-3 and few-shot-CoT GPT-3 randomly select two in-context examples from the training data to build the prompt. Our \textsc{PromptPG}\xspace is built on top of few-shot GPT-3 with a different selection strategy: (a) in the training stage, the agent learns to select two examples from 20 candidates and is evaluated on 160 training examples to calculate the reward; (b) in the test stage, the agent with an optimal policy chooses two examples from 20 candidates for each test example. The candidates are randomly selected from the training set. Experiments for two few-shot GPT-3 baselines and our \textsc{PromptPG}\xspace are repeated three times, and the average accuracy is reported in Table \ref{tab:results}. More implementation details can be found in Appendix \ref{appx:details}. \subsection{Experimental Results} Table~\ref{tab:results} demonstrates the results of different baselines and our method on the \textsc{TabMWP}\xspace{} dataset. Benefiting from pre-training on the tabular corpus, the TAPEX baseline performs better on average than UnifiedQA with a similar model size, which is only pre-trained on unstructured textual data. Increasing the model size can improve the prediction accuracy for both UnifiedQA and TAPEX. Fine-tuned on \textsc{TabMWP}\xspace{}, the baseline models can significantly improve the prediction performance on the average and all aggregated accuracy metrics. \begin{figure}[t!] \centering \renewcommand\tabcolsep{3.9pt} \renewcommand1.5{1.1} \resizebox{1.0\linewidth}{!}{ \begin{tabular}{lcccccccccccl} \toprule \multirow{3}{*}{\textbf{Method}} & \multirow{2}{*}{\textbf{Training}} & \multirow{2}{*}{\textbf{Selection}} & \multicolumn{2}{c}{\textbf{Question Types}} & \multicolumn{5}{c}{\textbf{Answer Types}} & \multicolumn{2}{c}{\textbf{Grades}} & \multirow{3}{*}{\textbf{~Avg.}} \\ \cmidrule(lr){4-5} \cmidrule(lr){6-10} \cmidrule(lr){11-12} & \textbf{Data} & \textbf{Strategy} & FREE & MC & INT & DEC & EXTR & BOOL & OTH & 1-6 & 7-8 & \\ \midrule \rowcolor[rgb]{0.93,0.93,0.93} \multicolumn{13}{l}{\textit{Heuristic Baselines}} \\ Heuristic guess & - & - & 6.71 & 39.81 & 8.37 & 0.26 & 30.80 & 51.22 & 26.67 & 17.55 & 12.27 & 15.29 \\ Human performance & - & - & \underline{84.61} & \underline{93.32} & \underline{84.95} & \underline{83.29} & \underline{97.18} & \underline{88.69} & \underline{96.20} & \underline{94.27} & \underline{81.28} & \underline{90.22} \\ \rowcolor[rgb]{0.93,0.93,0.93} \multicolumn{13}{l}{\textit{pre-trained Baselines}} \\ UnifiedQA$_{\textsc{Small}}$ & - & - & 1.18 & 43.62 & 1.37 & 0.43 & 38.70 & 49.78 & 37.14 & 15.57 & 7.65 & 12.18 \\ UnifiedQA$_{\textsc{Base}}$ & - & - & 4.60 & 43.02 & 5.28 & 1.97 & 37.08 & 50.11 & 38.10 & 17.14 & 11.11 & 14.56 \\ UnifiedQA$_{\textsc{Large}}$ & - & - & 4.48 & \underline{48.80} & 5.19 & 1.72 & \underline{48.33} & \underline{50.33} & \underline{40.00} & 19.78 & 10.87 & 15.96 \\ TAPEX$_{\textsc{Base}}$ & - & - & 7.32 & 39.76 & 8.68 & \underline{2.06} & 35.06 & 47.11 & 20.95 & 18.67 & 11.81 & 15.73 \\ TAPEX$_{\textsc{Large}}$ & - & - & \underline{8.80} & 46.59 & \underline{10.62} & 1.72 & 46.91 & 48.11 & 30.48 & \underline{22.65} & \underline{13.18} & \underline{18.59} \\ \rowcolor[rgb]{0.93,0.93,0.93} \multicolumn{13}{l}{\textit{fine-tuned Baselines}} \\ UnifiedQA$_{\textsc{Small}}$ & 7,686 & - & 22.27 & 51.31 & 27.27 & 2.83 & 52.28 & 48.11 & 69.52 & 35.85 & 21.71 & 29.79 \\ UnifiedQA$_{\textsc{Base}}$ & 7,686 & - & 34.02 & 70.68 & 40.74 & 7.90 & 84.09 & 55.67 & 73.33 & 53.31 & 30.46 & 43.52 \\ UnifiedQA$_{\textsc{Large}}$ & 7,686 & - & 48.67 & \underline{82.18} & 55.97 & \underline{20.26} & 94.63 & \underline{68.89} & \underline{79.05} & 65.92 & 45.92 & 57.35 \\ TAPEX$_{\textsc{Base}}$ & 7,686 & - & 39.59 & 73.09 & 46.85 & 11.33 & 84.19 & 61.33 & 69.52 & 56.70 & 37.02 & 48.27 \\ TAPEX$_{\textsc{Large}}$ & 7,686 & - & \underline{51.00} & 80.02 & \underline{59.92} & 16.31 & \textbf{95.34} & 64.00 & 73.33 & \underline{67.11} & \underline{47.07} & \underline{58.52} \\ \rowcolor[rgb]{0.93,0.93,0.93} \multicolumn{13}{l}{\textit{Prompting Baselines w/ GPT-3}} \\ Zero-shot & - & - & 53.57 & 66.67 & 55.55 & 45.84 & 78.22 & 55.44 & 54.29 & 63.37 & 48.41 & 56.96 \\ Zero-shot-CoT & - & - & 54.36 & 66.92 & 55.82 & 48.67 & \underline{78.82} & 55.67 & 51.43 & 63.62 & 49.59 & 57.61 \\ Few-shot (2-shot) & 2 & Random & 54.69 & 64.11 & 58.36 & 40.40 & 75.95 & 52.41 & 53.02 & 63.10 & 49.16 & 57.13 \\ Few-shot-CoT (2-shot) & 2 & Random & \underline{60.76} & \underline{69.09} & \underline{60.04} & \underline{63.58} & 76.49 & \underline{61.19} & \textbf{67.30} & \underline{68.62} & \underline{55.31} & \underline{62.92} \\ \rowcolor[rgb]{0.93,0.93,0.93} \multicolumn{13}{l}{\textit{\textbf{\textsc{PromptPG}\xspace w/ GPT-3 (Ours)}}} \\ Few-shot-CoT (2-shot) & 160+20 & Dynamic & \textbf{66.17} & \textbf{74.11} & \textbf{64.12} & \textbf{74.16} & 76.19 & \textbf{72.81} & 65.71 & \textbf{71.20} & \textbf{64.27} & \textbf{68.23}$_{5.31\uparrow}$ \\ \bottomrule \end{tabular} } \captionof{table}{Evaluation results of various baselines and our method on \textsc{TabMWP}\xspace{}. Training Data: number of used training data; Selection Strategy: strategy of selecting in-context examples for few-shot GPT-3; FREE: \textit{free-text} questions; MC: \textit{multi-choice} questions; INT: integer answers; DEC: decimal answers; EXTR: extractive text answers; BOOL: Boolean text answers; OTH: other text answers.} \vspace{-2mm} \label{tab:results} \end{figure} Without any example provided to GPT-3, zero-shot GPT-3 achieves a comparable accuracy as the best fine-tuned baselines UnifiedQA$_{\textsc{Large}}$ and TAPEX$_{\textsc{Large}}$, showing its surprisingly good generalization ability on \textsc{TabMWP}\xspace. Provided with two randomly sampled in-context examples as the prompt, few-shot GPT-3 gets an improvement of 0.17\%. Generating the multi-step solution before the answer, the few-shot-CoT GPT-3 model reports the best performance among all of these baseline models, with an accuracy of 62.92\%. Unlike few-shot-CoT GPT-3 randomly selecting the in-context examples, our proposed \textsc{PromptPG}\xspace learns to select performing examples with the help of policy gradient. \textsc{PromptPG}\xspace establishes a state-of-the-art performance on the \textsc{TabMWP}\xspace{} dataset: it surpasses the best baseline few-shot-CoT GPT-3 by 5.31\% on average. \textsc{PromptPG}\xspace shows its consistent advantages on two question types, two grade groups, and most of the answer types. \textbf{Heuristic guess and human performance.} The accuracy of \textit{multi-choice} questions by heuristic guess is 39.81\%, which aligns with the fact that there are 2.88 options on average. The accuracy for \textit{free-text} questions is considerably low since the inputs of \textsc{TabMWP}\xspace problems do not have direct clues for the answers. Humans outperform all benchmarks consistently across question types, answer types, and grade groups, with a 21.99\% average accuracy advantage over our best performing \textsc{PromptPG}\xspace. This gap is to be filled by future research on semi-structured mathematical reasoning. \textbf{Problem types and difficulty.} Among all the baselines, we find it is easier for models to answer \textit{multi-choice} questions than \textit{free-text} questions. Questions with the boolean (BOOL) and other (OTH) answer types tend to have lower accuracy scores than the extractive (EXTR) answer type, because the former ones need the abilities of fact verification and language understanding on diverse options, respectively. It is also not surprising for us to find that all the models perform worse on problems in grades 7-8 than in a lower-level group of 1-6. \subsection{Ablation Study} Here, we will study how different factors have an effect on the performances of baselines and our method on \textsc{TabMWP}\xspace{}. Experiments are conducted on 1,000 development examples. \textbf{Blind study of the dataset.} We evaluate the information gain of each component of the \textsc{TabMWP}\xspace problems by removing it from model inputs. To eliminate the impact and variance caused by example selection, the study is conducted using the zero-shot GPT-3 model. As shown in Table \ref{tab:blind}, there is a dramatic decline when either the tabular context (T) or the question text (Q) is missing from the inputs. For example, T$\rightarrow$A and Q$\rightarrow$A only attain an average accuracy of 6.10\% and 7.00\%, respectively, and their accuracies are near to zero on the \textit{multi-choice} questions. Taking both tabular and textual data as inputs (TQ$\rightarrow$A), the model significantly beats the heuristic guess. With the complete input information (TQ(C)$\rightarrow$A), the full model achieves the best performance. The blind study shows that our \textsc{TabMWP}\xspace{} is robust and reliable in distribution, and all input components are indispensable parts that provide necessary information for answering the questions. \begin{figure}[h!] \centering \small \renewcommand\tabcolsep{4.0pt} \resizebox{1.0\linewidth}{!}{ \begin{tabular}{lcccccccccccc} \toprule \textbf{Model} & \textbf{Format} & FREE & MC & INT & DEC & EXTR & BOOL & OTH & 1-6 & 7-8 & \textbf{Avg.} \\ \midrule Heuristic guess & TQ(C)$\rightarrow$A & 7.31 & 40.36 & 9.20 & 0.00 & 34.44 & 47.32 & 50.00 & 17.99 & 13.96 & 16.40 \\ \midrule Zero-shot GPT-3 & T$\rightarrow$A & 8.28 & 0.36 & 10.24 & 0.67 & 0.66 & 0.00 & 0.00 & 9.41 & 1.02 & 6.10 \\ Zero-shot GPT-3 & Q$\rightarrow$A & 9.24 & 1.09 & 10.94 & 2.68 & 1.32 & 0.89 & 0.00 & 10.23 & 2.03 & 7.00 \\ Zero-shot GPT-3 & T(C)$\rightarrow$A & 8.28 & 41.82 & 10.24 & 0.67 & 36.42 & 50.89 & 25.00 & 23.60 & 8.12 & 17.50 \\ Zero-shot GPT-3 & Q(C)$\rightarrow$A & 9.10 & 33.09 & 10.94 & 2.01 & 25.17 & 44.64 & 25.00 & 21.29 & 7.11 & 15.70 \\ Zero-shot GPT-3 & TQ$\rightarrow$A & 55.31 & 68.36 & 56.60 & 50.34 & 79.47 & 54.46 & 58.33 & 66.34 & 47.46 & 58.90 \\ Zero-shot GPT-3 (full model) & TQ(C)$\rightarrow$A & 54.76 & 72.00 & 56.42 & 48.32 & 76.82 & 66.07 & 66.67 & 67.00 & 47.97 & 59.50 \\ \bottomrule \end{tabular} } \captionof{table}{Blind studies on \textsc{TabMWP}\xspace{}. T: tabular context; Q: question; C: choice options; A: answer.} \label{tab:blind} \end{figure} \begin{figure}[ht] \begin{minipage}{0.48\textwidth} \vspace{-2mm} \centering \includegraphics[width=0.85\textwidth]{figures/fig_acc_diff_train_num.pdf} \caption*{(a) Accuracy w.r.t. different numbers of training examples, given 20 candidate examples.} \end{minipage} \hfill \begin{minipage}{0.48\textwidth} \centering \includegraphics[width=0.85\textwidth]{figures/fig_acc_diff_cand_num_font16.pdf} \caption*{(b) Accuracy w.r.t. different numbers of candidates, given 80 and 160 training examples.} \end{minipage} \vspace{-2mm} \caption{Accuracy w.r.t. different numbers of training and candidate examples. Experiments are conducted on 1,000 development instances, and each setting is repeated with four random seeds.} \vspace{-2mm} \label{fig:ablation} \end{figure} \textbf{Number of training examples.} We study the effect of different numbers of training examples on our dynamic prompt learning in Figure~\ref{fig:ablation} (a). With more training examples, the prediction accuracy first gradually increases to a peak of around 160 training examples. After that, the accuracy goes down with a growing variance. We reckon it is because the policy gradient algorithm can benefit from the scaling-up training data but fails to exploit more examples efficiently. \textbf{Number of candidate examples.} In Figure~\ref{fig:ablation} (b), we investigate how different numbers of candidate examples can affect policy learning performance. With the increasing candidate number, it is observed that the prediction accuracy will first go up and then go down after a threshold, given 80 or 160 training examples. It is probably because when the candidate pool is too small, the policy gradient algorithm has a limited action space to explore enough problem types. In contrast, too many candidates could make the algorithm hard to learn an optimal policy in a large search space. \begin{wraptable}{r}{0.44\textwidth} \vspace{-3.0mm} \centering \fontsize{9.0pt}{\baselineskip}\selectfont \renewcommand\tabcolsep{3.0pt} \renewcommand1.5{0.88} \begin{tabular}{lc} \toprule \textbf{Selection strategy} & \textbf{Acc. (\%)} \\ \midrule Same question type & 66.2 $\pm$ 0.60 \\ Same answer type & 67.9 $\pm$ 0.38 \\ Same grade level & 67.9 $\pm$ 1.87 \\ \midrule Most complex (\# of table cells) & 64.0 $\pm$ 0.42 \\ Most complex (\# of ques. words) & 68.2 $\pm$ 0.26 \\ \midrule Random selection & 65.2 $\pm$ 4.01 \\ Nearest neighbor & 68.2 $\pm$ 0.29 \\ \midrule \textbf{\textsc{PromptPG}\xspace (Ours}) & \textbf{70.9 $\pm$ 1.27}\\ \bottomrule \end{tabular} \vspace{-2mm} \caption{Evaluation results w.r.t. different strategies for selecting in-context examples.} \label{tab:selection} \vspace{-2mm} \end{wraptable} \textbf{Different selection strategies.} In Table~\ref{tab:selection}, we compare the proposed \textsc{PromptPG}\xspace with random selection and other heuristic-based example selection strategies for the few-shot-CoT GPT-3 model. Compared to random selection, selecting the same question or answer type of examples helps the model to take the task-relevant examples as the prompt, thus improving the accuracy and reducing the variance. Choosing the most complex examples does not boost the prediction performance consistently. The most semantically similar examples, as a kind of nearest neighbor search of the test example, help construct the performing and stable prompt for GPT-3. \textsc{PromptPG}\xspace shows its effectiveness in selecting optimal in-context examples over other strategies and largely reduces the instability caused by randomness. \subsection{Case Study} We conduct the case study in Appendix \ref{appx:case_study}. We visualize the two in-context examples selected by strategies of our \textsc{PromptPG}\xspace, nearest neighbor search, and random selection, in Figure \ref{fig:selected_exp_promptpg}, \ref{fig:selected_exp_nearest}, and \ref{fig:selected_exp_random}, respectively. The nearest neighbor search strategy selects the ``superficially'' similar examples to the test example. Instead, \textsc{PromptPG}\xspace tends to select examples that have multiple reasoning steps in the solution and similar abilities in mathematical reasoning, which results in higher prediction accuracy. Successful examples in Figure \ref{fig:accurate_1} - \ref{fig:accurate_5} show that \textsc{PromptPG}\xspace is able to generate reasonable reasoning steps to predict correct answers for a wide range of \textsc{TabMWP}\xspace problems. Failure examples in Figure \ref{fig:wrong_1} - \ref{fig:wrong_6} suggest that \textsc{PromptPG}\xspace has limitations when solving problems provided with complex tabular contexts or requiring a high-level ability of mathematical reasoning. \section{Related Work} \subsection{Math Word Problems} The task of solving Math Word Problems (MWPs) is to predict the answer given a natural language description of a math problem. There have been great efforts in developing datasets for MWPs, including Dolphin18K \citep{huang2016well}, DRAW-1K \citep{upadhyay2017annotating}, Math23K \citep{wang2017deep}, MathQA \citep{amini2019mathqa}, ASDiv \citep{miao2020diverse}, and SVAMP \citep{patel2021nlp}. However, these datasets only involve the textual modality, and most are limited to a small data scale. Some recent datasets like DVQA \citep{kafle2018dvqa}, Geometry3K \citep{lu2021inter} and IconQA \citep{lu2021iconqa} introduce math problems with diagrams as the visual context, where the system needs to perform mathematical reasoning over multi-modal information. To the best of our knowledge, our dataset \textsc{TabMWP}\xspace is the first dataset that requires mathematical reasoning over heterogeneous information from both the textual question and the tabular context. To solve MWPs, one popular line of previous methods is to generate the intermediate expressions and execute them to get the final answers \citep{huang2017learning,roy2017unit,amini2019mathqa}. Inspired by the recent progress achieved by GPT-3 in solving MWPs \citep{wei2022chain,wang2022self,kojima2022large}, we evaluate \textsc{TabMWP}\xspace using GPT-3 models in zero-shot and few-shot learning manners. \subsection{Table QA Datasets} Table Question Answering (Table QA) refers to the task of answering questions about tabular data. Numerous datasets have been developed for Table QA. For example, TabMCQ \citep{jauhar2016tabmcq} is an early dataset collected from grade exams. Datasets like WTQ \citep{pasupat2015compositional}, WikiSQL \citep{zhong2017seq2sql}, and SQA \citep{iyyer2017search} contain semi-structured tables from Wikipedia, while Spider \citep{yu2018spider} collects structured tables sourced from databases. Recent work aims at introducing datasets that require multi-hop reasoning between the textual and tabular data: HybridQA \citep{chen2020hybridqa}, OTTQA \citep{chen2020open}, MultiModalQA \citep{talmor2020multimodalqa}, AIT-QA \citep{katsis2021ait}, and FeTaQA \citep{nan2022fetaqa}. Datasets most related to our \textsc{TabMWP}\xspace{} dataset are FinQA \citep{chen2021finqa}, TAT-QA \citep{zhu2021tat}, and MultiHiertt \citep{zhao2022multihiertt} because they need numerical reasoning on financial reports with tabular data. Note that 77.6\% of questions in TAT-QA can be solvable without mathematical reasoning and 50.0\% of questions in FinQA are not table-must to be answered. In contrast, our proposed \textsc{TabMWP}\xspace collects questions where both mathematical reasoning and tabular context are necessary. \subsection{Prompt Learning for Language Models} Large pre-trained language models, such as GPT-3 \citep{chen2020big}, have shown their remarkable ability of few-shot learning on a wide range of downstream tasks \citep{houlsby2019parameter,brown2020language,lu2022learn}. Given a few in-context examples as demonstrations, GPT-3 can generalize to unseen test examples without parameter updating. For example, \cite{wei2022chain} randomly select different in-context examples from the training set and formulate their corresponding prompt with a test sample. However, recent studies show that few-shot GPT-3 highly depends on the selection of in-context examples and could be unstable, varying from the near chance to near state-of-the-art performance \citep{zhao2021calibrate,liu2022makes}. To mitigate the volatility of selecting in-context examples, \cite{lu2022fantastically} propose retrieving relevant examples that are semantically similar to the test sample. Other possible strategies could be using brute-force permutation search or relying on manually designed heuristics like choosing the most complex examples. Inspired by reinforcement learning's ability to search for an optimal action policy, we propose applying the policy gradient strategy \citep{sutton1998introduction} to learn to select in-context examples more efficiently and stably without designing human-designed heuristics. \section{Conclusion} In this paper, we propose \textsc{TabMWP}\xspace{}, the first large-scale dataset for math word problems in tabular contexts. \textsc{TabMWP}\xspace{} contains 38,431 open-domain problems with two question types and three answer types, and each problem is annotated with a multi-step solution. We evaluate \textsc{TabMWP}\xspace{} using state-of-the-art QA and TableQA methods in both pre-trained and fine-tuned settings, as well as the large pre-trained language model GPT-3. We further propose a novel approach, \textsc{PromptPG}\xspace, for few-shot GPT-3, which utilizes policy gradient to learn to select in-context examples from the training data and construct the performing prompt for the test example. Experimental results show that \textsc{PromptPG}\xspace outperforms existing strong baselines by a large margin of 5.31\% and reduces the accuracy volatility compared to random selection. To the best of our knowledge, it is the first work that applies reinforcement learning to select in-context examples for the few-shot GPT-3 model. \section{Acknowledge} We would like to thank Zhou Yu and Jiuxiang Gu for insightful discussions on dataset collection. We thank Chenhao Mu and Yao Fu for constructive suggestions in developing baselines and experiments. The work does not relate to Liang Qiu's position at Amazon Alexa. \section{Submission of conference papers to ICLR 2023} ICLR requires electronic submissions, processed by \url{https://openreview.net/}. See ICLR's website for more instructions. If your paper is ultimately accepted, the statement {\tt {\textbackslash}iclrfinalcopy} should be inserted to adjust the format to the camera ready requirements. The format for the submissions is a variant of the NeurIPS format. Please read carefully the instructions below, and follow them faithfully. \subsection{Style} Papers to be submitted to ICLR 2023 must be prepared according to the instructions presented here. Authors are required to use the ICLR \LaTeX{} style files obtainable at the ICLR website. Please make sure you use the current files and not previous versions. Tweaking the style files may be grounds for rejection. \subsection{Retrieval of style files} The style files for ICLR and other conference information are available online at: \begin{center} \url{http://www.iclr.cc/} \end{center} The file \verb+iclr2023_conference.pdf+ contains these instructions and illustrates the various formatting requirements your ICLR paper must satisfy. Submissions must be made using \LaTeX{} and the style files \verb+iclr2023_conference.sty+ and \verb+iclr2023_conference.bst+ (to be used with \LaTeX{}2e). The file \verb+iclr2023_conference.tex+ may be used as a ``shell'' for writing your paper. All you have to do is replace the author, title, abstract, and text of the paper with your own. The formatting instructions contained in these style files are summarized in sections \ref{gen_inst}, \ref{headings}, and \ref{others} below. \section{General formatting instructions} \label{gen_inst} The text must be confined within a rectangle 5.5~inches (33~picas) wide and 9~inches (54~picas) long. The left margin is 1.5~inch (9~picas). Use 10~point type with a vertical spacing of 11~points. Times New Roman is the preferred typeface throughout. Paragraphs are separated by 1/2~line space, with no indentation. Paper title is 17~point, in small caps and left-aligned. All pages should start at 1~inch (6~picas) from the top of the page. Authors' names are set in boldface, and each name is placed above its corresponding address. The lead author's name is to be listed first, and the co-authors' names are set to follow. Authors sharing the same address can be on the same line. Please pay special attention to the instructions in section \ref{others} regarding figures, tables, acknowledgments, and references. There will be a strict upper limit of 9 pages for the main text of the initial submission, with unlimited additional pages for citations. \section{Headings: first level} \label{headings} First level headings are in small caps, flush left and in point size 12. One line space before the first level heading and 1/2~line space after the first level heading. \subsection{Headings: second level} Second level headings are in small caps, flush left and in point size 10. One line space before the second level heading and 1/2~line space after the second level heading. \subsubsection{Headings: third level} Third level headings are in small caps, flush left and in point size 10. One line space before the third level heading and 1/2~line space after the third level heading. \section{Citations, figures, tables, references} \label{others} These instructions apply to everyone, regardless of the formatter being used. \subsection{Citations within the text} Citations within the text should be based on the \texttt{natbib} package and include the authors' last names and year (with the ``et~al.'' construct for more than two authors). When the authors or the publication are included in the sentence, the citation should not be in parenthesis using \verb|\citet{}| (as in ``See \citet{Hinton06} for more information.''). Otherwise, the citation should be in parenthesis using \verb|\citep{}| (as in ``Deep learning shows promise to make progress towards AI~\citep{Bengio+chapter2007}.''). The corresponding references are to be listed in alphabetical order of authors, in the \textsc{References} section. As to the format of the references themselves, any style is acceptable as long as it is used consistently. \subsection{Footnotes} Indicate footnotes with a number\footnote{Sample of the first footnote} in the text. Place the footnotes at the bottom of the page on which they appear. Precede the footnote with a horizontal rule of 2~inches (12~picas).\footnote{Sample of the second footnote} \subsection{Figures} All artwork must be neat, clean, and legible. Lines should be dark enough for purposes of reproduction; art work should not be hand-drawn. The figure number and caption always appear after the figure. Place one line space before the figure caption, and one line space after the figure. The figure caption is lower case (except for first word and proper nouns); figures are numbered consecutively. Make sure the figure caption does not get separated from the figure. Leave sufficient space to avoid splitting the figure and figure caption. You may use color figures. However, it is best for the figure captions and the paper body to make sense if the paper is printed either in black/white or in color. \begin{figure}[h] \begin{center} \fbox{\rule[-.5cm]{0cm}{4cm} \rule[-.5cm]{4cm}{0cm}} \end{center} \caption{Sample figure caption.} \end{figure} \subsection{Tables} All tables must be centered, neat, clean and legible. Do not use hand-drawn tables. The table number and title always appear before the table. See Table~\ref{sample-table}. Place one line space before the table title, one line space after the table title, and one line space after the table. The table title must be lower case (except for first word and proper nouns); tables are numbered consecutively. \begin{table}[t] \caption{Sample table title} \label{sample-table} \begin{center} \begin{tabular}{ll} \multicolumn{1}{c}{\bf PART} &\multicolumn{1}{c}{\bf DESCRIPTION} \\ \hline \\ Dendrite &Input terminal \\ Axon &Output terminal \\ Soma &Cell body (contains cell nucleus) \\ \end{tabular} \end{center} \end{table} \section{Default Notation} In an attempt to encourage standardized notation, we have included the notation file from the textbook, \textit{Deep Learning} \cite{goodfellow2016deep} available at \url{https://github.com/goodfeli/dlbook_notation/}. Use of this style is not required and can be disabled by commenting out \texttt{math\_commands.tex}. \centerline{\bf Numbers and Arrays} \bgroup \def1.5{1.5} \begin{tabular}{p{1in}p{3.25in}} $\displaystyle a$ & A scalar (integer or real)\\ $\displaystyle {\bm{a}}$ & A vector\\ $\displaystyle {\bm{A}}$ & A matrix\\ $\displaystyle {\tens{A}}$ & A tensor\\ $\displaystyle {\bm{I}}_n$ & Identity matrix with $n$ rows and $n$ columns\\ $\displaystyle {\bm{I}}$ & Identity matrix with dimensionality implied by context\\ $\displaystyle {\bm{e}}^{(i)}$ & Standard basis vector $[0,\dots,0,1,0,\dots,0]$ with a 1 at position $i$\\ $\displaystyle \text{diag}({\bm{a}})$ & A square, diagonal matrix with diagonal entries given by ${\bm{a}}$\\ $\displaystyle {\textnormal{a}}$ & A scalar random variable\\ $\displaystyle {\mathbf{a}}$ & A vector-valued random variable\\ $\displaystyle {\mathbf{A}}$ & A matrix-valued random variable\\ \end{tabular} \egroup \vspace{0.25cm} \centerline{\bf Sets and Graphs} \bgroup \def1.5{1.5} \begin{tabular}{p{1.25in}p{3.25in}} $\displaystyle {\mathbb{A}}$ & A set\\ $\displaystyle \mathbb{R}$ & The set of real numbers \\ $\displaystyle \{0, 1\}$ & The set containing 0 and 1 \\ $\displaystyle \{0, 1, \dots, n \}$ & The set of all integers between $0$ and $n$\\ $\displaystyle [a, b]$ & The real interval including $a$ and $b$\\ $\displaystyle (a, b]$ & The real interval excluding $a$ but including $b$\\ $\displaystyle {\mathbb{A}} \backslash {\mathbb{B}}$ & Set subtraction, i.e., the set containing the elements of ${\mathbb{A}}$ that are not in ${\mathbb{B}}$\\ $\displaystyle {\mathcal{G}}$ & A graph\\ $\displaystyle \parents_{\mathcal{G}}({\textnormal{x}}_i)$ & The parents of ${\textnormal{x}}_i$ in ${\mathcal{G}}$ \end{tabular} \vspace{0.25cm} \centerline{\bf Indexing} \bgroup \def1.5{1.5} \begin{tabular}{p{1.25in}p{3.25in}} $\displaystyle {a}_i$ & Element $i$ of vector ${\bm{a}}$, with indexing starting at 1 \\ $\displaystyle {a}_{-i}$ & All elements of vector ${\bm{a}}$ except for element $i$ \\ $\displaystyle {A}_{i,j}$ & Element $i, j$ of matrix ${\bm{A}}$ \\ $\displaystyle {\bm{A}}_{i, :}$ & Row $i$ of matrix ${\bm{A}}$ \\ $\displaystyle {\bm{A}}_{:, i}$ & Column $i$ of matrix ${\bm{A}}$ \\ $\displaystyle {\etens{A}}_{i, j, k}$ & Element $(i, j, k)$ of a 3-D tensor ${\tens{A}}$\\ $\displaystyle {\tens{A}}_{:, :, i}$ & 2-D slice of a 3-D tensor\\ $\displaystyle {\textnormal{a}}_i$ & Element $i$ of the random vector ${\mathbf{a}}$ \\ \end{tabular} \egroup \vspace{0.25cm} \centerline{\bf Calculus} \bgroup \def1.5{1.5} \begin{tabular}{p{1.25in}p{3.25in}} $\displaystyle\frac{d y} {d x}$ & Derivative of $y$ with respect to $x$\\ [2ex] $\displaystyle \frac{\partial y} {\partial x} $ & Partial derivative of $y$ with respect to $x$ \\ $\displaystyle \nabla_{\bm{x}} y $ & Gradient of $y$ with respect to ${\bm{x}}$ \\ $\displaystyle \nabla_{\bm{X}} y $ & Matrix derivatives of $y$ with respect to ${\bm{X}}$ \\ $\displaystyle \nabla_{\tens{X}} y $ & Tensor containing derivatives of $y$ with respect to ${\tens{X}}$ \\ $\displaystyle \frac{\partial f}{\partial {\bm{x}}} $ & Jacobian matrix ${\bm{J}} \in \mathbb{R}^{m\times n}$ of $f: \mathbb{R}^n \rightarrow \mathbb{R}^m$\\ $\displaystyle \nabla_{\bm{x}}^2 f({\bm{x}})\text{ or }{\bm{H}}( f)({\bm{x}})$ & The Hessian matrix of $f$ at input point ${\bm{x}}$\\ $\displaystyle \int f({\bm{x}}) d{\bm{x}} $ & Definite integral over the entire domain of ${\bm{x}}$ \\ $\displaystyle \int_{\mathbb{S}} f({\bm{x}}) d{\bm{x}}$ & Definite integral with respect to ${\bm{x}}$ over the set ${\mathbb{S}}$ \\ \end{tabular} \egroup \vspace{0.25cm} \centerline{\bf Probability and Information Theory} \bgroup \def1.5{1.5} \begin{tabular}{p{1.25in}p{3.25in}} $\displaystyle P({\textnormal{a}})$ & A probability distribution over a discrete variable\\ $\displaystyle p({\textnormal{a}})$ & A probability distribution over a continuous variable, or over a variable whose type has not been specified\\ $\displaystyle {\textnormal{a}} \sim P$ & Random variable ${\textnormal{a}}$ has distribution $P$\\% so thing on left of \sim should always be a random variable, with name beginning with \r $\displaystyle \mathbb{E}_{{\textnormal{x}}\sim P} [ f(x) ]\text{ or } \mathbb{E} f(x)$ & Expectation of $f(x)$ with respect to $P({\textnormal{x}})$ \\ $\displaystyle \mathrm{Var}(f(x)) $ & Variance of $f(x)$ under $P({\textnormal{x}})$ \\ $\displaystyle \mathrm{Cov}(f(x),g(x)) $ & Covariance of $f(x)$ and $g(x)$ under $P({\textnormal{x}})$\\ $\displaystyle H({\textnormal{x}}) $ & Shannon entropy of the random variable ${\textnormal{x}}$\\ $\displaystyle D_{\mathrm{KL}} ( P \Vert Q ) $ & Kullback-Leibler divergence of P and Q \\ $\displaystyle \mathcal{N} ( {\bm{x}} ; {\bm{\mu}} , {\bm{\Sigma}})$ & Gaussian distribution % over ${\bm{x}}$ with mean ${\bm{\mu}}$ and covariance ${\bm{\Sigma}}$ \\ \end{tabular} \egroup \vspace{0.25cm} \centerline{\bf Functions} \bgroup \def1.5{1.5} \begin{tabular}{p{1.25in}p{3.25in}} $\displaystyle f: {\mathbb{A}} \rightarrow {\mathbb{B}}$ & The function $f$ with domain ${\mathbb{A}}$ and range ${\mathbb{B}}$\\ $\displaystyle f \circ g $ & Composition of the functions $f$ and $g$ \\ $\displaystyle f({\bm{x}} ; {\bm{\theta}}) $ & A function of ${\bm{x}}$ parametrized by ${\bm{\theta}}$. (Sometimes we write $f({\bm{x}})$ and omit the argument ${\bm{\theta}}$ to lighten notation) \\ $\displaystyle \log x$ & Natural logarithm of $x$ \\ $\displaystyle \sigma(x)$ & Logistic sigmoid, $\displaystyle \frac{1} {1 + \exp(-x)}$ \\ $\displaystyle \zeta(x)$ & Softplus, $\log(1 + \exp(x))$ \\ $\displaystyle || {\bm{x}} ||_p $ & $L^p$ norm of ${\bm{x}}$ \\ $\displaystyle || {\bm{x}} || $ & $L^2$ norm of ${\bm{x}}$ \\ $\displaystyle x^+$ & Positive part of $x$, i.e., $\max(0,x)$\\ $\displaystyle \bm{1}_\mathrm{condition}$ & is 1 if the condition is true, 0 otherwise\\ \end{tabular} \egroup \vspace{0.25cm} \section{Final instructions} Do not change any aspects of the formatting parameters in the style files. In particular, do not modify the width or length of the rectangle the text should fit into, and do not change font sizes (except perhaps in the \textsc{References} section; see below). Please note that pages should be numbered. \section{Preparing PostScript or PDF files} Please prepare PostScript or PDF files with paper size ``US Letter'', and not, for example, ``A4''. The -t letter option on dvips will produce US Letter files. Consider directly generating PDF files using \verb+pdflatex+ (especially if you are a MiKTeX user). PDF figures must be substituted for EPS figures, however. Otherwise, please generate your PostScript and PDF files with the following commands: \begin{verbatim} dvips mypaper.dvi -t letter -Ppdf -G0 -o mypaper.ps ps2pdf mypaper.ps mypaper.pdf \end{verbatim} \subsection{Margins in LaTeX} Most of the margin problems come from figures positioned by hand using \verb+\special+ or other commands. We suggest using the command \verb+\includegraphics+ from the graphicx package. Always specify the figure width as a multiple of the line width as in the example below using .eps graphics \begin{verbatim} \usepackage[dvips]{graphicx} ... \includegraphics[width=0.8\linewidth]{myfile.eps} \end{verbatim} or \begin{verbatim} \usepackage[pdftex]{graphicx} ... \includegraphics[width=0.8\linewidth]{myfile.pdf} \end{verbatim} for .pdf graphics. See section~4.4 in the graphics bundle documentation (\url{http://www.ctan.org/tex-archive/macros/latex/required/graphics/grfguide.ps}) A number of width problems arise when LaTeX cannot properly hyphenate a line. Please give LaTeX hyphenation hints using the \verb+\-+ command. \subsubsection*{Author Contributions} If you'd like to, you may include a section for author contributions as is done in many journals. This is optional and at the discretion of the authors. \subsubsection*{Acknowledgments} Use unnumbered third level headings for the acknowledgments. All acknowledgments, including those to funding agencies, go at the end of the paper.

a3dd4b58f2c8b3ca4274c874349e3306845184ee

\section{Introduction} In the past decades, as important properties, the $\partial_{}\bar{\partial}$--lemma and its variation has been extensively studied in complex geometry, symplectic geometry and generalized complex geometry; see \cite{DGMS75,BK98,Merkulov98,Man99,Cav06,TY12-1,TY12-2, AT13, AT15,TY16,YY20} etc. Among these works, a profound one of Deligne, Griffiths, Morgan, and Sullivan is that if a compact complex manifold satisfies the $\partial_{}\bar{\partial}$--lemma, then it is formal. In symplectic geometry, a variation of $\partial_{}\bar{\partial}$--lemma, called $d\delta$--lemma, also attracts many attention: for a compact symplectic manifold, it satisfies the $d\delta$--lemma if and only if the Hard Lefschetz condition holds, and in this case the symplectic manifold is formal, and its de Rham cohomology admits a structure of Frobenius manifold. In this paper we consider the holomorphic Poisson manifold. Originally, Poisson structures arise from Hamiltonian system of classical dynamics. In many situations, the Poisson structures are actually holomorphic; see \cite{BZ99,Hit06,L-GSX08,Go10,Hit12,CFP16,CGP15} etc. In particular, holomorphic Poisson structures are closely related to generalized complex geometry. On the one hand, a holomorphic Poisson structure naturally defines a generalized complex structure of special type; see \cite{Gu11}. On the other hand, by Bailey's local classification theorem of generalized complex structures, each generalized complex manifold is locally equivalent to the product of a symplectic manifold and a holomorphic Poisson manifold; see \cite{Ba13}. We refer the readers to \cite{Hit03,Hit06,Fu05,L-GSX08,Go10,CSX10,Gu11,BX15,BCV19} and references therein for more results on the applications of holomorphic Poisson structures in generalized complex geometry and the relationships with other geometries. Assume that $(M,\pi)$ is a compact holomorphic Poisson manifold. The Koszul--Brylinski operator $\partial_{\pi}:=\iota_{\pi}\partial-\partial\iota_{\pi}$, studied independently by Koszul \cite{Kos84} and Brylinski \cite{Bry88}, has many properties. The $k$--th Koszul--Brylinski homology $H_{k}(M, \pi)$ of $(M,\pi)$ is defined as the $k$-th hypercohomology of holomorphic Koszul--Brylinski complex $$ \cdots \stackrel{\partial_{\pi}}{\to}\Omega_{M}^{p+1} \stackrel{\partial_{\pi}}{\to} \Omega_{M}^{p} \stackrel{\partial_{\pi}}{\to} \Omega_{M}^{p-1} \stackrel{\partial_{\pi}}{\to} \cdots. $$ Moreover, as a BV operator, $\partial_{\pi}$ generates a Lie bracket $[-,-]_{\partial_\pi}$ (defined by \ref{Lie str on forms1}). Most notably, there exists a holomorphic version of Evens--Lu--Weinstein duality for $H_{\bullet}(X,\pi)$, which is a generalization of Serre duality for Dolbeault cohomology; see \cite[Theorem 4.4]{Sti11}. Furthermore, there is a canonical Fr\"{o}hlicher--type spectral sequence, called the Dolbeault--Koszul--Brylinski spectral sequence (see Definition \ref{Dol-Poi-spectral-seq}), which converges to $H_{\bullet}(X,\pi)$. However, in general, it is not easy to compute the holomorphic Koszul--Brylinski homology for a specific holomorphic Poisson manifold. As far as we know, only some particular class are calculated(see \cite{Hon19,HX11,Sti11}). Essentially, the Koszul--Brylinski homology and the algebraic structures on it naturally depend both on the Poisson structures and complex structures, but the Poisson structures are some kind complicated. Thus, a natural question, motivated by Brylinski \cite{Bry88}, arises now is: \begin{quest}\label{pro} What conditions on a holomorphic Poisson manifold $(M,\pi)$ ensure the degeneracy of $E_1$--page of the Dolbeault--Koszul--Brylinski spectral sequence? \end{quest} Motivated by the works of $\partial_{}\bar{\partial}$--lemma above and the close relations between holomorphic Poisson geometry and symplectic geometry, generalized complex geometry, we consider the $\partial_{}\bar{\partial}$--lemma and $\partial_{\pi}\bar{\partial}$--lemma on holomorphic Poisson manifold. Our first main result states as follows. \begin{theorem}\label{main-theorem1} Let $(M,\pi)$ be a holomorphic Poisson manifold. If $M$ satisfies the $\partial_{}\bar{\partial}$--lemma or $\partial_{\pi}\bar{\partial}$--lemma, then its Dolbeault--Koszul--Brylinski spectral sequence degenerates at $E_{1}$--page: $$ H_{k}(M, \pi) \cong \bigoplus_{p-q=n-k}H^{p,q}_{\bar{\partial}}(M). $$ \end{theorem} Notice that if $M$ is a projective manifold or K\"{a}hler manifold, then it automatically satisfies the $\partial\bar{\partial}$--lemma, and hence the Theorem \ref{main-theorem1} is applicable to these situations. The Lie bracket $[-,-]_{\partial_\pi}$ on the Dolbeault complex of $M$ which is generated by the Koszul--Brylinski operator $\partial_{\pi}$, is in fact compatible with the Dolbeault operator, i.e., the Dolbeault complex of $M$ admits a DGLA structure. Analogous to the symplectic case, we consider the formality of such DGLA. \begin{theorem}\label{main-theorem2} Suppose that $(M,\pi)$ is a holomorphic Poisson manifold. If $(M,\pi)$ satisfies the $\partial_{\pi}\bar{\partial}$--lemma, then the DGLA $(A_M^{\bullet,\bullet},\bar{\partial},[-,-]_{\partial_\pi})$ is formal. \end{theorem} Actually, the DGLA $(A_{M}^{\bullet,\bullet},\bar{\partial},[-,-]_{\partial_\pi})$ is closely relevant with the Kodaira--Spencer DGLA $(A^{0,\bullet}(M,\wedge^\bullet T_{M}),\bar{\partial},[-,-]_{})$ which controls the deformations of complex structure of $M$; see Proposition \ref{deformation}. Therefore it is worth seeking the Maurer--Cartan elements of the DGLA $(A_M^{\bullet,\bullet},\bar{\partial},[-,-]_{\partial_\pi})$. \begin{theorem}\label{main-theorem3} Let $(M,\pi)$ be a holomorphic Poisson manifold. If $(M,\pi)$ satisfies the $\partial_{}\bar{\partial}$--lemma or $\partial_{\pi}\bar{\partial}$--lemma, then for any $[\alpha]\in H_{\bar{\partial}}^{1,1}(M)$, there exists an Maurer--Cartan element $\alpha_t$ whose $\alpha_1$ is a representative of $[\alpha]$. In this case, $\pi^\sharp[\alpha]\in H^1(M,\mathcal{T}_{M})$ is tangent to a deformation of complex structure. \end{theorem} It is noteworthy that the fact that on a holomorphic Poisson manifold $(X,\pi)$, for any $[\alpha]\in H_{\bar{\partial}}^{1,1}(M)$, $\pi^\sharp[\alpha]\in H^1(M,\mathcal{T}_{M})$ is tangent to a deformation of complex structure, is proved by Hitchin \cite{Hit12} under the assumption of $\partial_{}\bar{\partial}$--lemma or $H^2(M,\CO_M)=0$. Later, Hitchin's result is generalized by Fiorenza and Manetti (cf. \cite[Theorem 6.3]{FM12}) under the assumption that the natural map $H^2_{dR}(M,\mathbb{C})\to H^2(M,\CO_M)$ is surjective. This paper is organized as follows. In Section [\ref{c2}] we first review some basics of holomorphic Poisson manifold and discuss some results on $\partial_{}\bar{\partial}$--lemma and $\partial_{\pi}\bar{\partial}$--lemma. We devote Section [\ref{c3}] to the formality of the Dolbeault complex. In Section [\ref{c4}] we study the Maurer--Cartan equation of the DGLA on the Dolbeault complex. Finally, some examples are stated in Section [\ref{c6}]. \subsection*{Acknowledgments} The author would like to thank the School of Mathematics of Sichuan University and Tianyuan Mathematical Center in Southwest China for the hospitalities during the spring of 2022. In particular, the author would like to thank the referee for introducing the example of Nakamura manifold in the subsection [\ref{Nak}] to him. This work is partially supported by the National Nature Science Foundation of China (No. 12126309, 12126354, 12171351), the Natural Science Foundation of Chongqing (No. CSTC2020JCYJ--MSXMX0160), and the Scientific and Technological Research Program of Chongqing Municipal Education Commission (Grant No. KJQN202201108). \section{Preliminaries}\label{c2} \subsection{Basics of holomorphic Poisson manifold} In this subsection, we recall some basic facts on holomorphic Poisson manifolds. Let $M$ be a complex manifold and let $\CO_M$ be its structure sheaf (i.e., the sheaf of holomorphic functions), $\Omega_{M}^{p}$ be the sheaf of holomorphic $p$--forms, $\mathcal{T}_{M}$ be the sheaf of holomorphic vector fields. \begin{definition} A complex manifold $M$ is called a {\it holomorphic Poisson manifold} if $M$ admits a section $\pi\in \Gamma(M, \wedge^{2}\mathcal{T}_{M})$ with $\bar{\partial}\pi=0, [\pi,\pi]_{SN}=0$ where $[-,-]_{SN}$ means the Schouten bracket. \end{definition} Such $\pi$ is called a {\it holomorphic Poisson bi--vector field} of $M$. It induces a sheaf morphism $\pi^\sharp:\Omega_{M}^{1}\to\mathcal{T}_{M}$ via $$\pi^\sharp(df)(dg)=\pi(df,dg).$$ The {\it Koszul--Brylinski operator} of $(M, \pi)$ on the sheaf $\Omega_{M}^{\bullet}$ of holomorphic forms of $M$ is defined as $$ \partial_{\pi}:= \iota_{\pi}\circ \partial-\partial\circ \iota_{\pi}, $$ where $\partial$ is the Dolbeault operator and $\iota_{\pi}$ is the contraction operator with respect to $\pi$. One can verify the following lemma by direct calculations. \begin{lemma}\label{identities} Let $(M, \pi)$ be a holomorphic Poisson manifold. Then we have the following identities: \begin{itemize} \item[(1)]$\bar{\partial}\iota_{\pi}-\iota_{\pi}\bar{\partial}=0,$ $\bar{\partial}\partial_{\pi}+\partial_{\pi}\bar{\partial}=0;$ \item[(2)]$\partial\partial_{\pi}+\partial_{\pi}\partial=0;$ \item[(3)]$\partial_{\pi}\iota_{\pi}-\iota_{\pi}\partial_{\pi}=0,$ $\partial_{\pi}^2=0$. \end{itemize} \end{lemma} Moreover, one can check that the Koszul--Brylinski operator $\partial_{\pi}$ is a BV operator, i.e., for any $\alpha\in \Omega_{M}^{k}$ and $\beta\in \Omega_{M}^{l}$, \begin{equation}\label{Lie str on forms1} [\alpha, \beta]_{\partial_{\pi}} =(-1)^{k}\Big(\partial_{\pi}(\alpha\wedge \beta)-(\partial_{\pi}\alpha) \wedge \beta -(-1)^{k} \alpha\wedge (\partial_{\pi}\beta)\Big) \end{equation} is a Gerstenhaber bracket (a Poisson bracket with degree $-1$) on $\Omega_{M}^{\bullet}$. Equivalently, the bracket $[-, -]_{\partial_{\pi}}$ is obtained by Leibniz rule via \begin{equation}\label{Lie str on forms2} [\alpha, \beta]_{\partial_{\pi}} :=L_{\pi^{\sharp}(\alpha)}\beta-L_{\pi^{\sharp}(\beta)}\alpha-\partial(\pi(\alpha,\beta)), \; \forall \; \alpha,\beta\in \Omega_{M}^{1}. \end{equation} Thus, there is a holomorphic Koszul--Brylinski complex $$ 0\to \Omega_{M}^{n} \stackrel{\partial_{\pi}}{\to} \cdots \stackrel{\partial_{\pi}}{\to}\Omega_{M}^{s+1} \stackrel{\partial_{\pi}}{\to} \Omega_{M}^{s} \stackrel{\partial_{\pi}}{\to} \Omega_{M}^{s-1} \stackrel{\partial_{\pi}}{\to} \cdots\stackrel{\partial_{\pi}}{\to}\mathcal{O}_M\to 0. $$ \begin{definition} Let $(M, \pi)$ be a holomorphic Poisson manifold. The {\it $k$--th Koszul--Brylinski homology} of $(M,\pi)$ is defined as the $k$--th hypercohomology of its holomorphic Koszul--Brylinski complex, that is to say, \begin{equation*} H_{k}(M, \pi):=\mathbb{H}^{k}(M, (\Omega_{M}^{\bullet},\partial_{\pi})). \end{equation*} \end{definition} \begin{lemma}\label{KB-equal1} Suppose $(M, \pi)$ is a holomorphic Poisson manifold. Then its holomorphic Koszul--Brylinski complex admits a resolution which is the total complex of the double complex $(\A_{M}^{\bullet,\bullet}, \partial_{\pi},\bar{\partial})$, where $\A_{M}^{p,q}$ is the sheaf of $(p,q)$--forms on $M$. \end{lemma} \begin{proof} The lemma is followed by the fact that $\A_{M}^{p,\bullet}$ is a fine resolution of $\Omega_{M}^{p}$ and $\partial_{\pi}$ commutes with $\bar{\partial}$; also see \cite[Theorem 5.1]{Sti11}. \end{proof} \begin{remark}For a complex $n$--dimensional manifold $M$, one has $$ H_{k}(M, \pi=0) \cong \bigoplus\limits_{q-p+n=k}H^{p,q}_{\bar{\partial}}(M). $$ \end{remark} Actually the Koszul--Brylinski homology and the Dolbeault cohomology are closely related with spectral sequences. \begin{definition}\label{Dol-Poi-spectral-seq} Let $(M, \pi)$ be a compact holomorphic Poisson manifold of complex dimension $n$. The following spectral sequence associated to the double complex $(\Gamma(M, \mathcal{A}_{M}^{\bullet,\bullet}), \partial_{\pi},\bar{\partial})$, \begin{equation}\label{Dol-Poisson-spec.-seq} E_{1}^{s,t}:=H_{\bar{\partial}}^{n-s,t}(M) \Longrightarrow H_{n-s+t}(M,\pi), \end{equation} is called the {\it Dolbeault--Koszul--Brylinski spectral sequence of $(M, \pi)$}. \end{definition} In general, the Dolbeault--Koszul--Brylinski spectral sequence \eqref{Dol-Poisson-spec.-seq} does not degenerate at $E_{1}$--page (see examples in Section \ref{c6}). An equivalent condition of the $E_{1}$--degeneracy of Dolbeault--Koszul--Brylinski spectral sequence is given as follows. \begin{lemma}[{\cite[Lemma 5.7]{CCYY22}}]\label{deg} Let $(M, \pi)$ be a compact holomorphic Poisson manifold of complex dimension $n$. Then the {\it Dolbeault--Koszul--Brylinski spectral sequence} of $(M, \pi)$ degenerates at $E_{1}$--page if and only if for any $0\leq k\leq 2n$, $$ \sum_{p-q=n-k} \dim_{\mathbb{C}}\, H_{\bar{\partial}}^{p,q}(M)=\dim_{\mathbb{C}} H_{k}(M,\pi). $$ \end{lemma} By a result of Sti\'{e}non \cite[Theorem 6.4]{Sti11}, the holomorphic Evens--Lu--Weinstein pairing on the holomorphic Koszul--Brylinski homology is non--degenerate. More precisely, if $(M, \pi)$ is a compact holomorphic Poisson manifold of complex dimension $n$, then for $0\leq k\leq 2n$, there is an isomorphism \begin{equation}\label{Serre-Poincare-duality} H_{2n-k}(M,\pi)\cong H_{k}(M,\pi). \end{equation} Meanwhile, in the dual aspect, there exists a {\it holomorphic Lichnerowicz--Poisson complex} $(\wedge^{\bullet}\mathcal{T}_{M},b_{\pi})$: \begin{equation*} 0\to \CO_M\stackrel{b_{\pi}}{\to} \cdots \stackrel{b_{\pi}}{\to}\wedge^{s-1}\mathcal{T}_{M} \stackrel{b_{\pi}}{\to} \wedge^{s}\mathcal{T}_{M} \stackrel{b_{\pi}}{\to} \wedge^{s+1}\mathcal{T}_{M} \stackrel{b_{\pi}}{\to} \cdots\stackrel{b_{\pi}}{\to}\wedge^{n}\mathcal{T}_{M} \to 0 \end{equation*} where $b_{\pi}(-)=[\pi,-]_{SN}$. The $k$--th hypercohomology of $(\wedge^{\bullet}\mathcal{T}_{M},b_{\pi})$ is called the $k$--th \emph{holomorphic Lichnerowicz--Poisson cohomology}, i.e., $$ H^{k}(M, \pi):=\mathbb{H}^{k}(M, (\wedge^{\bullet}\mathcal{T}_{M},b_{\pi})). $$ If $M$ admits a holomorphic volume form $\omega\in \Gamma(M,\Omega^n_M)$, that is to say, $M$ is a Calabi--Yau manifold, then there is a natural morphism of sheaves $$ \iota_{(-)}\omega: \wedge^{s}\mathcal{T}_{M} \to \Omega_{M}^{n-s}. $$ However, it does not induce a morphism of sheaf complexes between $(\wedge^{\bullet}\mathcal{T}_{M},b_{\pi})$ and $(\Omega_{M}^{\bullet},\partial_{\pi})$ since generally $\iota_{(-)}\omega$ does not commutative with the differentials. This motivates the following definition. \begin{definition}[{cf. \cite{We97,BZ99}}] A holomorphic Poisson manifold $(X,\pi)$ is called {\it unimodular} if there is a holomorphic volume form $\omega$ such that the morphism $\iota_{(-)}\omega$ induces a morphism of sheaf complexes from $(\wedge^{\bullet}\mathcal{T}_{M},b_{\pi})$ to $(\Omega^{\bullet}_{M},\partial_{\pi})$. \end{definition} An equivalent condition of a holomorphic Poisson manifold $(M,\pi)$ being unimodular is $\partial_\pi\omega=0$, or the modular vector field, introduced by Weinstein \cite{We97} and Brylinski--Zuckerman \cite{BZ99}, vanishes. In particular, we have \begin{proposition}[{\cite[Proposition 4.7]{Sti11}}]\label{dual} If the $n$--dimensional holomorphic Poisson manifold $(M,\pi)$ is unimodular, then for any $k\in \mathbb{Z}$, there is an isomorphism $$ H_{k}(M, \pi)\cong H^{2n-k}(M, \pi). $$ \end{proposition} \subsection{$\partial_{}\bar{\partial}$--lemma and $\partial_{\pi}\bar{\partial}$--lemma} In this subsection, we consider the $E_{1}$--degeneracy of Dolbeault--Koszul--Brylinski spectral sequence of a holomorphic Poisson manifold $(M, \pi)$ under the assumption of $\partial_{}\bar{\partial}$--lemma or $\partial_{\pi}\bar{\partial}$--lemma. Let $A_{M}^{s,t}:=\Gamma(M, \mathcal{A}_{M}^{s,t})$ be the space of differential $(s,t)$--forms on $M$. \begin{theorem}\label{deg1} Let $(M,\pi)$ be a holomorphic Poisson manifold. If $M$ satisfies the $\partial_{}\bar{\partial}$--lemma, then its Dolbeault--Koszul--Brylinski spectral sequence degenerates at $E_{1}$--page, i.e. $$ H_{k}(M, \pi) \cong \bigoplus_{p-q=n-k}H^{p,q}_{\bar{\partial}}(M). $$ \end{theorem} \begin{proof} Recall that a compact complex manifold $M$ satisfies the \emph{$\partial\bar{\partial}$--lemma}, if the equation $$\ker\,\partial\cap\ker\,\bar{\partial}\cap\mathrm{im}\,d=\mathrm{im}\,\partial\bar{\partial}$$ holds for the double complex $(A^{\bullet,\bullet}_{M}, \partial, \bar{\partial})$ (cf. \cite{DGMS75}). For any class $[\alpha]\in H^{\bullet,\bullet}_{\bar{\partial}}(M)$, take $\beta=\partial_{}\alpha$, then we have the following $$ \left\{ \begin{array}{ll} \bar{\partial}\beta=\bar{\partial}\partial_{}\alpha=-\partial_{}\bar{\partial}\alpha=0,&\\ \partial_{}\beta=\partial_{}^2\alpha=0, & \\ \beta=d\alpha, & \end{array} \right. $$ i.e. $\beta\in\ker\partial_{}\cap \ker\, \bar{\partial}\cap \mathrm{im} \, d$. If $M$ satisfies the $\partial_{}\bar{\partial}$--lemma, then there exists a $\gamma$ on $M$ such that $\beta=\partial_{}\bar{\partial}\gamma$. Let $\tilde{\alpha}=\alpha-\bar{\partial}\gamma$. Then we have that $[\alpha]=[\tilde{\alpha}]$ in $H^{\bullet,\bullet}_{\bar{\partial}}(M)$ and $\partial_{}\tilde{\alpha}=0$. This means that we can always choose the $\partial$--closed representatives of the Dolbeault cohomology classes in $H^{\bullet,\bullet}_{\bar{\partial}}(M)$. Now let $[\alpha]\in H^{\bullet,\bullet}_{\bar{\partial}}(M)$ such that $\partial(\alpha)=0$. If we write $\eta=\partial_{\pi}(\alpha)$, then by the Lemma \ref{identities}, we have $$ \left\{ \begin{array}{ll} \bar{\partial}\eta=0,&\\ \partial_{}\eta=0, & \\ \eta=(d\iota_{\pi}-\iota_{\pi}d)(\alpha)=d(\iota_{\pi}\alpha), & \end{array} \right. $$ Once again, by the $\partial_{}\bar{\partial}$--lemma of $M$, there exists a $\zeta$ on $M$ such that $\partial_{\pi}(\alpha)=\partial_{}\bar{\partial}(\zeta)=-\bar{\partial}\partial_{}(\zeta)$. Equivalently, this means the differential of the $E_1$--page of the Dolbeault--Koszul--Brylinski spectral sequence of $(M, \pi)$ is zero. Consequently, we have that the Dolbeault--Koszul--Brylinski spectral sequence of $(M, \pi)$ degenerates at $E_{1}$--page. Thus $$ H_{k}(M, \pi) \cong H_{k}(M, 0) \cong \bigoplus_{p-q=n-k}H^{p,q}_{\bar{\partial}}(M), $$ and the proof is completed. \end{proof} It is worth noting that if $M$ is a projective manifold or K\"{a}hler manifold, then it automatically satisfies the $\partial\bar{\partial}$--lemma. Hence a corollary of the Theorem \ref{deg1} is the following. \begin{corollary}\label{Kahler} Let $(M,\pi)$ be a compact holomorphic Poisson manifold. If $M$ is a projective manifold or K\"{a}hler manifold, then the Dolbeault--Koszul--Brylinski spectral sequence of $(M, \pi)$ degenerates at $E_{1}$--page. \end{corollary} Moreover, combined with the Proposition \ref{dual} and Theorem \ref{deg1}, we have \begin{corollary} Let $(M,\pi)$ be a unimodular holomorphic Poisson manifold of complex dimension $n$. If $M$ satisfies the $\partial\bar{\partial}$--lemma, then $$ H^{k}(M, \pi) \cong \bigoplus_{p-q=k-n}H^{p,q}_{\bar{\partial}}(M). $$ \end{corollary} \begin{example}\label{p-x} Suppose $\pi_{\mathbb{P}^{n}}$ is a holomorphic Poisson bi--vector field on $\mathbb{P}^{n}$. By the Corollary \eqref{Kahler}, we have $$ H_{k}(\mathbb{P}^{n},\pi_{\mathbb{P}^{n}})= \left\{ \begin{array}{ll} \mathbb{C}^{n+1},&k=n,\\ 0, & k\neq n, \end{array} \right.$$ since its Hodge numbers are $h^{p,q}=\delta_{pq}$. Suppose $\pi$ is a holomorphic Poisson bi--vector field on $M=\mathbb{P}^{m}\times \mathbb{P}^{n}$. By the Corollary \eqref{Kahler} and the K\"{u}nneth's formula for Dolbeault cohomology(cf. \cite[Corollary 19]{CFGU00}), we have $$ H_{k}(M,\pi)= \left\{ \begin{array}{ll} \mathbb{C}^{(m+1)(n+1)},&k=m+n,\\ 0, & k\neq n. \end{array} \right. $$ \end{example} Motivated by the $\partial\bar{\partial}$--lemma, following the work \cite{DGMS75}, we have the following definition. \begin{definition}[{$\partial_{\pi}\bar{\partial}$--lemma}] Let $(M,\pi)$ be a holomorphic Poisson manifold. We say that $(M,\pi)$ satisfies the $\partial_{\pi}\bar{\partial}$--lemma if $$ \ker\, \partial_{\pi}\cap \ker\, \bar{\partial}\cap \mathrm{im} \,(\partial_{\pi}+\bar{\partial}) = \mathrm{im}\,\partial_{\pi} \bar{\partial}. $$ \end{definition} An equivalent description of $\partial_{\pi}\bar{\partial}$--lemma, which are the special case of \cite[Lemma 5.15]{DGMS75}, states as follows. \begin{lemma Let $(M,\pi)$ be a holomorphic Poisson manifold. Then the following conditions are equivalent: \begin{itemize} \item[(1)] $(M,\pi)$ satisfies the $\partial_{\pi}\bar{\partial}$--lemma; \item[(2)] $\ker\bar{\partial} \cap \mathrm{im}\partial_{\pi}= \mathrm{im}\partial_{\pi}\bar{\partial} , \ker\partial_{\pi} \cap \mathrm{im}\bar{\partial}= \mathrm{im}\partial_{\pi}\bar{\partial}$; \item[(3)] $\ker\bar{\partial} \cap \ker\partial_{\pi} \cap (\mathrm{im}\partial_{\pi}+\mathrm{im}\bar{\partial})= \mathrm{im}\partial_{\pi}\bar{\partial}.$ \end{itemize} \end{lemma} Another two closely related cohomologies we need in this paper are the Bott--Chern cohomology and the Aeppli cohomology. These two cohomologies are the special case of Angella and Tomassini \cite{AT15}. \begin{definition} Let $(M,\pi)$ be a holomorphic Poisson manifold. The {\it $(p,q)$--th Bott--Chern cohomology} of $(M,\pi)$ is defined as $$ H_{BC}^{p,q}(M,\pi):= \frac{\ker\,\partial_{\pi}\cap \ker\,\bar{\partial}}{\mathrm{im}\, \partial_{\pi}\bar{\partial}}, $$ while the {\it $(p,q)$--th Aeppli cohomology} of $(M,\pi)$ is defined as $$ H_{A}^{p,q}(M,\pi):= \frac{\ker\,\partial_{\pi}\bar{\partial}}{\mathrm{im}\, \partial_{\pi}+\mathrm{im}\, \bar{\partial}}. $$ \end{definition} One can check that the identity map induces natural morphisms \begin{equation}\label{morphisms} \vcenter{ \xymatrix@R=0.5cm{ & H_{BC}^{\bullet,\bullet}(M,\pi) \ar[dl]^{} \ar[d]^{}\ar[dr]^{} & \\ H_{\partial_{\pi}}^{\bullet,\bullet}(M) \ar[dr]_{} & H_{\bullet}(M, \pi)\ar[d]^{} & H_{\bar{\partial}}^{\bullet,\bullet}(M) \ar[dl]^{} \\ & H_{A}^{\bullet,\bullet}(M,\pi) & } } \end{equation} since $\ker\, \partial_{\pi}\cap \ker\, \bar{\partial}\subset\ker\,(\partial_{\pi}+\bar{\partial}) \subset\ker\,\partial_{\pi}\bar{\partial}$ and $\mathrm{im}\,\partial_{\pi} \bar{\partial} \subset \mathrm{im}\,(\partial_{\pi}+ \bar{\partial}) \subset \mathrm{im}\,\partial_{\pi}+ \mathrm{im}\,\bar{\partial}$. Generally, each morphism in \ref{morphisms} is neither injective nor surjective. The following theorem is an application of \cite[Theorem 1 \& Theorem 2 \& Lemma 2.4]{AT15}. \begin{theorem}\label{lem1} Let $(M,\pi)$ be a holomorphic Poisson manifold. Then \begin{itemize} \item[(1)] $\sum\limits_{p+q=k}\big(\mathrm{dim}_{\mathbb{C}} H_{BC}^{p,q}(M,\pi) +\mathrm{dim}_{\mathbb{C}}H_{A}^{p,q}(M,\pi)\big) \geq 2\mathrm{dim}_{\mathbb{C}}H_{k}(M,\pi)$; \item[(2)]the identity holds if and only if $(M,\pi)$ satisfies the $\partial_{\pi}\bar{\partial}$--lemma. In this case, all morphisms in the diagram \ref{morphisms} are isomorphic. \end{itemize} \end{theorem} For more characterizations of the $\partial_{\pi}\bar{\partial}$--lemma we refer to \cite[Proposition 5.17]{DGMS75}. Especially, an analogous result to Theorem \ref{deg1} whose proof is obtained by replacing the operator $\partial_{}$ with $\partial_{\pi}$ states as follows. \begin{theorem}\label{deg2} Let $(M,\pi)$ be a compact holomorphic Poisson manifold. If $(M,\pi)$ satisfies the $\partial_{\pi}\bar{\partial}$--lemma, then the Dolbeault--Koszul--Brylinski spectral sequence $E_{\bullet}$ degenerate at the first page, or $$ H_{k}(M, \pi) \cong \bigoplus_{p-q=n-k}H^{p,q}_{\partial_{\pi}}(M). $$ \end{theorem} \begin{remark} One can check that for a complex manifold $M$ with trivial holomorphic Poisson bi--vector field $\pi$, it does not satisfy the $\partial_{\pi}\bar{\partial}$--lemma, but its Dolbeault--Koszul--Brylinski spectral sequence degenerates at the first page. In Section \ref{c6} more examples whose Dolbeault--Koszul--Brylinski spectral sequence degenerate at $E_{1}$--page but do not satisfy the $\partial_{\pi}\bar{\partial}$--lemma are given. \end{remark} \section{Formality of the Dolbeault complex}\label{c3} Recall for a holomorphic Poisson manifold $(M, \pi)$, the Koszul--Brylinski operator $\partial_{\pi}$ is a BV operator, and it generates a Lie bracket $[\alpha, \beta]_{\partial_{\pi}}$. In this section we consider its formality properties with respect to variant differentials. \begin{lemma}\label{DGL} Suppose $(M, \pi)$ is a holomorphic Poisson manifold. Then $(A_{M}^{\bullet,\bullet},\partial,[-,-]_{\partial_\pi})$, $(A_{M}^{\bullet,\bullet},\bar{\partial},[-,-]_{\partial_\pi})$ and $(A_{M}^{\bullet,\bullet},d,[-,-]_{\partial_\pi})$ are three differential Gerstenhaber algebras. \end{lemma} \begin{proof} Note $(A_{M}^{\bullet,\bullet},[-,-]_{\partial_\pi})$ is a Lie algebra. By the Lemma \ref{identities}, all three differential $\partial, \bar{\partial}$ and $d$ are derivations with respect to the bracket $[-,-]_{\partial_\pi}$, thus we have the lemma. \end{proof} For any $k\geq1$, the map \begin{equation}\label{contraction} \iota_{\pi}^k: (A_{M}^{\bullet,\bullet},\partial_{\pi},\bar{\partial}) \to (A_{M}^{\bullet-2k,\bullet},\partial_{\pi},\bar{\partial}) \end{equation} is a well--defined morphism of double complexes since by the Lemma \ref{identities}, $\iota_{\pi}^{k}\partial_{\pi}=\partial_{\pi}\iota_{\pi}^{k}$ and $\iota_{\pi}^{k}\bar{\partial}=\bar{\partial}\iota_{\pi}^{k}$. Thus the operator $$e^{\iota_{\pi}}:=\sum_{k=0}\frac{1}{k!}\iota_{\pi}^k,$$ is well--defined on $(A_{M}^{\bullet,\bullet},\partial_{\pi},\bar{\partial})$ with inverse $e^{-\iota_{\pi}}$. \begin{lemma}\label{f1} For any natural number $k$, $\iota_{\pi}^k\partial=\partial\iota_{\pi}^k+k\iota_{\pi}^{k-1}\partial_{\pi}$. Moreover, we have that \begin{equation}\label{f2} e^{\iota_{\pi}}\partial =(\partial+\partial_\pi)e^{\iota_{\pi}}. \end{equation} \end{lemma} \begin{proof} Inductively, $$ \iota_{\pi}^{k+1}\partial =\iota_{\pi}(\partial\iota_{\pi}^k+k\iota_{\pi}^{k-1}\partial_{\pi}) =\iota_{\pi}\partial\iota_{\pi}^k+k\iota_{\pi}^{k}\partial_{\pi} =\partial\iota_{\pi}^{k+1}+(k+1)\iota_{\pi}^{k}\partial_{\pi}. $$ Thus, we have \begin{center} $ e^{\iota_{\pi}}\partial =\sum\limits_{k=0}\frac{1}{k!}\iota_{\pi}^k\partial =\sum\limits_{k=0}\frac{1}{k!}(\partial\iota_{\pi}^k+k\iota_{\pi}^{k-1}\partial_{\pi}) =(\partial+\partial_\pi)e^{\iota_{\pi}}, $ \end{center} and the lemma is proved. \end{proof} \begin{theorem}[{\cite[Corollary 2]{ST08}} \& {\cite[Theorem 3.2]{FM12}}] The DGLAs $(A_M^{\bullet,\bullet},\partial,[-,-]_{\partial_\pi})$ and $(A_M^{\bullet,\bullet},d,[-,-]_{\partial_\pi})$ are formal and quasi--isomorphic to abelian Lie algebras. \end{theorem} \begin{proof} For any $\partial$--closed forms $\alpha\in A_M^{i,k-i},\ \beta\in A_M^{j,l-j}$, \begin{eqnarray*} [\alpha,\beta]_{\partial_{\pi}} &=& (-1)^k\big(\partial_{\pi}(\alpha\wedge \beta)-(\partial_{\pi}\alpha) \wedge \beta -(-1)^{k} \alpha\wedge (\partial_{\pi}\beta)\big)\\ &=& \partial\big((-1)^k\iota_{\pi}(\alpha\wedge \beta)-(-1)^k(\iota_{\pi}\alpha) \wedge \beta - \alpha\wedge (\iota_{\pi}\beta)\big), \end{eqnarray*} i.e. $[\alpha,\beta]_{\partial_{\pi}}$ is $\partial$--exact. Combined with the Lemma \eqref{f1}, we have that $(A_M^{\bullet,\bullet},\partial,[-,-]_{\partial_\pi})$ is formal and quasi--isomorphic to the abelian Lie algebra $(H^{\bullet,\bullet}_\partial(M),0)$. Set $d_{\pi}:=\iota_{\pi}d-d\iota_{\pi}=\partial_{\pi}+\bar{\partial}_{\pi}$. According to a result by Sharygin--Talalaev \cite[Lemma 5]{ST08}, the Lie bracket $$ [\alpha,\beta]_{d_{\pi}}= (-1)^k\big(d_{\pi}(\alpha\wedge \beta)-(d_{\pi}\alpha) \wedge \beta -(-1)^{k} \alpha\wedge (d_{\pi}\beta)\big), \alpha\in A_{M}^{k},\ \beta\in A_{M}^{l}, $$ associated to $d_{\pi}$ is $d$--exact if both $\alpha$ and $\beta$ are $d$--closed. Since the Poisson bi--vector field $\pi$ is holomorphic, one have that the Dolbeault operator $\bar{\partial}$ commutes with the operator $\iota_{\pi}$. Equivalently, we get $\bar{\partial}_{\pi}=\iota_{\pi}\bar{\partial}-\bar{\partial}\iota_{\pi}=0$. Therefore we in fact have that $[\alpha,\beta]_{\partial_{\pi}}=[\alpha,\beta]_{d_{\pi}}$ is $d$--exact, and the equation \eqref{f2} becomes as \begin{eqnarray*} e^{\iota_{\pi}}d&=& e^{\iota_{\pi}}(\partial+\bar{\partial}) \\ &=& e^{\iota_{\pi}}\partial+e^{\iota_{\pi}}\bar{\partial} \\ &=& (\partial+\partial_\pi)e^{\iota_{\pi}} +\bar{\partial}e^{\iota_{\pi}}\\ &=& (d+\partial_\pi)e^{\iota_{\pi}}. \end{eqnarray*} Therefore we conclude that $(A_M^{\bullet,\bullet},d,[-,-]_{\partial_\pi})$ is formal and quasi--isomorphic to the abelian Lie algebra $(H^{\bullet}_{dR}(M),0)$. \end{proof} We are now in a position to give the proof of Theorem \ref{main-theorem2}. \begin{proof}[Proof of Theorem \ref{main-theorem2}] Under the assumption that $(M,\pi)$ satisfies the $\partial_{\pi}\bar{\partial}$--lemma, we claim the morphisms of DGLAs in the diagram \begin{equation*}\label{formality} \begin{tikzpicture}[scale=0.3] \draw (-10,0) node[left] {$(A_{M}^{\bullet,\bullet},\bar{\partial},[-,-]_{\partial_\pi})$}; \draw[-latex] (-8,0)-- (-10,0) node[right] {\quad\, $(\ker\partial_{\pi},\bar{\partial},[-,-]_{\partial_\pi})$}; \draw(6.5,0) node[right] {$(\ker\bar{\partial},\partial_{\pi},[-,-]_{\partial_\pi})$}; \draw[-latex] (18,0)-- (20,0) node[right] {$(A_{M}^{\bullet,\bullet},\partial_{\pi},[-,-]_{\partial_\pi}) $}; \draw (3.5,-5) node[left] {$(H_{\partial_{\pi}}^{\bullet,\bullet}(M),0,[-,-]_{\partial_\pi}) $}; \draw(6.5,-5) node[right] {$(H_{\bar{\partial}}^{\bullet,\bullet}(M),0,[-,-]_{\partial_\pi}) $}; \draw (12.5,-10) node[left] {$(H_{BC}^{\bullet,\bullet}(M,\pi),0,[-,-]_{\partial_\pi})$}; \draw[-latex] (-2.8,-0.5)-- (-2.8,-4); \draw[-latex] (3,-8.5)--(-3,-6); \draw[-latex] (13,-1)-- (13,-4); \draw[-latex] (7,-8.5)--(13,-6); \draw (-2.5,-2) node[left] {$p_1$}; \draw (13,-2) node[left] {$p_2$}; \draw (-8,1) node[left] {$i_1$}; \draw (20,1) node[left] {$i_2$}; \end{tikzpicture} \end{equation*} are all well--defined and quasi--isomorphic. In fact, by the symmetry of the two operators $\partial_{\pi}$ and $\bar{\partial}$ in $\partial_{\pi}\bar{\partial}$--lemma and the Theorem \ref{lem1}, we only need to check the morphisms $i_1$ and $p_1$ are well--defined and quasi--isomorphic. Indeed, by the facts that $\partial_{\pi}$ is commutative with $\bar{\partial}$ and both $\partial_{\pi}, \bar{\partial}$ are derivations of the bracket $[-,-]_{\partial_\pi}$, we obtain that both $i_1$ and $p_1$ are well--defined. Following \cite[Proposition 9.7.1]{Man99}, we prove that both $i_1$ and $p_1$ are quasi--isomorphisms. If $\partial_{\pi}(\alpha)=0, i_1(\alpha)=\alpha=\bar{\partial} (\beta)$ for some $\beta$, then $\alpha\in \ker\partial_{\pi}\cap \mathrm{im}\bar{\partial}$. By the $\partial_{\pi}\bar{\partial}$--lemma, there exists a $\gamma$ such that $\alpha=\bar{\partial}\partial_{\pi}(\gamma)$. This means that for any element $\alpha$ of $\ker\partial_{\pi}$, if its image $i_1(\alpha)$ is $\bar{\partial}$--exact, then $\alpha$ itself is $\bar{\partial}$--exact in $\ker\partial_{\pi}$. By definition, we conclude that $H(i_1)$ is injective. Further, if $\bar{\partial} (\alpha)=0$, then $\partial_{\pi}(\alpha)\in \ker\bar{\partial} \cap \mathrm{im}\partial_{\pi}$. By the $\partial_{\pi}\bar{\partial}$--lemma, there exists a $\beta$ such that $\partial_{\pi}(\alpha)=\partial_{\pi}\bar{\partial}(\beta)$. Equivalently, $\alpha-\bar{\partial}\beta\in \ker \partial_{\pi}$ and $[\alpha-\bar{\partial}\beta]=[\alpha]$ in $H_{\bar{\partial}}^{\bullet,\bullet}(M)$. That is to say, $H(i_1)$ is surjective. Meanwhile, if $\alpha\in \mathrm{im}\partial_{\pi}$ and $\bar{\partial}(\alpha)=0$, then $\alpha\in \ker\bar{\partial}\cap \mathrm{im}\partial_{\pi}$. Once again, by the $\partial_{\pi}\bar{\partial}$--lemma, there exists a $\gamma$ such that $\alpha=\bar{\partial}\partial_{\pi}(\gamma)$. In other words, $H(p_1)$ is injective. At last, the morphism $H(p_1)$ is surjective since if $\partial_{\pi}(\alpha)=0$, then $\bar{\partial}\alpha\in \ker\partial_{\pi}\cap \mathrm{im}\bar{\partial}$. Once again, by the $\partial_{\pi}\bar{\partial}$--lemma, there exists a $\beta$ such that $\bar{\partial}\alpha=\bar{\partial}\partial_{\pi}(\beta)$. Equivalently, the $\bar{\partial}$--closed element $\alpha-\partial_{\pi}(\beta)$ is cohomologous to $\alpha$ in $H_{\partial_{\pi}}^{\bullet,\bullet}(M)$. Thus, if $(M,\pi)$ satisfies the $\partial_{\pi}\bar{\partial}$--lemma, then the DGLA $(A_M^{\bullet,\bullet},\bar{\partial},[-,-]_{\partial_\pi})$ is quasi--isomorphic to $(H_{\bar{\partial}}^{\bullet,\bullet}(M),0,[-,-]_{\partial_\pi})$. By definition, this means that $(A_M^{\bullet,\bullet},\bar{\partial},[-,-]_{\partial_\pi})$ is formal. \end{proof} \section{Maurer--Cartan elements}\label{c4} In this section we consider the Maurer--Cartan equation of the DGLA $(A_M^{\bullet,\bullet}[[t]]=A_M^{\bullet,\bullet}\otimes \mathbb{C}[[t]],\bar{\partial},[-,-]_{\partial_\pi})$: \begin{equation}\label{MCE1} \bar{\partial}\alpha_t+\frac{1}{2}[\alpha_t,\alpha_t]_{\partial_\pi}=0. \end{equation} Naturally, due to the degree reason, the solutions (called Maurer--Cartan elements) of such equation (if exists) lie in $A_M^{1,1}[[t]]$. If we write $\alpha_t=\sum\limits_{i=1}^\infty\alpha_it^i$, then the Maurer--Cartan equation is equivalent to a system of equations \begin{equation}\label{MCE2} \begin{cases} \bar{\partial}\alpha_1=0 , & \\ \bar{\partial}\alpha_k+\frac{1}{2}\sum\limits_{i=1}^{k-1}[\alpha_i,\alpha_{k-i}]_{\partial_\pi}=0, & k\geq 2. \end{cases} \end{equation} \begin{theorem}\label{Solution_of_MCE1} Let $(M,\pi)$ be a holomorphic Poisson manifold. If $(M,\pi)$ satisfies the $\partial_{}\bar{\partial}$--lemma or $\partial_{\pi}\bar{\partial}$--lemma, then for any $[\alpha]\in H_{\bar{\partial}}^{1,1}(M)$, there exists a Maurer--Cartan element $\alpha_t$ whose $\alpha_1$ is a representative of $[\alpha]$. \end{theorem} \begin{proof} We here only prove the theorem under the assumption of $\partial_{\pi}\bar{\partial}$--lemma since the case that $(M,\pi)$ satisfies the $\partial_{}\bar{\partial}$--lemma can be obtained with total same strategy by replacing $\partial_{\pi}$ with $\partial$. Assume that $(M,\pi)$ satisfies the $\partial_{\pi}\bar{\partial}$--lemma. To prove the theorem, it is sufficient to find $\alpha_2,\alpha_3,\cdots$ such that they satisfy the Maurer--Cartan equation \ref{MCE2}. For any class $[\alpha]\in H^{\bullet,\bullet}_{\bar{\partial}}(M)$, take $\beta=\partial_{\pi}\alpha$, then we have the following $$ \left\{ \begin{array}{ll} \bar{\partial}\beta=\bar{\partial}\partial_{\pi}\alpha=-\partial_{\pi}\bar{\partial}\alpha=0,&\\ \beta=\partial_{\pi}\alpha, & \end{array} \right. $$ i.e. $\beta\in\ker\partial_{\pi}\cap \ker\, \bar{\partial} \cap (\mathrm{im} \,\partial_{\pi}+\mathrm{im} \,\bar{\partial})$. By the $\partial_{\pi}\bar{\partial}$--lemma of $(M,\pi)$, there exists a $\gamma$ on $M$ such that $\beta=\partial_{\pi}\bar{\partial}\gamma$. Let $\tilde{\alpha}=\alpha-\bar{\partial}\gamma$. Then we have that $[\alpha]=[\tilde{\alpha}]$ in $H^{\bullet,\bullet}_{\bar{\partial}}(M)$ and $\partial_{\pi}\tilde{\alpha}=0$. Therefore, in what follows we always choose the $\partial_{\pi}$--closed representatives of the Dolbeault cohomology classes in $H^{1,1}_{\bar{\partial}}(M)$. Let $[\alpha_1]\in H_{\bar{\partial}}^{1,1}(M)$ such that $\partial_{\pi}(\alpha_1)=0$., then $$ \left\{ \begin{array}{ll} \bar{\partial}[\alpha_1,\alpha_1]_{\partial_\pi} =[\bar{\partial}\alpha_1,\alpha_1]_{\partial_\pi} -[\alpha_1,\bar{\partial}\alpha_1]_{\partial_\pi}=0,&\\ {[\alpha_1,\alpha_1]_{\partial_\pi}} =\partial_\pi(\alpha_1\wedge\alpha_1) -\partial_\pi(\alpha_1)\wedge\alpha_1 -\alpha_1\wedge\partial_\pi(\alpha_1)=\partial_\pi(\alpha_1\wedge\alpha_1),& \end{array} \right. $$ This means that $[\alpha_1,\alpha_1]_{\partial_\pi}$ is $\bar{\partial}$--closed and $\partial_\pi$--exact. By the $\partial_{\pi}\bar{\partial}$--lemma of $(M,\pi)$, there exists a $(2,1)$--form $\zeta_2$ such that $[\alpha_1,\alpha_1]_{\partial_\pi}=\partial_\pi\bar{\partial}\zeta_2$. Therefore if take $\beta_2=\frac{1}{2}\zeta_2,\alpha_2=\partial_\pi\beta_2$, then we have that $$\bar{\partial}\alpha_2+\frac{1}{2}[\alpha_1,\alpha_1]_{\partial_\pi}=0.$$ Inductively, suppose that we already found $\partial_\pi$--exact forms $\alpha_2=\partial_\pi\beta_2,\cdots,\alpha_k=\partial_\pi\beta_k$ satisfying the Maurer--Cartan equation \ref{MCE2}. Let $$ \gamma_k=[\alpha_1,\alpha_{k}]_{\partial_\pi}+\cdots+[\alpha_i,\alpha_{k+1-i}]_{\partial_\pi} +\cdots+[\alpha_k,\alpha_{1}]_{\partial_\pi}. $$ Note the degree of the bracket $[-,-]_{\partial_\pi}$ is $-1$, and the symmetry of $[-,-]_{\partial_\pi}$ indicates that $[\alpha_i,\alpha_{j}]_{\partial_\pi}=[\alpha_j,\alpha_{i}]_{\partial_\pi}$ and $[\bar{\partial}\alpha_i,\alpha_{j}]_{\partial_\pi}=-[\alpha_j,\bar{\partial}\alpha_{i}]_{\partial_\pi}$. When $k=2l\geq4$, we have that \begin{eqnarray*} \bar{\partial}\gamma_k &=& 2\cdot\sum\limits_{i=1}^{l}\bar{\partial} [\alpha_i,\alpha_{2l+1-i}]_{\partial_\pi} \\ &=& 2\cdot\sum\limits_{i=1}^{l}\big( [\bar{\partial}\alpha_i,\alpha_{2l+1-i}]_{\partial_\pi} +[\bar{\partial}\alpha_{2l+1-i},\alpha_i]_{\partial_\pi}\big)\\ &=&\sum\limits_{s+t=i}\sum\limits_{i=2}^{l} [[\alpha_s,\alpha_{t}]_{\partial_\pi},\alpha_{2l+1-i}]_{\partial_\pi} +\sum\limits_{p+q=2l+1-i}\sum\limits_{i=1}^{l} [[\alpha_p,\alpha_{q}]_{\partial_\pi},\alpha_i]_{\partial_\pi}\\ &=& \sum\limits_{p+q+r=2l+1}[[\alpha_p,\alpha_{q}]_{\partial_\pi},\alpha_r]_{\partial_\pi}\\ &=& 0. \end{eqnarray*} Analogous, when $k=2l-1\geq3$, we have that \begin{eqnarray*} \bar{\partial}\gamma_k &=& 2\cdot\sum\limits_{i=1}^{l-1}\bar{\partial} [\alpha_i,\alpha_{2l-i}]_{\partial_\pi} +\bar{\partial} [\alpha_l,\alpha_{l}]_{\partial_\pi}\\ &=& 2\cdot\sum\limits_{i=1}^{l-1}\big( [\bar{\partial}\alpha_i,\alpha_{2l-i}]_{\partial_\pi} +[\bar{\partial}\alpha_{2l-i},\alpha_i]_{\partial_\pi} \big) + 2\cdot[\bar{\partial}\alpha_l,\alpha_{l}]_{\partial_\pi}\\ &=&\sum\limits_{s+t=i}\sum\limits_{i=2}^{l-1} [[\alpha_s,\alpha_{t}]_{\partial_\pi},\alpha_{2l-i}]_{\partial_\pi} +\sum\limits_{p+q=2l-i}\sum\limits_{i=1}^{l-1}[[\alpha_p,\alpha_{q}]_{\partial_\pi},\alpha_i]_{\partial_\pi} +\sum\limits_{u+v=l}[[\alpha_u,\alpha_{v}]_{\partial_\pi},\alpha_l]_{\partial_\pi}\\ &=& \sum\limits_{p+q+r=2l}[[\alpha_p,\alpha_{q}]_{\partial_\pi},\alpha_r]_{\partial_\pi}\\ &=& 0. \end{eqnarray*} Hence we always have that $\gamma_k$ is $\bar{\partial}$--closed. Moreover, $$ \gamma_k =[\alpha_1,\partial_\pi\beta_{k}]_{\partial_\pi}+ \sum\limits_{i=2}^{k} [\partial_\pi\beta_i,\alpha_{k+1-i}]_{\partial_\pi} =\partial_\pi\big(-[\alpha_1,\beta_{k}]_{\partial_\pi}+\sum\limits_{i=2}^{k} [\beta_i,\alpha_{k+1-i}]_{\partial_\pi}\big), $$ This means that $\gamma$ is also $\partial_\pi$--exact. By the $\partial_{\pi}\bar{\partial}$--lemma of $(M,\pi)$, there exists a $(2,1)$--form $\zeta_{k+1}$ such that $\gamma=\partial_\pi\bar{\partial}\zeta_{k+1}$. Therefore if take $\beta_{k+1}=\frac{1}{2}\zeta_{k+1},\alpha_{k+1}=\partial_\pi\beta_{k+1}$, then we have that $$\bar{\partial}\alpha_{k+1}+\frac{1}{2}\sum\limits_{i=1}^{k}[\alpha_1,\alpha_1]_{\partial_\pi}=0.$$ So the theorem is proved. \end{proof} Recall for a holomorphic Poisson manifold $(M, \pi)$, the holomorphic Poisson bi--vector field $\pi$ induces a sheaf morphism $\pi^\sharp:\Omega_{M}^{1}\to \mathcal{T}_{M}$ by contraction with $\pi$. More generally, for any $p\geq1$, $\pi$ induces a sheaf morphism $ \pi^\sharp:\Omega_{M}^{p}\to\wedge^p\mathcal{T}_{M} $ and $$ \pi^\sharp:\A_{M}^{p,q}\to\A^{0,q}(M,\wedge^p\mathcal{T}_{M}) $$ which is given locally by $$ \pi^\sharp:dz_{i_1}\wedge\cdots\wedge dz_{i_p}\wedge d\hat{z}_{i_1}\wedge\cdots\wedge d\hat{z}_{i_q} \mapsto (-1)^pX_{z_{i_1}}\wedge\cdots\wedge X_{z_{i_p}}\otimes d\hat{z}_{i_1}\wedge\cdots\wedge d\hat{z}_{i_q}. $$ Here $X_f$ is the Hamiltonian vector field with respect to $f$. Such $\pi^\sharp$ connects the DGLA $(A_{M}^{\bullet,\bullet},\bar{\partial},[-,-]_{\partial_\pi})$ with the Kodaira--Spencer DGLA $(A^{0,\bullet}(M,\wedge^\bullet T_{M}),\bar{\partial},[-,-]_{})$. More precisely, we have \begin{proposition}\label{deformation} Let $(M, \pi)$ be a holomorphic Poisson manifold, then the holomorphic Poisson bi--vector field $\pi$ induces a map of DGLAs $$ \pi^\sharp:(A_{M}^{\bullet,\bullet},\bar{\partial},[-,-]_{\partial_\pi}) \to (A^{0,\bullet}(M,\wedge^\bullet T_{M}),\bar{\partial},[-,-]_{}). $$ \end{proposition} \begin{proof} By definition, we have that $\pi^\sharp$ is homomorphism of algebras $A_{M}^{\bullet,\bullet}$ and $A^{0,\bullet}(M,\wedge^\bullet T_{M})$. Since $\pi$ is holomorphic, one can obtain that $\pi^\sharp$ is commutative with $\bar{\partial}$. At last, note these two DGLAs are Gerstenhaber algebras, and for any $1$--forms $\alpha,\beta$, $$\pi^\sharp[\alpha,\beta]_{\partial_\pi} =[\pi^\sharp(\alpha),\pi^\sharp(\beta)]_{}, $$ thus the map $\pi^\sharp$ is a homomorphism of Lie algebras and the proposition is proved. \end{proof} A classical result of Maurer--Cartan equation is that a map of DGLAs induces a map of Maurer--Cartan elements. By this fact, based on the Proposition \ref{deformation}, a corollary of the Theorem \ref{Solution_of_MCE1} states: \begin{corollary}\label{def} If a holomorphic Poisson manifold $(X,\pi)$ satisfies the $\partial_{}\bar{\partial}$--lemma or $\partial_{\pi}\bar{\partial}$--lemma, then for any $[\alpha]\in H_{\bar{\partial}}^{1,1}(M)$, $\pi^\sharp[\alpha]\in H^1(M,\mathcal{T}_{M})$ is tangent to a deformation of complex structure. \end{corollary} \section{Examples}\label{c6} In this section, we discuss some examples with different properties in the viewpoints of Poisson geometry. More precisely we consider some special nilmanifolds and solvmanifolds. Let $G$ be a complex nilpotent Lie group with Lie algebra $\mathfrak{g}$ whose complexification is $\mathfrak{g}_{\mathbb{C}}:=\mathfrak{g}\mathfrak{}\otimes_{\mathbb{R}}\mathbb{C}$, and let $H$ be a co--compact discrete subgroup of $G$. Suppose $M=G/H$ is the associated nilmanifold endowed with a $G$--left--invariant complex structure $J$ and an $G$--left--invariant holomorphic Poisson bi--vector field $\pi$. Then for any $p$, there exists a natural inclusion of complexes $$ i: (\wedge^{p,\bullet}\mathfrak{g}_{\mathbb{C}}^*,\bar{\partial}) \hookrightarrow (\Gamma(M,\A_M^{p,\bullet}),\bar{\partial}). $$ \begin{lemma}[{\cite[Lemma 6.1]{CCYY22}}]\label{nil} If the map $i$ defined above is a quasi--isomorphism, then the $k$--th total cohomology of double complex $(\wedge^{\bullet,\bullet}\mathfrak{g}_{\mathbb{C}}^*,\partial_\pi,\bar{\partial})$ is isomorphic to $H_k(M,\pi)$ for any $k$. \end{lemma} \begin{remark} A result of Sakane (see \cite[Theorem 1]{Sak76}) states that if a nilmanifold is complex parallelizable (i.e. the holomorphic tangent bundle is holomorphically trivial), the inclusion $i$ is a quasi--isomorphism. For more conditions such that the inclusion $i$ is a quasi--isomorphism, the reader may refer to Angella \cite[Theorem 3.2]{Angella}. \end{remark} \subsection{Iwasawa manifold} Let $\mathrm{H}(3; \mathbb{C})$ be the complex Heisenberg Lie group: $$ \mathrm{H}(3; \mathbb{C})= \Bigg\{ \small{\left(\begin{array}{ccc} 1 & z_{1} & z_{2}\\ 0 & 1 & z_{3} \\ 0 & 0 & 1 \end{array} \right)}\mid z_{i}\in \mathbb{C} \Bigg\}\subset \mathrm{GL}(3; \mathbb{C}). $$ As complex manifolds, $\mathrm{H}(3; \mathbb{C})$ is isomorphic to $\mathbb{C}^{3}$. Consider the discrete group $\mathrm{G}_3:=\mathrm{GL}(3; \mathbb{Z}[i])\cap \mathrm{H}(3; \mathbb{C})$, where $\mathbb{Z}[i]=\{a+bi\mid a,b\in \mathbb{Z}\}$ is the Gaussian integers. The left multiplication gives a natural $\mathrm{G_3}$--action on $\mathrm{H}(3; \mathbb{C})$, and a correspondent faithful $\mathrm{G}_3$--action on $\mathbb{C}^{3}$ given by $ (a_{1}, a_{2}, a_{3})\cdot(z_{1}, z_{2}, z_{3}):=(z_{1}+a_{1}, z_{2}+a_{1}z_{3}+a_{2}, z_{3}+a_{3}), $ where $a_{1}, a_{2}, a_{3}\in \mathbb{Z}[i]$. Consequently, the associated $\mathrm{G}_3$--quotient space $$ \mathbb{I}_3:=\mathbb{C}^{3}/ \mathrm{G}_3 $$ is a compact complex threefold, which is called {\it Iwasawa manifold}. A basis of the space $(\mathfrak{g}_{\mathbb{C}}^*)^{1,0}$ of left--invariant holomorphic differential forms of $\mathrm{H}(3; \mathbb{C})$ is $$w_1=dz_1, w_2=dz_2-z_1dz_3, w_3=dz_3.$$ Therefore $dw_1=dw_3=0, dw_2=-w_1\wedge w_3, \bar{\partial}(\bar{w_1})=\bar{\partial}(\bar{w_3})=0, \bar{\partial}(\bar{w_2})=-\bar{w_1}\wedge \bar{w_3}.$ Dually, a basis of Lie algebra $\mathfrak{g}_{\mathbb{C}}^{1,0}$ of left--invariant holomorphic vector fields of $\mathrm{H}(3; \mathbb{C})$ is $$X_1=\frac{\partial}{\partial z_1}, X_2=\frac{\partial}{\partial z_2}, X_3=\frac{\partial}{\partial z_3}+z_1\frac{\partial}{\partial z_2}$$ with relations $[X_1,X_2]=[X_2,X_3]=0, [X_1,X_3]=X_2$. As a consequent, $\mathbb{I}_3$ is complex parallelizable, hence it is a Calabi--Yau manifold. Furthermore, a $\mathrm{H}(3; \mathbb{C})$--left--invariant holomorphic bi--vector field is in the form of $\pi=c_1X_1\wedge X_2+c_2X_1\wedge X_3+c_3X_2\wedge X_3$. One can check that $[\pi,\pi]=0$ if and only if $c_2=0$. Note $\pi$ is the linear combination of two compatible Poisson bi--vector fields $\pi_{12}=X_1\wedge X_2$ and $\pi_{23}=X_2\wedge X_3$. But on $\wedge^{\bullet,\bullet}\mathfrak{g}^*_{\mathbb{C}}$, $\partial_{\pi_{12}}=\partial_{\pi_{23}}=0$ which indicates $\partial_\pi=0.$ Hence by Lemma \ref{nil}, the Dolbeault--Koszul--Brylinski spectral sequence of $(\mathbb{I}_3,\pi)$ degenerates at $E_1$--page. However $(\mathbb{I}_3,\pi)$ neither satisfies the $\partial_{}\bar{\partial}$--lemma nor satisfies the $\partial_{\pi}\bar{\partial}$--lemma. Indeed, by the structure equations, we have that $$-w_1\wedge\bar{w_1}\wedge \bar{w_3}\notin\mathrm{im}\partial_{}\bar{\partial},$$ while $$ -w_1\wedge\bar{w_1}\wedge \bar{w_3} =(\partial_{}+\bar{\partial})(w_1\wedge\bar{w_2}) \in\ker\, \partial_{}\cap \ker\, \bar{\partial}\cap \mathrm{im} \,(\partial_{}+\bar{\partial}) $$ which indicates that $(\mathbb{I}_3,\pi)$ does not satisfy the $\partial_{}\bar{\partial}$--lemma (see also \cite{Angella13}). Meanwhile, since $\mathrm{im}\partial_{\pi}\bar{\partial}=\emptyset$, $$ -\bar{w_1}\wedge \bar{w_3}=(\partial_{\pi}+\bar{\partial})(\bar{w_2}) \in\ker\, \partial_{\pi}\cap \ker\, \bar{\partial}\cap \mathrm{im} \,(\partial_{\pi}+\bar{\partial}) \neq\emptyset. $$ Therefore $\mathbb{I}_3$ does not satisfy the $\partial_{\pi}\bar{\partial}$--lemma. \subsection{A six--dimensional complex nilmanifold} Motivated by the construction of Iwasawa manifold, we consider complex nilpotent Lie group $$ G=\Bigg\{A= \small{\left(\begin{array}{cccc} 1 &z_1 & z_{2} & z_{3}\\ 0 &1 & z_{4} & z_{5}\\ 0 & 0 & 1 & z_{6} \\ 0& 0 & 0 & 1 \end{array} \right)}\mid z_{i}\in \mathbb{C} \Bigg\}\subset \mathrm{GL}(4; \mathbb{C}). $$ As complex manifolds, $G$ is isomorphic to $\mathbb{C}^{6}$. Consider the discrete group $H:=\mathrm{Gl}(4; \mathbb{Z}[i])\cap G$. Analogously, the left multiplication gives a natural $H$--action on $G$. The associated $H$--quotient space $ \mathbb{I}_6:=G/ H $ is a 6--dimensional compact complex manifold. Moreover, a basis of the space $(\mathfrak{g}_{\mathbb{C}}^*)^{1,0}$ of left--invariant holomorphic differential forms of $G$ is given by: \begin{eqnarray*} &&w_1 =dz_1,\;\; w_2=dz_2-z_1dz_4,\;\; w_3=dz_3-z_1dz_5+(z_1z_4-z_2)dz_6, \\ && w_4 =dz_4,\;\; w_5=dz_5-z_4dz_6,\;\; w_6=dz_6 \qquad \qquad\qquad \end{eqnarray*} with structure equations $$ \begin{cases} dw_1=dw_4=dw_6=0, & \\ dw_2=-w_1\wedge w_4, & \\ dw_3=-w_1\wedge w_5-w_2\wedge w_6, & \\ dw_5=-w_4\wedge w_6 . & \end{cases} $$ Dually, the Lie algebra $\mathfrak{g}_{\mathbb{C}}^{1,0}$ of left--invariant holomorphic vector fields of $G$ has a basis \begin{eqnarray*} &&X_{1}=\frac{\partial}{\partial z_1}, \;\; X_2=\frac{\partial}{\partial z_2}, \;\; X_3=\frac{\partial}{\partial z_3},\qquad \qquad\qquad \\ &&X_4=\frac{\partial}{\partial z_4}+z_1\frac{\partial}{\partial z_2}, \;\; X_5=\frac{\partial}{\partial z_5}+z_1\frac{\partial}{\partial z_3},\;\; X_6=\frac{\partial}{\partial z_6}+z_2\frac{\partial}{\partial z_3}+z_4\frac{\partial}{\partial z_5} \end{eqnarray*} whose only non--trivial relations are $[X_1,X_4]=X_2, [X_1,X_5]=X_3=[X_2,X_6], [X_4,X_6]=X_5$. Consequently, $\mathbb{I}_6$ is holomorphically parallelizable, a $6$--dimensional non--K\"{a}hler Calabi--Yau manifold since it also does not satisfy the $\partial_{}\bar{\partial}$--lemma: $$-w_1\wedge\bar{w_1}\wedge \bar{w_4}\notin\mathrm{im}\partial_{}\bar{\partial}$$ while $$ -w_1\wedge\bar{w_1}\wedge \bar{w_4} =(\partial_{}+\bar{\partial})(w_1\wedge\bar{w_2}) \in\ker\, \partial_{}\cap \ker\, \bar{\partial}\cap \mathrm{im} \,(\partial_{}+\bar{\partial}). $$ The Hodge diamond of $\mathbb{I}_6$ states as(see \cite[Appendix A]{CCYY22}): \begin{center} \footnotesize{\begin{tikzpicture}[scale=0.08] \draw (30,30) node[left] {$1$}; \draw (25,25) node[left] {$6$}; \draw (35,25) node[left] {$3$}; \draw (20,20) node[left] {$15$}; \draw (30,20) node[left] {$18$}; \draw (40,20) node[left] {$5$}; \draw (15,15) node[left] {$20$}; \draw (25,15) node[left] {$45$}; \draw (35,15) node[left] {$30$}; \draw (45,15) node[left] {$6$}; \draw (10,10) node[left] {$15$}; \draw (20,10) node[left] {$60$}; \draw (30,10) node[left] {$75$}; \draw (40,10) node[left] {$36$}; \draw (50,10) node[left] {$5$}; \draw (5,5) node[left] {$6$}; \draw (15,5) node[left] {$45$}; \draw (25,5) node[left] {$100$}; \draw (35,5) node[left] {$90$}; \draw (45,5) node[left] {$30$}; \draw (55,5) node[left] {$3$}; \draw (0,0) node[left] {$1$}; \draw (10,0) node[left] {$18$}; \draw (20,0) node[left] {$75$}; \draw (30,0) node[left] {$120$}; \draw (40,0) node[left] {$75$}; \draw (50,0) node[left] {$18$}; \draw (60,0) node[left] {$1$}; \draw (5,-5) node[left] {$3$}; \draw (15,-5) node[left] {$30$}; \draw (25,-5) node[left] {$90$}; \draw (35,-5) node[left] {$100$}; \draw (45,-5) node[left] {$45$}; \draw (55,-5) node[left] {$6$}; \draw (10,-10) node[left] {$5$}; \draw (20,-10) node[left] {$36$}; \draw (30,-10) node[left] {$75$}; \draw (40,-10) node[left] {$60$}; \draw (50,-10) node[left] {$15$}; \draw (15,-15) node[left] {$6$}; \draw (25,-15) node[left] {$30$}; \draw (35,-15) node[left] {$45$}; \draw (45,-15) node[left] {$20$}; \draw (20,-20) node[left] {$5$}; \draw (30,-20) node[left] {$18$}; \draw (40,-20) node[left] {$15$}; \draw (25,-25) node[left] {$3$}; \draw (35,-25) node[left] {$6$}; \draw (30,-30) node[left] {$1$}; \end{tikzpicture}} \end{center} Here we only consider some special holomorphic Poisson structures on $\mathbb{I}_6$ given by $G$--left--invariant holomorphic bi--vector fields: $$ \pi_{1}=X_2\wedge X_3,\;\;\; \pi_{2}=X_1\wedge X_3. $$ Akin to the Iwasawa manifold, the holomorphic Koszul--Brylinski homology of $\mathbb{I}_6$ can be computed in terms of the total cohomology of the double complex $(\wedge^{\bullet,\bullet}\mathfrak{g}^{\ast}_{\mathbb{C}},\partial_\pi,\bar{\partial})$. For the simplicity, we write $w^{i_{1}\cdots i_{p}\bar{j_{1}}\cdots \bar{j_{q}}} =w^{i_{1}}\wedge \cdots \wedge w^{i_{p}}\wedge w^{\bar{j_{1}}}\wedge\cdots\wedge w^{\bar{j_{q}}}$, for any $1\leq p,q\leq6$. \subsubsection{$\partial_{\pi}\bar{\partial}$--lemma on $(\mathbb{I}_{6},\pi_1)$} Consider the holomorphic Poisson manifold $(\mathbb{I}_{6},\pi_1)$. The only possible candidates of non--closed elements are \begin{eqnarray*} \partial_{\pi_1}w^{23i_{1}\cdots i_{p-2}\bar{j_{1}}\cdots \bar{j_{q}}}&=& (\iota_{\pi_1}\circ \partial-\partial\circ \iota_{\pi_1})w^{23i_{1}\cdots i_{p-2}\bar{j_{1}}\cdots \bar{j_{q}}}\\ &=&\iota_{\pi_1}(w^{23}\wedge \partial w^{i_{1}\cdots i_{p-2}\bar{j_{1}}\cdots \bar{j_{q}}}) -\partial w^{i_{1}\cdots i_{p-2}\bar{j_{1}}\cdots \bar{j_{q}}}\\ &=&\partial w^{i_{1}\cdots i_{p-2}\bar{j_{1}}\cdots \bar{j_{q}}} -\partial w^{i_{1}\cdots i_{p-2}\bar{j_{1}}\cdots \bar{j_{q}}}=0. \end{eqnarray*} This means the Dolbeault--Koszul--Brylinski spectral sequence of $(M,\pi_1)$ degenerates at $E_1$--page. Analogous to $\mathbb{I}_3$, $(\mathbb{I}_6,\pi_1)$ does not satisfy the $\partial_{\pi}\bar{\partial}$--lemma since $\mathrm{im}\partial_{\pi}\bar{\partial}=\emptyset$, while $$ -\bar{w_1}\wedge \bar{w_4}=(\partial_{\pi}+\bar{\partial})(\bar{w_2}) \in\ker\, \partial_{\pi}\cap \ker\, \bar{\partial}\cap \mathrm{im} \,(\partial_{\pi}+\bar{\partial}) \neq\emptyset. $$ \subsubsection{$\partial_{\pi}\bar{\partial}$--lemma on $(\mathbb{I}_{6},\pi_2)$} Consider the holomorphic Poisson manifold $(\mathbb{I}_{6},\pi_2)$. Note that $\mathfrak{g}^{6,0}=\langle w^{123456}\rangle$, $\partial_{\pi_2} \mathfrak{g}^{6,0}=0$, $\bar{\partial} \mathfrak{g}^{6,0}=0$, thus we have \begin{center} $H_{0}(X, {\pi_2})=\langle [w^{123456}]\rangle=\mathbb{C}.$ \end{center} Furthermore, observe that $\mathfrak{g}^{5,0} =\langle w^{23456},w^{13456},w^{12456},w^{12356},w^{12346},w^{12345}\rangle$, $\bar{\partial} \mathfrak{g}^{5,0}=0$, and the only non--closed monomial of $\mathfrak{g}^{5,0}$ under $\partial_{\pi_{2}}$ is $ \partial_{\pi_{2}} w^{12356} =-w^{1456}. $ Meanwhile, since $\mathfrak{g}^{6,1} =\langle w^{123456\bar{1}},w^{123456\bar{2}},w^{123456\bar{3}}, w^{123456\bar{4}},w^{123456\bar{5}},w^{123456\bar{6}}\rangle$, $\partial_{\pi_2} \mathfrak{g}^{6,1}=0$, and the only non--closed monomials of $\mathfrak{g}^{6,1}$ under $\bar{\partial}$ are: $$\bar{\partial}w^{123456\bar{2}}=-w^{123456\bar{1}\bar{4}}, \quad \bar{\partial}w^{123456\bar{3}}=-w^{123456\bar{1}\bar{5}}-w^{\bar{2}\bar{6}},\quad \bar{\partial}w^{123456\bar{5}}=-w^{123456\bar{4}\bar{6}}, $$ Consequently, we have that \begin{multline*} H_{1}(\mathbb{I}_{6},{\pi_2})= \\ \langle [w^{23456}],[w^{13456}],[w^{12456}],[w^{12346}],[w^{12345}], [w^{123456\bar{1}}],[w^{123456\bar{4}}],[w^{123456\bar{6}}]\rangle=\mathbb{C}^{8}. \end{multline*} But $6+3>8$, by the Lemma \ref{deg}, this means the Dolbeault--Koszul--Brylinski spectral sequence of $(\mathbb{I}_6,\pi_2)$ does not degenerate at $E_1$--page. \subsubsection{Two submanifolds of $(\mathbb{I}_{6},\pi_2)$} Let $$\Gamma_1= \Bigg\{\small{\left(\begin{array}{cccc} 1 &z_{1} & z_{2} & z_{3}\\ 0 &1 & z_{4} & a_{24}\\ 0 & 0 & 1 & a_{34} \\ 0& 0 & 0 & 1 \end{array} \right)}\mid z_{i}\in \mathbb{C},a_{24},a_{34}\in\mathbb{Z}[i] \Bigg\}. $$ Naturally $(Y_1=\Gamma_1/H,\pi_2|_{Y_1}=X_1\wedge X_3)$ is a 4--dimensional nilmanifold, a closed holomorphic Poisson submanifold of $\mathbb{I}_{6}$. With the same arguments, one can check that the Dolbeault--Koszul--Brylinski spectral sequence of $(Y_1,\pi_3|_{Y_1})$ do not degenerate at $E_1$ page. Take $$\Gamma_2= \Big\{A= \small{\left(\begin{array}{cccc} 1 &z_{1} & z_{2} & z_{3}\\ 0 &1 & a_{23} & a_{24}\\ 0 & 0 & 1 & a_{34} \\ 0& 0 & 0 & 1 \end{array} \right)}\mid z_{i}\in \mathbb{C},a_{23},a_{24},a_{34}\in\mathbb{Z}[i] \Big\}. $$ Naturally $(Y_2=\Gamma_2/H,\pi_2|_{Y_2}=X_1\wedge X_3)$ is a complex 3--torus, a closed holomorphic Poisson submanifold of $\mathbb{I}_{6}$. Hence by Corollary \ref{Kahler}, the Dolbeault--Koszul--Brylinski spectral sequence of $(Y_2,\pi_3|_{Y_2})$ degenerates at $E_1$ page while the one of $(\mathbb{I}_{6}, \pi_2)$ does not. \subsection{Nakamura manifold}\label{Nak} In this subsection, based on \cite[Example 3.4]{AK17}, we consider a special solvmanifold: the (holomorphically parallelizable) Nakamura manifold. Consider the complex Lie group $G:=\mathbb{C}\ltimes_\phi\mathbb{C}^2$ where $\phi(z)= \left(\begin{array}{cccc} e^z &0 \\ 0 &e^{-z} \end{array} \right)\in SL(2,\mathbb{C}). $ There exist $a+\sqrt{-1}b\in \mathbb{C}$ and $c+\sqrt{-1}d\in \mathbb{C}$ such that $\mathbb{Z}(a+\sqrt{-1}b)+\mathbb{Z}(c+\sqrt{-1}d)$ is a lattice in $\mathbb{C}$ and $\phi(a+\sqrt{-1}b)$ and $\phi(a+\sqrt{-1}b)$ are conjugate to elements of $SL(4,\mathbb{Z})$, where we regard $SL(2,\mathbb{C})\subset SL(4,\mathbb{R})$. Hence there exists a lattice $\Gamma=(\mathbb{Z}(a+\sqrt{-1}b)+\mathbb{Z}(c+\sqrt{-1}d))\ltimes_\phi\Gamma_{\mathbb{C}^2}$ of $G$, where $\Gamma_{\mathbb{C}^2}$ is a lattice of $\mathbb{C}^2$. The holomorphically parallelizable solvmanifold $M=G/\Gamma$ is called (holomorphically parallelizable) Nakamura manifold. Denote by $\mathfrak{g}_{\mathbb{C}}$ the complexification of the Lie algebra $\mathfrak{g}$ of $G$. Once taking the coordinate $(z_1,z_2,z_3)$ of $G$, we obtain a basis $\{ X_1=\frac{\partial}{\partial z_1}, X_2=e^{z_1}\frac{\partial}{\partial z_2}, X_3=e^{-z_1}\frac{\partial}{\partial z_3} \}$ of $(\mathfrak{g}_{\mathbb{C}})^{1,0}$ with Lie bracket $$ [X_1,X_2]=X_2, [X_1,X_3]=-X_3, [X_2,X_3]=0. $$ Meanwhile, the dual basis of the space $(\mathfrak{g}_{\mathbb{C}}^*)^{1,0}$ is $\{ w_1=dz_1, w_2=e^{-z_1}dz_2, w_3=e^{z_1}dz_3 \}$ with the structure equations $$ dw_1=0, dw_2=-w_1\wedge w_2, dw_3=w_1\wedge w_3. $$ Therefore the $G$--left--invariant holomorphic bi--vector fields $\pi_{12}=X_1\wedge X_2$ and $\pi_{23}=X_2\wedge X_3$ are Poisson. Moreover, there exists a finite--dimensional subcomplex $B_\Gamma^{\bullet}$ of $(A^{0,\bullet}_M,\bar{\partial})$ such that the inclusions $\iota: B_\Gamma^{\bullet}\hookrightarrow (A^{0,\bullet}_M,\bar{\partial})$ and $$\iota: \wedge^{\bullet}(\mathfrak{g}_{\mathbb{C}}^*)^{1,0}\otimes B_\Gamma^{\bullet} \hookrightarrow (A^{\bullet,\bullet}_M,\bar{\partial})$$ are quasi--isomorphisms. Analogous to the Lemma \ref{nil}, we have that the holomorphic Koszul--Brylinski homology of $(M,\pi)$ can be computed in terms of the total cohomology of the double complex $(\wedge^{\bullet}(\mathfrak{g}_{\mathbb{C}}^*)^{1,0}\otimes B_\Gamma^{\bullet}, \partial_{\pi}, \bar{\partial})$ if $\pi$ is $G$--left--invariant. However, the Dolbeault cohomology of $M$ depends on the lattices $\Gamma$. More precisely, the subcomplex $B_\Gamma^{\bullet}$ depends on two distinguished cases: \begin{itemize} \item[(1)]Both $b\in \mathbb{Z}\pi$ and $d\in \mathbb{Z}\pi$. In this case, $ B_\Gamma^{\bullet}=\wedge^\bullet \mathbb{C}\langle d\bar{z_1}, e^{-z_1}d\bar{z_2},e^{z_1}d\bar{z_3}\rangle $ ; \item[(2)]Either $b\notin \mathbb{Z}\pi$ or $d\notin \mathbb{Z}\pi$. In this case, \begin{eqnarray*} B_\Gamma^{1} &=& \mathbb{C}\langle d\bar{z_1}\rangle, \\ B_\Gamma^{2} &=& \mathbb{C}\langle d\bar{z_2}\wedge d\bar{z_3}\rangle, \\ B_\Gamma^{3} &=& \mathbb{C}\langle d\bar{z_1}\wedge d\bar{z_2}\wedge d\bar{z_3}\rangle. \end{eqnarray*} \end{itemize} Note for both two cases, the Dolbeault operator $\bar{\partial}$ on $B_\Gamma^{\bullet}$ is trivial. Therefore we have that the Nakamura manifold $M$ always do not satisfy the $\partial_{}\bar{\partial}$--lemma since $\mathrm{im}\partial_{}\bar{\partial}=\emptyset$, while $$ -w_1\wedge w_2\wedge \bar{w_1} =(\partial_{}+\bar{\partial})(w_2\wedge\bar{w_1}) \in\ker\, \partial_{}\cap \ker\, \bar{\partial}\cap \mathrm{im} \,(\partial_{}+\bar{\partial}). $$ Furthermore, if we consider $(M, \pi_{12}=X_1\wedge X_2)$, one can check that $(M, \pi_{12})$ do not satisfy the $\partial_{\pi}\bar{\partial}$--lemma since $\mathrm{im}\partial_{\pi_{12}}\bar{\partial}=\emptyset$, while $$ -1=(\partial_{\pi_{12}}+\bar{\partial})(w_2) \in\ker\, \partial_{\pi}\cap \ker\, \bar{\partial}\cap \mathrm{im} \,(\partial_{\pi}+\bar{\partial}) \neq\emptyset. $$ However, if we consider $(M, \pi_{23}=X_2\wedge X_3)$, then by the direct calculations, we have that on $\wedge^{\bullet}(\mathfrak{g}_{\mathbb{C}}^*)^{1,0}\otimes B_\Gamma^{\bullet}$, both differentials $\partial_{\pi}$ and $\bar{\partial}$ are trivial, hence $(M, \pi_{23})$ satisfies the $\partial_{\pi}\bar{\partial}$--lemma.

34fb058eac4b9429a6b2482666e5b9210249726b

\section*{Abstract} {\bf GeV-scale dark matter particles with strong coupling to baryons evade the standard direct detection limits as they are efficiently stopped in the overburden and, consequently, are not able to reach the underground detectors. On the other hand, it has been shown that it is possible to probe this parameter space taking into account the flux of dark matter particles boosted by interactions with cosmic rays. We revisit these bounds paying particular attention to interactions of the relativistic dark matter particles in the Earth's crust. The effects of nuclear form factors, inelastic scattering and extra dependence of the cross section on transferred momentum (e.g., due to presence of light mediators) are studied and are found to be crucial for answering the question as to whether the window for GeV-scale strongly interacting dark matter is closed or not. } \section{Introduction} \label{sec:intro} Direct detection experiments are trying to shed light on the nature of dark matter (DM) but, although great progress was made in past years \cite{LZ:2022ufs,PandaX-4T:2021bab,XENON:2018voc}, their reach in the space of DM mass and couplings is still limited. In particular, DM interactions with nucleons are probed by experiments looking for collisions of DM with nuclei in the detector which requires a certain minimal DM kinetic energy to trigger a detectable signal. Given the fact that halo DM particles reach the Earth at velocities of the order of $\sim 10^{-3}\,c$, such particles don't attain sufficient kinetic energy if their mass is too low. Hence, sub-GeV DM is typically not probed by standard direct detection experiments. Another ``blind spot'' for standard direct detection experiments is a result of the fact that if DM interacts too strongly with nuclei, it is efficiently stopped in the Earth's crust and does not reach the underground detectors. This latter issue was addressed, e.g., by the dedicated CRESST surface run~\cite{CRESST:2017ues} in the context of direct detection. Additionally, strongly coupled sub-GeV DM has been further constrained by the possible effects on structure formation~\cite{Rogers:2021byl,Maamari:2020aqz} or on the cooling of gas clouds near the Galactic Centre~\cite{Bhoonah:2018gjb}. Nonetheless, state of the art probes may still leave room for strongly interacting DM candidates like the stable ``sexaquark'' state with a mass around 2~GeV~\cite{Farrar:2020zeo}. In this work we concentrate on yet another constraint on DM with strong couplings to baryons. Namely, it was shown in Ref.~\cite{Bringmann:2018cvk} that collisions of cosmic ray (CR) nuclei with such DM in the Galactic halo result in a flux of relativistic DM particles coming to Earth (CRDM flux). These particles can trigger detectable signal in standard direct detection experiments despite their sub-GeV mass. In this way, the Xenon-1T limits~\cite{XENON:2018voc} were reinterpreted to constrain the spin-independent DM-nucleon scattering cross sections roughly between $10^{-31}$ and $10^{-28}\,$cm$^2$ for DM masses up to about 2~GeV~\cite{Bringmann:2018cvk}. It is worth stressing that also in the CRDM case, the upper boundary of the excluded region is set by the fact that DM coupled too strongly to nucleons cannot reach the underground detectors. In this text, we focus on the attenuation of the CRDM flux in the Earth's crust, and we show that a more precise treatment (described in section~\ref{sec:att}) leads to extension of CRDM limits to larger DM masses. The consequence of this is to close the parameter space for DM-nucleon cross sections exceeding $10^{-30}\,$cm$^2$ (see section~\ref{sec:limits}). As discussed in section~\ref{sec:conclusion}, we checked that our conclusions hold for a range of generic DM scenarios such as those where interactions with nucleons proceed via light mediators. While the main results are highlighted in this text, the technical details of the modeling and particle physics scenarios can be found in \cite{Alvey:2022pad}. For the analysis performed in this work we used the numerical tool {\sf DarkSUSY}~\cite{Bringmann:2018lay} and the updated routines will be included in the next public release of this code. \section{Attenuation of the CRDM flux in the Earth's crust} \label{sec:att} The evolution of the DM kinetic energy $T_\chi^z$ at depth $z$ can be described by the energy loss equation: \begin{equation} \label{eq:eloss} \frac{dT_\chi^z}{dz}=-\sum_N n_N\int_0^{\omega_\chi^\mathrm{max}}\!\!\!\mathrm{d}\omega_\chi\,\frac{\mathrm{d} \sigma_{\chi N}}{\mathrm{d}\omega_\chi} \omega_\chi\,, \end{equation} where the sum runs over the nuclei $N$ in the overburden, each with a number density $n_N$ and a differential cross section $\mathrm{d} \sigma_{\chi N}/\mathrm{d}\omega_\chi$ describing the scattering with DM particles in terms of the kinetic energy lost by the DM particle, $\omega_\chi$. It is this cross section that has to be treated more precisely in order to obtain realistic predictions as to the parameter space that is excluded by the non-observation of the CRDM component in detectors like Xenon-1T. In particular, we concentrate on detailed modeling of the nuclear form factors in the elastic contribution to $\mathrm{d} \sigma_{\chi N}/\mathrm{d}\omega_\chi$ and on the effect of including inelastic scattering in following sections~\ref{sec:ff} and~\ref{sec:inel}, respectively. \subsection{Effect of nuclear form factors}\label{sec:ff} For the calculation of the elastic contribution to the DM-nucleus scattering cross section, we follow the approach of standard direct detection experiments that translate their observations into limits on the spin-independent DM-nucleon cross section $\sigma_{SI}$ using its following relation to the differential DM-nucleus cross section: \begin{equation}\label{eq:siconst} \left.\frac{\mathrm{d} \sigma_{\chi N}}{\mathrm{d}\omega_\chi}\right|_{\mathrm{el}} = A^2\frac{\mu_{\chi N}^2}{\mu_{\chi p}^2} \times \frac{\sigma_{\rm SI}}{\omega_\chi^{\rm max}} \times G^2(Q^2)\,. \end{equation} Here $\mu$ refers to reduced mass of the given 2-particle system, $\omega_\chi^{\rm max}$ is the maximum energy lost by the DM particle\footnote{In the case of elastic scattering, $\omega_\chi$ is equal to the kinetic energy of the recoiled nucleus $T_N$ and for a given nucleus mass, $\omega_\chi^{\rm max}=T_N^{\rm max}$ is uniquely determined by the DM kinetic energy and mass.} and $A$ is the atomic mass number of the nucleus. The factor $A^2$ then captures the coherent enhancement of the scattering cross section characteristic for spin-independent couplings of DM to nucleons (under the assumption of an equal coupling of DM to protons and neutrons). Nuclear form factor $G(Q^2)$, on the other hand, expresses the loss of coherence across the nucleus for large momentum transfers $Q^2 = 2m_N \omega_\chi$. Since, especially for heavy nuclei, $G$ is a steeply falling function of $Q^2$, the form factors lead to a significant reduction of the elastic cross section in the case of relativistic CRDM particles, i.e., reduced attenuation of the CRDM flux. The effect of nuclear form factors in the attenuation part was neglected in the initial study~\cite{Bringmann:2018cvk}, but was added in a later re-analysis~\cite{Xia:2021vbz}. In our work we identify the importance of the form factors for setting CRDM limits, and compared to~\cite{Xia:2021vbz}, include the more accurate model-independent form factors \cite{Duda:2006uk}. On the other hand, we point out below that the almost vanishing cross section at large momentum transfers, as obtained when considering only the contribution of Eq.~\eqref{eq:siconst}, is unphysical since the additional contribution of inelastic scattering becomes relevant for CRDM particles scattering on nuclei. \subsection{Effect of inelastic scattering}\label{sec:inel} Although the CRDM flux peaks for kinetic energies between 10 and 100~MeV depending on DM mass (see~\cite{Alvey:2022pad} for details), a significant amount of DM particles with kinetic energies larger than 100~MeV may arrive on Earth. For these, the contribution of inelastic scattering may easily dominate the right-hand side of Eq.~\eqref{eq:eloss}. Indeed, by comparison with analogous processes in case of neutrino-nucleus scattering (see, e.g., \cite{Formaggio:2012cpf} for a review), DM particles with kinetic energies $T_\chi \gtrsim 0.1~$GeV are expected to effectively scatter off individual nucleons for large energy transfers $\omega_\chi$ (via so-called quasi-elastic scattering), for $T_\chi \gtrsim 0.3~$GeV the excitation of hadronic resonances becomes possible and, finally, for kinetic energies of a few GeV, deep inelastic scattering becomes the relevant contribution at large $\omega_\chi$ where the individual quarks and gluons are resolved. Calculation of the corresponding cross sections has to take into account the effect of the nuclear environment (like the nuclear potential or spin statistics) and cannot, hence, be easily performed analytically. For this reason, we estimate the inelastic contribution to $\mathrm{d} \sigma_{\chi N}/\mathrm{d}\omega_\chi$ by first using the numerical code \texttt{GiBUU}~\cite{Buss:2011mx} to calculate neutrino-nucleus cross sections. We then rescale the result appropriately in order to take into account the properties of DM scattering \cite{Alvey:2022pad}. Irrespective of the precise implementation, we concluded that the inelastic scattering leads to a significant increase of the stopping power in soil compared to what was assumed in~\cite{Xia:2021vbz} and, hence, a significant reduction in the size of the excluded region, as described below. \section{Exclusion limits}\label{sec:limits} In the section, we present the results for the case of a ``constant'' DM-nucleus cross section~\eqref{eq:siconst}.\footnote{The differential cross section where the only dependence on the transferred momentum comes from the nuclear form factors is a good approximation for contact interactions in the highly non-relativistic limit. For the CRDM particles, full $Q^2$-dependence of the cross section may influence the final results, see the comments in section~\ref{sec:conclusion}.} As can be seen in Fig.~\ref{figConst}, CRDM limits including both the effect of form factors and inelastic scattering in the attenuation part are stronger than the conservative limits of~\cite{Bringmann:2018cvk}, especially for heavier DM where the form factors play significant role. On the other hand, the results of~\cite{Xia:2021vbz} clearly overestimate the excluded region. One can see that the true limits correspond roughly to the case where inelastic scattering is neglected, but at the same time, the CRDM flux is artificially cut at kinetic energies around $0.2\,$GeV. This confirms our findings related to the attenuation of the CRDM flux, namely that DM particles with $\mathcal{O}(0.1)\,$GeV kinetic energies become efficiently stopped by inelastic scattering. In the right panel of Fig.~\ref{figConst} we compare the CRDM excluded region to limits based on the Lyman-$\alpha$ forest~\cite{Rogers:2021byl}, the Milky Way satellite population~\cite{Maamari:2020aqz}, gas clouds in the Galactic Centre region~\cite{Bhoonah:2018gjb}, the XQC experiment~\cite{McCammon:2002gb,Mahdawi:2018euy}, and a recently analysed storage dewar experiment~\cite{Neufeld:2019xes,Xu:2021lmg}. We also present the limits based on the CRESST surface run~\cite{CRESST:2017ues,Emken:2018run} (solid green lines), together with the alternative limits based on the assumption of a thermalization efficiency of $\epsilon_{\rm th}=2$\,\%~\cite{Mahdawi:2018euy} and $\epsilon_{\rm th}=1$\,\%~\cite{Xu:2020qjk} (green dashed and dash-dotted lines, respectively), which is significantly more pessimistic than the one adopted in the CRESST analysis. \begin{figure \centering \includegraphics[width=\textwidth]{figure_four_wl.pdf} \caption{ Limits on a constant spin-independent DM-nucleon scattering cross section as a function of the DM mass (solid lines). In the left panel, dash-dotted lines show the excluded region that results when assuming a constant cross section in the attenuation part (as in Ref.~\cite{Bringmann:2018cvk}). Dashed lines show the effects of adding form factors in the attenuation part, but no inelastic scattering, resulting in limits similar to those derived in Ref.~\cite{Xia:2021vbz}. For the latter case, for comparison, we also show the effect of artificially cutting the incoming CRDM flux at the indicated energies. In the right panel, we compare the CRDM limits to other published constraints. } \label{figConst} \end{figure} \section{Discussion and conclusions}\label{sec:conclusion} As can be seen in Fig.~\ref{figConst}, irrespective of the thermalization efficiency assumed for the CRESST experiment, there is no parameter space left unconstrained for DM-nucleon cross sections exceeding $10^{-30}\,$cm$^2$ in the entire MeV to GeV DM mass range.\footnote{Note that strongly coupled DM lighter than about 10~MeV is nonetheless excluded by the BBN constraints~\cite{Krnjaic:2019dzc}.} Of course, the simplified DM-nucleus cross section assumed here is in contrast to the more complex dependence of the cross section on the transferred momentum and invariant mass that is expected for realistic sub-GeV DM models. Several of these more generic DM scenarios were considered in our study~\cite{Alvey:2022pad}, namely, DM interacting via scalar or vector mediators with MeV-to-GeV masses and finite-size DM. The extra $Q^2$-suppression of the cross section leads to a decrease in the CRDM flux in certain cases, consequently, a tiny open parameter space can appear for a narrow range of mediator masses, but only if the CRESST thermalization efficiency was indeed as low as 2\,\%. We note that such an assumption is not supported by data or simulations \cite{florian}. Our conclusion is, hence, that there is generically no room to hide for sub-GeV DM with DM-nucleon cross sections exceeding $10^{-30}\,$cm$^2$. \section*{Acknowledgements} HK thanks James Alvey and Torsten Bringmann for a fruitful collaboration and the organizers of the IDM conference for the opportunity to present this work. HK is supported by the ToppForsk-UiS Grant No. PR-10614.

21f16f080d3e31c5117888442c0e55ee30f5c843

\section{Proofs} \printProofs \end{document} \section{Background of Optimal Transport} \label{sec:ot} This section introduces the background of optimal transport (OT), including both the static and dynamic formulations. Of special importance is the dynamic formulation, which is closely related to the rectified flow approach. The readers can find systematic introductions to OT in a collection of excellent textbooks \cite{villani2021topics, figalli2021invitation, ambrosio2021lectures, peyre2019computational, ollivier2014optimal, santambrogio2015optimal, villani2009optimal}. \paragraph{Static formulations} The optimal transport problem was first formulated by Gaspard Monge in 1781 when he studied the problem of how to redistribute mass, e.g., a pile of soil, with minimal effort. Monge's problem can be formulated as \bbb \label{equ:m} \inf_{T} \e{c(T(X_0)- X_0)} ~~~~s.t.~~~~ \law(T(X_0))=\tg_1, ~~~ \law(X_0)=\tg_0, \eee where we minimize the $c$-transport cost in the set of deterministic couplings $(X_0,X_1)$ that satisfy $X_1 = T(X_0)$ for a transport mapping $T\colon \RR^d\to \RR^d$. The Monge–Kantorovich (MK) problem in \eqref{equ:mk} is the relaxation of \eqref{equ:m} to the set of all (deterministic and stochastic) couplings of $\tgg$. The two problems are equivalent when the optimum of \eqref{equ:mk} is achieved by a deterministic coupling, which is guaranteed if $\tg_0$ is an absolutely continuous measure on $\RR^d$. A key feature of the MK problem is that it is a linear programming w.r.t. the law of the coupling $(X_0,X_1)$, and yields a dual problem of form: % \bbb \label{equ:dmk} \sup_{\mu, \nu} \tg_1(\mu) - \tg_0(\nu) ~~s.t.~~ \mu(x_1) - \nu(x_0)\leq c(x_1-x_0),~~~~\forall (x_0,x_1), \eee where we write $\tg_1(\mu) \defeq \int \mu(x) \d \tg_1(x)$, and $\mu,\nu$ are optimized in all functions from $\RR^d$ to $\RR$. For any coupling $(X_0,X_1)$ of $\tgg$, and $(\mu,\nu)$ satisfying the constraint in \eqref{equ:dmk}, it is easy to see that \bbb \label{equ:dualderive} \E[c(X_1-X_0)] \geq \E[\mu(X_1) - \nu(X_0)] = \tg_1(\mu) - \tg_0(\nu). \eee As the left side of \eqref{equ:dualderive} only depends on $(X_0,X_1)$ and the right side only on $(\mu,\nu)$, one can show that $(X_0,X_1)$ is $c$-optimal and $(\mu,\nu)$ solves \eqref{equ:dmk} iff $\mu(X_0) + \nu(X_1) = c(X_1-X_0)$ holds with probability one, which provides a basic optimality criterion. % Many existing OT algorithms are developed by exploiting the primal dual relation of \eqref{equ:mk} and \eqref{equ:dmk} (see e.g., \cite{korotin2022neural}), but have the drawback of yielding minimax problems that are challenging to solve in practice. % If $c$ is strictly convex, the % optimal transport map of \eqref{equ:m} is unique (almost surely) and yields a form of \bb % T(x)=x + \dd c^*(\dd \nu(x)), ~~~~ \ee where $c^*$ is the convex conjugate function of $c$, and $\nu$ is an optimal solution of \eqref{equ:dmk}, which is $c$-convex in that $\nu(x) = \sup_{y}\left \{ -c(y-x) + \mu(y)\right \}$ with $\mu$ the associated solution. In the canonical case of quadratic cost $c(x) = \frac{1}{2}\norm{x}^2$, we can write $T(x) = \dd \phi(x)$, where $\phi(x) \defeq \frac{1}{2}\norm{x}^2 + \nu(x)$ is a convex function. \paragraph{Dynamic formulations} Both the MK and Monge problems can be equivalently framed in dynamic ways as finding continuous-time processes that transfer $\tg_0$ to $\tg_1$. Let $\{x_t \colon t\in[0,1]\}$ be a smooth path connecting $x_0$ and $x_1$, whose time derivative is denoted as $\dot x_t$. For convex $c$, by Jensen's inequality, we can represent the cost $c(x_1-x_0)$ in an integral form: $$ c({x_1 - x_0}) = c\left (\int_0^1 \dot x_t \dt \right) = \inf_{x} \int_0^1 c({\dot x_t}) \dt, $$ where the infimum is attained when $x_t$ is the linear interpolation (geodesic) path: $x_t = t x_1 + (1-t)x_0$. Hence, the MK optimal transport problem \eqref{equ:mk} is equivalent to \bbb \label{equ:qt} \inf_{\vv X} \E\left [ \int_0^1 c(\dot X_t) \dt \right ] ~~~~~s.t.~~~~ \law(X_0) = \tg_0, ~~\law(X_1) = \tg_1, \eee where we optimize in the set of time-differentiable % stochastic processes $\vv X = \{X_t \colon t\in[0,1]\}$. The optimum of \eqref{equ:qt} is attained by $X_t = t X_1 + (1-t)X_0$ when $(X_0,X_1)$ is a $c$-optimal coupling of \eqref{equ:mk}, which is known as the \emph{displacement interpolation} \citep{mccann1997convexity}. We call the objective function in \eqref{equ:qt} the path-wise $c$-transport cost. The Monge problem can also be framed in a dynamic way. Assume the transport map $T$ can be induced by an ODE model $\d X_t = v_t(X_t)\dt $ such that $X_1 = T(X_0)$. Then the Monge problem is equivalent to % \bbb \label{equ:cm} \inf_{v,\traj X} \E\left [ \int _0^1 c({v_t(X_t)}) \dt \right] ~~~~~s.t.~~~~~ \d X_t = v_t(X_t)\dt ,~~~~~ \law(X_0) = \tg_0, ~~~~~ \law(X_1) = \tg_1, \eee which is equivalent to restricting $\vv X$ in \eqref{equ:qt} to the set of processes that can be induced by ODEs. Assume that $X_t$ following $\d X_t = v_t(X_t)\dt $ yields a density function $\varrho_t$. Then it is well known that $\varrho_t$ satisfies the continuity equation: % $$ \dot \varrho_t + \div(v_t\varrho_t) = 0. $$ Hence, we can rewrite \eqref{equ:cm} into an optimization problem on $(v, \varrho)$, yielding the celebrated \emph{{\bbformula} formula} \cite{benamou2000computational}: \bbb \label{equ:bb} \inf_{v, \varrho} \int_0^1 \int c(v_t(x)) \varrho_t(x)\d x \dt ~~~~s.t.~~~~ \dot \varrho_t + \div(v_t \varrho_t) = 0,~~~~\rho_0 = \d\tg_0/\dx, ~~~~\rho_1 = \d \tg_1/\dx, \eee where $\d \pi_i/\d x$ denotes the density function of $\tg_i$. The key idea of \eqref{equ:cm} and \eqref{equ:bb} is to restrict the optimization of \eqref{equ:qt} to the set of deterministic processed induced by ODEs, which significantly reduces the search space. Intuitively, Jensen's inequality $\E[c(Z)]\geq c(\E[Z])$ shows that we should be able to reduce the expected cost of a stochastic process by ``marginalizing'' out the randomness. In fact, we will show that, for a differentiable stochastic process $\vv X$, % its ($c$-)rectified flow yields no larger path-wise $c$-transport cost in \eqref{equ:qt} than $\vv X$ (see Lemma~\ref{thm:rectdual} and Theorem~\ref{thm:optmultimarg}). However, all the dynamic formulations above are still highly challenging to solve in practice. We will show that $c$-rectified flow can be viewed as a special coordinate descent like approach to solving \eqref{equ:qt} (Section~\ref{sec:crectifyOptView}). % \section{Introduction} The Monge–Kantorovich (MK) optimal transport (OT) problem concerns finding an optimal coupling between two distributions $\tg_0,\tg_1$: % \bbb \label{equ:mk} \inf_{(X_0,X_1)} % \E[c(X_1 - X_0)],~~~~s.t.~~~~ \law(X_0)=\tg_0,~~ \law(X_1) = \tg_1, \eee where we seek to find (the law of) an optimal coupling $(X_0,X_1)$ of $\tg_0$ and $\tg_1$, for which marginal laws of $X_0,X_1$ equal $\tg_0,\tg_1$, respectively, to minimize $\E[c(X_1-X_0)]$, called the $c$-transport cost, for a cost function $c$. Theories, algorithms, and applications of optimal transport have attracted a vast literature; see, for example, the monographs of \cite{villani2021topics, villani2009optimal, ambrosio2021lectures, santambrogio2015optimal, peyre2019computational} for overviews. % Notably, OT has been growing into a popular and powerful technique in machine learning, for key tasks such as learning generative models, transfer learning, and approximate inference \citep[e.g.,][]{peyre2019computational, arjovsky2017wasserstein, solomon2014wasserstein, el2012bayesian, courty2014domain, marzouk2016introduction}. The OT problem should be treated differently depending on whether $\tg_0,\tg_1$ are discrete or continuous measures. In this work, we focus on the continuous case when $\tg_0,\tg_1$ are high dimensional absolutely continuous measures on $\RR^d$ that are observed through empirical observations, a setting called data-driven OT in \cite{trigila2016data}. A well known result in OT \citep[e.g.,][]{villani2009optimal} shows that, if $\tg_0$ is continuous, the optimization in \eqref{equ:mk} can be restricted to the set of deterministic couplings satisfying $X_1 = T(X_0)$ for some continuous transport mapping $T\colon \RR^d\to\RR^d$, which is often approximated in practice with deep neural networks % \citep[e.g.,][]{makkuva2020optimal, korotin2021neural, korotin2022neural, huang2020convex}. However, continuous OT remains highly challenging computationally. One major difficulty is to handle the coupling constraints of $\law(X_0) = \tg_0$ and $\law(X_1)=\tg_1$, which are infinite dimensional when $\tgg$ are continuous. As a result, \eqref{equ:mk} can not be solved as a ``clean" unconstrained optimization problem. There are essentially two types of approaches to solving \eqref{equ:mk} in the literature. One uses Lagrange duality to turn \eqref{equ:mk} into a certain minimax game, and the other one approximates the constraint with an integral (often entropic-like) penalty function. However, the minimax approaches suffer from convergence and instability issues and are difficult to solve in practice, while the regularization approach can not effectively enforce the infinite-dimensional coupling constraints. \paragraph{This work} We present a different approach to continuous OT that re-frames \eqref{equ:mk} into a sequence of simple unconstrained nonlinear least squares optimization problems, which monotonically reduce the transport cost of a coupling while automatically preserving the marginal constraints. Different from the minimax and regularization approaches that enforce the constraints from outside, our method is an \emph{interior} approach which starts from a valid coupling (typically the naive independent coupling), and traverses inside the constraint set to decrease the transport cost. Such an interior approach is non-trivial and has not been realized before, because there exists no obvious unconstrained parameterization of the set of couplings of $\tgg$. % Our method is made possible by leveraging \emph{rectified flow} \cite{rectified}, a recent approach to constructing (non-optimal) transport maps for generative modeling and domain transfer. What makes rectified flow special is that it provides a simple procedure that turns a given coupling into a new one that obeys the same marginal laws, while yielding no worse transport cost w.r.t. \emph{all} convex functions $c$ simultaneously. Despite this attractive property, as pointed out in \cite{rectified}, rectified flow can not be used to optimize any fixed cost $c$, as it is essentially a special \emph{multi-objective} optimization procedure that targets no specific cost. Our method is a variant of rectified flow that targets a user-specified cost function $c$ and hence yields a new approach to the OT problem \eqref{equ:mk}. \paragraph{Rectified flow} We provide a high-level overview of the rectified flow of \cite{rectified} and the main results of this work. For a given coupling $(X_0,X_1)$ of $\tgg$, the \emph{rectified flow} induced by $(X_0,X_1)$ is the time-differentiable process $\vv Z = \{Z_t \colon t\in[0,1]\}$ over an artificial notion of time $t\in[0,1]$, that solves the following ordinary differential equation (ODE): \bbb \label{equ:ztflow_0} \d \Z_t = v^X_t(\Z_t) \dt,~~~~t\in[0,1],~~~~\text{starting from~~~~} \Z_0 = X_0, && % \eee where $v^X\colon \RR^d \times [0,1]\to \RR^d$ is a time-dependent velocity field defined as the solution of \bbb \label{equ:opt0} \inf_v \int_0^1 \e{ \norm{X_1-X_0 - v(X_t, t)}^2 } \dt, && X_t = t X_1 + (1-t)X_0, \eee and $X_t$ is the linear interpolation between $X_0$ and $X_1$. Eq~\eqref{equ:opt0} is a least squares regression problem of predicting the line direction of $(X_1-X_0)$ from every space-time point $(X_t, t)$ on the linear interpolation path, yielding a solution of % $$ v^X_t(z) = \E\left [X_1-X_0~|~ X_t=z \right ], $$ which is the average of direction $(X_1-X_0)$ for all lines that pass point $X_t = z$ at time $t$. The (conditional) expectations $\E[\cdot]$ above are w.r.t. the randomness of $(X_0,X_1)$. We assume that the solution of \eqref{equ:ztflow_0} exists and is unique, and hence $v^X_t(z)$ is assumed to exist at least on the trajectories of the ODE. The start-end pair $(Z_0,Z_1)$ induced by $\vv Z$ is called the \emph{rectified coupling} of $(X_0,X_1)$, and we denote it by $(Z_0,Z_1)=\mixupmap((X_0,X_1))$. % In practice, the expectation $\E[\cdot]$ is approximated by empirical observations of $(X_0, X_1)$, and $v$ is approximated by a parametric family, such as deep neural networks. In this case, the optimization in Eq~\eqref{equ:opt0} can be solved conveniently with off-the-shelf stochastic optimizers such as stochastic gradient descent (SGD), without resorting to minimax algorithms or expensive inner loops. % This makes rectified flow attractive for deep learning applications as these considered in \cite{rectified}. The importance of $(\Z_0,\Z_1) = \map((X_0,X_1))$ is justified by two key properties: 1) \emph{$(\Z_0,\Z_1)$ shares the same marginal laws as $(X_0,X_1)$ and is hence a valid coupling of $\tgg$; } 2) \emph{$(\Z_0,\Z_1)$ yields no larger convex transport costs than $(X_0,X_1)$, that is, $\E[c(\Z_1-\Z_0)] \leq \E[c(X_1-X_0)]$, for \emph{every} convex function $c\colon\RR^d\to \RR$.} Hence, it is natural to recursively apply the $\map$ mapping, that is, $(Z_0^{k+1}, Z_1^{k+1}) = \map((Z^k_0,Z_1^k))$ starting from $ (Z_0^0, Z_1^0) = (X_0, X_1)$, yielding a sequence of couplings that is monotonically non-increasing in terms of all convex transport costs. % The initialization can be taken to be the independent coupling $(Z_0^0,Z_1^0) \sim \tg_0\times \tg_1$, or any other couplings that can be constructed from marginal (unpaired) observations of $\tgg$. In practice, each step of $\map$ is empirically approximated by first drawing samples of $(Z_0^k, Z_1^k)$ from the ODE with drift $v^k$, and then constructing the next flow $v^{k+1}$ from the optimization in \eqref{equ:opt0}. Although this process accumulates errors, it was shown that one or two iterations are sufficient for practical applications \citep{rectified}. Note that the $\map$ procedure is ``cost-agnostic" in that it does not dependent on any specific cost $c$. Although the recursive $\map$ update is monotonically non-increasing on the transport cost for all convex $c$, it does not necessarily converge to the optimal coupling for any pre-specified $c$, as the update would stop whenever two cost functions are conflicting with each other. In \cite{rectified}, a coupling $(X_0,X_1)$ is called \emph{straight} if it is a fixed point of $\map$, that is, $(X_0,X_1) = \map((X_0,X_1))$. It was shown that rectifiable couplings that are optimal w.r.t. a convex $c$ must be straight, but the opposite is not true in general. One exception is the one dimension case ($d=1$), for which all convex functions $c$ (whose $c$-optimal coupling exists) share a common optimal coupling that is also straight. But this does not hold when $d\geq 2$. % \paragraph{$c$-Rectified flow} In this work, we modify the $\map$ procedure so that it can be used to solve \eqref{equ:mk} given a user-specified cost function $c$. We show that this can be done easily by properly restricting the optimization domain of $v$ and modifying the loss function in \eqref{equ:opt0}. The case of quadratic loss $c(x) = \frac{1}{2}\norm{x}^2$ is particularly simple, for which we simply need to restrict the $v$ to be a gradient field $v_t = \dd f_t$ in the optimization of \eqref{equ:opt0}. For more general convex $c$, we need to restrict $v$ to have a form of $v_t(x) = \dd c^*(\dd f_t(x))$, with $f$ minimizing the following loss function: \bbb\label{equ:minvc} \inf_{f} \int_0^1 \e{ c^*(\dd f( X_t)) - (X_1-X_0) \tt \dd f(X_t) + c(X_1-X_0) }\dt, % \eee where $c^*$ denotes the conjugate function of $c$. Obviously when $c(x)=\frac{1}{2}\norm{x}^2$, \eqref{equ:minvc} reduces to \eqref{equ:opt0} with $v = \dd f$. The loss function in \eqref{equ:minvc} is closely related to \emph{Bregman divergence} \citep[e.g.,][]{banerjee2005clustering} and the so-called \emph{matching loss} \citep[e.g.,][]{auer1995exponentially}. We call $\vv Z =\{Z_t\colon t\in[0,1]\}$ that follows $\d Z_t = \dd c^*(\dd f_t(Z_t)) \dt $ with $Z_0 = X_0$ and $f$ solving \eqref{equ:minvc} the $c$-rectified flow of $(X_0,X_1)$, and the corresponding $(Z_0,Z_1)$ the $c$-rectified coupling of $(X_0,X_1)$, denoted as $(Z_0,Z_1) = \crectify((X_0,X_1))$. Similar to the original rectified coupling, the $c$-rectified coupling $(Z_0,Z_1)$ also share the same marginal laws as $(X_0,X_1)$ and hence is a coupling of $\tgg$. In addition, $(Z_0,Z_1)$ yields no larger transport cost than $(X_0,X_1)$ w.r.t. $c$, that is, $\E[c(Z_1-Z_0)] \leq \E[c(X_1-X_0)]$. But this only holds for the specific $c$ that is used to define the flow, rather than all convex functions like $\rectify$. More importantly, recursively performing $\crectify$ allows us to find $c$-optimal couplings that solve the OT problem \eqref{equ:mk}. Under mild conditions, we have \bb (X_0,X_1) = \crectify((X_0,X_1)) &&\iff && \text{$(X_0,X_1)$ is $c$-optimal in \eqref{equ:mk}} &&\iff && \ell^*_{X,c} = 0, \ee where $\ell^*_{X,c}$ denotes the minimum value of the loss function in \eqref{equ:minvc}, which provides a criterion of $c$-optimality of a given coupling without solving the OT problem. Moreover, when following the recursive update $(Z_0^{k+1},Z_1^{k+1}) = \crectify((Z_0^{k}, Z_1^{k}))$, the $\ell^*_{Z^k,c} $ is guaranteed to decay to zero with $\min_{k\leq K} \ell^*_{Z^k,c} = \bigO{1/K}$. \input{texfiles/OT_background} \section{Rectified Flow: An Optimization-Based View} \label{sec:rectopt} We introduce rectified flow of \cite{rectified} from an optimization-based perspective: we show that rectified flow can be viewed as the solution of a special constrained dynamic optimization problem, which allows us to gain more understanding of rectified flow and motivates the development of $c$-rectified flow. Following \cite{rectified}, for a time-differentiable stochastic process $\vv X = \{X_t \colon t\in[0,1]\}$, % its expected velocity field $v^\X$ is defined as \bbb \label{equ:gvxzte} v^{\vv X}_t(z) = \E[\dot X_t ~|~X_t =z]. \eee where $\dot X_t$ denotes the time derivative of $X_t$. Obviously, $v^{\vv X}$ is the solution of \bbb \label{equ:infvLx} \inf_{v}\left\{ L_{\vv X}(v) \defeq \int_0^1 \e{\norm{\dot X_t - v_t(X_t)}^2}\dt \right \}, \eee where the optimization is on the set of all measurable velocity fields $v\colon \RR^d \to \RR^d$. The importance of $v^{\vv X}$ lies on the fact that it characterizes the time-evolution of the marginal laws $\rho_t \defeq \law(X_t)$ of $\vv X$, through the continuity equation in the distributional sense: \bbb \label{equ:conddeq} \partial_t \rho_t + \div (v^{\vv X}_t \rho_t)=0,~~~~ \rho_0 = \law(X_0),~~~~ t\in[0,1]. \eee Precisely, Equation~\eqref{equ:conddeq} should be interpreted by its weak and integral form: \bbb \label{equ:weakconddeq} \rho_t(h) - \rho_0(h) - \int_0^t \rho_t(\dd h \tt v^\X_s) \d s =0, ~~~~ \rho_0=\law(X_0), ~~~~{\forall h \in \Cc},~~~~ t\in[0,1], \eee where {$\rho_t(h)\defeq \int h(x) \d \rho_t(x)$} and $\Cc$ denotes the set of continuously differentiable functions on $\RR^d$ with compact support. Hence, if the solution of Eq~\eqref{equ:conddeq}-\eqref{equ:weakconddeq} is unique, then the marginal laws $\{\law(X_t)\}_t$ of $\vv X$ are uniquely determined by $v^\X$ and the initial $\law(X_0)$. We define the rectified flow of $\vv X$, denoted by $\vv Z = \rectflow(\vv X)$, as the ODE driven by $v^{\vv X}$: \bbb \label{equ:zofvx} \d Z_t = v_t^{\vv X}(Z_t) \dt,~~~~ Z_0 = X_0, ~~~~ t\in[0,1]. \eee Moreover, the rectified flow of a coupling $(X_0,X_1)$ is defined as the rectified flow of $\vv X$ when $\vv X$ is the linear interpolation of $(X_0,X_1)$. \begin{mydef} A stochastic process $\vv X$ is called rectifiable if $v^\vv X$ exists and is locally bounded, and Equation~\eqref{equ:zofvx} has an unique solution. A coupling $(X_0,X_1)$ is called rectifiable if its linear interpolation process $\vv X$, following $X_t = t X_1 +(1-t) X_0 $, is rectifiable. In this case, we call $\vv Z= \rectflow(\vv X)$ the rectified flow of $(X_0,X_1)$, and write it (with an abuse of notation) as $\vv Z = \rectflow((X_0,X_1))$. The corresponding $(Z_0,Z_1)$ is called the rectified coupling of $(X_0,X_1)$, denoted as $(Z_0,Z_1) = \map((X_0,X_1))$. \end{mydef} By the definition in \eqref{equ:zofvx}, we have $v^{\vv Z} =v^\X$, and hence the marginal laws $\law(Z_t)$ of $\vv Z$ are governed by the same continuity equation \eqref{equ:conddeq}-\eqref{equ:weakconddeq}, which is a well known fact. As shown in \citep{kurtz2011equivalence}, Equation~\eqref{equ:zofvx} has an unique solution iff Equation~\eqref{equ:weakconddeq} has an unique solution, which implies that $\vv Z$ and $\vv X$ share the same marginal laws. We also assumed that the solution of \eqref{equ:infvLx} is unique; if not, results in the paper hold for all solutions of \eqref{equ:infvLx}. \begin{thm}[Theorem~\nn{3.3} of \cite{rectified}] Assume that $\vv X$ is rectifiable. We have% \bb \vv Z = \rectflow(\vv X) &&\Rightarrow&& v^{\vv X} = v^{\vv Z} &&\Rightarrow&& \law(X_t) = \law(Z_t), ~~\forall t\in[0,1]. \ee \end{thm} Hence, rectified flow turns a rectifiable % stochastic process into a flow while preserving the marginal laws. \paragraph{A {optimization} view of rectified flow} We show that the rectified flow $\vv Z$ of $\vv X$ achieves the minimum of the path-wise $c$-transport cost % in the set of time-differentiable stochastic processes whose expected velocity field equals $v^\X$. This explains that the property of non-increasing convex transport costs of rectified flow/coupling. \begin{lem} \label{thm:rectdual} The rectified flow $\vv Z = \rectflow(\vv X_t)$ in \eqref{equ:zofvx} attains the minimum of \bbb \label{equ:pathopt} \inf_{\vv Y} \left\{ F_c(\vv Y) \defeq \int_0^1 \e{ c(\dot Y_t) } \dt , ~~~~s.t.~~~~ v^{\vv Y} = v^\X \right\}, \eee which holds for \emph{any} convex functions $c\colon \RR^d\to \RR$. \end{lem} \begin{proof} For any stochastic process $\vv Y$ with $v^{\vv X}_t(z) = v^{\vv Y}_t(z) = \E[\dot Y_t | Y_t=z] $, we have \bb F_c(\vv Y) & = \int_0^1\E[c(\dot Y_t)] \dt \\ & \geq\int_0^1 \E[c(\E[\dot Y_t|Y_t])] \dt \ant{Jensen's inequality} \\ & = \int_0^1 \E[c(v^{\vv Y}(Y_t))] \dt \\ & = \int_0^1 \E[c(v^{\vv X}(X_t))] \dt \ant{$v^\X=v^\Y$, and hence $\law(X_t)=\law(Y_t)$} \\ & = \int_0^1 \E[c(v^{\vv X}(Z_t))] \dt \ant{$\law(X_t) = \law(Z_t)$} \\ & = \int_0^1 \E[c(\dot Z_t)\dt = \vv F_c(\vv Z). \ee \end{proof} Lemma~\ref{thm:rectdual} suggests that the rectified flow decreases the path-wise $c$-transport cost: $F_c(\vv Z) \leq F_c(\vv X)$, for all convex $c$. Note that $\e{c(Z_1-Z_0)}\leq F_c(\vv Z) $ by Jensen's inequality, and $\e{c(X_1-X_0)} = F_c(\vv X)$ if $\vv X$ is the linear interpolation of $(X_0,X_1)$. Hence, in this case, we have % $$ \E[c(Z_1-Z_0)] \leq F_c(\vv Z) \leq F_c(\vv X) = \E[c(X_1-X_0)], $$ which yields a proof of Theorem~3.2 of \cite{rectified} that the rectified coupling $(Z_0,Z_1)$ yields no larger convex transport costs than $(X_0,X_1)$. % \paragraph{A primal-dual relation} Let us generalize the least squares loss $L_\X(v)$ in \eqref{equ:infvLx} to a a Bregman divergence based loss: \bb \tilde L_{\X,c}(v) \defeq \int_0^1 \e{\bcb{\dot X_t; ~ v_t(X_t)}}\dt, && \bc(x;y) = c(x) - c(y) - (x-y)\tt \dd c(y), \ee where $\bc(\cdot;\cdot)$ is the Bregman divergence w.r.t. $c$. The least squares loss $L_\X $ is recovered with $c(x) = \frac{1}{2} \norm{x}^2$. Rectified flow can be alternatively implemented by minimizing $\tilde L_{\X,c}$ with a differentiable strictly convex $c$, as in this case the minimum of $\tilde L_{\X,c}$ is also attended by $v^\X(z) = \E[\dot X_t|X_t=z]$. The $c$-rectified flow is obtained if we minimize $\tilde L_{\X,c}$ with $v$ restricted to be a form of $v = \dd c^*\circ \dd f_t$. See more in Section~\ref{sec:crectify}. In the following, we show that the optimization in \eqref{equ:pathopt} can be viewed as the dual problem \eqref{equ:gvxzte}. \begin{thm} For % any differentiable convex function $c$, and {rectifiable process $\vv X$}, we have $$ \tilde \ell^*_{\vv \X,c} \defeq \inf_{v} \tilde L_{\vv X,c} (v) = \sup_{\vv Y}\left\{ F_c(\vv X) - F_c(\vv Y) ~~ s.t.~~ v^\Y=v^\X \right \}, % $$ and the optima above are achieved when $v = v^{\X}$ and $\vv Y = \rectflow(\vv X)$. \end{thm} \begin{proof} Let $\var_c(\dot X_t~|~X_t )\defeq \E[c(\dot X_t) - c(\E [\dot X_t|X_t])~|~X_t] $. For any $v$, we have \bb \tilde L_{\X,c}(v) &= \int_0^1 \E[c(\dot X_t) - c(v(X_t)) - (\dot X_t - v(X_t)) \dd c(v(X_t))] \dt \\ & = \int_0^1 \E[c(\dot X_t) - c(v(X_t)) - (v^\X(X_t) - v(X_t)) \dd c(v(X_t)) ] \dt \ant{$v^\X(X_t)=\E[\dot X_t |X_t]$} \\ & \geq \int_0^1 \E[c(\dot X_t) - c(v^\X(X_t)) ] \dt \ant{$c(v^\X) \geq c(v) + (v^\X-v) \dd c(v)$} \\ & = \int_0^1 \var_c(\dot X_t~|~X_t) \dt , \ee The inequality is tight when $v = v^\X$, which attains the minimum of $\tilde L_{\X,c}$. Write $ R_{\vv X,c}(\vv Y) = F_c(\vv X) - F_c(\vv Y)$. We know that $\vv Z = \map(\vv X)$ attains the maximum of $R_{\vv X,c}(\vv Y)$ subject to $v^{\vv Y} = v^{\vv X}$ by Lemma~\ref{thm:rectdual}. In addition, \bb R_{\vv X, c}(\vv Z) & = \int_0^1 \E[c(\dot X_t) - c(\dot Z_t) ] \dt \\ & = \int_0^1 \E[c(\dot X_t) - c(v^\X_t(Z_t)])] \dt \\ & = \int_0^1 \E[c(\dot X_t) - c(v^\X_t(X_t)])] \dt \ant{$\law(Z_t)=\law(X_t),\forall t $}\\ & = \int_0^1 \E[c(\dot X_t) - c(\E[\dot X_t|X_t])] \dt \\ % & = \int_0^1 \E[\var_c(\dot X_t~|~X_t)] \dt. % \ee This concludes the proof. % \end{proof} \paragraph{Straight couplings} The $\tilde \ell^*_{\vv X,c}=\int_0^1\var_c(\dot X_t|X_t)\dt $ above provides a measure of how much the different paths of $\X $ intersect with each other. If $c$ is strictly convex and $\tilde \ell^*_{\vv X,c} = 0$, we have $\dot X_t = \E[\dot X_t|X_t]$ almost surely, meaning that there exist no two paths that go across a point along two different directions. In this case, $\vv X$ is a fixed point of $\rectflow(\cdot)$, that is, $\vv X = \vv Z = \map(\vv X)$, because we have $\d X_t =\dot X_t \d t = \E[\dot X_t |X_t] \dt = v^\X(X_t ) \dt $, which is the same Equation \eqref{equ:zofvx} that defines $\vv Z$. Similarly, if $\vv X $ is the linear interpolation of the coupling $(X_0,X_1)$, then $\tilde \ell^*_{\X,c} =0$ with strictly convex $c$ if and only if $(X_0,X_1)$ is a fixed point of the $\map$ mapping, that is, $(X_0,X_1) = \map((X_0,X_1))$, following \cite{rectified}. Such couplings are called \emph{straight}, or \emph{fully rectified} in \cite{rectified}. Obtaining straight couplings is useful for learning fast ODE models because the trajectories of the associated rectified flow $\vv Z$ are straight lines and hence can be calculated in closed form without iterative numerical solvers. See \cite{rectified} for more discussion. % Moreover, \cite{rectified} showed that rectifiable $c$-optimal couplings must be straight. In the one dimensional case ($d=1$), the straight coupling, if it exists, is unique and attains the minimum of $\E[c(X_1-X_0)]$ for all convex functions for which $c$-optimal coupling exists. For higher dimensions ($d\geq 2$), however, straight couplings are not unique, and the specific straight coupling obtained at the convergence of the recursive $\map$ update (i.e. $(Z_0^{k+1},Z_1^{k+1})=\map((Z_0^{k},Z_1^{k}))$) is implicitly determined by the initial coupling $(Z_0^0,Z_1^0)$, and is not expected to be optimal w.r.t. any pre-fixed $c$. The following counter example shows a somewhat stronger negative result: there exist straight couplings that are not optimal w.r.t. all second order differentiable convex functions with invertible Hessian matrices. \begin{theoremEnd}[proof at the end, no link to proof]{exa} \label{exa:toycounterexp} Take $\tg_0 = \tg_1 = \normal(0,I)$. Hence, for $c(x)=\norm{x}^p$ with $p>0$, the $c$-optimal mapping is the trivial identity coupling $(X_0,X_0)$ with $X_0\sim \tg_0$. However, consider the coupling $(X_0,AX_0)$, where $A$ is a non-identity and non-reflecting rotation matrix (namely $A\tt A=I$, $\det(A)=1$, $A \neq I$ and $A$ does not have $\lambda =-1$ as an eigenvalue). Then $ (X_0, A X_0)$ is a straight coupling of $\tg_0$ and $\tg_1$, but it is not $c$-optimal for all second order differentiable convex function $c$ whose Hessian matrix is invertible everywhere. {See Appendix for the proof.} It is the rotation transform that makes $(X_0,AX_0)$ sub-optimal, which is removed in the proposed $c$-rectified flow in Section~\ref{sec:crectify} via a Helmholtz like decomposition. \end{theoremEnd} \begin{proofEnd} i) The fact that $A\tt A =I$ and $\tg_0=\tg_1=\normal(0,I)$ ensures that $AX_0\sim \tg_1$ and hence $(X_0, A X_0)$ is a coupling of $\tg_0$ and $\tg_1$. Let $X_t = t A X_0 + (1-t) X_0$ be the linear interpolation of the coupling. Related, we have $\dot X_t = A X_0 - X_0 $. Canceling $X_0$ yields that \bbb \label{dotxAIcexample} \dot X_t = (A-I) (t A + (1-t)I)^{-1}X_t, \eee where we use the fact that $t A + (1-t)I$ is convertible for $t\in[0,1]$, which we prove as follows: if $t A + (1-t)I$ is not invertible, then $A$ must have $\lambda =-\frac{1-t}{t}$ as one of its eigenvalues; but as a rotation matrix, all eigenvalues of $A$ must have a norm of $1$, which means that we must have $t=0.5$ and $\lambda = -1$. This, however, is excluded by the assumption that $A$ is non-reflecting (and hence $\lambda \neq -1$). Equation~\eqref{dotxAIcexample} shows that $\dot X_t$ is uniquely determined by $X_t$. Hence, we have $\int_0^1 \e{\var(\dot X_t|X_t) } \dt = 0$, which implies that $(X_0,AX_0)$ is a straight coupling by Theorem~\nn{3.6} of \cite{rectified}. 2) Let $c$ be a second order differentiable convex function whose Hessian matrix $\dd^2 c(x)$ is invertible everywhere. Let $c^*$ be the convex conjugate of $c$, then $c^*$ is also second order differentiable and $\dd c(\dd c^*(x)) = x$, and $\dd^2 c^*(x) = \dd^2 c(x)^{-1}.$ If $(X_0,A X_0)$ is a $c$-optimal coupling, there must exists a function $ \phi\colon \RR^d\to \RR$, such that \bbb \label{equ:axcstartphix} A x = x + \dd c^*(\dd \phi(x)),~~\forall x, \eee where $c^*$ is the convex conjugate of $c$. Equation~\eqref{equ:axcstartphix} is equivalent to $\dd c(Ax - x) = \dd \phi(x)$, which means that $\dd \phi $ is continuously differentiable. Taking gradient on both sides of \eqref{equ:axcstartphix} gives \bbb \label{equ:aihxbx} A -I = H_x B_x, &&& H_x = \dd^2 c^*(\dd\phi(x)),~~~~ B_x = \dd^2 \phi(x), \eee where $H_x, B_x$ are both symmetric and $H_x$ is positive definite, and hence Then $H_x B_x$ is a diagonalizable (all its eigenvalues are real) by Lemma~\ref{thm:diagmat}. However, $A-I$ is not diagonalizable because $A$ must have complex eigenvalues as a non-reflecting, non-identity rotation matrix. Hence, \eqref{equ:aihxbx} can not hold. \begin{lem}\label{thm:diagmat} Assume that $A,B$ are two real symmetric matrices and $A$ is positive definite. Then $AB$ is diagonalizable (on the real domain), that is, there exists an invertible matrix $P$, such that $P^{-1} AB P$ is a diagonal matrix. \end{lem} \begin{proof} This is a standard result in linear algebra. Because $A$ is positive definite, there exists an invertible symmetric matrix $C$, such that $CC = A.$ Then, $AB = CCB$, and it is similar to $CBC^{-1}$, which is symmetric and hence diagonalizable. \end{proof} \end{proofEnd} \section{Differentiable Processes with Equivalent Marginal Laws} \label{sec:marginal0} The marginal preserving property of rectified flow is due to the property of $v^{\vv Z} = v^\X$ by construction. However, we show in this section that $v^{\vv X} = v^{\vv Z}$ is only a sufficient condition: two differentiable processes $\vv X$ and $\vv Z$ can have the same marginal laws even if $r\defeq v^\X -v^{\vv\Z} \neq 0$. This is because $r $, as illustrated in Example~\ref{exa:toycounterexp}, can be a rotation-only vector field (in a generalized sense shown below) that introduces rotation components into the dynamics without modifying the marginal distributions. Therefore, the constraint of $v^{\vv Y}=v^\X$ in the optimization problem \eqref{equ:pathopt} may be too restrictive. A natural relaxation of \eqref{equ:pathopt} would be \bbb \label{equ:pathoptmlaw} \inf_{\vv Y} \left\{ F_c(\vv Y) \defeq \e{ \int_0^1 c(\dot Y_t) \dt} , ~~~~s.t.~~~~ \law(Y_t) = \law(X_t),~~\forall t\in[0,1]\right\}, \eee which yields a dynamic OT problem with % a continuum of marginal constraints. In Section~\ref{sec:crectify}, we show that the solution of \eqref{equ:pathoptmlaw} yields our $c$-rectified flow that solve the OT problem at the fixed point. Solving \eqref{equ:pathoptmlaw} allows us to remove the rotational components of $v^\X$, which is what what renders rectified flow non-optimal. In this section, we first characterize the necessary and sufficient condition for having equivalent marginal laws. \begin{mydef} A time-dependent vector field $r \colon \RR^d\times [0,1]\to \RR^d$ is said to be $\X$-marginal-preserving if \bbb \label{equ:edhxrt0} \int_0^t\E[\dd h(X_t) \tt r_t(X_t) ] =0,~~~\forall t\in[0,1], ~~~~ h \in \Cc. \eee \end{mydef} Equation~\eqref{equ:edhxrt0} is equivalent to saying that $\E[\dd h(X_t) \tt r_t(X_t) ] =0$ holds almost surely assuming that $t$ is a random variable following $\uniform([0,1])$ (i.e., $t$-almost surely). Let $\rho_t = \law(X_t)$ and it yields a density function $\varrho_t$. Using integration by parts, we have $$ 0= \E[\dd h(X_t) \tt r_t(X_t) ] = \int \dd h(x) \tt r_t(x)\varrho_t(x) \d x = - \int h(x) \div(r_t(x) \varrho_t(x)) \d x, ~~~~\forall h \in \Cc, $$ which gives $\div(r_t \varrho_t)=0$. This says that $r_t \varrho_t$ is a rotation-only (or divergence-free) vector field in the classical sense. \begin{lem}\label{thm:marginalrot} Let $\vv X$ and $\vv Y$ be two stochastic processes with the same initial distributions $\law(X_0) = \law(Y_0)$. Assume that $\vv X$ is rectifiable, and $v^{\vv Y}_t(z):=\E[\dot Y_t |Y_t = z]$ exists and is locally bounded. % Then $\X$ and $\Y$ share the same marginal laws at all time, that is, $\law(X_t) = \law(Y_t)$, $\forall t\in[0,1],$ if and only if $v^\X - v^\Y$ is $\vv Y$-marginal-preserving. % \end{lem} \begin{proof} Taking any $h$ in $\Cc$, % we have for $t\in[0,1]$ \bb \E[h(X_t)] - \E[h(X_0)] & = \int_0^t \E[\dd h(X_s)\tt \dot X_s ] \d s \\ & =\int_0^t \E[\dd h(X_s)\tt v^\X_s(X_s)] \d s \ant{$v^\X_s(X_s) = \E[\dot X_s |X_s]$} . \ee This suggests that the marginal law $\rho_t\defeq \law(X_t)$ satisfies \bbb \label{equ:tgth0} \rho_t(h) - \rho_ 0 (h) - \int_0^t \rho_s (\dd h \tt v^\X_s) \d s =0, ~~~ \forall h \in \Cc , \eee where we define $\rho_t (h) = \int h(x) \d \rho_t(x)$. Equation~\eqref{equ:tgth0} is formally written as the continuity equation: \bbb \label{equ:contieq0} \dot \rho_t + \div (g_t^\X \rho_t) = 0. \eee Similarly, $\tilde\rho_t\defeq \law(Y_t)$ satisfies \bbb \label{equ:tgtph0} \tilde\rho_t(h) - \tilde\rho_ 0 (h) - \int_0^t \tilde\rho_s (\dd h \tt v^\Y_s) \d s =0, ~~~ \forall h, \eee If $v_t^\X - v_t^\Y$ is $\law(Y_t)$-preserving for $\forall t\in[0,1]$, we have \bb \E[h(Y_t)] - \E[h(Y_0)] & =\int_0^t \E[\dd h(Y_s)\tt \dot Y_s] \d s \\ & = \int_0^t \E[\dd h(Y_s)\tt v^\Y_s(Y_s)] \d s \\ % & = \int_0^t \E[\dd h(Y_s)\tt v^\X_s(Y_s)] \d s + \int_0^t \E[\dd h(Y_s) \tt (v^\Y_s(Y_s) - v^\X_s(Y_s))] \d s \\ & =\int_0^t \E[\dd h(Y_s) \tt v^\X_s(Y_s)] \d s \ant{$v^\X-v^\Y$ is $\vv Y$-preserving}, \ee which suggests that $\tilde\rho_t \defeq \law(Y_t)$ solves the same continuity equation \eqref{equ:contieq0}, starting from the same initialization as $\law(X_0) = \law(Y_0)$. Hence, we have $\rho_t = \tilde\rho_t$ if the solution of \eqref{equ:contieq0} is unique, which is equivalent to the uniqueness of the solution of $\d Z_t = \vofX_t(Z_t)$ in \eqref{equ:zofvx} following Corollary 1.3 of \cite{kurtz2011equivalence}. On the other hand, if $\rho_t = \law(X_t) = \law (Y_t) = \tilde\rho_t$, following \eqref{equ:tgth0} and \eqref{equ:tgtph0}, we have for any $h \in \Cc$, \bb 0 & = \int_0^t \tilde\rho_t (\dd h\tt v^{\X}_s) - \tilde\rho_t(\dd h\tt v^{\Y}_s) \d s = \int_0^t \dd h(Y_s) \tt (v^\X(Y_s) -v^\Y(Y_s))\d s, \ee which is the definition of $\vv Y$-marginal-preserving following \eqref{equ:edhxrt0}. \end{proof} \section{$c$-Rectified Flow} % \label{sec:crectify} We introduce $c$-rectified flow, a $c$-dependent variant of rectified flow that guarantees to minimize the $c$-transport cost when applied recursively. This section is organized as follows: Section~\ref{sec:crectifyDefine} defines and discusses the $c$-rectified flow of a differentiable stochastic process $\vv X$, which we show yields the solution of the infinite-marginal OT problem~\eqref{equ:pathoptmlaw}. Section~\ref{sec:crectifyCoupling} considers the $c$-rectified flow of a coupling $(X_0,X_1)$, which we show is non-increasing on the $c$-transport cost. Section~\ref{sec:crectifyFixed} proves that the fixed points of $\crectify$ are $c$-optimal. Section~\ref{sec:crectifyOptView} interprets $c$-rectified flow as an alternating direction descent method for the dynamic OT problem \eqref{equ:qt}, and a majorize-minimization (MM) algorithm for the static OT problem \eqref{equ:mk}. Section~\ref{sec:hj} discusses a key lemma relating $c$-optimal couplings and its associated displacement interpolation with Hamilton-Jacobi equation. \subsection{$c$-Rectified Flow of Time-Differentiable Processes $\vv X$} \label{sec:crectifyDefine} For a {convex} cost function $c\colon \RR^d\to \RR$ and a time-differentiable process $\vv X$, the $c$-rectified flow of $\vv X$, denoted as $\vv Z = \crectflow(\vv X)$, is defined as the solution of \bbb \label{equ:zgxft} \d \Z_t =g^{\X,c}_t(\Z_t) \dt,~~~~ \Z_0 = X_0,~~~~\text{with}~~~~ g^{\X,c}_t(z) = \dd c^*(\dd f^{\X,c}_t(z)), ~~~~ t\in[0,1], \eee where $c^*(x) \defeq \sup_{y}\{ x\tt y - c(y)\}$ is the convex conjugate of $c$, and $f^{\X,c}\colon \RR^d\times [0,1] \to \RR$ is the optimal solution of \bbb \label{equ:bregloss} \inf_{f} \left\{ L_{\X,c} (f) \defeq \int_0^1 \E\left [ \mcb{\dot X_t;~ \dd f_t( X_t)} \right ] \dt \right\}, \eee where $\mc~ \colon \RR^d\times \RR^d\to [0,+\infty)$ is a loss function defined as % $$ \mcb{x;y} = c(x) - x\tt y + c^*(y). $$ Note that we have $\mcb{x;y}\geq 0$ for $\forall x,y$ following the definition of the conjugate $c^*$ (or the Fenchel-Young inequality). Losses of form $\mcb{x;y}$ {is equivalent to the so called \emph{matching loss} proposed for learning generalized linear models \cite{auer1995exponentially}.} % Compared with the original rectified flow, the difference of $c$-rectified flow is i) restricting the velocity field to a form of $g_t = \dd c^*\circ \dd f_t$, and ii) replacing the quadratic objective function to the matching loss. These two changes combined yield a Helmholtz like decomposition of $v^\X$ as we show below, allowing us to remove the ``{rotation-only}" component of $v^\X$ % and obtain $c$-optimal couplings at fixed points. \paragraph{Bregman divergence, Helmholtz decomposition, marginal preserving} We can equivalently write \eqref{equ:bregloss} using Bergman divergence associated with $c$, that is, % $$\bcb{x;y}\defeq c(x)-c(y)- \nabla c(y)\tt (x-y).$$ Then it is easy to see that $\mcb{x;y} = \bcb{x; \dd c^*(y)}$, by using the fact that $\dd c(\dd c^*(y)) = y$ and $c^*(y) = y \tt \dd c^*(y) - c(\dd c^*(y))$. Hence, $\mc$ and $\bc$ are equivalent up to the monotonic transform $\dd c^*$ on $y$. The minimum $\bcb{x;y}=0$ is achieved when $y = x$, while $\mcb{x;y}=0$ is achieved when $\dd c^*(y)=x$. Therefore, \eqref{equ:bregloss} is equivalent to \bbb \label{equ:bregloss2} \inf_{f} \int_0^1 \E\left [ \bcb{\dot X_t;~ g_t( X_t))} \right ] \dt , && \text{with~~~~} g_t = \dd c^*\circ \dd f_t. \eee Moreover, the generalized Pythagorean theorem of Bregman divergence (e.g., \citep{banerjee2005clustering}) gives \bbb \label{equ:pythabreg} \e{\bcb{\dot X_t; ~~ g_t}~|~ X_t} = \bcb{\e{\dot X_t|X_t};~~ g_t} + \e{\bcb{\dot X_t; ~~ \e{\dot X_t|X_t}}}. \eee Because $v^\X(X_t)= \e{\dot X_t|X_t}$ and the last term of \eqref{equ:pythabreg} is independent with $g_t$ , we can further reframe \eqref{equ:bregloss} into % \bbb \label{equ:bregvloss} \min_{f} \int_0^1 \E \left [ \bcb{v^\X_t( X_t); ~~ g_t(X_t))} \right ] \dt, && \text{with~~~~} g_t = \dd c^*\circ \dd f_t, \eee which can be viewed as projecting the expected velocity $v^\X_t$ to the set of functions of form $g_t = \dd c^*\circ \dd f_t$, w.r.t. the Bregman divergence. This yields an orthogonal decomposition of $v_t^\X$: \bbb \label{equ:helm} v^{\X}_t = \dd c^*\circ \dd f^{\X,c}_t + r^{\X,c}_t, \eee where $r^{\X,c}_t = v^{\X,c}_t - \dd c^*\circ \dd f^{\X,c}_t$ is the residual term. The key result below shows that $r^{\X,c}$ is $\X$-marginal-preserving, which ensures that the $c$-rectified flow preserves the marginals of $\X$. \begin{mydef} We say that $\vv X$ is $c$-rectifiable if $v^\X$ exists, the minimum of \eqref{equ:bregloss} exists and is attained by a locally bounded function $f^{\vv X, c}$, and % the solution of % Equation~\eqref{equ:zgxft} exists and is unique. % \end{mydef} \begin{thm}\label{thm:marginalgood} Assume that $\X$ is $c$-rectifiable, and $c^*\defeq \sup_y \{x\tt y - c(y)\}$ and $c^*\in\Cone$. We have i) $v^\X- g^{\X, c}$ is $\X$-marginal-preserving. ii) $\vv Z = \crectify(\vv X)$ preserves the marginal laws of $\vv X$, that is, $\law(Z_t) =\law(X_t)$, $\forall t\in[0,1]$. \end{thm} \begin{proof} i) By $v^\X_t(z) = \E[\dot X_t|X_t=z]$, the loss function in \eqref{equ:bregloss} is equivalent to % \bb L_{\X,c}(f) & = \int_0^1 \e{c^*(\dd f_t(X_t)) - \E[\dot X_t~|~X_t ] \tt \dd f_t(X_t ) + c(\dot X_t) } \dt \\ & = \int_0^1 \e{c^*(\dd f_t(X_t)) - v^\X_t(X_t) \tt \dd f_t(X_t) + c(\dot X_t)} \dt. \ee By Euler-Lagrange equation, we have $$ \int_0^1 \e{(\dd c^*(\dd f_s^{\X,s}(X_s) ) - v^\X(X_s))\tt \dd g_s(X_s) } \d s = 0, ~~~~ \forall g: ~g_s\in \Cc. $$ Taking $g_s = h$ if $s<t$ and $g_s = 0 $ if $s>t$ yields that $r^{\X,c}(x) =\dd c^*(\dd f_s^{\X,c}(X_s) ) - v^\X(X_s)$ is $\vv X$-marginal-preserving following \eqref{equ:edhxrt0}. ii) Note that $\vv Z$ is rectifiable if $\vv X$ is $c$-rectifiable. % Applying Lemma~\ref{thm:marginalrot} yields the result. % \end{proof} For the quadratic cost $c(x) = c^*(x) = \frac{1}{2} \norm{x}^2$, the $\dd c^*$ is the identity mapping, and \eqref{equ:helm} reduces to the {\helm} decomposition, which represents a velocity field into the sum of a gradient field and a divergence-free field. Hence, \eqref{equ:helm} yields a generalization of {\helm} decomposition, in which a monotonic transform $\dd c^*$ is applied on the gradient field component. We call \eqref{equ:helm} a \emph{Bregman {\helm} decomposition}. \paragraph{Remark: score matching} In some special cases, $v^\X$ may already be a gradient field, and hence the rectified flow and $c$-rectified flow coincide for $c(x) = \frac{1}{2} \norm{x}^2$. One example of this is when $X_t = \alpha_t X_1 + \beta_t \xi$ for some time-differentiable functions $\alpha_t$ and $\beta_t$, and $\xi \sim \normal(0,I)$, satisfying $\alpha_1 = 1, \beta_1 = 0$, and $X_0 = \alpha_0 X_1 + \beta_0 \xi$. In this case, one can show that \bb v_t^\X(z) = \E[\dot \alpha_t X_1+\dot \beta_t X_0~|~X_t=z] = \dd f_t(z), &&\text{with} && f_t(z) = \eta_t\log \varrho_t(z) + \frac{\zeta_t}{2} \norm{z}^2, \ee where $\varrho_t$ is the density function of $X_t$ with $\varrho_t(z) \propto \int \phi \left ( \frac{z -\alpha_t x_1}{\beta_t}\right ) \d \tg_1(x_1)$ and $\phi(z) = \exp(-\norm{z}^2/2)$, % and $\eta_t = \beta_t^2 (\dot \alpha_t/\alpha_t - \dot \beta_t /\beta_t)$ and $\zeta_t = \dot \alpha_t /\alpha_t$. This case covers the probability flow ODEs \citep{song2020score} and denoising diffusion implicit models (DDIM) \citep{song2020denoising} with different choices of $\alpha_t$ and $\beta_t$ % as suggested in \cite{rectified}. When $\zeta_t = 0$, as the case of \cite{song2019generative}, $v_t^\X$ is proportional to $\dd\log \rho_t,$ the \emph{score function} of $\varrho_t$, and the least squares loss $L_\X(v)$ in \eqref{equ:infvLx} reduces to a time-integrated \emph{score matching} loss \citep{hyvarinen2005estimation, vincent2011connection}. However, $v_t^\X$ is generally not a score function or gradient function, especially in complicate cases when the coupling $(X_0,X_1)$ is induced from the previous rectified flow as we iteratively apply the rectification procedure. In these cases, it is necessary to impose the gradient form as we do in $c$-rectified flow. % \paragraph{$c$-Rectified flow solves Problem~\eqref{equ:pathoptmlaw}} We are ready to show that the $c$-rectified flow solves the optimization problem in \eqref{equ:pathoptmlaw}. Further, \eqref{equ:bregloss} forms a dual problem of \eqref{equ:pathoptmlaw}. % \begin{thm}\label{thm:optmultimarg} Under the conditions in Theorem~\ref{thm:marginalgood}, we have i) $\vv Z = \crectify(\vv X)$ attains the minimum of \eqref{equ:pathoptmlaw}. ii) Problem \eqref{equ:pathoptmlaw} and \eqref{equ:bregloss} has a strong duality: $$ \inf_f L_{\X,c}(f) = \sup_{\vv Y} \left \{F_c(\vv X) - F_c(\vv Y) \colon \law(Y_t) =\law(X_t), ~\forall t\in[0,1] \right \}. $$ As the optima above are achieved by $f^{\X,c}$ and $\vv Z$, we have $L_{\X,c}(f^{\X,c}) = F_c(\vv X) - F_c(\vv Z).$ \end{thm} \begin{proof} Write $R_{\X,c}( \vv Y) = F_c(\vv X) - F_c(\vv Y)$. First, we show that $ L_{\X,c}(f) \geq R_{\X,c}(\Y)$ % for any $f$ and $\vv Y$ that satisfies $\law(Y_t) = \law(X_t)$, $\forall t$: \bb & R_{\X,c}(\vv Y) \\ & = \E\left [ \int_0^1 c(\dot X_t) - c(\dot Y_t) \dt \right] \\ & \overset{(1)}{\leq} \E\left [ \int_0^1 c(\dot X_t) + c^*(\dd f_t(Y_t)) - \dot Y_t \tt \dd f_t(Y_t) \dt \right] \ant{Fenchel-Young inequality: $c(y)\geq x\tt y - c^*(x)$}\\ & = \E\left [ \int_0^1 c(\dot X_t) + c^*(\dd f_t(Y_t)) - v^\Y_t(Y_t) \tt \dd f_t(Y_t) \dt \right] \!\!\!\!\!\!\! \ant{$v^\Y_t(Y_t) =\E[\dot Y_t|Y_t]$}\\ & = \E\left [ \int_0^1 c(\dot X_t) + c^*(\dd f_t(X_t)) - v^\Y_t(X_t) \tt \dd f_t(X_t) \dt \right] \!\!\!\!\!\!\! \ant{$\law(X_t) = \law(Y_t)$} \\ & = \E\left [ \int_0^1 c(\dot X_t) + c^*(\dd f_t(X_t)) - v^\X_t(X_t) \tt \dd f_t(X_t) \dt \right] \!\!\!\!\!\!\!\ant{$v^\X-v^\Y$ is $\vv X$-marginal-preserving }\\ % & = L_{\X,c}(f). \ee Moreover, if we take $\vv Y = \vv \Z$ and $f = f^{\X,c}$, then the inequality in $\overset{(1)}{\leq}$ is tight because $\dot Z_t = \dd c^*(\dd f_t(Y_t))$ holds $t$-almost surely. Therefore, $R_{\X,c}(\vv Z) = L_{\X,c}(f^{\X,c}) \geq R^{\X, c} (Y) $, which suggests that $\vv Z$ attains the maximum of $R_{\X,c}$ (under the marginal constraints) and the strong duality holds. \end{proof} \subsection{$c$-Rectified Flow of Coupling $(X_0,X_1)$} \label{sec:crectifyCoupling} Similar to the case of rectified flow, the $c$-rectified flow/coupling of a coupling $(X_0,X_1)$ is defined as the $c$-rectified flow/coupling of its linear interpolation process. In the following, we show that the $c$-rectified coupling of a coupling % yields no larger {$c$-transport cost}. % \begin{mydef}\label{def:lincp} Let $\vv X$ be the linear interpolation of coupling $(X_0,X_1)$ in that $ X_t = t X_1 + (1-t) X_0,\forall t\in[0,1]$. % We say that $(X_0,X_1)$ is $c$-rectifiable if $\vv{X}$ is $c$-rectifiable, and call $\vv Z = \crectflow(\vv{ X})$ the $c$-rectified flow of $(X_0,X_1)$. We call the induced $(Z_0,Z_1)$ the $c$-rectified coupling of $(X_0,X_1)$, denoted as $(Z_0,Z_1) = \crectify((X_0,X_1))$. \end{mydef} Note that the $c$-transport cost $\E[c(X_1-X_0)]$ is related to {the path-wise $c$-transport cost $F_c(\vv X)$} via \bb F_{c}(\vv X) = \E[c(X_1-X_0)] + S_c(\vv X), && S_c(\vv X) \defeq \int_0^1 \E[c(\dot X_t) - c(X_1-X_0)] \dt, \ee where $S_c(\vv X)$ is a non-negative measurement of how close $\vv X$ is to be {geodesic}: We have $S_c(\vv X) \geq 0$ following Jensen's inequality $\int_0^1 c(\dot X_t)\dt \geq c(\int_0^1 \dot X_t\dt ) = c(X_1-X_0)$, and $S_c(\vv X) = 0$ if $X_t = t X_1 + (1-t) X_0$. Hence, when $\vv X$ is the linear interpolation of $(X_0,X_1)$, we have from Theorem~\ref{thm:optmultimarg} that \bbb \label{equ:straighexz} % \E[c(X_1-X_0)] - \E[c(Z_1-Z_0)] = S_c(\vv Z) + L_{\X,c}(f^{\X,c}) \geq 0. \eee which establishes that $(Z_0,Z_1)$ yields no larger transport cost than $(X_0,X_1)$. \begin{thm}\label{thm:cost0} Assume that $c$ is convex with conjugate $c^* \in \Cone$, % and the conditions in Definition~\ref{def:lincp} holds. Then Equation~\eqref{equ:straighexz} holds and $\E[c(Z_1-Z_0)] \leq \E[c(X_1-X_0)].$ \end{thm} Compared with the regular $\map$ mapping, the key difference here is that the monotonicity of $\crectify$ only holds for the specific $c$ that it employees, rather than all convex cost functions. More importantly, as we show below, recursively applying $\crectify$ yields optimal couplings w.r.t. $c$, a key property that the regular rectified flow misses. \subsection{Fixed Points of $c$-{$\map$} are $c$-Optimal} \label{sec:crectifyFixed} We show three key results regarding the optimality of fixed points of the $\crectify$ mapping: 1) A coupling $(X_0,X_1)$ is a fixed point of $\crectify$, that is, $(X_0,X_1) = \crectify((X_0,X_1))$, if and only if it is $c$-optimal; 2) Define $\ell^*_{X,c} = \inf_f L_{\X,c}(f)$ where $\vv X$ is the linear interpolation of $(X_0,X_1)$. Then $\ell^*_{X,c}$ yields an indication of $c$-optimality of $(X_0,X_1)$, that is, $L_{X,c}^*=0$, iff $(X_0,X_1)$ is $c$-optimal. 3) The minimum $\ell^*_{X,c}$ in the first $k$ iterations of $\crectify$ steps decreases with an $O(1/k)$ rate. \begin{thm}\label{thm:copt} Assume that $c$ is {convex} with conjugate $c^*$, and $c, c^*\in \Cone$ and $\vv X$ is the linear interpolation process of $(X_0,X_1)$. Assume that $(X_0,X_1)$ is a $c$-rectifiable coupling, and $f^{\X,c} \in C^{2,1}(\RR^d\times [0,1])$. Then the following statements are equivalent: i) $(X_0,X_1)$ is a fixed point of $\crectify$, that is, $(X_0,X_1) = \crectify(X_0,X_1)$. ii) $ \ell^*_{X,c}\defeq \inf_{f} L_{\X,c}(f) = L_{\X,c}(f^{\X,c}) = 0$, for $L_{\X,c}$ in \eqref{equ:bregloss}. iii) $(X_0,X_1)$ is a $c$-optimal coupling. \end{thm} \begin{proof} i) $\to$ ii). If $(Z_0,Z_1) = (X_0,X_1)$, we have $S_c(\vv Z)=0$ and $L_{\X,c}(f^{\X,c}) = 0$ following \eqref{equ:straighexz}. iii) $\to$ ii). If $(X_0,X_1)$ is $c$-optimal, we have $\E[c(X_1-X_0)] \leq \E[c(Z_1-Z_0)]$, which again implies that $L_{\X,c}(f^{\X,c}) =0$ following \eqref{equ:straighexz}. ii) $\to$ i) Note that \bb L_{\X,c}(f^{\X,c}) = \int_0^1\e{\bcb{\dot X_t; ~ g^{\X,c}_t(X_t)}} \dt \geq 0. \ee Therefore, $L_{\X,c}(f^{\X,c}) =0$ implies that $\dot X_t= g^{\X,c}_t(X_t)$ $t$-almost surely. Because $Z_t$ satisfies the same equation, whose solution is assumed to be unique, we have $\vv Z =\vv X$ and hence $(Z_0,Z_1) = (X_0,X_1)$. ii) $\to$ iii) Because $\vv X$ is the linear interpolation, we have $X_t = t X_1 + (1-t) X_0$, and {it simultaneously satisfies the ODE $\d X_t= g^{\X,c}_t(X_t) \dt $}. Using Lemma~\ref{lem:hjc2} shows that $(X_0,X_1)$ is $c$-optimal. \end{proof} Knowing that $L_{\X,c}(f^{\X,c})$ is an indication of $c$-optimality, we show below that it is guaranteed to converge to zero with recursive $\map$ updates. \begin{cor}\label{thm:oneoverk} Let $\vv Z^{k}$ be the $k$-th $c$-rectified flow of $(X_0,X_1)$, satisfying $\vv Z^{k+1} = \crectflow((Z_0^k,Z_1^k))$ and $(Z_0^0,Z_1^0) = (X_0,X_1)$. Assume each $(Z_0^k, Z_1^k)$ is $c$-rectifiable for $k=0,\ldots, K$. Then $$ \sum_{k=0}^K L_{\vv Z^k,c}(f^{\vv Z^k,c}) + S_c(\vv Z^{k+1}) \leq \E[c(X_1-X_0)]. $$ Therefore, if $\E[c(X_1-X_0)]<+\infty$, we have $\min_{k\leq K} L_{\vv\Z^k,c}(f^{\vv \Z^k,c}) + S_c(\vv Z^{k+1}) = \bigO{1/K}$. \end{cor} \begin{proof} Applying \eqref{equ:straighexz} to $(Z_0^k,Z_1^k)$ and $(Z_0^{k+1},Z_1^{k+1})$ yields $$ L_{\vv Z^k,c}(f^{\vv Z^k,c}) + S_c(\vv Z^{k+1}) = \E[c(Z_1^k-Z_0^k)] - \E[c(Z_1^{k+1}-Z_0^{k+1})]. $$ Summing it over $k=0,\ldots, K$, \bb \sum_{k=0}^K L_{\vv Z^k,c}(f^{\vv Z^k,c}) + S_c(\vv Z^{k+1}) & = \sum_{k=0}^K \E[c(Z_1^k-Z_0^k)] - \E[c(Z_1^{k+1}-Z_0^{k+1})] \\ & = \E[c(Z_1^0-Z_0^0)] - \E[c(Z_1^{K+1} - Z_0^{K+1})] \\ & \leq \E[c(X_1-X_0)]. \ee \end{proof} \subsection{ $c$-Rectified Flow as Optimization Algorithms} \label{sec:crectifyOptView} In this section, we draw more understanding on how iterative $c$-rectified flowing solves the static and dynamic OT problems. % We first show that $c$-rectified flow can be viewed as an alternative direction descent on the dynamic OT problem \eqref{equ:qt}, and then that $c$-rectified coupling as a majorize-minimization (MM) algorithm on the statistic OT problem~\eqref{equ:mk}. The results in this section are framed in terms of a general path-wise loss function $F_c(\vv Y)$, and hence provide a useful starting point for deriving $c$-rectified flow like approaches to more general optimization problems with coupling constraints. \paragraph{$c$-Rectified flow as alternative direction descent on \eqref{equ:qt}} The mapping $\vv Z^{k+1} = \crectflow(\vv Z^k)$ can be interpreted as an alternative direction descent procedure for the dynamic OT problem \eqref{equ:qt}: % \bbb & \vv X^k = \argmin_{\vv Y} \left \{ F_c(\vv Y) ~~~s.t.~~~ (Y_0,Y_1) =(Z_0^k,Z_1^k) \right\}, \label{equ:linF0} \\ & \vv \Z^{k+1} = \argmin_{\vv Y} \left \{ F_c(\vv Y) ~~~s.t.~~~ \law(Y_t) = \law(X_t^k), ~~ \forall t\in[0,1] \right\}. \label{equ:mpF0} \eee Here in \eqref{equ:linF0}, we minimize $F_c(\vv Y)$ in the set of processes whose start-end pair $(Y_0,Y_1)$ equals the coupling $(Z_0^k, Z_1^k)$ from $\vv Z^k$, which simply yields the linear interpolation $X_t^k = t Z^k_1 + (1-t)Z^k_0$ by Jensen's inequality. In \eqref{equ:mpF0}, we minimize $F_c(\vv Y)$ given the path-wise marginal constraint of $\law(Y_t)=\law(X_t^k)$ for all time $t\in[0,1]$, which yields the $c$-rectified flow following Theorem~\ref{thm:optmultimarg}. Note that the updates in both \eqref{equ:linF0} and \eqref{equ:mpF0} keep the start-end marginal laws $\law(Y_0)$ and $\law(Y_1)$ unchanged, and hence the algorithm stays inside the feasible set $\{\vv Y \colon \law(Y_0)=\tg_0, \law(Y_1)=\tg_1\}$ in \eqref{equ:qt} once it is initialized to be so. The updates in \eqref{equ:linF0}-\eqref{equ:mpF0} highlight a key difference between our method and the {\bbformula} approach~\eqref{equ:cm}-\eqref{equ:bb}: the key idea of {\bbformula} is to restrict the optimization domain to the set of deterministic, ODE-induced processes (a.k.a. flows), but our updates alternate between the deterministic $c$-rectified flow $\vv Z^k$ and the linear interpolation process $\vv X^k$, which is \emph{not} deterministic or ODE-inducable unless the fixed point is achieved. \paragraph{$c$-Rectified flow as an MM algorithm} The majorize-minimization (MM) algorithm \citep{hunter2004tutorial} is a general optimization recipe that works by finding a surrogate function that \emph{majorizes} the objective function. Let $F(X)$ be the objective concave function to be minimize. An MM algorithm consists of iterative update of form $X^{k+1} \in \argmin_Y F^+(Y~|~X^k)$, where $F^+$ is a majorization function of $F$ that satsifies $$ F(Y) = \min_{X} F^+(Y~|~X), ~~~~ \text{and the minimum is attained when $X = Y$}. $$ In this case, the MM update guarantees that % $F(X^k)$ is monotonically non-increasing: $$ F(X^{k+1}) \leq F^+(X^{k+1} | X^k) \leq F^+(X^k~|~X^k) = F(X^k). $$ One can also view MM as conducting coordinate descent on $(X,Y)$ for solving $\min_{X,Y} F^+(Y~|~X)$. In the following, we show that $(Z_0^{k+1},Z_1^{k+1}) = \crectify((Z_0^k,Z_1^k))$ can be interpreted as an MM algorithm for the static OT problem \eqref{equ:mk} for minimizing $\E[c(X_1-X_1)]$ in the set of couplings of $\tgg$. % The majorization function corresponding to $\crectify$ can be shown to be \bb F^+_c((Y_0,Y_1)~|~(X_0,X_1)) & = \inf_{\tilde{\vv Y}} \left \{ F_c(\tilde{\vv Y}) ~~~s.t.~~~ (\tilde Y_0,\tilde Y_1) = (Y_0, Y_1), ~~~ \vv Y\in \mathcal M_X \right \}, \\ & \text{with}~~~~\mathcal M_X = \{\vv Y \colon ~~ \law(Y_t) = \law(t X_1 + (1-t)X_0),~~\forall t\in[0,1]\}, \ee where $F^+_c((Y_0,Y_1)~|~(X_0,X_1))$ denotes the minimum value of $F_c(\tilde{ \vv Y}) $ for $\tilde{\vv Y}$ whose start-end points equal $(Y_0, Y_1)$, and yields the same marginal laws as that of the linear interpolation process of $(X_0,X_1)$. \begin{pro} i) $F^+_c$ yields a majorization function of the $c$-transport cost $\E[c(Y_1-Y_0)]$ in the sense that $$\displaystyle \E[c(Y_1-Y_0)]=\min_{(X_0,X_1)} \{ F^+_c((Y_0,Y_1)~|~(X_0,X_1)), {~~s.t.~~ (X_0,X_1) \in \Pi_{0,1}}\}, $$ and the minimum is attained by $(X_0,X_1) = (Y_0,Y_1)$, {where $\Pi_{0,1}$ denotes the set of couplings of $\tgg$.} ii) % $\crectify$ yields the MM update related $F^+$ in that $$ \displaystyle \crectify((X_0,X_1)) \in \argmin_{(Y_0,Y_1)\in \Pi_{0,1}}F^+_c((Y_0,Y_1)~|~(X_0,X_1)). $$ \end{pro} \begin{proof} i) For any coupling $(X_0,X_1)$ and $(Y_0,Y_1)$, we have $$ F^+_c((Y_0,Y_1)|(X_0,X_1)) \geq \inf_{\tilde{\vv Y}} \left \{ \vv F_c(\tilde{\vv Y}) ~~~s.t.~~~ (\tilde Y_0, \tilde Y_1) = (Y_0,Y_1) \right\} = \E[c(Y_1-Y_0)], $$ where the inequality holds because remove the constraint $\vv Y\in \mathcal M_X$. % In addition, it is obvious that the inequality above becomes equality when $(X_0,X_1) = (Y_0,Y_1)$. ii) Note that $$ \inf_{(Y_0,Y_1)}F^+_c((Y_0,Y_1)~|~(X_0,X_1)) = \inf_{\vv{ Y}} \left \{ F_c(\vv{ Y}) ~~~s.t.~~~ \vv Y\in \mathcal M_X \right \}, $$ whose minimum of the right side is attained by $\vv Y = \crectflow((X_0,X_1))$ following Theorem~\ref{thm:optmultimarg}. Hence, the minimum of the left side is attained by $(Y_0,Y_1) = \crectify((X_0,X_1))$. \end{proof} \subsection{Hamilton-Jacobi Equation and Optimal Transport} \label{sec:hj} The proof of Theorem~\ref{thm:copt} relies on a key lemma shows that if the trajectories of an ODE of form $\d X_t = \dd c^*(\dd f_t(X_t))\dt $ are geodesic in that $X_t = tX_1+(1-t) X_0$, then the induced coupling $(X_0,X_1)$ is an $c$-optimal coupling of its marginals. The proof of this lemma relies on Hamilton-Jacobi (HJ) equation, which provides a characterization of $f$ for an ODE $\d X_t = \dd c^*(\dd f_t(X_t))\dt $ whose trajectories are geodesic. The connection between HJ equation and optimal transport has been a classic result and can be found in, for example, \cite{villani2021topics,villani2009optimal}. \begin{lem}\label{lem:hjc2} Let $v_t(x) = \dd c^*(\dd f_t(x))$ where $c^* \in C^1(\RR^d)$ is a convex function $c$, % and $f \in C^{2,1}(\RR^d\times [0,1])$ and $\dd c^*$ is an injective mapping. Assume all trajectories of $\d x_t = v_t(x_t)\dt$ are geodesic paths in that $x_{t}=t x_1 +(1-t)x_0$. Then we have: i) There exists $\tilde f_t$ such that $\dd \tilde f_t=\dd f_t$ (and hence we can replace $f$ with $\tilde f$ in the assumption), such that the following Hamilton–Jacobi (HJ) equation holds \bbb \label{equ:hj} \partial_t \tilde f_t(x) + c^*(\dd \tilde f_t(x)) =0,~~~\forall x\in \RR^d, ~~ t\in[0,1], && \text{(HJ equation)}. \eee ii) % $f$ satisfies \bb f_t(y_t) = \inf_{y_0 \in \RR^d} \left\{ t c\left (\frac{y_t-y_0}{t} \right ) + f_0(y_0) \right \} , ~~\forall t \in [0,1], ~~~ y_t\in\RR^d, && \text{(Hopf-Lax formula)} \ee where the minimum is attained if $\{y_t\}$ follows the ODE $\d y_t = v_t(y_t)\dt $. iii) Assume a coupling $(X_0,X_1)$ of $\tg_0,\tg_1$ satisfies $\d X_t = v_t(X_t)\dt$. Then $(X_0, X_1)$ is a $c$-optimal coupling. % \end{lem} \begin{proof} i) Starting from any point $x_t = x \in\RR^d$ at time $t$, because the trajectories of $\d x_t = v_t(x_t) \dt $ are geodesic, we have $\dot x_t = v_t(x_t) =\const$ following the trajectory. Because $v_t(x) = \dd c^*(\dd f_t(x))$ and $\dd c^*$ is injective, we have $\dd f_t(x_t) = const$ as well. Hence, we have \bb 0 = \ddt \dd f_t(x_t) & = \partial_t \dd f_t(x_t) + \dd ^2f_t(x_t) \dot x_t \\ & = \partial_t\dd f_t(x_t) + \dd ^2f_t(x_t) \dd c^*(\dd f_t(x_t)). \ee On the other hand, define $h_t(x) = \partial_t f_t(x) + c^*(\dd f_t(x))$. Then we have \bb \dd_x h_t(x) & = \partial_t \dd f_t(x) + \dd^2 f_t(x_t) \dd c^*(\dd f_t(x)) = 0. \ee This suggests that $\dd_x h_t(x) =0$ everywhere and hence $h_t(x)$ does not depend on $x$. Define $\tilde f_t(x) = f_t(x) - \int_0^t h_t(x_0) \d t $, where $x_0$ is any fixed point in $\RR^d$. Then $$ \tilde h_t(x) \defeq \partial_t \tilde f_t(x) + c^*(\dd \tilde f_t(x)) = h_t(x) - h_t(x_0) = 0.$$ ii) Take any $y_0,y_1$ in $\RR^d$, let $y_t = t y_1 + (1-t) y_0$ be their linear interpolation. We have \bb & f_1(y_1) - f_0(y_0) \\ & = \int_0^1 (\partial_t f_t(y_t) + \dd f_t(y_t) \tt (y_t - y_0)) \dt \\ & = \int_0^1 \dd f_t(y_t)\tt (y_1 - y_0) - c^*(\dd f_t(y_t)) \dt \ant{$h_t = \partial f_t + c^*(\dd f_t) = 0$}\\ &\overset{(1)}{\leq} \int_0^1 c(y_1- y_0) \dt \ant{$c(x) + c^*(y) \geq x\tt y$}\\ & = c(y_1- y_0). % \ee The equality in $\overset{(1)}{\leq}$ is attained if $y_t$ follows the geodesic ODE $\d y_t = v_t(y_t) \dt $ as we have $y_1 -y_0 = \dd c^*(\dd f_t(y_t))$, $\forall t$ in this case. A similar derivation holds for $f_t$. iii) Note that i) gives that $c(y_1-y_0) \geq f_1(y_1) - f_0(y_0) $. For any coupling $(Y_0,Y_1)$ of $\tg_0,\tg_1$, we have \bb \E[c(Y_1-Y_0)] \geq \E[f_1(Y_1) - f_0(Y_0)] = \E[f_1(X_1) - f_0(X_0)] = \E[c(X_1- X_0)]. \ee Hence, $(X_0,X_1)$ is a $c$-optimal coupling. \end{proof} \paragraph{Connection to {\bbformula} Formula} The results in Lemma~\ref{lem:hjc2} can also formally derived from {\bbformula} problem \eqref{equ:bb}, as shown in the seminal work of \cite{benamou2000computational}. By introducing a Lagrangian multiplier $\lambda \colon \RR^d\times [0,1]\to \RR$ for the constraint of $\dot \varrho_t + \div (v_t \varrho_t) =0$, the problem in \eqref{equ:bb} can be framed into a minimax problem: $$ \inf_{v,\varrho} \sup_{\lambda } \left\{ \mathcal L(v,\varrho, \lambda) \defeq \int c(v_t) \varrho_t + \int \lambda_t (\dot \varrho_t + \div (v_t \varrho_t) ~~~~s.t.~~~~ \varrho \in \Gamma_{0,1} \right\}, $$ where $\mathcal L(v,\varrho, \lambda) $ is the Lagrangian function, and $\Gamma_{0,1}$ denotes the set of density functions $\{\varrho_t\}_t$ satisfying $\varrho_0=\d \tg_0/\dx, ~ \varrho_1 = \d \tg_1/\dx$. Note that the following integration by parts formulas: \bb \int \lambda_t \div (v_t\varrho_t)+ \dd \lambda_t \tt v_t\varrho_t = 0, && \int \lambda_t \dot \varrho_t + \dot \lambda_t \varrho_t= \lambda_1\varrho_1 - \lambda_0 \varrho_0, % \ee where we assume that $\lambda_t v_r \rho_t$ decays to zero sufficiently fast at infinity. We have $$ \mathcal L(v,\varrho, \lambda) = (\lambda_1 \varrho_1 - \lambda_0 \varrho_0) + \int (c\circ v_t) \rho_t - \dot \lambda_t \varrho_t - \dd \lambda_t \tt (v_t \varrho_t). $$ At the saddle points, the functional derivations of $\mathcal L$ equal zero, yielding \bb \frac{\delta \mathcal L}{\delta \varrho_t} = c(v_t) - \dot \lambda_t - \dd \lambda_t \tt v_t = 0, && \frac{\delta \mathcal L}{\delta v_t} = (\dd c(v_t) - \dd \lambda_t)\varrho_t = 0. \ee Assume $\varrho_t$ is positive everywhere and note that $\dd c^*(\dd c(x)) = x$, we have $v_t = \dd c^*(\dd \lambda_t)$, and hence $\dd \lambda_t\tt v_t - c( v_t) = c^*(\dd \lambda_t)$. Plugging it back to $\frac{\delta \mathcal L}{\delta \rho_t} =0$ yields that $\dot \lambda_t + c^*(\dd \lambda_t)=0$. Overall, the (formal) KKT condition of \eqref{equ:bb} is \bb & \dot \varrho_t + \div (v_t \varrho_t) = 0, ~~~\rho_0 = \d \tg_0/\dx, ~~\rho_1 = \d\tg_1/\dx \ant{coupling condition} \\ & v_t = \dd c^* (\dd \lambda_t) \ant{mapping is gradient of convex function} \\ & \dot \lambda_t + c^*\left ({\dd \lambda_t}\right ) = 0. \ant{Hamilton-Jacobi equation} \ee This matches the result in Lemma~\ref{lem:hjc2} with $\lambda_t = \tilde f_t$. \section{Discussion and Open Questions} \begin{enumerate} \item Corollary~\ref{thm:oneoverk} only bounds the surrogate measure $\ell^*_{Z^k, c}$. Can we directly bound the optimality gap on the $c$-transport cost $e^*_k = \E[c(Z_1^k-Z_0^k)] - \inf_{(Z_0,Z_1)} \E[c(Z_1-Z_0)]$? Can we find a certain type of strong convexity like condition, under which $e_k^*$ decays exponentially with $k$? \item For machine learning (ML) tasks such as generative models and domain transfer, the transport cost is not necessarily the direct object of interest. In these cases, as suggested in \cite{rectified}, rectified flow might be preferred because it is simpler and does not require to specify a particular cost $c$. Question: for such ML tasks, when would it be preferred to use OT with a specific $c$, and how to choose $c$ optimally? \item In practice, recursively applying the ($c$-)rectification accumulates errors because the training optimization for the drift field and the simulation of the ODE can not be conducted perfectly. How to avoid the error accumulation at each step? Assume $\{x_{1,i}\}_i\sim \tg_1$, and $\{z_{0,i}^k, z_{1,i}^k\}_i$ is obtained by solving the ODE of the $k$-th $c$-rectified flow starting from $z_{0,i}^k \sim \tg_0$. As we increase $k$, $\{z_{0,i}^k\}_i$ may yield increasingly bad approximation of $\tg_1$ due to the error accumulation. One way to fix this is to adjust $\{z_{1,i}^k\}$ to make it closer to $\{x_{1,i}^k\}_i$ at each step. This can be done by reweighting/transporting $\{z_{1,i}^k\}_i$ towards $\{x_{1,i}^k\}_i$ by minimizing certain discrepancy measure, or replacing each $z_{1,i}^k$ with $x_{\sigma(i)}^k$ where $\sigma$ is a permutation that yields a one-to-one matching between $\{z_1\datai\}$ and $\{x_1\datai\}_i$. The key and challenging part is to do the adjustment in a good and fast way, ideally with a (near) linear time complexity. \item With or without the adjustment step, build a complete theoretical analysis on the statistical error of the method. \item In what precise sense is rectified flow solving a multi-objective variant of optimal transport? \end{enumerate}

69cec91fa9aa643619acb008dd49277ae7c058ae

\section{Introduction} In recent years, many explanations methods have been developed for explaining machine learning models, with a strong focus on local analysis, i.e., generating explanations for individual prediction (see \citep{molnar2022} for a survey). Among this plethora of methods, one of the most prominent and active techniques are Counterfactual Explanations \citep{Wachter2017CounterfactualEW}. Unlike popular local attribution methods, e.g., SHAP \citep{lundberg2020local2global} and LIME \citep{ribeiro2016why}, which highlight the importance score of each feature, Counterfactuals Explanations (CE) describe the smallest modification to the feature values that changes the prediction to a desired target. Although CE are intuitive and user-friendly by giving recourse in some scenarios (e.g., loan application), they have many shortcomings in practice. Indeed, most counterfactual methods rely on a gradient-based algorithm or heuristics approaches \citep{survey_counterfactual}, thus can fail to identify the most natural explanations and lack guarantees. Most algorithms either do not guarantee sparse counterfactuals (changes in the smallest number of features) or do not generate in-distribution samples (see \citep{counterfactual_r1, CHOU202259} for a survey on counterfactuals methods). Although some works \citep{optimalce_vidal, face_counterfactual, prototype_basedce} try to solve the plausibility/sparsity problem, the suggested solutions are not entirely satisfactory. \\ In another direction, many papers \citep{dice, Karimi2020ModelAgnosticCE, diverce_ce} encourages the generation of diverse counterfactuals in order to find actionable recourse \citep{Ustun2019ActionableRI}. Actionability is a vital desideratum, as some features may be non-actionable, and generating many counterfactuals increases the chance of getting actionable recourse. However, the diversity of CE makes the explanations less intelligible, and the synthesis of various CE or local explanations, in general, is yet to be comprehensively solved \citep{rethinkinxai}. In addition, recently \cite{himanoisycounterfactuals} highlights a new problem of local CE called: \textit{noisy responses to prescribed recourses}. Indeed, in real-world scenarios, some individuals may not be able to implement exactly the prescribed recourses, and they show that most CE methods fail in this noisy environment. Therefore, we propose to reverse the usual way of explaining with counterfactual by computing \textit{Counterfactual rules}. We introduce a new line of counterfactuals: we build interpretable policies for changing a decision with a given probability that ensure the stability of the deduced recourse. These policies are optimal (in sparsity) and faithful to the data distribution. Their computation comes with statistical guarantees as they use a consistent estimator of the conditional distribution. Our proposal is to find a general policy or rule that permits changing the decision while fixing some features instead of generating many counterfactual samples. One of the main challenges is to identify the (minimal) set of features that provide the best promising directions for changing the decision to the desired output. We also show this approach can be extended for finding a collection of regional counterfactuals, such that we have a global counterfactual policy for analyzing a model. An example of the counterfactual rules that we introduce is given in figure \ref{fig:oce}. \begin{figure}[ht] \centering \includegraphics[scale=0.4]{figures/lcr_rcl_illustration.png} \caption{Illustration of the local and regional Counterfactuals Rules that we introduced on a dataset with 4 variables: Age, Salary, Sex, and HoursPerWeek. The Counterfactual Rules define intervals on the minimal subset of features to change the decision of a model prediction in the local counterfactual rule or the decision of a rule that applies on a sub-population in the regional counterfactual rule. In Blue, we have the proposed rules to change the decision.} \label{fig:oce} \end{figure} \section{Motivation and Related works} Most of the methods that propose Counterfactuals Explanations are based on the approach of the seminal work of \cite{wachter2017counterfactual}: the counterfactuals are generated by optimizing a cost, but this procedure does not account directly the plausibility of the counterfactual examples (see \citep{counterfactual_r1} for classification of CE methods). Indeed, a major shortcoming is that the adverse decision needed for obtaining the counterfactual is not designed to be feasible or representative of the underlying data distribution. However, some recent studies proposed ad-hoc plausibility constraint in the optimization, using for instance an outlier score \citep{dace}, an Isolation Forest \citep{optimalce_vidal} or a density-weighted metrics \citep{face_counterfactual} to generate in-distribution samples. In another direction, \cite{prototype_basedce} proposes to use an autoencoder that penalizes out-of-distribution candidates. Instead of relying on ad-hoc constraints, we propose CE that gives plausible explanations by design. Indeed, for each observation, we identify the variables and associated ranges of values that have the highest probability of changing the prediction. We can compute this probability with a consistent estimator of the conditional distribution $P(Y | \boldsymbol{X}_S)$. As a consequence, the sparsity of the counterfactuals is not encouraged indirectly by adding a penalty term ($\ell_0$ or $\ell_1$) as existing works \citep{dice}. Our approach is inspired by the concept of \textit{Same Decision Probability (SDP)} (introduced in \citep{Chen2012TheSP}) that can be used for identifying the smallest subset of features to guarantee (with a given probability) the stability of a prediction. This minimal subset is called \textit{Sufficient Explanations}. In \citep{amoukou2021consistent}, it has been shown that the \textit{SDP} and the \textit{Sufficient Explanations} can be estimated and computed efficiently for identifying important local variables in any classification and regression models. For counterfactuals, we are interested in the dual set: we want the minimal subset of features that have a high probability of changing the decision (when the other features are fixed). Another limitation of the current CE is their local nature and the multiplicity of the explanations produced. While some papers \citep{dice, Karimi2020ModelAgnosticCE, diverce_ce} promote the generation of diverse counterfactual samples to ensure actionable recourse, such diverse explanations should be summarized to be intelligible \citep{rethinkinxai}, but the compilation of local explanations is often a very difficult problem. To address this problem, we do not generate counterfactual samples, but we build a rule \textit{Counterfactual Rules} (CR) from which we can derive counterfactuals. Contrary to classic CE which gives the nearest instances with a desired output, we find the most effective rule for each observation (or group of similar observations) that changes the prediction to the desired target. This local rule easily aggregates similar counterfactuals. For example, if $\boldsymbol{x} = \{ \texttt{Age=20, Salary=35k, HoursWeek=25h, Sex=M}, \dots \}$ with \texttt{Loan=False}, fixing the variables \texttt{Age} and \texttt{Sex} and changing the \texttt{Salary} and \texttt{HoursWeek} change the decision. Therefore, instead of given multiples combination of \texttt{Salary} and \texttt{HoursWeek} (e.g. 35k and 40h or 40k and 55h, \dots) that result in many instances, the counterfactual rule gives the range of values: \texttt{IF HoursWeek $\in \texttt{[35h, 50h]}$, Salary $\in$ [40k, 50k], and the} \textbf{remaining features are fixed} \texttt{THEN Loan=True}. It can be extended at a regional scale, e.g., given a rule $\textbf{R} = \{\texttt{IF Salary} \in \texttt{[35k, 20k], Age} \in \texttt{[20, 80] THEN Loan=False}\}$, the regional Counterfactual Rule (CR) could be $\{ \texttt{\textbf{IF } Salary} \in \texttt{[40k, 50k],} \texttt{ HoursWeek} \in \texttt{[35h, 50h] and the} \textbf{ remaining rules are fixed} \texttt{ THEN Loan=True}\}$. The main difference between a local and a global CR is that the Local-CR explain a single instance by fixing the remaining feature values (not used in the CR) ; while a regional-CR is defined by keeping the remaining variables in a given interval (not used in the regional-CR). Moreover, by giving ranges of values that guarantee a high probability of changing the decision, we partly answer the problem of \textit{noisy responses to prescribed recourses} \citep{himanoisycounterfactuals} so long as the perturbations are within our ranges. Although the \textit{Local Counterfactual Rule} is new, the \textit{Regional Counterfactual Rule} can be related to some recent works. Indeed, \cite{rawal2020beyond} proposed Actionable Recourse Summaries (AReS), a framework that constructs global counterfactual recourses in order to have a global insight of the model and detect unfair behavior. While AReS is similar to the Regional Counterfactual Rule, we emphasize some significant differences. Our methods can address regression problems and deal with continuous features. Indeed, AReS needs to discretize the continuous features, inducing a trade-off between speed and performance as noticed by \citep{globalce}. Thus, too few bins result in unrealistic recourse, while too many bins result in excessive computation time. In addition, AReS uses a greedy heuristic search approach to find global recourse, which might produce sub-optimal recourse. As we have already mentioned, the changes we provide overcome these two limitations because the consistency of our counterfactual is controlled by an estimation of the probability of changing the decision, and because we favor changes of a minimum number of features. Another global CE framework has been introduced in \citep{cet4} to ensure transparency: the Counterfactual Explanation Tree (CET) partitions the input space with a decision tree and assigns an appropriate action for changing the decision of each subspace. Therefore, it gives a unique action for changing the decision of multiple instances. In our case, we offer more flexibility in the counterfactual explanations because we provide a range of possible values that guarantee a change with a given probability. In our approach, we do not make any assumption about the cost of changing the feature nor the causal structure. If we have such information, then we can add it as additional post-processing such that it can be made more explicit and more transparent for the final user as required for trustworthy AI. \section{Minimal Counterfactual Rules} We assume that we have an i.i.d sample $\mathcal{D}_n = \{(\boldsymbol{X}_i,Y_i)_{i=1,\dots,n}\}$ such that $(\X, Y) \sim P_{(\X, Y)}$ where $\X \in \mathcal{X}$ (typically $\mathcal{X}=\mathbb{R}^p$) and $Y \in \mathcal{Y}$. The output $\mathcal{Y}$ can be discrete or continuous. We want to explain the predictor $f:\mathbb{R}^p \mapsto \mathcal{Y}$, that has been learned with the dataset $\mathcal{D}_n$. We use uppercase letters for random variables and lowercase letters for their value assignments. For a given subset $S \subset [p]$, $\XS = (X_i)_{i \in S}$ denotes a subgroup of features, and we write $\x=(\xs,\xsb)$ (with some abuse of notation). For an observation $(\x,y=f(\x))$, we have a target set $\YSt \subset \mathcal{Y}$, such that $y\notin \YSt$. For the simple case of classification problem, $\YSt = \{ y^\star\}$ is the standard singleton such that $y^\star\in \mathcal{Y}$ is different of $y$. Contrary to standard approaches, our definition of the counterfactual deals also with the regression case by considering $\YSt = [a,b]\subset \mathbb{R}$; our definitions and computations of counterfactuals are the same for both classification and regression. We remind that the classic CE problem (defined only for classification) is to find a function $\A: \mathcal{X} \mapsto \mathcal{X}$, such that for all observations $\x \in \mathcal{X}$, $f(\x)\neq y^\star$, and we have $f(\A(\x))=y^\star$. With standard CE, the function is defined point-wise by solving an optimisation program. Most often $\A(\cdot)$ is not a real function, as $\A(x)$ may be in fact a collection of (random) values $\{\x_1^\star,\dots,\x_p^\star\}$. A more recent point of view was proposed by \cite{cet4}, and it defines $\A$ as a decision tree, where in each leaf $L$, the best perturbation $a_L$ is predicted and add it to all the instances $\x \in L$. \\ Our approach is hybrid, because we do not propose a single action for each subspace of $\mathcal{X}$ or sub-group of population, but we give sets of possible perturbations. Indeed, a \emph{Local Counterfactual Rule} (Local-CR) for $\YSt$ and observation $\x$ (with $f(\x)\notin \YSt$) is a rectangle $C_{S}(\boldsymbol{x};\YSt) = \prod_{i\in S} [a_i, b_i], a_i, b_i \in \overline{\mathbb{R}}$ such that for all perturbations of $\x=\left(\xs,\xsb \right)$ obtained as $\boldsymbol{x}^\star = \left(\zs,\xsb \right)$ with $\zs \in C_{S}(\x;\YSt)$ and $\boldsymbol{x}^\star$ an in-distribution sample, then $f\left( \boldsymbol{x}^\star\right)$ is in $\YSt$ with a high probability.\\ Similarly, a \emph{Regional Counterfactual Rule} (Regional-CR) $C_S(\boldsymbol{R}; \YSt)$ is defined for $\YSt$ and a rectangle $\boldsymbol{R}=\prod_{i=1}^{d} [a_i, b_i], a_i, b_i \in \overline{\mathbb{R}}$, if for all observations $\x=(\xs,\xsb) \in \boldsymbol{R}$, the perturbations obtained as $\boldsymbol{x}^\star = (\zs,\xsb)$ with $\zs \in C_S(\boldsymbol{R},\YSt)$ and $\boldsymbol{x}^\star$ an in-distribution sample are such that $f\left( \boldsymbol{x}^\star\right)$ is in $\YSt$ with high probability.\\ We build such rectangles sequentially, first, we propose to find the best directions $S \subset [p]$ that offers the best probability of change. Then, we find the best intervals $[a_i, b_i], i \in S$ that change the decision to the desired target. A central tool in this approach is the Counterfactual Decision Probability. \begin{definition} \label{def:cdp}\textbf{Counterfactual Decision Probability (CDP).} The Counterfactual Decision Probability of the subset $S\subset\left\llbracket 1,p\right\rrbracket $, w.r.t $\boldsymbol{x}=\left(\boldsymbol{x}_{S},\boldsymbol{x}_{\bar{S}}\right)$ and the desired target $\YSt$ (s.t. $f(\x)\notin \YSt)$ i \[CDP_{S}\left(\YSt; \boldsymbol{x}\right)=P\left(f(\X) \in \YSt\left|\boldsymbol{X}_{\bar{S}}=\boldsymbol{x}_{\bar{S}}\right.\right). \nonumber\] \end{definition} The $CDP$ of the subset S is the probability that the decision changes to the desired target $\YSt$ by sampling the features $\XS$ given $\boldsymbol{X}_{\bar{S}} = \boldsymbol{x}_{\bar{S}}$. It is related to the Same Decision Probability $SDP_{S}(\mathscr{Y}; \boldsymbol{x}) = P\left(f(\X) \in \mathscr{Y} \vert \XS=\xs \right)$ used in \citep{amoukou2021consistent} for solving the dual problem of selecting the most local important variables for obtaining and maintaining the decision $f(\x) \in \mathscr{Y}$ (where $f(\x)\in\mathscr{Y}\subset \mathcal{Y}$). The set $S$ is called the Minimal Sufficient Explanation. Indeed, we have $CDP_S(\YSt; \boldsymbol{x}) = SDP_{\bar{S}}(\YSt; \boldsymbol{x})$. The computation of these probabilities is challenging and discussed in Section 4. We now focus on the minimal subset of features $S$ such that the model makes the desired decision with a given probability $\pi$ \begin{definition} \label{def:minimal_countset}(\textbf{ Minimal Divergent Explanations}). Given an instance $\boldsymbol{x}$ and a desired target $\YSt$, $S$ is a Divergent Explanation for probability $\pi>0$, if $CDP_{S}\left(\YSt;\boldsymbol{x}\right)\geq\pi$, and no subset $Z$ of $S$ satisfies $CDP_{Z}\left(\YSt;\boldsymbol{x}\right)\geq\pi$. Hence, a Minimal Divergent Explanation is a Divergent Explanation with minimal size. \end{definition} The set minimizing this probability is not unique, and we can have several Minimal Divergent Explanations. Note that the probability $\pi$ represents the minimum level required for a set to be chosen for generating counterfactuals, and its value should be as high as possible and depends on the use case. We have now enough material to define our main criterion for building a Local Counterfactual Rule (Local-CR): \begin{definition}\label{def:local_counterfactual_rule} (\textbf{Local Counterfactual Rule}). Given an instance $\boldsymbol{x}$, a desired target $\YSt \not\owns f(\x)$ , a Minimal Divergent Explanation $S$, the rectangle $C_{S}(\boldsymbol{x}; \YSt) = \prod_{i\in S} [a_i, b_i], a_i, b_i \in \overline{\mathbb{R}}$ is a Local Counterfactual Rule with probability $\pi_C$ if \begin{align} \label{eq:crp_instance} CRP_S(\YSt,\x, C_S(\x;y^\star)) \triangleq P( f(\X) \in \YSt \; | \boldsymbol{X}_S \in C_S(\boldsymbol{x};\YSt), \boldsymbol{X}_{\bar{S}} = \boldsymbol{x}_{\bar{S}}) \geq \pi_C. \end{align} The $CRP_S$ is the Counterfactual Rule Probability. \end{definition} The higher the probability $\pi_C$ is, the better the relevance of the rule $C_S(\x; \YSt)$ is, for this instance. Given a set $S$, we seek for the maximal rectangle in the direction $S$ satisfying Definition \ref{eq:crp_instance}. In practice, we can observe that the Local-CR $C_{S}(\cdot;\YSt)$ for neighbors $\x,\x'$ are often quite close, because the Minimal Divergent Explanations are similar and the corresponding rectangles often overlaps. Hence, this motivates a generalisation of these Local-CR to hyperrectangle $\boldsymbol{R} = \prod_{i=1}^{d} [a_i, b_i], a_i, b_i \in \overline{\mathbb{R}}$ regrouping similar observations. We denote $\text{supp}(\boldsymbol{R}) = \{i : [a_i, b_i] \neq \overline{\mathbb{R}}\}$ the support of the rectangle, and we extend the Local-CR to Regional Counterfactual Rules (Regional-CR). In order to do it, we denote $\boldsymbol{R}_{\bar{S}} = \prod_{i \in \bar{S}} [a_i, b_i]$ as the rectangle with intervals of $\boldsymbol{R}$ in $\text{supp}(\boldsymbol{R}) \cap \bar{S}$ and we also defines the corresponding Counterfactual Decision Probability CDP (Definition \ref{def:cdp}) for rule $\boldsymbol{R}$ and subset $S$ as $CDP_S(\YSt; \boldsymbol{R}) = P\left(f(\X) \in \YSt \left|\boldsymbol{X}_{\bar{S}} \in \boldsymbol{R}_{\bar{S}}\right.\right)$. Therefore, we can also compute the Minimal Divergent Explanation for rule $\boldsymbol{R}$ using Definition \ref{def:minimal_countset} with the CDP for rules. \begin{definition}\label{def:regional_rule} (\textbf{Regional Counterfactual Rule}). Given any rectangle $\boldsymbol{R}$, a desired target $\YSt$, a Minimal Divergent Explanation $S$ of $R$, the rectangle $C_S(\boldsymbol{R}; y^\star) = \prod_{i\in S} [a_i, b_i]$ is a Regional Counterfactual Rule with probability $\pi_C$ if \begin{align} \label{eq:crp_rule} CRP_S(\YSt; \boldsymbol{R}, C_S(\boldsymbol{R}, \YSt)) \triangleq P( f(\X) \in \YSt \; | \boldsymbol{X}_S \in C_S(\boldsymbol{R},\YSt), \XSb \in \boldsymbol{R}_{\bar{S}} )\geq \pi_C. \end{align} $CRP_S(\YSt; \boldsymbol{R}, C_S(\boldsymbol{R}))$ is the corresponding Counterfactual Rule Probability for rule $\boldsymbol{R}$. \end{definition} \paragraph{Remarks: } Local-CR and regional-CR differ slightly: for local, we condition by $\boldsymbol{X}_{\bar{S}} = \boldsymbol{x}_{\bar{S}}$ in Eq. \ref{eq:crp_instance}, while for regional, we condition by $\boldsymbol{X}_{\bar{S}} \in \boldsymbol{R}_{\bar{S}}$. For computing regional-CR, we can start for a rectangle generated by any method, such as \citep{bayesianRuleListRudin, OptimalDecisionTreeRudin}. The only condition is that it contains a homogeneous group, i.e. with almost the same output. However, by default we use as initial rules the Sufficient Rules derived in \citep{amoukou2021consistent} as it handles regression problem. The Sufficient Rules are minimal support rectangles define for a given output $\mathscr{Y}$ as $C_S(\mathscr{Y}) = \Pi_{i\in S} [a_i,b_i]$ such that $\forall \x \in \mathcal{X}, \xs \in C_S(\mathscr{Y})$, $P(f(\X) \in \mathscr{Y} \vert \XS = \xs) \geq \pi$. \section{Estimation of the $CDP$ and $CRP$} In order to compute the probabilities $CDP_S$ and $CRP_S$ for any $S$, we use a dedicated Random Forest (RF) $m_{k, n}$ that learns the model $f$ to explain. Indeed, the conditional probabilities $CDP_S$ and $CRP_S$ can be easily computed from a RF by combining the Projected Forest algorithm \citep{benard2021shaff} and the Quantile Regression Forest \citep{meinshausen2006quantile}: hence we can estimate consistently the probabilities $CDP_S(\mathscr{Y}^\star; \boldsymbol{x})$. We adapt the approach used in \citep{amoukou2021consistent} and remind for the sake of completeness, the computation of the estimate of $SDP_S$. \subsection{Projected Forest and $CDP_S$} The estimator of the $SDP_S$ is built upon a learned Random Forest \citep{breiman1984classification}. A Random Forest (RF) is a predictor consisting of a collection of $k$ randomized trees (see \citep{Loh2011ClassificationAR} for a detailed description of decision tree). For each instance $\boldsymbol{x}$, the predicted value of the $j$-th tree is denoted $m_n(\boldsymbol{x}, \Theta_j)$ where $\Theta_j$ represents the resampling data mechanism in the $j$-th tree and the successive random splitting directions. The trees are then averaged to give the prediction of the forest as: \begin{align} \label{eq:random_forest_baggin_estimator} \small m_{k, n}(\boldsymbol{x}, \Theta_{1:k}, \mathcal{D}_n) = \frac{1}{k} \sum_{l=1}^{k} m_n(\boldsymbol{x}; \Theta_l, \mathcal{D}_n) \end{align} However, the RF can also be view as an adaptive nearest neighbor predictor. For every instance $\boldsymbol{x}$, the observations in $\mathcal{D}_n$ are weighted by $w_{n, i}(\boldsymbol{x}; \Theta_{1:k}, \mathcal{D}_n)$, $i=1, \dots, n$. Therefore, the prediction of RF can be rewritten as\[ \small m_{k, n}(\boldsymbol{x}, \Theta_{1:k}, \mathcal{D}_n) = \sum_{i=1}^{n} w_{n, i}(\boldsymbol{x}; \Theta_{1:k}, \mathcal{D}_n) Y_i. \nonumber \] This emphasizes the central role played by the weights in the RF's algorithm, see \citep{meinshausen2006quantile, amoukou2021consistent} for detailed description of the weights. Therefore, it naturally gives estimators of other quantities e.g., Cumulative hazard function \citep{ishwaran2008random}, Treatment effect \citep{wager2017estimation}, conditional density \citep{du2021wasserstein}. For instance, \cite{meinshausen2006quantile} showed that we can used the same weights to estimate the Conditional Distribution Function with the following estimator: \begin{align} \widehat{F}(y | \boldsymbol{X} = \boldsymbol{x}, \Theta_{1:k}, \mathcal{D}_n) = \sum_{i=1}^{n} w_{n, i}(\boldsymbol{x}; \Theta_{1:k}, \mathcal{D}_n) \mathds{1}_{Y_{i} \leq y} \label{eq:estimator_boostrap_quantile} \end{align} In another direction, \cite{benard2021shaff} introduced the Projected Forest algorithm \citep{benard2021mda, benard2021shaff} that aims to estimate $E[Y | \boldsymbol{X}_S]$ by modifying the RF's prediction algorithm. \paragraph{Projected Forest:} To estimate $E[Y | \XS = \xs]$ instead of $E[Y | \boldsymbol{X} = \x]$ using a RF, \cite{benard2021interpretable} suggests to simply ignore the splits based on the variables not contained in $S$ from the tree predictions. More formally, it consists of projecting the partition of each tree of the forest on the subspace spanned by the variables in S. The authors also introduced an algorithmic trick that computes the output of the Projected Forest efficiently without modifying the initial tree structures. We drop the observations down in the initial trees, ignoring the splits which use a variable not in $S$: when it encounters a split involving a variable $i \notin S$, the observations are sent both to the left and right children nodes. Therefore, each instance falls in multiple terminal leaves of the tree. To compute the prediction of $\xs$, we follow the same procedure, and gather the set of terminal leaves where $\xs$ falls. Next, we collect the training observations which belong to every terminal leaf of this collection, in other words, we keep only the observations that fall in the intersection of the leaves where $\xs$ falls. Finally, we average their outputs $Y_i$ to generate the estimation of $E[Y | \XS = \xs]$. Notice that the author show that this algorithm converges asymptotically to the true projected conditional expectation $E[Y | \XS = \xs]$. As the RF, the PRF gives also a weight to each observation. The associated PRF is denoted $ m_{k, n}^{(\xs)}(\xs) = \sum_{i=1}^{n} w_{n, i}(\xs) Y_i$. Therefore, as the weights of the original forest was used to estimate the CDF in equation \ref{eq:estimator_boostrap_quantile}, \cite{amoukou2021consistent} used the weights of the Projected Forest Algorithm to estimate the $SDP$ as $\widehat{SDP}_{S}\left(\mathscr{Y};\boldsymbol{x}\right) = \sum_{i=1}^{n} w_{n, i}(\xs) \mathds{1}_{Y_i \in \mathscr{Y}}$. The idea is essentially to replace $Y_i$ by $\mathds{1}_{Y_i \in \mathscr{Y}}$ in the Projected Forest equation defined above. The authors also show that this estimator converges asymptotically to the true $SDP_S$. Therefore, we can estimate the $CDP$ with the following estimator \begin{equation} \label{eq:CDP_estimator} \widehat{CDP}_{S}\left(\YSt;\boldsymbol{x}\right) = \sum_{i=1}^{n} w_{n, i}(\xsb) \mathds{1}_{Y_i \in \YSt}. \end{equation} \paragraph{Remarks:} Note that we only give the estimator of the $CDP_S$ of an instance $\x$. The estimator of the $CDP_S$ of a rule $R$ will be discussed in the next section as it is related to the estimator of the $CRP_S$. \subsection{Regional RF and $CRP_S$} In this section, we focus on the estimation of the $CRP_S(\YSt, \x, C_S(\x;\YSt)) = P(f(\X) \in \YSt \; | \boldsymbol{X}_S \in C_S(\boldsymbol{x}; \YSt), \boldsymbol{X}_{\bar{S}} = \boldsymbol{x}_{\bar{S}})$ and $CRP_S(\YSt, \boldsymbol{R}, C_S(\boldsymbol{R};\YSt)) = P(f(\X) \in \YSt \; | \boldsymbol{X}_S \in C_S(\boldsymbol{R};\YSt), \boldsymbol{X}_{\bar{S}} \in \boldsymbol{R}_{\bar{S}})$. For simplicity, we remove the dependency of the rectangles in $\YSt$. Based on the previous Section, we already know that the estimators using the RF will be in the form of $\widehat{CRP}_{S}\left(\YSt,\boldsymbol{x}, C_S(\boldsymbol{x})\right) = \sum_{i=1}^{n} w_{n, i}(\x) \mathds{1}_{Y_i \in \YSt}$, thus we only need to find the right weighting. The main challenge is that we have a condition based on a region, e.g., $\XS \in C_S(\boldsymbol{x})$ or $\boldsymbol{X}_{\bar{S}} \in \boldsymbol{R}_{\bar{S}}$ (regional-based) instead of condition of type $\XS = \xs$ (fixed value-based) as usually. However, we introduced a natural generalization of the RF algorithm to make predictions when the conditions are both regional-based and fixed value-based. Thus, the case where there are only regional-based conditions are naturally derived. \paragraph{Regional RF to estimate $CRP_S(\YSt,\x, C_S(\x)) = P(f(\X) \in \YSt \; | \boldsymbol{X}_S \in C_S(\boldsymbol{x}), \boldsymbol{X}_{\bar{S}} = \boldsymbol{x}_{\bar{S}})$:} The algorithm is based on a slight modification of RF. Its works as follow: we drop the observations in the initial trees, if a split used variable $i \in \bar{S}$, i.e., fixed value-based condition, we use the classic rules of RF, if $x_i \leq t$, the observations go to the left children, otherwise the right children. However, if a split used variable $i \in S$, i.e, regional-based condition, we use the rectangles $C_S(\boldsymbol{x}) = \prod_{i=1}^{|S|} [a_i, b_i]$. The observations are sent to the left children if $b_i \leq t$, right children if $a_i > t$ and if $t \in [a_i, b_i]$ the observations are sent both to the left and right children. Therefore, we use the weights of the Regional RF algorithm to estimate the $CRP_S$ as in equation \ref{eq:CDP_estimator}, the estimator is $\widehat{CRP}_S(y^\star; \boldsymbol{x}, C_S(\boldsymbol{x})) = \sum_{i=1}^{n} w_{n, i}(\x) \mathds{1}_{Y_i = y^\star}$. A more detailed version of the algorithm is provided and discussed in Appendix. To estimate the $CDP$ of a rule $CDP_{S}\left(\YSt; \boldsymbol{R}\right)=P\left(f(\X) \in \YSt \left|\boldsymbol{X}_{\bar{S}}\in \boldsymbol{R}_{\bar{S}} \right.\right)$, we just have to apply the projected Forest algorithm to the Regional RF, i.e., when a split involving a variable outside of $\bar{S}$ is met, the observations are sent both to the left and right children nodes, otherwise we use the Regional RF split rule, i.e., if an interval of $\boldsymbol{R}_{\bar{S}}$ is below $t$, the observations go to the left children, otherwise the right children and if $t$ is in the interval, the observations go to the left and right children. The estimator of the $CRP_S(\YSt; \boldsymbol{R}, C_S(\boldsymbol{R}))$ for rule is also derived from the Regional RF. Indeed, it is a special case of the Regional RF algorithm where there are only regional-based conditions. \section{Learning the Counterfactual Rules} We compute the Local and Regional CR using the estimators of the previous section. First, we find the Minimal Divergent Explanation in the same way as Minimal Sufficient Explanation can be found \citep{amoukou2021consistent}. As the exploration of all possible subsets is exponential, we search the Minimal Divergent Subset among the $K=10$ most frequently selected variables in the RF $m_{k,n}$ used to estimate the probabilities $CDP_S, CRP_S$ ($K$ is an hyper-parameter to select according to the use case and computational power). We can also use any importance measure. Given an instance $\boldsymbol{x}$ or rectangle $\boldsymbol{R}$ (and set $\YSt$) and their corresponding Minimal Divergent Explanation S, we want to find a rule $C_S(\boldsymbol{x}) = \prod_{i \in S} [a_i, b_i]$ s.t. given $\boldsymbol{X}_{\bar{S}} = \boldsymbol{x}_{\bar{S}}$ or $\boldsymbol{X}_{\bar{S}} \in \boldsymbol{R}_{\bar{S}}$ and $\XS \in C_S(\boldsymbol{x})$, the probability that $Y \in \YSt$ is high. More formally, we want: $P(f(\X) \in \YSt | \XS \in C_S(\boldsymbol{x}), \boldsymbol{X}_{\bar{S}} = \boldsymbol{x}_{\bar{S}})$ or $P(f(\X) \in \YSt| \XS \in C_S(\boldsymbol{x}), \boldsymbol{X}_{\bar{S}} \in \boldsymbol{R}_{\bar{S}} )$ above $\pi_C$. The computation of the rectangles $C_S(\boldsymbol{x}) = \prod_{i\in S|} [a_i, b_i]$ relies heavily on our use of RF and on the algorithmic trick of the projected RF. Indeed, the rectangles defining the rules arise naturally from RF, while AReS \citep{rawal2020beyond} relies on binned variables to generate candidate rules and tests all these possible rules for choosing an optimal one. We overcome the computational burden and the challenge of choosing the number of bins. \begin{figure}[!htb] \minipage{0.30\textwidth} \includegraphics[width=\linewidth]{figures/tree_partition.png} \caption{The partition of the RF learned to classify the toy data (Green/Blue stars). Its has 10 leaves. The explainee $\boldsymbol{x}$ is the Blue triangle in leaf 5. }\label{fig:forest_part} \endminipage\hfill \minipage{0.30\textwidth} \includegraphics[width=\linewidth]{figures/projected_partition.png} \caption{The partition of the projected Forest when we condition on $X_0$, i.e., ignoring the splits based on $X_1$ (the dashed lines).}\label{fig:projected_part} \endminipage\hfill \minipage{0.30\textwidth}% \includegraphics[width=\linewidth]{figures/counterfactual_rule.png} \caption{The optimal CR for $\boldsymbol{x}$ when we condition given $X_0=x_0$ is the Green region, its corresponds to the union of leaf 3 and 4 of the forest}\label{fig:cr} \endminipage \end{figure} To illustrate the idea, we use a two-dimensional data $(X_0, X_1)$ with label Y represented as Green/Blue stars in figure \ref{fig:forest_part}. We fit a Random Forest to classify this dataset and show its partition in figure \ref{fig:forest_part}. The explainee $\boldsymbol{x}$ is the Blue triangle observation. By looking at the different cells/leaves of the RF, we can guess that the Minimal Divergent Explanation of $\boldsymbol{x}$ is $S = X_1$. Indeed, in figure \ref{fig:projected_part}, we observe the leaves of the Projected Forest when we do not condition on $S = X_1$, thus projected the RF's partition only on the subspace $X_0$. Its consists of ignoring all the splits in the other directions (here the $X_1$-axis), thus $\boldsymbol{x}$ falls in the projected leaf 2 (see figure \ref{fig:projected_part}) and its $CDP$ is $CDP_{X_1}(\text{Green}; \boldsymbol{x})=\frac{10 \text{ Green}}{10\text{ Green} + 17\text{ Blue}} = 0.58$. Finally, the problem of finding the optimal rectangle $C_S(\boldsymbol{x}) = [a_i, b_i]$ in the direction of $X_1$ s.t. the decision changes can be easily solved by using the leaves of the RF. In fact, by looking at the leaves of the RF (figure \ref{fig:forest_part}) of the observations that belong in the projected RF leaf 2 (figure \ref{fig:projected_part}) where $\boldsymbol{x}$ falls, we see in figure \ref{fig:cr} that the optimal rectangle to change the decision given $X_0 = x_0$ or being in the projected RF leaf 2 is the union of the intervals on $X_1$ of the leaf 3 and 4 of the RF (see the Green region of figure \ref{fig:cr}). Given an instance $\boldsymbol{x}$ and its Minimal Divergent Explanation $S$, the first step is the collect of the observations which belong to the leaf of the Projected Forest given $\bar{S}$ where $\boldsymbol{x}$ falls. It corresponds to the observations that has positive weights in the computation of the $CDP_S(\YSt; \boldsymbol{x}) = \sum_{i=1}^{n} w_{n, i}(\boldsymbol{x}_{\bar{S}}) \mathds{1}_{Y_i \in \YSt}$, i.e., $\{\x_i: w_{n, i}(\boldsymbol{x}_{\bar{S}}) >0\}$. Then, we used the partition of the original forest to find the possible leaves $C_S(\boldsymbol{x})$ in the direction $S$. The possible leaves is among the RF's leaves of the collected observations $\{\boldsymbol{x}_i: w_{n,i}(\boldsymbol{x}_{\bar{s}}) >0\}$. Let denote $L(\boldsymbol{x}_i)$ the leaves of the observations $\x_i$ with $w_{n, i}(\boldsymbol{x}_{\bar{S}}) >0$. A possible leaf is a leaf $L(\boldsymbol{x}_i)$ s.t. $CRP_S(\YSt, \boldsymbol{x}, L(\boldsymbol{x}_i)_S) = P( f(\X) \in \YSt | \XS \in L(\x_i)_S, \boldsymbol{X}_{\bar{S}} = \boldsymbol{x}_{\bar{S}}) \geq \pi_C$. Finally, we merge all the neighboring possible leaves to get the largest rectangle, and this maximal rectangle is the counterfactual rule. Note that the union of the possible leaves is not necessary a connected space, thus we can have multiple counterfactual rules. We apply the same idea to find the regional CR. Given a rule $\boldsymbol{R}$ and its Minimal Divergent Explanation $S$, we used the Projection given $\boldsymbol{X}_{\bar{S}} \in \boldsymbol{R}_{\bar{S}}$ to find the compatible observations and their leaves and combine the possible ones to obtain the regional CR that has $CRP_S(\YSt, \boldsymbol{R}, C_S(\boldsymbol{R})) \geq \pi_C$. For example, if we consider the leaf 5 of the original forest as a rule: \texttt{If $\boldsymbol{X} \in $ Leaf 5, then predict Blue}. Its Minimal Divergent Explanation is also $S=X_1$. The R-CR would also be the Green region in figure \ref{fig:cr}. Indeed, if we satisfy the $X_0$ condition of the leaf 5 and $X_1$ condition of the leaf 3 and 4, then the decision change to Green. \section{Experiments} To demonstrate the performance of our framework, we conduct two experiments on real-world datasets. The first consists of showing how we can use the \textit{Local Counterfactual Rules} for explaining a regression model. In the second experiment, we compare our approaches with the 2 baselines methods in classification problem: (1) \textbf{CET} \citep{cet4}, which partition the input space using a decision tree and associate a vector perturbation for each leaf, (2) \textbf{AReS} \citep{rawal2020beyond} performs an exhaustive search for finding global counterfactual rules, but we used the implementation of \cite{cet4} that adapts the algorithm for returning counterfactuals samples instead of rules. We compare the methods only in classification problem as most prior works do not deal regression problem. In all experiments, we split our dataset into train ($75\%$) - test ($25\%$), and we learn a model $f$, a LightGBM \textit{(estimators=50, nb leaves=8)}, on the train set that is the explainee. We learn $f$'s predictions on the train set with an approximating RF $m_{nb,n}$ \textit{(estimators=20, max depth=10)}: \textbf{that} will be used to generate the CR with $\pi=0.9$. The used parameters for \textbf{AReS}, \textbf{CET} are \textit{max rules=8, bins=10} and \textit{max iterations=1000, max leaf=8, bins=10} respectively. Due to page limitation, the detailed parameters of each method are provided in Appendix. \paragraph{Sampling CE using the Counterfactual Rules:} Notice that our approaches cannot be directly compare with the baseline methods since they all return counterfactual samples while we give rules (range of vector values) that permit to change the decision with high probability. However, we adapt the CR to generate also counterfactual samples using a generative model. For example, given an instance $\x = (\xs, \x_{\bar{S}})$, target $\YSt$ and its counterfactual rule $C_S(\x; \YSt)$, we want to find a sample $x^\star = (\boldsymbol{z}_S, \x_{\bar{S}})$ with $\boldsymbol{z}_S \in C_S(\x, \YSt)$ s.t $\x^\star$ is an in-distribution sample and $f(\x^\star) \in \YSt$. Instead of using a complex conditional generative model as \citep{modeling_td, sdv} that can be difficult to calibrate, we use an energy-based generative approach \citep{ebmduvenaud, yanebm}. The core idea is to find $\boldsymbol{z}_S \in C_S(\x, y^\star)$ s.t. $\x^\star$ maximize a given energy score to ensure that it is an in-distribution sample. As an example of an energy function, we use the negative outlier score of an Isolation Forest \citep{liu2008isolation}. We use Simulated Annealing (see \citep{review_simulated_annealing} for a review) to maximize the negative outlier score using the information of the counterfactual rules $C_S(\x; \YSt)$. In fact, the range values given by the CR $C_S(\x; \YSt)$ reduce the search space for $\boldsymbol{z}_S$ drastically. We used the training set $\mathcal{D}_n$ to find the possible values i.e., we defined $P_i$, $P_S$ as the list of values of the variable $i \in S$ found in $\mathcal{D}_n$ and $P_S = \{ \boldsymbol{z}_S = (z_1, \dots, z_S): \boldsymbol{z}_S \in C_S(\x, y^\star), z_i \in P_i\}$ the possible values of $\boldsymbol{z}_S$ respectively. Then, we sample $\boldsymbol{z}_S$ in the set $P_S$ and use Simulated Annealing to find a $\x^\star$ that maximizes the negative outlier score. Note that the algorithm is the same for sampling CE with the Regional-CR. A more detailed version of the algorithm is provided in Appendix. Finally, we compare the methods on unseen observations using three criteria. \textit{Correctness} is the average number of instances for which acting as prescribed change to the desired prediction. \textit{Plausibility} is the average number of inlier (predict by an Isolation Forest) in the counterfactual samples. \textit{Sparsity} is the average number of features that have been changed, and especially for the global counterfactual methods (AReS, Regional-CR) that do not ensure to cover all the instances, we compute \textit{Coverage} that corresponds to the average number of unseen observations we cover. \paragraph{Local counterfactual rules for regression:} We give recourse for the \textbf{California House Price} dataset \citep{california_data} derived from the 1990 U.S. census. We have information about each district (demography, \dots), and the goal is to predict the median house value of each district. To illustrate the efficiency of the Local-CR, we select all the observations in the test set having a price lower than $100k$ (1566 houses), and we aim to find the recourse that permit to increase their price : we want the price $y$ to be in the interval $\YSt=[200k, 250k]$. For each instance $\x$, we compute the Minimal Divergent Explanation $S$, the Local-CR $C_S(\x; [200k, 250k])$ and a CE using the Simulated Annealing as described above. We succeed in changing the decision of all the observations $(\textit{Correctness}=1)$ and most of them passed the outlier test with $\textit{Plausibility}=0.92$. On top of that, our Local-CR have sparse support ($\textit{Sparsity}=4.45$). For example, the Local-CR of the instance $\x =$ \texttt{(Longitude=-118.2, latitude=33.8, housing median age=26, total rooms=703, total bedrooms=202, population=757, households=212, median income=2.52)} is $C_S(\x, [200k, 250k]) =$\texttt{ (total room $\in [2132, 3546],$ total bedrooms $\in [214, 491]$)}. It means if \texttt{total room and total bedrooms} satisfy the conditions in $C_S(\x, [200k, 250k])$ and the remaining features of $\x$ is fixed, then the probability that the price is in $[200k, 250k]$ is 0.97. \paragraph{Comparisons of Local-CR and Regional-CR with baselines (AReS, CET):} We use 3 real-world datasets: \textbf{Diabetes} \citep{diabetes} contains diagnostic measurements and aims to predict whether or not a patient has diabetes, \textbf{Breast Cancer Wisconsin (BCW)} \citep{UCI} consists of predicting if a tumor is benign or not using the characteristic of the cell nuclei, and \textbf{Compas} \citep{compasdata} was used to predict recidivism, and it contains information about the criminal history, demographic attributes. During the evaluation, we observe that \textbf{AReS, CET} are very sensitive to the number of bins and the maximal number of rules or actions as noticed by \citep{globalce}. A bad parameterization gives completely useless explanations. Moreover, a different model needs to be trained for each class to be accurate, while we only need to have a RF that has good precision. In table \ref{tab:results}, we notice that the Local and Regional-CR succeed in changing decisions with a high accuracy in all datasets, outperforming \textbf{AReS} and \textbf{CET} with a large margin on \textbf{BCW}, and \textbf{Diabetes}. Moreover, we notice that the baselines struggle to change at the same time the positive and negative class, (e.g. CET has \textit{Acc}=1 in the positive class, and 0.21 for the negative class on \textbf{BCW}) or when they have a good \textit{Acc}, the CE are not plausible. For instance, CET has \textit{Acc}=0.98 and \textit{Psb}=0 on \textbf{Compas}, meaning that all the CE are outlier. Regarding the coverage of the global CE, CET covers all the instances as it partitions the space, but we observe that \textbf{AReS} has a smaller \textit{Coverage}$=\{0.43, 0.44, 0.81\}$ than the Regional-CR which has $\{1, 0.7, 1\}$ for \textbf{BCW, Diabetes, and Compas} respectively. To sum up, the CR is easier to train and provides more accurate and plausible rules than the baselines methods. \begin{table}[ht!] \caption{Results of the \textit{Correctness} (Acc), \textit{Plausibility}, and \textit{Sparsity} (Sprs) of the different methods. We compute each metric according to the positive (Pos) and negative (Neg) class.} \label{tab:results} \resizebox{\columnwidth}{!}{% \begin{tabular}{ccccccccccccccccccc} \cline{2-19} & \multicolumn{6}{c}{\textbf{COMPAS}} & \multicolumn{6}{c}{\textbf{BCW}} & \multicolumn{6}{c}{\textbf{Diabetes}} \\ \cline{2-19} & \multicolumn{2}{c}{Acc} & \multicolumn{2}{c}{Psb} & \multicolumn{2}{c|}{Sps} & \multicolumn{2}{c}{Acc} & \multicolumn{2}{c}{Psb} & \multicolumn{2}{c|}{Sps} & \multicolumn{2}{c}{Acc} & \multicolumn{2}{c}{Psb} & \multicolumn{2}{c}{Sps} \\ \cline{2-19} & Pos & Neg & Pos & Neg & Pos & \multicolumn{1}{c|}{Neg} & Pos & Neg & Pos & Neg & Pos & \multicolumn{1}{c|}{Neg} & Pos & Neg & Pos & Neg & Pos & Neg \\ \textbf{L-CR} & 1 & 0.9 & 0.87 & 0.73 & 2 & \multicolumn{1}{c|}{4} & 1 & 1 & 0.96 & 1 & 9 & \multicolumn{1}{c|}{7} & 0.97 & 1 & 0.99 & 0.8 & 3 & 4 \\ \textbf{R-CR} & 0.9 & 0.98 & 0.74 & 0.93 & 2 & \multicolumn{1}{c|}{3} & 0.89 & 0.9 & 0.94 & 0.93 & 9 & \multicolumn{1}{c|}{9} & 0.99 & 0.99 & 0.9 & 0.87 & 3 & 4 \\ \textbf{AReS} & 0.98 & 1 & 0.8 & 0.61 & 1 & \multicolumn{1}{c|}{1} & 0.63 & 0.34 & 0.83 & 0.80 & 4 & \multicolumn{1}{c|}{3} & 0.73 & 0.60 & 0.77 & 0.86 & 1 & 1 \\ \textbf{CET} & 0.85 & 0.98 & 0.7 & 0 & 2 & \multicolumn{1}{c|}{2} & 1 & 0.21 & 0.6 & 0.80 & 8 & \multicolumn{1}{c|}{2} & 0.84 & 1 & 0.60 & 0.20 & 6 & 6 \end{tabular}% } \end{table} \section{Conclusion} Most current works that generate CE are implicit through an optimization process or a brunch of random samples, thus lacking guarantees. For this reason, we rethink CE as \textit{Counterfactual Rules}. For any individual or sub-population, it gives the simplest policies that change the decision with high probability. Our approach learns robust, plausible, and sparse adversarial regions where the observations should be moved. We make central use of Random Forests, which give consistent estimates of the interest probabilities and naturally give the counterfactual rules we want to extract. In addition, it permits us to deal with regression problems and continuous features. Consequently, our methods are suitable for all datasets where tree-based model performs well (e.g., tabular data). A prospective work is to evaluate the robustness of our methods to noisy human responses, i.e., when the prescribed recourse is not implemented exactly, and to refine the methodology for selecting the threshold probabilities $\pi$ and $\pi_C$. \newpage \section{Regional RF detailed} In this section, we give a simple application of the Regional RF algorithm to better understand how it works. Recall that the regional RF is a generalization of the RF's algorithm to give prediction even when we condition given a region, e.g., to estimate $E(f(\X) \; | \boldsymbol{X}_S \in C_S(\boldsymbol{x}), \boldsymbol{X}_{\bar{S}} = \boldsymbol{x}_{\bar{S}})$ with $C_{S}(\boldsymbol{x}) = \prod_{i=1}^{|S|} [a_i, b_i], a_i, b_i \in \bar{\mathbb{R}}$ a hyperrectangle. The algorithm works as follows: we drop the observations in the initial trees, if a split used variable $i \in \bar{S}$, a fixed value-based condition, we used the classic rules i.e., if $x_i \leq t$, the observations go to the left children, otherwise the right children. However, if a split used variable $i \in S$, regional-based condition, we used the hyperrectangle $C_S(\boldsymbol{x}) = \prod_{i=1}^{|S|} [a_i, b_i]$. The observations are sent to the left children if $b_i \leq t$, right children if $a_i > t$ and if $t \in [a_i, b_i]$ the observations are sent both to the left and right children. To illustrate how it works, we use a two dimensional variables $\X \in \mathbb{R}^2$, a simple decision tree $f$ represented in figure \ref{fig:tree_example}, and want to compute for $\x = [1.5, 1.9],$ $E(f(\X) | \boldsymbol{X}_1 \in [2, \; 3.5], \boldsymbol{X}_{0} = 1.5)$. We assume that $P(X_1 \in [2, \; 3.5] \; | X_0 = 1.5) >0$ and denoted $T_1$ as the set of the values of the splits based on variables $X_1$ of the decision tree. One way of estimating this conditional mean is by using Monte Carlo sampling. Therefore, there are two cases : \begin{figure}[ht!] \centering \includegraphics[scale=0.5]{figures/illustration_neurips.png} \caption{Representation of a simple decision tree (right figure) and its associated partition (left figure). The gray part in the partition corresponds to the region $[2, \; 3.5] \times [1, 2]$} \label{fig:tree_example} \end{figure} \begin{itemize} \item If $\forall t \in T_1,$ $t \leq 2$ or $t > 3$, then all the observations sampled s.t. $\Tilde{X}_i \sim \mathcal{L} (\X \; |\boldsymbol{X}_1 \in [2, \; 3.5], \boldsymbol{X}_{0} = 1.5)$ follow the same path and fall in the same leaf. The Monte Carlo estimator of the decision tree $E(f(\X) | \boldsymbol{X}_1 \in [2, \; 3.5], \boldsymbol{X}_{0} = 1.5)$ is equal to the output of the Regional RF algorithm. \begin{itemize} \item For instance, a special case of the case above is: if $\forall t \in T_1, t \leq 2$, and we sample using $\mathcal{L} (\X \; |\boldsymbol{X}_1 \in [2, \; 3.5], \boldsymbol{X}_{0} = 1.5)$, then all the observations go to the right children when they encounters a node using $X_1$ and fall in the same leaf. \end{itemize} \item If $\exists \; t \in T_1$ and $t \in [2, \; 3.5]$, then the observations sampled s.t. $\Tilde{X}_i \sim \mathcal{L} (\X \; |\boldsymbol{X}_1 \in [2, \; 3.5], \boldsymbol{X}_{0} = 1.5)$ can fall in multiple terminal leaf depending on if their coordinates $x_1$ is lower than $t$. Following our example, if we generate samples using $\mathcal{L} (\X \; |\boldsymbol{X}_1 \in [2, \; 3.5], \boldsymbol{X}_{0} = 1.5)$, the observations will fall in the gray region of figure \ref{fig:tree_example}, and thus can fall in node 4 or 5. Therefore, the true estimate is: \begin{align} & E(f(\X) | \boldsymbol{X}_1 \in [2, \; 3.5], \boldsymbol{X}_{0} = 1.5 ) \nonumber\\ & = p(X_1 \leq 2.9\; | X_0=1.5)*E[f(\X)\;| \X \in L_4] + p(X_1 > 2.9\; | X_0=1.5)*E[f(\X)\; |\X \in L_5] \label{fig:weighted_mean} \end{align} \end{itemize} Concerning the last case $(t \in [2, \; 3.5])$, we need to estimate the different probabilities $p(X_1 \leq 2.9\; | X_0=1.5), p(X_1 > 2.9\; | X_0=1.5)$ to compute $E(f(\X) | \boldsymbol{X}_1 \in [2, \; 3.5], \boldsymbol{X}_{0} = 1.5 )$, but these probabilities are difficult to estimate in practice. However, we argue that we can ignore these splits, and thus do no need to fragment the query region using the leaves of the tree. Indeed, as we are no longer interest in a point estimate but regional (population mean) we do not need to go to the level of the leaves. We propose to ignore the splits of the leaves that divide the query region. For instance, the leaves 4 and 5 split the region $[2, \; 3.5]$ in two cells, by ignoring these splits we estimate the mean of the gray region by taking the average output of the leaves 4 and 5 instead of computing the mean weighted by the probabilities as in Eq. \ref{fig:weighted_mean}. Roughly, it consists to follow the classic rules of a decision tree (if the region is above or below a split) and ignore the splits that are in the query region, i.e., we average the output of all the leaves that are compatible with the condition $\boldsymbol{X}_1 \in [2, \; 3.5], \boldsymbol{X}_{0} = 1.5$. We think that it leads to a better approximation for two reasons. First, we observe that the case where t is in the region and thus divides the query region does not happen often. Moreover, the leaves of the trees are very small in practice, and taking the mean of the observations that fall in the union of leaves that belong to the query region is more reasonable than computing the weighted mean and thus trying to estimate the different probabilities $p(X_1 \leq 2.9\; | X_0=1.5), p(X_1 > 2.9\; | X_0=1.5)$. \section{Additional experiments} In table \ref{tab:add_exp}, we compare the \textit{Correctness} (Acc), \textit{Plausibility} (Psb), and \textit{Sparsity} (Sprs) of the different methods on additonal real-world datasets: FICO \citep{helocdata}, NHANESI \citep{nhanes}. We observe that the L-CR, and R-CR outperform the baseline methods by a large margin on \textit{Correctness} and \textit{Plausibility}. The baseline methods still struggle to change at the same time the positive and negative class. In addition, AReS and CET give better sparsity, but their counterfactual samples are less plausible than the ones generated by the CR. \begin{table}[ht!] \caption{Results of the \textit{Correctness} (Acc), \textit{Plausibility}, and \textit{Sparsity} (Sprs) of the different methods. We compute each metric according to the positive (Pos) and negative (Neg) class.} \label{tab:add_exp} \resizebox{\columnwidth}{!}{% \begin{tabular}{ccccccccccccc} \cline{2-13} & \multicolumn{6}{c}{\textbf{FICO}} & \multicolumn{6}{c}{\textbf{NHANESI}} \\ \cline{2-13} & \multicolumn{2}{c}{Acc} & \multicolumn{2}{c}{Psb} & \multicolumn{2}{c|}{Sps} & \multicolumn{2}{c}{Acc} & \multicolumn{2}{c}{Psb} & \multicolumn{2}{c}{Sps} \\ \cline{2-13} & Pos & Neg & Pos & Neg & Pos & \multicolumn{1}{c|}{Neg} & Pos & Neg & Pos & Neg & Pos & Neg \\ \textbf{L-CR} & 0.98 & 0.94 & 0.98 & 0.99 & 5 & \multicolumn{1}{c|}{5} & 0.99 & 0.98 & 0.98 & 0.97 & 5 & 6 \\ \textbf{R-CR} & 0.90 & 0.94 & 0.98 & 0.99 & 9 & \multicolumn{1}{c|}{8.43} & 0.86 & 0.95 & 0.96 & 0.99 & 7 & 7 \\ \textbf{AReS} & 0.34 & 0.01 & 0.85 & 0.86 & 2 & \multicolumn{1}{c|}{1} & 0.06 & 1 & 0.87 & 0.92 & 1 & 1 \\ \textbf{CET} & 0.76 & 0 & 0.76 & 0.60 & 2 & \multicolumn{1}{c|}{2} & 0 & 0.40 & 0.82 & 0.56 & 0 & 5 \end{tabular}% } \end{table} \section{Simulated annealing to generate counterfactual samples using the Counterfactual Rules} \begin{lstlisting}[language=Python, caption=The simulated annealing algorithm to generate samples that satisfy the condition CR] import numpy as np def generate_candidate(x, S, x_train, C_S, n_samples): """ Generate sample by sampling marginally between the features value of the training observations. Args: x (numpy.ndarray)): 1-D array, an observation S (list): contains the indices of the variables on which to condition x_train (numpy.ndarray)): 2-D array represent the training samples C_S (numpy.ndarray)): 3-D (#variables x 2 x 1) representing the hyper-rectangle on which to condition n_samples (int): number of samples Returns: The generated samples """ x_poss = [x_train[(C_S[i, 0] <= x_train[:, i]) * (x_train[:, i] <= C_S[i, 1]), i] for i in S] x_cand = np.repeat(x.reshape(1, -1), repeats=n_samples, axis=0) for i in range(len(S)): rdm_id = np.random.randint(low=0, high=x_poss[i].shape[0], size=n_samples) x_cand[:, S[i]] = x_poss[i][rdm_id] return x_cand def simulated_annealing(outlier_score, x, S, x_train, C_S, batch, max_iter, temp, max_iter_convergence): """ Generate sample X s.t. X_S \in C_S using simulated annealing and outlier score. Args: outlier_score (lambda functon): outlier_score(X) return a outlier score. If the value are negative, then the observation is an outlier. x (numpy.ndarray)): 1-D array, an observation S (list): contains the indices of the variables on which to condition x_train (numpy.ndarray)): 2-D array represent the training samples C_S (numpy.ndarray)): 3-D (#variables x 2 x 1) representing the hyper-rectangle on which to condition batch (int): number of sample by iteration max_iter (int): number of iteration of the algorithm temp (double): the temperature of the simulated annealing algorithm max_iter_convergence (double): minimun number of iteration to stop the algorithm if it find an in-distribution observation Returns: The generated sample, and its outlier score """ best = generate_candidate(x, S, x_train, C_S, n_samples=1) best_eval = outlier_score(best)[0] curr, curr_eval = best, best_eval it = 0 for i in range(max_iter): x_cand = generate_candidate(curr, S, x_train, C_S, batch) score_candidates = outlier_score(x_cand) candidate_eval = np.max(score_candidates) candidate = x_cand[np.argmax(score_candidates)] if candidate_eval > best_eval: best, best_eval = candidate, candidate_eval it = 0 else: it += 1 # check convergence if best_eval > 0 and it > max_iter_convergence: break diff = candidate_eval - curr_eval t = temp / np.log(float(i + 1)) metropolis = np.exp(-diff / t) if diff > 0 or rand() < metropolis: curr, curr_eval = candidate, candidate_eval return best, best_eval \end{lstlisting} \section{Parameters detailed} In this section, we give the different parameters of each method. For all methods and datasets, we first used a greedy search given a set of parameters. For AReS, we use the following set of parameters: \begin{itemize} \item max rule = $\{4, 6, 8\}$, max rule length $=\{4, 8 \}$, max change num $= \{2, 4, 6\}$, \item minimal support $= 0.05$, discretization bins = $\{ 10, 20\}$, \item $\lambda_{acc} = \lambda_{cov} = \lambda_{cst} = 1$. \end{itemize} For CET, we search in the following set of parameters: \begin{itemize} \item max iterations $ = \{500, 1000\}$, \item max leaf size $= \{ 4, 6, 8, -1\}$, \item $\lambda = 0.01, \gamma = 1 $. \end{itemize} Finally, for the Counterfactual Rules, we used the following parameters: \begin{itemize} \item nb estimators = $\{20, 50 \}$, max depth= $\{8, 10, 12\}$, \item $\pi=0.9$, $\pi_C=0.9$. \end{itemize} We obtained the same optimal parameters for all datasets: \begin{itemize} \item AReS: max rule $= 4$, max rule length$= 4$, max change num $= 4$, minimal support $= 0.05$, discretization bins = $10$, $\lambda_{acc} = \lambda_{cov} = \lambda_{cst} = 1$ \item CET: max iterations $= 1000$, max leaf size $=-1$, $\lambda = 0.01, \gamma = 1 $ \item CR: nb estimators$= 20$, max depth$=10$, $\pi=0.9$, $\pi_C=0.9$ \end{itemize} The code and the results can be found at \url{https://github.com/anoxai/counterfactual_rules}. \newpage

0098b30cc361131ccdd764aabcdc041c67bdd4dc

\section{Introduction} Combinatorial optimization problems on graphs are widespread in operation research, with applications in planning and logistics. Their study is strongly related to algorithm theory and computational complexity theory. The most representative example of such discrete variational problems is the travelling salesperson problem (TSP) \cite{shmoys1985traveling}: given a set of cities and distances between each pair of them, one asks for the shortest route that visits each city exactly once and returns to the origin city (i.e.\ a tour). Like many related combinatorial problems and despite its straightforward formulation, the TSP belongs to the class of NP-hard problems. In practical terms, computing an exact solution becomes computationally intractable as known algorithms perform exponentially many steps in the number of cities. In real-world situations, there is quite often the need to solve many similar instances of a given combinatorial optimization problem. In that case, additional structure, including geometry and randomness, can be exploited. The Euclidean formulation of the TSP, i.e., when cities are points in $\mathbb{R}^d$ and distances are given by the Euclidean distance, is still NP-hard \cite{papadimitriou1977euclidean}, but Karp \cite{karp1977probabilistic} observed that solutions to random instances, i.e., when cities are sampled independently and uniformly, can be efficiently approximated via a partitioning scheme. His proof relies upon the seminal work by Beardwood, Halton and Hammersely \cite{beardwood1959shortest}, where precise asymptotics for optimal costs of a random instance of the problem were first established: given i.i.d.\ points $(X_i)_{i=1}^n$ distributed according to a probability density $\rho$ on $\mathbb{R}^d$, denoting the length $ \mathcal{C}_{\mathsf{TSP}}((X_i)_{i=1}^n)$ of the (random) solution to the TSP cycling through such points satisfies the $\mathbb{P}$-a.s.\ limit \begin{equation}\label{eq:bhh} \lim_{n \to \infty} n^{\frac{1}{d}-1} \mathcal{C}_{ \mathsf{TSP}}((X_i)_{i=1}^n) = \beta_{\operatorname{BHH}} \int_{\mathbb{R}^d} \rho^{1-\frac{1}{d}},\end{equation} where $\beta_{\operatorname{BHH}} = \beta_{\operatorname{BHH}}(d) \in (0, \infty)$ is a constant depending on the dimension $d$ only. The scaling $n^{1-1/d}$ is intuitively explained by the fact that the $n$ cities are connected through paths of typical length $n^{-1/d}$ (as if they were on a regular grid). Building upon these ideas, several authors \cite{papadimitriou1978probabilistic, steele1981subadditive, steele1997probability, yukich2006probability} contributed towards establishing a general theory to obtain limit results of BHH-type, i.e., as in \eqref{eq:bhh}, for a wide class of random Euclidean combinatorial optimization problems. The theory allows also for more general weights than the Euclidean length, including $p$-th powers of the Euclidean distance, a variant often motivated by modelling needs. If $0<p<d$, with a minimal modification of the techniques one obtains BHH-type results as in \eqref{eq:bhh}, with the scaling replaced by $n^{1-p/d}$, the constant $\beta_{\operatorname{BHH}}$ now depending on $p$, $d$ and the specific combinatorial optimization problem, and the integrand $\rho^{1-1/d}$ replaced by $\rho^{1-p/d}$. For $p \ge d$, the situation becomes subtler and \eqref{eq:bhh} is known for the TSP only if $p=d$, see \cite{yukich1995asymptotics} and \cite[Section 4.3]{yukich2006probability}. Despite the wide applicability of this theory, several classical problems such as those formulated over two random sets of points, are not covered and require different mathematical tools. The Euclidean assignment problem, also called bipartite matching, is certainly the most representative among these: given two sets of $n$ points $(x_i)_{i=1}^n$, $(y_j)_{j=1}^n \subseteq \mathbb{R}^d$, one defines the matching cost functional as \[ \mathsf{M}^p\bra{ (x_i)_{i=1}^n, (y_j)_{j=1}^n } = \min_{\sigma} \sum_{i=1}^n |x_i - y_{\sigma(i)}|^p,\] where the minimum is taken among all the permutations $\sigma$ over $n$ elements. This is often interpreted in terms of optimal planning for the execution of a set of jobs at positions $y_j$'s to be assigned to a set of workers at the positions $x_i$'s. Although the assignment problem belongs to the P complexity class, i.e., an optimal $\sigma$ can be found in a polynomial number of steps (with respect to $n$) the analysis of random instances shows some interesting behavior in low dimensions. Indeed, if $(X_i)_{i=1}^n$, $(Y_j)_{j=1}^n$ are i.i.d.\ and uniformly distributed on the cube $(0,1)^d$, it is known \cite{dudley1969speed, AKT84, talagrand1992matching, dobric1995asymptotics} that \footnote{ The notation $A\lesssim B$ means that there exists a constant $C>0$, such that $A\le C B$, where $C$ depends on the dimension $d$, $p$ and possibly other quantities tacitly considered as fixed, e.g.\ a domain $\Omega \subseteq \mathbb{R}^d$ or a probability density $\rho$. We use the notation $\lesssim_q$ to indicate the dependence on the parameter $q$. We write $A\sim B$ if both $A\lesssim B$ and $B\lesssim A$.} \begin{equation*}\label{eq:boundE} \mathbb{E}\lt[ \mathsf{M}^1( (X_i)_{i=1}^n, (Y_j)_{j=1}^n ) \rt]\sim\begin{cases} \sqrt{n} &\textrm{ for } d=1\\ \sqrt{ n \log n} & \textrm{ for } d=2\\ n^{1 -\frac{1}{d}} & \textrm{ for } d\ge 3. \end{cases} \end{equation*} In particular, for $d\in\cur{1,2}$ the cost is asymptotically larger than the heuristically motivated $n^{1-1/d}$. This exceptional scaling is intuitively due to local fluctuations of the distributions of the two families of points. Inspired by the combinatorial approach in \cite{BoutMar} for the random Euclidean bipartite matching problem in dimension $d\ge 3$, Barthe and Bordenave \cite{BaBo} first proposed a general theory to establish results of BHH-type \eqref{eq:bhh} for a wide class of random Euclidean combinatorial optimization problems over two sets of $n$ points. Let us point out that the equality in \eqref{eq:bhh} is actually only proven for uniform measures while in general only upper and lower bounds (which are conjectured to coincide) are known. In case of $p$-th power weighted distances, the theory developed in \cite{BaBo} applies in the range $0<p<d/2$, which appears quite naturally in their arguments. The difficulty to go beyond the threshold $p=d/2$ is that \eqref{eq:bhh} cannot hold without additional hypothesis on the density $\rho$. For example, because of fluctuations a necessary condition is connectedness of the support of $\rho$. Nevertheless, in the case of the Euclidean bipartite matching problem, it was recently proved \cite{goldman2021convergence} that if $\rho$ is the uniform measure on the unit cube with $d \ge 3$ and $p \ge 1$, then \begin{equation}\label{eq:matching-bhh-cube} \lim_{n \to \infty}n^{\frac{p}{d}-1} \mathbb{E}\lt[ \mathsf{M}^p( (X_i)_{i=1}^n, (Y_j)_{j=1}^n ) \rt] = \beta_{\mathsf{M}}.\end{equation} Here $\beta_{\mathsf{M}} \in (0, \infty)$ depends on $d$ and $p$ only. The proof is a combination of classical subadditivity arguments -- that originate from \cite{beardwood1959shortest} -- and tools from the theory of optimal transport. In particular, the defect in subadditivity is estimated using the connection between Wasserstein distances and negative Sobolev norms. In this context, the use of this type of estimates can be traced back to a recent PDE ansatz proposed in statistical physics \cite{CaLuPaSi14}. Since then, it has been successfully used in the mathematical literature \cite{AmStTr16,Le17,holden2018gravitational, goldman2021quantitative, BobLe19, ledoux2019optimal, goldman2022fluctuation, huesmann2021there, chen2022asymptotics}, even beyond the case of i.i.d.\ points \cite{wang2019limit, jalowy2021wasserstein, huesmann2022wasserstein, borda2021berry}. We refer to \cite{caracciolo2015scaling, BeCa, benedetto2021random} for further statistical physics literature. In fact, the technique in \cite{goldman2021convergence} is quite robust and coarser estimates can be used, avoiding the use of PDEs. Still, the results apply only for the Euclidean bipartite matching problem thanks to its connection with optimal transport. The main purpose of this paper is to show that for a quite general class of bipartite combinatorial problems it is actually possible to rely on the good bounds for the matching problem to obtain the analog of \eqref{eq:bhh} provided $p<d$. This is inspired by \cite{capelli2018exact} where a similar idea is used for the TSP and the $2$-factor problem when $p=d=2$. As alluded to, an important open question left from the theory developed in \cite{BaBo} (see also \cite{DeScSc13}) is the existence of a limit in \eqref{eq:bhh} for general densities. The only result in this direction is \cite{ambrosio2022quadratic}, which established for $p=d=2$ that the limit of the expected cost (suitably renormalized) exists if $\Omega$ is a bounded connected open set, with Lipschitz boundary and $\rho$ is H\"older continuous and uniformly strictly positive and bounded from above on $\Omega$. This settled a conjecture from \cite{benedetto2021random} and, more importantly for our purposes, combined subadditivity and PDE arguments with a Whitney-type decomposition to take into account the structure of $\Omega$ and its boundary. While we do not address this question here, some of the ideas from \cite{ambrosio2022quadratic} are further developed in this work. \subsection{Main result} Our aim is to establish limit results for the cost of a wide class of Euclidean combinatorial optimization problems of two random point sets, in the range $d/2 \le p <d$ for any dimension $d \ge 3$. This overcomes the limitations of \cite{BaBo}, showing that in higher dimensions bipartite problems behave much more similarly to non-bipartite ones. Our general theorem can be stated as follows (a precise description of all the assumptions and notation is given in \cref{sec:notation}). \begin{theorem}\label{thm:main} Let $d \ge 3$, $p\in [1,d)$ and let $\mathsf{P} = (\mathcal{F}_{n,n})_{n \in \mathbb{N}}$ be a combinatorial optimization problem over complete bipartite graphs such that assumptions \ref{as:isomorphism}, \ref{as:spanning}, \ref{as:bddegree}, \ref{ass:local-merging} and \ref{as:growth} hold and write $\C_{\pP}^p( (x_i)_{i=1}^n, (y_j)_{j=1}^n)$ for the optimal cost of the problem over the two sets of $n$ points $(x_i)_{i=1}^n$, $(y_j)_{j=1}^n \subseteq \mathbb{R}^d$, with respect to the Euclidean distance raised to the power $p$. Then, there exists $\beta_{\mathsf{P}}\in (0, \infty)$ depending on $p$, $d$ and $\mathsf{P}$ only such that the following hold. Let $\Omega \subseteq \mathbb{R}^d$ be a bounded open set and assume that it is either convex or has $C^2$ boundary. Let $\rho$ be a H\"older continuous probability density on $\Omega$, uniformly strictly positive and bounded from above. Given i.i.d.\ random variables $(X_i)_{i=1}^\infty$, $(Y_j)_{j=1}^\infty$ with common law $\rho$ we have $\mathbb{P}$-a.s.\ that \begin{equation}\label{eq:limit-mean-main} \limsup_{n \to \infty} n^{\frac{p}{d}-1} \C_{\pP}^p\bra{ (X_i)_{i=1}^n, (Y_j)_{j=1}^n} \le \beta_{\mathsf{P}} \int_{\Omega} \rho^{1-\frac{p}{d}}.\end{equation} Moreover, if $\rho$ is the uniform density and $\Omega$ is either a cube or has $C^2$ boundary, then the above is a $\mathbb{P}$-a.s.\ limit and equality holds. \end{theorem} Our assumptions \ref{as:isomorphism}, \ref{as:spanning}, \ref{as:bddegree}, \ref{ass:local-merging} and in particular \ref{as:growth} are slightly stronger than those introduced in \cite[Section 5.3]{BaBo}, but it is not difficult to show that all the specific examples discussed in \cite{BaBo} satisfy them. In particular, our result apply to the TSP, the minimum weight connected $k$-factor problem and the $k$-bounded degree minimum spanning tree. It is thus fair to say that for compactly supported densities, \cref{thm:main} extends the main results in \cite{BaBo}. \begin{remark} Let us point out that \eqref{eq:limit-mean-main} also holds in expectation (see \cref{prop:limit-mean-iid}). \end{remark} \begin{remark} Arguing as in \cite{BaBo} (see also \cite{ambrosio2022quadratic}) and considering a ``boundary'' variant of $\mathsf{P}$ it should be possible to adapt the proof of \cref{thm:main} to show that there exists $\beta_{\mathsf{P}}^b>0$ such that \[ \beta_{\mathsf{P}}^b \int_{\Omega} \rho^{1-\frac{p}{d}}\le \liminf_{n \to \infty} n^{\frac{p}{d}-1} \C_{\pP}^p\bra{ (X_i)_{i=1}^n, (Y_j)_{j=1}^n}. \] However since we are currently not able to prove that $\beta_{\mathsf{P}}^b=\beta_{\mathsf{P}}$ we decided to leave it aside. \end{remark} \begin{remark} In fact our result applies, at least in expectation, to any $p-$homogeneous bi-partite functional $\mathbf{C}$ satisfying the subadditivity inequality \eqref{Sp} (which is similar to the condition $(\mathcal{S}_p)$ from \cite{BaBo}) and the growth condition \eqref{Rp} (somewhat reminiscent of condition $(\mathcal{R}_p)$ from \cite{BaBo}). See \cref{rem:phom}. \end{remark} Of course, our result applies in particular for the Euclidean assignment problem. \begin{corollary}\label{cor:matching} For $d \ge 3$, $p\in [1,d)$, let $\Omega \subseteq \mathbb{R}^d$ be a cube or a bounded connected open set with $C^2$ boundary and let $\rho$ be a H\"older continuous probability density on $\Omega$, uniformly strictly positive and bounded from above. Then, given i.i.d.\ $(X_i)_{i=1}^\infty$, $(Y_j)_{j=1}^\infty$ with common law $\rho$, we have $\mathbb{P}$-a.s.\ that \[ \limsup_{n \to \infty}n^{\frac{p}{d}-1} \mathsf{M}^p\bra{ (X_i)_{i=1}^n, (Y_j)_{j=1}^n} \le \beta_{\mathsf{M}} \int_{\Omega} \rho^{1-\frac{p}{d}},\] with $\beta_{\mathsf{M}}$ as in \eqref{eq:matching-bhh-cube}. Moreover, if $\rho$ is the uniform density and $\Omega$ has $C^2$ boundary, then the above is a $\mathbb{P}$-a.s.\ limit and equality holds. \end{corollary} \begin{remark} In the case of the matching problem, combining ideas from this paper and \cite{goldman2021convergence} the conclusion of \cref{cor:matching} could be extended to every $p\ge 1$ (at least in expectation). \end{remark} \subsection{Comments on the proof technique} Our proof leverages on the techniques developed for the bipartite matching problem, in particular \cite{goldman2021convergence, ambrosio2022quadratic} to carefully estimate the defects in a geometric subadditivity argument. Comparing the approach in \cite{BaBo}, which works if $p<d/2$, with that in \cite{goldman2021convergence}, which holds instead for any $p$, a crucial difference is that the errors due to local oscillations in the two distributions of points are mitigated in the latter by spreading them evenly across all the points. This is possible since the optimal transport relaxation allows for general couplings as well as continuous densities, rather than discrete matchings only. The overall strategy is thus to find a suitable replacement for such operation in the purely combinatorial setting. The starting point is \cref{prop:partition} where we prove a subadditivity inequality. The problem is then to estimate the defect in subadditivity. This is achieved by combining the following three key observations. The first one is to bound from above the cost of the problem over any two point sets $(x_i)_{i=1}^n$, $(y_j)_{j=1}^n$ by the sum of a term of order $n^{1-p/d}$ plus the bipartite matching cost between the two point sets. This is stated as an assumption (\ref{as:growth}), but can be easily checked on many specific problems (\cref{lem:capelli}): being an upper bound, it usually suffices to combine an optimal matching with the solution to an additional non-bipartite combinatorial optimization problem, such as the TSP, to build a feasible solution. This approach was first successfully used in \cite{capelli2018exact} (see also \cite{ambrosio2022quadratic}) for the random bipartite TSP in the case $p=d=2$, where one can simply argue that the main contribution comes from the logarithmic corrections in the matching cost. The second key observation is that for point sets mostly made of i.i.d.\ points (while much less is assumed on the remaining ones), it is still possible to obtain good bounds for the matching cost. We refer to \cref{sec:ot} for the precise statements, but the underlying idea is strongly related to bounds for the optimal transport cost in terms of the negative Sobolev norms -- thus relying again on the PDE ansatz originally introduced in the statistical physics literature. The third observation is that, in order to ensure that a small fraction of i.i.d.\ uniformly distributed points can indeed be found in the subadditivity defect terms, it is enough to keep them out of the optimization procedure on the smaller scales. As usual with those arguments, the proof of existence of the limit is performed first on the Poisson version of the random problem, so to retain a fraction of points we perform a thinning procedure. Besides these main ideas, plenty of technical modifications with respect to the arguments in \cite{BaBo} and \cite{goldman2021convergence, ambrosio2022quadratic} are required, e.g.\ in order to establish improved subadditivity inequalities (\cref{prop:partition}) and to extend the Whitney-type decomposition argument from \cite{ambrosio2022quadratic} to $p\neq 2$. \subsection{Further questions and conjectures} Our results raise several questions about costs and properties of solutions to Euclidean random combinatorial optimization problems over two point sets. We list here a few which we believe are worth exploring. \begin{enumerate}[label=\emph{\arabic*}.] \item Existence of a limit in \eqref{eq:limit-mean-main} for non-uniform densities is rather easy to conjecture, but so far our techniques do not improve upon \cite{BaBo}, hence the problem remains largely open. \item Our techniques break down if $p \ge d$, but it is natural to conjecture that \cref{thm:main} should hold also in that range. In fact, the correct rate $n^{1-p/d}$ could follow directly from (\ref{as:growth}) combined with the corresponding result for the matching problem. \item In this work we considered only the case of compactly supported densities $\rho$. It would be interesting to investigate the case where the support is $\mathbb{R}^d$. To the best of our knowledge, the only results available so far in this direction are \cite{Le17,ledoux2019optimal} where the correct rates are established for the Gaussian density in the case of the matching problem. \item The assumptions in \cite{BaBo} are slightly different than ours, although the specific problems considered therein satisfy both. It would be interesting to find examples which satisfy only one set of these, or possibly simplify even more our assumptions. \item Many problems, such as the bounded degree minimum spanning tree, but also the bipartite matching problem itself, can be naturally formulated also for two families of points with different number of elements: it could be of interest to investigate limit results also in those cases. \item The cases $d \in \cur{1,2}$ are necessarily excluded by our analysis, since subadditivity arguments do not apply already for the random bipartite matching problem. It is however already an open question, whether the additional logarithmic correction indeed appears in the asymptotic rates for many other problems. As an example, we mention that for the Euclidean minimum spanning tree over two random point sets (without any uniform bound on the degree) no logarithmic corrections appear \cite{correddu2021minimum}, but the maximum degree is unbounded, hence it is not covered by our results. \item In the deterministic literature, for the TSP and other NP-hard Euclidean combinatorial optimization problems, polynomial time approximation schemes are known \cite{arora2003approximation} for any (fixed) dimension $d$, as the number of points grows. Can our approach lead to similar schemes for problems on two families of points, possibly under some mild regularity assumption on their spatial distributions? \end{enumerate} \subsection{Structure of the paper} In \cref{sec:notation} we first introduce some general notation. We then discuss Whitney-type decompositions, Sobolev spaces as well as recall useful known facts on the Optimal Transport problem, and possibly some novel ones (\cref{prop:density-helps}). We close the section with a variant of the standard subadditivity (Fekete-type) arguments, suited for our purposes together with some simple concentration inequalities. \cref{sec:cop} is devoted to the combinatorial optimization problems we consider, discussing in particular the main assumptions that we require and some useful consequences. In \cref{sec:poisson} we establish a variant our main result in the case of Poisson point processes and in \cref{sec:main} we use it to deduce \cref{thm:main}. These two sections in fact rely upon the novel bounds for the Euclidean assignment problem that we finally establish in \cref{sec:ot}. \section{Notation and preliminary results}\label{sec:notation} \subsection{General notation} Given $n \in \mathbb{N}$, we write $[n] = \cur{1, \ldots, n}$ and $[n]_1 = \cur{ (1,i)}_{i=1}^n$, $[n]_2 = \cur{(2,i)}_{i=1}^n$, which easily allows to define two disjoint copies of $[n]$. Given a finite set $A$, we write $|A|$ for the number of its elements, while, if $A \subseteq \mathbb{R}^d$ is infinite, $|A|$ denotes its Lebesgue measure. Given a metric space $(\Omega, \mathsf{d})$, $x \in \Omega$, $A\subseteq \Omega$, we write $\mathsf{d}(x, A) = \min_{y \in A}\cur{ \mathsf{d}(x,y)}$ and $\operatorname{diam}(A) = \sup_{x,y \in A} \mathsf{d}(x,y)$. We endow every set $\Omega \subseteq \mathbb{R}^d$ with the Euclidean distance. A partition $\cur{\Omega_k}_{k=1}^K$ of a set $\Omega$ is always intended up to a set of Lebesgue measure zero. A rectangle $R \subseteq \mathbb{R}^d$ is a subset of the form $R = \prod_{i=1}^d (x_i,x_i+L_i)$, and is said to be of moderate aspect ratio if for every $i,j$, $L_i/L_j\le 2$. If $L_i= L$ for every $i$, then $R = Q$ is a cube of side length $L$. We write $Q_L = (0,L)^d$. We write $I_\Omega$ for the indicator function of a set $\Omega$. \subsection{Families of points} Given a set $\Omega$, we consider finite ordered families of points ${\bf x} = \bra{x_i}_{i=1}^n \subseteq \Omega$, with $n \in \mathbb{N}$, letting ${\bf x} = \emptyset$ if $n=0$. For many purposes the order will not be relevant, but we thus may allow e.g.\ for repetitions (which will be probabilistically negligible anyway). Given a family ${\bf x} \subseteq \mathbb{R}^d$, we write $\mu^{{\bf x}} = \sum_{i=1}^n \delta_{x_i}$ for the associated empirical measure and, for every (Borel) $\Omega \subseteq \mathbb{R}^d$, we let ${\bf x}(\Omega) = \mu^{{\bf x}}(\Omega)$. In the special case $\Omega = \mathbb{R}^d$, we simply write $|{\bf x}| = {\bf x}(\mathbb{R}^d) = \mu^{{\bf x}}(\mathbb{R}^d)$ for the total number of points (counted with multiplicity). We also write ${\bf x}_ \Omega$ for its restriction to $\Omega$, i.e., the family of all points $x_i \in \Omega$, so that ${\bf x} = {\bf x}_{\Omega}$ if ${\bf x} \subseteq \Omega$ (conventionally, we naturally re-index it over $i=1, \ldots, {\bf x}(\Omega)$ with the order inherited from that in ${\bf x}$). Given ${\bf x} = \bra{x_i}_{i=1}^n$, ${\bf y}= \bra{y_j}_{j=1}^m\subseteq \mathbb{R}^d$, their union is ${\bf x} \cup {\bf y} = (x_1, \ldots, x_n, y_1, \ldots, y_m)$. Strictly speaking, the union should be called concatenation, since the operation is not commutative, in general. \subsection{Whitney partitions} We recall following partitioning result \cite[Lemma 5.1]{ambrosio2022quadratic}. \begin{lemma}\label{lem:decomp} Let $\Omega\subset \mathbb{R}^d$ be a bounded domain with Lipschitz boundary and let $\mathcal{Q} = \{Q_i\}_i$ be a Whitney partition of $\Omega$. Then, for every $\delta>0$ sufficiently small, letting $\mathcal{Q}_\delta=\{Q_i \ : \ \operatorname{diam}(Q_i) \ge \delta\}$, there exists a finite family $\mathcal{R}_\delta=\{\Omega_j\}_j$ of disjoint open sets such that: \begin{enumerate}[label=(\roman*), series=partition] \item \label{partition-1} $(\Omega_k)_{k=1}^K = \mathcal{Q}_\delta \cup \mathcal{R}_\delta$ is a partition of $\Omega$, \item \label{partition-2}$ |\Omega_k| \sim \operatorname{diam}(\Omega_k)^d$ for every $k=1, \ldots, K$, \item \label{partition-3}if $\Omega_k \in \mathcal{Q}_\delta$, then $\operatorname{diam}(\Omega_k) \sim \mathsf{d}(x, \Omega^c)$ for every $x \in \Omega_k$, \item \label{partition-4}if $\Omega_k \in \mathcal{R}_\delta$, then $\operatorname{diam}(\Omega_k)\sim \delta$ and $\mathsf{d}(x, \Omega^c) \lesssim \delta$, for every $x \in \Omega_k$. \end{enumerate} Here all the implicit constants depend only on the initial partition $\mathcal{Q}$ (and not on $\delta$). \end{lemma} For later use, we collect some useful bounds related to these partitions. \begin{lemma}\label{lem:bound-partition} Let $\Omega\subset \mathbb{R}^d$ be a bounded domain with Lipschitz boundary and let $\mathcal{Q} = \{Q_i\}_i$ be a Whitney partition of $\Omega$. Then, for every $\delta>0$ sufficiently small, letting $(\Omega_k)_{k=1}^K = \mathcal{Q}_\delta \cup \mathcal{R}_\delta$ as in \cref{lem:decomp}, one has that $|\mathcal{R}_\delta| \lesssim \delta^{1-d}$ and the following holds: \begin{enumerate} \item For every $\alpha \in \mathbb{R}$, \begin{equation}\label{eq:whitney-general-q} \sum_{k=1}^K \operatorname{diam}(\Omega_k)^{\alpha} \lesssim_\alpha \begin{cases} 1 & \text{if $\alpha>d-1$,}\\ |\log \delta|& \text{if $\alpha=d-1$,}\\ \delta^{1-d-\alpha} & \text{if $\alpha<d-1$.} \end{cases} \end{equation} \item For every $k =1,\ldots, K$, and $x \in \Omega_k$, \begin{equation}\label{eq:before-last-claim} \sum_{j\, : \, \mathsf{d}(x, \Omega_j) \le \operatorname{diam}(\Omega_k)} |\Omega_j|^{-\frac{1}{2}} \lesssim \delta^{1-\frac{d}{2}} \operatorname{diam}(\Omega_k)^{d-1}, \end{equation} and \begin{equation}\label{lastclaim} \sum_{j \, : \, \mathsf{d}(x,\Omega_j)>\operatorname{diam}(\Omega_k)} \frac{|\Omega_j|^{\frac{1}{2}}}{\mathsf{d}(x, \Omega_j)^{d-1}} \lesssim \delta^{1-\frac{d}{2}} |\log \bra{\operatorname{diam}(\Omega_k)}|, \end{equation} \end{enumerate} In all the inequalities the implicit constants depend upon $\mathcal{Q}$ (and $\alpha$ in \eqref{eq:whitney-general-q}) only. \end{lemma} By property \emph{\ref{partition-2}}, inequality \eqref{eq:whitney-general-q} also holds for the sum $\sum_{k=1}^K |\Omega_k|^{\alpha}$, with $\alpha d$ instead of $\alpha$. \begin{proof} Since $\partial \Omega$ is Lipschitz, it follows from properties \ref{partition-1}, \ref{partition-2} and \ref{partition-4} that, for every $x \in \Omega$ and $r \ge s \ge \delta$, \begin{equation}\label{eq:uniform-bound-omega-k} \abs{ \cur{k\, : \, \Omega_k\subseteq B(x, r), \operatorname{diam}(\Omega_k) \in [s, 2s)}} \lesssim (r/s)^{d-1},\end{equation} with the implicit constant depending on $\mathcal{Q}$ only. It follows that $|\mathcal{R}_\delta| \lesssim \delta^{1-d}$ and, for every $\ell \le |\log_2 \delta|$, the number of cubes $\Omega_k \in \mathcal{Q}_\delta$ with $\operatorname{diam}(\Omega_k) \in [2^{-\ell}, 2^{-\ell+1})$ is estimated by $ 2^{\ell (d-1)}$. Therefore, for $\alpha \in \mathbb{R}$, \[ \begin{split} \sum_{k=1}^K \operatorname{diam}(\Omega_k)^{\alpha} & \lesssim \sum_{\Omega_k \in \mathcal{Q}_{\delta}} \operatorname{diam}(\Omega_k)^{\alpha} + \sum_{\Omega_k \in \mathcal{R}_\delta} \operatorname{diam}(\Omega_k)^{\alpha} \\ & \lesssim \sum_{\ell \le |\log_2 \delta|} \abs{\cur{ \Omega_k \in \mathcal{Q}_{\delta} : \operatorname{diam}(Q_k) \in [2^{-\ell}, 2^{-\ell+1})}} 2^{-\ell \alpha} + |\mathcal{R}_\delta| \cdot \delta^{\alpha}\\ & \lesssim \sum_{\ell \le |\log_2 \delta|} 2^{\ell(d-1)}\cdot 2^{-\ell \alpha} + \delta^{1-d} \cdot \delta^{\alpha}.\end{split} \] Since $\ell$ is also bounded from below in the summation (e.g.\ by $-|\log_2\operatorname{diam}(\Omega)|$), we obtain \eqref{eq:whitney-general-q}. In order to prove \eqref{eq:before-last-claim} and \eqref{lastclaim} we further notice that, given $\Omega_k$, $\Omega_j$ and $x \in \Omega_k$, we have that, for some constant $C = C(\mathcal{Q})$, \begin{equation}\label{eq:omega-j-contained-ball} \Omega_j \subseteq B(x, C \max\cur{\mathsf{d}(x, \Omega_j), \operatorname{diam}(\Omega_k)}).\end{equation} Indeed, if $\Omega_j \in \mathcal{R}_\delta$, then $\operatorname{diam}(\Omega_j)\lesssim \delta\lesssim \operatorname{diam}(\Omega_k)$, hence \eqref{eq:omega-j-contained-ball} holds. If instead $\Omega_j \in \mathcal{Q}_\delta$, then we can find $y \in \Omega_j$ with $|x-y|\le 2 \mathsf{d}(x,\Omega_j)$, so that, by the triangle inequality, \[ \mathsf{d}(y, \Omega^c) \le |x-y| + \mathsf{d}(x, \Omega^c) \lesssim \max\cur{\mathsf{d}(x, \Omega_j), \operatorname{diam}(\Omega_k)} \] and by property \ref{partition-3} we obtain that $\operatorname{diam}(\Omega_j) \lesssim \max\cur{\mathsf{d}(x, \Omega_j), \operatorname{diam}(\Omega_k)}$, yielding again the desired inclusion. We now prove \eqref{eq:before-last-claim}. Let $\ell_k \le |\log_2 \delta|$ be such that $\operatorname{diam}(\Omega_{k}) \in [2^{-\ell_k}, 2^{-\ell_k+1})$. Combining \eqref{eq:omega-j-contained-ball} and \eqref{eq:uniform-bound-omega-k}, we see that, for every $\ell \le |\log_2\delta|$, there are at most $2^{(\ell-\ell_k)(d-1)}$ sets $\Omega_j$ such that $\mathsf{d}(x, \Omega_j) \le \operatorname{diam}(\Omega_k)$ and $\operatorname{diam}(\Omega_j) \in [2^{-\ell}, 2^{-\ell+1})$. Therefore, \[ \begin{split} \sum_{j\, : \, \mathsf{d}(x, \Omega_j) \le \operatorname{diam}(\Omega_k)} |\Omega_j|^{-\frac{1}{2}} & \lesssim \sum_{\ell \le |\log_2\delta|} 2^{-\ell \frac{d}{2}} 2^{(\ell-\ell_k)(d-1)} \\ & \lesssim 2^{-\ell_k(d-1)} \sum_{\ell \le |\log_2\delta|} 2^{\ell(\frac{d}{2}-1)} \lesssim \delta^{\frac{d}{2}-1} \operatorname{diam}(\Omega_k)^{d-1}, \end{split}\] recalling that $\ell$ is bounded from below by a constant depending on $\mathcal{Q}$ only. This proves \eqref{eq:before-last-claim}. To prove \eqref{lastclaim}, we split dyadically, \begin{equation}\label{eq:final-claim-step}\begin{split} \sum_{j \, : \, \mathsf{d}(x,\Omega_j)>\operatorname{diam}(\Omega_k)} \frac{|\Omega_j|^{\frac{1}{2}}}{d(x,\Omega_j)^{d-1}} & \lesssim \sum_{\ell \le \ell_k} \frac{1}{(2^{-\ell})^{d-1}} \sum_{ j\, : \, d(x,\Omega_j) \in [2^{-\ell}, 2^{-\ell+1})} |\Omega_j|^{\frac{1}{2}} \\ & \stackrel{\eqref{eq:omega-j-contained-ball}}{\lesssim} \sum_{\ell \le \ell_k} 2^{\ell(d-1)} \sum_{ \Omega_j \subset B(x,C 2^{-\ell})} |\Omega_j|^{\frac{1}{2}}. \end{split}\end{equation} Let us also notice that, if $\Omega_j \subseteq B(x, C2^{-\ell})$, then necessarily $\delta \le \operatorname{diam}(\Omega_j) \lesssim 2^{-\ell}$ (since $\operatorname{diam}(\Omega_j)^d \sim |\Omega_j|$). Thus for $\ell'$ with $2^{-\ell'} \sim 2^{-\ell}$, \[\begin{split} \sum_{ \Omega_j \subset B(x, C 2^{-\ell})} |\Omega_j|^{\frac{1}{2}} & \lesssim \sum_{\ell' \le u \le |\log_2 \delta|} 2^{-u\frac{d}{2}} \abs{\cur{ \Omega_j \subseteq B(x, C 2^{-\ell}) \, : \operatorname{diam}(\Omega_j) \in [2^{-u}, 2^{-u+1}) }}\\ & \stackrel{\eqref{eq:uniform-bound-omega-k}}{\lesssim} \sum_{\ell' \le u \le |\log_2 \delta|} 2^{-u\frac{d}{2}} \cdot 2^{(u-\ell) (d-1) }= 2^{-\ell(d-1)} \sum_{\ell' \le u \le |\log_2 \delta|} 2^{u(\frac{d}{2}-1)}\\ & \lesssim 2^{-\ell(d-1)} \delta^{1-\frac{d}{2}}. \end{split}\] In the last line we used once again that $\ell'$ is bounded from below. Using this bound in \eqref{eq:final-claim-step}, we conclude that \[ \sum_{j \, : \, \mathsf{d}(x,\Omega_j)>\operatorname{diam}(\Omega_k)} \frac{|\Omega_j|^{\frac{1}{2}}}{d(x,\Omega_j)^{d-1}} \le \sum_{\ell \le \ell_k} 2^{\ell(d-1)} \cdot 2^{-\ell(d-1)} \delta^{1-\frac{d}{2}} \lesssim \delta^{1-\frac{d}{2}} |\log \bra{ \operatorname{diam}(\Omega_k)}|, \] where in the last line we used again that $\ell$ is bounded from below. This concludes the proof of \eqref{lastclaim}. \end{proof} \subsection{Sobolev norms} Given a bounded domain $\Omega \subseteq \mathbb{R}^d$ with Lipschitz boundary and $p \in (1,\infty)$, with H\"older conjugate $q = p/(p-1)$, we write $\| f \|_{L^p(\Omega)}$ for the Lebesgue norm of $f$, and \[ \|f\|_{W^{-1,p}(\Omega)}=\sup_{|\nabla \phi|_{L^{q}(\Omega)}\le 1} \int_{\Omega} f \phi=\inf_{\textrm{div} \xi=f} \|\xi\|_{L^p(\Omega)} \] for the negative Sobolev norm. We notice in particular that if $\|f\|_{W^{-1,p}(\Omega)}<\infty$ then $\int_\Omega f=0$. In this case we may also restrict the supremum to functions $\phi$ having also average zero. When it is clear from the context, we will drop the explicit dependence on $\Omega$ in the norms. Let us recall that we can bound the $W^{-1,p}$ norm by the $L^p$ norm. We give here a proof based on the embedding $L^{pd/(p+d)}\subset W^{-1,p}$ (for $p>d/(d-1)$) which is an elementary alternative to the PDE arguments used in \cite[Lemma 3.4]{goldman2021convergence}. \begin{lemma} Let $\Omega$ be a bounded domain with Lipschitz boundary and let $f:\Omega\to \mathbb{R}$ such that $\int_\Omega f=0$. Then, for every $p>d/(d-1)$, \begin{equation}\label{eq:Lp} \|f\|_{W^{-1,p}(\Omega)}\lesssim |\Omega|^{\frac{1}{d}} \|f\|_{L^p(\Omega)}. \end{equation} Moreover, the implicit constant depends on $\Omega$ only through the corresponding constant for the Sobolev embedding. \end{lemma} \begin{proof} Let $q$ be the H\"older conjugate of $p$, $q^*$ the Sobolev conjugate of $q$ and $p^*=pd/(p+d)$ the H\"older conjugate of $q^*$. We then have for every $\phi$ with $\|\nabla \phi\|_{L^q(\Omega)}\le 1$, \[ \int_{\Omega} f \phi\le \lt(\int_{\Omega} |f|^{p^*}\rt)^\frac{1}{p^*} \lt(\int_{\Omega} |\phi|^{q^*}\rt)^{\frac{1}{q^*}}\lesssim \lt(\int_{\Omega} |f|^{p^*}\rt)^\frac{1}{p^*} \lt(\int_{\Omega} |\nabla \phi|^{q}\rt)^{\frac{1}{q}}\le \lt(\int_{\Omega} |f|^{p^*}\rt)^\frac{1}{p^*}. \] Using that $p^*<p$ and H\"older inequality concludes the proof of \eqref{eq:Lp}. \end{proof} As in \cite{ambrosio2022quadratic}, \eqref{eq:Lp} will however not be precise enough when estimating the error in subadditivity in the case of general densities and domains. We will instead rely on gradient bounds for the Green kernel $(G(x,y))_{x,y\in \Omega}$ of the Laplacian with Neumann boundary conditions to obtain sharper estimates. See \cite{AmStTr16,AmGlau,Le17,goldman2022fluctuation} for related results. Let us however point out that in our case we will not rely on any stochastic cancellation in the form of Rosenthal inequality \cite{rosenthal1970subspaces} but will instead use a purely deterministic estimate. We will assume that \begin{equation}\label{eq:green-kernel-bound} \abs{ \nabla _x G(x,y) } \lesssim |x-y|^{1-d}, \quad \text{for every $x$, $y \in \Omega$,}\end{equation} where the implicit constant depends uniquely on $\Omega$. \begin{remark}\label{rem:green}This condition is satisfied for instance if $\Omega$ is $C^2$ or convex, see e.g.\ \cite{wang2013gradient}. Notice that since it is a local condition it also holds for $Q\backslash \Omega$ with $\Omega$ a $C^2$ open set with $d(\partial Q,\partial \Omega)>0$. \end{remark} \begin{remark}\label{eq:Riesz} Let us point out that as in \cite{Le17}, instead of \eqref{eq:green-kernel-bound} it would have been enough to have $L^p$ bounds (for the same $p$ as for the cost $\C_{\pP}^p$) on the Riesz transform for the Neumann Laplacian. From the available results for the Dirichlet Laplacian \cite{JerKenInhom,Shen}, we expect that for every Lipschitz domain there is $p>3$ (depending on the domain) for which these bounds hold. In particular, this would allow to extend the validity of Theorem \ref{thm:main} to every Lipschitz domain when $d=3$. However, since we were not able to find in the literature the corresponding results for the case of Neumann boundary conditions we kept the stronger hypothesis \eqref{eq:green-kernel-bound}. \end{remark} We then have \begin{lemma}\label{lem:W1pWhitney} Let $\Omega\subset \mathbb{R}^d$ be a bounded domain with Lipschitz boundary, such that \eqref{eq:green-kernel-bound} holds and let $\rho$ be a density bounded above and below on $\Omega$. For $\delta>0$ sufficiently small, let $(\Omega_k)_{k=1}^K = \mathcal{Q}_\delta \cup \mathcal{R}_\delta$ as in \cref{lem:decomp}. If there exists $h>0$ such that $|b_k|\le h^{\frac{1}{2}} |\Omega_k|^{\frac{1}{2}}$ for $k=1,\cdots, K$, then for every $p\ge 1$, \begin{equation}\label{estimW1pWhitney} \lt\|\sum_{k=1}^K \frac{b_k}{\rho(\Omega_k)}( I_{\Omega_k} -\rho(\Omega_k))\rho\rt\|_{W^{-1,p}(\Omega)}\lesssim \delta^{1-\frac{d}{2}}h^{\frac{1}{2}}. \end{equation} \end{lemma} \begin{proof} Set $$ B_k = \frac{ b_k}{\rho(\Omega_k)}, \qquad f_k= \bra{I_{\Omega_k} - \rho(\Omega_k)} \rho.$$ Let then $\phi_k$ denotes the solution to the equation $\Delta \phi_k = f_k$, with null Neumann boundary conditions on $\Omega$ and use as competitor $\xi=\sum_{k=1}^K B_k \nabla \phi_k$ in the definition of the $W^{-1,p}$ norm. We get, \begin{equation}\label{normW1pA} \nor{ \sum_{k=1}^K B_k f_k}_{W^{-1,p}(\Omega)}^p \le \int_{\Omega} \abs{ \sum_{k=1}^K B_k {\nabla \phi_k}}^p \lesssim h^{\frac{p}{2}} \int_{\Omega} \bra{\sum_{k=1}^K |\Omega_k|^{-\frac{1}{2}} \abs{\nabla \phi_k}}^p. \end{equation} To bound the last term, we use the integral representation in terms of the Green's function, \[ \phi_k = \int_{\Omega} G(x,y) f_k(y) d y, \] to obtain that, for every $x\in \Omega$, \begin{equation}\label{boundgradphii} |\nabla \phi_k(x)|\lesssim \min\cur{\operatorname{diam}(\Omega_k), \frac{|\Omega_k|}{\mathsf{d}(x,\Omega_k)^{d-1}}}. \end{equation} Indeed, by \eqref{eq:green-kernel-bound}, \[\begin{split} |\nabla \phi_k(x)|& \lesssim \int_{\Omega_k} \frac{dy}{|x-y|^{d-1}}+ |\Omega_k| \int_{\Omega} \frac{dy}{|x-y|^{d-1}} \le \int_{\cur{|y|\le \operatorname{diam}(\Omega_k)}} \frac{dy}{|y|^{d-1}} +|\Omega_k|\\ & \lesssim \operatorname{diam}(\Omega_k). \end{split}\] Moreover, for $x\notin \Omega_k$, we get directly from \eqref{eq:green-kernel-bound}, \[ |\nabla \phi_k(x)|\lesssim \frac{|\Omega_k|}{\mathsf{d}(x,\Omega_k)^{d-1}}. \] For any $k=1, \ldots, K$ and $x \in \Omega_k$, we then estimate \[ \begin{split} \sum_{j=1}^K |\Omega_j|^{-\frac{1}{2}} \abs{\nabla \phi_j(x)} \stackrel{\eqref{boundgradphii}}{\lesssim} & \sum_{j \, : \, \mathsf{d}(x, \Omega_j) \le \operatorname{diam}(\Omega_k)} |\Omega_j|^{-\frac{1}{2}} \operatorname{diam}(\Omega_k) \\ & + \sum_{j \, : \, \mathsf{d}(x, \Omega_j) > \operatorname{diam}(\Omega_k)} \frac{ |\Omega_j|^{-\frac{1}{2}}}{\mathsf{d}(x, \Omega_j)^{d-1}}\\ & \lesssim \delta^{1-\frac{d}{2}}\bra{\operatorname{diam}(\Omega_k)^{d} + \log(\operatorname{diam}(\Omega_k))}, \end{split}\] having used inequalities \eqref{eq:before-last-claim} and \eqref{lastclaim} from \cref{lem:bound-partition}. Therefore, we can split the integration \[\begin{split} \int_{\Omega} \bra{\sum_{j=1}^K |\Omega_j|^{-\frac{1}{2}} \abs{\nabla \phi_j}}^p & = \sum_{k=1}^K \int_{\Omega_k} \bra{\sum_{j=1}^K |\Omega_j|^{-\frac{1}{2}} \abs{\nabla \phi_j}}^p\\ & \lesssim \delta^{p(1-\frac{d}{2})} \sum_{k=1}^K \bra{|\Omega_k|^{p+1} + |\Omega_k| \abs{\log(\operatorname{diam}(\Omega_k))}^p}\\ & \lesssim \delta^{p(1-\frac{d}{2})}, \end{split} \] estimating e.g.\ $\abs{\log(\operatorname{diam}(\Omega_k))}^p \lesssim \operatorname{diam}(\Omega_k)^{-\frac{1}{2}}$ and using \eqref{eq:whitney-general-q} (with $\alpha = (p+1)d$ and $\alpha = d-1/2>d-1$). In combination with \eqref{normW1pA} this concludes the proof of \eqref{estimW1pWhitney}. \end{proof} \subsection{Optimal Transport} Given two positive Borel measures $\mu$, $\lambda$ on $\mathbb{R}^d$ with $\mu(\mathbb{R}^d) = \lambda(\mathbb{R}^d) \in (0, \infty)$ and finite $p$-th moments, the optimal transport cost of order $p\ge 1$ between $\mu$ and $\lambda$ is defined as the quantity \[ \mathsf{W}^p(\mu, \lambda) = \min_{\pi\in\Gamma(\mu,\lambda)} \int_{\mathbb{R}^d\times\mathbb{R}^d} {|x-y|}^p d \pi(x,y),\] where $\Gamma(\mu, \lambda)$ is the set of couplings between $\mu$ and $\lambda$, i.e., finite Borel measures $\pi$ on the product $\mathbb{R}^d\times \mathbb{R}^d$ such that their marginals are respectively $\mu$ and $\lambda$. Notice that if $\mu(\mathbb{R}^d) = \lambda(\mathbb{R}^d) = 0$ then $\mathsf{W}^p(\mu, \lambda) = 0$, while if $\mu(\Omega) \neq \lambda(\Omega)$, we conveniently extend the definition setting $\mathsf{W}^p(\mu, \lambda) = \infty$. Let us recall that the triangle inequality for the Wasserstein distance of order $p$ (which is defined as the the $p$-th root of $\mathsf{W}^p(\mu, \lambda)$) yields \begin{equation}\label{eq:triangle} \mathsf{W}^p( \mu, \nu) \lesssim \mathsf{W}^p(\mu, \lambda) + \mathsf{W}^p(\nu, \lambda), \end{equation} A straightforward, but useful subadditivity inequality is \begin{equation}\label{eq:convexity} \mathsf{W}^p\bra{ \sum_{k} \mu_k, \sum_{k} \nu_k} \le \sum_{k} \mathsf{W}^p(\mu_k, \nu_k). \end{equation} valid for any (countable) family of measures $(\mu_k, \nu_k)_{k}$. To keep notation simple, we write \[ \mathsf{W}^p_{\Omega} (\mu, \lambda) = \mathsf{W}^p (\mu\mathop{\raisebox{-.127ex}{\reflectbox{\rotatebox[origin=br]{-90}{$\lnot$}}}}\Omega, \lambda\mathop{\raisebox{-.127ex}{\reflectbox{\rotatebox[origin=br]{-90}{$\lnot$}}}} \Omega).\] and, if a measure is absolutely continuous with respect to Lebesgue measure, we only write its density. For example, $\mathsf{W}^p_{\Omega} \bra{ \mu, \mu(\Omega)/|\Omega| }$ denotes the transportation cost between $\mu \mathop{\raisebox{-.127ex}{\reflectbox{\rotatebox[origin=br]{-90}{$\lnot$}}}} \Omega$ to the uniform measure on $\Omega$ with total mass $\mu(\Omega)$. For $q \ge p$, Jensen inequality gives \begin{equation}\label{eq:jensen} \mathsf{W}^p_{\Omega} (\mu, \nu) \le \mu(\Omega)^{1-\frac{p}{q}} \bra{ \mathsf{W}_{\Omega}^q (\mu, \nu)}^{\frac{p}{q}}.\end{equation} Our arguments make substantial use of two crucial properties of the optimal transport cost. The first one \cite[Lemma 3.1]{goldman2021convergence} is a simple consequence of \eqref{eq:triangle} and \eqref{eq:convexity}. \begin{lemma} For every $p\ge 1$, there exists a constant $C>0$ depending only on $p$ such that the following holds. Let $\Omega \subseteq \mathbb{R}^d$ be Borel and $(\Omega_k)_{k\in \mathbb{N}}$ be a countable Borel partition of $\Omega$. Then, for finite measures $\mu$, $\lambda$, and $\varepsilon\in(0,1)$, we have the inequality \begin{equation}\label{eq:mainsub} \mathsf{W}^p_{\Omega}\lt(\mu, \alpha \lambda \rt)\le (1+\varepsilon)\sum_i \mathsf{W}^p_{\Omega_k}\lt(\mu, \alpha_k \lambda \rt)+ \frac{C}{\varepsilon^{p-1}}\mathsf{W}^p_\Omega\lt(\sum_k \alpha_k \chi_{\Omega_k} \lambda , \alpha \lambda \rt). \end{equation} where $\alpha = \mu(\Omega)/\lambda(\Omega)$ and $\alpha_k = \mu(\Omega_k)/\lambda(\Omega_k)$. \end{lemma} The second one is \cite[Lemma 2.2]{ambrosio2022quadratic} which gives an upper bound for the Wasserstein distance in terms of a negative Sobolev norm. It follows from the Benamou-Brenier formulation of the optimal transport problem (see also \cite[Corollary 3]{peyre2018comparison}). \begin{lemma}\label{lem:peyre} Assume that $\Omega \subseteq \mathbb{R}^d$ is a bounded connected open set with Lipschitz boundary. If $\mu$ and $\lambda$ are measures on $\Omega$ with $\mu(\Omega) = \lambda(\Omega)$, absolutely continuous with respect to the Lebesgue measure and $\inf_\Omega \lambda>0$, then, for every $p\ge 1$, \begin{equation}\label{eq:estimCZ} W_{\Omega}^p(\mu,\lambda)\lesssim \frac{1}{\inf_{\Omega} \lambda^{p-1}}\nor{ \mu - \lambda}_{W^{-1,p}(\Omega)}^p. \end{equation} \end{lemma} As in many recent works on the matching problem, we will use this inequality to improve on the trivial bound \begin{equation}\label{eq:w-trivial} \mathsf{W}^p_{\Omega} (\mu, \lambda) \le \operatorname{diam}(\Omega)^p \mu(\Omega),\end{equation} which holds as soon as $\mu(\Omega) = \nu(\Omega)$. Much of our effort in the proofs will be ultimately to deal with an intermediate situation, where the measures can be decomposed as the sum of a ``good'' part, i.e., absolutely continuous with smooth density and a ``bad'' remainder about which not much can be assumed. We prove here a general inequality which could also be of independent interest. \begin{proposition}\label{prop:density-helps} Let $\Omega \subseteq \mathbb{R}^d$ be a bounded Lipschitz domain, $\rho$ be a density bounded above and below on $\Omega$, $\mu$ be any finite measure on $\Omega$ and $h >0$. Then, for every $p>d/(d-1)$, \begin{equation}\label{claim} \mathsf{W}^p_\Omega\bra{ \mu +h\rho, \alpha \rho} \lesssim \frac{1}{h^{\frac{p}{d}}} \mu(\Omega)^{1+\frac{p}{d}}, \end{equation} where $\alpha= \frac{\mu(\Omega)}{\rho(\Omega)} +h$. Moreover, this inequality is invariant by rescaling of $\Omega$. \end{proposition} \begin{proof} By scaling we may assume that $|\Omega|=1$. Notice that, by the trivial bound \[ \mathsf{W}^p(\mu +h\rho, \alpha \rho)\lesssim \mu(\Omega), \] we can assume that $\mu(\Omega) \ll h$. Let $P_t$ be the heat semi-group with null Neumann boundary conditions on $\Omega$ and set $\mu_t=P_t\mu$. By triangle inequality \eqref{eq:triangle} and \eqref{eq:estimCZ}, we have \[\begin{split} \mathsf{W}^p(\mu +h\rho, \alpha \rho)&\lesssim \mathsf{W}^p(\mu +h\rho, \mu_t+h\rho)+\mathsf{W}^p(\mu_t+h\rho, \alpha \rho)\\ &\lesssim \mathsf{W}^p\bra{ \mu , \mu_t}+\frac{1}{\eta^{p-1}}\nor{\mu_t- \frac{\mu(\Omega)}{\rho(\Omega)}\rho}^p_{W^{-1,p}}\\ &\lesssim t^\frac{p}{2} \mu(\Omega) +\frac{1}{\eta^{p-1}}\nor{ \mu_t- \frac{\mu(\Omega)}{\rho(\Omega)}\rho}^p_{W^{-1,p}}. \end{split}\] We now estimate the last term. For this let $q$ be the H\"older conjugate exponent of $p$, i.e., $q=p/(p-1) \in (1, d)$ and $q^* = qd/(d-q)$ be the Sobolev conjugate of $q$. We first use triangle inequality and the fact that $\rho$ is bounded from above and below to estimate \[ \begin{split} \nor{ \mu_t- \frac{\mu(\Omega)}{\rho(\Omega)}\rho}_{W^{-1,p}} & \le \nor{ \mu_t- \mu(\Omega)}_{W^{-1,p}} + \mu(\Omega)\nor{ 1- \frac{1}{\rho(\Omega)}\rho}_{W^{-1,p}}\\ & \lesssim \nor{ \mu_t- \mu(\Omega)}_{W^{-1,p}} + \mu(\Omega). \end{split} \] Using that the Sobolev embedding is equivalent to ultra-contractivity i.e. if $\int_\Omega \phi=0$, \[ \|\phi_t\|_{L^\infty(\Omega)}\lesssim t^{-\frac{d}{2q_*}} \|\phi\|_{L^{q^*}(\Omega)}\lesssim t^{-\frac{d}{2q_*}} \|\nabla \phi\|_{L^{q}(\Omega)} , \] we finally estimate for every $\phi$ with $\|\nabla \phi\|_{L^{q}(\Omega)}\le 1$ and $ \int_{\Omega} \phi=0$, \[ \int_{\Omega}\phi(\mu_t-\mu(\Omega))= \int_\Omega \phi_t d\mu\le \mu(\Omega) \|\phi_t\|_{L^\infty(\Omega)}\le \mu(\Omega) t^{-\frac{d}{2q_*}}.\] Therefore, by taking the supremum over $\phi$ we find \[ \nor{ \mu_t- \mu(\Omega)}_{W^{-1,p}}\lesssim \mu(\Omega )t^{-\frac{d}{2q_*}}. \] Taking the $p$-th power we find for $t\le 1$, \[\begin{split} \mathsf{W}^p(\mu +h\rho, \alpha \rho)& \lesssim \mu(\Omega)\lt[t^{\frac{p}{2}} +t^{-\frac{pd}{2q^*}}\lt(\frac{\mu(\Omega)}{h}\rt)^{p-1}\rt]\\ & =\mu(\Omega)\lt[t^{\frac{p}{2}} +t^{-\frac{p}{2}\lt(\frac{d}{q}-1\rt)}\lt(\frac{\mu(\Omega)}{h}\rt)^{p-1}\rt] \end{split}.\] Optimizing in $t$ we find $t^{\frac{p}{2}}=\lt(\frac{\mu(\Omega)}{h}\rt)^{\frac{(p-1)q}{d}}$ which satisfies $t\ll1$ if $\mu(\Omega)\ll h$. Since $(p-1)q=p$, this concludes the proof of \eqref{claim}. \end{proof} \begin{remark} Since by H\"older inequality it will be enough for us to apply \cref{prop:density-helps} for $p$ arbitrarily close to $d$, the condition $p>d/(d-1)$ will not be a limitation for us. Let us however mention that, in the critical case $p=d/(d-1)$ one can argue similarly, relying instead on the Moser-Trudinger inequality \cite[Remark 1.4]{Cianchi}, to obtain (in the case $\rho=1$ and $\Omega=Q$ a cube for simplicity) $$ \mathsf{W}^p_Q\bra{ \mu +h, \frac { \mu(Q)}{|Q|} + h} \lesssim \mu(Q) \lt|\log \lt(\frac{\mu(Q)}{|Q|h}\rt)\rt| \lt(\frac{\mu(Q)}{h}\rt)^{\frac{1}{d-1}}.$$ If instead $p<d/(d-1)$, using the same proof as above but with the inclusion $W^{1,q}(Q) \subseteq L^\infty(Q)$ and letting $t \to 0$ gives the estimate $$ \frac{1}{|Q|^{\frac{p}{d}}}\mathsf{W}^p_Q\bra{ \mu + h, \frac { \mu(Q)}{|Q|} + h} \lesssim \mu(Q)\lt(\frac{ \mu(Q)}{|Q|h}\rt)^{p-1}.$$ \end{remark} We close this section with the following result easily adapted from \cite[Proposition 2.4]{ambrosio2022quadratic} which helps in particular to reduce the transport problem from H\"older to constant densities. \begin{proposition}\label{prop:map-heat-semigroup} For $d\ge 1$, $\alpha \in (0,1)$ and $\rho_0>0$, there exists $C = C(\rho_0, d, \alpha)>0$ such that the following holds: for any $\rho\in C^\alpha( (0,1)^d)$ with \[ \int_{(0,1)^d} \rho = 1 \quad \text{and} \quad \rho_0 \le \rho\le \rho_0^{-1},\] there exists $T: (0,1)^d \to (0,1)^d$ such that $T_{\sharp} \rho = 1$, with \[ \operatorname{Lip} T, \operatorname{Lip} T^{-1} \le 1 + C \nor{ \rho -1}_{C^\alpha}. \ \end{proposition} \subsection{A subadditivity lemma} We will need a slight variant of the usual convergence results for subadditive functions, see e.g.\ \cite{steele1997probability, BoutMar}. \begin{lemma}\label{lem:sub} Let $\alpha, \beta, c>0$, $f: [1, \infty) \to [0, \infty)$ be continuous and such that the following holds: for every $\eta\in (0,1/2]$, there exists $C(\eta)>0$ such that, for every $m \in \mathbb{N}\setminus\cur{0}$ and $L \ge C(\eta)$, \begin{equation}\label{eq:monotonicity} f(m L) \le f( L (1-\eta)) + c \eta^{\alpha} + C(\eta)L^{-\beta}.\end{equation} Then $\lim_{L\to \infty} f(L) \in [0, \infty)$ exists. \end{lemma} \begin{proof} We use the following fact: for any open interval $(a,b) \subseteq [0, \infty)$, the union $$\bigcup_{m=1} ^\infty (ma, mb) \supseteq (A, +\infty)$$ contains a half-line, for some $A>0$. Indeed, one has $(ma, mb) \cap ((m+1)a, (m+1)b)\neq \emptyset$ if $mb > (m+1)a$, which holds for every $m> a/(b-a)$. First, we show that $f$ is uniformly bounded. Let $\eta = 1/2$ and use the fact that both $f(L)$ and $L^{-\beta}$ are continuous for $L \in [1/2,2]$, hence bounded, so that by \eqref{eq:monotonicity}, for every $m \ge 1$, $L \in [1,2]$, \[ f(m L) \le \sup_{\ell \in [1,2]} \bra{f(\ell/2) + c2^{-\alpha} + C(1/2) \ell^{-\beta}} < \infty.\] since $\bigcup_{m= 1}^\infty [m, 2m] = [1, \infty)$, it follows that $f$ is uniformly bounded on $[1, \infty)$. To show that the limit exists (and is finite) we argue that \[ \limsup_{L \to \infty} f(L) \le \liminf_{L \to \infty} f(L).\] Given $\varepsilon\ll 1$, let $\eta = \eta(\varepsilon) \in (0,1/2]$ such that $c\eta^{\alpha}= \varepsilon$ and $L_{\varepsilon}>0$ such that $C(\eta) L_{\varepsilon}^{-\beta}=\varepsilon$, so that, for every $L \ge L_\varepsilon$, $$ C(\eta) L^{-\beta} \le \varepsilon.$$ Let then $L^* > \max \cur{ L_\varepsilon, C(\eta)}$ be such that \[ f(L^*) < \liminf_{L \to \infty} f(L) + \varepsilon.\] By continuity of $f$, there exists $a < L^* <b$ with $a>\max \cur{ L_\varepsilon, C(\eta)}$ such that the same inequality holds for $L \in (a,b)$. For every $m \ge 1$, and $L \in (a/(1-\eta), b/(1-\eta))$, we have $L \ge \max\cur{L_\varepsilon, C(\eta)}$ and $L(1-\eta) \in (a,b)$, hence using \eqref{eq:monotonicity} we obtain \[ f(m L) \le f(L (1-\eta)) + c \eta^{\alpha} + C(\eta) L^{-\beta} \le \liminf_{L \to \infty} f(L) + 3 \varepsilon .\] Using that $\cup_{m =1}^\infty (ma/(1-\eta), mb/(1-\eta))$ contains a half-line $(A,+\infty)$, it follows that $$ \limsup_{L\to \infty} f(L) \le \liminf_{L \to \infty} f(L) + 3 \varepsilon,$$ and the thesis follows letting $\varepsilon \to 0$. \end{proof} \subsection{Concentration inequalities} We close this section by recalling some standard concentration inequalities. Let us start with a general definition. \begin{definition}\label{def:concen} We say that a random variable $X$ with $\mathbb{E}\sqa{X}=h$ satisfies (algebraic) concentration if for every $q\ge 1$ there exists $C(q)\in (0, \infty)$ such that \[ \mathbb{E}\sqa{|X-h|^q}\le C(q) |h|^{\frac{q}{2}}. \] \end{definition} We then have \begin{lemma} Poisson, binomial and hypergeometric random variables satisfy concentration. More precisely, if : \begin{itemize} \item[i)] $N$ is a Poisson random variable with parameter $n \ge 1$ then, for every $q\ge 1$, \begin{equation}\label{eq:momentPoi} \mathbb{E}\lt[|N-n|^q\rt]\lesssim_q n^{\frac{q}{2}}. \end{equation} Hence, for every $\gamma \in (0,1)$, \begin{equation}\label{eq:density-bound-below-Poi} \mathbb{P} \bra{ N < \gamma n \quad \text{or} \quad N > (1+\gamma) n} \lesssim_{q,\gamma} (1-\gamma)^{-2q} n^{-q}. \end{equation} \item[ii)] $B$ is a binomial random variable with parameters $n$ and $p\in (0,1)$ (so that $\mathbb{E}\sqa{B} = np$) then, for every $q \ge 1$, \begin{equation}\label{eq:binomial-concentration} \mathbb{E}\lt[|B-np|^q\rt]\lesssim_q n^{\frac{q}{2}}. \end{equation} \item[iii)] $H$ is a hypergeometric random variables counting the number of red marbles extracted in $z$ draws without replacement from an urn containing $u$ marbles, $r$ of which are red (so that $\mathbb{E}\sqa{H} = z r/u$) then, for every $q \ge 1$, \begin{equation}\label{eq:concentration-hyper} \mathbb{E}\sqa{ \abs{ H - zr/u}^q } \lesssim_q r^{\frac{q}{2}}.\end{equation} \end{itemize} \end{lemma} \begin{proof} We only prove concentration in the hypergeometric case, since it is classical for both Poisson and binomial random variables. We may assume that $r \ge 1$, otherwise there is nothing to prove since $H = \mathbb{E}\sqa{H} = 0$. From \cite[Theorem 1]{hush2005concentration}, we have, for $\lambda \ge 2$, $$ \mathbb{P}\bra{ \abs{ H - \mathbb{E}\sqa{ H}} \ge \lambda } \le 2 \exp\bra{ -\alpha \lambda^2 },$$ where $$ \alpha = \min\cur{ \frac{1}{z+1} + \frac{1}{u-z+1}, \frac{1}{r+1} + \frac{1}{u-r+1} } \ge \frac{u+2 }{(r+1)(u-r+1)} \ge \frac{1}{r+1}.$$ As usual, writing $$ \mathbb{E}\sqa{ \abs{ H - \mathbb{E}\sqa{ H}}^q } = \int_0^\infty \mathbb{P}\bra{ \abs{ H - \mathbb{E}\sqa{ H}} \ge \lambda } p\lambda^{p-1} d \lambda, $$ yields the bound $$ \mathbb{E}\sqa{ \abs{ H - \mathbb{E}\sqa{ H}}^q } \lesssim_q 1+ \alpha^{-\frac{q}{2}} \lesssim_q 1 + (r+1)^{\frac{q}{2}},$$ which is bounded from above by $r^{q/2}$, since $r \ge 1$. \end{proof} \section{Combinatorial optimization problems over bipartite graphs}\label{sec:cop} \subsection{Graphs} Although we are interested in random combinatorial optimization over Euclidean bipartite graphs, it is useful to recall some general terminology. A (finite, undirected) graph $G = (V, E)$ is defined by a finite set $V = V_G$ of vertices (or nodes) and a set of edges $E = E_G$, which is a collection of unordered pairs $e = \cur{x,y} \subseteq V$ with $x \neq y$. A graph $G'$ is a subgraph of $G$ and we write $G' \subseteq G$, if $V_{G'} \subseteq V_G$ and $E_{G'}\subseteq E_G$. The induced subgraph over a subset of vertices $V' \subseteq V_G$ is defined as the subgraph $G'$ with $V_{G'} = V'$ and all the edges from $E_G$ connecting vertices in $V'$. It will be useful to denote by $\emptyset$ the empty graph, i.e., $V = \emptyset$, $E = \emptyset$, which is a subgraph of any graph $G$ Given a vertex $x \in V$, its neighborhood in $G$ is the set \[ \mathcal{N}_G(x) = \cur{ y \in V\, : \cur{x,y} \in E}.\] The degree of $x$ in $G$, $\deg_G(x)$, is the number of elements in $\mathcal{N}_G(x)$. Given $\kappa \in \mathbb{N}$, a graph $G$ is $\kappa$-regular if $\deg_G(x) = \kappa$ for every $x \in V_G$. We say that a subgraph $G' \subseteq G$ spans $V_G$ if $V_{G'} = V_G$ and $\mathcal{N}_{G'}(x) \neq \emptyset$ for ever $x \in V_{G'}$. We say that two subgraphs $G_1$, $G_2$ of $G$ are disjoint if $V_{G_1} \cap V_{G_2} =\emptyset$. A graph $G$ is connected if it cannot be decomposed as the union of two disjoint subgraphs $G = G_1 \cup G_2$, i.e., $V_G = V_{G_1} \cup V_{G_2}$ with both $V_{G_1}$, $V_{G_2} \neq \emptyset$, $V_{G_1} \cap V_{G_2} = \emptyset$ and $E_{G} =E_{G_1} \cup E_{G_2}$. Given $\kappa \in \mathbb{N}$, $\kappa \ge 1$, we say that a graph $G$ is $\kappa$-connected if any subgraph $G' \subseteq G$ obtained by removing from $G$ $(\kappa-1)$-edges is still connected. A cycle is a connected $2$-regular graph, a tree is a connected graph which contains no cycles as subgraphs Given two graphs $G_1$, $G_2$ and an injective function $\sigma: V_{G_1} \to V_{G_2}$, we let $\sigma(E_1) = \cur{ \cur{\sigma(x), \sigma(y)} \, : \, \cur{x,y} \in E_{G_1}}$. If $\sigma(E_1) \subseteq E_2$, then we say that $G_1$ embeds into $G_2$ via $\sigma$. If $\sigma$ is bijective and $\sigma(E_1) = E_2$, then we say that $G_1$ is isomorphic to $G_2$ via $\sigma$. A graph $G$ is complete if $E_G$ consists of all the pairs $\cur{x,y} \subseteq V$ with $x \neq y$. The complete graph over $V = [n]$ is commonly denoted by $\mathcal{K}_n$. Any complete graph $G$ with $n$ vertices is isomorphic to $\mathcal{K}_n$. We say that the graph $G$ is bipartite over a partition $V = X \cup Y$ (i.e., $X \cap Y = \emptyset$), if every $e\in E$ can be written as $e= \cur{x,y}$ with $x\in X$, $y \in Y$. A graph is complete bipartite if it is bipartite over a partition $V = X \cup Y$ and every pair $\cur{x,y}$ with $x\in X$, $y \in Y$ is an edge. For any $n, m \in \mathbb{N}$, any two complete bipartite graphs with $X$ having $n$ elements and $Y$ having $m$ elements are isomorphic. To fix a representative, we define $\mathcal{K}_{n,m}$ as the complete bipartite graph over the vertex set $V = [n]_1 \cup [m]_2$. We introduce a weight function on edges $w: E \to [0, \infty)$, $w(e) = w(x,y)$. The total weight of $G$ is then \[ w(G) = \sum_{e \in E} w(e).\] A subgraph $G' \subseteq G$ of a weighted graph is always intended with the restriction of $w$ on $E'$. Notice that for the empty graph $\emptyset \subseteq G$ we have $w(\emptyset) = 0$. We are interested in geometric realizations of graphs, where vertices are in correspondence with points in a metric space $(\Omega, \mathsf{d})$, and the weight function is a power of the distance between the corresponding points, with a fixed exponent $p>0$. Since we consider only complete and complete bipartite graphs, we introduce the following notation. Given ${\bf x} = (x_i)_{i=1}^n \subseteq \Omega$, we let $\mathcal{K}({\bf x})$ be the complete graph $\mathcal{K}_n$ endowed with the weight function $w(i,j) = \mathsf{d}(x_i, x_j)^p$. Similarly, given ${\bf x} = (x_i)_{i=1}^n$, ${\bf y} = (y_j)_{j=1}^m \subseteq \Omega$, we let $\mathcal{K}({\bf x}, {\bf y})$ denote the complete bipartite graph $\mathcal{K}_{n,m}$ endowed with the weight function $w((1,i), (2,j)) = \mathsf{d}(x_i, y_j)^p$. Notice that the points in ${\bf x}$ and ${\bf y}$ may not be all distinct, but this will in fact occur with probability zero. If all the points are distinct, then we can and will identify the vertex set directly with the set of points ${\bf x}$ for $\mathcal{K}({\bf x})$, and with the set of points in ${\bf x} \cup {\bf y}$ for $\mathcal{K}({\bf x}, {\bf y})$. With this convention, if ${\bf x} = {\bf x}^0 \cup {\bf x}^1$, ${\bf y} = {\bf y} ^0 \cup {\bf y}^1$, then both $\mathcal{K}({\bf x}^0, {\bf y}^0)$ and $\mathcal{K}({\bf x}^1, {\bf y}^1)$ are naturally seen as subgraphs of $\mathcal{K}({\bf x}, {\bf y})$. \subsection{Combinatorial problems} A combinatorial optimization problem $\mathsf{P}$ on weighted graphs is informally defined by prescribing, for every graph $G$, a set of subgraphs $G' \subseteq G$, also called feasible solutions $\mathcal{F}_{G}$, and, after introducing a weight $w$, by minimizing $w(G')$ over all $G'\in \mathcal{F}_G$. Our aim is to study problems on random geometric realizations of complete bipartite graphs $\mathcal{K}_{n,n}$, thus it is sufficient to define a combinatorial optimization problem over complete bipartite graphs as a collection of feasible solutions $\mathsf{P} = ( \mathcal{F}_{n,n})_{ n \in \mathbb{N}}$, with $\mathcal{F}_{n,n}$ being the feasible solutions on $\mathcal{K}_{n,n}$. We will mostly consider problems $\mathsf{P}$ that satisfy the following assumptions: \begin{enumerate}[label=A\arabic*, series=A] \item \label{as:isomorphism} \emph{(isomorphism)} if $\sigma$ is any isomorphism of $\mathcal{K}_{n,n}$ into itself and $G \in \mathcal{F}_{n,n}$, then $\sigma(G) = (\sigma(V_G), \sigma(E_G)) \in \mathcal{F}_{n,n}$; \item \label{as:spanning} \emph{(spanning)} for every $n \in \mathbb{N}$, $\mathcal{F}_{n,n}$ is not empty and there exists $\mathsf{c}_{\operatorname{A2}} >0$ such that, for $n < \mathsf{c}_{\operatorname{A2}}$, $\mathcal{F}_{n,n} = \cur{ \emptyset}$ while for $n \ge \mathsf{c}_{\operatorname{A2}}$, every $G \in \mathcal{F}_{n,n}$ spans $\mathcal{K}_{n,n}$;\ \item \label{as:bddegree} \emph{(bounded degree)} there exists $\mathsf{c}_{\operatorname{A3}} >0$ such that, for every $n \in \mathbb{N}$ and every feasible solution $G \in \mathcal{F}_{n,n}$, one has $\deg_G(x) \le \mathsf{c}_{\operatorname{A3}}$ for every $x \in G$. \end{enumerate} Given $\mathsf{P} = ( \mathcal{F}_{n,n})_{n \in \mathbb{N}}$, we canonically extend it to graphs $\mathcal{K}_{n,m}$, with $n \neq m$, defining $\mathcal{F}_{n,m}$ as the collection of all graphs $\sigma(G)$ where $G\in \mathcal{F}_{z,z}$, $z = \min\cur{n,m}$ and $\sigma$ is an isomorphism of $\mathcal{K}_{n,m}$ into itself. In the geometric setting, i.e., when $\mathcal{K}_{n,m}$ is mapped into $\mathcal{K}({\bf x},{\bf y})$ with ${\bf x} = (x_i)_{i=1}^n$, ${\bf y} = (y_j)_{j=1}^m\subseteq \Omega$, with $(\Omega, \mathsf{d})$ metric space, we introduce the following notation for the cost of a problem $\mathsf{P}$: $$ \C_{\pP}^p ({\bf x}, {\bf y}) = \min_{G \in \mathcal{F}_{n,m}} \sum_{\cur{(1,i),(2,j)}\in E_G} \mathsf{d}(x_i,y_j)^p .$$ Recalling the definition of $\mathcal{F}_{n,m}$ if $n \neq m$, we also have the identity \begin{equation}\label{eq:min-subgraphs} \C_{\pP}^p ({\bf x}, {\bf y}) = \min_{\substack{{\bf x}' \subseteq {\bf x}, {\bf y}' \subseteq {\bf y} \\ |{\bf x}'| = |{\bf y}'| = \min\cur{|{\bf x}|, |{\bf y}|}}} \C_{\pP}^p ({\bf x}', {\bf y}').\end{equation} \begin{remark} Assumption \ref{as:spanning} ensures that, if $\min\cur{|{\bf x}|, |{\bf y}|} < \mathsf{c}_{\operatorname{A2}}$, then $\C_{\pP}^p({\bf x}, {\bf y}) = 0$. \end{remark} \begin{remark} If $(\Omega', \mathsf{d}')$ is a metric space and $S: \Omega \to \Omega'$ is Lipschitz, i.e., for some constant $\operatorname{Lip} S$ one has $\mathsf{d}'(S(x), S(y)) \le (\operatorname{Lip} S) \mathsf{d}(x, y)$ for every $x$, $y\in \Omega$, then writing $S({\bf x}) = (S(x_i))_{i=1}^n$, $S({\bf y}) = (S(y_j))_{j=1}^m$, we clearly have the inequality \begin{equation}\label{eq:lipschitz-bound-deterministic} \C_{\pP}^p( S({\bf x}), S({\bf y}) ) \le (\operatorname{Lip} S)^p \C_{\pP}^p( {\bf x}, {\bf y} ). \end{equation} \end{remark} \begin{remark} Similar definitions and assumptions may be given in the non-bipartite case, thus defining combinatorial optimization problems $\mathsf{P} = ( \mathcal{F}_{n})_{n \in \mathbb{N}}$ over complete graphs, as a collection of feasible solutions $\mathcal{F}_{n}$ over the complete graph $\mathcal{K}_{n}$. \end{remark} \subsection{Examples} Let us introduce some fundamental examples of these problems. \subsubsection*{Assignment problem} The minimum weight bipartite matching problem, also called assignment problem, is defined letting $\mathcal{F}_{n,n}$ be the set of perfect matchings in $\mathcal{K}_{n,n}$, i.e., spanning subgraphs induced by a collection of edges which have no vertex in common (if $n=0$ we simply let $\mathcal{F}_{n,n} = \cur{\emptyset}$). Feasible solutions are in correspondence with permutations $\sigma$ over $[n]$, letting $$ E_{\sigma} = \cur{ \cur{(1,i), (2,\sigma(i))} \, : \, i \in [n] }.$$ When $n \neq m$, e.g.\ $n \le m$, the same correspondence holds with the set of injective maps $\sigma:[n] \to [m]$. Therefore, given a weight $w$ on $\mathcal{K}_{n,m}$, the cost of the assignment problem is \[\min_{\sigma} \sum_{i=1}^n w\bra{ (1,i), (2,\sigma(i))}.\] In the geometric case, i.e., on the weighted graph $\mathcal{K}({\bf x}, {\bf y})$ with ${\bf x} = (x_i)_{i=1}^n$, ${\bf y}=(y_j)_{j=1}^m \subseteq \Omega$ and $w\bra{ (1,i), (2,j)}=\mathsf{d}(x_i, y_{j} )^p$, this expression becomes \[ \mathsf{M}^p({\bf x}, {\bf y}) = \min_{\sigma} \sum_{i=1}^n \mathsf{d}(x_i, y_{\sigma(i)} )^p.\] If $n>m$, then one simply exchanges the roles of $n$ and $m$. \begin{remark} If $n=m$, Birkhoff's theorem ensures equivalence between the bipartite matching problem and the optimal transport between the associated empirical measures $\mu^{\bf x} = \sum_{i=1}^n \delta_{x_i}$, $\mu^{\bf y} = \sum_{j=1}^n \delta_{y_j}$, i.e., \begin{equation}\label{eq:wass-equals-matching} \mathsf{M}^p({\bf x}, {\bf y}) = \mathsf{W}^p( \mu^{\bf x}, \nu^{\bf y}). \end{equation} Therefore, using the triangle inequality \eqref{eq:triangle}, we can bound from above as follows: \begin{equation} \label{eq:matching-below-wasserstein} \mathsf{M}^p({\bf x}, {\bf y}) \lesssim \mathsf{W}^p\bra{ \mu^{\bf x}, n \lambda} + \mathsf{W}^p\bra{ \mu^{\bf y}, n \lambda}, \end{equation} for every probability measure $\lambda$ on $\mathbb{R}^d$. \end{remark} \subsubsection*{Travelling salesperson problem} The travelling salesperson problem (TSP) is usually defined on a general graph by prescribing as feasible solutions the cycles visiting each vertex exactly once (also called Hamiltonian cycles). In the complete bipartite case $\mathcal{K}_{n,n}$, such cycles exist for every $n \ge 2$, and assumptions \ref{as:isomorphism}, \ref{as:spanning} and \ref{as:bddegree} are also clearly satisfied (letting $\mathcal{F}_{n,n} = \cur{\emptyset}$ if $n\in \cur{0,1}$). Similarly as in the case of the assignment problem, feasible solutions are in this case in correspondence with pairs of permutations $\sigma$, $\tau$ over $[n]$, letting \begin{equation}\label{eq:e-sigma-tau} E_{\sigma, \tau} = \cur{ \cur{(1,\sigma(i)), (2,\tau(i))}, \cur{(1,\sigma(i), (2,\tau(i+1)) )} \, : \, i \in \cur{1, \ldots, n}},\end{equation} where we conventionally let $\tau(n+1) = \tau(1)$ (we will always use summation $\operatorname{mod} n$ in such cases). In words, $\sigma$ and $\tau$ prescribe the order at which the vertices are visited by the cycle. When $n \neq m$, e.g.\ $n \le m$, the same correspondence holds with injective maps $\sigma$, $\tau$ from $[n]$ into $[m]$. Therefore, given a weight $w$ on $\mathcal{K}_{n,m}$, the cost of the TSP reads \[\min_{\sigma, \tau} \sum_{i=1}^n w\bra{(1,\sigma(i)), (2,\tau(i))} + w\bra{(1,\sigma(i)), (2,\tau(i+1)) }.\] In the geometric case, i.e., on the weighted graph $\mathcal{K}({\bf x}, {\bf y})$ with ${\bf x} = (x_i)_{i=1}^n$, ${\bf y}=(y_j)_{j=1}^m \subseteq \Omega$, this becomes \[ \mathbf{C}_{\mathsf{TSP}}^p({\bf x}, {\bf y}) = \min_{\sigma, \tau} \sum_{i=1}^n \mathsf{d}(x_{\sigma(i)}, y_{\tau(i)} )^p + \mathsf{d}(x_{\sigma(i)}, y_{\tau(i+1)} )^p.\] If $n>m$, then one simply exchanges the roles of $n$ and $m$. The non-bipartite version of the TSP , i.e., on $\mathcal{K}_n$, feasible solutions to the TSP are in correspondence with permutations $\sigma$ over $[n]$, letting $$ E_{\sigma} = \cur{ \cur{\sigma(i), \sigma(i+1)} \, : \, i \in [n]}.$$ In the geometric case ${\bf x} = (x_i)_{i=1}^n \subseteq \Omega$, it becomes \[ \mathbf{C}_{\mathsf{TSP}}^p({\bf x}) = \min_{\sigma} \sum_{i=1}^n \mathsf{d}(x_{\sigma(i)}, x_{\sigma(i+1)} )^p. \] \subsubsection*{Connected $\kappa$-factor problem} The TSP can be generalized in many directions. For example, since a cycle is a connected graph such that every vertex has degree $2$, i.e., it is $2$-regular, we may instead define as feasible solutions $\kappa$-regular spanning connected subgraphs, for a fixed $\kappa \in \mathbb{N}$, $\kappa \ge 2$. This defines a non-empty set of feasible solutions $\mathcal{F}_{n,n}$ over $\mathcal{K}_{n,n}$ if $n \ge \kappa$ (otherwise we let $\mathcal{F}_{n,n} = \cur{\emptyset}$) and assumptions \ref{as:isomorphism}, \ref{as:spanning} and \ref{as:bddegree} are easily seen to be satisfied. We refer to such problem as the (minimum weight) connected $\kappa$-factor problem. A simpler variant is to require that feasible solutions are $\kappa$-regular but not necessarily connected: this is simply known as (minimum weight) $\kappa$-factor problem. Let us notice that, for $\kappa=1$, this reduces to the assignment problem. Back to the the connected $\kappa$-factor problem, a simple fact worth noticing, that we will use below, is that any connected $\kappa$-regular bipartite graph $G$ is $2$-connected, i.e., it remains connected even after removing a single edge. Assume that $V_G = X \cup Y$, with $X \cap Y = \emptyset$ and by contradiction let $x\in X$, $y \in Y$ be such that $\cur{x,y} \in E_G$ and the subgraph $G' \subseteq G$ with edge set $E_{G'} = E_G \setminus \cur{x,y}$ is not connected: there are two disjoint subgraphs $G_1'$, $G_2'$ with $x \in V_{G_1'}$, $y \in V_{G_2'}$ with $G' = G_1' \cup G_2'$. All the vertices in $G_1'$ have degree $\kappa$, except for $x$, whose degree is $\kappa-1$. However, if we let $n_X = |V_{G_1'} \cap X|$, $n_Y = |V_{G_1'} \cap Y|$, then using the fact that the graph $G_1'$ is bipartite we can count the number of edges as the sum of the degrees of the vertices in $V_{G_1'} \cap X$ or equivalently of those in $V_{G_1'} \cap Y$, which leads to the identity $\kappa n_X - 1 = \kappa n_Y$, from which $\kappa(n_X - n_Y ) = 1$, which gives a contradiction. \subsubsection*{$\kappa$-bounded degree minimum spanning tree} The minimum weight spanning trees (MST) problem is defined by letting feasible solutions be all spanning subgraphs that are trees, i.e., connect and acyclic, whose existence on any given connected graph is guaranteed by standard algorithms. This problem however may not have uniformly bounded degree, thus assumption \ref{as:bddegree} may not be satisfied. Therefore, we restrict the set of feasible solutions to spanning trees over $\mathcal{K}_{n,n}$ such that that each vertex degree is less than or equal to some fixed $\kappa \ge 2$ (letting $\mathcal{F}_{0,0}=\cur{\emptyset}$). This problem, known as the $\kappa$-bounded degree minimum spanning tree ($\kappa$-MST), satisfies assumptions \ref{as:isomorphism}, \ref{as:spanning} and \ref{as:bddegree}: notice in particular that removing any edge from a Hamiltonian cycle, i.e., a feasible solution for the TSP, gives a $2$-bounded degree minimum spanning tree. We remark here that the $\kappa$-MST problem may be also directly defined over graphs $\mathcal{K}_{n,m}$, with $n \neq m$, with a non trivial set of feasible solutions (provided that $|n-m|$ is not too large). However, also in this case we follow our the general convention, so that if $n \neq m$, the set $\mathcal{F}_{n,m}$ does not contain spanning trees of $\mathcal{K}_{n,m}$ but only spanning trees over subgraphs isomorphic to $\mathcal{K}_{z,z}$ with $z = \min\cur{n,m}$. A simple fact that we will use below is that any $G \in \mathcal{F}_{n,n}$ contains at least one leaf (i.e., a vertex with degree $1$) in $[n]_1$ and one in $[n]_2$. This is because more generally any spanning tree over $\mathcal{K}_{n,n}$ contains at least one leaf in $[n]_1$ and one in $[n]_2$. Indeed, assume by contradiction that there are no leaves in $[n]_1$. Then, since the tree spans, all the vertices in $[n]_1$ must have degree at least $2$ (the graph is connected, hence every vertex has at least degree $1$) and since no edges connect pairs of vertices in $[n]_1$, these are all distinct, hence the tree contains at least $2n$ edges, which contradicts the well-known fact that any tree (not necessarily bipartite) over $2n$ vertices must have $2n-1$ edges. In order to perform our analysis, we introduce two further assumptions that we discuss in the following subsections. \subsection{Local merging} Our analysis relies on a key subadditivity inequality, that ultimately follows by a stability assumption with respect to local merging operations, besides assumptions \ref{as:isomorphism} and \ref{as:bddegree}. Let us give the following general definition. \begin{definition}[gluing] Given a graph $G$ and two disjoint subgraphs $G_1$, $G_2 \subseteq G$, we say that $G' \subseteq G$ is obtained by gluing at $x_1 \in V_{G_1}$, $x_2 \in V_{G_2}$ if $V_{G'} = V_{G_1}\cup V_{G_2}$, \[ (E_{G_1} \cup E_{G_2}) \setminus E_{G'} \subseteq \mathcal{N}_{G_1}(x_1) \cup \mathcal{N}_{G_2}(x_2)\] and \[ E_{G'} \setminus (E_{G_1} \cup E_{G_2}) \subseteq \cur{ \cur{x_1, y} : y \in \mathcal{N}_{G_2}(x_2)} \cup \cur{ \cur{x_2, y} : y \in \mathcal{N}_{G_2}(x_1)}.\] \end{definition} In words, gluing at $x_1$, $x_2$ means that the two subgraphs are joined by (possibly) removing and adding edges connecting $x_2$ to vertices from the neighborhood of $x_1$ in $G_1$, and similarly $x_1$ to vertices from the the neighborhood of $x_2$ in $G_2$. In particular, we have that $\mathcal{N}_{G'}(x) = \mathcal{N}_{G_1}(x)$ for every $x\in V_{G_1} \setminus \bra{ \mathcal{N}_{G_1}(x_1)\cup \cur{x_1}}$, and similarly $\mathcal{N}_{G'}(x) = \mathcal{N}_{G_2}(x)$ for every $x\in V_{G_2} \setminus \bra{ \mathcal{N}_{G_2}(x_2)\cup \cur{x_2}}$. Back to combinatorial optimization problems over bipartite graphs, our assumption is, loosely speaking, that any two (non empty) feasible solutions $G \in \mathcal{F}_{n,n}$, $G' \in \mathcal{F}_{n',n'}$, can be glued together yielding a feasible solution $G \in \mathcal{F}_{n+n', n+n'}$. In fact, we also allow adding up to $\mathsf{c}$ edges, but only connecting vertices of $G$, where $\mathsf{c} \in \mathbb{N}$ is a constant (depending only on the problem $\mathsf{P}$). Before giving a precise formulation of the assumption, we notice that $G$ and $G'$ are in general not disjoint: what we mean is that $G'$ must be suitably ``translated''. Precisely, given $n \in \mathbb{N}$, we introduce the map $$ \tau_n: V_{G'} \to [n+n']_1 \cup [n+n']_2$$ defined as $$ \tau\bra{ (1,i)} = (1, n+i), \quad \tau\bra{ (2,j)} = (2,n+j),$$ so that $G$, $\tau( G') \subseteq \mathcal{K}_{n+n', n+n'}$ are disjoint. We consider therefore combinatorial optimization problems $\mathsf{P}$ over bipartite graphs which satisfy the following assumption: \begin{enumerate}[label=A\arabic*, series=A, resume] \item\label{ass:local-merging} \emph{(local merging)} there exists $\mathsf{c}_{\operatorname{A4}} \ge 0$ such that, for every $n$, $n' \in \mathbb{N}$, and $G\in \mathcal{F}_{n,n}$, $G'\in \mathcal{F}_{n',n'}$ with both $G\neq \emptyset$ and $G' \neq \emptyset$, one can find $G'' \in \mathcal{F}_{n+n', n+n'}$ obtained by gluing $G$ and $\tau (G')$ at the vertices $(1,1)$, $(1,n+1)$ and possibly adding up to $\mathsf{c}_{\operatorname{A4}}$ edges from those of $\mathcal{K}_{n,n}$. \end{enumerate} The reason why we also allow up to $\mathsf{c}_{\operatorname{A4}}$ additional edges is to include some problems where connectedness may be destroyed by gluing, such as the $\kappa$-MST. This should be compared with the \emph{merging} assumption \cite[(A4)]{BaBo}, where a bounded number of edges from the whole $\mathcal{K}_{n+n', n+n'}$ instead is allowed to be added to the union $G \cup \tau(G')$ (with our notation). Notice however that, in our case, since the extra edges are from $\mathcal{K}_{n,n}$ it remains true that \begin{equation}\label{eq:neighbour-same} \text{ $\mathcal{N}_{G''}(x) = \mathcal{N}_{\tau (G')}(x)$, for every $x\in V_{\tau (G')} \setminus \bra{ \mathcal{N}_{\tau (G')}((1,n+1))\cup \cur{(1, n+1)}}$,}\end{equation} which is a key condition that we use below. All the problems described in the previous section satisfy \ref{ass:local-merging}. \begin{lemma} The TSP, the connected $\kappa$-factor problem (as well as the non connected one) and the $\kappa$-MST over complete bipartite graphs satisfy assumption \ref{ass:local-merging}. \end{lemma} \begin{proof} Let $G \in \mathcal{F}_{n,n}$, $G'\in\mathcal{F}_{n', n'}$ be both non empty. Then (e.g.\ by assumption \ref{as:spanning}) $\deg_{G}(1,1) \ge 1$ but also $\deg_{\tau (G')}(1, n+1) \ge 1$. The basic idea is to pick $y \in \mathcal{N}_{G}(1,1)$, $y' \in \mathcal{N}_{G'}(1,n+1)$, remove the edges $\cur{(1,1), y}$, $\cur{(1,n+1), y'}$ and add instead $\cur{ (1,1), y'}$, $\cur{(1,n+1), y'}$. This operation does not change the vertex degrees, in particular at $(1,1)$ and $(1,n+1)$. For the TSP and more generally the connected $\kappa$-factor problem, the resulting graph $G''$ is connected, because after removing a single edge, both graphs $G$ and $\tau (G')$ are still connected, and adding the new edges has the effect of connecting the two graphs (hence in this case $\mathsf{c}_{\operatorname{A4}} = 0$). For the $\kappa$-bounded degree MST, we use the fact that the tree $G \in \mathcal{F}_{n,n}$ must have at least one leaf in the set of $[n]_1$ and one in the set $[n]_2$. Therefore, we obtain a connected tree (with degree bounded by $\kappa$) if we add also one edge connecting two such leaves (hence is this case $\mathsf{c}_{\operatorname{A4}} = 1$). \end{proof} \subsection{Subadditivity inequality} Using all the assumptions introduced so far, in particular \ref{ass:local-merging}, we establish a fundamental subadditivity inequality. \begin{proposition}[Approximate subadditivity]\label{prop:partition} Let $\mathsf{P}$ be a combinatorial optimization problem over bipartite graphs satisfying assumptions \ref{as:isomorphism}, \ref{as:spanning}, \ref{as:bddegree} and \ref{ass:local-merging}. For a metric space $(\Omega,\mathsf{d})$ and a finite partition $\Omega = \cup_{k=1}^K \Omega_k$, $K \in \mathbb{N}$, \begin{enumerate}[label=\emph{\roman*})] \item \label{prop-partition-x0-assumption} let ${\bf x}^0$, ${\bf y}^0 \subseteq \Omega$ be such that $\min\cur{|{\bf x}^0|, |{\bf y}^0|} \ge \max\cur{\mathsf{c}_{\operatorname{A2}}, K}$, \item for every $k =1, \ldots, K$, let ${\bf x}^k$, ${\bf y}^k \subseteq \Omega_k$ with $|{\bf x}^k| = |{\bf y}^k| =n_k$, with either $n_k \ge \mathsf{c}_{\operatorname{A2}}$ or $n_k = 0$ (i.e., both families are empty) \item let ${\bf z}=(z_k)_{k=1}^K$ with $z_k \in \Omega_k$, for every $k=1,\ldots, K$. \end{enumerate} Then, the following inequality holds: \begin{equation}\label{eq:sub} \C_{\pP}^p\bra{ {\bf x}^0 \cup \bigcup_{k=1}^K {\bf x}^k, {\bf y}^0 \cup \bigcup_{k=1}^K {\bf y}^k } - \sum_{k=1}^K \C_{\pP}^p ({\bf x}^k, {\bf y}^k) \lesssim \C_{\pP}^p({\bf x}^0, {\bf y}^0)+ \mathsf{M}^p({\bf z}, {\bf x}^0)+ \sum_{k=1}^K \operatorname{diam}(\Omega_k)^p. \end{equation} The implicit constant depends only upon $p$, $\mathsf{c}_{\operatorname{A2}}$, $\mathsf{c}_{\operatorname{A3}}$ and $\mathsf{c}_{\operatorname{A4}}$ (in particular not on $K$). \end{proposition} \begin{remark}\label{rem:remove-z-points} The role played by the points ${\bf z}$ is quite marginal, and indeed if ${\bf x}^0(\Omega_k) >0$ for every $k$, then by choosing $z_k \in {\bf x}^0_{\Omega_k}$, the term $\mathsf{M}^p({\bf z}, {\bf x}^0)$ vanishes. \end{remark} \begin{proof} Recalling \eqref{eq:min-subgraphs}, up to replacing ${\bf x}^0$, ${\bf y}^0$ with subsets ${\bf x}'$, ${\bf y}'$ with $|{\bf x}'| = |{\bf y}'| = \min\cur{|{\bf x}^0|, |{\bf y}^0|}$, we may also assume that $|{\bf x}^0| = |{\bf y}^0|$. For every $k=1, \ldots, K$ let $G_k \subseteq \mathcal{K}({\bf x}^k, {\bf y}^k)$ be a minimizer for $\mathsf{P}$. If $n_k=0$, then $G_k =\emptyset$. Otherwise, $n_k \ge \mathsf{c}_{\operatorname{A2}}$, and by assumption \ref{as:spanning} it is in particular non-empty and using Markov inequality, we can choose $x^k \in {\bf x}^k$ such that $$ \sum_{y \in \mathcal{N}_{G_k} (x^k)} \mathsf{d}(x^k, y)^p \le \frac{ 4\C_{\pP}^p({\bf x}^k, {\bf y}^k)}{|{\bf x}^k|}\lesssim \operatorname{diam}(\Omega_k)^p.$$ For the last estimate we used that $\deg_{x^k}(G_k) \le \mathsf{c}_{\operatorname{A3}}$. Similarly, let $G_0\subseteq \mathcal{K}({\bf x}^0, {\bf y}^0)$ be a (also non-empty) minimizer for $\C_{\pP}^p({\bf x}^0, {\bf y}^0)$ and let $\sigma :\{1, \ldots, K\}\to \{ 1, \ldots, |{\bf x}^0|\}$ be an optimal matching between ${\bf z}$ and ${\bf x}^0$. We iteratively use assumptions \ref{as:isomorphism} and \ref{ass:local-merging} to define feasible solutions $$ \tilde{G}_k \subseteq \mathcal{K}\bra{ {\bf x}^0 \cup \bigcup_{i=1}^{k} {\bf x}^i, {\bf y}^0 \cup \bigcup_{i=1}^{k} {\bf y}^i}.$$ We begin by letting $\tilde{G}_0 = G_0$. For $k =1, \ldots, K$, having already defined $\tilde{G}_{k-1}$, if $n_k = 0$, then we simply let $\tilde{G}_k = \tilde{G}_{k-1}$. Otherwise, we obtain a feasible solution $\tilde{G}_k$ by gluing $G_{k}$ with $\tilde{G}_{k-1}$ at the vertices $x^k$, $x_{k}^0$ and adding up to $\mathsf{c}_{\operatorname{A4}}$ edges from $\mathcal{K}({\bf x}^k, {\bf y}^k)$. The fact that we can glue at any such pair of vertices is due to assumption \ref{as:isomorphism}: up to isomorphisms we can assume that $x^k$ corresponds to the abstract graph vertex $(1,1)$ and that $x_{k}^0$ to $(1, n_k+1)$. % This construction gives the following inequality between the graph weights, if $n_k \neq 0$: \begin{equation}\label{eq:gluing-error} \begin{split} w( \tilde{G}_k ) - w( \tilde{G}_{k-1}) - w(G_k) \le & \mathsf{c}_{\operatorname{A4}} \operatorname{diam}(\Omega_k)^p \\ & + \sum_{y \in \mathcal{N}_{G_k}(x^k) } \mathsf{d}(x_{\sigma(k)}^0, y)^p + \sum_{y \in \mathcal{N}_{\tilde{G}_{k-1}}(x_{\sigma(k)}^0) } \mathsf{d}(x^k, y)^p, \end{split} \end{equation} while if $n_k = 0$, we simply have $w(\tilde{G}_k) = w(\tilde{G}_{k-1})$. We bound from above the last two terms in \eqref{eq:gluing-error} as follows: first, \[\begin{split} \sum_{y \in \mathcal{N}_{G_k}(x^k) } \mathsf{d}(x_{\sigma(k)}^0, y)^p & \lesssim \sum_{y \in \mathcal{N}_{G_k}(x^k) } \mathsf{d}( x_{\sigma(k)}^0, z_k)^p + \mathsf{d}( z_k, x^k)^p + \mathsf{d}(x^k, y)^p\\ & \lesssim \mathsf{d}( x_{\sigma(k)}^0, z_k)^p + \mathsf{d}( z_k, x^k)^p +\sum_{y \in \mathcal{N}_{G_k}(x^k) } \mathsf{d}(x^k, y)^p \\ & \lesssim \mathsf{d}( x_{\sigma(k)}^0, z_k)^p + \operatorname{diam}(\Omega_k)^p, \end{split}\] where we used that $\deg_{x^k}(G_k) \le \mathsf{c}_{\operatorname{A3}}$. To bound the last term, we notice that each step in the construction we are locally merging at different points in ${\bf x}^0$: since no such points are adjacent because the graph is bipartite, using \eqref{eq:neighbour-same} by induction yields \[ \mathcal{N}_{\tilde{G}_{k-1}}(x_{\sigma(k)}^0)=\mathcal{N}_{G_0}(x_{\sigma(k)}^0),\] which in particular contains at most $\mathsf{c}_{\operatorname{A3}}$ elements, since $G_0$ is feasible. Therefore, \[ \begin{split} \sum_{y \in \mathcal{N}_{\tilde{G}_{k-1}}(x_{\sigma(k)}^0)} \mathsf{d}(x^k, y)^p & = \sum_{y \in \mathcal{N}_{G_0}(x_{\sigma(k)}^0)} \mathsf{d}(x^k, y)^p \\ & \lesssim \sum_{y \in \mathcal{N}_{G_0}(x_{\sigma(k)}^0)} \mathsf{d}(x^k, z_k)^p +\mathsf{d}(z_k, x_{\sigma(k)}^0)^p + \mathsf{d}(x_{\sigma(k)}^0, y)^p\\ & \lesssim \operatorname{diam}(\Omega_k)^p + \mathsf{d}(z_k, x_{\sigma(k)}^0)^p + \sum_{y \in \mathcal{N}_{G_0}(x^0_{\sigma(k)})}\mathsf{d}(x_{\sigma(k)}^0, y)^p.\end{split}\] Summing \eqref{eq:gluing-error} upon $k=1,\ldots, K$, we obtain \eqref{eq:sub} because $$ \sum_{k=1}^K \mathsf{d}(z_k, x_{\sigma(k)}^0)^p = \mathsf{M}^p({\bf z}, {\bf x}^0)$$ and, being all the points $x^0_{\sigma(k)}$ different, \[ \sum_{k=1}^K \sum_{y \in \mathcal{N}_{G_0}(x^0_{\sigma(k)})}\mathsf{d}(x_{\sigma(k)}^0, y)^p \le \sum_{x\in {\bf x}^0} \sum_{y \in \mathcal{N}_{G_0}(x)}\mathsf{d}(x, y)^p = \C_{\pP}^p({\bf x}^0, {\bf y}^0). \qedhere \] % % % \end{proof} \subsection{Growth/regularity} The last assumption that we introduce for a combinatorial optimization problem $\mathsf{P}$ over bipartite graphs is a general upper bound for the cost when specialized to a geometric graph in the Euclidean cube $(0,1)^d$: \begin{enumerate}[label=A\arabic*, series=A, resume] \item\label{as:growth} (growth/regularity) There exists $\mathsf{c}_{\operatorname{A5}} \ge 0$ such that, for every ${\bf x}, {\bf y} \subseteq (0,1)^d$, we have \begin{equation} \label{eq:upper-bound-deterministic} \C_{\pP}^p({\bf x}, {\bf y}) \le \mathsf{c}_{\operatorname{A5}} \bra{ \min\cur{|{\bf x}|^{1-\frac{p}{d}}, |{\bf y}|^{1-\frac{p}{d}}}+ \mathsf{M}^p({\bf x},{\bf y})}.\end{equation} \end{enumerate} \begin{remark}\label{remgrowth} Notice that if $\Omega\subset (0,1)^d$ then \eqref{eq:upper-bound-deterministic} applies in particular for ${\bf x},{\bf y} \subseteq \Omega$. By scaling we obtain that for every bounded set $\Omega$ and every ${\bf x},{\bf y} \subseteq \Omega$, \[\C_{\pP}^p({\bf x}, {\bf y}) \le \mathsf{c}_{\operatorname{A5}} \bra{ \operatorname{diam}(\Omega)^p\min\cur{|{\bf x}|^{1-\frac{p}{d}}, |{\bf y}|^{1-\frac{p}{d}}}+ \mathsf{M}^p({\bf x},{\bf y})}.\] \end{remark} Using \eqref{eq:min-subgraphs}, we obtain at once that in order to establish that a given problem $\mathsf{P}$ satisfies \eqref{eq:upper-bound-deterministic} it is enough to consider the case where ${\bf x}$, ${\bf y} \subseteq (0,1)^d$ have the same number of elements. Notice that this assumption seems slightly different with respect to the previous ones, as it explicitly refers to the cost for Euclidean realizations of the graph, instead of feasible solutions, and relies as well on the assignment problem. In fact, the constant $\mathsf{c}_{\operatorname{A5}}$ depends upon the problem $\mathsf{P}$ but also on the dimension $d$ and the exponent $p$, which however will be fixed in our derivations so we avoid to explicitly state it. It is well known that quite general arguments, such as the space-filling curve heuristics \cite[Chapter 2]{steele1997probability}, lead to an upper bound in terms of $n^{1-p/d}$ for non-bipartite combinatorial optimization problems over $n$ points in a cube, under very mild assumptions, including those introduced above. Simple examples show that similar bounds cannot hold for their bipartite counterparts, which explains the second term in the right-hand side of \eqref{eq:upper-bound-deterministic}. To establish it in our examples we follow the strategy from \cite{capelli2018exact}, where limit results for the random Euclidean bipartite TSP for $p=d=2$ were first obtained. \begin{lemma} \label{lem:capelli} The TSP, the connected $\kappa$-factor problem (as well as the non-connected one) and the $\kappa$-MST problems over complete bipartite graphs satisfy assumption \ref{as:growth} (with a constant $\mathsf{c}_{\operatorname{A5}}$ depending on $\kappa$, $p$, $d$ only). \end{lemma} \begin{proof} Let us first observe that the cost of the $\kappa$-MST problem is always bounded from above by the cost of the minimum weight connected $\kappa$-factor problem, since given any connected $\kappa$-factor, one can extract from it a MST whose degree at every vertex is then bounded by $\kappa$. Therefore it is sufficient to check that assumption \ref{as:growth} holds with $\mathsf{P}$ being the connected $\kappa$-factor problem, for any $\kappa \ge 2$ (the case $\kappa=2$ being the TSP). For $(\Omega, \mathsf{d})$ a general metric space and ${\bf x}, {\bf y} \subseteq \Omega$ we establish first the bound \begin{equation}\label{eq:c-bip-c-mono} \C_{\pP}^p({\bf x}, {\bf y}) \lesssim \mathbf{C}_{\mathsf{TSP}}({\bf x}) + \mathsf{M}^p({\bf x}, {\bf y}).\end{equation} Combining this with the fact that when $(\Omega,\mathsf{d})$ is the unit cube $(0,1)^d$ with the Euclidean distance, $\mathbf{C}_{\mathsf{TSP}}({\bf x}) \lesssim |{\bf x}|^ {1-p/d}$ (a well-known fact, proved e.g.\ via space-filling curves) this would conclude the proof of \eqref{eq:upper-bound-deterministic}. \\ Assume without loss of generality that $|{\bf x}| = |{\bf y}| = n \ge \kappa$ and let $\rho$ be a permutation over $[n]$ that induces an optimal assignment between ${\bf x}$ and ${\bf y}$. Consider then an optimizer for the TSP over $\mathcal{K}({\bf x})$, which we also identify with a permutation $\sigma$ over $[n]$. We then define the feasible solution $G \in \mathcal{F}_{n,n}$ for the connected $\kappa$-factor problem whose edge set is $$ E_{G} = \cur{ \cur{ (1, \sigma(i)), (2, \rho(\sigma(i+\ell))} \, : \, i \in [n], \ell \in \cur{0,1,\ldots, \kappa-1}},$$ which generalizes $E_{\sigma, \tau}$ from \eqref{eq:e-sigma-tau} with $\tau= \rho \circ \sigma$ in the $\kappa=2$ case, and as in \eqref{eq:e-sigma-tau} we use the summation $\operatorname{mod} n$, i.e., $i+ \ell = i + \ell - n$ if $i+\ell>n$. Clearly, any vertex has degree $\kappa$ and the graph is connected, since $E_{G} \supset E_{\sigma, \tau}$. In follows that \[ \C_{\pP}^p({\bf x}, {\bf y}) \le \sum_{i=1}^n \sum_{\ell = 0}^{\alpha-1} \mathsf{d}(x_{\sigma(i)}, y_{\rho( \sigma(i+\ell))})^p.\] Using the triangle inequality for every $i$ and $\ell$, we bound from above \[\mathsf{d}(x_{\sigma(i)}, y_{\rho( \sigma(i+\ell))})^p \lesssim \sum_{j = 0}^{\ell-1} \mathsf{d}(x_{\sigma(i+j)}, x_{\sigma(i+j-1)})^p +\mathsf{d} ( x_{\sigma(i+\ell)}, y_{\rho( \sigma(i+\ell))})^p. \] Summation upon $i$ (keeping $\ell$ fixed) gives \[ \sum_{j = 0}^{\ell-1} \sum_{i=1}^n \mathsf{d}(x_{\sigma(i+j)}, x_{\sigma(i+j-1)})^p + \sum_{i=1}^n \mathsf{d} ( x_{\sigma(i+\ell)}, y_{\rho( \sigma(i+\ell))})^p \lesssim \ell \operatorname{TSP}({\bf x}) + \mathsf{M}^p({\bf x}, {\bf y}),\] hence, after summing upon $\ell = 0, \ldots, \kappa-1$, we obtain \eqref{eq:c-bip-c-mono}. \end{proof} \section{Convergence results for Poisson point processes}\label{sec:poisson} \subsection{Point processes} We define a point process on $\mathbb{R}^d$ as a random finite family of points $\mathcal{N} = (X_i)_{i=1}^N \subseteq \mathbb{R}^d$, i.e. a $N$-uple of random variables with values in $\mathbb{R}^d$, where the total number of points $N$ is also random and a.s.\ finite (if $N=0$, then $\mathcal{N} = \emptyset$). We extend the notation for families of points to point processes (naturally defined for each realization of the random variables): for a process $\mathcal{N} = \bra{X_i}_{i=1}^N$, write $\mu^\mathcal{N} := \sum_{i=1}^N \delta_{X_i}$ and, given a Borel $\Omega \subseteq \mathbb{R}^d$, let $\mathcal{N}(\Omega) = \mu^{\mathcal{N}}(\Omega)$ be the (random) number of variables belonging to $\Omega$, while $\mathcal{N}_{\Omega}$ denotes its restriction to $\Omega$, i.e., the collection of the variables such that $X_i \in \Omega$ (naturally re-indexed over $i=1, \ldots, \mathcal{N}(\Omega)$, with the order inherited from the original process). Given two point processes $\mathcal{N} = (X_i)_{i=1}^N$, $\mathcal{M} = (Y_j)_{j=1}^M$, their union is $\mathcal{N} \cup \mathcal{M} = (X_1, \ldots, X_N, Y_1, \ldots, Y_M)$. Given a finite Borel measure $\lambda$ on $\mathbb{R}^d$, a Poisson point process $\mathcal{N}^\lambda$ with intensity $\lambda$ can be constructed from a random collection of i.i.d.\ variables $(X_i)_{i=1}^\infty$ with common law $\lambda/\lambda(\mathbb{R}^d)$ and, after introducing a further independent Poisson variable $N^{\lambda}$ with mean $\lambda(\mathbb{R}^d)$, by considering only the first $N^{\lambda}$ variables, i.e., $$ \mathcal{N}^\lambda := (X_i)_{i=1}^{N^{\lambda}}.$$ A key property of a Poisson point process (with intensity $\lambda$) is that, given any countable Borel partition $\mathbb{R}^d = \cup_{k} \Omega_k$, the variables $(\mathcal{N}^\lambda(\Omega_k))_k$ are independent Poisson variables, each with mean $\lambda(\Omega_k)$ and, conditionally upon their value, the points in each $\Omega_k$ are i.i.d.\ variables with common probability law $\lambda\mathop{\raisebox{-.127ex}{\reflectbox{\rotatebox[origin=br]{-90}{$\lnot$}}}} \Omega_k / \lambda(\Omega_k)$. This property can be summarized by stating that the restrictions $(\mathcal{N}^\lambda_{\Omega_k})_{k}$ are independent Poisson point processes, with each $\mathcal{N}^\lambda_{\Omega_k}$ having intensity given by the restriction $\lambda\mathop{\raisebox{-.127ex}{\reflectbox{\rotatebox[origin=br]{-90}{$\lnot$}}}} \Omega_k$. We will use the well-known thinning operation, which apparently dates back to R\'enyi \cite{renyi1956characterization}, to split a Poisson point process $\mathcal{N}^\lambda$ with intensity $\lambda$ into two independent Poisson point processes, each containing approximatively a given fraction of points: for $\eta \in [0,1]$, the $\eta$-thinning of a Poisson point process $\mathcal{N}^{\lambda} = (X_i)_{i=1}^{N^\lambda}$ defines the two processes $$\mathcal{N}^{(1-\eta)\lambda} = (X_i)_{i=1}^{N^{(1-\eta)\lambda}} \quad \text{and} \quad \mathcal{N}^{\eta \rho} = (X_{N^{(1-\eta)\lambda}+i})_{i=1}^{N^\lambda-N^{(1-\eta)\lambda}},$$ where $N^{(1-\eta)\lambda} = \sum_{i=1}^{N} Z_i$ is defined using a further sequence of i.i.d.\ Bernoulli random variables $(Z_i)_{i=1}^\infty$ with $\mathbb{P}(Z_i=1)=1-\eta$ (independent from the variables $(X_i)_i$ and $N^{\lambda}$). Clearly, $\mathcal{N}^{\lambda} = \mathcal{N}^{(1-\eta)\lambda}\cup \mathcal{N}^{\eta \lambda}$, and it is straightforward to prove that both are independent and Poisson point processes with intensities respectively $(1-\eta)\lambda$ and $\eta \lambda$. \subsection{Statement} The aim of this section is to prove the analogue of \cref{thm:main} for Poisson point processes (instead of i.i.d.\ points). \begin{theorem}\label{thm:limit-poisson} Let $d \ge 3$, $p\in [1,d)$ and let $\mathsf{P} = (\mathcal{F}_{n,n})_{n \in \mathbb{N}}$ be a combinatorial optimization problem over complete bipartite graphs such that assumptions \ref{as:isomorphism}, \ref{as:spanning}, \ref{as:bddegree}, \ref{ass:local-merging} and \ref{as:growth} hold. Then, there exists $\beta_{\mathsf{P}} \in (0, \infty)$ (depending on $p$ and $d$) such that the following holds. Let $\Omega \subseteq \mathbb{R}^d$ be a bounded domain with Lipschitz boundary and such that \eqref{eq:green-kernel-bound} holds. Let $\rho$ be a H\"older continuous probability density on $\Omega$, uniformly strictly positive and bounded from above. For every $n\in (0, \infty)$, let $\mathcal{N}^{n\rho}$, $\mathcal{M}^{n\rho}$ be independent Poisson point processes with intensity $n \rho$ on $\Omega$. Then, \begin{equation}\label{eq:limsup-poisson} \limsup_{n \to \infty} n^{\frac{p}{d}-1}\mathbb{E}\sqa{ \C_{\pP}^p\bra{ \mathcal{N}^{n\rho}, \mathcal{M}^{n \rho}}} \le \beta_{\mathsf{P}} \int_{\Omega} \rho^{1-\frac{p}{d}}.\end{equation} Moreover, if $\rho$ is the uniform density and $\Omega$ is a cube or its boundary is $C^2$, then the limit exists and equals the right-hand side. \end{theorem} After having introduced some general notation and proved some basic facts, we split the proof into four main cases. We deal first with the case of a uniform density on a cube and establish existence of the limit via subadditivity. Then, we consider H\"older densities on a cube and move next to general domains. Finally, we establish existence of the limit for uniform densities on domains with $C^2$ boundary. \subsection{General facts}\label{sec:strategy} Although each case has its distinctive features, the underlying strategy is common and relies on \cref{prop:partition} in combination with a preliminary application of the thinning operation. To avoid repetitions and introduce a general notation, we give a description of the construction and show a first lemma which uses the fundamental ideas upon which we elaborate in the next sections. Let $\mathcal{N}$, $\mathcal{M}$ be two independent Poisson point processes on $\Omega$ with common intensity given by a finite measure $\lambda$. In our applications, $\lambda$ is Lebesgue measure or $\lambda = n \rho$, but for simplicity here we omit to specify it. We apply the $\eta$-thinning to $\mathcal{N} = \mathcal{N}^{1-\eta} \cup \mathcal{N}^{\eta}$, obtaining independent Poisson point processes with respective intensities $(1-\eta)\lambda$, $\eta \lambda$, and similarly to $\mathcal{M} = \mathcal{M}^{1-\eta} \cup \mathcal{M}^{\eta}$. Given a finite Borel partition $\Omega = \bigcup_{k=1}^K \Omega_k$, for each $k=1, \ldots, K$, we pick a minimizer $G_k\subseteq \mathcal{K}\bra{\mathcal{N}^{1-\eta}_{\Omega_k}, \mathcal{M}^{1-\eta}_{\Omega_k}}$ for the problem $$ \C_{\pP}^p( \mathcal{N}^{1-\eta}_{\Omega_k}, \mathcal{M}^{1-\eta}_{\Omega_k}).$$ Writing \[ Z_k = \min\cur{ |\mathcal{N}^{1-\eta}_{\Omega_k}|, |\mathcal{M}^{1-\eta}_{\Omega_k}|},\] we notice that $G_k = \emptyset$ if and only if $Z_k < \mathsf{c}_{\operatorname{A2}}$ (by \cref{rem:uniqueness-minimizer} for $p>1$, $G_k$ is a.s.\ unique. For $p=1$ we can consider a measurable selection). We define point processes $\mathcal{U}$, $\mathcal{V}$ on $\Omega$ by setting $\mathcal{U}_{\Omega_k} \subseteq \mathcal{N}^{1-\eta}_{\Omega_k}$, $\mathcal{V}_{\Omega_k} \subseteq \mathcal{M}^{1-\eta}_{\Omega_k}$, given by all the points, respectively in $\mathcal{N}^{1-\eta}_{\Omega_k}$ and $\mathcal{M}^{1-\eta}_{\Omega_k}$, which do not belong to the set of vertices of $G_k$. In particular, if $G_k = \emptyset$, then $\mathcal{U}_{\Omega_k} = \mathcal{N}^{1-\eta}_{\Omega_k}$, $\mathcal{V}_{\Omega_k} = \mathcal{M}^{1-\eta}_{\Omega_k}$. Notice that by construction the $K$ pairs of processes $\bra{ (\mathcal{U}_{\Omega_k}, \mathcal{V}_{\Omega_k})}_{k=1}^K$ are independent, but for any $k$ the two processes $\mathcal{U}_{\Omega_k}$, $\mathcal{V}_{\Omega_k}$ are not in general independent. For later use, we prove: \begin{lemma} For every $k=1, \ldots, K$ such that \begin{equation}\label{eq:condition-omega-k-not-too-small} \lambda(\Omega_k) > 4 \mathsf{c}_{\operatorname{A2}}, \end{equation} we have, for every $q \ge 1$, \begin{equation}\label{eq:bound-moment-q-remainder-uv} \mathbb{E}\sqa{ |\mathcal{U}_{\Omega_k}|^q +|\mathcal{V}_{\Omega_k}|^q}\lesssim_{q} \lambda(\Omega_k)^{\frac{q}{2}}.\end{equation} \end{lemma} \begin{proof} In the event \begin{equation*}\label{eq:fraction-points} A_k = \cur{Z_k \ge (1-\eta)\lambda(\Omega_k)/2}, \end{equation*} since $\eta \in (0,1/2)$ we have that $Z_k \ge \mathsf{c}_{\operatorname{A2}}$ hence by assumption \ref{as:spanning}, every feasible solution (in particular the optimal solution $G_k$) spans a subgraph of $\mathcal{K}_{ \mathcal{N}^{1-\eta}(\Omega_k), \mathcal{M}^{1-\eta}(\Omega_k)}$ isomorphic to $\mathcal{K}_{Z_k, Z_k}$, so that $$ |\mathcal{U}_{\Omega_k}| \le |\mathcal{N}^{1-\eta}_{\Omega_k}| - Z_k \le \abs{ |\mathcal{M}^{1-\eta}_{\Omega_k}| - |\mathcal{N}^{1-\eta}_{\Omega_k}|}.$$ Using \eqref{eq:momentPoi}, we have $$ \mathbb{E}\sqa{ |\mathcal{U}_{\Omega_k}|^q I_{A_k} } \le \mathbb{E}\sqa{ \abs{ |\mathcal{M}^{1-\eta}_{\Omega_k}| - |\mathcal{N}^{1-\eta}_{\Omega_k}|}^q} \lesssim_q \lambda(Q_L)^{\frac{q}{2}}.$$ By the union bound and \eqref{eq:density-bound-below-Poi} with $n = (1-\eta) \lambda(\Omega_k)$, $\gamma =1/2$, we have \begin{equation*}\label{eq:estimate-few-points-Ak} \begin{split} \mathbb{P}( A^c_k) & \le \mathbb{P}(|\mathcal{N}^{1-\eta}_{\Omega_k}| < (1-\eta) \lambda(\Omega_k)/2) + \mathbb{P}(|\mathcal{M}^{1-\eta}_{\Omega_k}| < (1-\eta) \lambda(\Omega_k)/2) ) \\ & \lesssim_q \lambda(\Omega_k)^{-q}. \end{split}\end{equation*} Therefore, \[\begin{split} \mathbb{E} \sqa{ |\mathcal{U}_{\Omega_k}|^q I_{A^c_k}} & \le \mathbb{E} \sqa{ |\mathcal{N}^{1-\eta}_{\Omega_k}|^q I_{A^c_k}} \le \mathbb{E}\sqa{|\mathcal{N}^{1-\eta}_{\Omega_k}|^{2q} }^{\frac{1}{2}} \mathbb{P}(A^c_k)^{\frac{1}{2}} \\ & \lesssim_q \lambda(\Omega_k)^{\frac{q}{2}}. \end{split}\] Arguing similarly for $|\mathcal{V}_{\Omega_k}|$, we obtain \eqref{eq:bound-moment-q-remainder-uv} \end{proof} For $k = 1, \ldots, K$, we define \[ {\bf x}^k = \mathcal{N}^{1-\eta}_{\Omega_k} \setminus \mathcal{U}_{\Omega_k}, \quad {\bf y}^k = \mathcal{M}^{1-\eta}_{\Omega_k} \setminus \mathcal{V}_{\Omega_k},\] so that by construction $|{\bf x}^k| = |{\bf y}^k| =n_k$, with \[ n_k = \begin{cases}Z_k & \text{if $Z_k \ge \mathsf{c}_{\operatorname{A2}}$,}\\ 0 & \text{otherwise.} \end{cases}\] Moreover, since the optimizer $G_k$ is a feasible solution in $\mathcal{K}({\bf x}^k, {\bf y}^k)$, we have $$ \C_{\pP}^p(\mathcal{N}^{1-\eta}_{\Omega_k}, \mathcal{M}^{1-\eta}_{\Omega_k}) = \C_{\pP}^p( {\bf x}^k, {\bf y}^k).$$ We then let ${\bf x}^0 = \mathcal{N}^\eta \cup \mathcal{U}$, ${\bf y}^0 = \mathcal{M}^\eta \cup \mathcal{V}$. In the event \begin{equation}\label{eq:event-many-points} \cur{ \min\cur{|\mathcal{N}^\eta|, |\mathcal{M}^\eta|} \ge \min\cur{K, \mathsf{c}_{\operatorname{A2}}}}, \end{equation} \cref{prop:partition} applies for any choice of points ${\bf z} = (z_k)_{k=1}^K$ with $z_k \in \Omega_k$, yielding the inequality \begin{equation}\label{eq:subadditive-random} \C_{\pP}^p\bra{ \mathcal{N}, \mathcal{M}} - \sum_{k=1}^K \C_{\pP}^p(\mathcal{N}^{1-\eta}_{\Omega_k}, \mathcal{M}^{1-\eta}_{\Omega_k}) \lesssim \C_{\pP}^p(\mathcal{N}^\eta \cup \mathcal{U}, \mathcal{M}^\eta \cup \mathcal{V})+ \mathsf{M}^p(\mathcal{N}^\eta \cup \mathcal{U}, {\bf z}) + \sum_{k=1}^K \operatorname{diam}(\Omega_k)^p. \end{equation} By \cref{rem:remove-z-points}, if also \begin{equation}\label{eq:event-many-points-each-omegak} \cur{ \min_{k=1,\ldots, K} \min\cur{ |\mathcal{N}^\eta_{\Omega_k}|, |\mathcal{M}^{\eta}_{\Omega_k}|} \ge 1}\end{equation} then the term $\mathsf{M}^p(\mathcal{N}^\eta \cup \mathcal{U}, {\bf z})$ can be removed in \eqref{eq:subadditive-random}. Once \eqref{eq:subadditive-random} is established, the next step is to take expectation and carefully estimate the ``error terms'' in the right-hand side. To convey the main ideas, we start with the simplest case when $K$ is kept fixed as we let $n \to \infty$ in the intensity of the process $\lambda = n \rho$. \begin{lemma}\label{lem:subadditivity-finite-partition} With the notation and assumptions of \cref{thm:limit-poisson}, fix $K \in \mathbb{N}$ and consider a Borel partition $\Omega = \bigcup_{k=1}^K \Omega_k$. Then, \begin{equation}\label{eq:limsup-poisson-finite-partition} \limsup_{n \to \infty} n^{\frac{p}{d}-1} \mathbb{E}\sqa{ \C_{\pP}^p\bra{ \mathcal{N}^{n\rho}, \mathcal{M}^{n \rho}}} \le \sum_{k=1}^K \limsup_{n \to \infty} n^{\frac{p}{d}-1}\mathbb{E}\sqa{ \C_{\pP}^p\bra{ \mathcal{N}^{n\rho}_{\Omega_k}, \mathcal{M}^{n \rho}_{\Omega_k}}} .\end{equation} \end{lemma} \begin{proof} We can assume that each $\Omega_k$ is not negligible. Then, condition \eqref{eq:condition-omega-k-not-too-small} with $\lambda = n \rho$ holds if $n$ is sufficiently large. Letting \begin{equation}\label{eq:A-finiteK} A = \bigcap_{k=1}^K \cur{ \min\cur{|\mathcal{N}^{n\eta\rho}_{\Omega_k}|, |\mathcal{M}^{n\eta\rho}_{\Omega_k}|} \ge \mathsf{c}_{\operatorname{A2}}} =\bigcap_{k=1}^K A_k,\end{equation} we have that both \eqref{eq:event-many-points} and \eqref{eq:event-many-points-each-omegak} hold $A$. By the union bound in combination with \eqref{eq:density-bound-below-Poi}, we estimate, for every $q \ge 1$, \begin{equation}\label{eq:union-bound} \mathbb{P}(A^c) \le \sum_{k=1}^K \mathbb{P}(A^c_k) \lesssim_{q,\eta, K} n^{-q}.\end{equation} Combined with the trivial inequality $\C_{\pP}^p\bra{ \mathcal{N}^{n\rho}, \mathcal{M}^{n\rho}} \lesssim |\mathcal{N}^{n\rho}|$ we obtain that \[ \mathbb{E}\sqa{ \C_{\pP}^p\bra{ \mathcal{N}^{n\rho}, \mathcal{M}^{n\rho}} I_{A^c} } \le \mathbb{E}\sqa{ |\mathcal{N}^{n\rho}|^2}^{\frac{1}{2}} \mathbb{P}(A^c)^{\frac{1}{2}} \lesssim_{q,K} n^{\frac{1-q}{2}},\] which is infinitesimal if $q>1$ (even without dividing by $n^{1-p/d}$). Therefore, \begin{equation*} \label{eq:Ac-does-not-matter} \limsup_{n \to \infty} n^{\frac{p}{d}-1}\mathbb{E}\sqa{ \C_{\pP}^p\bra{ \mathcal{N}^{n\rho}, \mathcal{M}^{n \rho}} I_A} = \limsup_{n \to \infty} n^{\frac{p}{d}-1}\mathbb{E}\sqa{ \C_{\pP}^p\bra{ \mathcal{N}^{n\rho}, \mathcal{M}^{n \rho}} } \end{equation*} and we only need to prove the following inequality, for fixed $\eta$, \begin{equation*}\label{eq:limsup-poisson-finite-partition-eta} \limsup_{n \to \infty} n^{\frac{p}{d}-1}\mathbb{E}\sqa{ \C_{\pP}^p\bra{ \mathcal{N}^{n\rho}, \mathcal{M}^{n \rho}} I_A} - \sum_{k=1}^K \limsup_{n \to \infty} n^{\frac{p}{d}-1}\mathbb{E}\sqa{ \C_{\pP}^p\bra{ \mathcal{N}^{n\rho}_{\Omega_k}, \mathcal{M}^{n \rho}_{\Omega_k}}} \lesssim_{K} \eta^{1-\frac{p}{d}}, \end{equation*} and finally let $\eta\to 0$ to obtain the thesis. To this aim, we multiply \eqref{eq:subadditive-random} by $I_A$ and take expectation, obtaining the inequality \begin{multline}\label{eq:exp-sub-explicit} \mathbb{E}\sqa{ \C_{\pP}^p\bra{ \mathcal{N}^{n\rho}, \mathcal{M}^{n\rho}} I_{A} } - \sum_{k=1}^K \mathbb{E}\sqa{ \C_{\pP}^p\bra{ \mathcal{N}^{(1-\eta)n\rho}_{\Omega_k}, \mathcal{M}^{(1-\eta)n\rho}_{\Omega_k}} } \\ \lesssim \mathbb{E}\sqa{ \C_{\pP}^p\bra{ \mathcal{N}^{\eta n\rho}\cup \mathcal{U}, \mathcal{M}^{\eta n\rho}\cup \mathcal{V}}}+K. \end{multline} Since, for each $k=1, \ldots, K$, \begin{equation*}\label{eq:exp-sub-finite-partition} \begin{split} \limsup_{n \to \infty} n^{\frac{p}{d}-1}\mathbb{E}&\sqa{ \C_{\pP}^p\bra{ \mathcal{N}^{(1-\eta)n\rho}_{\Omega_k}, \mathcal{M}^{(1-\eta)n\rho}_{\Omega_k}} } \\ & = (1-\eta)^{1-\frac{p}{d}} \limsup_{n \to \infty} \bra{(1-\eta)n}^{\frac{p}{d}-1}\mathbb{E}\sqa{ \C_{\pP}^p\bra{ \mathcal{N}^{(1-\eta)n\rho}_{\Omega_k}, \mathcal{M}^{(1-\eta)n\rho}_{\Omega_k}} }\\ & \le \limsup_{n \to \infty} n^{\frac{p}{d}-1}\mathbb{E}\sqa{ \C_{\pP}^p\bra{ \mathcal{N}^{n\rho}_{\Omega_k}, \mathcal{M}^{n\rho}_{\Omega_k}} }, \end{split}\end{equation*} we need to focus only on the terms in the right-hand side of \eqref{eq:exp-sub-explicit}. Since the last term is constant, we are left with the proof of \begin{equation}\label{tofinishfinite} \limsup_{n \to \infty} n^{\frac{p}{d}-1}\mathbb{E}\sqa{ \C_{\pP}^p\bra{ \mathcal{N}^{\eta n\rho}\cup \mathcal{U}, \mathcal{M}^{\eta n\rho}\cup \mathcal{V}} }\lesssim \eta^{1-\frac{p}{d}}. \end{equation} We first notice that by \eqref{eq:bound-moment-q-remainder-uv} and H\"older inequality we have for every $q\ge 1$, \begin{equation}\label{momentuVfinite} \mathbb{E}\sqa{ |\mathcal{U}|^q +|\mathcal{V}|^q}\lesssim_{q} K^{\frac{q}{2}}n^{\frac{q}{2}}. \end{equation} We now use assumption~\ref{as:growth} so that \begin{equation}\label{eq:last-term-finite-partition} \mathbb{E}\sqa{ \C_{\pP}^p\bra{ \mathcal{N}^{\eta n\rho}\cup \mathcal{U}, \mathcal{M}^{\eta n\rho}\cup \mathcal{V}}} \lesssim \mathbb{E}\sqa{ | \mathcal{N}^{\eta n\rho}\cup \mathcal{U}|^{1-\frac{p}{d}}} + \mathbb{E}\sqa{ \mathsf{M}^p\bra{ \mathcal{N}^{\eta n\rho}\cup \mathcal{U}, \mathcal{M}^{\eta n\rho}\cup \mathcal{V}}}. \end{equation} To estimate the first term in the right-hand side, we use H\"older inequality and \eqref{momentuVfinite} with $q=1$, \begin{equation}\label{eq:last-term-easy-finite-partition} \begin{split} \mathbb{E}\sqa{| \mathcal{N}^{\eta n\rho}\cup \mathcal{U}|^{1-\frac{p}{d}}} & \lesssim \mathbb{E}\sqa{ |\mathcal{N}^{\eta n \rho}|}^{1-\frac{p}{d}} + \mathbb{E}\sqa{ |\mathcal{U}|}^{1-\frac{p}{d}}\\ & \lesssim n^{1-\frac{p}{d}}\bra{\eta^{1-\frac{p}{d}} + C_K n^{-\frac{1}{2}(1-\frac{p}{d})}}.\end{split}\end{equation} \noindent For the second term, thanks to \eqref{momentuVfinite} we may use \cref{prop:density-helps-matching} with $H=n^{1/2}$ and $h=\min\cur{\mathbb{E}\sqa{|\mathcal{N}^{\eta n\rho}|}, \mathbb{E}\sqa{|\mathcal{M}^{\eta n\rho}|}}\sim n\eta$ so that for some $\alpha<2$ and $\beta>0$ \[ \mathbb{E}\sqa{ \mathsf{M}^p\bra{ \mathcal{N}^{\eta n\rho}\cup \mathcal{U}, \mathcal{M}^{\eta n\rho}\cup \mathcal{V}}}\lesssim n^{1-\frac{p}{d}}\bra{\eta^{1-\frac{p}{d}} + C_{K,\eta} n^{-\frac{\beta}{2} (2-\alpha)}}. \] Plugging this and \eqref{eq:last-term-easy-finite-partition} in \eqref{eq:last-term-finite-partition} concludes the proof of \eqref{tofinishfinite}. \end{proof} \begin{remark}\label{rem:limit-case-finite-partition} We notice that the proof above yields also the inequality \begin{equation}\begin{split}\label{eq:liminf-poisson-finite-partition} \liminf_{n \to \infty} n^{\frac{p}{d}-1}\mathbb{E}\sqa{ \C_{\pP}^p\bra{ \mathcal{N}^{n\rho}, \mathcal{M}^{n \rho}}} & \le \liminf_{n \to \infty} n^{\frac{p}{d}-1}\mathbb{E}\sqa{ \C_{\pP}^p\bra{ \mathcal{N}^{n\rho}_{\Omega_1}, \mathcal{M}^{n \rho}_{\Omega_1}}} \\ & \quad + \sum_{k=2}^K \limsup_{n \to \infty} n^{\frac{p}{d}-1} \mathbb{E}\sqa{ \C_{\pP}^p\bra{ \mathcal{N}^{n\rho}_{\Omega_k}, \mathcal{M}^{n \rho}_{\Omega_k}}} .\end{split}\end{equation} This follows by repeating the argument only along a subsequence $n_\ell \to \infty$ such that \[ \lim_{\ell \to \infty} n_{\ell}^{\frac{p}{d}-1}\mathbb{E}\sqa{ \C_{\pP}^p\bra{ \mathcal{N}^{n_\ell \rho}_{\Omega_1}, \mathcal{M}^{n_\ell \rho}_{\Omega_1}}} = \liminf_{n \to \infty} n^{\frac{p}{d}-1}\mathbb{E}\sqa{ \C_{\pP}^p\bra{ \mathcal{N}^{n\rho}_{\Omega_1}, \mathcal{M}^{n \rho}_{\Omega_1}}}.\] \end{remark} \subsection{Uniform density on a cube}\label{sec:cube-constant-density} In this section we consider the case of a uniform measure on a cube. Up to rescaling (see \eqref{eq:homogeneity}) it is equivalent to consider two independent Poisson point processes $\mathcal{N}_{Q_L}$ and $\mathcal{M}_{Q_L}$ with intensity one on $Q_L$ and prove that \[ f(L) =\frac{1}{|Q_L|} \mathbb{E}\sqa{ \C_{\pP}^p\bra{ \mathcal{N}_{Q_{L}}, \mathcal{M}_{Q_L}}} \] has a limit as $L\to \infty$. \begin{proposition}\label{prop:uniform} Let $d \ge 3$, $p\in [1,d)$ and let $\mathsf{P} = (\mathcal{F}_{n,n})_{n \in \mathbb{N}}$ be a combinatorial optimization problem over complete bipartite graphs such that assumptions \ref{as:isomorphism}, \ref{as:spanning}, \ref{as:bddegree}, \ref{ass:local-merging} and \ref{as:growth} hold. Then, there exists $\beta_{\mathsf{P}} \in (0, \infty)$ (depending on $p$ and $d$) such that \begin{equation}\label{eq:limit-cube} \lim_{L \to \infty} f(L) = \beta_{\mathsf{P}}. \end{equation} \end{proposition} \begin{proof} We split the proof into several steps. In the first two steps we establish basic properties of $f$, before moving to the main argument. This follows the strategy of the previous section and ultimately relies upon an application of \cref{lem:sub}.\\ \medskip \setcounter{proof-step}{0} \noindent{\emph{Step \stepcounter{proof-step}\arabic{proof-step}. Continuity and upper bound.}} Writing $z = \min\cur{n,m}$, we first notice that by Assumption \ref{as:growth} and \eqref{eq:matching-iid} of \cref{prop:matching-iid}, \[\mathbb{E}\sqa{ \C_{\pP}^p \bra{ (X_i)_{i=1}^n, (Y_j)_{j=1}^m}} \lesssim z^{1-\frac{p}{d}} + \mathbb{E}\sqa{ \mathsf{M}^p\bra{ (X_i)_{i=1}^n, (Y_j)_{j=1}^m}} \lesssim z^{1-\frac{p}{d}}.\] This proves on the one hand that $f$ is bounded from above as \begin{equation}\label{eq:uniform-bound-f} f(L) \lesssim L^{p-d} \mathbb{E}\sqa{ \min\cur{ |\mathcal{N}_{Q_L}|, |\mathcal{M}_{Q_L}|}^{1-\frac{p}{d}} }\lesssim L^{p-d} \mathbb{E}\sqa{ |\mathcal{N}_{Q_L}| }^{1-\frac{p}{d}} \lesssim 1.\end{equation} On the other hand, combining it with dominated convergence it also gives continuity of $f$ thanks to the representation formula \begin{equation*}\label{eq:conditioning} \begin{split} f(L) & = \sum_{n, m=0}^\infty \frac{1}{L^d} \mathbb{E}\sqa{\C_{\pP}^p(\mathcal{N}_{Q_{L}}, \mathcal{M}_{Q_L}) \Big | |\mathcal{N}_{Q_L}|=n, |\mathcal{M}_{Q_L}|=m} e^{-2L^d} \frac{ (L^d)^{n+m}}{n! m!}\\ & = L^{p-d} \sum_{n, m=0}^\infty \mathbb{E}\sqa{\C_{\pP}^p \bra{ (X_i)_{i=1}^n, (Y_j)_{j=1}^m}} e^{-2L^d} \frac{ (L^d)^{n+m}}{n! m!} \end{split}\end{equation*} where $(X_i)_{i=1}^n$, $(Y_j)_{j=1}^m$ are i.i.d.\ points on $Q_1$. We also notice that by a simple scaling argument, if $\mathcal{N}^{\lambda}$, $\mathcal{M}^{\lambda}$ are independent Poisson processes of intensity $\lambda>0$ on $Q_L$ then \begin{equation}\label{eq:homogeneity} \frac{1}{|Q_L|} \mathbb{E}\sqa{ \C_{\pP}^p\bra{ \mathcal{N}_{Q_{L}}^{\lambda}, \mathcal{M}_{Q_L}^{\lambda} }} = \frac{\lambda^{1-\frac{p}{d}} }{|Q_{\lambda L}|} \mathbb{E}\sqa{ \C_{\pP}^p\bra{ \mathcal{N}_{Q_{\lambda^{\frac{1}{d}} L}}^1, \mathcal{M}_{Q_{\lambda^{\frac{1}{d}} L}}^1}} = \lambda^{1-\frac{p}{d}} f(\lambda^{\frac{1}{d}} L).\end{equation} Combined with \eqref{eq:uniform-bound-f}, it yields that for any cube $Q$ and $\lambda>0$, \begin{equation}\label{eq:general-upper-bound-cube} \mathbb{E}\sqa{ \C_{\pP}^p\bra{ \mathcal{N}_{Q}^{\lambda}, \mathcal{M}_{Q}^{\lambda} }} \lesssim |Q| \lambda^{1-\frac{p}{d}}. \end{equation} \noindent{\emph{Step \stepcounter{proof-step}\arabic{proof-step}. Lower bound.}} The spanning assumption \ref{as:spanning} yields that, if e.g.\ $\mathsf{c}_{\operatorname{A2}} \le n \le m$, then $$ \C_{\pP}^p\bra{ (X_i)_{i=1}^n, (Y_j)_{j=1}^m} \ge \sum_{i=1}^n \min_{j=1, \ldots, m} |X_i - Y_j|^p.$$ The following classical lower bound, e.g.\ proved in \cite[Chapter 2]{steele1997probability}, $$ \mathbb{E}\sqa{ \min_{j=1, \ldots, m} |X_i - Y_j|^p } \gtrsim m^{-\frac{p}{d}}$$ entails that $$ \C_{\pP}^p\bra{ (X_i)_{i=1}^n, (Y_j)_{j=1}^m} \gtrsim m^{-\frac{p}{d}} \cdot n.$$ Writing $Z = \min\cur{|\mathcal{N}_{Q_L}|, |\mathcal{M}_{Q_L}|}$, we deduce that $$ f(L) \gtrsim L^{p-d} \mathbb{E}\sqa{ \max\cur{|\mathcal{N}_{Q_L}|,|\mathcal{M}_{Q_L}|}^{-\frac{p}{d}} Z I_{\cur{ Z\ge \mathsf{c}_{\operatorname{A2}}}}}.$$ Let $$ A = \cur{ |Q_L|/2 \le Z \le \max\cur{|\mathcal{N}_{Q_L}|, |\mathcal{M}_{Q_L}|} \le 3|Q_L|/2}.$$ By \eqref{eq:density-bound-below-Poi} with $\eta=1/2$, we have $\mathbb{P}(A) \to 1$ as $L \to \infty$. Therefore id $L$ is large enough, \[ f(L) \gtrsim L^{p-d} \mathbb{E}\sqa{ \max\cur{|\mathcal{N}_{Q_L}|,|\mathcal{M}_{Q_L}|}^{-\frac{p}{d}} Z I_{A}} \gtrsim L^{p-d} \mathbb{E}\sqa{ L^{d-p} I_{A}} \gtrsim 1. \] In the remaining steps we prove the following claim. There exists $\beta=\beta(p,d)>0$ such that for every $\eta\in(0,1/2)$, there exists $C(\eta)>0$ such that, for every $m \in \mathbb{N}$, $m \ge 1$ and $L \ge C(\eta)$, \begin{equation}\label{toprovemonotonicity} f(mL)-f( (1-\eta)L)\lesssim \eta^{1-\frac{p}{d}} + C(\eta) L^{-\beta}. \end{equation} This would conclude the proof of \eqref{eq:limit-cube} by \cref{lem:sub}. \noindent{\emph{Step \stepcounter{proof-step}\arabic{proof-step}. Partitioning and exclusion of the event in which few points are sampled.}} Using the notation from \cref{sec:strategy}, we partition $\Omega = Q_{mL}$ into $K=m^d$ cubes $Q_i = Q_L + L z_i \subseteq Q_{mL}$ with $z_i\in \mathbb{Z}^d$ and two independent Poisson processes $\mathcal{N}$, $\mathcal{M}$ of unit intensity on $Q_{mL}$. We first reduce to the event $$ A = \cur{ \min\cur{ |\mathcal{N}^{\eta}_{Q_{mL}}|, |\mathcal{M}^{\eta}_{Q_{mL}}|} \ge \eta |Q_{mL}|/2 },$$ which contains \eqref{eq:event-many-points} provided $L$ is sufficiently large (depending on $\eta$ only, not on $m$). We first argue that $A^c$ is of small probability. Indeed, using a union bound we find that for every $q \ge 1$, \[ \mathbb{P}(A^c) \le \mathbb{P}\bra{ |\mathcal{N}^{\eta}_{Q_{mL}}| < |Q_{mL}|/2 } +\mathbb{P}\bra{ |\mathcal{M}^{\eta}_{Q_{mL}}| < |Q_{mL}|/2 } \stackrel{\eqref{eq:density-bound-below-Poi}}{\lesssim_{\eta,q}} |Q_{mL}|^{-q}.\] If $A^c$ holds, we use the trivial bound that follows from Assumption \ref{as:bddegree}: $$ \C_{\pP}^p( \mathcal{N}_{Q_{mL}}, \mathcal{M}_{Q_{mL}}) \lesssim |\mathcal{N}_{Q_{mL}}| |Q_{mL}|^{\frac{p}{d}},$$ so that for any given $\beta>0$ and provided we choose $q$ sufficiently large. \begin{equation}\label{eq:bad-set} \begin{split} \frac 1 {|Q_{mL}|} \mathbb{E}\sqa{ \C_{\pP}^p( \mathcal{N}_{Q_{mL}}, \mathcal{M}_{Q_{mL}}) I_{A^c}} & \lesssim |Q_{mL}|^{\frac{p}{d}-1} \mathbb{E}\sqa{ |\mathcal{N}_{Q_{mL}}|^2}^{\frac{1}{2}} \mathbb{P}(A^c)^{\frac{1}{2}} \\ & \lesssim_{\eta,q} |Q_{mL}|^{\frac{p}{d}-1} |Q_{mL}| \cdot |Q_{mL}|^{-q} \lesssim_\eta L^{-\beta}. \end{split}\end{equation} \noindent If $A$ holds, letting ${\bf z}$ be the set of centres of the $m^d$ cubes, inequality \eqref{eq:subadditive-random} reads \begin{equation}\label{eq:sub-real} \begin{split} \C_{\pP}^p( \mathcal{N}_{Q_{mL}}, \mathcal{M}_{Q_{mL}}) - \sum_{i=1}^{m^d} \C_{\pP}^p(\mathcal{N}^{1-\eta}_{Q_i}, \mathcal{M}^{1-\eta}_{Q_i}) & \lesssim \C_{\pP}^p(\mathcal{N}^\eta\cup \mathcal{U}, \mathcal{M}^\eta \cup \mathcal{V}) \\ & \quad + \mathsf{M}^p(\mathcal{N}^\eta \cup \mathcal{U}, {\bf z}) + m^d L^p. \end{split}\end{equation} Notice that by the properties of the Poisson point process, the law of $\C_{\pP}^p(\mathcal{N}^{1-\eta}_{Q_i}$, $\mathcal{M}^{1-\eta}_{Q_i})$ equals that of $\C_{\pP}^p(\mathcal{N}^{1-\eta}_{Q_L}$, $\mathcal{M}^{1-\eta}_{Q_L})$. In particular \[\begin{split} \frac{1}{|Q_L|} \mathbb{E}\sqa{\C_{\pP}^p(\mathcal{N}^{1-\eta}_{Q_i}, \mathcal{M}^{1-\eta}_{Q_i})I_A} & \le \frac{1}{|Q_L|} \mathbb{E}\sqa{\C_{\pP}^p(\mathcal{N}^{1-\eta}_{Q_i}, \mathcal{M}^{1-\eta}_{Q_i})} \\ & = \frac{1}{|Q_L|} \mathbb{E}\sqa{\C_{\pP}^p(\mathcal{N}^{1-\eta}_{Q_L}, \mathcal{M}^{1-\eta}_{Q_L})}\\ & \stackrel{\eqref{eq:homogeneity}}{=} (1-\eta)^{1-\frac{p}{d}}f((1-\eta)^{\frac{1}{d}}L) \le f((1-\eta)^{\frac{1}{d}}L).\end{split}\] We thus obtain from \eqref{eq:sub-real}, \begin{multline*} \frac{1}{|Q_{mL}|} \mathbb{E}\sqa{ \C_{\pP}^p( \mathcal{N}_{Q_{mL}}, \mathcal{M}_{Q_{mL}}) I_A} - f((1-\eta)^{\frac{1}{d}}L) \lesssim \frac{1}{|Q_{mL}|}\mathbb{E}\sqa{ \C_{\pP}^p(\mathcal{N}^\eta\cup \mathcal{U}, \mathcal{M}^\eta \cup \mathcal{V})}\\ + \frac{1}{|Q_{mL}|} \mathbb{E}\sqa{ \mathsf{M}^p(\mathcal{N}^\eta \cup \mathcal{U}, {\bf z}) I_A} + L^{p-d}. \end{multline*} In the final two steps we prove that \begin{equation}\label{claimMz} \frac{1}{|Q_{mL}|} \mathbb{E}\sqa{ \mathsf{M}^p(\mathcal{N}^\eta \cup \mathcal{U}, {\bf z}) I_A}\lesssim L^{p-d} \end{equation} and \begin{equation}\label{claimNM} \frac{1}{|Q_{mL}|}\mathbb{E}\sqa{ \C_{\pP}^p(\mathcal{N}^\eta\cup \mathcal{U}, \mathcal{M}^\eta \cup \mathcal{V})}\lesssim \eta^{1-\frac{p}{d}} + C(\eta) L^{-\beta}. \end{equation} In combination with \eqref{eq:bad-set} this would conclude the proof of \eqref{toprovemonotonicity}.\\ \noindent{\emph{Step \stepcounter{proof-step}\arabic{proof-step}. Proof of \eqref{claimMz}.}} On $A$, we have $|\mathcal{N}^{\eta}_{Q_{mL}}| \ge m^d$, thus (randomly) choosing $m^d$ points from $\mathcal{N}^{\eta}$, we find after relabelling a family $(X_i)_{i=1}^{m^d}$ of points i.i.d.\ and uniformly distributed on $Q_{mL}$. Recalling that ${\bf z}$ denotes the set of centres of the $m^d$ cubes $Q_i$ we can bound \[ \mathbb{E}\sqa{ \mathsf{M}^p(\mathcal{N}^\eta \cup \mathcal{U}, {\bf z}) I_A} \le \mathbb{E}\sqa{ \mathsf{M}^p((X_i)_{i=1}^{m^d}, {\bf z})}.\] We then use \eqref{eq:matching-below-wasserstein} with $n=m^d$ and $\lambda$ the uniform density on the cube $Q_{mL}$, so that \[ \mathbb{E}\sqa{ \mathsf{M}^p((X_i)_{i=1}^{m^d}, {\bf z})} \lesssim \mathbb{E}\sqa{ \mathsf{W}^p_{Q_{mL}}\bra{ \sum_{i=1}^{m^d} \delta_{X_i}, \frac{m^d}{|Q_{mL}|}}} + \mathbb{E}\sqa{ \mathsf{W}^p_{Q_{mL}}\bra{ \mu^{\bf z}, \frac{m^d}{|Q_{mL}|}}} \lesssim m^d L^p,\] having used \eqref{eq:transport-iid} to bound the first term (the second term is trivially estimated by transporting the mass on each cube $Q_i$ to its center). This proves \eqref{claimMz}.\\ \medskip \noindent{\emph{Step \stepcounter{proof-step}\arabic{proof-step}. Proof of \eqref{claimNM}.}} We use Assumption \ref{as:growth} (on $Q_{mL}$ instead of $Q_1$, see \cref{remgrowth}), so that \[ \begin{split} \mathbb{E}\sqa{ \C_{\pP}^p(\mathcal{N}^\eta\cup \mathcal{U}, \mathcal{M}^\eta \cup \mathcal{V}) } &\lesssim (mL)^p\mathbb{E}\sqa{ (|\mathcal{N}^\eta_{Q_{mL}}|+|\mathcal{U}|)^{1-\frac{p}{d}} } + \mathbb{E}\sqa{ \mathsf{M}^p(\mathcal{N}^\eta\cup \mathcal{U}, \mathcal{M}^\eta \cup \mathcal{V})}. \end{split}\] We further bound the first contribution using H\"older inequality \begin{equation}\label{eq:first-contribution-cube}\begin{split} \mathbb{E}\sqa{ (|\mathcal{N}^\eta_{Q_{mL}}|+|\mathcal{U}|)^{1-\frac{p}{d}} }& \le \mathbb{E}\sqa{ |\mathcal{N}^{\eta}_{Q_{mL}}|}^{1-\frac{p}{d}}+ \mathbb{E}\sqa{ |\mathcal{U}|}^{1-\frac{p}{d}} \\ & \lesssim \eta^{1-\frac{p}{d}} (mL)^{d-p} + \mathbb{E}\sqa{ |\mathcal{U}|}^{1-\frac{p}{d}}. \end{split}\end{equation} To proceed further, let us recall that in \cref{sec:strategy} we argued that $\bra{ \bra{\mathcal{U}_{Q_i}, \mathcal{V}_{Q_i}}}_{i=1}^{m^d}$ are independent (and also independent from $\mathcal{N}^{\eta}$, $\mathcal{M}^{\eta}$). Moreover, since the law of each $\bra{\mathcal{N}^{1-\eta}_{Q_i}, \mathcal{M}^{1-\eta}_{Q_i}}$ coincides with that of $\bra{\mathcal{N}^{1-\eta}_{Q_L}, \mathcal{M}^{1-\eta}_{Q_L}}$ (up to a translation by $-Lz_i$, since $Q_i = Q_L+Lz_i$) it follows that the same property holds for the processes $\bra{\mathcal{U}_{Q_i}, \mathcal{V}_{Q_i}}$: their law coincides with that of $\bra{\mathcal{U}_{Q_L}, \mathcal{V}_{Q_L}}$ (also up to translating by $-Lz_i$). Using \eqref{eq:bound-moment-q-remainder-uv} with $q=1$, we obtain \[ \mathbb{E}\sqa{ |\mathcal{U}|} = m^d \mathbb{E}\sqa{ |\mathcal{U}_{Q_L}|} \lesssim m^d L^{\frac{d}{2}},\] thus \eqref{eq:first-contribution-cube} yields \[ \frac{(mL)^p}{|Q_{mL}|}\mathbb{E}\sqa{(|\mathcal{N}^\eta_{Q_{mL}}|+|\mathcal{U}|)^{1-\frac{p}{d}} } \lesssim \eta^{1-\frac{p}{d}} + L^{\frac{p-d}{2}}.\] Combining this with \cref{thm:bound-matching}, concludes the proof of \eqref{claimNM}. \end{proof} \subsection{H\"older density on a cube} In this section, we still assume that $\Omega = Q$ is a cube, but consider the case of a general H\"older continuous density $\rho$, uniformly bounded from above and below. Up to rescaling and translation, it is sufficient to consider the case $\Omega = (0,1)^d$. The proof of \eqref{eq:limsup-poisson} in this case is obtained by combining the case of constant density treated above together with \cref{lem:subadditivity-finite-partition} and the following claim: there exists a constant $C = C(\rho)>0$ such that, for $r<C$ and for every cube $Q \subseteq (0,1)^d$ with side length $r$ the following inequality holds: \begin{equation}\label{eq:inequality-cube-holder-iid} \mathbb{E}\sqa{ \C_{\pP}^p\bra{ \mathcal{N}^{n \rho}_{Q}, \mathcal{M}^{n \rho}_{Q} }}\le (1+C^{-1} r^\alpha) \mathbb{E}\sqa{ \C_{\pP}^p\bra{ \mathcal{N}^{n \rho(Q)/r^d}, \mathcal{M}^{n \rho(Q)/r^d}}}, \end{equation} where $\mathcal{N}^{n \rho(Q)/r^d}$, $\mathcal{M}^{n \rho(Q)/r^d}$ are two independent Poisson point processes with constant intensity $n \rho(Q)/r^d$ on the cube $(0,r)^d$, and $\alpha$ denotes the H\"older exponent of $\rho$. Indeed, assume that the claim holds and let us prove \eqref{eq:limsup-poisson}. Given any $r< C(p,\rho)$ of the form $r = 1/K^{1/d}$, we consider a partition of $(0,1)^d = \bigcup_{k=1}^K Q_k$ into $K$ disjoint sub-cubes of side length $r$, so that \[\begin{split} \limsup_{n \to \infty} n^{\frac{p}{d}-1} \mathbb{E} & \sqa{ \C_{\pP}^p\bra{ \mathcal{N}^{n \rho}, \mathcal{M}^{n \rho} }} \\ & \stackrel{\eqref{eq:limsup-poisson-finite-partition}}{\le} \sum_{k=1}^{K} \limsup_{n \to \infty}n^{\frac{p}{d}-1}\mathbb{E}\sqa{ \C_{\pP}^p\bra{ \mathcal{N}^{n \rho}_{Q_k}, \mathcal{M}^{n \rho}_{Q_k} }}\\ & \stackrel{\eqref{eq:inequality-cube-holder-iid}}{\le} (1+C^{-1} r^\alpha)\sum_{k=1}^{K} \limsup_{n \to \infty}n^{\frac{p}{d}-1}\mathbb{E}\sqa{ \C_{\pP}^p\bra{ \mathcal{N}^{n \rho(Q_k)/r^d}, \mathcal{M}^{n \rho(Q_k)/r^d} }}\\ & = \beta_{\mathsf{P}} (1+C^{-1} r^\alpha) \sum_{k=1}^{K} \rho(Q_k)^{1-\frac{p}{d}} r^p, \end{split}\] where the last line follows from \eqref{eq:limsup-poisson} in the case of a cube and constant intensity. Letting $K \to \infty$, we have that $r\to 0$ and the easily seen convergence \[ \lim_{K \to \infty} \sum_{k=1}^{K} \rho(Q_k)^{1-\frac{p}{d}} r^p = \lim_{K \to \infty}\int_{(0,1)^d} \sum_{k=1}^K I_{Q_k} \bra{ \frac{ \rho(Q_k)}{r^d}}^{-\frac{p}{d}} \rho = \int_{(0,1)^d} \rho^{1-\frac{p}{d}}.\] This would conclude the proof of \eqref{eq:limsup-poisson} also in this case. We now prove \eqref{eq:inequality-cube-holder-iid} for which we closely follow \cite[Lemma 2.5]{ambrosio2022quadratic}. Up to translating, we may assume that $Q = (0,r)^d$. We write $\rho_0 = \min_{(0,1)^d} \rho$ and define $\rho^r(x) = \rho(rx)r^d/\rho(Q)$ for $x \in (0,1)^d$, so that $\int_{(0,1)^d} \rho^r = 1$, and for every $x$, $y\in (0,1)^d$, \[ \rho^r(x) - \rho^r(y) \le \frac{ \nor{\rho}_{C^\alpha}}{\bar{\rho}} r^\alpha |x-y|^\alpha,\] thus $\nor{ \rho^r - 1 }_{C^\alpha}\lesssim r^\alpha$ if $r$ is sufficiently small. We define $S: Q \to Q$ as $S(x) = r T^{-1}(x/r)$, where $T$ is the map provided by \cref{prop:map-heat-semigroup}. It holds $\operatorname{Lip} S = \operatorname{Lip} T^{-1}$, and $S_\sharp 1/r^d = \rho/\rho(Q)$. Therefore, $S\bra{ \mathcal{N}^{n\rho(Q)/r^d}}= (S(X_i))_{i=1}^{N^{n\rho(Q)/r^d}(Q)}$, which is a Poisson point process on $Q$ with intensity $n \rho$, i.e., it has the same law as $\mathcal{N}^{n \rho}_Q$, and similarly $S\bra{\mathcal{M}^{n\rho(Q)/r^d}}$ has the same law as $\mathcal{M}^{n\rho}_Q$. Therefore, \[\begin{split} \mathbb{E}\sqa{ \C_{\pP}^p\bra{ \mathcal{N}^{n \rho}_{Q}, \mathcal{M}^{n \rho}_{Q} }}& = \mathbb{E}\sqa{ \C_{\pP}^p\bra{ S(\mathcal{N}^{n \rho(Q)/r^d}), S(\mathcal{M}^{n \rho(Q)/r^d}) }}\\ & \stackrel{\eqref{eq:lipschitz-bound-deterministic}}{\le} (\operatorname{Lip} S)^p \mathbb{E}\sqa{ \C_{\pP}^p\bra{ \mathcal{N}^{n \rho(Q)/r^d}, \mathcal{M}^{n \rho(Q)/r^d}}}. \end{split}\] This proves the claim since $(\operatorname{Lip} S)^p = (\operatorname{Lip} T^{-1})^p \le 1+C r^\alpha$ if $r$ is sufficiently small. \begin{remark}\label{rem:general-domain-holder-density} Let us notice that the fact that $\Omega$ is a cube is not used in the proof of \eqref{eq:inequality-cube-holder-iid}, which therefore holds true for every bounded domain $\Omega$ and H\"older continuous density $\rho$ uniformly bounded from above and below. In particular, combining \eqref{eq:inequality-cube-holder-iid} with \eqref{eq:general-upper-bound-cube} we obtain that there exists $C=C(\rho)>0$ such that, for every cube $Q \subseteq \Omega$ with side length $r<C$, \begin{equation}\label{eq:inequality-cube-holder-general} \mathbb{E}\sqa{ \C_{\pP}^p\bra{ \mathcal{N}^{n \rho}_{Q}, \mathcal{M}^{n \rho}_{Q} }}\lesssim |Q| n^{1-\frac{p}{d}}, \end{equation} where the implicit constant depends on $p$, $d$ and $\rho$ only. \end{remark} \subsection{General density on a domain} We prove \eqref{eq:limsup-poisson} for a domain $\Omega$ and a H\"older density $\rho$. The main difficulty here is that since we rely on the result established in the previous section we need to partition $\Omega$ into cubes. This is accomplished relying on the Whitney-type decomposition provided by \cref{lem:decomp}. We begin by fixing a Whitney decomposition $\mathcal{Q} =(Q_i)_i$ such that every cube $Q_i$ has side length $r< C$, where $C=C(\rho)$ is as in \cref{rem:general-domain-holder-density}. Then, by \cref{lem:decomp}, for every sufficiently small $\delta>0$ we have a finite Borel partition of $\Omega = \bigcup_{k=1}^K \Omega_k$, whose elements are collected into the two disjoint sets $\mathcal{Q}_{\delta}$, $\mathcal{R}_{\delta}$. We fix $\eta \in (0,1/2)$ and use the construction from \cref{sec:strategy}. We set $\delta=n^{-\gamma}$ for $\gamma>0$ to be fixed below. The first constraint is that \eqref{eq:condition-omega-k-not-too-small} holds with $\lambda=n\rho$ so that we need $n\delta^d\gg1$, i.e. \begin{equation}\label{eq:gamma-first-condition} \gamma d<1. \end{equation} We first reduce to the case when there are many points in each $\Omega_k$. Defining the event $A$ as in \eqref{eq:A-finiteK} and arguing as in \eqref{eq:union-bound} gives here, for every $q>0$, the inequality \[ \begin{split} \mathbb{P}(A^c) & \lesssim_{q,\eta} \sum_{k=1}^K (n |\Omega_k|)^{-q} \lesssim_{q,\eta} n^{-q} \delta^{1-d-dq} = n^{-q(1-d\gamma)+(d-1)\gamma }, \end{split}\] where we used \eqref{eq:whitney-general-q} with $\alpha = -qd$ in the second inequality. Under the assumption \eqref{eq:gamma-first-condition} this is infinitesimal provided $q$ is chosen sufficiently large. Arguing exactly as before we can thus reduce ourselves to the case where $A$ holds. In that case, both \eqref{eq:event-many-points} and \eqref{eq:event-many-points-each-omegak} hold and thus by \eqref{eq:subadditive-random} \begin{equation}\label{eq:exp-sub-explicit-2} \begin{split} \mathbb{E}\sqa{ \C_{\pP}^p\bra{ \mathcal{N}^{n\rho}, \mathcal{M}^{n\rho}} I_{A} } & - \sum_{k=1}^K \mathbb{E}\sqa{ \C_{\pP}^p\bra{ \mathcal{N}^{(1-\eta)n\rho}_{\Omega_k}, \mathcal{M}^{(1-\eta)n\rho}_{\Omega_k}} } \\ & \qquad \lesssim \mathbb{E}\sqa{ \C_{\pP}^p\bra{ \mathcal{N}^{\eta n\rho}\cup \mathcal{U}, \mathcal{M}^{\eta n\rho}\cup \mathcal{V}}} + \sum_{k=1}^K \operatorname{diam}(\Omega_k)^p. \end{split}\end{equation} We start by considering the left-hand side of \eqref{eq:exp-sub-explicit-2}. For $\Omega_k\in \mathcal{R}_\delta$ we use the simple bound $ \C_{\pP}^p\bra{ \mathcal{N}^{(1-\eta)n\rho}_{\Omega_k}, \mathcal{M}^{(1-\eta)n\rho}_{\Omega_k}} \lesssim \operatorname{diam}(\Omega_k)^p |\mathcal{N}^{(1-\eta)n\rho}_{\Omega_k}|$, to estimate \[\begin{split} n^{\frac{p}{d}-1}\sum_{\Omega_k \in \mathcal{R}_\delta} \mathbb{E}\sqa{ \C_{\pP}^p\bra{ \mathcal{N}^{(1-\eta)n\rho}_{\Omega_k}, \mathcal{M}^{(1-\eta)n\rho}_{\Omega_k}} } & \lesssim n^{\frac{p}{d}-1} \delta^p \sum_{\Omega_k \in \mathcal{R}_\delta} \mathbb{E}\sqa{ |\mathcal{N}^{(1-\eta)n\rho}_{\Omega_k}|} \\ & \lesssim n^{\frac{p}{d}-1} \delta^p \cdot \delta^{1-d} \cdot n \delta^d = n^{-\gamma + \frac{p}{d}(1-d\gamma)}. \end{split}\] This tends to zero provided $\gamma d >p/(p+1)$ which is in particular true if (recall that $p<d$) \begin{equation}\label{eq:gamma-second-condition} \gamma d > d/(d+1). \end{equation} Notice that this condition is compatible with \eqref{eq:gamma-first-condition}. Under condition \eqref{eq:gamma-second-condition} we thus have \begin{multline*} \limsup_{n\to \infty} n^{1-\frac{p}{d}} \sum_{k=1}^K \mathbb{E}\sqa{ \C_{\pP}^p\bra{ \mathcal{N}^{(1-\eta)n\rho}_{\Omega_k}, \mathcal{M}^{(1-\eta)n\rho}_{\Omega_k}} } \\ =\limsup_{n\to \infty} \sum_{\Omega_k\in \mathcal{Q}_\delta} n^{1-\frac{p}{d}} \mathbb{E}\sqa{ \C_{\pP}^p\bra{ \mathcal{N}^{(1-\eta)n\rho}_{\Omega_k}, \mathcal{M}^{(1-\eta)n\rho}_{\Omega_k}} }. \end{multline*} Since every $\Omega_k\in \mathcal{Q}_\delta$ is a cube, we may combine \eqref{eq:limsup-poisson} in $\Omega_k$ together with the precise limit procedure, justified by the domination given in \eqref{eq:inequality-cube-holder-general} (this is why each cube $Q_i$ in the Whitney partition has side length $r<C$), to obtain \begin{equation*}\label{eq:main-contribution-G-delta} \limsup_{n\to \infty} n^{1-\frac{p}{d}} \sum_{k=1}^K \mathbb{E}\sqa{ \C_{\pP}^p\bra{ \mathcal{N}^{(1-\eta)n\rho}_{\Omega_k}, \mathcal{M}^{(1-\eta)n\rho}_{\Omega_k}} } \le (1-\eta)^{1-\frac{p}{d}} \int_\Omega \rho^{1-\frac{p}{d}}.\end{equation*} \noindent We now turn to the right-hand side of \eqref{eq:exp-sub-explicit-2}. The last term is easily estimated using directly \eqref{eq:whitney-general-q} with $\alpha=p$. In particular, if $p<d-1$ we notice that \[ n^{\frac{p}{d}-1} \sum_{k=1}^K \operatorname{diam}(\Omega_k)^p\lesssim n^{\frac{p}{d}-1} \delta^{1-(d-p)}=(n\delta^d)^{-(1-\frac{p}{d})} \delta \] which goes to zero if \eqref{eq:gamma-first-condition} holds. We finally estimate the first term in the right-hand side of \eqref{eq:exp-sub-explicit-2}. We argue as in \eqref{eq:last-term-finite-partition} and \eqref{eq:last-term-easy-finite-partition} which we combine with \cref{prop:density-helps-matching-partition} to obtain that for every $\varepsilon>0$, \begin{multline}\label{almostpartition} n^{1-\frac{p}{d}}\mathbb{E}\sqa{ \C_{\pP}^p\bra{ \mathcal{N}^{\eta n\rho}\cup \mathcal{U}, \mathcal{M}^{\eta n\rho}\cup \mathcal{V}}}\lesssim \mathbb{E}\sqa{|\mathcal{U}|/n}^{1-\frac{p}{d}}+\eta^{1-\frac{p}{d}}\\ +C(\eta,\varepsilon,\gamma) n^{\varepsilon}\bra{ \bra{\max\cur{n^{\frac{p}{d}}\delta^{p+1},n^{\frac{2}{d}} \delta^3}}^{\alpha} +\bra{n\delta^d}^{-\beta}}. \end{multline} Using \eqref{eq:whitney-general-q} with $\alpha = d/2<d-1$ we have \[ \mathbb{E}\sqa{|\mathcal{U}|/n}^{1-\frac{p}{d}}\lesssim \bra{\sum_{k=1}^K (|\Omega_k|/n)^{\frac{1}{2}}}^{1-\frac{p}{d}} \lesssim n^{\frac{1}{2}} \delta^{1-\frac{d}{2}} = \bra{\delta (n\delta^d)^{-\frac{1}{2}}}^{1-\frac{p}{d}}. \] Under condition \eqref{eq:gamma-first-condition} this term goes to zero. Regarding the term inside brackets in \eqref{almostpartition} we notice that if $q=\max\cur{p,2}$, then under condition \eqref{eq:gamma-first-condition}, \[ \max\cur{n^{\frac{p}{d}}\delta^{p+1},n^{\frac{2}{d}} \delta^3}=n^{-\gamma+ \frac{q}{d}(1-d\gamma)}. \] In particular, as above this term goes to zero under condition \eqref{eq:gamma-second-condition}.\\ We can thus choose first $\gamma$ satisfying both \eqref{eq:gamma-first-condition} and \eqref{eq:gamma-second-condition} and then $\varepsilon=\varepsilon(\alpha,\beta,\gamma,p)>0$ such that \[ \lim_{n\to \infty} n^{\varepsilon}\bra{ \bra{\max\cur{n^{\frac{p}{d}}\delta^{p+1},n^{\frac{2}{d}} \delta^3}}^{\alpha} +\bra{n\delta^d}^{-\beta}}=0. \] With this choice we find \[ \limsup_{n\to \infty }n^{1-\frac{p}{d}}\mathbb{E}\sqa{ \C_{\pP}^p\bra{ \mathcal{N}^{\eta n\rho}\cup \mathcal{U}, \mathcal{M}^{\eta n\rho}\cup \mathcal{V}}}\lesssim \eta^{1-\frac{p}{d}}, \] from which we conclude the proof of \eqref{eq:limsup-poisson} after sending $\eta\to 0$. \subsection{Uniform density on a domain} In this last case, we assume that $\Omega$ is a bounded domain with $C^2$ boundary and $\rho = I_{\Omega}/|\Omega|$ is uniform. After a simple rescaling, it is more convenient to argue with Poisson point processes $\mathcal{N}^n_{\Omega}$, $\mathcal{M}^{n}_\Omega$ with constant intensity $n$ (on $\Omega$) so that the thesis reduces to \[ \lim_{n \to \infty} n^{\frac{p}{d}-1}\mathbb{E}\sqa{ \C_{\pP}^p\bra{ \mathcal{N}^{n}_{\Omega}, \mathcal{M}^{n}_{\Omega}}} = \beta_{\mathsf{P}} |\Omega|.\] Since the boundary of $\Omega$ is $C^2$, we can apply the result from the previous section and obtain the upper bound \[ \limsup_{n \to \infty} n^{\frac{p}{d}-1}\mathbb{E}\sqa{ \C_{\pP}^p\bra{ \mathcal{N}^{n}_{\Omega}, \mathcal{M}^{n}_{\Omega}}} \le \beta_{\mathsf{P}} |\Omega|.\] To prove the corresponding lower bound, we follow closely the argument of \cite[Theorem 24]{BaBo}: we fix a cube $Q$ sufficiently large so that $\Omega \subseteq Q$ and introduce a Poisson point process $\mathcal{N}^n_{Q}$ with intensity $n$ on $Q$. For $k=2,\ldots, K$, let $\Omega_k$ be the connected components of $Q\backslash \Omega$ so that $Q\backslash \Omega=\cup_{k=2}^K\Omega_k$. Notice that for every $k$ either $\partial \Omega_k$ is $C^2$ or is the union of $\partial Q$ and a $C^2$ surface. In particular each $\Omega_k$ satisfies \eqref{eq:green-kernel-bound}. Using \eqref{eq:liminf-poisson-finite-partition} with the decomposition $Q = \Omega \cup \bigcup_{k=2}^K \Omega_k$, we obtain \[ \begin{split} \beta_{\mathsf{P}} |Q|=\liminf_{n\to \infty}n^{\frac{p}{d}-1}\mathbb{E}\sqa{ \C_{\pP}^p\bra{ \mathcal{N}^{n}_{Q}, \mathcal{M}^{n}_Q}} & \le \liminf_{n \to \infty} n^{\frac{p}{d}-1}\mathbb{E}\sqa{ \C_{\pP}^p\bra{ \mathcal{N}^{n}_{\Omega}, \mathcal{M}^{n}_{\Omega}}} \\ & \quad + \sum_{k=2}^K\limsup_{n \to \infty} n^{\frac{p}{d}-1}\mathbb{E}\sqa{ \C_{\pP}^p\bra{ \mathcal{N}^{n}_{ \Omega_k}, \mathcal{M}^{n}_{ \Omega_k}}}. \end{split}\] Now for every $k$, using \eqref{eq:limsup-poisson} we have \[ \limsup_{n \to \infty} n^{\frac{p}{d}-1}\mathbb{E}\sqa{ \C_{\pP}^p\bra{ \mathcal{N}^{n}_{ \Omega_k}, \mathcal{M}^{n}_{ \Omega_k}}} \le \beta_{\mathsf{P}}|\Omega_k|.\] Therefore, \[ \liminf_{n \to \infty} n^{\frac{p}{d}-1}\mathbb{E}\sqa{ \C_{\pP}^p\bra{ \mathcal{N}^{n}_{\Omega}, \mathcal{M}^{n}_{\Omega}}} \ge \beta_{\mathsf{P}} |Q| - \beta_\mathsf{P} \sum_{k=2}^K |\Omega_k| = \beta_{\mathsf{P}} | Q|,\] which is the desired conclusion. \section{Proof of main result}\label{sec:main} From \cref{thm:limit-poisson}, we deduce our main result \cref{thm:main}. We follow a relatively standard strategy, using de-Poissonization and concentration of measure arguments with the necessary adjustments to deal with our setting. First, we argue that \cref{thm:limit-poisson} yields similar convergence in the case of a deterministic number of independent points. \begin{proposition}\label{prop:limit-mean-iid} Let $d \ge 3$, $p\in [1,d)$ and let $\mathsf{P} = (\mathcal{F}_{n,n})_{n \in \mathbb{N}}$ be a combinatorial optimization problem over complete bipartite graphs such that assumptions \ref{as:isomorphism}, \ref{as:spanning}, \ref{as:bddegree}, \ref{ass:local-merging} and \ref{as:growth} hold. Then, with $\beta_{\mathsf{P}}\in (0, \infty)$ given by \cref{thm:limit-poisson} the following hold. Let $\Omega \subseteq \mathbb{R}^d$ be a bounded domain with Lipschitz boundary and such that \eqref{eq:green-kernel-bound} holds and let $\rho$ be a H\"older continuous probability density on $\Omega$, uniformly strictly positive and bounded from above. Given i.i.d.\ random variables $(X_i)_{i=1}^\infty$, $(Y_j)_{j=1}^\infty$ with common law $\rho$, we have \begin{equation}\label{eq:limit-mean-main-iid} \limsup_{n \to \infty} n^{\frac{p}{d}-1} \mathbb{E}\sqa{ \C_{\pP}^p\bra{ (X_i)_{i=1}^n, (Y_j)_{j=1}^n}} \le \beta_{\mathsf{P}} \int_{\Omega} \rho^{1-\frac{p}{d}}.\end{equation} Moreover, if $\rho$ is the uniform density and $\Omega$ is either a cube or has $C^2$ boundary, the limit exists and is equal to the right-hand side. \end{proposition} \begin{remark}\label{rem:phom} The only properties we used to established \cref{prop:limit-mean-iid} are the subadditivity property \eqref{eq:sub}, the growth condition \eqref{eq:upper-bound-deterministic} as well as the $p-$homogeneity of the problem. In particular it holds for every bipartite $p-$homogeneous functional $\mathbf{C}$ satisfying \begin{itemize} \item For every $\Omega\subset \mathbb{R}^d$ and every partition $\Omega = \cup_{k=1}^K \Omega_k$, $K \in \mathbb{N}$, if ${\bf x}^0$, ${\bf y}^0 \subseteq \Omega$ are such that $\min\cur{|{\bf x}^0|, |{\bf y}^0|} \ge \max\cur{\mathsf{c}_{\operatorname{A2}}, K}$, for every $k =1, \ldots, K$, ${\bf x}^k$, ${\bf y}^k \subseteq \Omega_k$ are such that $|{\bf x}^k| = |{\bf y}^k| =n_k$, with either $n_k \ge \mathsf{c}_{\operatorname{A2}}$ or $n_k = 0$ and ${\bf z}=(z_k)_{k=1}^K$ with $z_k \in \Omega_k$, for every $k=1,\ldots, K$ then \begin{equation}\label{Sp} \mathbf{C}\bra{ {\bf x}^0 \cup \bigcup_{k=1}^K {\bf x}^k, {\bf y}^0 \cup \bigcup_{k=1}^K {\bf y}^k } - \sum_{k=1}^K \mathbf{C}({\bf x}^k, {\bf y}^k) \lesssim \mathbf{C}({\bf x}^0, {\bf y}^0)+ \mathsf{M}^p({\bf z}, {\bf x}^0)+ \sum_{k=1}^K \operatorname{diam}(\Omega_k)^p. \end{equation} \item There exists $\mathsf{c}_{\operatorname{A5}} \ge 0$ such that, for every ${\bf x}, {\bf y} \subseteq (0,1)^d$, we have \begin{equation} \label{Rp} \mathbf{C}({\bf x}, {\bf y}) \le \mathsf{c}_{\operatorname{A5}} \bra{ \min\cur{|{\bf x}|^{1-\frac{p}{d}}, |{\bf y}|^{1-\frac{p}{d}}}+ \mathsf{M}^p({\bf x},{\bf y})}.\end{equation} \end{itemize} \end{remark} \begin{proof} The proof is similar to the proof of \cref{lem:subadditivity-finite-partition}. We set ${\bf x}=(X_i)_{i=1}^{n}$ and ${\bf y}=(Y_j)_{j=1}^{n}$. Let $\eta\in(0,1/2)$ and consider two independent copies $\mathcal{N}^{(1-\eta)n\rho}$ and $\mathcal{M}^{(1-\eta)n\rho}$ of Poisson point processes with intensity $(1-\eta)n\rho$ on $\Omega$. We claim that \begin{equation}\label{claimdepoi} \limsup_{n\to \infty} n^{\frac{p}{d}-1} \mathbb{E}\sqa{ \C_{\pP}^p\bra{ {\bf x}, {\bf y}}}- \limsup_{n\to \infty} n^{\frac{p}{d}-1} \mathbb{E}\sqa{ \C_{\pP}^p\bra{ \mathcal{N}^{(1-\eta)n\rho}, \mathcal{M}^{(1-\eta)n \rho}}}\lesssim \eta^{1-\frac{p}{d}}. \end{equation} By \cref{thm:limit-poisson}, this would conclude the proof of \eqref{eq:limit-mean-main-iid} since $\eta$ is arbitrary. We introduce the random variables $N=\max\cur{n-|\mathcal{N}^{(1-\eta)n\rho}|,0}$ and $M=\max\cur{n-|\mathcal{N}^{(1-\eta)n\rho}|,0}$ and notice that by the concentration properties of Poisson random variables, also $N$ and $M$ have the concentration property. Moreover, the event \[ A=\cur{ |N-\eta n|\le \eta n/2}\cap \cur{|M- \eta n|\le \eta/2} \] is of overwhelming small probability and thus arguing exactly as in the proof of \cref{lem:subadditivity-finite-partition} we have \[ \limsup_{n\to \infty} n^{\frac{p}{d}-1} \mathbb{E}\sqa{ \C_{\pP}^p\bra{ {\bf x}, {\bf y}}}=\limsup_{n\to \infty} n^{1-\frac{p}{d}} \mathbb{E}\sqa{ \mathbf{C}_\mathsf{P}( {\bf x}, {\bf y} ) I_A}. \] We let $\mathcal{N}= (X_i)_{i=n-N+1}^n$ and $\mathcal{M}=(Y_j)_{j=n-M+1}^n$ so that in $A$, ${\bf x}=\mathcal{N}^{(1-\eta)n\rho}\cup \mathcal{N}$, ${\bf y}=\mathcal{M}^{(1-\eta)n\rho}\cup \mathcal{M}$ and $\min\cur{|\mathcal{N}|,|\mathcal{M}|}\gtrsim \eta n$. \\ In $A$ we let ${\bf x}^1\subset \mathcal{N}^{(1-\eta)n\rho}$ and ${\bf y}^1\subset \mathcal{M}^{(1-\eta)n\rho}$ be such that $|{\bf x}^1|=|{\bf y}^1|$ and $$ \mathbf{C}_\mathsf{P}( \mathcal{N}^{(1-\eta)n\rho}, \mathcal{M}^{(1-\eta)n\rho} ) = \mathbf{C}_\mathsf{P}({\bf x}^1, {\bf y}^1 ).$$ We then set $\mathcal{U}=\mathcal{N}^{(1-\eta)n\rho}\backslash {\bf x}^1$, $\mathcal{V}=\mathcal{M}^{(1-\eta)n\rho}\backslash {\bf y}^1$, ${\bf x}^0=\mathcal{U}\cup \mathcal{N}$ and ${\bf y}^0=\mathcal{V}\cup \mathcal{M}$. Using \cref{lem:sub} on $\Omega$ with $K=1$, i.e.\ a trivial partition, we find that in $A$, \begin{multline*} \mathbf{C}_\mathsf{P}({\bf x},{\bf y})-\mathbf{C}_\mathsf{P}( \mathcal{N}^{(1-\eta)n\rho}, \mathcal{M}^{(1-\eta)n\rho} )\lesssim \mathbf{C}_\mathsf{P}({\bf x}^0,{\bf y}^0) +1\\ \stackrel{\eqref{eq:upper-bound-deterministic}}{\lesssim} \min\cur{ |{\bf x}^0|^{1-\frac{p}{d}}, |{\bf y}^0|^{1-\frac{p}{d}} } + \mathsf{M}^p({\bf x}^0, {\bf y}^0) + 1. \end{multline*} Multiplying by $I_A$, taking expectation and arguing exactly as in \eqref{tofinishfinite} (using in particular \cref{prop:density-helps-matching}) we conclude the proof of \eqref{claimdepoi}.\\ With a similar argument one can prove that \[ \liminf_{n\to \infty} n^{\frac{p}{d}-1} \mathbb{E}\sqa{ \C_{\pP}^p\bra{ \mathcal{N}^{n\rho}, \mathcal{M}^{n \rho}}}\le \liminf_{n\to \infty} n^{\frac{p}{d}-1} \mathbb{E}\sqa{ \C_{\pP}^p\bra{ {\bf x}, {\bf y}}}, \] which concludes the proof of \cref{prop:limit-mean-iid}. \end{proof} To conclude the proof of \cref{thm:main}, we prove a concentration bound, which improves \eqref{eq:limit-mean-main-iid} to complete convergence. The argument requires minimal assumptions on the combinatorial optimization problem and relies essentially on the validity of a Poincar\'e inequality. \begin{proposition}\label{prop:concentration} Let $d \ge 3$, $p\in [1,d)$ and let $\mathsf{P} = (\mathcal{F}_{n,n})_{n \in \mathbb{N}}$ be a combinatorial optimization problem over complete bipartite graphs such that assumptions \ref{as:bddegree} and \ref{as:growth} hold. Let $\Omega \subseteq \mathbb{R}^d$ be a bounded domain with Lipschitz boundary and let $\rho$ be a probability density on $\Omega$, uniformly strictly positive and bounded from above. Let $(X_i)_{i=1}^\infty$, $(Y_j)_{j=1}^\infty$ be i.i.d.\ random variables with common law $\rho$. For every $q \ge 2$ and $\varepsilon>0$, \begin{equation} \label{eq:concentration-cost-n} \mathbb{P}\bra{ n^{\frac{p}{d}-1}\abs{ \C_{\pP}^p\bra{ (X_i)_{i=1}^n, (Y_j)_{j=1}^n} - \mathbb{E}\sqa{\C_{\pP}^p\bra{ (X_i)_{i=1}^n, (Y_j)_{j=1}^n}}} > \varepsilon} \le_q \frac{1}{ \varepsilon^q n^{\frac{\alpha q}{2}}},\end{equation} with $$ \alpha = \begin{cases} 1-2/d & \text{if $p \in [1,2)$,}\\ 1-p/d & \text{if $p \ge 2$.}\end{cases}$$ In particular, complete (hence $\mathbb{P}$-a.s.) convergence holds: $$ \lim_{n \to \infty} n^{\frac{p}{d}-1}\abs{ \C_{\pP}^p\bra{ (X_i)_{i=1}^n, (Y_j)_{j=1}^n} - \mathbb{E}\sqa{\C_{\pP}^p\bra{ (X_i)_{i=1}^n, (Y_j)_{j=1}^n}}} = 0.$$ \end{proposition} \begin{remark}[Poincar\'e inequality] We first recall that for every Lipschitz function $F: \Omega^{2n} \to \mathbb{R}$ we have the following $L^q$-Poincar\'e inequality, \begin{equation} \label{eq:poincare-q} \mathbb{E}\sqa{ \abs{ F\bra{ (X_i)_{i=1}^n, (Y_j)_{j=1}^n} - \mathbb{E}\sqa{ F\bra{ (X_i)_{i=1}^n, (Y_j)_{j=1}^n }}}^q } \lesssim_q \mathbb{E}\sqa{ \abs{ \nabla F \bra{ (X_i)_{i=1}^n, (Y_j)_{j=1}^n}}^q}. \end{equation} Here $\abs{\nabla F}$ denotes the usual Euclidean norm of the gradient. We stress the fact that the implicit constant in \eqref{eq:poincare-q} does not depend upon $n$. Inequality \eqref{eq:poincare-q} is a consequence of well-known facts: first, the assumptions on $\Omega$ yield the $L^2$-Poincar\'e inequality with respect to the uniform measure, $$ \int_{\Omega} \abs{ u - \frac{1}{|\Omega|}\int_\Omega u }^2 \lesssim \int_\Omega |\nabla u|^2. $$ Using that the constant $c = \int_{\Omega} u \rho$ minimizes $\int_{\Omega} \abs{ u - c }^2 \rho$ and that $\rho$ is bounded from above and below, and, we obtain the weighted version $$ \int_{\Omega} \abs{ u - \int_\Omega u \rho }^2 \rho \le \int_{\Omega} \abs{ u - \frac{1}{|\Omega|}\int_\Omega u }^2\rho \le C \int_\Omega |\nabla u|^2 \rho.$$ for some $C = C(\rho, \Omega) \in (0, \infty)$. Then, a standard tensorization argument \cite[Corollary 5.7]{ledoux2001concentration} entails that the inequality holds also on the product space $\Omega^{2 n}$, endowed with the product measure $\rho^{\otimes 2n}$, with the same constant $C$. This yields \eqref{eq:poincare-q} with $q=2$. The general case $q\ge 2$ follows finally from the chain rule. Preliminarily, we notice that if $\mu$ is a probability measure on $\mathbb{R}^D$, then the validity of the inequality \begin{equation}\label{eq:poincare-mean} \int \abs{ u - \int u d \mu}^q d \mu \lesssim \int \abs{\nabla u}^qd \mu \end{equation} for every Lipschitz function $u: \mathbb{R}^D \to \mathbb{R}$ is equivalent to \begin{equation}\label{eq:poincare-median} \int \abs{ u - m_u}^q d \mu \lesssim \int \abs{\nabla u}^qd \mu, \end{equation} where $m_u$ denotes a median of (the law) of $u$, i.e.\ any $m \in \mathbb{R}$ such that $\mu( u \le m) \ge 1/2$ and $\mu( u \ge m) \ge 1/2$. Indeed, \[ \abs{ m_u - \int u d \mu} \le \abs{ \int (m_u - u) d \mu } \le \int \abs{ m_u - u} d \mu.\] Since $m_u$ can be characterized as a minimizer for $c \mapsto \int \abs{ u - c}d\mu$, we also have \[ \int \abs{ m_u - u} d \mu \le \int \abs{ u - \int u d \mu} d \mu.\] Using Jensen's inequality, we obtain \[ \abs{ m_u - \int u d \mu}^q \le \min \cur{ \int \abs{ m_u - u}^q d \mu, \int \abs{ u - \int u d \mu}^q d \mu.}\] Then, assuming that \eqref{eq:poincare-mean} or \eqref{eq:poincare-median} holds, using the triangle inequality and the bound above, we obtain the validity of the other inequality. To conclude, we assume that \eqref{eq:poincare-median} holds for $q=2$ and argue that it also holds for any $q \ge 2$. Up to adding a suitable constant, we can assume that $u$ is Lipschitz with $m_u = 0$. We then consider the Lipschitz function $v = \abs{u}^{q/2} \operatorname{sign}(u)$ (recall that in our case the support of $\mu = \rho^{\otimes 2n}$ is bounded, hence we can assume that also $u$ is bounded), so that $m_v = 0$ and apply the $q=2$ case of \eqref{eq:poincare-median}: \[\begin{split} \int \abs{u}^q d \mu & = \int \abs{v}^2 d \mu \lesssim \int |\nabla v|^2 d \mu \lesssim \int \abs{u}^{q-2} |\nabla u|^2 d \mu \\ & \lesssim \bra{ \int \abs{u}^q d \mu}^{1-2/q} \bra{ \int \abs{\nabla u}^q d \mu}^{2/q} . \end{split}\] Dividing both sides by $\bra{ \int \abs{u}^q d \mu}^{1-2/q}$ yields the desired conclusion. \end{remark} \begin{proof}[Proof of \cref{prop:concentration}] The second statement follows choosing $q$ sufficiently large in \eqref{eq:concentration-cost-n} so that the right-hand side in \eqref{eq:concentration-cost-n} is summable. We thus focus on the proof of \eqref{eq:concentration-cost-n}. Given a feasible $G \subseteq \mathcal{K}_{n,n}$, i.e., $G \in \mathcal{F}_{n,n}$, and ${\bf x} = (x_i)_{i=1}^n$, ${\bf y} = (y_j)_{j=1}^n \subseteq \Omega$, write $$ w_G({\bf x}, {\bf y}) = \sum_{ \cur{(1,i), (2,j)} \in E_G } | x_i - y_j|^p.$$ Since $p \ge 1$, $w_G$ is Lipschitz with a.e.\ derivative given by \[ \nabla_{x_i} w_G({\bf x}, {\bf y}) = \sum_{ (2,j) \in \mathcal{N}_G((1,i))} p |x_i-y_j|^{p-2} (x_i-y_j),\] and \[ \nabla_{y_j} w_G({\bf x}, {\bf y}) = -\sum_{ (1,i) \in \mathcal{N}_G((1,j))} p |x_i-y_j|^{p-2} (x_i-y_j).\] Notice also that $w_G$ is differentiable at every $({\bf x},{\bf y})$ such that $x_i\neq y_j$ for every $i,j$. Since $G \in \mathcal{F}_{n,n}$, assumption \ref{as:bddegree} yields that the sums above contain at most $\c_3$ terms, hence we bound, using Cauchy-Schwarz inequality, $$ \abs{ \nabla_{x_i} w_G({\bf x}, {\bf y})}^2 \lesssim \sum_{ (2,j) \in \mathcal{N}_G((1,i))} |x_i-y_j|^{2(p-1)},$ and similarly $$ \abs{ \nabla_{y_j} w_G({\bf x}, {\bf y})}^2 \lesssim \sum_{ (2,j) \in \mathcal{N}_G((1,i))} |x_i-y_j|^{2(p-1)}.$$ Summing upon $i$ and $j\in\cur{1,\ldots, n}$, we obtain, for the Euclidean norm of the gradient, the inequality $$ \abs{ \nabla w_G({\bf x}, {\bf y})}^2 \lesssim \sum_{ \cur{(1,i),(2,j)} \in E_G} |x_i-y_j|^{2(p-1)}.$$ If $p \ge 2$, we simply bound each term $|x_i-y_j|^{2(p-1)} \le \operatorname{diam}(\Omega)^{p-2} |x_i-y_j|^{p}$, obtaining $$ \abs{ \nabla w_G({\bf x}, {\bf y})}^2 \lesssim \sum_{ \cur{(1,i),(2,j)} \in E_G} |x_i-y_j|^{p} = w_G({\bf x}, {\bf y}).$$ If $p \in [1,2)$, we use H\"older inequality and the fact that $|E_G| \lesssim n$ (again by assumption \ref{as:bddegree}), to obtain $$ \abs{ \nabla w_G({\bf x}, {\bf y})}^2 \lesssim \bra{ \sum_{ \cur{(1,i),(2,j)} \in E_G} |x_i-y_j|^{p}}^{\frac{1}{r}} n^{1-\frac{1}{r}} = w_G({\bf x}, {\bf y})^{\frac{1}{r}} n^{1-\frac{1}{r}},$$ with $r = p/(2(p-1))$. \noindent Using the trivial bound $w_G({\bf x},{\bf y}) \lesssim n$, it follows in particular that each $w_G({\bf x}, {\bf y})$ has a Lipschitz constant bounded independently of $G$ (although the bound depends upon $n$). Therefore, also $$ \C_{\pP}^p\bra{ {\bf x}, {\bf y}} = \inf_{G \in \mathcal{F}_{n,n}} w_G({\bf x}, {\bf y}),$$ is Lipschitz, hence differentiable at Lebesgue a.e.\ $({\bf x}, {\bf y})$, by Rademacher theorem. Let $({\bf x},{\bf y})$ be a point of differentiability for both $w_G$ and $\C_{\pP}^p$ (which holds for Lebesgue a.e. point). Let $G = G({\bf x}, {\bf y}) \in \mathcal{F}_{n,m}$ be any minimizer for the problem on the graph $\mathcal{K}({\bf x},{\bf y})$ (which is a.e. unique if $p>1$ by Remark \cref{rem:uniqueness-minimizer}). For every $({\bf x}', {\bf y}')$, we have the inequality \[ \C_{\pP}^p({\bf x}',{\bf y}')\le w_G({\bf x}', {\bf y}'), \] with equality at $({\bf x},{\bf y})$, hence we obtain the identities, \begin{equation}\label{eq:identity-derivative-cpp} \nabla_{x_i} \C_{\pP}^p({\bf x},{\bf y})= \nabla_{x_i} w_G({\bf x}, {\bf y}), \quad \nabla_{y_j} \C_{\pP}^p({\bf x},{\bf y})= \nabla_{y_j} w_G({\bf x}, {\bf y}). \end{equation} Therefore, \[ \abs{ \nabla \C_{\pP}^p\bra{ {\bf x}, {\bf y}} }^2 \lesssim \begin{cases} \C_{\pP}^p\bra{ {\bf x}, {\bf y}}^{\frac{1}{r}} n^{1-\frac{1}{r}} & \text{if $p \in [1,2)$,} \\ \C_{\pP}^p\bra{ {\bf x}, {\bf y}} & \text{if $p \ge 2$.} \end{cases} \] If now ${\bf x}=(X_i)_{i=1}^n$ and ${\bf y}=(Y_j)_{j=1}^n$, combining this with \eqref{eq:poincare-q} and \eqref{eq:upper-bound-deterministic} yields \[ \begin{split} \mathbb{E}&\sqa{ \abs{ \C_{\pP}^p\bra{ {\bf x}, {\bf y} } - \mathbb{E}\sqa{ \C_{\pP}^p\bra{ {\bf x}, {\bf y}}}}^q } \lesssim_q \mathbb{E}\sqa{ \abs{ \nabla \C_{\pP}^p\bra{ {\bf x}, {\bf y}}}^{q}}\\ &\qquad \qquad \qquad \qquad \lesssim \begin{cases} \vspace{1em} \mathbb{E}\sqa{ \bra{ n^{1-\frac{p}{d}}+ \mathsf{M}^p \bra{ {\bf x}, {\bf y}}}^{\frac{q}{2r}}} n^{(1-\frac{1}{r})\frac{q}{2}} & \text{if $p \in [1,2)$,}\\ \mathbb{E}\sqa{ \bra{n^{1-\frac{p}{d}}+ \mathsf{M}^p \bra{ {\bf x}, {\bf y}}}^{\frac{q}{2}}} & \text{if $p\ge 2$.} \end{cases} \end{split}\] By the equivalence between $\mathsf{M}^p$ and $\mathsf{W}^p$ (recall \eqref{eq:wass-equals-matching}), the triangle inequality \eqref{eq:triangle} and \eqref{eq:jensen} and finally using \eqref{eq:matching-iid} with $qp$ instead of $p$, we bound from above $$ \mathbb{E}\sqa{ \bra{ \mathsf{M}^p \bra{ (X_i)_{i=1}^n, (Y_j)_{j=1}^n}}^{\frac{q}{2}}} \lesssim n^{(1-\frac{p}{d})\frac{q}{2}}.$$ If $p \ge 2$, we conclude at once that $$ \mathbb{E}\sqa{ \abs{ \C_{\pP}^p\bra{ {\bf x}, {\bf y} } - \mathbb{E}\sqa{ \C_{\pP}^p\bra{ {\bf x}, {\bf y}}}}^q } \lesssim_q n^{(1-\frac{p}{d})\frac{q}{2}},$$ hence \eqref{eq:concentration-cost-n} by Markov inequality. If $p \in [1,2)$, we bound similarly and obtain, after simple computations, $$ \mathbb{E}\sqa{ \abs{ \C_{\pP}^p\bra{ {\bf x}, {\bf y} } - \mathbb{E}\sqa{ \C_{\pP}^p\bra{ {\bf x}, {\bf y}}}}^q } \lesssim n^{\bra{ (1-\frac{p}{d})(1-\frac{1}{p})+\frac{1}{p}-\frac{1}{2}}q},$$ which leads to the corresponding case of \eqref{eq:concentration-cost-n} by Markov inequality. \end{proof} \begin{remark}[uniqueness of minimizers]\label{rem:uniqueness-minimizer} If $p>1$, for Lebesgue a.e.\ $({\bf x}, {\bf y})$, the minimizer $G \in \mathcal{F}_{n,m}$ for the problem on $\mathcal{K}({\bf x}, {\bf y})$ is unique. This in particular yields that it is unique a.s., when ${\bf x} = (X_i)_{i=1}^n$, ${\bf y} = (Y_j)_{j=1}^m$ are random i.i.d.\ with a common density $\rho$. For simplicity, we argue in the case of $|{\bf x}| = |{\bf y}|$ only, but the same result holds in general. Let $({\bf x}, {\bf y})$ be a differentiability point for $\C_{\pP}^p({\bf x}, {\bf y})$ with $X_i\neq Y_j$ for every $i,j$. Notice that by the previous proof this holds a.s. . Let $G, G' \in \mathcal{F}_{n,n}$ be minimizers for the problem on $\mathcal{K}({\bf x}, {\bf y})$, so that by \eqref{eq:identity-derivative-cpp} we obtain that, for every $i\in [n]$, $\nabla_{x_i} w_G ({\bf x}, {\bf y}) = \nabla_{x_i} w_{G'} ({\bf x}, {\bf y})$, i.e., $$ \sum_{ (2,j) \in \mathcal{N}_G((1,i))} \abs{x_i - y_j}^{p-2} (x_i-y_j) = \sum_{ (2,j) \in \mathcal{N}_{G'}((1,i))} \abs{x_i - y_j}^{p-2} (x_i-y_j) $$ Assuming that $E_G \neq E_{G'}$, we can find $i$, $j \in [n]$ such that $(2,j) \in \mathcal{N}_{G'}((1,i)) \setminus \mathcal{N}_G((1,i))$ (up to exchanging the roles of $G$ and $G'$). Then, \begin{equation}\label{eq:identity-direct}\begin{split} \abs{x_i-y_j}^{p-2}(x_i-y_j)= & \sum_{(2,k) \in \mathcal{N}_{G}((1,i)} \abs{x_i-y_k}^{p-2}(x_i-y_k) \\ & - \sum_{ (2,k)\in \mathcal{N}_{G'}((1,i)) \setminus \cur{(2,j)} } \abs{x_i-y_k}^{p-2}(x_i-y_k). \end{split} \end{equation} We notice that the right-hand side above is a function $U({\bf x}, {\bf y})$ which however does not depend on the variable $y_j$. The map $$z \in \mathbb{R}^d \mapsto \abs{z}^{p-2} z \in \mathbb{R}^d$$ is invertible, with a Borel inverse which we denote by $f$, hence we can rewrite \eqref{eq:identity-direct} equivalently as the identity $$ y_j = x_i - f\bra{ U ({\bf x}, {\bf y})},$$ where right-hand side is a Borel function of $({\bf x}, {\bf y})$ which does not depend on $y_j$. This identity however cannot hold on a set of positive Lebesgue measure. \end{remark} \section{Bounds for the Euclidean assignment problem}\label{sec:ot} In this section we establish some novel upper bounds for the random Euclidean assignment problem, in the case of not necessarily i.i.d.\ uniformly distributed points. \subsection{Matching of i.i.d.\ points} We begin with a general upper bound for the Wasserstein distance between the empirical measure of i.i.d.\ points and the corresponding common law when $d\ge 3$ and $p\ge 1$. As a consequence, we also obtain a similar bound for the Euclidean assignment problem. We derive the general case of a H\"older continuous law bounded above and below on an open connected set with Lipschitz boundary $\Omega$ from the case of the uniform law on a cube $Q \subseteq \mathbb{R}^d$. In that case, it is a well-known result, marginally discussed in \cite{AKT84}, where the focus is on the $d=2$ case. However, we point out that the case $d\ge 3$, $p\ge d/2$ was, to our knowledge, not explicitly covered in the literature until the proof provided by \cite{Le17}, which clearly extends to any $p\neq 2$ (see also \cite{goldman2021convergence}). \begin{proposition}\label{prop:matching-iid} Let $d \ge 3$, $p \ge 1$ and $\Omega$ be a bounded connected open set with Lipschitz boundary. For every H\"older continuous density $\rho: \Omega\mapsto \mathbb{R}$ bounded above and below and independent sequences $(X_{i})_{i=1}^\infty$, $(Y_j)_{j=1}^\infty$ of i.i.d.\ random variables with common law $\rho$, \begin{equation}\label{eq:transport-iid} \mathbb{E}\sqa{ \mathsf{W}^p\bra{\sum_{i=1}^n \delta_{X_i}, n\rho } } \lesssim |\Omega|^{\frac{p}{d}} n^{1-\frac{p}{d}}, \end{equation} and therefore \begin{equation}\label{eq:matching-iid} \mathbb{E}\sqa{ \mathsf{M}^p\bra{ (X_i)_{i=1}^n, (Y_j)_{j=1}^m} } \lesssim |\Omega|^{\frac{p}{d}} \min\cur{n,m}^{1-\frac{p}{d}}. \end{equation} \end{proposition} \begin{proof} Inequality \eqref{eq:matching-iid} follows from \eqref{eq:transport-iid} assuming e.g.\ $n \le m$ and \eqref{eq:matching-below-wasserstein} with $\lambda = \rho$. Hence, we focus on the proof of \eqref{eq:transport-iid}. By Jensen inequality \eqref{eq:jensen}, it is enough to prove this bound for large $p$ so that we may assume without loss of generality that $p>d/(d-1)$. We then set $\mu= \frac{1}{n} \sum_{i=1}^n \delta_{X_i}$. We first prove the statement in the case $\Omega=Q$ is a cube. By scaling we may assume that $Q=(0,1)^d$ is the unit cube. By Proposition \ref{prop:map-heat-semigroup}, there is a bi-Lipschitz map $T: Q\mapsto Q$ with Lipschitz constant depending only on $\rho$ such that $T\sharp \rho=1$. Then, $X_i'=T(X_i)$ are i.i.d.\ uniformly distributed on $Q$ Letting $\mu'=T\#\mu=\frac{1}{n}\sum_{i=1}^n \delta_{X'_i}$ we have \[ \mathsf{W}^p(\mu, \rho)\lesssim \mathsf{W}^p(\mu',1) \] and the statement follows from \cite{Le17}. Consider now $\Omega$ a general bounded connected open set with Lipschitz boundary. We say that $\Omega$ is well-partitioned if there exists convex polytopes $(\Omega_k)_{k=1}^K$ covering $\Omega$, with $|\Omega_k\cap \Omega_{k'}|=0$ for $k\neq k'$ and such that each $\Omega_k$ is bi-Lipschitz homeomorphic to a cube. By \cite{Garcia-Slepcev}, every connected and Lipschitz domain is bi-Lipschitz homeomorphic to a well-partitioned and smooth domain so that arguing exactly as above we may assume that $\Omega$ itself is smooth and well-partitioned. Let $T_k: \Omega_k\mapsto Q_k$ be Lipschitz homeomorphisms between $\Omega_k$ and some cubes $Q_k$. We then define $\rho_k=T_k\#\rho$, $n_k=\mu(\Omega_k)$ and $\mu_k= \frac{n}{n_k}T_k\#\mu$. Notice in particular that we may write $\mu_k=\frac{1}{n_k}\sum_{i=1}^{n_k} \delta_{Y_i}$ where $(Y_i)_{i=1}^\infty$ are i.i.d.\ with common law $\rho_k/\rho_k(Q_k)$ and that $n_k$ is a Binomial random variable with parameters $n$ and $\rho_k(Q_k)=\rho(\Omega_k)$. Using \eqref{eq:mainsub} with $\varepsilon=1$ we thus find \[ \begin{split} \mathsf{W}^p(\mu,\rho) & \lesssim \sum_{k=1}^K \mathsf{W}^p_{\Omega_k}\bra{\mu, \frac{n_k}{n\rho(\Omega_k)}\rho} +\mathsf{W}^p\bra{\sum_{k=1}^K \frac{n_k}{n \rho(\Omega_k)} \rho I_{\Omega_k},\rho}\\ & \stackrel{\eqref{eq:estimCZ}}{\lesssim} \sum_{k=1}^K \frac{n_k}{n}\mathsf{W}^p_{Q_k}\bra{\mu_k, \rho_k} +\nor{\sum_{k=1}^K \bra{\frac{n_k}{n \rho(\Omega_k)}-1} I_{\Omega_k} \rho}_{W^{-1,p}(\Omega)}^p\\ & \stackrel{\eqref{eq:Lp}}{\lesssim}\sum_{k=1}^K \frac{n_k}{n}\mathsf{W}^p_{Q_k}\bra{\mu_k, \rho_k} +\sum_{k=1}^K |\Omega_k| \lt|\frac{n_k}{n \rho(\Omega_k)}-1\rt|^p. \end{split}\] Taking the expectation and using the concentration properties of binomial random variables \eqref{eq:binomial-concentration} we find \[ \mathbb{E}\sqa{\mathsf{W}^p(\mu,\rho)}\lesssim \sum_{k=1}^K\mathbb{E}\sqa{ \frac{n_k}{n}\mathsf{W}^p_{Q_k}\bra{\mu_k, \rho_k}} + \frac{1}{n^{\frac{p}{2}}}. \] By the first part of the proof and the concentration properties of Binomial random variables we get \[ \mathbb{E}\sqa{ \frac{n_k}{n}\mathsf{W}^p_{Q_k}\bra{\mu_k, \rho_k}}\lesssim \frac{1}{n^{\frac{p}{d}}} \] which concludes the proof of \eqref{eq:matching-iid} since $p/2>p/d$. \end{proof} \begin{remark}\label{rem:matching} By translation and scaling invariance, when $\Omega=Q$ and $\rho= \frac{1}{|\Omega|} I_{Q}$ is the uniform measure of a cube $Q\subset \mathbb{R}^d$, the implicit constant in \eqref{eq:matching-iid} does not depend on $Q$. \end{remark} \subsection{Matching with a fraction of i.i.d.\ points} In this section we extend the bound \eqref{eq:matching-iid} for the matching to the case where most of the points are still i.i.d.\ but essentially no assumption is made on the remaining points. This is used in \cref{thm:limit-poisson} and in the de-Poissonization procedure (see Proposition \ref{prop:limit-mean-iid}). Just like in \cref{thm:limit-poisson} we will have to consider three different situations. Let us however set some common notation. Letting $\mathcal{N}$, $\mathcal{M}$, $\mathcal{U}$ and $\mathcal{V}$ be point processes on $\Omega$ ($\mathcal{N}$ and $\mathcal{M}$ will contain the i.i.d.\ points), we want to estimate \[ \mathbb{E}\sqa{ \mathsf{M}^p( \mathcal{U} \cup \mathcal{N}, \mathcal{V} \cup \mathcal{M}) }. \] Setting \[Z=\min\cur{|\mathcal{U}|+|\mathcal{N}|,|\mathcal{V}|+|\mathcal{M}|},\] we want to construct two (random) subsets $\mathcal{S} \subseteq \mathcal{U} \cup \mathcal{N}$, $\mathcal{T} \subseteq \mathcal{V} \cup \mathcal{M}$, both containing $Z$ points, so that \begin{equation}\label{triangleMp} \mathsf{M}^p\bra{ \mathcal{U} \cup \mathcal{N}, \mathcal{V} \cup \mathcal{M} } \le \mathsf{W}^p\bra{ \mu^{\mathcal{S}}, \nu^{\mathcal{T}}}\stackrel{\eqref{eq:triangle}}{\lesssim} \mathsf{W}^p\bra{ \mu^{\mathcal{S}}, Z\rho} +\mathsf{W}^p\bra{ \nu^{\mathcal{T}}, Z\rho}, \end{equation} where $\mu^{\mathcal{S}}, \mu^{\mathcal{T}}$ are the associated empirical measures. We then separately estimate the two terms on the right-hand side of \eqref{triangleMp}. Since the construction is completely symmetric, we detail it only for $\mathcal{S} \subseteq \mathcal{U} \cup \mathcal{N}$. It is given as the union of two sets, a ``good'' set $\mathcal{G} \subseteq \mathcal{N}$ and a ``bad'' set $\mathcal{B} \subseteq \mathcal{U}$. We first define the set $\mathcal{G}$ by sampling without replacement \[ |\mathcal{G}|=\min\cur{|\mathcal{N}|,Z} \] points from $\mathcal{N}$. Similarly, the set $\mathcal{B}$ is constructed by sampling without replacement \[ |\mathcal{B}|=\max\cur{Z-|\mathcal{N}|,0} \] points from $\mathcal{U}$. Notice that \begin{equation}\label{Z} Z=|\mathcal{G}|+|\mathcal{B}| \end{equation} and that when conditioned on $|\mathcal{G}|$, the points in $\mathcal{G}$ are still i.i.d.\ with common law $\rho$. We then write $\mu^{\mathcal{S}} = \mu^{\mathcal{G}} + \mu^{\mathcal{B}}$ for the associated empirical measure. Using the triangle inequality \eqref{eq:triangle} and \eqref{eq:convexity}, we then split the estimate in two: \[\begin{split} \mathsf{W}^p (\mu^{\mathcal{S}}, Z\rho ) & \lesssim \mathsf{W}^p \bra{ \mu^{\mathcal{G}} + \mu^{\mathcal{B}}, |\mathcal{G}|\rho + \mu^{\mathcal{B}}} + \mathsf{W}^p \bra{|\mathcal{G}|\rho + \mu^{\mathcal{B}}, Z\rho}\\ & \lesssim \mathsf{W}^p \bra{ \mu^{\mathcal{G}},|\mathcal{G}|\rho}+ \mathsf{W}^p \bra{|\mathcal{G}|\rho + \mu^{\mathcal{B}}, Z\rho }. \end{split}\] Taking expectation we find \begin{equation}\label{commonstartpoint} \mathbb{E}\sqa{ \mathsf{W}^p (\mu^{\mathcal{S}}, Z\rho )}\lesssim \mathbb{E}\sqa{\mathsf{W}^p \bra{ \mu^{\mathcal{G}},|\mathcal{G}|\rho}}+ \mathbb{E}\sqa{\mathsf{W}^p \bra{|\mathcal{G}|\rho + \mu^{\mathcal{B}}, Z\rho }}. \end{equation} To estimate the first term in the right-hand side, we will rely on \eqref{eq:transport-iid}. It is in the estimate of the last term that we need to argue differently depending on the cases. In the first one (see \cref{prop:density-helps-matching}), since we have a good control on the moments of $|\mathcal{U}|$ we can directly appeal to \cref{prop:density-helps}. In the two other cases (see \cref{prop:density-helps-matching-partition,thm:bound-matching}) we need to combine it with a localization argument.\\ We start with the first case. \begin{proposition}\label{prop:density-helps-matching} Let $d \ge 3$, $p \ge 1$ and $\Omega$ be a bounded domain with Lipschitz boundary, $\rho: \Omega\mapsto \mathbb{R}$ be a H\"older continuous density bounded above and below and $(X_{i})_{i=1}^\infty$, $(Y_j)_{j=1}^\infty$ be independent sequences of i.i.d.\ random variables with common law $\rho$. \\ Then, there exists $\alpha=\alpha(p,d)<2$ and $\beta=\beta(p)>0$ such that the following holds. Let $M, N\in \mathbb{N}$ be random variables satisfying concentration (recall \cref{def:concen}) and set $\mathcal{N}=(X_{i})_{i=1}^N$, $\mathcal{M} = (Y_{j})_{j=1}^M$ and $h=\min\cur{\mathbb{E}\sqa{M},\mathbb{E}\sqa{N}}$. Then, for every point processes $\mathcal{U}$, $\mathcal{V}$, for which there exists $1\le H\le h$ such that for every $q\ge 1$ \begin{equation}\label{hyp:UVhelps} \mathbb{E}\sqa{ |\mathcal{U}|^q+|\mathcal{V}|^q}\le C(q) H^q \end{equation} for some $C(q)>0$, we have \begin{equation*}\label{eq:density-helps-random} \mathbb{E}\sqa{ \mathsf{M}^p( \mathcal{U} \cup \mathcal{N}, \mathcal{V} \cup \mathcal{M}) } \lesssim h^{1-\frac{p}{d}}\bra{1+ \bra{\frac{H^\alpha}{h}}^\beta}. \end{equation*} Here the implicit constant depends only on $p$, $d$, the constants involved in the concentration properties of $M,N$ and $(C(q))_{q\ge 1}$ from \eqref{hyp:UVhelps}. \end{proposition} \begin{proof} Starting from \eqref{triangleMp} and \eqref{commonstartpoint} we first estimate by \eqref{eq:matching-iid} and H\"older inequality, \[ \mathbb{E}\sqa{\mathsf{W}^p \bra{ \mu^{\mathcal{G}},|\mathcal{G}|\rho}}\lesssim \mathbb{E}\sqa{|\mathcal{G}|^{1-\frac{p}{d}}}\le \mathbb{E}\sqa{|\mathcal{G}|}^{1-\frac{p}{d}}. \] Since $|\mathcal{G}|\le \min\cur{M,N} +|\mathcal{V}|$, by \eqref{hyp:UVhelps} with $q=1$ and $H\le h$, we have $\mathbb{E}\sqa{|\mathcal{G}|}\lesssim h$ and thus \[ \mathbb{E}\sqa{\mathsf{W}^p \bra{ \mu^{\mathcal{G}},|\mathcal{G}|\rho}}\lesssim h^{1-\frac{p}{d}}. \] We are then left with the proof of \begin{equation}\label{toprovehelps} \mathbb{E}\sqa{\mathsf{W}^p \bra{|\mathcal{G}|\rho + \mu^{\mathcal{B}}, Z\rho }}\lesssim h^{1-\frac{p}{d}} \bra{\frac{H^\alpha}{h}}^\beta. \end{equation} We first single out the event \[ A=\cur{|\mathcal{G}|\ge h/2} \] and claim that for $q\ge 1$ \begin{equation}\label{Achelps} \mathbb{P}\sqa{A^c}\lesssim_q h^{-q}. \end{equation} Indeed, since $A^c\subset \{N\le \mathbb{E}\sqa{N}/2\}\cup\{M\le \mathbb{E}\sqa{M}/2\}$, \eqref{Achelps} follows by combining a union bound together with the concentration properties of $M$ and $N$. Since \[ \mathsf{W}^p \bra{|\mathcal{G}|\rho + \mu^{\mathcal{B}}, Z\rho }\le \mathsf{W}^p\bra{ \mu^{\mathcal{B}}, |\mathcal{B}|\rho}\lesssim |\mathcal{B}|\le |\mathcal{U}|, \] we then find \[ \mathbb{E}\sqa{\mathsf{W}^p \bra{|\mathcal{G}|\rho + \mu^{\mathcal{B}}, Z\rho }I_{A^c}}\lesssim \mathbb{E}\sqa{|\mathcal{U}|^2}^{\frac{1}{2}} \mathbb{P}\sqa{A^c}^{\frac{1}{2}}\stackrel{\eqref{hyp:UVhelps}\&\eqref{Achelps}}{\lesssim_q} h^{-q} H. \] By taking $q$ large enough, in order to prove \eqref{toprovehelps} it is therefore sufficient to show \begin{equation}\label{toprovehelpsA} \mathbb{E}\sqa{\mathsf{W}^p \bra{|\mathcal{G}|\rho + \mu^{\mathcal{B}}, Z\rho }I_A}\lesssim h^{1-\frac{p}{d}} \bra{\frac{H^\alpha}{h}}^\beta. \end{equation} We start with the case $p>d/(d-1)$. By \cref{prop:density-helps}, \begin{multline*} \mathbb{E}\sqa{\mathsf{W}^p \bra{|\mathcal{G}|\rho + \mu^{\mathcal{B}}, Z\rho }I_A}\lesssim\mathbb{E}\sqa{\frac{|\mathcal{B}|^{1+\frac{p}{d}}}{|\mathcal{G}|^{\frac{p}{d}}} I_A} \lesssim h^{-\frac{p}{d}}\mathbb{E}\sqa{ |\mathcal{U}|^{1+\frac{p}{d}} I_A}\stackrel{\eqref{hyp:UVhelps}}{\lesssim} h^{1-\frac{p}{d}} \frac{H^{1+\frac{p}{d}}}{h}. \end{multline*} This proves \eqref{toprovehelpsA} in this case with $\alpha=1+p/d$ and $\beta=1$.\\ If now $p<d/(d-1)<2$, we use Jensen's inequality \eqref{eq:jensen} to obtain \begin{multline*} \mathbb{E}\sqa{\mathsf{W}^p \bra{|\mathcal{G}|\rho + \mu^{\mathcal{B}}, Z\rho }I_A}\lesssim\mathbb{E}\sqa{ Z^{1-\frac{p}{2}} \bra{\mathsf{W}^2\bra{|\mathcal{G}|\rho + \mu^{\mathcal{B}}, Z\rho }I_A}^{\frac{p}{2}}}\\ \le \mathbb{E}\sqa{ Z}^{1-\frac{p}{2}} \mathbb{E}\sqa{\mathsf{W}^2\bra{|\mathcal{G}|\rho + \mu^{\mathcal{B}}, Z\rho }I_A}^{\frac{p}{2}}. \end{multline*} Recalling \eqref{Z} we find $\mathbb{E}\sqa{ Z}\lesssim h+H\lesssim h$. Using finally \eqref{toprovehelpsA} with $p=2$ we conclude that \[ \mathbb{E}\sqa{\mathsf{W}^p \bra{|\mathcal{G}|\rho + \mu^{\mathcal{B}}, Z\rho }I_A}\lesssim h^{1-\frac{p}{2}} \bra{h^{1-\frac{2}{d}} \bra{\frac{H^\alpha}{h}}^\beta}^{\frac{p}{2}}= h^{1-\frac{p}{d}} \bra{\frac{H^\alpha}{h}}^{\beta\frac{p}{2}}. \] This proves \eqref{toprovehelpsA} also in this case. \end{proof} We now consider the case when the moment bounds for $\mathcal{U}$ and $\mathcal{V}$ are only valid after restricting on a Whitney-type decomposition from \cref{lem:decomp}. \setcounter{proof-step}{0} \begin{proposition}\label{prop:density-helps-matching-partition} Let $d \ge 3$, $p \ge 1$ and $\Omega \subseteq \mathbb{R}^d$ be a bounded connected open set with Lipschitz boundary and such that \eqref{eq:green-kernel-bound} holds. Fix a Whitney partition $\mathcal{Q} = (Q_i)_i$, and for $\delta >0$ let $(\Omega_k)_{k=1}^K = \mathcal{Q}_\delta \cup \mathcal{R}_\delta$ be given by \cref{lem:decomp}. Let finally $\rho$ be a H\"older continuous probability density on $\Omega$, bounded above and below.\\ Then, there exist $\alpha=\alpha(p,d)>0$ and $\beta=\beta(p,d) >0$ such that the following holds. For every $\eta\in (0,1)$, $\varepsilon>0$ and $\gamma\in( 0, 1/d)$, there exists $C(\eta,\varepsilon,\gamma)$ such that for every Poisson point processes $\mathcal{N}^{\eta n \rho}, \mathcal{M}^{\eta n \rho}$ with intensity $\eta n\rho$ and every point processes $\mathcal{U}$ and $\mathcal{V}$ on $\Omega$ such that \begin{equation}\label{hypmomUVhelpmatching} \mathbb{E}\lt[|\mathcal{U}_{\Omega_k}|^q+ |\mathcal{V}_{\Omega_k}|^q\rt]\lesssim_q (n|\Omega_k|)^{\frac{q}{2}} \qquad \forall q>0, \end{equation} if $\delta= n^{-\gamma}$ then \begin{multline*} n^{\frac{p}{d}-1}\mathbb{E}\sqa{ \mathsf{M}^p( \mathcal{U} \cup \mathcal{N}^{\eta n\rho}, \mathcal{V} \cup \mathcal{M}^{\eta n\rho}) }\lesssim \eta^{1-\frac{p}{d}}\\ +C(\eta,\varepsilon,\gamma) n^{\varepsilon}\bra{ \bra{\max\cur{n^{\frac{p}{d}}\delta^{p+1},n^{\frac{2}{d}} \delta^3}}^{\alpha} +\bra{n\delta^d}^{-\beta}}. \end{multline*} \end{proposition} \begin{proof} Using the notation from the beginning of this section, we start as above from \eqref{triangleMp} and \eqref{commonstartpoint} and estimate by \eqref{eq:matching-iid}, \[ \mathbb{E}\sqa{\mathsf{W}^p \bra{ \mu^{\mathcal{G}},|\mathcal{G}|\rho}}\lesssim \mathbb{E}\sqa{|\mathcal{G}|}^{1-\frac{p}{d}}. \] Since $|\mathcal{G}|\le |\mathcal{N}^{n\eta \rho}|$ we get \[ \mathbb{E}\sqa{\mathsf{W}^p \bra{ \mu^{\mathcal{G}},|\mathcal{G}|\rho}}\lesssim (\eta n)^{1-\frac{p}{d}}. \] In order to conclude the proof it is thus enough to show \begin{equation}\label{mainclaimdensityhelps} n^{\frac{p}{d}-1} \mathbb{E}\sqa{\mathsf{W}^p\bra{ \mu^{\mathcal{B}}+|\mathcal{G}|\rho, Z\rho}}\le C(\eta,\varepsilon,\gamma) n^{\varepsilon}\bra{ \bra{\max\cur{n^{\frac{p}{d}}\delta^{p+1},n^{\frac{2}{d}} \delta^3}}^{\alpha} +\bra{n\delta^d}^{-\beta}}. \end{equation} \setcounter{proof-step}{0} \noindent{\emph{Step \stepcounter{proof-step}\arabic{proof-step}. Reduction to a ``good'' event.}} We let \[ A=\{|\mathcal{G}|\in[\eta n/2, 3\eta n]\}\cap \bigcap_{k=1}^K \cur{ \max\cur{ |\mathcal{U}_{\Omega_k}|, |\mathcal{V}_{\Omega_k}| }\le (n |\Omega_k|)^{\frac{1}{2}} \cdot n^{\varepsilon} }. \] and claim that \begin{equation}\label{estimAcdensityhelp} n^{\frac{p}{d}-1}\mathbb{E}\sqa{\mathsf{W}^p\bra{ \mu^{\mathcal{B}}+|\mathcal{G}|\rho, Z\rho} I_{A^c}}\le C(\eta, \varepsilon,\gamma)\bra{n\delta^d}^{-\beta}. \end{equation} We first prove that for every $q>0$, \begin{equation}\label{estimPAcdensityhelp} \mathbb{P}\sqa{A^c}\lesssim_q C(\eta) n^{-q} +\delta^{1-d} n^{-\varepsilon q}. \end{equation} To prove this we use a union bound and split \begin{multline*} \mathbb{P}\sqa{A^c}\le \mathbb{P}\sqa{|\mathcal{G}|\notin [\eta n/2, 3\eta n] }\\ +\sum_{k=1}^K \mathbb{P}\sqa{|\mathcal{U}_{\Omega_k}|\ge (n |\Omega_k|)^{\frac{1}{2}} \cdot n^{\varepsilon} }+ \sum_{k=1}^K \mathbb{P}\sqa{|\mathcal{V}_{\Omega_k}|\ge (n |\Omega_k|)^{\frac{1}{2}} \cdot n^{\varepsilon} }. \end{multline*} Regarding the first term we notice that \begin{multline*} \{|\mathcal{G}|\notin[\eta n/2, 3\eta n]\}\subset \{ |\mathcal{N}^{\eta n \rho}|< \eta n/2\}\cup \{ |\mathcal{N}^{\eta n \rho}|> 3\eta n\} \cup \{ |\mathcal{M}^{\eta n \rho}|< \eta n/2\}. \end{multline*} Using once more a union bound and \eqref{eq:density-bound-below-Poi}, we find \[ \mathbb{P}\sqa{|\mathcal{G}|\notin [\eta n/2, 3\eta n] }\lesssim_q C(\eta) n^{-q}. \] Regarding the two sums, by \eqref{hypmomUVhelpmatching}, we have for every $k\in [1,K]$ \[ \mathbb{P}\sqa{|\mathcal{U}_{\Omega_k}|\ge (n |\Omega_k|)^{\frac{1}{2}} \cdot n^{\varepsilon} }\lesssim_q n^{-\varepsilon q} \] and similarly for $\mathcal{V}$. Since $K\lesssim \delta^{1-d}$ by \eqref{eq:whitney-general-q} this concludes the proof of \eqref{estimPAcdensityhelp}.\\ We now turn to \eqref{estimAcdensityhelp}. As above by the bound $\mathsf{W}^p\bra{ \mu^{\mathcal{B}}+|\mathcal{G}|\rho, Z\rho}\lesssim |\mathcal{U}|$ and Cauchy-Schwarz, we have \[ \mathbb{E}\sqa{\mathsf{W}^p\bra{ \mu^{\mathcal{B}}+|\mathcal{G}|\rho, Z\rho} I_{A^c}}\lesssim \mathbb{E}\sqa{|\mathcal{U}|^2}^{\frac{1}{2}}\mathbb{P}\sqa{A^c}^{\frac{1}{2}}. \] Using once more Cauchy-Schwarz together with \eqref{hypmomUVhelpmatching} with $q=2$ we have \[ \mathbb{E}\sqa{|\mathcal{U}|^2}^{\frac{1}{2}}\lesssim K^{\frac{1}{2}} n^{\frac{1}{2}} \] so that by \eqref{estimPAcdensityhelp} and $K\lesssim \delta^{1-d}$ \[ n^{\frac{p}{d}-1} \mathbb{E}\sqa{\mathsf{W}^p\bra{ \mu^{\mathcal{B}}+|\mathcal{G}|\rho, Z\rho} I_{A^c}}\lesssim_q C(\eta) \delta^{\frac{1}{2}(1-d)} (n^{-q} +\delta^{1-d} n^{-\varepsilon q})^{\frac{1}{2}}n^{\frac{p}{d}-\frac{1}{2}}. \] Since $\delta=n^{-\gamma}$, this concludes the proof of \eqref{estimAcdensityhelp} provided we choose $q$ large enough depending on $\varepsilon$ and $\gamma$. \medskip \noindent In the remaining two steps we prove that in $A$, \begin{equation}\label{mainclaimdensityhelpsA} n^{\frac{p}{d}-1}\mathsf{W}^p\bra{ \mu^{\mathcal{B}}+|\mathcal{G}|\rho, Z\rho}\lesssim_\eta n^{\varepsilon}\bra{ \bra{\max\cur{n^{\frac{p}{d}}\delta^{p+1},n^{\frac{2}{d}} \delta^3}}^{\alpha} +\bra{n\delta^d}^{-\beta}}. \end{equation} After taking expectation and in combination with \eqref{estimAcdensityhelp} this would conclude the proof of \eqref{mainclaimdensityhelps}. From this point all the estimates are deterministic.\\ \medskip \noindent{\emph{Step \stepcounter{proof-step}\arabic{proof-step}. Estimate for $p>d/(d-1)$.}} We first use \eqref{eq:mainsub}, e.g.\ with $\varepsilon=1$, to obtain \begin{equation}\label{eq:subadd-epsilon-1 \mathsf{W}^p_{\Omega} \bra{ \mu^{\mathcal{B}}+ |\mathcal{G}|\rho, Z\rho } \lesssim \sum_{k=1}^K \mathsf{W}^p_{\Omega_k}\bra{ \mu^{\mathcal{B}} + |\mathcal{G}|\rho, \alpha_k \rho} + \mathsf{W}^p_{\Omega}\bra{ \sum_{k=1}^K \alpha_k I_{\Omega_k} \rho, Z\rho }, \end{equation} with \begin{equation}\label{eq:alpha_k} \alpha_k = \frac{ \mu^{\mathcal{B}}(\Omega_k)}{\rho(\Omega_k)} + |\mathcal{G}|.\end{equation} We bound the terms in the right-hand side separately. For the sum of ``local'' terms, we estimate differently according to $\Omega_k\in \mathcal{R}_\delta$ or $\Omega_k\in \mathcal{Q}_\delta$. In the first case we use the naive bound \[ \begin{split} \mathsf{W}^p_{\Omega_k}\bra{ {\mu^{\mathcal{B}} + |\mathcal{G}|\rho}, \alpha_k \rho} & \stackrel{\eqref{eq:convexity}}{\le} \mathsf{W}^p_{\Omega_k}\bra{ \mu^{\mathcal{B}} , \frac{\mu^{\mathcal{B}}(\Omega_k)}{\rho(\Omega_k)} \rho} \stackrel{\eqref{eq:w-trivial}}{\le} \operatorname{diam}(\Omega_k)^p |\mathcal{U}_{\Omega_k}| \\ & \lesssim n^{\frac{1}{2}+\varepsilon} \delta^{p + \frac{d}{2}}.\end{split}\] Since $K\lesssim \delta^{1-d}$ we find \begin{equation}\label{eq:second-term-matching-whitney}n^{\frac{p}{d}-1} \sum_{\Omega_k \in \mathcal{R}_\delta} \mathsf{W}^p_{\Omega_k}\bra{ \mu^{\mathcal{B}} + |\mathcal{G}| \rho, \alpha_k \rho} \lesssim n^\varepsilon n^{\frac{p}{d}} \delta^{1+p} (n\delta^{-d})^{ - \frac{1}{2}}\le n^\varepsilon n^{\frac{p}{d}} \delta^{1+p}. \end{equation} If $\Omega_k \in \mathcal{Q}_\delta$ is a cube, we use instead \cref{prop:density-helps} with $\mu^{\mathcal{B}}$ instead of $\mu$ and $|\mathcal{G}|$ instead of $h$, so that \[ \begin{split} \mathsf{W}^p_{\Omega_k}\bra{ {\mu^{\mathcal{B}} + |\mathcal{G}| \rho}, \alpha_k\rho} & \lesssim_\eta \frac{\mu^{\mathcal{B}}(\Omega_k)^{1+\frac{p}{d}}}{n^{\frac{p}{d}}} \lesssim_\eta n^{-\frac{p}{d}} |\mathcal{U}_{\Omega_k}|^{1+\frac{p}{d}} \\ & \lesssim_\eta n^{(1+\frac{p}{d})\varepsilon} n^{\frac{1}{2} (1-\frac{p}{d})} |\Omega_k|^{\frac{1}{2}(1+\frac{p}{d})}. \end{split}\] Summing this inequality yields \begin{equation*}\begin{split} n^{\frac{p}{d}-1}\sum_{\Omega_k \in \mathcal{Q}_\delta} \mathsf{W}^p_{\Omega_k}\bra{ \mu^{\mathcal{B}} + |\mathcal{G}| \rho, \alpha_k \rho} & \lesssim_\eta n^{(1+\frac{p}{d})\varepsilon} n^{-\frac{1}{2} (1-\frac{p}{d})} \sum_{k=1}^K |\Omega_k|^{\frac{1}{2}(1+\frac{p}{d})} \\ & \stackrel{ \eqref{eq:whitney-general-q}}{\lesssim_\eta} n^{2 \varepsilon} n^{-\frac{1}{2} (1-\frac{p}{d})}\max\cur{1,\delta^{\frac{1}{2}(d-2-p)}}. \end{split}\end{equation*} Notice that since $n \delta^d\ge 1$, \[ n^{-\frac{1}{2} (1-\frac{p}{d})}\max\cur{1,\delta^{\frac{1}{2}(d-2-p)}}\le (n\delta^d)^{-\frac{1}{2} (1-\frac{p}{d})} \] so that \begin{equation}\label{eq:third-term-matching-whitney} n^{\frac{p}{d}-1}\sum_{\Omega_k \in \mathcal{Q}_\delta} \mathsf{W}^p_{\Omega_k}\bra{ \mu^{\mathcal{B}} + |\mathcal{G}| \rho, \alpha_k \rho} \lesssim_\eta n^{2 \varepsilon} (n\delta^d)^{-\frac{1}{2} (1-\frac{p}{d})}. \end{equation} We then consider the last term in \eqref{eq:subadd-epsilon-1}. Using \cref{lem:peyre} with $Z \rho \gtrsim \eta n$ in place of $\lambda$ (recall that we assume here that $A$ holds), we get \begin{equation}\label{firststepglobalmatchingwhitney} \mathsf{W}^p_{\Omega}\bra{ \sum_{k=1}^K \alpha_k I_{\Omega_k} \rho, Z\rho } \lesssim_\eta n^{1-p} \nor{ \sum_{k=1}^K \alpha_kI_{\Omega_k} \rho - Z\rho}_{W^{-1,p}(\Omega)}^p. \end{equation} Recalling \eqref{eq:alpha_k} and that $Z = \mu^{\mathcal{B}}(\Omega) + |\mathcal{G}|$, we can rewrite \[ \sum_{k=1}^K \alpha_kI_{\Omega_k} \rho - Z\rho = \sum_{k=1}^K \frac{ \mu^{\mathcal{B}}(\Omega_k)}{\rho(\Omega_k)} \bra{I_{\Omega_k}- \rho(\Omega_k)} \rho. \] By \eqref{estimW1pWhitney} of \cref{lem:W1pWhitney} with $h= n^{1+2\varepsilon}$ we thus have in $A$ \[ \nor{ \sum_{k=1}^K \alpha_kI_{\Omega_k} \rho - Z\rho}_{W^{-1,p}(\Omega)}\lesssim n^{\varepsilon} \delta^{1-\frac{d}{2}} n^{\frac{1}{2}}. \] Combining this with \eqref{firststepglobalmatchingwhitney} we get that in $A$, \[n^{\frac{p}{d}-1} \mathsf{W}^p_{\Omega}\bra{ \sum_{k=1}^K \alpha_k I_{\Omega_k} \rho, Z\rho } \lesssim n^{p\varepsilon} (n\delta^d)^{-p\frac{(d-2)}{2d}}. \] Inserting this estimate, \eqref{eq:second-term-matching-whitney} and \eqref{eq:third-term-matching-whitney} in \eqref{eq:subadd-epsilon-1} we finally obtain (notice that $p(d-2)>d-p$ for $p>d/(d-1)$) that in $A$, \[ n^{\frac{p}{d}-1}\mathsf{W}^p\bra{ \mu^{\mathcal{B}}+|\mathcal{G}|\rho, Z\rho}\lesssim_\eta n^{\max\cur{p,2} \varepsilon} \bra{ n^{\frac{p}{d}} \delta^{1+p}+(n\delta^d)^{-\frac{(d-p)}{2d}}}. \] Up to replacing $\varepsilon$ by $\max\cur{p,2} \varepsilon$, this concludes the proof of \eqref{mainclaimdensityhelpsA} if $p>d/(d-1)$.\\ \medskip \noindent{\emph{Step \stepcounter{proof-step}\arabic{proof-step}. Estimate for $p\le d/(d-1)$.}} Since $2>d/(d-1)\ge p$, we may use Jensen's inequality \eqref{eq:jensen} to infer that in $A$, \[\begin{split} n^{\frac{p}{d}-1} \mathsf{W}^p\bra{ \mu^{\mathcal{B}}+|\mathcal{G}|\rho, Z\rho}& \le n^{\frac{p}{d}-1} Z^{1-\frac{p}{2}}\bra{\mathsf{W}^2\bra{\mu^{\mathcal{B}}+|\mathcal{G}|\rho, Z\rho}}^{\frac{p}{2}}\\ & \lesssim (\eta +n^{-\frac{1}{2}})^{1-\frac{p}{2}}\bra{n^{\frac{2}{d}-1}\mathsf{W}^2\bra{\mu^{\mathcal{B}}+|\mathcal{G}|\rho, Z\rho}}^{\frac{p}{2}}\\ & \lesssim \bra{n^{\frac{2}{d}-1}\mathsf{W}^2\bra{\mu^{\mathcal{B}}+|\mathcal{G}|\rho, Z\rho}}^{\frac{p}{2}}. \end{split}\] Using \eqref{mainclaimdensityhelpsA} for $p=2$ concludes the proof of \eqref{mainclaimdensityhelpsA} also in this case. \end{proof} Finally, we consider the case of a cube $Q_{mL}$ decomposed into cubes of sidelength $L$. The difficulty compared to the previous two cases is to obtain bounds which are independent of $m$. This is achieved using the additional independence for the point processes $\mathcal{U}$, $\mathcal{V}$. While we believe that a direct proof combining Green kernel bounds in the spirit of the proof of \cref{lem:W1pWhitney} together with a Rosenthal type inequality for the (non independent) random variables $\mu^{\mathcal{B}}(Q_i)$ should be possible we give a more elementary proof based on subadditivity and concentration. \begin{proposition}\label{thm:bound-matching} Let $d \ge 3$, $\eta \in (0,1/2)$, $L \ge 1$ and $m \in \mathbb{N}\setminus\cur{0}$. Let $\mathcal{U}$, $\mathcal{V}$ be point processes on $Q_{mL}$ such that the restrictions $(\mathcal{U}_{Q_i}, \mathcal{V}_{Q_i})_{i}$ on all sub-cubes $Q_i= Q_L + Lz_i\subseteq Q_{mL}$, with $z_i \in \mathbb{Z}^d$, are independent copies (translated by the vector $Lz_i$) of the pair of processes $(\mathcal{U}_{Q_L}, \mathcal{V}_{Q_L})$ and such that for every $q\ge 1$, there exists $C(q)>0$ such that \begin{equation}\label{hypUV} \mathbb{E}\sqa{ |\mathcal{U}_{Q_i}|^q +|\mathcal{V}_{Q_i}|^q}\le C(q) L^{d\frac{q}{2}}.\end{equation} Let $\mathcal{N}^{\eta}$, $\mathcal{M}^{\eta}$ independent Poisson processes on $Q_{mL}$ with constant intensity $\eta$, also independent from $(\mathcal{U},\mathcal{V})$. Then, for every $p \in [1, d)$, there exists $C(\eta) = C(\eta,p,d, (C(q))_{q\ge 1})>0$ and $\alpha= \alpha(p,d)>0$, such that, if $L \ge C(\eta)$, \[\mathbb{E}\sqa{ \frac{1}{|Q_{m L}|} \mathsf{M}^p\bra{ \mathcal{U} \cup \mathcal{N}^\eta, \mathcal{V} \cup \mathcal{M}^{\eta} }} \lesssim \eta^{1-\frac{p}{d}} + \frac{ C(\eta)}{L^\alpha}.\] \end{proposition} \begin{remark} Let us preliminarily notice that, for any $R \subseteq Q_{mL}$ that is the disjoint union of $k$ cubes among the cubes $Q_i = Q_L + Lz_i$, $z_i \in \mathbb{Z}^d$, we have the upper bound, if $q \ge 1$, \begin{equation} \label{eq:moment-ur}\begin{split} \mathbb{E}\sqa{ |\mathcal{U}_R|^{q}} & = k^{q} \mathbb{E}\sqa{ \bra{\frac 1 k \sum_{Q_i \subseteq R} |\mathcal{U}_{Q_i}|}^{q}} \le k^{q} \mathbb{E}\sqa{\frac 1 k \sum_{Q_i \subseteq R} |\mathcal{U}_{Q_i}|^{q}} \\ & \lesssim k^{q} L^{d\frac{q}{2}} \lesssim (kL^d)^{q} = |R|^{q}. \end{split}\end{equation} In particular, we have \begin{equation} \label{eq:moment-qml} \mathbb{E}\sqa{ |\mathcal{U}|^{q} } \lesssim m^{dq} L^{d\frac{q}{2}} \lesssim |Q_{mL}|^q. \end{equation} Moreover, by Rosenthal inequalities \cite{rosenthal1970subspaces}, if $q \ge 2$, \[\begin{split} \mathbb{E}\sqa{ \abs{ |\mathcal{U}_R| - \mathbb{E}\sqa{|\mathcal{U}_R|}}^q } & = \mathbb{E}\sqa{ \abs{ \sum_{Q_i\subseteq R} \bra{ |\mathcal{U}_{Q_i}| - \mathbb{E}\sqa{|\mathcal{U}_{Q_i}|}} }^q} \\ & \lesssim k \mathbb{E}\sqa{ \abs{ |\mathcal{U}_{Q_L}| - \mathbb{E}\sqa{|\mathcal{U}_{Q_L}|}}^q } + k^{\frac{q}{2}} \mathbb{E}\sqa{ \abs{ |\mathcal{U}_{Q_L}| - \mathbb{E}\sqa{|\mathcal{U}_{Q_L}|}}^2 }^{\frac{q}{2}} \\ & \lesssim k L^{d\frac{q}{2}} + k^{\frac{q}{2}} L^{d\frac{q}{2}} \lesssim |R|^{\frac{q}{2}}. \end{split}\] We will use all these bounds in the proof below. \end{remark} \begin{proof}[Proof of \cref{thm:bound-matching}] For simplicity, we write throughout the proof $Q$ instead of $Q_{mL}$. As in the previous two proofs, we start from \eqref{triangleMp} and \eqref{commonstartpoint} (with $\rho=I_Q/|Q|$) and estimate by \eqref{eq:matching-iid}, see also \cref{rem:matching}, \[ \mathbb{E}\sqa{\frac{1}{|Q|}\mathsf{W}^p_Q \bra{ \mu^{\mathcal{G}},\frac{|\mathcal{G}|}{|Q|}}}\lesssim \mathbb{E}\sqa{|\mathcal{G}|}^{1-\frac{p}{d}} |Q|^{\frac{p}{d}-1}. \] Since $|\mathcal{G}|\le |\mathcal{N}^\eta|$ we get \[ \mathbb{E}\sqa{\frac{1}{|Q|}\mathsf{W}^p_Q \bra{ \mu^{\mathcal{G}},\frac{|\mathcal{G}|}{|Q|}}}\lesssim \eta^{1-\frac{p}{d}}. \] In order to conclude the proof it is thus enough to show \setcounter{proof-step}{0} \begin{equation}\label{eq:bound-bad-part} \mathbb{E}\sqa{ \frac{1}{|Q|} W_{Q}^p \bra{ \frac{ |\mathcal{G}|}{|Q|} + \mu^{\mathcal{B}}, \frac{Z}{|Q|} } } \lesssim \frac{C(\eta)}{L^\alpha}.\end{equation} We split the proof into several steps. We first consider the case $p\ge 2\ge d/(d-1)$.\\ \medskip \noindent{\emph{Step \stepcounter{proof-step}\arabic{proof-step}. Concentration bounds for $\mu^\mathcal{B}$.}} In this intermediate step, we collect some facts about $\mu^{\mathcal{B}}(R)$, where $R \subseteq Q$ is a disjoint union of $k$ cubes $Q_i = Q_L + Lz_i$, $z_i \in \mathbb{Z}^d$. First of all, the construction of $\mathcal{B}$ ensures that $\mathbb{E}\sqa{ \mu^{\mathcal{B}}(Q_i)}$ does not depend on $Q_i$ (one could in fact prove that $(\mu^{\mathcal{B}}(Q_i))_{i}$ is an exchangeable sequence). We deduce that \begin{equation} \label{eq:constant-mean} \mathbb{E}\sqa{ \mu^{\mathcal{B}}(Q_i)} = \frac{ \mathbb{E}\sqa{ \mu^{\mathcal{B}}(Q)}}{m^d}, \quad \text{hence} \quad \frac{\mathbb{E}\sqa{ \mu^{\mathcal{B}}(R)}}{|R|} = \frac{ \mathbb{E}\sqa{ \mu^{\mathcal{B}}(Q)}}{|Q|}.\end{equation} Indeed, when conditioned on $Z=z$, $|\mathcal{N}^\eta|=n$, $|\mathcal{U}| = u_Q$, and the number of points $|\mathcal{U}_R|=u_R \le u_Q$, $\mu^{\mathcal{B}}(R)$ is the number of ``successes'' in the random sampling procedure, without replacement which we used to define $\mathcal{B}$, with $b = \max\cur{z-n,0}$ draws from an urn containing $ u_Q$ marbles, $u_R$ of which have the desired feature (their extraction defines a success). This is explicitly given by a hypergeometric distribution with parameters $\bra{u_Q, u_R, b}$: given $s_R \le u_R$, \[\mathbb{P}\bra{ \mu^{\mathcal{B}}(R) = s_R | B}= { u_R \choose s_R} {u_Q-u_R \choose b - s_R} / { u_Q \choose b},\] where for brevity we write $$ B = \cur{ Z=z, |\mathcal{N}^\eta|=n_Q, |\mathcal{U}| = u_Q, |\mathcal{U}_R| = u_R}.$$ Specializing to $R = Q_i$, we see that this quantity does not depend on $Q_i$, since $|\mathcal{U}_{Q_i}|$ are i.i.d.\ variables, hence the joint laws of the variables $(Z, |\mathcal{N}^\eta|, |\mathcal{U}|, |\mathcal{U}_{Q_i}| )$ involved in the definition of the law of $\mu^{\mathcal{B}}(Q_i)$ do not depend on $i$. \noindent Using the concentration inequality \eqref{eq:concentration-hyper} for hypergeometric random variables, we have $$ \mathbb{E}\sqa{\abs{ \mu^{\mathcal{B}}(R) - \mathbb{E}\sqa{\mu^{\mathcal{B}}(R)|B}}^p |B} \lesssim \bra{u_R}^{\frac{p}{2}},$$ from which we find, thanks to \eqref{eq:moment-ur} (recall that $p\ge 2$), \begin{equation}\label{eq:density-fluctation} \mathbb{E}\sqa{ \abs{ \mu^\mathcal{B}(R)- \mathbb{E}\sqa{ \mu^\mathcal{B}(R)}} ^p } \lesssim |R|^{\frac{p}{2}}.\end{equation} \noindent{\emph{Step \stepcounter{proof-step}\arabic{proof-step}. Subadditivity bound.}} Using \eqref{eq:constant-mean} and \eqref{eq:density-fluctation} above, we are in a position to follow closely the main argument of \cite[Proposition 5.4]{goldman2021convergence}. We define, for a rectangle $R \subseteq Q$ that is a union of cubes $Q_i$'s, \[ f(R)= \mathbb{E}\lt[\frac{1}{|R|}\mathsf{W}^p_R\bra{ \mu^\mathcal{B} + \frac{|\mathcal{G}|}{|Q|}, \frac{\mu^{\mathcal{B}}(R)}{|R|} + \frac{|\mathcal{G}|}{|Q|} }\rt]. \] We say that $\mathcal{R}$ is an admissible partition of $R$ if it is made of rectangles satisfying the following conditions. Each $R_k\in \mathcal{R}$ is a union of cubes $Q_i$, it is of moderate aspect ratio and $3^{-d}|R|\le |R_j|\le |R|$. We claim that there exists $C_\eta=C(d,p,\eta)>0$ such that for every admissible partition $\mathcal{R}$ of $R$ and every $\varepsilon\in(0,1)$, we have \begin{equation}\label{onestep} f(R)\le (1+\varepsilon)\sum_i \frac{|R_i|}{|R|} f(R_i) +\frac{C_\eta}{\varepsilon^{p-1}} \frac{1}{|R|^{\frac{p(d-2)}{2d}}}. \end{equation} Setting \[ \alpha = \frac{ \mu^{\mathcal{B}}(R)}{|R|}+\frac{ |\mathcal{G}|}{|Q|}, \quad \alpha_i = \frac{ \mu^{\mathcal{B}}(R_i)}{|R_i|}+\frac{|\mathcal{G}|}{|Q|} \] and using \eqref{eq:mainsub}, this reduces to \begin{equation}\label{toproveonestep} \mathbb{E}\sqa{ \frac{1}{|R|}\mathsf{W}^p_{R}\bra{\sum_{i} \alpha_i I_{R_i},\alpha}} \le \frac{C_\eta}{ |R|^{\frac{p(d-2)}{2d}}}. \end{equation} \noindent First, we single out the event $$ A= \cur{ \min\cur{ |\mathcal{N}^\eta|, |\mathcal{M}^\eta|} \ge \eta |Q| /2}.$$ Notice that on $A$, we have $\alpha\gtrsim \eta$. By the concentration bound \eqref{eq:density-bound-below-Poi}, for every $q\ge 1$, $\mathbb{P}(A^c) \lesssim_q (\eta |Q|)^{-q}\le (\eta |R|)^{-q}$. Therefore, if $A^c$ holds, we can use the trivial bound \[\begin{split}\frac{1}{|R|}\mathsf{W}^p_{R}\bra{\sum_{i} \alpha_i I_{R_i},\alpha}& \le \frac{1}{|R|}\mathsf{W}^p_{R}\bra{\sum_{i} \frac{ \mu^{\mathcal{B}}(R_i)}{|R_i|} I_{R_i},\frac{ \mu^{\mathcal{B}}(R)}{|R|}} \\ & \le |R|^{\frac{p}{d}-1} \mu^{\mathcal{B}}(R) \le |R|^{\frac{p}{d}-1} |\mathcal{U}_R|. \end{split}\] Using Cauchy-Schwarz inequality and \eqref{eq:moment-ur} with $q=2$, we get for any $q \ge 1$, \[ \mathbb{E} \sqa{ \frac{1}{|R|}\mathsf{W}^p_{R}\bra{\sum_{i} \alpha_i I_{R_i},\alpha} I_{A^c}} \lesssim_{\eta,q} |R|^{\frac{p}{d}-q}, \] which is estimated by the right-hand side of \eqref{toproveonestep} provided we choose $q$ large enough.\\ \noindent If $A$ holds, we use \eqref{eq:estimCZ} in combination with \eqref{eq:Lp} (recall that for rectangles of moderate aspect ratio the Sobolev constant is uniformly bounded) to get \begin{align*} \frac{1}{|R|}\mathsf{W}^p_{R}\bra{\sum_{i} \alpha_i I_{R_i},\alpha}&\lesssim \frac{ |R|^{\frac{p}{d}-1} }{ \alpha^{p-1}} \sum_{i} |R_i| \abs{\alpha_i - \alpha }^p\\ &\lesssim \eta^{1-p} |R|^{\frac{p}{d}}\sum_{i} \abs{\alpha_i - \alpha }^p. \end{align*} We thus have \[\begin{split} \mathbb{E}\sqa{\frac{1}{|R|}\mathsf{W}^p_{R}\bra{\sum_{i} \alpha_i I_{R_i},\alpha}I_A} & \lesssim \eta^{1-p} |R|^{\frac{p}{d}}\sum_{i}\mathbb{E}\sqa{\abs{\alpha_i - \alpha }^pI_A}\\ & \le \eta^{1-p} |R|^{\frac{p}{d}}\sum_{i}\mathbb{E}\sqa{\abs{\alpha_i - \alpha }^p}.\end{split}\] Using that $\alpha_i-\alpha=\frac{\mu^{\mathcal{B}}(R_i)}{|R_i|}-\frac{\mu^{\mathcal{B}}(R)}{|R|}$, \eqref{eq:constant-mean} and triangle inequality we have \[ \begin{split} \sum_{i}\mathbb{E}\sqa{\abs{\alpha_i - \alpha }^p} & \lesssim \sum_i \mathbb{E}\sqa{\abs{\frac{\mu^{\mathcal{B}}(R_i)}{|R_i|}- \mathbb{E}\sqa{\frac{\mu^{\mathcal{B}}(R_i)}{|R_i|}} }^p} + \mathbb{E}\sqa{\abs{\frac{\mu^{\mathcal{B}}(R)}{|R|}- \mathbb{E}\sqa{\frac{\mu^{\mathcal{B}}(R)}{|R|}} }^p}\\ & \stackrel{\eqref{eq:density-fluctation}}{\lesssim} |R|^{-\frac{p}{2}}. \end{split}\] This proves \[ \mathbb{E}\sqa{\frac{1}{|R|}\mathsf{W}^p_{R}\bra{\sum_{i} \alpha_i I_{R_i},\alpha}I_A}\lesssim \frac{\eta^{1-p}}{ |R|^{\frac{p(d-2)}{2d}}}, \] concluding the proof of \eqref{toproveonestep}. \noindent{\emph{Step \stepcounter{proof-step}\arabic{proof-step}. Dyadic approximation.}} Starting from the cube $Q=Q_{mL}$, we build a sequence of finer and finer partitions of $Q_{mL}$ by rectangles of moderate aspect ratios that are unions of sub-cubes $Q_i$'s. We let $\mathcal{R}_0=\{Q_{mL}\}$ and define $\mathcal{R}_k$ inductively as follows. Let $R\in \mathcal{R}_k$. Up to translation we may assume that $R=\prod_{i=1}^d (0, m_i L)$ for some $m_i\in \mathbb{N}$. We then split each interval $(0,m_i L)$ into $(0,\lfloor\frac{m_i}{2}\rfloor L)\cup(\lfloor\frac{m_i}{2}\rfloor L, m_i L)$. It is readily seen that this induces an admissible partition of $R$. Let us point out that when $m_i=1$ for some $i$, the corresponding interval $(0,\lfloor\frac{m_i}{2}\rfloor L)$ is empty. This procedure stops after a finite number of steps $K$ once $\mathcal{R}_K=\{Q_L+z_i, z_i\in [0,m-1]^d \cap \mathbb{Z}^d\}$. It is also readily seen that $2^{K-1}<m\le 2^K$ and that for every $k\in [0,K]$ and every $R\in \mathcal{R}_k$ we have $|R|\sim (2^{K-k} L)^d$. \\ We prove via a downward induction the existence of $\Lambda_\eta>0$ such that for every $k\in [0,K]$ and every $R\in \mathcal{R}_{k}$, \begin{equation}\label{induction} f(R)\le f(Q_L)+ \Lambda_\eta(1+f(Q_L)) L^{-\frac{d-2}{2}} \sum_{j=K-k}^K 2^{- j\frac{d-2}{2}}. \end{equation} The statement is clearly true for $k=K$, since the law of the point process on each cube $Q_i = Q_L + z_i$ is the same, hence $f(Q_i) = f(Q_L)$. Assume that it holds true for $k+1$. Let $R\in \mathcal{R}_{k}$. Applying \eqref{onestep} with $\varepsilon= (2^{K-k} L)^{-(d-2)/2}\ll1$, we get \[ \begin{split} f(R)&\le (1+ \varepsilon) \sum_{R_i\in \mathcal{R}_{k+1}, R_i\subset R} \frac{|R_i|}{|R|} f(R_i) + \frac{C_\eta}{\varepsilon^{p-1}} \frac{1}{|R|^{\frac{p(d-2)}{2d}}}\\ &\stackrel{\eqref{induction}}{\le} (1+\varepsilon) \lt(f(Q_L)+ \Lambda_\eta(1+f(Q_L))L^{-\frac{d-2}{2}} \sum_{j=K-k+1}^K 2^{- j\frac{d-2}{2}}\rt) \\ & \qquad \qquad + C_\eta(2^{K-k} L)^{-\frac{d-2}{2}}\\ &\le f(Q_L)+ \Lambda_\eta(1+f(Q_L))L^{-\frac{d-2}{2}}\cdot\\ & \qquad \qquad \cdot \lt[\sum_{j=K-k+1}^K 2^{- j\frac{d-2}{2}}+2^{-(K-k)\frac{d-2}{2}}\lt( \frac{C_\eta+1}{\Lambda_\eta}+L^{-\frac{d-2}{2}} \sum_{j=K-k+1}^K 2^{- j\frac{d-2}{2}} \rt)\rt]. \end{split}\] If $L$ is large enough (depending on $\eta$) then $$ \bra{ \sum_{j=K-k+1}^K 2^{- j\frac{d-2}{2}}}\eta^{1-3p} L^{-\frac{(d-2)}{2}}\lesssim_\eta \bra{ \sum_{j=0}^\infty 2^{- j\frac{d-2}{2}}} L^{-\frac{(d-2)}{2} } \le \frac{1}{2}.$$ Finally, choosing $\Lambda_\eta\ge 2(C+1)$ yields \eqref{induction}. Applying \eqref{induction} to $R=Q_{mL}$ and using that $\sum_{j\ge 0} 2^{- j\frac{d-2}{2}}<\infty$, we get \begin{equation}\label{conclusiondyadic} f(Q_{mL})\le f(Q_L)+ \Lambda_\eta(1+f(Q_L)) \frac{1}{L^{\frac{d-2}{2}}}. \end{equation} \noindent{\emph{Step \stepcounter{proof-step}\arabic{proof-step}. Conclusion in the case $p\ge 2$.}} We finally claim that \begin{equation}\label{claimfQL} f(Q_L)\le \frac{C_\eta}{L^{\frac{1}{2}(d-p)}}. \end{equation} Arguing verbatim as in the proof of \eqref{toproveonestep} of {\it Step 5}, we see that it is enough to assume that we are in the event $A= \cur{ \min\cur{ |\mathcal{N}^\eta|, |\mathcal{M}^\eta|} \ge \eta |Q| /2}$. Since in this case $|\mathcal{G}|/|Q| \gtrsim \eta$, \cref{prop:density-helps} yields \[ \begin{split} \mathbb{E}\sqa{ \frac{1}{|Q_L|} \mathsf{W}^p_{Q_L}\bra{ \mu^\mathcal{B} + \frac{|\mathcal{G}|}{|Q|}, \frac{\mu^\mathcal{B}(Q_L)}{|Q_L|}+ \frac{|\mathcal{G}|}{|Q|}} I_A } & \lesssim \frac{1}{|Q_L| \eta^{\frac{p}{d}}} \mathbb{E}\sqa{ \bra{ \mu^\mathcal{B}(Q_L)}^{1+\frac{p}{d}} I_A }\\ & \le \frac{1}{|Q_L| \eta^{\frac{p}{d}}} \mathbb{E}\sqa{ \bra{ |\mathcal{U}_{Q_L}|}^{1+\frac{p}{d}} } \\& \stackrel{\eqref{hypUV}}{\lesssim} \frac{L^{\frac{d+p}{2}}}{L^d \eta^{\frac{p}{d}}} \lesssim \frac{1}{\eta^{\frac{p}{d}}L^{\frac{d-p}{2}}}. \end{split}\] This proves \eqref{claimfQL}. Inserting this into \eqref{conclusiondyadic} finally gives (recall that $p>2$) \[ f(Q)\le \frac{C_\eta}{L^{\frac{1}{2}(d-p)}}. \] This concludes the proof of \eqref{eq:bound-bad-part} with $\alpha=(d-p)/2$ when $p\ge 2$.\\ \noindent{\emph{Step \stepcounter{proof-step}\arabic{proof-step}. The case $p\le 2$.}} If $p\le 2$, we argue as in the previous two proofs and use \eqref{eq:jensen} to obtain (recall that $Z=|\mathcal{G}|+|\mathcal{B}|\le |\mathcal{U}|+|\mathcal{N}^\eta|$) \begin{multline*} \mathbb{E}\sqa{ \frac{1}{|Q|} \mathsf{W}^p_{Q} \bra{ \frac{ |\mathcal{G}|}{|Q|} + \mu^{\mathcal{B}}, \frac{ Z}{|Q|} } } \le \mathbb{E}\sqa{ \bra{ \frac{Z}{|Q|}}^{1-\frac{p}{2}} \bra{ \frac{1}{|Q|} \mathsf{W}_{Q}^2 \bra{ \frac{ |\mathcal{G}|}{|Q|} + \mu^{\mathcal{B}}, \frac{ Z}{|Q|} } }^{\frac{p}{2}} }\\ \le \bra{ \frac { \mathbb{E}\sqa{Z}}{|Q|}}^{1-\frac{p}{2}} \mathbb{E}\sqa{ \frac{1}{|Q|} \mathsf{W}_{Q}^2 \bra{ \frac{ |\mathcal{G}|}{|Q|} + \mu^{\mathcal{B}}, \frac{ Z}{|Q|} } }^{\frac{p}{2}}\\ \lesssim \bra{L^{-\frac{d}{2}}+\eta }^{1-\frac{p}{2}}\bra{ \frac{C(\eta)}{L^\alpha}}^{\frac{p}{2}} \lesssim \bra{ \frac{C(\eta)}{L^\alpha}}^{\frac{p}{2}}, \end{multline*} where in the last step we used \eqref{eq:moment-qml} and \eqref{eq:bound-bad-part} with $p=2$. This concludes the proof of \eqref{eq:bound-bad-part} for any $p<d$. \qedhere \end{proof} \printbibliography \end{document} \end{document}

a982d3098ed5fdcb38342b440ad71ce7b558595f

\section{\label{sec:level1}Introduction} Nonlinear superconducting devices are the basic components of superconducting quantum electronics. For example, inductive nonlinearities provide the anharmonicity essential for the superconducting circuits to operate as qubits \cite{krantz2019quantum, blais2021circuit}. The Josephson junction (JJ) in particular has been widely employed as the source of nonlinearity for superconducting qubits \cite{nakamura1999coherent, mooij1999josephson, van2000quantum, martinis2002rabi, yu2002coherent, vion2002manipulating, koch2007charge, manucharyan2009fluxonium, steffen2010high}. Additionally, Superconducting Quantum Intereference Devices (SQUIDs), which employ JJ nonlinearity, are used for making tunable couplers between qubits \cite{wallquist2006selective, osborn2007frequency, sandberg2008tuning, palacios2008tunable, niskanen2007quantum, chen2014qubit, mckay2016universal, weber2017coherent}, which provide high on-off ratio leading to lower gate error \cite{yan2018tunable, sete2021floating}, and they also enable parametric entangling gates which have high fidelity \cite{roth2017analysis, caldwell2018parametrically, reagor2018demonstration, sete2021parametric}. Such nonlinearity-based tunability has also been implemented in quantum memories for on-demand storing and transferring photons by tuning the coupling to a microwave resonator \cite{yin2013catch, pierre2014storage, sardashti2020voltage}. One of the most important devices that exploits JJ nonlinearity is the Josephson parametric amplifier (JPA), which is widely used for near-quantum-limited measurements \cite{zimmer1967parametric, yurke1989observation, yamamoto2008flux, roy2015broadband, jeffrey2014fast}. JPAs can be used either in the degenerate (phase-sensitive) mode where they can in principle operate without adding any noise \cite{caves1982quantum, clerk2010introduction}, or in the nondegenerate (phase-preserving) mode where they ideally add half a photon of noise to the signal \cite{clerk2010introduction, caves2012quantum}. The former is used for squeezing coherent radiation or squeezing noise below the vacuum level (single mode squeezing) \cite{movshovich1990observation, castellanos2008amplification, zhong2013squeezing}, while the latter can be used for two-mode squeezing \cite{bergeal2010phase, eichler2011observation}. Applications of strongly squeezed states have recently extended beyond sensing to secure quantum communication \cite{pogorzalek2019secure, fedorov2021experimental} and fault tolerant quantum information processing \cite{fukui2018high}. Recently, there has been an increasing interest in the kinetic inductance (KI) nonlinearity of superconducting nanowire (NW) devices made from NbTiN, NbN, and granular Al materials, a nonlinearity which is distributed along the NW \cite{maleeva2018circuit}, as an alternative to JJ-based devices. Nonlinearity distribution provide higher critical current, which can lead to a higher dynamic range and the ability to frequency-multiplex several devices \cite{eom2012wideband, vissers2016low}. Compared to JJ arrays in which the nonlinearity is also distributed over the whole array \cite{castellanos2007widely, castellanos2008amplification, eichler2014controlling, zhou2014high, planat2019understanding}, KI NWs naturally possess the diluted nonlinearity without any explicit junction formation and are relatively simple to fabricate. In addition, higher-order nonlinearities have recently been identified as a limiting factor for gain, quantum efficiency, and squeezing in localized JJ-based devices \cite{kochetov2015higher, boutin2017effect}. The limitations of KI devices are less explored, however, recent work on KI devices has demonstrated nearly 50 dB of gain and 26 dB deamplification of coherent states\cite{parker2021near}. Superconducting quantum electronic elements are rapidly finding applications in areas such as spin-based quantum systems \cite{kobayashi2021engineering, schaal2020fast}, hybrid spin-photon systems \cite{mi2018coherent, borjans2020resonant, harvey2022coherent}, hybrid semiconductor-superconductor systems including Majorana zero modes \cite{mourik2012signatures,de2015realization}, and hybrid magneto-opto-mechanical systems \cite{zoepfl2020single}. What all of these systems have in common is magnetic fields that can easily reach 0.1~T to 1~T or higher. It is well known that even mT magnetic fields, such as the fields present near circulators in conventional setups \cite{lecocq2020microwave}, are detrimental to conventional JJ-based superconducting devices. Therefore, magnetic-field resilient nonlinear superconducting devices would enable the integration of superconducting electronics with these novel quantum systems, allowing them to benefit from tunability, amplification, and squeezing provided by superconducting nonlinear devices. \begin{figure*} \centering \includegraphics [width=0.585\textwidth] {SetupChip.pdf} \includegraphics [width=0.35\textwidth] {Fig1d.pdf} \caption{Device design and measurement. (a) Equivalent circuit diagram of the split-ring resonator in the hanger configuration. The nonlinear kinetic inductance is equivalent to an array of Josephson junctions and the split gap is represented by a capacitance between the terminals of the resonator. The resonator is capacitively coupled to the feedline. The inset is a cartoon for the highly disordered superconducting film. (b) SEM (left) and optical (middle and right) images showing the fabricated device on different scales. Several split-ring resonators are fabricated on each chip in the hanger configuration. The circuit diagram in (a) corresponds to the middle panel. (c) Schematic of measurement setup. Signals generated by VNA and/or microwave sources are combined and attenuated by 50 dB then sent to one of three chips placed at 8 mK. The output is amplified by a low-noise HEMT amplifier of 40 dB gain at 4 K and measured by VNA at room temperature. Short DC pulses are used to select the chip through latching microwave switches. (d) Resonance frequency and intrinsic quality factor versus magnetic field in the linear regime for the 0.1~$\mu$m wide KI NW resonator.} \label{fig:Overall} \end{figure*} Recently, magnetic field compatible nonlinearity has been explored in InAs-based Josephson Junctions \cite{pita2020gate} and granular Al NW junctions \cite{winkel2020implementation}, pushing the magnetic field range of operation up to 100s of mT. Although the resilience of linear KI resonator devices to magnetic fields is well established \cite{samkharadze2016high, kroll2019magnetic, yu2021magnetic}, the KI nonlinearity and devices based on it have yet to receive experimental attention at high magnetic fields. Here, we investigate, in magnetic fields as high as 3~T, the Kerr four-wave mixing (4WM) nonlinearity coefficient of NW KI resonators (without DC current bias) and their behaviour as degenerate and nondegenerate parametric amplifiers, for three different KI NW widths 0.1~$\mu$m, 0.3~$\mu$m, and 1~$\mu$m. We find that increasing magnetic field causes an increase in internal quality factor and parabolic decrease in resonance frequency, consistent with previous reports \cite{zollitsch2019tuning}. For the 0.1 $\mu$m width, the quality factor increases from $30,000$ at $B=0$ to $70,000$ at 2~T. We also characterize the nonlinearity by measuring the self-Kerr coefficients in magnetic fields. The coefficients for the two 1~$\mu$m-wide wires exhibit a near-monotonic increase with increasing magnetic field up to 3~T, by ~20\% and ~60\%, substantially larger than theoretical predictions. For the narrower wires, the dependence on magnetic field is not monotonic. Next, we demonstrate both nondegenerate and degenerate parametric amplification at 0 and 2~T, for the 0.1 $\mu$m width. Similar saturation gains are observed at $\sim$2~T and 0~T, but we find that the saturation gain is reached at a smaller pump power at 2~T, because of the combined effect of the significant increase in internal quality factor, and the small increase in Kerr coefficients, at 2~T. {\color{black}For the 1~$\mu$m-wide resonators, the saturation gain decreases by 6 dB when magnetic field increased from 0 to 2 T.} \section{\label{sec:level1} Methods} Our resonators are made of a $t=10$ nm thick NbTiN superconducting film sputtered on a float-zone silicon substrate having $20$~k$\Omega\cdot$cm resistivity. The film is patterned by electron-beam lithography (EBL) into several split-ring NW resonators coupling capacitively to a CPW feedline of 50 $\Omega$ characteristic impedance, in the hanger configuration \cite{gao2008experimental, khalil2012analysis, megrant2012planar, samkharadze2016high, niepce2019high, basset2019high, yu2021magnetic}. The NbTiN film is etched by reactive ion etcher (RIE) in SF$_6$/O$_2$ plasma. The nonlinear KI in such a nanowire superconducting film is often envisioned as the network of nanometer-scale superconducting islands that act as interconnected JJs. Figure \ref{fig:Overall}(a) depicts the equivalent circuit diagram of a single capacitively coupled split-ring resonator as an array of JJ shunted by capacitors. This circuit model accurately describes the dispersion relation and nonlinearity of high KI resonators \cite{maleeva2018circuit}. The nonlinearity (JJ-array like behaviour) arises from the strong disorder nature of the NbTiN film, which promotes the KI to be stronger than conventional magnetic inductance in our devices. Small junctions are formed between superconducting grains, as illustrated in the inset of Fig.\ref{fig:Overall}(a). The critical temperature and sheet resistance of the film are measured to be $T_c=9.7$ K and $R_{\square} =340$ $\Omega/\square$, which gives sheet kinetic inductance $L_{\square} =48$ pH/$\square$ according to the relation $L_{\square} \approx \hbar R_{\square} / \pi \Delta_0$ \cite{annunziata2010tunable}, where $\Delta_0=1.764 k_B T_c$ is the superconducting energy gap at zero Kelvin and $k_B$ is Boltzmann constant. The critical current density of the film is measured to be $J_c=15$ mA/$\mu$m$^2$. Resonators of three widths 0.1, 0.3 and 1 $\mu$m are fabricated on three chips using the same source film. The estimated characteristic impedance of the three sets of resonators is $Z_r\approx3$, 1.5 and 0.7 k$\Omega$, respectively. The chips with the 0.1 $\mu$m and 1 $\mu$m resonators are etched at the same time, and the chip with the 0.3 $\mu$m resonators is etched separately. All chips are tested in a dilution refrigerator at 8 mK. The resonators are designed to have resonance frequencies in the range between 3 and 6 GHz, by varying their length $l_r$, and to have coupling quality factors between $10^3$ and $10^5$, by varying their distance from the feedline. A chip similar to the measured ones is shown in Fig.\ref{fig:Overall}(b) at three different scales. The transmission through the feedline is measured using a Vector Network Analyzer (VNA), whose signal is attenuated by total of 50 dB at the input of the feedline and the output is amplified by a low-noise high-electron-mobility transistor (HEMT) amplifier at the 4K stage, as shown by the measurement setup in Fig.\ref{fig:Overall}(c). \section{Results and Discussion} \subsection{Resonators Characterization in Linear Regime} The measured transmission spectrum near resonance is fitted to the model in Refs. \cite{swenson2013operation, anferov2020millimeter, basset2019high}, which allows us to extract the low-power resonance frequency $\omega_{0}$, the intrinsic quality factor $Q_i$, the coupling quality factor $Q_c$ and the dimensionless power-dependent nonlinearity parameter $\xi$, as explained in Appendix \ref{app:NLResp}. The resonators are first characterized in the linear regime (i.e. $\xi \approx 0$) to determine the dominant source of intrinsic loss. At zero magnetic field, $Q_i$ exhibits the typical behavior for a resonator dominated by two-level-system (TLS) loss \cite{wang2009improving, sage2011study, sandberg2012etch}, with $Q_i \approx 4,000$ for unsaturated TLS and $30,000$ for totally saturated TLS for the 0.1 $\mu$m resonator (see appendix \ref{app:Losses}). By ramping up the magnetic field we see two effects. Figure \ref{fig:Overall}(d) shows $\omega_0$ and $Q_i$ as a function of magnetic field $B$ at an excitation power that is high enough to saturate TLS but not too high to maintain the linear behavior. The resonance frequency decreases quadratically with the field strength according to $\Delta \omega_0/\omega_0 = -(\pi/48)[De^2t^2B^2/(\hbar k_B T_c)]$ \cite{samkharadze2016high, kroll2019magnetic, yu2021magnetic}, where $D$ is the electronic diffusion constant and $e$ is the electron charge constant. By fitting the data to this relation we get $D\approx 5$ cm$^2$s$^{-1}$, which is the same order of magnitude as the previously reported value for NbTiN \cite{samkharadze2016high}. The dip in $Q_i$ at $B \approx 0.2$ T, also observed before\cite{samkharadze2016high} is because of the coupling to the magnetic centers in the Si substrate. For magnetic fields above this dip, $Q_i$ increases with $B$ up to 2 T. A similar behavior has been reported before for narrow NbTiN resonators \cite{samkharadze2016high, zollitsch2019tuning}. This indicates that the resonator losses are dominated by the magnetic centers in the substrate even at zero magnetic field \cite{zollitsch2019tuning}, and that losses coming from magnetic vortices are negligible. The observed enhancement in the internal quality factor improves the performance of linear devices, and as we will see, can improve the performance of nonlinear KI devices in magnetic field. \subsection{Kerr Nonlinearity} In our resonators, the dominant nonlinearity is 4WM (the Kerr effect). Three-wave mixing (3WM) is anticipated to be negligible because no DC current or flux is applied, so it is not explored. Different applications have different requirements on the strength of the nonlinearity. For example, in KI detectors the nonlinearity is a side effect and the Kerr coefficient should be as low as possible, while for qubits, large Kerr coefficients are preferable since they isolate the computation space, \textit{e.g.}, Kerr coefficients of 100-400 MHz are typical for transmons \cite{blais2021circuit}. Parametric amplification falls between these two limits, where Kerr coefficient has to be higher than the nonlinear loss to maximize gain, but should be lower than the coupling loss to the feedline to maximize the dynamic range\cite{eichler2014controlling}, as discussed in next section. The nonlinearity of the KI resonator was studied in detail in Ref. \cite{yurke2006performance}. The Hamiltonian of the resonator is \begin{equation} \label{eqn:Hamil} \begin{split} H = & \sum\limits_{n} \hbar \omega_n \left(a_n^\dagger a_n + \frac{1}{2}\right) + \sum\limits_{n} \frac{\hbar K_n}{2} \left( a_n^\dagger a_n \right)^2 \\ & + \sum\limits_{n,m \ne n} \frac{\hbar K_{mn}}{2} \left( a_m^\dagger a_m a_n^\dagger a_n \right) \end{split} \end{equation} where $\omega_n$ and $K_n$ are the resonance frequency and self-Kerr coefficient of mode $n$, respectively, $K_{mn}$ is the cross-Kerr coefficient between modes $m$ and $n$, and $a_n$ ($a_n^\dagger$) is the annihilation (creation) operator of the intra-resonator field of mode $n$. \begin{figure*} \centering \includegraphics [width=0.34\textwidth] {Fig2a.pdf} \includegraphics [width=0.29\textwidth] {Fig2b,c.pdf} \includegraphics [width=0.34\textwidth] {Fig2d.pdf} \caption{Kerr nonlinearity measurement. (a) $S_{21}$ measurements (dots) and fits (lines) for a resonator of 0.1 $\mu$m width at different powers; -90 dBm ($\xi=-0.03 \ll \xi_b$, negligible nonlinearity), -85 dBm ($\xi=-0.3 < \xi_b$, below bifurcation) and -83 dBm ($\xi=-0.46 > \xi_b$, above bifurcation). Lines are separated by 10 dB for clarity. The dashed part of the line at -83 dBm denotes the unstable states above bifurcation. (b) Shift in resonance frequency by varying input power versus estimated number of photons in the resonator for three resonators of different widths. Markers denote measurement points while solid lines are linear fits for extracting the self-Kerr coefficients: $K_0/2\pi = -0.584$, $-0.122$ and $-0.0108$ Hz for $w_r=0.1$, 0.3 and 1 $\mu$m, respectively. (c) Shift in resonance frequency $\omega_1/2\pi\approx 6.11$ GHz ($\omega_0/2\pi\approx 3.08$ GHz) of the first excited (fundamental) mode when pumping at the fundamental (first excited) mode versus estimated number of pump photons in the resonator for a resonator of width 0.1 $\mu$m. Markers denote measurement points while solid lines are linear fits for extracting the cross-Kerr coefficients: $K_{01}/2\pi = -1.406$ Hz and $K_{10}/2\pi = -1.398$ Hz. The pump power level at the first excited mode used for calculating $K_{10}$ is estimated relative to that at the fundamental mode from the relative values of the self-Kerr coefficients $K_0$ and $K_1$ (see text for details). (d) Extracted self-Kerr coefficients versus magnetic field for six resonators of widths 0.1, 0.3 and 1 $\mu$m, two of each width.} \label{fig:Nonlinear} \end{figure*} We characterize the nonlinearity of our resonators by experimentally measuring their self- and cross-Kerr coefficients. This requires estimating the average number of photons in the resonator corresponding to a certain driving power at the feedline using the extracted fitting parameters. Figure \ref{fig:Nonlinear}(a) displays the transmission response at three different powers. At low power ($\xi \approx 0$) the resonance frequency is $\omega_r = \omega_0$ and the resonator operates in the linear regime (top curve in Fig.\ref{fig:Nonlinear}(a)). By driving at a higher power and at a frequency closer to $\omega_r$, more photons couple to the resonator driving it into the nonlinear regime. As the average number of photons in the resonator increases, $\omega_r$ red-shifts resulting in the asymmetric resonance lineshape (middle curve in Fig.\ref{fig:Nonlinear}(a)). Above a certain power, corresponding to the onset of bifurcation ($\xi_b=-2/\sqrt{27}$ ignoring nonlinear loss \cite{anferov2020millimeter}), the resonator goes into the bistability regime, where the resonator has three valid states at the same frequency with different amplitudes; two of which are stable and one is unstable (bottom curve in Fig.\ref{fig:Nonlinear}(a)). This bistability feature has been used for digital threshold quantum detection in Josephson bifurcation amplifier, where the sensitivity is only limited by quantum fluctuations \cite{siddiqi2004rf, mallet2009single, vijay2009invited}. The red-shift in $\omega_r$ at increasing power is proportional to the average number of photons in the resonator, $n_{\text{ph}}$, such that $\Delta \omega_r=K_0 n_\text{ph}/2$, where $K_0$ is self-Kerr coefficient depending on the critical current and kinetic inductance of the resonator. The value of $n_{\text{ph}}$ is determined for each input power as explained in Appendix \ref{app:NLResp}. We do the measurement at different powers and extract $K_0$ from the slope of the linear fit of $\Delta \omega_r$ versus $n_{\text{ph}}$. In this work, we focus on the trends with magnetic field and NW width, rather than the exact values of Kerr coefficients, so that uncertainties in $n_{\text{ph}}$, which is difficult to precisely determine, do not impact interpretation of the data. Figure \ref{fig:Nonlinear}(b) presents $\Delta \omega_r$ versus $n_{\text{ph}}$ for three resonators of widths 0.1, 0.3 and 1 $\mu$m. The decrease in self-Kerr parameter $|K_{0}|$ as the resonator width $w_r$ increases comes from the dependence of $K_{0}$ on the resonator critical current $I_c=J_cw_rt$ and kinetic inductance $L_k=L_{\square}l_r/w_r$, where $|K_{0}| \propto \omega_0^2/L_k I_c^2$ \cite{yurke2006performance, anferov2020millimeter}. Next, we extract the cross-Kerr coefficients which describe the strength of the intermode coupling in the resonator. This coupling can be used in circuit quantum electrodynamics for nondemolition measurements \cite{buks2006dephasing, helmer2009quantum, suchoi2010intermode, kumar2010exploiting}. One approach is to measure the number of photons in a certain mode, called signal mode, by driving another mode, called detector mode, and detecting its dephasing \cite{buks2006dephasing, suchoi2010intermode} or frequency shift \cite{kumar2010exploiting}. The sensitivity of this scheme is determined by the magnitude of the cross-Kerr coefficient between the two modes \cite{buks2006dephasing}, where according to (\ref{eqn:Hamil}) the resonance shift of mode $n$ induced by pumping at mode $m$ is $\Delta \omega_r^{(n)} = K_{mn} n_\text{ph}^{(m)}/2$. To extract $K_{mn}$ for $m\ne n$, we apply a two-tone technique \cite{tancredi2013bifurcation, weissl2015kerr, krupko2018kerr, yu2021magnetic} by pumping mode $m$ with high power using a MW source ($\omega_p \sim \omega_m$) while probing mode $n$ with low power using the VNA, as depicted in Fig. \ref{fig:Overall}(c). We sweep the VNA frequency around $\omega_n$ for different pump powers at $\omega_m$. The shifts in the resonance frequency of the fundamental mode (i.e. $n=0$) of a resonator of width 0.1 $\mu$m when pumping at the first excited mode (i.e. $m=1$) and vice versa (i.e. $n=1$ and $m=0$) are shown in Fig. \ref{fig:Nonlinear}(c) as a function of the number of pump photons in the resonator. A meaningful conclusion on the relation between $K_{01}$ and $K_{10}$, or between either of them and $K_0$, requires the power levels at $\omega_0$ and $\omega_1$ to be estimated accurately with respect to each other. The ratio between the self-Kerr coefficients can be used for this purpose since we know that $|K_n| \propto \omega_n^2$ \cite{yurke2006performance}. We take the power level at $\omega_0$ as a reference and calibrate the estimated power at $\omega_1$ such that $K_{1}/K_{0}=(\omega_1/\omega_0)^2=4$. The extracted values for $K_{01}$ and $K_{10}$ after power calibration are almost equal, as shown in Fig. \ref{fig:Nonlinear}(c), which agrees with previous theoretical predictions \cite{yurke2006performance}. We also find that the ratio $K_0/K_{01} \approx 0.4$, which is close to 0.375 calculated in \cite{maleeva2018circuit}. The nonlinearity of the NW resonators is measured as a function of applied in-plane magnetic field (Fig. \ref{fig:Overall}(d)). The value of $|K_{0}|$ for two different 1~$\mu$m wide NW resonators monotonically increases by around $20\%$ and $60\%$ for magnetic fields up to $\sim 3$ T, the highest magnetic field we studied on those chips (Fig.~\ref{fig:Nonlinear}d, bottom). Notably, Kerr coefficients extracted from multiple 0.1 and 0.3 $\mu$m NW resonators do not exhibit a monotonic upward trend in $|K_0|$. Rather, $|K_0|$ increases for some values of $B$, and decreases in other regions, and exhibits a larger relative scatter in best-fit values compared to the 1 $\mu$m NWs (Fig.~\ref{fig:Nonlinear}(d), top and middle). Despite the presence of some scattering of data, the overall trend of for $|K_0|$ to increase between 0 and 2 T (2.5 T) for the 0.1 $\mu$m (0.3 $\mu$m) resonators. We compare the measured variation of the Kerr coefficient $K_0$ with magnetic field $B$ to the variation calculated from the theory of superconductivity. The Kerr coefficient depends on $T_c$, which decreases with increasing magnetic field due to breaking the time-reversal degeneracy of Cooper pairs \cite{tinkham2004introduction}. The relation between $K_0$ and $T_c$ originates from the dependence of $K_0$ on the superconducting parameters $I_c$ and $L_{\square}$, which both vary with $T_c$ \cite{annunziata2010tunable}, and it is given by (see Appendix \ref{app:KVsB}) \begin{equation}\label{eqn:Del_K} \frac{\Delta K_0}{K_0} = - \frac{\Delta T_c}{T_c} = \frac{\pi}{24} \frac{D e^2 t^2}{\hbar k_B T_c} B^2. \end{equation} According to this relation, the expected change in $K_0$ at $B=2$~T is only $\sim$0.1\%. Therefore, the observed increase in $|K_0|$ for the $1$~$\mu$m NWs (Fig.~\ref{fig:Nonlinear}(d), bottom) is more than 100 times larger than the theoretically predicted increase in $|K_0|$. It is difficult to pinpoint the cause for this enhancement compared to theoretical predictions. We speculate that it may reflect inhomogeneity present in the sputtered NbTiN film. For example, weak links, which may be viewed as sections of the NW with smaller critical currents and temperatures, could contribute to the enhancement of the Kerr coefficient. The less systematic variation in Kerr coefficients with magnetic field for the narrower wires may also be a signature of film or nanowire etching inhomogeneity. \subsection{Nondegenerate Parametric Amplification} \begin{figure*} \centering \begin{minipage}[b]{.6\linewidth} {\includegraphics [width=1\textwidth]{Fig3a.pdf}} {\includegraphics [width=1\textwidth]{Fig3d.pdf}} \end{minipage}% \hfill \begin{minipage}[b]{.4\linewidth} {\includegraphics [width=1\textwidth] {Fig3b,c.pdf}} \end{minipage}% \caption{Nondegenrate Parametric Amplification. (a) Measurement and fitting of signal gain versus reduced pump-resonance detuning, $\tilde{\delta}=(\omega_p-\omega_0) /(\kappa_c+\kappa_i)$, at different pump powers for a 0.1 $\mu$m-wide resonator. The highest-power line exhibits the bifurcation effect, where the solid part represents the fitting of the measured states, the dashed part represents the none-measured stable states, and the dotted part represents the unstable states. The inset shows the frequency configuration of the process (see text for details). (b), (c) Measured signal gain versus pump power, $P_p$, and reduced pump-resonance frequency detuning, $\tilde{\delta}$, at $B=0$ and 1.7 T, respectively, for the 0.1 $\mu$m resonator. The measurement is taken at downward pump frequency sweeping at each pump power. The marked-lines highlight the maximum gain for each pump power. (d) Maximum gain versus pump power for two resonators of widths 0.1 and 1 $\mu$m at zero and high and magnetic fields. The gain is maximized over the pump frequency. The two lines of the 0.1 $\mu$m resonator correspond to the lines highlighted in (b) and (c). The inset is the maximum gain over both pump frequency and power of the two resonators versus magnetic field. The solid lines are guide for the eye.} \label{fig:NDParaAmp} \end{figure*} A KI resonator can be operated as a parametric amplifier when pumped by a high-power tone. Here we consider a 4WM process, where the signal is slightly detuned from the pump within the linewidth of the same mode and the amplification occurs by the conversion of two pump photons into signal and idler photons. The performance of such parametric amplifiers was analyzed theoretically \cite{yurke2006performance} and measured experimentally \cite{tholen2007nonlinearities}. In the stiff-pump approximation, where the pump power at the output is assumed to be equal to the pump power at the input, the signal gain in the hanger configuration (Fig. ~\ref{fig:Overall}(a)) is given by \cite{anferov2020millimeter} \begin{equation} \label{eqn:gs} g_s = 1 - \left(\frac{e^{i\phi}}{\cos \phi} \right) \left(\frac{\kappa_c}{\kappa_c+\kappa_i} \right) \frac{1/2 -i\left(\tilde{\delta}-2\xi n-\tilde{\Delta}\right)}{2\left(\lambda_+ + i\tilde{\Delta}\right)\left(\lambda_- + i\tilde{\Delta}\right)} \end{equation} where $\kappa_c = \omega_0/Q_c$ and $\kappa_i = \omega_0/Q_i$ are the coupling and intrinsic loss rates, respectively, $\xi$ is the reduced power-dependent nonlinearity parameter, $\eta$ is the reduced power-dependent nonlinear loss, $n$ is the reduced number of photons in the cavity, $\tilde{\delta}=(\omega_p - \omega_0)/(\kappa_c+\kappa_i)$ is the reduced pump-resonance detuning, $\tilde{\Delta}=(\omega_s - \omega_p)/(\kappa_c+\kappa_i)$ is the reduced signal-pump detuning, $\phi$ is a phase representing the asymmetry in the feedline and $\lambda_{\pm} = 1/2 \pm \sqrt{(\xi n)^2+ - (\tilde{\delta}-2\xi n)^2}$. We measure the performance of the KI resonators as nondegenerate (phase insensitive) parametric amplifiers in magnetic field. We stimulate a 4WM process by supplying a pump tone detuned from the resonance frequency by $\delta = \omega_p-\omega_0$ and a low-power signal detuned from the pump tone by $\Delta = \omega_s-\omega_p$, where $|\Delta| \ll \kappa_c+\kappa_i$. Inside the resonator two pump photons at $\omega_p$ are converted into two photons at $\omega_s$ and $2\omega_p-\omega_s$, leading to the amplification of the signal and the generation of a new tone (idler) at frequency $\omega_i = 2\omega_p-\omega_s$, as illustrated in the inset of Fig. \ref{fig:NDParaAmp}(a). The amplified output power of the signal is then measured by a spectrum analyzer using a measurement bandwidth well below $|\Delta|$. We define the signal gain at $\omega_s=\omega_p+\Delta$ as the ratio of the signal output power at $\omega_s$ when pump is on to the signal output power at $\omega'_s$ when pump is off, where $\omega'_s$ is far from resonance. The signal output power at $\omega'_s$ when pump is off is used as an approximation for the signal input power at $\omega_s$ (the signal output power at $\omega_s$ when pump is off does not reflect the signal input power at $\omega_s$ when $\omega_s$ is close to the resonance frequency). The gain data for $\Delta=10$ kHz for a resonator of width 0.1 $\mu$m at $B=0$ T is fitted with the expression in (\ref{eqn:gs}) and plotted in Fig. \ref{fig:NDParaAmp}(a) at different pump powers. The gain increases with pump power until the onset of bifurcation where the system enters the bistability regime. The parametric amplifier exhibits large gain when operated at pump power and frequency near the bifurcation point. The condition for large amplification is that the system can access the bifurcation regime, which is satisfied when the nonlinear (i.e. two-photon) loss rate is smaller than $|K_0|/\sqrt{3}$ \cite{yurke2006performance}. \begin{figure} \centering \includegraphics [width=1\linewidth] {FigDR.pdf} \includegraphics [width=1\linewidth] {FigNL.pdf} \includegraphics [width=1\linewidth] {FigNF.pdf} \caption{Parametric Amplifier Characterization. (a), (b) Gain as a function of signal power $P_s$ for the 1 $\mu$m and the 0.1 $\mu$m resonators, respectively. In (a), the data denoted by {\smaller$\CIRCLE$} and {\smaller$\blacksquare$} at $B=0$ T correspond to different pump powers ($P_p =-75.3$ and $-75.8$ dBm, respectively), hence different gain values. (c) Noise power referred to the input of the 0.1 $\mu$m KI PA at $B=0$ T in the presence of a coherent tone when pump is on at 9 dB gain and when it is off. (d), (e) Noise figure as a function of gain for the 1 $\mu$m and the 0.1 $\mu$m KI PAs, respectively, at zero and high magnetic fields.} \label{fig:DRNF} \end{figure} \begin{figure*} \centering \includegraphics [width=0.195\textwidth] {Fig4a.pdf} \includegraphics [width=0.39\textwidth] {Fig4b,c.pdf} \includegraphics [width=0.39\textwidth] {Fig4d.pdf} \caption{Degenerate Parametric Amplification. (a) Schematic of the output signal phase space without (top panel) and with (bottom panel) pump tone. The former is a coherent state whose amplitude is independent of phase, while the later is a squeezed state whose amplitude depends on its phase relative to the pump phase. The inset shows the frequency configuration of the experiment. The signal is amplitude-modulated at $\omega_M$ and the gain is measured at the modulation sideband. (b),(c) Measured signal gain as a function of the relative phase between the signal and pump tones at $B=0$ and $2$ T, respectively. The points at which gain is above (below) zero indicate the amplification (deamplification). (d) Maximum gain versus pump power. The gain is maximized over phase and frequency at each power value. The solid lines are guide for the eye.} \label{fig:DParaAmp} \end{figure*} Figures \ref{fig:NDParaAmp}(b) and (c) show the full measurement of the gain versus pump power and pump frequency for the same resonator of width 0.1 $\mu$m at 0 and 1.7~T, respectively. In both cases, gain is observed in the NW resonator. The marked points which are connected by solid lines in Fig.~\ref{fig:NDParaAmp}(b) and (c) indicate the frequency where the maximum gain is obtained for each pump power. The maximum gain is plotted in Fig. \ref{fig:NDParaAmp}(d) versus the pump power for the $0.1$~$\mu$m and $1$~$\mu$m NW resonators. At zero magnetic field, the maximum gains of the two resonators are almost equal, $\sim$12.5 dB. However, to reach this gain, the 1 $\mu$m resonator requires higher pump power due its lower Kerr coefficient. This verifies that the maximum gain is insensitive to the Kerr nonlinearity, so that high gains can still be achieved for resonators with arbitrarily small Kerr-coefficients if pumped by sufficiently high power. However, as mentioned before, pumping with high power is not desirable since it heats up the resonator by generating thermal quasi-particles. Above a certain threshold, the quasi-particle losses will dominate the total losses of the resonator decreasing the intrinsic quality factor and increasing the noise level. {\color{black}On the other hand, the gain of the two resonators is quite different at high magnetic field. The maximum gain of the 0.1 $\mu$m resonator is almost not affected at 1.7 T, except for a small shift in the power at which it is observed, while the maximum gain of the 1 $\mu$m is significantly suppressed at 2 T. The measured maximum gain is plotted versus magnetic field in the inset of Fig. \ref{fig:NDParaAmp}(d). The 0.1 $\mu$m resonator exhibits a stable gain over the whole measurement range of $B$, which is limited by the microwave switches in our setup, while the gain of the 1 $\mu$m resonator degrades gradually with $B$. To confirm this result, another 1 $\mu$m resonator was measured and similar degradation for its gain versus $B$ was observed (not shown).} The shift in the power at which the maximum gain is observed at high field, (see Fig. \ref{fig:NDParaAmp}(d) main panel), is due to the fact that the resonators have higher $Q_i$ and slightly higher $|K_0|$ in magnetic field, as illustrated in Fig. \ref{fig:Overall}(d) and Fig. \ref{fig:Nonlinear}(d), respectively. The higher $Q_i$ translates into a larger average number of photons in the resonator given a certain pump power, while the higher $|K_0|$ translates into a larger $\xi$ given a certain number of photons in the resonator. {\color{black}Most importantly, these results demonstrate the sensitivity of the gain levels obtained at high magnetic field to the resonator width, where narrower KI devices can maintain their high gains up to higher values of magnetic field.} For operating points near the bifurcation, gain can additionally be limited by pump depletion, where as the signal output power approaches the pump power, the gain begins to saturate. This defines the dynamic range of the device and it is limited by the pump power at which the bifurcation occurs. For a lossless device ($\kappa_i \ll \kappa_c$), the bifurcation power is inversely proportional to $|K_0| Q_c^2$ (see Appendix \ref{app:NLResp}). Therefore lower $|K_0|$ and $Q_c$ gives larger dynamic range. However, this requires operating at higher pump power leading to the excitation of thermal quasi-particles and lowering $Q_i$ \cite{eichler2014controlling}. {\color{black}We determine the dynamic range of our devices by measuring their gain as a function of input signal power $P_s$, and observing the signal power at which 1-dB compression of the gain occurs. In Fig. \ref{fig:DRNF}(a) we show the data for the 1 $\mu$m resonator for two different gains of 11 and 6 dB, determined by pump power and frequency, at 0 T, and for a gain of 6 dB at 2 T. The 11 dB gain gives lower 1-dB compression point, \textit{i.e.} smaller dynamic range, $\sim-110$ dBm compared to $-105$ dBm for the 6 dB gain, which agrees with both theory and experiment in literature \cite{eichler2014controlling, parker2021near}. The presence of magnetic field does not affect the 1-dB compression point as long as the device operates at a similar gain, as shown in Fig. \ref{fig:DRNF}(a) for the case of 6 dB gain with and without field. The 0.1 $\mu$m KI PA also exhibits a dynamic range that is insensitive to magnetic field, where the 1-dB compression point occurs at $\sim-118$ dBm at both 0 and 1.7 T, as shown in Fig. \ref{fig:DRNF}(b) for a gain of 11 dB. The smaller dynamic range of the 0.1 $\mu$m KI PA compared to the 1 $\mu$m one is attributed to the larger Kerr coefficient of the 0.1 $\mu$m resonator.} {\color{black}We also characterize the noise properties of our KI PAs. Figure \ref{fig:DRNF}(c) shows the noise power referred to the input of the 0.1 $\mu$m KI PA in the presence of a coherent input signal at $\sim-120$ dBm. The input referred noise decreases by roughly the same amount of the gain when the resonator is pumped for amplification, improving the signal-to-noise ratio (SNR) by the same factor. Similar results are obtained at high magnetic field and for the 1 $\mu$m resonator. In addition, we measure the noise figure as a function of gain. We define the noise figure as the ratio between the SNR when pump is off, SNR$_\text{off}$, and the SNR when pump is on at a certain gain, SNR$_\text{on}(G)$, where in this definition the SNR$_\text{off}$ is used as an indication of the SNR at the input of the KI PA. The results for the 1 $\mu$m and the 0.1 $\mu$m KI PAs are shown in Figs. \ref{fig:DRNF}(d) and (e), respectively, at zero and high magnetic fields. The noise figure decreases linearly with gain for both KI PAs over the whole range of amplification, even with the existence of the magnetic field. This means that the gains of these resonators are not high enough to reach the saturation point of SNR$_\text{on}(G)$, where the noise at the mixing chamber, either from quantum fluctuations or added noise by the KI PA, dominates over the noise added by the HEMT or thermal noise at higher stages. Despite this limitation in our devices, we can conclude that the magnetic field did not significantly enhance the added noise of the 0.1 $\mu$m KI PA up to 12 dB gain.} {\color{black}These results show that, other than the gain suppression of the 1 $\mu$m resonator, the high magnetic field does not affect the characteristics of our KI PAs. Therefore, KI devices of sufficiently narrow widths, $\sim$0.1 $\mu$m, are capable of operating efficiently as PAs at high magnetic fields up to $\sim$2 T.} \subsection{Degenerate Parametric Amplification} Finally we explore the phase-sensitive amplification of our NW resonators in magnetic field. Here the device operates in the degenerate mode, where the pump and signal tones are set to the same frequency ($\Delta = 0$), so the idler is generated at the same frequency. This results in squeezing the output phase space, with squeezing level and direction determined by the power and phase of the pump tone, respectively, as illustrated by the schematic in Fig. \ref{fig:DParaAmp}(a). The magnitude of the output signal is then either amplified or deamplified depending on its phase relative to the pump phase. Similarly the variance of the output signal undergoes the same effect, which can be used for squeezing thermal and vacuum noise. Squeezing level up to 26 dB has been recently achieved using a KI-based degenerate parametric amplifier \cite{parker2021near} operating in zero magnetic field. In order to probe the gain of our NW resonators in the degenerate regime, we amplitude-modulate the signal with a sinusoidal wave of small frequency $\omega_M$ and measure the gain at the sideband with measurement bandwidth much less than $\omega_M$. We use $\omega_M=20$ kHz, and we adjust the pump frequency for maximum gain. Similar to the case of nondegenerate amplification, the gain here is obtained by comparing the output sideband power when the pump is on to that when the pump is off at a reference point far from resonance. Figures \ref{fig:DParaAmp}(b) and (c) show the measured gain versus $\Delta \phi = \phi_s-\phi_p$ for a resonator of width 0.1 $\mu$m at 0 and 2 T with estimated pump power at the resonator $-81$ and $-83$ dBm, respectively. In both figures the gain oscillates at phase period $\pi$, such that the gain varies from a maximum value of 7.5 dB at $\Delta \phi=0$ to a minimum value of -7.5 dB at $\Delta \phi = \pm\pi/2$. In Fig. \ref{fig:DParaAmp}(d) the maximum gain per phase difference and pump frequency is plotted as a function of pump power at 0 and 2 T. The gain peaks at the bifurcation power with almost the same value of $\sim16$ dB at 0 and 2 T. Similar to the nondegenerate case, there is a power shift between the measurements at 0 and 2 T, due to the combination of higher $Q_i$ and slightly higher $|K_0|$ at 2~T. These results demonstrate the ability of KI resonators to operate as phase-sensitive parametric amplifiers at high magnetic field without compromising their amplification or squeezing capabilities. \section{Conclusion} We experimentally investigate the ability of superconducting NW KI resonators to function as nonlinear devices in high magnetic fields for the first time. By characterizing the Kerr nonlinear coefficients of the resonators, we show that their value generally increases for in-plane magnetic fields up to 3 T. The Kerr coefficients of the 1 $\mu$m wide NW resonators increase by more than 100 times more than anticipated. We speculate that this could be due to inhomogeneities in the 10 nm thick NbTiN film used to build the devices, which are revealed through the application of magnetic fields on the NW; a topic that deserves further study. By stimulating a 4WM process, we operate the KI resonators as parametric amplifiers in high magnetic field, and measure their phase-preserving and phase-sensitive gains. {\color{black}Phase-preserving gains above 12 dB are obtained for KI resonators of widths 0.1 and 1 $\mu$m at 0 T, limited by our device design. At magnetic fields around 2 T, the gain of the 1 $\mu$m KI PA decreases by about 5 dB, while the gain of the 0.1 $\mu$m maintains its value. Both the 1 $\mu$m and the 0.1 $\mu$m KI PAs exhibit robust performance in terms of dynamic range and noise figure around 2 T.} Phase-sensitive gain of 16 dB is obtained at 2 T for the 0.1 $\mu$m KI PA. We also observe signal deamplification at 2 T, which implies the potential of these devices for noise squeezing. These results show that KI-based nonlinear devices, including parametric amplifiers, can function efficiently at high magnetic field. This would enable systems requiring high magnetic field for operation, such as spin and majorana based quantum systems, to benefit from the wide range of applications that superconducting nonlinear devices can offer. Parametric amplification in particular could significantly enhance the qubit measurement fidelity in these quantum systems. \begin{acknowledgments} This work was undertaken with support from the Stewart Blusson Quantum Matter Institute (SBQMI), the National Science and Research Council of Canada through the Discovery Grant scheme, the Canadian Foundation for Innovation through the John Edwards Leaders Foundation scheme, and the Canada First Research Excellence Fund, Quantum Materials and Future Technologies Program. MK acknowledges financial support from the SBQMI QuEST fellowship program. MK and JS acknowledge Shabir Barzanjeh and Tim Duty for helpful suggestions and feedback on the manuscript. The authors acknowledge CMC Microsystems for the provision of computer aided design tools that were essential to obtain the results presented here. This research was supported in part through computational resources provided by Advanced Research Computing at the University of British Columbia. \end{acknowledgments}

40ffcc2132c241f49a65626129c70f096aa6edc4

\section*{Introduction} Computer simulations are a vital tool in biochemistry and biophysics. They complement and even replace physical experiments for investigating important biomolecular processes such as protein-ligand binding \cite{shukla2018identification,hempel2021synergistic,novack2022binding}, protein folding\cite{piana2014assessing,li2018finding,sugita1999replica}, and membrane permeation \cite{de2001water,venable2019molecular,gupta2020permeation,sugita2021large}. However, many of these processes involve complex transitions that are so slow with respect to the simulation time scale, that they might not be observed even once within a simulation time span \cite{voelz2010molecular, voelz2012slow}. Enhanced sampling methods modify the Hamiltonian of a system in such a way that one is able to sample the thermodynamics of slow processes more efficiently, for example by adding biasing potentials (e.g. umbrella sampling\cite{torrie1977nonphysical,bartels1999adaptive}, accelerated molecular dynamics\cite{hamelberg2004accelerated,miao2015gaussian}), changing state parameters such as the temperature (e.g. replica exchange\cite{swendsen1986replica,sugita1999replica})\cite{hansmann1997parallel,marinari1992simulated}, or metadynamics \cite{grubmuller1995predicting,barducci2011metadynamics,voter1997hyperdynamics}. In order to un-bias the samples from such simulations, reweighting estimators combine data sampled at different thermodynamic states to compute observables in the state of interest. Important estimators include the weighted histogram analysis method (WHAM) \cite{ferrenberg1989optimized,kumar1992weighted,souaille2001extension} and bin-less WHAM, also known as multistate Bennett acceptance ratio (MBAR) \cite{shirts2008statistically,kong2003theory,tan2012theory,zhang2015stochastic,bartels2000analyzing}. When combining the estimation of different thermodynamic states with Markov state models, one obtains a family of estimators that include the discrete transition-based analysis method (dTRAM) \cite{wu2014statistically}, the dynamic weighted historam analysis method (DHAM) \cite{rosta2015free} and the transition-based analysis method (TRAM)\cite{wu2016multiensemble}. In this paper we concentrate on MBAR and TRAM. MBAR is a statistically optimal method to combine data from multiple thermodynamic states \cite{kong2003theory,bartels2000analyzing,shirts2008statistically}. It is a generalization of the Bennett acceptance ratio method \cite{bennett1976efficient} to an arbitrary number of states, has been shown to minimize the estimator variance \cite{shirts2008statistically} and works under the assumption that all states were sampled in equilibrium. The latter condition can be difficult to achieve for molecular systems with rare events. TRAM \cite{wu2016multiensemble} combines reweighting from different thermodynamic states with Markov State Model (MSM) \cite{schutte1999direct,swope2004describing,singhal2005error,sriraman2005coarse,noe2007hierarchical,chodera2007automatic,prinz2011markov} theory to enable free energy estimation from trajectories that are too short to have produced samples that reflect the equilibrium distribution. The trajectory samples only need to be in a local equilibrium, but not in global equilibrium, allowing for more arbitrary choices of initial conditions and use of short trajectories. Moreover, TRAM extracts kinetic information in the form of an MSM in addition to thermodynamic information in the form of free energy estimates. Both MBAR and TRAM were originally formulated in the form of self-consistent equations that can be solved iteratively \cite{shirts2008statistically,wu2016multiensemble}. Depending on the system and data under consideration, these self-consistent algorithms can take hours or even days to converge \cite{zhang2015stochastic,ding2019fast,tan2016locally}. To accelerate convergence, MBAR has been adapted and optimized in various ways since its inception \cite{zhang2015stochastic,zhang2017stratified,ding2019fast,jia2021free,ferguson2017bayeswham,tan2016locally}, but for TRAM the self-consistent algorithm remains the only available solver as yet. In this work, we develop SATRAM, a stochastic approximation for TRAM, which converges faster to a given accuracy threshold than TRAM, decreasing estimation time up to an order of magnitude. By extension, we also develop SAMBAR, a stochastic approximation for MBAR, which is a special case of SATRAM in the same way that MBAR is a special case of TRAM: they are identical when all data comes from a single Markov state. \section*{Theory} Our system of interest is a reference ensemble governed by a dimensionless potential function $u(x)$ ($x$ being a configuration state), which has units of thermal energy, $\beta=(k_B T)^{-1}$, $T$ being the temperature. In the canonical ensemble, $u(x)$ is equal to $\beta U(x)$, where $U$ is the potential energy function. The equilibrium distribution of the system is given by the Boltzmann distribution \begin{equation} \mu(x) = e^{f-u(x)}, \end{equation} where $f$, the free energy, acts as a constant that normalizes $\mu(x)$, and is equal to the negative log of the partition function. Now we consider the case where we have collected samples from $K$ simulations that were performed at different thermodynamic states. Each thermodynamic state can be related to our unbiased (reference) ensemble as $u^k(x) = u(x) + b^k(x)$, $k \in \{1, ... K\}$ being the thermodynamic state index. The term $b^k(x)$ can result from various thermodynamic changes such as an added umbrella potential, alchemical potentials or a change in temperature, but will from here on simply be referred to as the \textit{bias} energy. The equilibrium distribution for each state can be written as a reweighting of the unbiased distribution as \begin{equation} \label{eq:unbiased_probability} \mu^k(x) = e^{f^k - b^k(x)}\mu(x), \end{equation} where the $f^k$ are the free energies of the different ensembles, and again function as normalizing factors. Our goal is to estimate the $f^k$ so that we can compare the free energies of the biased states, and/or for the purpose of being able to reweight our samples back to the unbiased distribution using Eq. (\ref{eq:unbiased_probability}). \subsection*{MBAR} MBAR is equivalent to binless WHAM \cite{shirts2008statistically,kong2003theory,bartels2000analyzing} and can be derived as a generalization of the Bennett Acceptance Ratio (BAR) \cite{bennett1976efficient} to multiple states. MBAR thus reduces to BAR when only two states are considered. The self-consistent MBAR equations are given by \begin{equation} f^k = -\ln\sum_{x\in X} \frac{\exp[-b^k(x)]}{\sum_{l=1}^K N^l \;\exp[f^l - b^l(x)]}, \label{eq:MBAR_self_consistent} \end{equation} where $N^k$ are the number of samples taken at thermodynamic state $k$, and $X$ is the set of all samples of size $|X| = N = \sum_k N^k$. These equations can be solved by iterating until self-consistency is achieved. Since the equations depend on all samples, the computational cost for a single iteration scales as $O(KN)$, i.e., proportional to the total number of samples $N$ and total number of thermodynamic states $K$. The same scaling behavior also applies to numerical algorithms that are based on maximum-likelihood formulations of MBAR \cite{ding2019fast}. \subsection*{TRAM} The transition-based reweighting analysis method, TRAM \cite{wu2016multiensemble}, combines MSMs and MBAR and generalizes both. TRAM assumes that the trajectories sampled in each thermodynamic states can be approximated by a Markov model, thereby relaxing MBAR's assumption that the data samples from global equilibrium and allowing for shorter trajectories. This is modeled by the assumption that within each Markov state, data is distributed according to a local equilibrium distribution $\mu_i^k(x)$, where $i$ is the Markov state index and $k$ the thermodynamic state index. The local equilibrium distribution can be reweighted to the global equilibrium distribution, $\mu(x)$ by the relation \begin{equation} \mu_i^k(x) = \mu(x) e^{f_i^k -b^k(x)}, \end{equation} where $f_i^k$ is the free energy of configuration state $i$ in thermodynamic state $k$, and $b^k(x)$ is the bias energy of sample $x$ evaluated at thermodynamic state $k$. The TRAM likelihood thus consists of two parts. It combines the MSM likelihood of sampling a specific sequence of discrete states ($L_{\mathrm{MSM}}$), with the local equilibrium likelihood of sampling the continuous configurations within those states ($L_{\mathrm{LEQ}})$. The total likelihood is the product over all thermodynamic states, and given by \cite{wu2016multiensemble} \begin{equation} L_{\mathrm{TRAM}} = \prod_{k=1}^K \underbrace{\left( \prod_{i,j}(p_{ij}^k)^{c_{ij}^k} \right)}_{L^k_{\mathrm{MSM}}} \underbrace{\left( \prod_{i=1}^m \prod_{x \in X_i^k} \mu(x) e^{f_i^k -b^k(x)} \right)}_{L^k_{\mathrm{LEQ}}}, \end{equation} where $K$ is the number of thermodynamic states, and $m$ the number of configuration states. The MSM term in the likelihood is the product of the $p_{ij}^k$, the probability of transitioning from configuration state $i$ to configuration state $j$ in thermodynamic state $k$, to the power of $c_{ij}^k$, the observed number of transitions, for all pairs of configuration states $i$ and $j$. The local-equilibrium term is formed by reweighting the aforementioned local equilibrium distribution $\mu_i^k(x)$, where $X_i^k$ is the set of samples sampled from thermodynamic state $k$ that fall into Markov state $i$. The TRAM solution is obtained by maximizing this likelihood with appropriate normalization constraints and with the detailed balance constraints $e^{f_i^k} p^k_{ij}=e^{f_j^k} p^k_{ji}$. \citet{wu2016multiensemble} show that the maximum-likelihood estimate for $f_i^k$ and $v_i^k$ satisfies the following equations: \begin{align} \sum_j \frac{c_{ij}^k + c_{ji}^k}{\mathrm{exp}[f_j^k - f_i^k] v_j^k + v_i^k} = 1, \quad \text{for all } \;i,\; k, \label{eq:TRAM_1}\\ \sum_{x \in X_i} \frac{\mathrm{exp}[f_i^k-b^k(x)]}{\sum_l R_i^l \mathrm{exp}[f_i^l - b^l(x)]} = 1, \quad \text{for all }\; i, \;k. \label{eq:TRAM_2} \end{align} The $v_i^k$ are a matrix of Lagrange multipliers that arise by reducing the TRAM problem from a constrained problem to an unconstrained problem. The $R_i^k$ are effective state counts given by \begin{equation} R_i^k = \sum_j \frac{(c_{ij}^k + c_{ji}^k)v_j^k}{v_j^k+\mathrm{exp}[f_i^k -f_j^k]v_i^k} + N_i^k - \sum_j c_{ji}^k, \label{eq:reduced_sample_counts} \end{equation} The computational bottleneck in these equations are the equations \cref{eq:TRAM_2} which, like the MBAR equations, depend on all samples and thereby also have a computational complexity of order $O(KN)$. TRAM is equivalent to MBAR when all samples fall into a single Markov state. \subsection*{Related work} Before introducing the stochastic methods SAMBAR and SATRAM, we note that there are various existing implementations and optimizations available for MBAR. Perhaps the most well-known implementation is pymbar \cite{shirts2008statistically}, which combines self-consistent iteration of the MBAR equations with treating MBAR as an optimization problem for maximum precision. The FastMBAR \cite{ding2019fast} package implements MBAR as an optimization problem only and runs faster owing to its GPU-accelerated Quasi-Newton solver but does not boast the extensive functionality of pymbar, such as the computation of state overlap and observables. Both make use of scipy's \cite{2020SciPy-NMeth} optimization algorithms. An existing stochastic solver of the MBAR equations is RE-SWHAM \cite{zhang2015stochastic}, which uses a post-hoc replica-exchange algorithm that re-samples the data obtained from multiple simulations. Stratified UWHAM and its stochastic counterpart stratified RE-SWHAM build on this by being able to handle systems out of equilibrium \cite{zhang2017stratified}. This is done by expanding the set of thermodynamic states by splitting data that was sampled out of equilibrium into multiple thermodynamic states that are (artificially) in equilibrium, after which MBAR can be applied. Local WHAM \cite{tan2017optimally} and its stochastic counterpart resample data using generalized serial tempering, and approximate the free energies by only considering neighbouring states, thereby reducing computational complexity. Similarly, divide-and-conquer MBAR (DC-MBAR) \cite{jia2021free} only calculates the free energy differences between states which have sufficient overlap, which reduces the computational complexity from $O(KN)$ to $O(N).$ For TRAM, the original implementation is available as part of the PyEMMA software package \cite{scherer_pyemma_2015}. Additionally, the Deeptime \cite{hoffmann2021deeptime} functionality has recently been extended with a parallelized TRAM implementation. The authors are not aware of any stochastic implementation, or any other optimization attempt with respect to TRAM. \subsection*{Stochastic approximation} Both MBAR and TRAM are formulated as nonlinear equation systems, Eqs. \eqref{eq:MBAR_self_consistent} and \eqref{eq:TRAM_1}-\eqref{eq:TRAM_2}, respectively, which we here solve by stochastic approximation. To introduce this formalism, let $M(x)$ be a function and $a$ a constant such that the equation \begin{equation} M(x) = a \end{equation} has a unique solution at $x = \hat{x}$. \citet{robbins1951stochastic} showed that if we can observe a random variable, $N(x)$, for which $\mathbb{E}[N(x)] = M(x)$, we can approximate $\hat{x}$ by iteratively solving \begin{equation} \label{eq:SA} x_{n+1} = x_n - \eta_n(N(x_n) - a), \end{equation} where $n$ is the iteration number, and $\eta$ the learning rate which controls the size of the parameter update. It can be shown that, under mild assumptions, $x$ will almost surely converge to the optimal value $\hat{x}.$ In machine learning practice, the random variable $N(x)$ is usually the gradient of an objective function evaluated over a small randomly chosen subset of the entire dataset, called a \textit{batch}. Evaluating the objective function with respect to the batch is computationally cheaper than evaluating it over the entire dataset, so that stochastic iterations are faster than their deterministic counterparts. Stochastic approximation was previously used by \citet{tan2017optimally} to derive the locally weighted histogram analysis method, a stochastic approximation to WHAM. \subsection*{SAMBAR} We apply the stochastic approximation from Eq. \eqref{eq:SA} to MBAR, see supplementary information for the full derivation. The resulting stochastic method iteratively draws a random batch of samples of size $|B|$ and performs the update \begin{equation} f^k = f^k - \eta \left( \frac{1}{|B|} \sum_{x\in B} \frac{N \; \exp[f^k -b^k(x)]}{\sum_{l=1}^S N^l \;\exp[f^l - b^l(x)]}\right), \end{equation} where $\eta$ is the learning rate. The batchwise algorithm reduces the computational complexity for one parameter update to $O(K)$. The expression between the brackets in fact equals the gradient of the MBAR likelihood of observing the batch of samples, i.e. equivalent to performing batch-wise gradient descent. When $|B|=N$, this is equivalent to solving MBAR as a convex optimization problem as shown in \citet{shirts2008statistically,ding2019fast}. \subsection*{SATRAM} Analogously, we also derive stochastic approximation for the TRAM equations, see supplementary information. SATRAM updates the free energies $f_i^k$ and Lagrange multipliers $v_i^k$ as \begin{align} f_i^k &:= f_i^k - \eta \left(\frac{1}{|B|} \sum_{x \in B} 1_{x\in S_i} \frac{\mathrm{exp}[f_i^k-b^k(x)]}{\sum_l R_i^l\; \mathrm{exp}[f_i^l - b^l(x)]} \right), \label{eq:SATRAM_1}\\ v_i^k &:= (1-\eta)\cdot v_i^k + \eta \cdot \frac{1}{N} \sum_j \frac{(c_{ij}^k + c_{ji}^k)v_i^k}{\mathrm{exp}[f_j^k - f_i^k] v_j^k + v_i^k} \label{eq:SATRAM_2}, \end{align} where the $R_i^k$ are computed as \begin{equation} R_i^k = \frac{1}{N} \left(\sum_j \frac{(c_{ij}^k + c_{ji}^k)v_j^k}{v_j^k+\mathrm{exp}[f_i^k -f_j^k]v_i^k} + N_i^k - \sum_j c_{ji}^k \right). \end{equation} The computational complexity of the update of the $f_i^k$ has been reduced to $O(K)$. To avoid a large drift in the energies, a normalization step is executed after each free energy update, which consists of shifting all $f_i^k$ by a constant so that the minimum of the free energies equals zero. When all samples are binned into a single Markov state all indices $i$ and $j$ refer to the single state and most factors in $R_i^k = R^k$ cancel out so that $R_i^k$ reduces to $\frac{N^k}{N}$. The SATRAM equations thus reduce to SAMBAR in the same way that TRAM reduces to MBAR when considering only one Markov state. \subsection*{Algorithmic details} \paragraph{Initialization} The runtime until convergence of (SA)TRAM can be improved by starting estimation from the mean of the bias energies, i.e. $f_i^k = \langle b^k(x) \rangle, \; \forall k, i$ (see also the implementation notes in the supplementary information). \paragraph{Batch size scheduling} The hyperparameters in the free energy update in Eq. \eqref{eq:SATRAM_1} are the learning rate $\eta$ and the batch size $b=|B|$. Setting the learning rate to $\eta=\sqrt{b/N}$ and increasing the batch size over time was found to be more robust than varying the learning rate (see supplementary information). The initial batch size $b_0$ is doubled every $p$ epochs. After each doubling of the batch size, the learning rate is again set to $\eta=\sqrt{b/N}$. When the batch size reaches the total dataset size, SATRAM reverts to a deterministic implementation. The learning rate may be decreased over time to achieve convergence to arbitrary accuracy, though in all our experiments decreasing the learning rate was not necessary to converge to well below chemical accuracy. \paragraph{Finalization} For all configuration states $i$ that were \textit{not} sampled in thermodynamic state $k$, $R_i^k=0$. TRAM estimates a finite free energy for these empty states based on the estimate in states where $R_i^l > 0$, $l \neq k$, but SATRAM does not, since the division by $R_i^k$ in \cref{eq:SATRAM_1} would cause numerical errors for empty states. Although this does not affect computed observables, it does affect the estimated $f^k$, where $f^k =-\mathrm{ln}\sum_i \mathrm{exp}[-f_i^k]$. To estimate the free energy of these empty states, \begin{equation} f_i^k = -\mathrm{ln} \sum_{x \in X_i} \frac{\mathrm{exp}[-b^k(x)]}{\sum_l R_i^l \mathrm{exp}[f_i^l - b^l(x)]} \label{eq:SATRAM_finalize} \end{equation} is computed once as a finalization step after SATRAM has converged. This equation is equal to the second step of the TRAM algorithm, see Eq. (19) in \citet{wu2016multiensemble}. The full SATRAM algorithm is summarized in \cref{alg:SATRAM}. \begin{algorithm} \caption{The SATRAM Algorithm}\label{alg:SATRAM} \begin{algorithmic} \STATE \textbf{parameters:} initial batch size $b$, doubling interval $p$, initialized $f_i^k$ and $v_i^k$ \WHILE{($n <$ max iterations) and (convergence criterion not reached)} \STATE $B \gets$ random batch of samples, $|B|=b$ \STATE update $f_i^k$ using \cref{eq:SATRAM_1} with $B$ and $\eta = \sqrt{b/N}$ \STATE update $v_i^k$ using \cref{eq:SATRAM_2} \IF{$n$ mod $p == 0$} \STATE $b \gets 2\cdot b$ \ENDIF \STATE $n \gets n+1$ \ENDWHILE \STATE finalize $f_i^k$ using \cref{eq:SATRAM_finalize} \end{algorithmic} \end{algorithm} \section*{Results and discussion} We compare the performance of SATRAM and TRAM by applying them to three different sets of simulation data. The first two dataset both contain alanine dipeptide simulation data, generated by simulations at multiple temperatures and by umbrella sampling simulations respectively. The third dataset was generated by performing umbrella sampling of particles aggregating on a membrane. Since SAMBAR is a special case of SATRAM, its performance is briefly discussed for the parallel tempering alanine dipeptide data, in which a comparison to pymbar, a state-of-the-art MBAR solver, is included \cite{shirts2008statistically}. \subsection*{Using SATRAM to reweight from multiple temperatures} Alanine dipeptide is simulated in explicit solvent at 21 temperatures ranging from 300K to 500K, spaced 10K apart, see supplementary material for simulation details. To apply (SA)TRAM to the parallel tempering dataset, the trajectories are discretized into 40 Markov states using the KMeans++ algorithm \cite{arthur2006k} implemented in the Deeptime software package \cite{hoffmann2021deeptime}. To analyze convergence, we compare the respective outputs of (SA)TRAM to a ground truth which we refer to as $\hat{f}_i^k$. This is a converged TRAM estimate computed using Deeptime \cite{hoffmann2021deeptime} to a tolerance of $10^{-10}$. \begin{figure} \centering \includegraphics[width=0.85\textwidth]{images/PT/TRAM/combined_results.png} \caption{Results for the alanine dipeptide dataset generated by parallel tempering simulations. Data is clustered into 40 Markov states (using KMeans++ clustering). The SATRAM batch size is doubled every $p=10$ epochs. \textbf{a)} the free energy profile over the torsion angles estimated by TRAM (left) and SATRAM (center), and their difference (right), at 300 K, discretized to a grid of $50\times 50$ bins. Both estimates were run until convergence of the MAE to within 0.1 kcal/mol with respect to $\hat{f}^k$ (1330 epochs for TRAM, 489 for SATRAM). \textbf{b)} the estimated $f_i^k$ at $k=0$, i.e. at 300 K. In this state the estimated $f^k_i$ of SATRAM and TRAM show the largest discrepancy. \textbf{c)} Convergence of three metrics for TRAM, SATRAM with initial batch size $|B|_0=128$, and SATRAM with $|B|=N$. \textbf{i)} convergence of the JS-divergence of the probability distribution over the Markov states, $p_i$, w.r.t. $\hat{p}_i$. \textbf{ii)} convergence of the MAE of the $f^k$ w.r.t. $\hat{f}^k$. \textbf{iii)} convergence of the log-likelihood.} \label{fig:PT_TRAM} \end{figure} The parameters, $f_i^k$ and $v_i^k$, converge towards the ground truth for both SATRAM and TRAM (except for the $i,k$ that correspond to empty states in the case of SATRAM, which are computed in the finalization step). \cref{fig:PT_TRAM} (panels a and b) shows the final outputs of (SA)TRAM, after running both methods to within a maximum absolute error (MAE) in the $f^k$ of 0.1 kcal/mol with respect to $\hat{f}^k$, which is well below chemical accuracy. \cref{fig:PT_TRAM}.a show the free energy surface computed at $T=300K$ by TRAM and SATRAM, and the difference between their respective outputs. \cref{fig:PT_TRAM}.b shows the free energies per state, $f_i^k$, for $k=0$, i.e. at 300K. At this temperature, the maximum error between the $f_i^k$ estimated by TRAM and SATRAM is the largest compared to the other temperatures, which is reflected in a slight overestimation of the energy of Markov state at index 14 and 21, two respectively small clusters into which less than 0.2\% of all samples are binned. Not plotted are the $v_i^k$, which are identical for both estimators up to floating-point precision. The convergence of SATRAM is compared to TRAM, and SATRAM with a batch size $|B|=N$, in \cref{fig:PT_TRAM}.c, which shows the Jensen-Shannon divergence between the probability distributions over the Markov states $p_i=e^{-f_i}$ and $\hat{p}_i = e^{-\hat{f}_i}$ (\cref{fig:PT_TRAM}.c.i), the maximum absolute error between the $f^k$ and $\hat{f}^k$ (\cref{fig:PT_TRAM}.c.ii) and the log-likelihood (\cref{fig:PT_TRAM}.c.iii). For SATRAM the initial batch size was set to $|B|_0=128$, after which the batch size is doubled every $p=10$ epochs, so that $|B|=N$ after 120 epochs, from whereon the implementation is deterministic. From the lines showing SATRAM with $B=|N|$ (in green) we see that the convergence behaviour of the deterministic SATRAM implementation is similar to that of TRAM. SATRAM converges faster initially on all metrics in \cref{fig:PT_TRAM}c. The convergence of the MAE shows that SATRAM does not converge to the exact ground truth, which we attribute to accumulating numerical errors: since the highest energy state is $>6000$ kcal/mol, this is a relative error of the order $10^{-8}$. The convergence of SATRAM depends primarily on the learning rate/batch size schedule, the number of clusters, and the magnitude free energy difference. A smaller number of clusters will improve the performance of SATRAM over TRAM, and we found increasing the batch size to be more effective than decreasing the learning rate (see supplementary information). \subsection*{Using SAMBAR to reweight from multiple temperatures} Since MBAR is a special case of TRAM, it is perhaps unsurprising to see that the convergence behaviour of (SA)MBAR looks similar to that of (SA)TRAM. We compare SAMBAR with an initial batch size $|B|_0=128$ that doubles every $p=10$ epochs to MBAR, and SAMBAR with a batch size $|B|=N$. Additionally, the convergence behaviour of pymbar (using its default adaptive implementation) is compared. \cref{fig:PT_MBAR_convergence} shows that SAMBAR converges faster than MBAR on the same three metrics that were used to analyse the convergence of TRAM: the Jensen-Shannon divergence between the probability distributions over 40 clusters (\cref{fig:PT_MBAR_convergence}a) which correspond to the Markov states used for TRAM, the MAE between $f^k$ and $\hat{f}^k$ (\cref{fig:PT_MBAR_convergence}b), and the log-likelihood (\cref{fig:PT_MBAR_convergence}c). The MBAR ground truth, $\hat{f}^k$, is a converged MBAR estimate computed using pymbar \cite{shirts2008statistically}. \begin{figure} \centering \includegraphics[width=0.9\textwidth]{images/PT/MBAR/convergence.png} \caption{Convergence of three metrics for pymbar, MBAR, SAMBAR with initial batch size $|B|_0=128$, and SAMBAR with $|B|=N$ on the parallel tempering alanine dipeptide dataset. The SATRAM batch size is doubled every $p=10$ epochs. \textbf{a)} convergence of the JS-divergence of the probability distribution over 40 bins, $p_i$, w.r.t. $\hat{p}_i$. \textbf{b)} convergence of the MAE of the $f^k$ w.r.t. $\hat{f}^k$. \textbf{c)} convergence of the log-likelihood. } \label{fig:PT_MBAR_convergence} \end{figure} \cref{fig:PT_MBAR_convergence} shows that initially, SAMBAR also converges faster than pymbar, although we note that pymbar does not allow for initialization using the bias energies, and the free energies are initialized with zeros. For a fairer comparison of the runtime, the time it takes for pymbar to pre-process the data is not included. The convergence of SAMBAR may be improved with use of the ADAM \cite{kingma2014adam} optimizer, which results in a performance boost for SAMBAR, allowing it to converge significantly faster than pymbar. A more extensive comparison of SAMBAR, pymbar and FastMBAR is included in the SI. \subsection*{Using SATRAM to reweight from umbrella sampling} (SA)TRAM can also be applied to biased simulations, which we demonstrate by applying umbrella sampling to alanine dipeptide. 25 umbrella centers were spaced along both coordinates of interest, the torsion angles $\phi$ and $\psi$ in the protein backbone, giving a total $K=625$ thermodynamic states. For TRAM, the samples are discretized into 5 evenly spaced bins along both torsion angles, so that the number of Markov states $m$ equals $25$. An increasing batch size schedule is again employed for SATRAM, starting with an initial batch size $|B|_0 = 128$ and doubling every $p=10$ epochs. The dataset contains 625.000 samples, so that $|B|=N$ after 130 epochs. \cref{fig:US_TRAM} (panels a and b) shows that the estimates of TRAM and SATRAM converge to the same result for both the estimated free energy surface (\cref{fig:US_TRAM}.a), and the $f_i^k$ (plotted at 300K in \cref{fig:US_TRAM}.b). Not plotted are the $v_i^k$, which are identical for both estimators up to floating-point precision. The Jensen-Shannon divergence (\cref{fig:US_TRAM}.c.i), MAE (\cref{fig:US_TRAM}.c.ii) and log-likelihood (\cref{fig:US_TRAM}.c.iii) converge significantly faster for SATRAM than TRAM. \begin{figure} \centering \includegraphics[width=0.85\textwidth]{images/US/TRAM/combined_results.png} \caption{ Results for the the umbrella sampling dataset of alanine dipeptide with 25 Markov states (using box discretization with 5 bins per torsion angle). The SATRAM batch size is doubled every $p=10$ epochs. \textbf{a)} the free energy profile over the torsion angles estimated by TRAM (left) and SATRAM (center), and their difference (right), at 300 K, discretized to a grid of $50\times 50$ bins. Both estimates were run until convergence of the MAE to within 0.1 kcal/mol with respect to $\hat{f}^k$ (1179 epochs for TRAM, 79 for SATRAM). \textbf{b)} the estimated $f_i^k$ at $k=0$, i.e. at 300 K, in which the graphs of SATRAM and TRAM fully overlap. \textbf{c)} Convergence of three metrics for TRAM, SATRAM with initial batch size $|B|_0=128$, and SATRAM with $|B|=N$. \textbf{i)} convergence of the JS-divergence of the probability distribution over the Markov states, $p_i$, w.r.t. $\hat{p}_i$. \textbf{ii)} convergence of the MAE of the $f^k$ w.r.t. $\hat{f}^k$. \textbf{iii)} convergence of the log-likelihood. } \label{fig:US_TRAM} \end{figure} \subsection*{Using SATRAM to estimate free energy of membrane-mediated interactions} \begin{figure} \centering \includegraphics[width=0.78\textwidth]{images/MP/membrane_protein_aggregation_free_energies.png} \caption{Comparing TRAM and SATRAM in free energy analysis of the aggregation of membrane-bound proteins. \textbf{a)} simulation snapshot with peripheral protein particles shown in orange and membrane particles rendered transparent for clarity. The aggregation reaction coordinate $q$ is defined as the mean of projected pairwise distances between proteins particles. \textbf{b)} potential of the mean force along the reaction coordinate $q$. Results are shown for the given surface concentrations of proteins with three different stiffness values (plot titles). Outputs of the TRAM and SATRAM estimations coincide with the ground truth in the range of inspection (respectively, dashed and dot-dashed lines overlap the solid line). \textbf{c)} comparing free energies of thermodynamic states (i.e. different surface concentrations) between TRAM and SATRAM estimations. The range of free energies highlights the uncertainty in the bias potential. Colors match the legend given at the bottom of the figure. d: Convergence of TRAM and SATRAM with initial batch size $|B|_0=128$, with MAE of the $f^k$ w.r.t. $\hat{f}^k$ used as the metric. Each curve represents an instance of the inferred bias potential.} \label{fig:output_TRAM_SATRAM_MP} \end{figure} As a final example, SATRAM is applied to a model of peripheral proteins aggregating on the surface of a lipid bilayer membrane as result of membrane-mediated interactions by \citet{sadeghi2022investigating}. Their work used TRAM to infer the free energy profiles underlying the aggregation/dispersion process. The dataset contains time series at three different surface concentrations (copy number per unit area) $\Gamma_k,\ k=0,1,2,$ of proteins with three different stiffness values $Y_\mathrm{p}$ = 50, 100, 200 MPa. Simulations were performed using a dynamic membrane model \cite{sadeghi2018memmodel,sadeghi2020memhydro} that incorporates flexible peripheral proteins via a consistent force field masking \cite{sadeghi2021thermodynamics}. A reaction coordinate $q$ has been used to quantify aggregation (Fig. \ref{fig:output_TRAM_SATRAM_MP} and supplementary information) with larger $q$ indicating more clustered states. They proposed an effective bias potential of the form $b^k(q) = \xi_k \cdot b(q)$ such that the bias strength $\xi$ correlates with the surface concentration of proteins and $b\left(q\right)$ is a polynomial function specified by protein stiffness. To also estimate uncertainties originating from the assumptions on the bias potential, five independent instances of $\xi_k$ and $b(q)$ functions were derived (details in supplementary information). Fig. \ref{fig:output_TRAM_SATRAM_MP} shows SATRAM and TRAM results on free energy estimation in this system. SATRAM and TRAM consistently predict the potential of the mean force along the $q$, coincident with the high-accuracy TRAM estimate computed using Deeptime \cite{hoffmann2021deeptime} to a tolerance of $10^{-10}$ (Fig. \ref{fig:output_TRAM_SATRAM_MP}.b). Mean values of free energies of thermodynamic states, as well as their distributions (due to the uncertainty in the bias potential) match perfectly between SATRAM and TRAM (Fig. \ref{fig:output_TRAM_SATRAM_MP}.c). Convergence curves reveal that SATRAM consistently outperforms TRAM, and can achieve higher accuracy with less computational time. The case is most obvious for the stiffest protein, with the largest difference between free energies of thermodynamic states. This suggests that the stochasticity of SATRAM helps the most with the convergence where the objective function has a more rugged profile due to larger energy gradients. \section*{Conclusions} We have introduced SAMBAR and SATRAM, stochastic approximations to MBAR and TRAM respectively, where SAMBAR and MBAR are special cases of the more general SATRAM and TRAM. Like TRAM, SATRAM combines data sampled out of equilibrium in multiple thermodynamic states and estimates a multi-ensemble Markov model, but unlike TRAM, does this in a batch-wise manner. By repeatedly doubling the batch size a deterministic implementation is recovered when the batch size reaches the dataset size, allowing for convergence to arbitrary accuracy, but with an initial performance boost. We showed that SATRAM converges to chemical accuracy up to an order of magnitude faster than TRAM. The batch-wise formulation also allows for adding data on the fly, so that the free energies may be computed in parallel with the sampling algorithm. The SATRAM code is available on github at \href{https://github.com/noegroup/SATRAM}{https://github.com/noegroup/SATRAM}. \begin{acknowledgement} We thank Moritz Hoffmann, Phillip S. Hudson, Tim Hempel, and Yaoyi Chen for offering test data and helpful discussions. We gratefully acknowledge funding from European Commission (Grant No. ERC CoG 772230 “ScaleCell”), the BMBF (Berlin Institute for Learning and Data, BIFOLD), the Berlin Mathematics center MATH+ (AA1-6), the Deutsche Forschungsgemeinschaft DFG (SFB1114/A04,C03 and SFB958/A04), the NSF of China (Grant No. 12171367) and the Shanghai Municipal Science and Technology Commission (Grant No. 20JC1413500, 21JC1403700 and 2021SHZDZX0100). \end{acknowledgement} \nocite{eastman2017openmm,bgmol}

9698bb6c97bbea2a8fd2f7ae8ba3789dd1460e69

\section{Introduction} \label{intro} Abstract Meaning Representation (AMR) \cite{ban-AMR} parsing targets to transform a sentence into a directed acyclic graph, which represents the relations among different concepts. The original AMR does not provide concept-to-word alignment information, which hinders the trace-back from concept to input word and brings difficulties to AMR parsing. To solve the problem, based on Chinese AMR \cite{li-etal-2016-annotating} , \newcite{Li_Wen_Song_Qu_Xue_2019} further propose to add concept and relation alignment to the structure of Chinese AMR as shown in Figure~\ref{fig:camr}. Currently, while a majority of work is focusing on improving the performance of English AMR Parsing \cite{xu-seqpretrain,bevil-spring,HCL,bai-etal-2022-graph,chen-etal-2022-atp,drozdov-etal-2022-inducing}, those methods or models can not be directly applied to Chinese AMR Parsing since English AMR does not provide alignment information itself. To better reflect the full structure of Chinese AMR, CAMRP-2022 evaluation\footnote{https://github.com/GoThereGit/Chinese-AMR} firstly requires the AMR parser to generate explicit word alignment including concept and relation alignment which calls for novel models and algorithms. \begin{figure}[h] \centering \includegraphics[width=0.7\linewidth]{pic/camrp.pdf} \caption{An example of Chinese AMR. Red word-IDs under concepts denote concept alignment. Red words and word-IDs under relation denote relation alignment. } \label{fig:camr} \end{figure} We propose a two-stage method to conduct Chinese AMR Parsing with alignment generation\footnote{We participate in the closed-track evaluation where we can only use HIT-roberta\cite{cui-etal-2020-revisiting} as the pretrained language model}. In a nutshell, the method includes the Concept-Prediction and Relation-Prediction stages, which can be regarded as the process of graph formation. In the Concept-Prediction stage, we develop a hierarchical sequence tagging framework to deal with the concept generation and the complex multi-type concept alignment problem. In the Relation-Prediction stage, we utilize the biaffine network to predict relations and Relation-alignment simultaneously among predicted concepts. Our model ranks 2nd in the closed-track of the evaluation, achieving 0.7756 and 0.7074 Align-Smatch \cite{alignsmatch} F1 scores on the CAMR 2.0 test set and the blind-test set of CAMRP-2022 individually. \section{Method} Our methods includes the Concept-Prediction stage and Relation-Prediction stage. As illustrated in Figure~\ref{fig:piplines}, during training, both stages have individual input and output. During inference, the output concepts from the Concept-Prediction stage are passed to the Concept-Prediction stage to generate the full AMR graph. \begin{figure}[h] \centering \includegraphics[width=1\linewidth]{pic/pipeline2.pdf} \caption{Pipeline for the two-stage method. In the training phase, the relation prediction stage takes the concepts from the gold AMR graph as input. During the inference phase, the relation prediction stage takes the output of the concept predictor as input.} \label{fig:piplines} \end{figure} \subsection{Concept-Prediction} Different from English AMR where nodes or concepts can have arbitrary variable names, a large portion of concepts of Chinese AMR have standard variable names which denote the alignment to input words. Moreover, there are different alignment patterns which make generating the right alignment a complex problem. In following sections, we'll sequentially introduce the alignment rules we design for CAMRP-2022, the two-stage method and the model. \subsubsection{Multi-Type Concept Alignment Rule} We mainly design 6 different alignment rules for concepts according to different alignment patterns, which are \textbf{\textit{Direct Alignment}}, \textbf{\textit{Normalization Alignment}}, \textbf{\textit{Continuous Multi-word Alignment}}, \textbf{\textit{Discontinuous Multi-word Alignment}}, \textbf{\textit{Split Alignment}} and \textbf{\textit{Null-Aligned Concepts}}. The difference among alignment rules lies in how an abstract concept corresponds to the input words. We list three cases involving different rules as shown in Figure~\ref{fig:concept_align} as examples. \begin{enumerate} \item \textbf{Direct Alignment} is the easiest alignment where a concept directly corresponds to a certain word in the input without the need for any modification. \item \textbf{Normalization Alignment} exists when a concept still corresponds to one word in the input however needs to be ``normalized'' into the final concept. The normalization includes different situations like word sense disambiguation for predicate and Arabic numerals transformation for numerals in other languages. For example, as shown in Figure~\ref{fig:concept_align}, in case (a) the word ``\begin{CJK*}{UTF8}{gbsn}称为\end{CJK*}'' corresponds to the concept ``\begin{CJK*}{UTF8}{gbsn}称为-01\end{CJK*}'' after word sense disambiguation. In case (c), Chinese numeral ``\begin{CJK*}{UTF8}{gbsn}一\end{CJK*}'' would be mapped to concept ``1'' since all numeral concepts in CAMR are Arabic. \item \textbf{Continuous Multi-word Alignment} exists when multiple continued words in the input sentence are concatenated into the final concept, which usually happens for named entities. \item \textbf{Discontinuous Multi-word Alignment} means multiple discontinued words in the input sentence are joined into the final concept or preposition patterns like ``\begin{CJK*}{UTF8}{gbsn}在...上\end{CJK*}'' . \item \textbf{Split Alignment} denotes one word that could correspond to multiple concepts, which usually suggests the word corresponds to a sub-graph in the final AMR graph. \item \textbf{Null-Aligned Concepts} do not have alignment and could have arbitrary variable names. These concepts usually abstract away from syntactic features and do not directly correspond to certain word in the input sentence, making them harder for the model to predict. In fact, according to our experiment, our system could reach a 0.91 f1 score for aligned concepts' prediction but only a 0.70 f1 score for Null-Aligned concepts' prediction. \end{enumerate} \begin{figure}[t] \centering \includegraphics[width=1\linewidth]{pic/aligment_cases0928.pdf} \caption{Example of different Concept Alignment cases. Gold concepts denote all concepts in the gold AMR graph of the input sentence. The concepts can be divided into different categories according to the alignment rules. We use words in color to represent the unique alignments in each example.} \label{fig:concept_align} \end{figure} \begin{figure}[b] \centering \includegraphics[width=1\linewidth]{pic/pies0928.pdf} \caption{Statistics about the alignment between concepts and input words in CAMR 2.0.} \label{fig:concept_align_pie} \end{figure} As shown in Figure \ref{fig:concept_align_pie}, we further collect more statistics about the alignment between concepts and input words with the training set of the CAMR 2.0 dataset. From the perspective of input sentences, about 75\% of words in the input sentences are associated with certain concepts under one alignment rule. From the perspective of concepts, there are 83\% of concepts with alignment. For all concepts with alignment, a majority of them belongs to Direct(56\%) and Normalization(33\%) Alignments. \subsubsection{Sequence Tagging Framework} \begin{figure}[t] \centering \includegraphics[width=1\linewidth]{pic/TWO_MODELS.pdf} \caption{The Two-Stage Parsing Model. We use the same Concept Tagging model structure to conduct three hierarchical sequence tagging tasks, each with a different Tag Classifier. Relation Classification takes concepts as input and the Biaffine Classifier outputs the relation between every two concepts. In both models, words or concepts are first splitted into characters before feeding into the pretrained language model and we use the hidden state of the first character to represent the word in the last classifier layers. } \label{fig:models} \end{figure} In spite of the complex word alignment rules, we can see that a large portion of concepts are directly or indirectly aligned with a single word of input and one word can only correspond to one concept at most, which inspires us to adopt sequence tagging method. It can deal with concept prediction and Direct Alignment prediction simultaneously. Considering different alignment rules, we develop three sequence tagging rules to cover all possible situations. \paragraph{Model Structure} As depicted in Figure~\ref{fig:models}, we add a linear layer on the top of the Chinese RoBERTa model as a Tag Classifier. Adapting to character-based Chinese pretrained language model, Tag classification is conducted \textbf{on the first character's hidden state} for a word with multiple characters. During training, we use Cross-Entropy as the loss function and use the average loss of $N$ all tags in a sentence as the final loss, as described in Equation~\ref{eq:tagging}, \begin{equation} \begin{aligned} \text{TagCLS}(\mathbf{a}) & = [\mathbf{a};\mathbf{1}]\mathbf{W}, (\mathbf{a} \in \mathbb{R}^{1 \times d}, \mathbf{W} \in \mathbb{R}^{(d+1)\times c} ) \\ \text{Loss}&=\frac{1}{N}\sum_{i=1}^N CE(\text{TagCLS}(\mathbf{h_i}) ,\hat{\mathbf{h_i}}) \end{aligned} \label{eq:tagging} \end{equation} where $d$ denotes the hidden size, $c$ denotes the number of different tags, $N$ denotes the number of input words, $\mathbf{h_i}$ denotes the $i^{th}$ word's output hidden state from the encoder and $\hat{\mathbf{h_i}}$ denotes the one-hot vector of its gold tag. \paragraph{Surface Tagging} We design an 8-classes BIO tagging rule as the first step to process the input sentence. The eight classes are O, B-Single, B-Continuous-Multiword, I-Continuous-Multiword, B-Discontinuous-Multiword, I-Discontinuous-Multiword, B-Split, and B-Virtual. This tagging rule can cover 4 out of 6 alignment rules, which are Direct Alignment, Continuous Multi-word Alignment, Discontinuous Multi-word Alignment, and Split Alignment. Note that the B-Single tag is for both Direct Alignment and Normalization Alignment because they both correspond to one input word. As for B-Split, we use manually curated rules to split the word with Split Alignment. Note that B-Virtual is also added to label the virtual word for the later relation classification task. The F1-score of the Surface Tagging step can reach 91\% on the development set in our experiment. \paragraph{Normalization Alignment Tagging} Previous Surface Tagging can not recognize words that need normalization like word sense disambiguation so we introduce a 2-class Tagging rule to identify whether a word from the input sentence needs normalization before becoming a concept in the AMR graph. The labels can be collected directly from the gold AMR graph. If one concept is aligned to one identical word from the input sentence, then the word's label is negative. If the concept is aligned to a word different from itself, then the word's label is positive. The F1-score of Normalization Alignment Tagging can reach 0.95\% on the development set. After recognizing words needing normalization, we run a statistical normalization method as described in Appendix A. This step can cover and predict the Normalization Alignment. \paragraph{Null-Aligned Concept Tagging} For concepts that do not have alignment with input words, we define trigger words for those concepts and also use sequence tagging method. To be more specific, we first collect the dictionary of all Null-Aligned concepts in the training set and there are 184 different Null-Aligned concepts in total. The label of the input word is the class of Null-Aligned concept it triggers, or ``None'' if it triggers nothing. For Null-Aligned concept, we define the concepts that it has a direct relation to as its trigger concepts and the aligned word of the trigger concept as the trigger word. For example, as shown in Figure~\ref{fig:camr}, the Null-Aligned concept ``Event'' has direct relation to concept ``\begin{CJK*}{UTF8}{gbsn}钱塘江大潮\end{CJK*}''. According to the alignment information of the concept, the trigger words of the concept ``Event'' are x1 and x2. Since a Null-Aligned concept could have multiple concepts it has a direct relation to, we tried using the first or the last of the concepts. The experimental result shows that using the last concept is more effective with a 0.03 F1 improvement. There are nearly 5\% cases where the trigger concepts are all Null-Aligned concepts. Under such circumstances, we keep tracing back from the trigger concept until we reach the first concept with alignment and we regard this concept as the trigger concept. \subsection{Relation-Prediction} As shown in Figure~\ref{fig:models}, we design a RoBERTa-BiLSTM-Biaffine network to conduct relation prediction given the predicted concepts. All concepts are first split into characters before feeding into the RoBERTa model to extract hidden representations. After the RoBERTa model, all hidden states are fed into a one-layer BiLSTM network to better encode sequential information to the hidden states. At last, the \textbf{first hidden states of every two concepts} are fed into the biaffine network to get the relation between the two concepts. During training, we use Cross-Entropy as the loss function and use the average loss of $N\times N$ relations as the final loss, as described in Equation~\ref{eq:biaffine}, \begin{equation} \begin{aligned} \text{Biaffine}(\mathbf{a},\mathbf{b}) & = [\mathbf{a};\mathbf{1}]\mathbf{W}[\mathbf{b};\mathbf{1}]^T, (\mathbf{a} \in \mathbb{R}^{1 \times d}, \mathbf{b} \in \mathbb{R}^{1 \times d}, \mathbf{W} \in \mathbb{R}^{(d+1)\times c \times(d+1)}) \\ \text{Loss}&=\frac{1}{N^2}\sum_{i=1}^N \sum_{j=1}^N CE(\text{Biaffine}(\mathbf{h_i},\mathbf{h_j}), \hat{r}(\mathbf{h_i},\mathbf{h_j}))\\ \text{Relation}_{a,b} & = \arg \max \text{Biaffine}(\mathbf{a},\mathbf{b}) \end{aligned} \label{eq:biaffine} \end{equation} where $d$ denotes the hidden size, $c$ denotes the number of relations, $N$ denotes the number of input concepts, $\hat{r}$ denotes the one-hot vector for the gold relation and $\mathbf{h_i}$ denotes the $i^{th}$ output hidden state from the BiLSTM network. \paragraph{Relation-Alignment Prediction} On top of relation between concepts, another important feature of Chinese AMR is the relation alignment, which takes Chinese functional words' semantics into consideration in AMR graph. For example, as shown in Figure~\ref{fig:camr}, functional word ``\begin{CJK*}{UTF8}{gbsn}被\end{CJK*}'' is aligned to the ``arg1'' relation between concepts ``\begin{CJK*}{UTF8}{gbsn}称为-01\end{CJK*}'' and ``Event''. In fact, in the input Chinese sentence, the word ``\begin{CJK*}{UTF8}{gbsn}被\end{CJK*}'' is the marker of relation ``arg1'' between concept ``\begin{CJK*}{UTF8}{gbsn}钱塘江大潮\end{CJK*}'' and ``\begin{CJK*}{UTF8}{gbsn}称为-01\end{CJK*}''. We use the same model as relation prediction to align functional words with relations. To be more specific, concepts and functional words are both fed into the RoBERTa-BiLSTM-Biaffine network. As depicted in Figure~\ref{fig:relaton_example}, for any relation triples(concept1, concept2, relation), if the relation is aligned with functional word $w$, we create another triple(concept1, $w$, relation) for the model to predict. In this way, we can predict the relation and relation alignment simultaneously. After predicting all relations, if one concept is linked with one concept and one functional word with the same relation, the functional word will be aligned to the relation. \begin{figure}[t] \centering \includegraphics[width=0.8\linewidth]{pic/relation_table.pdf} \caption{An example of Relation Classification label matrix. The inputs are from the gold concepts of sentence ``\begin{CJK*}{UTF8}{gbsn}钱塘江大潮被称为天下奇观\end{CJK*}''. Each column denotes the start node of a relation while each row denotes end node of a relation. ``O'' denotes there is no relation between the two nodes. Relation in red denotes the relation alignment for functional words.} \label{fig:relaton_example} \end{figure} \subsection{Teacher Forcing in Training} Since our method has two stages, during inference the Relation-Prediction model takes the output of Concept-Prediction as input. To stabilize and prevent error propagation during training, we adopt the Teacher Forcing method, where we use the gold concepts and relations as the input of the relation prediction model. However, error propagation still exists in the inference phase. We will discuss the error propagation situation of our system in section~\ref{sec:error_p}. \section{Experiment} \begin{table}[t] \centering \resizebox{0.4\textwidth}{!}{ \begin{tabular}{lcc} \toprule Dataset Split & Sentences & Tokens \\ \midrule Train & 16576 & 386234\\ Development & 1789 & 41822\\ Test & 1713 & 39228 \\ Blind Test & 1999 & 36940 \\ \bottomrule \end{tabular}} \caption{The dataset description of CAMRP-2022. Note that the train, development and test splits are directly from CAMR 2.0 dataset while the blind test split is from this evaluation.} \label{tab:dataset} \end{table} \subsection{Dataset} \label{sec:dataset} The CAMRP-2022 evaluation uses the training, development, and test splits of the CAMR 2.0 dataset as its dataset and also involves an out-of-domain blind test set to measure the generalization performance of parsers. The statistics of the dataset are shown in Table~\ref{tab:dataset}. For concepts, there are 31941 different concepts in the training set. Among all concepts, there are 8443 different predicates that need to conduct word sense disambiguation and 184 Null-Aligned concepts. As for relations, there are 142 different relations and 841 relation alignment words. The top 5 most frequent relation alignment words are ``\begin{CJK*}{UTF8}{gbsn}的\end{CJK*}'',``\begin{CJK*}{UTF8}{gbsn}是\end{CJK*}'',``\begin{CJK*}{UTF8}{gbsn}和\end{CJK*}'',``\begin{CJK*}{UTF8}{gbsn}在\end{CJK*}'' and ``\begin{CJK*}{UTF8}{gbsn}对\end{CJK*}''. \subsection{Model} Both the Concept Tagging and Relation Classification model adopt the HIT-roberta-large\cite{cui-etal-2020-revisiting} pretrained model downloaded from HuggingFace model hub\footnote{https://huggingface.co/hfl/chinese-roberta-wwm-ext-large}. For Concept Tagging models, the output size of the tag classifier is 8, 2, 185 for Surface Tagging, Normalization Alignment Tagging, and Null-Aligned Concept Tagging individually and the dropout rate of the classifiers is 0.1 in all experiments. For the Relation Classification model, there is one BiLSTM layer and the hidden size is 4096. the dimension of Biaffine matrix is $4097\times142\times4097$. \subsection{Training Details} We use Adam as the optimizer and conduct hyper-parameter searches on batch-size (from 10 to 100) and learning rate (from 1e-5 to 1e-4 ) in all models. The optimal hyper-parameters for each model are listed in Appendix B. We train all models for 100 epochs with 1\% warmup steps and select the one with the best result on the development set as the final model. \subsection{Results} \label{sec:result} \begin{table}[t] \centering \resizebox{0.7\textwidth}{!}{ \begin{tabular}{lccc} \toprule Task(Dev) & Precision & Recall & F1 \\ \midrule Surface Tagging & 0.918 & 0.944 & 0.931 \\ Normalization Aligment Tagging & 0.878 & 0.878 & 0.878 \\ Null-Aligned Concept Tagging & 0.708 & 0.679 & 0.693 \\ Relation Classification (With Gold Concepts) & 0.751 & 0.737 & 0.744 \\ \midrule AlignSmatch & 0.778 & 0.766 & 0.768\\ \quad - Only Instance &0.830 & 0.833 &0.832\\ \quad - Only Attribute &0.928&0.954&0.941\\ \quad - Only Relation &0.614&0.556&0.583\\ \bottomrule \\ \toprule Task(Test) & Precision & Recall & F1 \\ \midrule AlignSmatch & 0.786 &0.765 &0.776 \\ \quad - Only Instance & 0.834 & 0.840 & 0.837\\ \quad - Only Attribute& 0.932 & 0.959 & 0.945\\ \quad - Only Relation & 0.628 & 0.570 & 0.598\\ \bottomrule \\ \toprule Task(Blind Test) & Precision & Recall & F1 \\ \midrule AlignSmatch& 0.715 & 0.696& 0.705\\ \quad - Only Instance & 0.768 & 0.775 & 0.772\\ \quad - Only Attribute & 0.866 & 0.901 & 0.883\\ \quad - Only Relation & 0.549 & 0.492 & 0.519\\ \bottomrule \end{tabular}} \caption{The fine-grained results of our model in CAMRP-2022. We report the overall and fine-grained AlignSmatch scores of our model on the development, test and blind test sets. We also report the results of each sub-task in the two-stage method on the development set.} \label{tab:result} \end{table} As shown in Table~\ref{tab:result}, we list the results of our trained 2-stage AMR Parser on the development, test, and blind-test set of CAMRP-2022. For the development set, we list the concrete results of all different sub-tasks in two stages along with the overall and fine-grained AlignSmatch scores. \paragraph{Sub-task Results} For the three sub-tasks of the Concept-Prediction stage, we can tell from Table~\ref{tab:result} that our model performs better in the Surface Tagging task with a 0.931 F1 score and Normalization Alignment Tagging task (0.878 F1) than in Null-Aligned Concept Tagging task (0.693 F1). It suggests that the model can better recognize concepts with alignment and there is a big performance drop when predicting concepts without alignment under the same sequence tagging framework. For the Relation Classification task, our model can reach 0.744 F1 when given gold concepts while only 0.583 F1 in inference when the concepts are generated by the Concept-Prediction stage instead of gold concepts. It reveals a train-inference discrepancy existing in the current method since the model might generate wrong concepts during the Concept Prediction stage in inference which would bias the Relation Prediction stage. \paragraph{AlignSmatch Results} As for the overall AlignSmatch scores, we can tell from the result of Development, Test and Blind-Test evaluations that there exists a domain shift. When looking at the fine-grained scores, the trend is consistent among three evaluation dataset that the performance of attribute or alignment prediction is better than instance prediction and far better than relation prediction. The trend indicates that the model generally outperforms in the first stage than in the second stage. Moreover, for relation prediction, we can see that the recall is about 5 points lower than the precision in all experiments and the gap is much bigger than instance or attribute prediction. The reason is that in the relation classification model there exists a performance gap in relation prediction and relation alignment prediction. Compared to relation prediction, a lot more relation alignments are not predicted while the relation-only score in AlignSmatch takes both relation and relation alignment into account, which makes the recall score lower. In fact, if we preclude relation alignment prediction in the relation-only score, the gap between precision and recall will be reduced to 2 points. It hints to us that we need to pay more attention to the relation alignment prediction to improve the overall performance. \section{Discussion} In this section, we summarize some problems that need to be addressed to improve the performance of the Chinese AMR parser. \subsection{Error Propagation in the Two-Stage Model} \label{sec:error_p} As pointed out in Section~\ref{sec:result}, there exists error propagation in the two-stage model. The direct evidence is that while the relation prediction could reach 0.744 F1 with gold concepts, this score drops to 0.583 when giving it the model predicted concepts. Error propagation also exists in the Concept-Prediction stage since Normalization Alignment needs both the correct result from Surface Tagging and Normalization Alignment Tagging. \subsection{Class Imbalance Problem} As pointed out in section~\ref{sec:dataset}, there exist severe class imbalance problems in both stages of the parsing task. As for the Concept-Prediction stage, the problem reflects the great differences in the distribution of different tags in the three tagging tasks, especially for the Null-Aligned Concept Tagging tasks. For the Relation-Prediction stage, a large portion of labels is ``None Relation'' as shown in Figure~\ref{fig:relaton_example}. We have tried some techniques like using weighted loss that assigns greater weight to the minority classes to handle the class imbalance problem. While this can greatly reduce the time required for the model to converge, it does not improve the final performance when all epochs are finished. \subsection{Improving the Null-Aligned Concept Prediction Performance} In our model, we use a trigger-based method to predict concepts without alignment. This method could cover nearly 95\% cases while the rest 5\% is neglected because they are mostly triggered by another Null-Aligned concept. Though we design methods to overcome the drawback by tracing back to the first aligned concept, the overall result of Null-Aligned Concept Prediction is still the lowest in the Concept-Prediction stage, which could lead to great bias for the next stage. A more natural method to predict those concepts might greatly improve this task. \section{Conclusion} In this paper, we provide a detailed description of the proposed two-stage Chinese AMR Parsing model which is the first to deal with the explicit word alignment problem for CAMRP-2022 evaluation. We also analyze the result and point out the limitation of the current method and some potential roads that might lead to improvement. Though straightforward, the method is far from perfect that it still calls for future exploration to reach a better result in the Chinese AMR Parsing task.

d8ba3f7c8515388fc5673d88cc4b55e95b187eaf

\section{Introduction} \label{sec:intro} The nature and geometry of the X-ray emission region in black hole binaries are still controversial, especially in the low/hard state, where most of the power is emitted in a spectrum quite unlike a standard disc (\citealt{Shakura_1973}). Spectral fitting alone is degenerate, with proposed geometries being a compact source on the spin axis (lamppost), extended emission along the jet direction (jet corona), extended coronal emission on top of an underlying accretion disc (sandwich), and extended coronal emission which replaces the accretion disc (truncated disc/hot inner flow) see e.g. \cite{Poutanen_2018}. The truncated disc/hot inner flow model has the advantage that it gives a framework to explain the evolution of the spectrum and its fast variability properties together (\citealt{Done_2007}), although there are persistent questions over the extent of disc truncation from modelling the reflected emission and its associated iron line (e.g. \citealt{Buisson_2019}; but see \citealt{Zdziarski_2021}). Another way to track the extent of the disc is the quasi-thermal emission arising from the same X-ray irradiation of the disc which gives rise to the iron line and reflected emission (\citealt{DeMarco_2015, Wang_2022}). Photons which are not reflected are reprocessed in the disc, producing a thermal reverberation signal. This gives a soft lag, where variations of soft photons follow those of hard photons with a light travel time delay. Reverberation size-scales do indeed point to a truncated disc, with a truncation radius which decreases as the source spectrum softens (\citealt{DeMarco_2021}). Perhaps, the most compelling evidence for a truncated disc is the new polarisation results for a low/hard state of Cyg~X-1. These rule out the X-ray emission region being aligned with the jet and instead require it to be aligned with the accretion flow (\citealt{Krawczynski_2022}). Truncated disc/hot inner flow models are thus strongly favoured, motivating our work in exploring how we can derive physical properties of the hot flow from X-ray spectral-timing data. The model of propagating fluctuations of mass accretion rate has given the most promising framework for the stochastic variability seen on 0.01--100 seconds (\citealt{Lyubarskii_1997, Kotov_2001}). Slower fluctuations of the local mass accretion rate stirred up at outer radii propagate down to modulate faster fluctuations at inner radii, producing correlated but lagged variability on multi-time-scales in the entire hot flow. The fluctuations are expected to be generated by the magnetorotational instability (MRI; \citealt{Balbus_1991, Balbus_1998}). Different disc responses to the fluctuations at different radii lead to the radial stratification of the variability time-scale (\citealt{Churazov_2001, Ingram_2013, Ingram_2016}). Given spectral inhomogeneity in the hot flow, i.e., softer (harder) photons are predominantly generated at outer (inner) radii, there arises the hard lag associated with the propagation time, where variations of hard photons lag behind those of soft photons. This picture is consistent with major timing properties observed, i.e., the broad-band variability and the hard lag (\citealt{Miyamoto_1988, Nowak_1999}). Indeed, \cite{Arevalo_2006, Ingram_2011, Ingram_2012} developed numerical models combining the propagating fluctuations process with spectrally-inhomogeneous hot flows, confirming general agreement with the variability properties observed by {\it RXTE}. However, accurately reproducing observed timing properties with the propagating fluctuations model has turned out very difficult. Plainly, observed power spectra and lags as a function of Fourier frequency in the low/hard state are too complex for models to fully capture. One of the variability features generally observed yet poorly understood is that power spectra are double-peaked (\citealt{Belloni_2002, Pottschmidt_2003, Axelsson_2005, Grinberg_2014}). The propagating fluctuations models normally produce a single peak in power spectra (\citealt{Arevalo_2006, Mahmoud_2018}). Reconstructing double-peaked power spectra within a single accretion flow appears to require fine-tuning of the flow properties (\citealt{Mahmoud_2018, Mahmoud_2018b}) A key recognition was made by \cite{Wilkinson_2009, Uttley_2011}, who found from {\it XMM-Newton} data that not only the hot flow but the disc generated considerable variability in the low/hard state. \cite{Rapisarda_2016} proposed that the fluctuations of mass accretion rate propagated from the variable disc to the hot flow. Their different variability time-scales due to their different physical properties naturally produce double-peaked power spectra (but see \citealt{Veledina_2016} for another potential mechanism). In this picture, even though the high-energy flux does not contain the disc emission, it does carry the imprint of the disc variability through the propagation. This idea also explains the wide range of time-scales on which the X-ray flux varies even when the truncation radius is only a few tens of gravitational radii (\citealt{Rapisarda_2016}). Another difficulty in accurate variability modelling is that there are other physical processes at work. Reverberation described above decreases the hard lags (\citealt{Mastroserio_2018}) and adds some of the high-frequency variability of the hot flow to the low-energy bands (\citealt{Kawamura_2022}). Compton scattering in the hot flow produces energy-dependent lags according to the number of scatterings seed photons experience (e.g. \citealt{Kylafis_2018, Garcia_2021}). There is no consensus on how important each process is at different time-scales and energy bands, and there can also be a contribution from the jet (though the extent of this is limited by the new polarisation results (\citealt{Krawczynski_2022})). The spectral-timing analysis is a powerful tool to constrain the accretion flow properties as it combines all the information from the energy spectrum and its fluctuations (power spectra) and causal connection (lags/leads e.g. \citealt{Axelsson_2018, Mahmoud_2019, Wang_2021, DeMarco_2021}). In our previous work \cite[hereafter K22]{Kawamura_2022}, we developed a spectral-timing model based on propagating fluctuations from a turbulent disc through a spectrally inhomogeneous (two Comptonisation regions) flow which generates variability at each radius. We also incorporated the reverberation from the variable Comptonisation components illuminating the disc to perform a self-consistent spectral-timing analysis. We applied the model to the recently discovered black hole transient MAXI~J1820+070 (\citealt{Kawamuro_2018, Tucker_2018}), which has been actively studied (e.g. \citealt{Kara_2019, Shidatsu_2019, Homan_2020, Bright_2020, Axelsson_2021, Ma_2021, You_2021, Tetarenko_2021, Wang_2022, Prabhakar_2022}) thanks to its exceptional brightness, low galactic absorption (\citealt{Uttley_2018}), and intensive monitoring by multiple telescopes such as {\it NICER}. K22 modelled the time-averaged energy spectrum for the {\it NICER} (0.5--10~keV) + {\it NuSTAR} (3--73~keV) and used this to develop a model for the variability below 10~keV seen in {\it NICER}. However, {\it NuSTAR} has less capability for fast timing, limiting the ability of K22 to investigate the propagating fluctuations process through the innermost parts of the hot flow, where higher energy photons are emitted. Better constraints on propagation require extending the bandpass for fast timing to higher energies. Here we use contemporaneous data from {\it Insight-HXMT} to test our model at higher energies (Section~\ref{sec:data}). The {\it Insight-HXMT} data show that our previous model fails to describe several key features of the energy-dependent power spectra and phase lags above 10~keV (Section~\ref{sec:old}). We consider several additional processes suggested in the context of other propagating fluctuations models (\citealt{Rapisarda_2017a, Mahmoud_2018b}), but none of these allows us to capture the energy dependence of variability (Section~\ref{sec:rev1}). Instead, we find that allowing the spectra to pivot is key to matching the data. This allows the energy spectra of the Comptonisation components to change their shape as well as normalisation in response to the propagating fluctuations (\citealt{Mastroserio_2018, Mastroserio_2019, Mastroserio_2021}). This is physically expected from Comptonisation models (\citealt{Veledina_2016, Veledina_2018}) and is seen in the data (\citealt{Malzac_2003, Gandhi_2008, Yamada_2013, Bhargava_2022}). We are able to give a good match to the power spectra and phase-lag spectra from 2.6~keV up to 48~keV by including spectral pivoting (Section~\ref{sec:rev2}). Additionally, we successfully perform a joint spectral-timing fit using our new model across the entire 2.6--48~keV bandpass (Section~\ref{sec:spec_timing}). We discuss the physical properties of the accretion flow, comparing them with theoretical hot flow models (Section~\ref{sec:discussion}), and conclude (Section~\ref{sec:conclusions}). All of the technical details of the model formalism are given in the appendices so that the main text stresses the physical aspects of the model. \section{Observation and data reduction} \label{sec:data} \begin{figure} \includegraphics[width=0.99\columnwidth]{figure/observation_data.pdf} \caption{Spectral-timing properties of MAXI~J1820+070 observed by {\it Insight-HXMT}. {\it Top}: Time-averaged energy spectrum. Red, green, and blue markers represent LE, ME, and HE telescopes, respectively. The dip around 22~keV (light green) is associated with fluorescent lines of silver generated within the ME detector. {\it Middle}: power spectra calculated for energy bands of 2.6--4.8~keV (black) and 35--48~keV (magenta). These energy bands are marked with shaded regions in the energy spectrum. The fundamental QPO and its second harmonic are marked with dashed lines. {\it Bottom}: Phase-lag spectrum between these energy bands. The lags are defined as positive if variations in higher energy bands lag behind those in lower energy bands (hard lags). The peak frequency of broad-band variability in the phase-lag spectra is marked with a dotted line. } \label{fig:observation_data} \end{figure} We investigate the bright low/hard state of MAXI J1820+070 observed by {\it Insight-HXMT}. The observation ID is P0114661003, and the observation was conducted from 2018-03-22 10:46:53 to 2018-03-24 02:49:49. The same data are studied in \cite{Wang_2020, Ma_2021, Yang_2022}. The observation time is slightly later than that we studied in K22 (Obs. ID: 1200120106; 2018-03-21), but there are simultaneous {\it NICER} data (Obs. ID: 1200120108; 2018-03-23) corresponding to these {\it Insight-HXMT} data. We checked that the energy spectrum, power spectra, and phase-lag spectra in these simultaneous {\it NICER} data are almost identical to K22. The observation data were processed with {\it Insight-HXMT} Data Analysis Software package (HXMTDAS) v2.04. The data were calibrated and screened under the same criteria as in \cite{Yang_2022}. The observation data were subdivided according to observation time. Since both spectral and variability properties do not change over the observation, we merged all data to achieve high signal-to-noise ratios. The energy spectrum obtained is shown in Fig.~\ref{fig:observation_data} (top). The different colours represent different telescopes (red: LE, green: ME, blue: HE; \citealt{Zhang_2020}). The energy spectrum from ME has a dip around 22~keV (light green), which is associated with silver fluorescent lines generated within the detector (\citealt{Li_2018}). Following \cite{You_2021}, we added 1.5~\% systematic errors to all spectral data. To study fast variability, we split background-subtracted light curves into segments of $256\,\si{s}$ with $1/128\,\si{s}$ time bins ($2^{15}$ points), where we avoided any data gaps. We only used the data where all telescopes were active to calculate power spectra and cross spectra using the same data points for every energy band. Following \cite{Uttley_2014, Ingram_2019}, we calculated the white-noise-subtracted power spectra and the cross spectra from segments and averaged them over different segments and logarithmically-spaced Fourier frequencies. We use the normalized power spectra for plots throughout the paper, such that their integral over frequency corresponds to the fractional variance (\citealt{Miyamoto_1991, Vaughan_2003}). Phase-lag spectra were calculated from the cross spectra, using the relation between the phase-lag spectrum $\phi (f)$ and cross spectrum $C(f)$, $\phi (f)=\tan ^{-1} (\Im[C(f)]/\Re[C(f)])$, where $\Re[\cdots]$ and $\Im[\cdots]$ denote the real and imaginary parts, respectively. The power spectra calculated for 2.6--4.8~keV and 35--48~keV light curves are shown in Fig.~\ref{fig:observation_data} (middle). These energy bands are marked in the energy spectrum with shaded regions. A fundamental QPO and its second harmonic exist around 0.036~Hz and 0.1~Hz (shown with dashed lines), in addition to the broad-band variability. The fundamental QPO appears to affect the phase lag between these two energy bands, creating a dip in the phase-lag spectrum around the corresponding frequency (\citealt{Ma_2021}), as shown in Fig.~\ref{fig:observation_data} (bottom). Positive lags mean hard lags throughout the paper. The effect of the second harmonic on the phase lag is not so clear between these energy bands. But we note that the bump at the corresponding frequencies is seen more clearly in phase-lag spectra for different choices of energy bands (\citealt{Ma_2021}). The phase lag from the broad-band variability has its peak at $\sim 1.2\,\si{Hz}$ (shown with a dotted line), which deviates from peak frequencies in the power spectra. This implies the complexity of the mechanism creating the broad-band variability since simple propagating fluctuations models lead to having the same peak frequency both in the power spectrum and cross spectrum (\citealt{Ingram_2013, Rapisarda_2016}). For all of data fits performed in this paper, we use {\tt XSPEC} 12.12.1 (\citealt{Arnaud_1996}). Variability data such as power spectra were formatted such that {\tt XSPEC} can import. We developed our model as an {\tt XSPEC} model. Being able to perform timing fits with the common tool in spectral fits is beneficial in performing spectral-timing fits. We ignore variability below $\sim 10^{-2}\,\si{Hz}$ because it behaves distinctly from other Fourier frequencies. \cite{Yang_2022} interpreted this low-frequency variability as the QPO sub-harmonic. \begin{figure*} \includegraphics[width=\linewidth]{figure/model_schematic_new.pdf} \caption{Overview of our model. Reverberation is not considered to focus on the propagating fluctuations process. (a) Assumed accretion flow geometry (height as a function of radius). The variable flow, where fluctuations are generated and propagate inwards, consists of radially stratified three spectral components. The characteristic radii are $(r_{\mathrm{in}}, r_{\mathrm{sh}}, r_{\mathrm{ds}}, r_{\mathrm{out}})=(6, 16, 32, 45)$. The outer stable disc does not contribute to the variability. (b) Generator frequency as a function of radius. It is modelled separately between the hot flow ($r<r_{\mathrm{ds}}$) and disc ($r_{\mathrm{ds}} \leq r$). The sample parameter values are $(B_{\mathrm{f}}, m_{\mathrm{f}})=(4, 1)$ and $(B_{\mathrm{d}}, m_{\mathrm{d}})=(0.03, 0.5)$. (c) Sample power spectra of the local mass accretion rate. The dashed lines represent intrinsic variability, where there is no interaction between different radii, at $r=45, 38, 33, 30, 22, 16, 12, 9, 6$ from left to right. The frequency, at which $f P(f)$ has its peak, corresponds to the local generator frequency $f_{\mathrm{gen}}(r)$. The red, green, and blue colours mean that the radii belong to the variable disc, soft Comptonisation, and hard Comptonisation regions, respectively. The solid lines represent correlated variability at $r=38$ (red), $22$ (green), $9$ (blue), where the inwards propagation of fluctuations is taken into account. (d) Time-averaged energy spectra. The disc, soft Comptonisation, hard Comptonisation, and their sum are colour-coded by red, green, blue, and black, respectively. To convert the variability of the mass accretion rate into that of the flux, a time-averaged spectrum is assign to each radius, according to which spectral region it constitutes. (e) Power spectra of the flux for 2.6--4.8~keV (black) and 35--48~keV (magenta). These energy bands are represented with shades of the same colours in Panel (d). (f) Phase-lag spectrum of the flux between 2.6--4.8~keV and 35--48~keV. } \label{fig:propagating_fluctuations_schematic} \end{figure*} \begin{table*} \caption{Summary of our model parameters. The variable flow is spectrally composed of the variable disc region and soft and hard Comptonisation regions from outer to inner. We call the entire Comptonisation regions the hot flow. } \begin{tabular}{llll} \hline Symbol &Meaning &Units &Default \\ \hline $M_{\mathrm{BH}}$ &Black hole mass. &$M_{\odot}$ &$8$ \\ $N_{\mathrm{r}}$ &Number of rings splitting the variable flow. & &$40$ \\ $r_{\mathrm{in}}$ &Inner radius of the hard Comptonisation. &$R_{\mathrm{g}}$ &$6$ \\ $r_{\mathrm{sh}}$ &Transition radius between the hard Comptonisation and soft Comptonisation. &$R_{\mathrm{g}}$ &$16$ \\ $r_{\mathrm{ds}}$ &Transition radius between the disc and soft Comptonisation. &$R_{\mathrm{g}}$ &$32$ \\ $r_{\mathrm{out}}$ &Outer radius of the variable disc. &$R_{\mathrm{g}}$ &$45$ \\ $F_{\mathrm{var, f}}$ ($F_{\mathrm{var, d}}$) &Fractional intrinsic variability per radial decade in the hot flow (variable disc). & &$0.8$ \\ $D$ &Damping factor. & &$0$ \\ $B_{\mathrm{f}}$ ($B_{\mathrm{d}}$) &Coefficient of the generator frequency in the hot flow (variable disc). & &$0.03$ \\ $m_{\mathrm{f}}$ ($m_{\mathrm{d}}$) &Power-law index of the generator frequency in the hot flow (variable disc). & &$0.5$ \\ $B^{(\mathrm{p})}_{\mathrm{f}}$ ($B^{(\mathrm{p})}_{\mathrm{d}}$) &Coefficient of the propagation frequency in the hot flow (variable disc). & &$0.03$ \\ $m^{(\mathrm{p})}_{\mathrm{f}}$ ($m^{(\mathrm{p})}_{\mathrm{d}}$) &Power-law index of the propagation frequency in the hot flow (variable disc). & &$0.5$ \\ $\gamma$ &Power-law index of the emissivity. & &3 \\ $b(r)$ &Inner boundary condition of the emissivity. & &$1-\sqrt{r_{\mathrm{in}}/r}$\\ $t_{0, \mathrm{h}}$ ($t_{0, \mathrm{s}}$) &Time delay of the top hat impulse response of reverberation for the hard (soft) Comptonisation$^a$. &$\si{s}$&$5.5\times 10^{-3}$\\ $\Delta t_{0, \mathrm{h}}$ ($\Delta t_{0, \mathrm{s}}$) &Time duration of the top hat impulse response of reverberation for the hard (soft) Comptonisation$^a$. &$\si{s}$&$10 \times 10^{-3}$\\ $S_{0}(E)$ &Fractional contribution of spectral components to the flux$^{b,c,d}$. & &0.5 \\ $\eta _{\mathrm{0, h}}$ ($\eta _{\mathrm{0, s}}$) &Constant term of the sensitivity of the hard (soft) Comptonisation to change in mass accretion rate$^e$. & &1 \\ $\eta _{\mathrm{1, h}}$ ($\eta _{\mathrm{1, s}}$) &Gradient term of the sensitivity of the hard (soft) Comptonisation to change in mass accretion rate$^e$. & &0 \\ \hline \multicolumn{4}{l}{{\it Notes.} $^a$ Parameters are required when the reverberation is considered.}\\ \multicolumn{4}{l}{$^b$ Each spectral component has its own parameter: $S_{0}(E)$ consists of $S_{\mathrm{d}}(E)$, $S_{\mathrm{s}}(E)$, $S^{(\mathrm{r})}_{\mathrm{s}}(E)$, $S_{\mathrm{h}}(E)$, and $S^{(\mathrm{r})}_{\mathrm{h}}(E)$ (see Section~\ref{sec:old}).}\\ \multicolumn{4}{l}{$^c$ $S_{0}(E)$ is replaced by $\eta (E) S_{0}(E)$ for timing fits when the spectral pivoting is included (Section~\ref{sec:rev2}).}\\ \multicolumn{4}{l}{$^d$ $S_{0}(E)$ is calculated from spectral models for spectral-timing fits (Section~\ref{sec:spec_timing}).}\\ \multicolumn{4}{l}{$^e$ Parameters are required for spectral-timing fits (Section~\ref{sec:spec_timing}).}\\ \end{tabular} \\ \label{tab:model_parameters} \end{table*} \section{Propagation and reverberation in our previous model} \label{sec:old} \subsection{Summary of our previous work} We start with connecting our current work to previous work (K22), during which we summarize our model. The overview of the model is shown in Fig.~\ref{fig:propagating_fluctuations_schematic}. Model parameters are summarized in Table~\ref{tab:model_parameters}, which also contains those introduced in the model updates. We give a more quantitative model summary in Appendix~\ref{sec:app_model_summary}. K22 used a spectral model of a disc as seed photons for soft and hard Comptonisation regions, together with their reflection. Such spectral models are quite degenerate in the {\it NICER} bandpass, so the simultaneous {\it NuSTAR} data were used to help constrain the components. This spectral model is coupled to a timing model by assuming geometry, where the variable disc, soft Comptonisation, hard Comptonisation regions are radially aligned across $r_{\mathrm{out}}$--$r_{\mathrm{ds}}$, $r_{\mathrm{ds}}$--$r_{\mathrm{sh}}$, $r_{\mathrm{sh}}$--$r_{\mathrm{in}}$ ($r_{\mathrm{in}}<r_{\mathrm{sh}}<r_{\mathrm{ds}}<r_{\mathrm{out}}$), respectively (see Fig.~\ref{fig:propagating_fluctuations_schematic} (a)). We use the lower-case letter $r$ as the radius from the central object in units of the gravitational radius $R_{\mathrm{g}}=GM_{\mathrm{BH}}/c^2$, where $G$, $M_{\mathrm{BH}}$, and $c$ are the gravitational constant, black hole mass, and speed of light in a vacuum. The timing model assumes that fluctuations are stirred up at the local generator time-scale at each radius and that these propagate down through the remaining inner parts of the flow in the propagation time-scale. In K22, we assumed that the generator time-scale is the same as the propagation time-scale, and call these time-scales the viscous time-scale altogether (\citealt{Lyubarskii_1997, Arevalo_2006, Ingram_2009}). However, \cite{Rapisarda_2017a} proposed from an observational point of view that these two time-scales are not exactly the same. Indeed, we will separate these time-scales as one of the model updates (Section~\ref{sec:rev1}). The viscous time-scale could be termed for both time-scales. To avoid confusion, we do not use the viscous time-scale but use the generator time-scale and propagation time-scale in the paper. The generator time-scale in a thin disc is well known to be related to the scale height $h$ via $f_{\mathrm{gen}}(r)=\alpha (h/r)^2 f_{\mathrm{K}}(r)$ where $f_{\mathrm{K}}(r)=(1/2\pi) r^{-3/2} (c/R_g)$ is the Keplarian frequency and $h$ is measured in units of $R_{\mathrm{g}}$ (\citealt{Shakura_1973, Novikov_1973, Churazov_2001, Ingram_2016}). However, neither the hot flow nor the turbulent edge of the thin disc is likely to be described by the untruncated thin disc properties. Hence in K22 we parameterised the generator time-scale as a power-law with radius, so that $f_{\mathrm{gen}}(r)=Br^{-m}f_{\mathrm{K}}(r)$. We allow this relationship to change between the very different physical regions so that there is a separate $B, m$ controlling the generator time-scale in the hot flow and turbulent disc: \begin{equation} f_{\mathrm{gen}} (r)= \begin{cases} B_{\mathrm{f}} r ^{-m_{\mathrm{f}}} f_{\mathrm{K}} (r) & (r_{\mathrm{in}} \leq r <r_{\mathrm{ds}}) ,\\ B_{\mathrm{d}} r ^{-m_{\mathrm{d}}} f_{\mathrm{K}} (r) & (r_{\mathrm{ds}} \leq r < r_{\mathrm{out}}), \end{cases} \label{eq:viscous_frequency} \end{equation} as shown in Fig.~\ref{fig:propagating_fluctuations_schematic} (b). Sample power spectra for the generated variability of the local mass accretion rate are shown with dashed lines in Fig.~\ref{fig:propagating_fluctuations_schematic} (c), where three radii are picked up from each spectral region. The functional form is a zero-centred Lorentzian with the cut-off frequency corresponding to the local generator frequency $f_{\mathrm{gen}}(r)$, which yields the peak at $f_{\mathrm{gen}}(r)$ in the $fP(f)$ representation. Slower fluctuations stirred up at larger radii modulate the faster fluctuations produced at smaller radii with a lag given by the propagation time (\citealt{Lyubarskii_1997}). K22 set the propagation speed from the generator time-scale as $v_{\mathrm{p}}(r)=rf_{\mathrm{gen}}(r)$. Sample power spectra for the variability, in which the inwards propagation of fluctuations is considered, are shown with solid lines in Fig.~\ref{fig:propagating_fluctuations_schematic} (c), where the middle radius out of the three selected radii is picked up for each spectral region. The analytic expression of the correlated variability (solid) using the intrinsic variability (dashed) is given in K22 Appendix~A (see \citealt{Ingram_2013} for the derivation). \cite{Rapisarda_2016} and K22 show how these two time-scales separated at the transition of the turbulent disc and hot flow makes the characteristic double hump shape of the power spectra. We set $(B_{\mathrm{d}}, m_{\mathrm{d}})=(0.03, 0.5)$ and $r_{\mathrm{out}}=45$ by assuming that the QPO is set by Lense-Thirring precession of the entire hot flow and the first bump in the power spectrum is set by the turbulent disc (\citealt{Ingram_2011}). The generator frequency at the outer edge is $f_{\mathrm{gen}}(r_{\mathrm{out}})=0.06\,\si{Hz}$, which corresponds to the observed low-frequency break (Fig.~\ref{fig:observation_data} (middle)). We assume that the local energy spectrum fluctuates in the same way as the local mass accretion rate at the corresponding radius. To convert the variability properties of mass accretion rate into those of the flux, the emission profile, how much each radius contributes to the flux, is needed (see Appendix~\ref{sec:app_model_summary} for details). We assign the time-averaged energy spectrum of the disc, soft Comptonisation, or hard Comptonisation to each radius according to which spectral region it constitutes. We note that what the model actually requires in calculating power spectra and cross spectra is the proportion of each spectral component in given energy bands, not the absolute flux. The sample time-averaged energy spectrum is shown in Fig.~\ref{fig:propagating_fluctuations_schematic} (d). Considering spectral hardening from the turbulent disc to soft and hard Comptonisation regions, high-energy photons tend to be emitted from inner regions, while low-energy photons are predominantly generated in outer regions. The analytic expressions of the power spectrum and cross spectrum for the flux are provided in K22 Appendix~A. Briefly, the model has radially-stratified variability and radially-stratified spectrum. The propagation of fluctuations produces correlated, lagged variability of mass accretion rate throughout the variable flow. This variability is encoded in radiation. The model provides energy-dependent flux variability because different energy bands have different emission profiles. The resultant power spectra calculated for the flux are compared between different energy bands in Fig.~\ref{fig:propagating_fluctuations_schematic} (e). Colours correspond to those of shades in the time-averaged energy spectrum (Fig.~\ref{fig:propagating_fluctuations_schematic} (d)). High-frequency variability ($\gtrsim 1\,\si{Hz}$) is more pronounced for the higher energy band because the higher-energy photons are more predominantly emitted from inner regions having higher-frequency variability. On the other hand, these power spectra are almost identical below $\sim 1\,\si{Hz}$, indicating that variability on these slow time-scales is propagated from outer regions rather than generated at their emission regions. This behaviour is consistent with the model assumption in the example that the hot flow, where most photons are radiated for both energy bands (Fig.~\ref{fig:propagating_fluctuations_schematic} (d)), has generator frequencies higher than $\sim 1\,\si{Hz}$ (Fig.~\ref{fig:propagating_fluctuations_schematic} (b)). The phase-lag spectrum calculated between these energy bands is shown in Fig.~\ref{fig:propagating_fluctuations_schematic} (f). Hard lags are primarily seen for 1--20~Hz as a result of the combination of the propagating fluctuations and spectral inhomogeneity. The little lags below $\sim 1\,\si{Hz}$ are attributed to the variability on these time-scales being propagated rather than generated for both energy bands, as described above. On the other hand, these energy bands lose lags above $\sim 20\,\si{Hz}$ because the variability on these fast time-scales is commonly dominated by the hard Comptonisation component for both the energy bands. The fluctuating soft and hard Comptonisation regions illuminate the outer disc and produce a reflected/reprocessed signal which lags behind the generated and propagated flow variability by the light travel time to the disc. The reflected emission itself is not strong, but photons which are not reflected heat the disc, giving a thermal reverberation signal which is strong at energies close to that of the disc emission ($\lesssim 2\,\si{keV}$; \citealt{Kara_2019}). This reverberation signal gives an independent check on the assumption of the disc truncation radius, and the fact that it is consistent (\citealt{DeMarco_2021}) gives strong supporting evidence for the underlying assumption that the QPO mechanism is Lense-Thirring precession. Our previous model includes the reverberation, along with the propagating fluctuations (K22). K22 showed that this model gave a fairly good fit to the energy dependence of the power spectrum across the {\it NICER} energy band (0.5--10~keV) and to the lags between the same fluctuations in different energy bands as a function of frequency. However, while the spectral components were built from {\it NuSTAR} data which extended above 10~keV, this instrument does not have a sufficient area to do high time-resolution studies, so the model prediction at higher energies could not be tested. This outburst of MAXI~J1820+070 was also monitored by {\it Insight-HXMT} (\citealt{Ma_2021, You_2021, Yang_2022}), which does have a sufficient effective area at high energies. As mentioned in the previous section, the {\it Insight-HXMT} data are not absolutely simultaneous with the {\it NICER}/{\it NuSTAR} dataset we used in K22, but they are very close in time, and spectral and variability properties are nearly constant during these periods. Hence we take the spectral-timing model of K22, use it to predict the higher energy behaviour, and compare it to the {\it Insight-HXMT} data. \begin{figure*} \includegraphics[width=\linewidth]{figure/psd_phase_fit_comp_nolorentz_new.pdf} \caption{Comparison of our model calculations between different versions. In the top and middle columns, the power spectra for 2.4--4.8~keV (black) and 35--48~keV (magenta) and the phase-lag spectrum between these energy bands are compared between the observation (stepped) and model (smooth), respectively. The bottom column shows the propagation frequency (solid) and generator frequency (dashed) used in the model calculations, with the Keplerian frequency (dash-dotted) for reference. {\it Left}: Predictions from the previous model from K22 (Section~\ref{sec:old}). The previous model was applied for the time-averaged spectrum from 0.5--73~keV and variability for 0.5--10~keV. {\it Mid-left}: Fitting with the previous model (Section~\ref{sec:old}). The time-averaged spectrum is ignored. {\it Mid-right}: Additional effects proposed in the context of other propagating fluctuations models (\citealt{Rapisarda_2017a, Mahmoud_2018b}) are included (Section~\ref{sec:rev1}). Eventually, the separate treatment of the generator time-scale and propagation time-scale is incorporated. {\it Right}: The change in spectral shapes, spectral pivoting, is taken into consideration (Section~\ref{sec:rev2}). The significant improvement of consistency of the model calculations to observation data points out the importance of the spectral variation in variability modelling at high energies.} \label{fig:psd_phase_fit_comp} \end{figure*} \begin{table} \caption{Model parameter values used in Figs.~\ref{fig:psd_phase_fit_comp}. Common parameter values used in all fitting are $M_{\mathrm{BH}}=8$, $N_{\mathrm{r}}=40$, $r_{\mathrm{in}}=6$, $r_{\mathrm{out}}=45$, $B_{\mathrm{d}}=B^{(\mathrm{p})}_{\mathrm{d}}=0.03$, $m_{\mathrm{d}}=m^{(\mathrm{p})}_{\mathrm{d}}=0.5$, $\gamma=3$, $b(r)=1-\sqrt{r_{\mathrm{in}}/r}$. The transition radii are fixed to $r_{\mathrm{sh}}=17.8$, $r_{\mathrm{ds}}=32.1$ for the left column and $r_{\mathrm{sh}}=16$, $r_{\mathrm{ds}}=32$ for other columns, although these differences are too subtle to become important. Other constraints are $F_{\mathrm{var, f}}=F_{\mathrm{var, d}}$ and $D=0$. The mark, `(f)', means that a value of the corresponding parameter is fixed.} \begin{tabular}{lllll} \hline Symbol&Left&Mid-left&Mid-right&Right$^{a}$\\ \hline $F_{\mathrm{var, d}}$ &$0.8$ (f) &$0.53$ &$0.59$ &$0.8$ (f) \\ $B_{\mathrm{f}}$ &$6$ (f) &$11.9$ &$342$ &$560$ \\ $m_{\mathrm{f}}$ &$1.2$ (f) &$1.50$ &$2.42$ &$2.73$ \\ $B^{(\mathrm{p})}_{\mathrm{f}}$ &$=B_{\mathrm{f}}$ &$=B_{\mathrm{f}}$ &$80.1$ &$166$ \\ $S_\mathrm{d}(2.6\textrm{--}4.8\,\si{keV})$ &$0.001$ (f) &$0$ (f) &$0$ (f) &$0$ (f) \\ $S_\mathrm{s}(2.6\textrm{--}4.8\,\si{keV})$ &$0.356$ (f) &$0.505$ &$0.471$ &$0.305$ \\ $S_\mathrm{h}(2.6\textrm{--}4.8\,\si{keV})$ &$0.330$ (f) &$=1-S_{\mathrm{s}}(E)$&$=1-S_{\mathrm{s}}(E)$ &$0.571$ \\ $S^{(\mathrm{r})}_\mathrm{s}(2.6\textrm{--}4.8\,\si{keV})$&$0.307$ (f) &$0$ (f) &$0$ (f) &$0$ (f) \\ $S^{(\mathrm{r})}_\mathrm{h}(2.6\textrm{--}4.8\,\si{keV})$&$0.006$ (f) &$0$ (f) &$0$ (f) &$0$ (f) \\ $S_\mathrm{d}(35\textrm{--}48\,\si{keV})$ &$0$ (f) &$0$ (f) &$0$ (f) &$0$ (f) \\ $S_\mathrm{s}(35\textrm{--}48\,\si{keV})$ &$0.213$ (f) &$0.348$ &$0.277$ &$-0.056$ \\ $S_\mathrm{h}(35\textrm{--}48\,\si{keV})$ &$0.474$ (f) &$=1-S_{\mathrm{s}}(E)$&$=1-S_{\mathrm{s}}(E)$ &$0.476$ \\ $S^{(\mathrm{r})}_\mathrm{s}(35\textrm{--}48\,\si{keV})$ &$0.134$ (f) &$0$ (f) &$0$ (f) &$0$ (f) \\ $S^{(\mathrm{r})}_\mathrm{h}(35\textrm{--}48\,\si{keV})$ &$0.179$ (f) &$0$ (f) &$0$ (f) &$0$ (f) \\ \hline \multicolumn{5}{l}{{\it Notes.} $^a$ $S_{0}(E)$ means $\eta (E)S_{0}(E)$ in this column.}\\ \end{tabular}\\ \label{tab:model_parameter_values} \end{table} \subsection{Comparison of our previous model to {\it Insight-HXMT} data} We compare the predictions of our previous model to the power spectra for the 2.6--4.8~keV and 35--48~keV bands and the phase-lag spectrum between these bands calculated from the {\it Insight-HXMT} observation data in Fig.~\ref{fig:psd_phase_fit_comp} (left). Model parameter values are summarised in Table~\ref{tab:model_parameter_values}, which also contains those for the rest of the columns in Fig.~\ref{fig:psd_phase_fit_comp}. The lower energy band is well reproduced by our previous model from K22, as expected, as it is within the {\it NICER} energy range over which K22 got good fits. However, the power spectrum at the higher energy band is clearly overestimated, and the phase lag between the two is completely wrong, peaking at too high a frequency with a lag which is too short to match the data. Next, we explore how well our previous model is capable of reproducing these observed timing properties. For this purpose, we ignore the time-averaged spectrum and attempt to minimize the sum of $\chi ^2$ values for the power spectra and phase-lag spectrum: \begin{equation} \begin{split} \sum_{k} \biggl\{ &\left( \frac{P_{\mathrm{data}}(E_1, f_{k})-P_{\mathrm{model}}(E_1, f_{k})}{\Delta P_{\mathrm{data}}(E_1, f_{k})} \right) ^2 \\ + &\left( \frac{P_{\mathrm{data}}(E_2, f_{k})-P_{\mathrm{model}}(E_2, f_{k})}{\Delta P_{\mathrm{data}}(E_2, f_{k})} \right) ^2 \\ + &\left( \frac{\phi_{\mathrm{data}}(E_1, E_2, f_{k})-\phi_{\mathrm{model}}(E_1, E_2, f_{k})}{\Delta \phi_{\mathrm{data}}(E_1, E_2, f_{k})} \right) ^2 \biggr\}, \\ \end{split} \end{equation} where $P_{\mathrm{data}}(E, f)$, $\Delta P_{\mathrm{data}}(E, f)$ are the observed power spectrum and its one-sigma error at frequency $f$ for energy $E$, $\phi_{\mathrm{data}}(E, E', f)$, $\Delta \phi_{\mathrm{data}}(E, E', f)$ the equivalents for the phase-lag spectrum between energy $E$ and $E'$, $P_{\mathrm{model}}(E, f)$ and $\phi (E, E', f)$ the modelled power spectrum and phase-lag spectrum, $E_{1}=2.6\textrm{--}4.8\,\si{keV}$, $E_{2}=35\textrm{--}48\,\si{keV}$, and $f_{k}$ the sampled Fourier frequency. The time-averaged energy spectrum is excluded from the fitting (we will come back to joint spectral-timing modelling in Section~\ref{sec:spec_timing}). Our model requires the fraction of each spectral component in calculating power spectra and cross spectra. We express this fraction as $S_{\mathrm{d}}(E)$, $S_{\mathrm{s}}(E)$, $S^{(\mathrm{r})}_{\mathrm{s}}(E)$, $S_{\mathrm{h}}(E)$, $S^{(\mathrm{r})}_{\mathrm{h}}(E)$ for the variable disc, soft Comptonisation and its reflection, hard Comptonisation and its reflection, respectively. Assuming that these spectral components give a complete set of X-ray emission for the energy band of interest, the sum of these fractions corresponds to unity: \begin{equation} S_{\mathrm{d}}(E)+S_{\mathrm{s}}(E)+S^{(\mathrm{r})}_{\mathrm{s}}(E)+S_{\mathrm{h}}(E)+S^{(\mathrm{r})}_{\mathrm{h}}(E)=1. \label{eq:spec_constraint} \end{equation} Whereas in Fig.~\ref{fig:psd_phase_fit_comp} (left), these fractions were calculated from the result of spectral fit in K22 for the self-consistent spectral-timing modelling, we will let them be independent of the time-averaged spectrum in order to focus on the variability properties for the time being. We fix $S_{\mathrm{d}}(E)=0$ because the disc emission is negligible above 2.6~keV. We also ignore the reverberation, i.e., $S^{(\mathrm{r})}_{\mathrm{s}}(E)=S^{(\mathrm{r})}_{\mathrm{h}}(E)=0$, to simplify the model. This is validated because our attempt is to capture the broad-band power spectrum and hard lags, which are expected to originate from the propagating fluctuations. The reverberation alters variability properties mildly, not drastically, for the energy range of interest. Finally, we only have the soft and hard Comptonisation components with the constraints of $S_{\mathrm{s}}(E)+S_{\mathrm{h}}(E)=1$. We do not include any models for the QPO features for simplicity. Ignoring the reverberation and QPOs does not affect the investigation of the capability of our previous model. We keep the black hole mass of $M_{\mathrm{BH}}=8M_{\odot}$ (\citealt{Torres_2020}) and emissivity profile, i.e., $\gamma=3$ and $b(r)=1-\sqrt{r_{\mathrm{in}}/r}$ (\citealt{Shakura_1973, Novikov_1973}), where we assume that radiation energy from the annulus ranging from $r$ to $r+\Delta r$ is proportional to $r^{-\gamma}b(r)2\pi r \Delta r$. In K22, the transition radii $r_{\mathrm{sh}}, r_{\mathrm{ds}}$ were calculated from the emissivity profile and spectral decomposition. However, we lack spectral decomposition. In addition, it turned out that model calculations are hardly affected by small changes in these parameters. Thus, we simply fix these transition radii to typical values, $r_{\mathrm{sh}}=16$ and $r_{\mathrm{sh}}=32$. We show the result of the joint fit to the power spectra for 2.6--4.8~keV and 35--48~keV and the phase-lag spectrum between these energy bands in Fig.~\ref{fig:psd_phase_fit_comp} (mid-left). Our model does not improve the fit even though we ignore the time-averaged energy spectrum. While the power spectrum for the higher energy band is preferentially modelled, the power spectrum for the lower energy band is completely underestimated, in turn. Plainly, while the previous model from K22 was designed to fit the data below 10~keV, it does not extrapolate to the higher energies, so does not adequately describe the physics of the propagation of fluctuations through the flow. This is important as K22 shows that the propagation speed is a key determinant of the nature of the hot flow, which can allow large-scale magnetically dominated flows (MAD) to be distinguished from those with turbulent dynamo (SANE) models. The poor applicability of our previous model to higher energy bands motivates our study to improve it. \section{Suppressing variability at high energies with a constant spectral shape} \label{sec:rev1} The major feature missing in the previous model for the power spectra is the strong suppression of fractional variability at high energies. The generation/propagation of fluctuations in the model, where whole fluctuations generated outer regions propagate down through the flow, always leads to an increase in variability with energy, as long as the spectrum hardens inwards. In contrast, the {\it Insight-HXMT} observation data show that plainly the high-energy broad-band power spectrum is a factor $\sim 3$ lower than the low-energy power spectrum at all frequencies (Fig.~\ref{fig:observation_data} (middle)). This decrease in fractional variability with energy was not seen in the {\it NICER} energy band ($\lesssim 10\,\si{keV}$; K22). But it has been seen before, in e.g. the {\it RXTE} data of other black hole binary low/hard states (e.g. \citealt{Nowak_1999, Axelsson_2018} for Cyg~X-1; \citealt{Malzac_2003} for XTE~J1118+480). In the context of other propagating fluctuations models, it was modelled by the damping of high-frequency fluctuations as they propagate inwards (\citealt{Arevalo_2006, Rapisarda_2017a}), and by decreasing the intrinsic variability power generated in the inner regions (\citealt{Mahmoud_2018b}). To implement these effects in our model, we introduce two new parameters. One is a damping parameter $D$, which allows the model to prevent high-frequency variability from propagating inwards. The variability is suppressed exponentially with $\mathrm{exp}(-D f \Delta t)$ via inwards propagation in the frequency domain, where $\Delta t$ is the propagation time. The damping effect is ignored if $D=0$. The other is for allowing the intrinsic variability amplitude to be different between the hot flow $F_{\mathrm{var, h}}$ and disc $F_{\mathrm{var, d}}$ (the previous model from K22 has $F_{\mathrm{var,f}}=F_{\mathrm{var,d}}$). Another observational feature that our previous model fails to capture is the discrepancy in the frequency at which the power spectra (in the $fP(f)$ representation) and phase-lag spectra peak. The power spectra and phase-lag spectra calculated by the model have a similar peak frequency. This observational property is also seen in the {\it RXTE} data (e.g. XTE J1550-564: \citealt{Rapisarda_2017a}), where the proposed solution was to allow the propagation time-scale to be different to the generator time-scale on which the fluctuations are generated. Following this, we separate these time-scales and define the propagation frequency with \begin{equation} f_{\mathrm{prop}} (r)= \begin{cases} B^{(\mathrm{p})}_{\mathrm{f}} r ^{-m^{(\mathrm{p})}_{\mathrm{f}}} f_{\mathrm{K}} (r) & (r_{\mathrm{in}} \leq r_{n} <r_{\mathrm{ds}}) ,\\ B^{(\mathrm{p})}_{\mathrm{d}} r ^{-m^{(\mathrm{p})}_{\mathrm{d}}} f_{\mathrm{K}} (r) & (r_{\mathrm{ds}} \leq r_{n} < r_{\mathrm{out}}), \end{cases} \label{eq:propagation_frequency} \end{equation} such that the propagation speed is provided by $v_{\mathrm{p}}(r)=rf_{\mathrm{prop}}(r)$. Since it is difficult to constrain disc parameters from the energy range of interest, we simply fix $B^{(\mathrm{p})}_{\mathrm{d}}=B_{\mathrm{d}}=0.03$ and $m^{(\mathrm{p})}_{\mathrm{d}}=m_{\mathrm{d}}=0.5$. To reduce the number of free parameters, we assume that $f_{\mathrm{prop}}(r)$ has the same radial dependence as $f_{\mathrm{gen}}(r)$, i.e., $m^{(\mathrm{p})}_{\mathrm{f}}=m_{\mathrm{f}}$. Eventually, we have only one additional parameter $B^{(\mathrm{p})}_{\mathrm{f}}$. The modified model formalism due to the damping effect is given in Appendix~\ref{sec:app_damping}. The implementation of other effects, that is, $F_{\mathrm{var, d}} \neq F_{\mathrm{var, f}}$ and $f_{\mathrm{gen}} (r) \neq f_{\mathrm{prop}}(r)$, is trivial. As in the last part of the previous section, we attempt to reproduce only variability properties based on the propagating fluctuations process. We keep those parameters fixed which are fixed in the previous fit. Even with all these additional effects, the model is not capable of matching the observation data. The damping parameter $D$ is pegged to its lower bound of zero, indicating that the damping described above is ineffective in improving the fit (\citealt{Mahmoud_2019}). This could be due to the following reason. Our model assumes that the intrinsic variability has a cut-off at the local generator frequency $f_{\mathrm{gen}}(r)$, as shown in Fig.~\ref{fig:propagating_fluctuations_schematic} (c). The high-frequency variability does not exist from the definition. We see that this assumption already includes some aspects of the damping. The MRI (\citealt{Balbus_1991, Balbus_1998}) is expected to produce variability up to faster than $f_{\mathrm{gen}}(r)$. However, it is variability slower than $f_{\mathrm{gen}}(r)$ that propagate inwards because the faster variability is viscously damped out (\citealt{Churazov_2001, Cowperthwaite_2014, Ingram_2016, Hogg_2016, Bollimpalli_2020, Turner_2021}). Our assumption on the intrinsic variability is an approximation of this physical picture. The damping parameter being pegged to zero indicates no need for additional damping effects. We did not find an improvement in the fits using separate variability amplitude between the variable disc region and hot flow region, either. Eventually, we perform the joint fit to the power spectra for 2.6--4.8~keV and 35--48~keV and the phase-lag spectrum between these energy bands by allowing $f_{\mathrm{gen}}(r) \neq f_{\mathrm{prop}}(r)$ for the hot flow and by removing any other additional effects, i.e., $F_{\mathrm{var, d}}=F_{\mathrm{var, f}}$ and $D=0$. The result is shown in Fig.~\ref{fig:psd_phase_fit_comp} (mid-right). Although we see a slightly better match to the observed phase-lag spectrum than the previous fit, the model still underestimates it considerably. More importantly, we still do not solve the essential issue: the model calculates larger variability for higher energy bands, inconsistent with the observation. \cite{Mahmoud_2018b, Mahmoud_2019} introduce more complex radial dependence for the intrinsic variability, emissivity and damping to capture energy-dependent variability properties. However, some assumptions involved with these complications remain to be tested. We do not explore the complex radial structure further and conclude that those additional effects implemented here are less effective than required by the high signal-to-noise ratio data obtained by {\it Insight-HXMT}. We note that the difficulty in reproducing the observation data here lies in joint fitting to the power spectra and phase-lag spectrum. It is possible to reproduce power spectra for these energy bands with the current model fairly well, ignoring the phase-lag spectrum. In this case, however, the lower-energy photons would come from inner regions, because inner regions are more variable than outer regions, and predict soft lags, which is completely in disagreement with the observed hard lags. This points to the importance of modelling cross spectra and power spectra. \begin{figure} \includegraphics[width=0.95\columnwidth]{figure/spectral_variation.pdf} \caption{Schematic picture of the spectral pivoting implemented as main model updates. The black and grey lines represent the spectra when mass accretion rates are the same as and higher than the average, respectively. The parameter $\eta (E)$ regulates how the spectrum responds to mass accretion rate fluctuations. The above picture shows the spectral pivoting, where $\eta(E_1), \eta(E_2)>0$ (positive correlation) and $\eta(E_3)<0$ (negative correlation).} \label{fig:spectral_variation} \end{figure} \begin{figure} \includegraphics[width=\columnwidth]{figure/psd_phase_fit_multi.pdf} \caption{Joint fit to six power spectra (top) and five phase-lag spectra (bottom) across 2.6--48~keV with our new model including the spectral pivoting. The power spectra are modelled with the sum of our new model for the broad-band variability and two Lorentzian functions for the QPOs. Each component is explicitly shown with dotted and dashed lines, respectively, for the highest energy band. In the calculation of phase-lag spectra, the lowest band of 2.6--4.8~keV is chosen as the reference band. The lower plot for each panel is the difference between data and model divided by one-sigma errors. The new model including the spectral pivot successfully reproduces the observation data. } \label{fig:psd_phase_fit_multi} \end{figure} \section{Spectral pivoting} \label{sec:rev2} So far, we have assumed that the spectral shape of each component does not vary in time. However, the constant spectral shape is unrealistic because mass accretion rate fluctuations make spectral parameters, e.g., the optical depth and electron temperature, vary on short time-scales (\citealt{Malzac_2003, Gandhi_2008, Yamada_2013, Bhargava_2022}). This oversimplification limits the model's flexibility to reproduce energy-dependent variability data. As the main updates of our model, we allow the spectral shape to fluctuate (\citealt{Veledina_2016, Veledina_2018, Mastroserio_2018, Mastroserio_2019, Mastroserio_2021}), along with the amplitude. Here, we give concise explanations of how the spectral pivoting is implemented and what the model gets to be able to handle with this update. More specific formalism is found in Appendix~\ref{sec:app_pivot}. The constant spectral shape means that the spectrum at every energy reacts to mass accretion fluctuations in the same way. We consider mass accretion rate and energy spectrum at a certain radius. By defining the average and difference from the average as $\dot{m}_0$ and $\Delta \dot{m}(t)$ for the mass accretion rate and as $S_0 (E)$ and $\Delta S(E, t)$ for the spectrum, the constant spectral shape is equivalent to $\Delta S (E, t)/S_0 (E)=\Delta \dot{m} (t)/\dot{m}_0$, which is independent of energy $E$. To let the spectral shape vary in time, we give the spectrum sensitivity to $\Delta \dot{m}(t)$ as a function of energy, $\eta (E)$, and redefine $\Delta S (E, t)/S_0 (E)=\eta(E) \Delta \dot{m} (t)/\dot{m}_0$, which now depends on energy. The amplitude of sensitivity parameter $|\eta (E)|$ regulates how sensitive the spectrum is to change of the mass accretion rate from its average, while its sign determines whether the spectrum reacts positively or negatively. The spectrum gets higher (lower) with an increase in mass accretion rate if $\eta (E) >0$ ($<0$). The energy at which $\eta (E)$ crosses zero, called the pivoting point, does not react to a change in mass accretion rate. The schematic picture of the spectral pivoting is shown in Fig.~\ref{fig:spectral_variation}. The decrease in $\eta (E)$ with energy, i.e., the spectrum being less sensitive to $\Delta \dot{m}(t)$ for higher energies, could let the power spectrum decrease with energy, as observed for MAXI~J1820+070, even if the mass accretion rate is more variable for central regions emitting higher-energy photons. In our implementation, there arises no lag between different energies from the spectral pivoting itself except for the phase lag of $\pi$ when $\eta (E_1)\eta (E_2)<0$. Our new model shares this feature of spectral pivoting with the model developed by \cite{Veledina_2016, Veledina_2018}. The new model returns to the previous one by setting $\eta (E)=1$. Each spectral component is expected to show its own sensitivity pattern. We give the sensitivity parameter to each spectral component, $\eta _{\mathrm{Y}}(E)\,(\mathrm{Y}={\mathrm{d, s, h}})$, where the lower subscripts stand the variable disc, soft Comptonisation, and hard Comptonisation, respectively. With the implementation of spectral pivoting, all $S_{\mathrm{Y}}(E)\,(\mathrm{Y}={\mathrm{d, s, h}})$ contained in the analytic expressions of power spectra and cross spectra is replaced by $\eta _{\mathrm{Y}}(E)S_{\mathrm{Y}}(E)$ (see Appendix~\ref{sec:app_pivot} for the derivation). This means that the model's flexibility is not bound by the constraint (\ref{eq:spec_constraint}) anymore because time-averaged spectra always appear as the product with their sensitivity. In addition, $\eta _{\mathrm{Y}}(E)S_{\mathrm{Y}}(E)$ can be negative in contrast to $0\leq S_{\mathrm{Y}}(E) \leq 1$. The spectral pivoting gives freedom to the model in this way. We attempt to fit the variability properties with the new model. We have $\eta _{\mathrm{Y}}(E) S_{\mathrm{Y}}(E)$ as model parameters, instead of $S_{\mathrm{Y}}(E)$. The negligible disc emission $S_{\mathrm{d}}(E)=0$ results in $\eta_{\mathrm{d}}(E)S_{\mathrm{d}}(E)=0$. We fix $D=0$, in which all intrinsic variability propagates inwards without any loss. We also fix $F_{\mathrm{var, d}}=F_{\mathrm{var, f}}$ to the typical value of $0.8$ because the sensitivity parameter $\eta (E)$ can regulate the variability amplitude. The simultaneous fit to the power spectra for 2.6--4.8~keV and 35--48~keV and the phase-lag spectrum between these energy bands with the new model is shown in Fig.~\ref{fig:psd_phase_fit_comp} (right). We see significant improvement in variability modelling by allowing the spectral shapes to vary in time. Our new model captures the energy-dependent variability, pointing to the importance of spectral pivoting in modelling variability at high energies. To study the variability for a continuous energy range, we split energy between 2.6--4.8~keV (LE) and 35--48~keV (HE) into four bands, i.e., 4.8--7~keV (LE), 7--11~keV (LE), 11--23~keV (ME), 23--35~keV (ME), where the telescopes used are specified in parenthesis, and attempt to reproduce power spectra for these six energy bands and phase-lag spectra with respect to the lowest energy band for the rest of five energy bands. We minimize \begin{equation} \begin{split} &\sum_{j, k} \left( \frac{P_{\mathrm{data}}(E_{j}, f_{k})-P_{\mathrm{model}}(E_{j}, f_{k})}{\Delta P_{\mathrm{data}}(E_{j}, f_{k})} \right) ^2 \\ + &\sum _{\substack{j, k\\(E_{j} \neq E_{\mathrm{r}})}} \left( \frac{\phi_{\mathrm{data}}(E_{\mathrm{r}}, E_{j}, f_{k})-\phi_{\mathrm{model}}(E_{\mathrm{r}}, E_{j}, f_{k})}{\Delta \phi_{\mathrm{data}}(E_{\mathrm{r}}, E_{\mathrm{j}}, f_{k})} \right) ^2 , \\ \end{split} \end{equation} through the fit, where $E_j$ is each energy band and $E_{\mathrm{r}}=2.6\textrm{--}4.8\,\si{keV}$ the reference band. For more complete modelling, we add two Lorentzian functions to model the QPOs in power spectra by using the {\tt XSPEC} model {\tt lorentz}. In all energy bands, we fix the centroid and width to $3.66 \times 10^{-2}\,\si{Hz}$, $1.20 \times 10^{-2}\,\si{Hz}$ for the QPO fundamental, $9.44 \times 10^{-2}\,\si{Hz}$, $1.16 \times 10^{-1}\,\si{Hz}$ for the second harmonic. Thus, the ${\tt lorentz}$ model has only one free parameter, the normalization. On the other hand, we do not use any additional models in phase-lag spectra due to the relatively small QPO features. The results of the joint fit to six power spectra and five phase-lag spectra are shown in Fig.~\ref{fig:psd_phase_fit_multi}. Each component forming power spectra is explicitly plotted with dashed (QPOs) and dotted (broad-band) lines only for the highest energy band of 35--48~keV. Model parameter values are summarized in Table~\ref{tab:model_parameter_values_multi}. We find that the new model matches observations well for all energy bands whilst keeping parameter values similar to those found in the joint fitting for 2.4--4.8~keV and 35--48~keV only (Fig.~\ref{fig:psd_phase_fit_comp} (right)). It is interesting to note that the spectral parameter for the soft Comptonisation component $\mu _{\mathrm{s}}(E) S_{\mathrm{s}}(E)$ decreases with energy and finally reaches a negative value at the highest energy band of 35--48~keV. This means that the soft Comptonisation component increases for an increase in mass accretion rate at low energies ($\lesssim 35\,\si{keV}$), whereas it decreases at high energies ($\gtrsim 35\,\si{keV}$), showing the pivoting point of $\sim 35\,\si{keV}$. Although the broad-band variability has been studied with {\it Insight-HXMT} observations (e.g. \citealt{Wang_2020, Yang_2022}), we succeeded in reproducing it with a physically motivated model for the first time. In addition, while propagating fluctuations models have been applied up to $\sim 35\,\si{keV}$ (\citealt{Mahmoud_2018, Mahmoud_2018b}) with {\it RXTE} observations, we extend the energy range up to 48~keV using {\it Insight-HXMT} observations with significantly improved residuals. Our successful modelling shows the propagating fluctuations scenario holds good up to high energy bands, keeping it the most plausible explanation for the aperiodic variability. \begin{table} \caption{Model parameter values used in Figs.~\ref{fig:psd_phase_fit_multi}. Fixed parameter values are the same as in Table~\ref{tab:model_parameter_values}.} \begin{tabular}{ll} \hline Symbol&Value\\ \hline $F_{\mathrm{var, d}}$ &$0.8$ (f) \\ $B_{\mathrm{f}}$ &$401$ \\ $m_{\mathrm{f}}$ &$2.59$ \\ $B^{(\mathrm{p})}_{\mathrm{f}}$ &$108$ \\ $\eta_{\mathrm{s}} S_\mathrm{s}(2.6\textrm{--}4.8\,\si{keV})$ &$0.315$ \\ $\eta_{\mathrm{h}} S_\mathrm{h}(2.6\textrm{--}4.8\,\si{keV})$ &$0.588$ \\ $\eta_{\mathrm{s}} S_\mathrm{s}(4.8\textrm{--}7\,\si{keV})$ &$0.242$ \\ $\eta_{\mathrm{h}} S_\mathrm{h}(4.8\textrm{--}7\,\si{keV})$ &$0.601$ \\ $\eta_{\mathrm{s}} S_\mathrm{s}(7\textrm{--}11\,\si{keV})$ &$0.189$ \\ $\eta_{\mathrm{h}} S_\mathrm{h}(7\textrm{--}11\,\si{keV})$ &$0.617$ \\ $\eta_{\mathrm{s}} S_\mathrm{s}(11\textrm{--}23\,\si{keV})$ &$0.116$ \\ $\eta_{\mathrm{h}} S_\mathrm{h}(11\textrm{--}23\,\si{keV})$ &$0.549$ \\ $\eta_{\mathrm{s}} S_\mathrm{s}(23\textrm{--}35\,\si{keV})$ &$0.049$ \\ $\eta_{\mathrm{h}} S_\mathrm{h}(23\textrm{--}35\,\si{keV})$ &$0.498$ \\ $\eta_{\mathrm{s}} S_\mathrm{s}(35\textrm{--}48\,\si{keV})$ &$-0.052$ \\ $\eta_{\mathrm{h}} S_\mathrm{h}(35\textrm{--}48\,\si{keV})$ &$0.484$ \\ \hline $\chi ^2 /\mathrm{d.o.f.}$ &$1095.0/457$ \\ \hline \end{tabular}\\ \label{tab:model_parameter_values_multi} \end{table} \section{Joint spectral-timing fit with spectral pivoting} \label{sec:spec_timing} \begin{figure*} \includegraphics[width=\linewidth]{figure/spec_psd_phase_fit_reflection.pdf} \caption{Joint fit to the time-averaged energy spectrum (left), five power spectra (middle), and four phase-lag spectra (right) across 2.6--48~keV with our new model including the spectral pivoting. In the left panel, the soft Comptonisation and its associated reflection are plotted with the green and light green lines, while the hard Comptonisation and its associated reflection are plotted with the blue and light blue lines. The black line shows their sum. The galactic absorption of $N_{\mathrm{H}}=1.4 \times 10^{21}\,\si{cm^{-2}}$ is considered for plotting spectral components. The colours of shades correspond to those in the power spectra and phase-lag spectra. In the mid panel, the power spectrum at the highest energy band is decomposed, as in Fig.~\ref{fig:psd_phase_fit_multi}. The bottom panels show residuals. The data-to-model ratio is used for the energy spectrum, while the difference between data and model divided by one-sigma errors is used for power spectra and phase-lag spectra. We successfully fit all data with the updated model including the spectral pivoting except the phase-lag spectra for the highest energy band. } \label{fig:spec_psd_phase_fit} \end{figure*} \begin{figure} \includegraphics[width=\columnwidth]{figure/spectral_pivoting_fit_reflection.pdf} \caption{Changes in energy spectrum due to changes in mass accretion rate derived from the spectral-timing fit in Fig.~\ref{fig:spec_psd_phase_fit}. Colours are the same as those in parameters Fig.~\ref{fig:spec_psd_phase_fit} (left). The energy spectra for the mass accretion rate on average and double the average are plotted with solid and dashed lines, respectively. The galactic absorption is not considered. The energy spectra change their shape as well as amplitude. } \label{fig:spectral_pivoting_fit_reflection} \end{figure} \begin{table} \caption{Model parameter values derived from the joint spectral-timing fit in Fig.~\ref{fig:spec_psd_phase_fit}. Fixed parameters related to spectrum are the seed photon temperature $kT_{\mathrm{seed, s}}=kT_{\mathrm{seed, h}}=0.2\,\si{keV}$, electron temperature $kT_{\mathrm{e, s}}=kT_{\mathrm{e, h}}=23\,\si{keV}$, Fe abundance $Z_{\mathrm{Fe}}=1.1$, inclination angle $i=66^{\circ}$, black hole spin $a^{*}=0$, electron density $N_{\mathrm{e}}=10 ^{20}\,\si{cm^{-3}}$, inner radius of reflection region $R_{\mathrm{in, s}}=45R_{\mathrm{g}}$, and outer radii of reflection region $R_{\mathrm{out, s}}=R_{\mathrm{out, h}}=1000R_{\mathrm{g}}$. The lower subscript `$\mathrm{s}$' (`$\mathrm{h}$') denotes the soft (hard) Comptonisation or its associated reflection component. Fixed parameters related to variability are the same as in Table~\ref{tab:model_parameter_values} in addition to the extra parameters about reverberation, $t_{0, \mathrm{s}}=t_{0, \mathrm{h}}=6\,\si{ms}$ and $\Delta t_{0, \mathrm{s}}=\Delta t_{0, \mathrm{h}}=10\,\si{ms}$. } \begin{tabular}{llll} \hline Component&Model&Symbol&Value\\ \hline \multicolumn{4}{c}{Spectral parameters}\\ \hline Soft Comptonisation&{\tt nthcomp} &$\Gamma_{\mathrm{s}}$ &$1.81$\\ and reflection & &$\mathrm{norm}_{\mathrm{s}}$ &$1.38$\\ &{\tt relxillCp} &$\log_{10} \xi _{\mathrm{s}}$ &$3.44$\\ & &$\mathrm{norm}^{(\mathrm{r})}_{\mathrm{s}}$&$0.0395$\\ Hard Comptonisation&{\tt nthcomp} &$\Gamma_{\mathrm{h}}$ &$1.50$\\ and reflection & &$\mathrm{norm}_{\mathrm{h}}$ &$2.11$\\ &{\tt relxillCp} &$R_{\mathrm{in, h}}$ &$78$\\ & &$\log_{10} \xi _{\mathrm{h}}$ &$1.70$\\ & &$\mathrm{norm}^{(\mathrm{r})}_{\mathrm{h}}$&$0.0328$\\ \hline \multicolumn{4}{c}{Variability parameters}\\ \hline Broad-band &{\tt our model} &$B_{\mathrm{f}}$ &$862$\\ & &$m_{\mathrm{f}}$ &$2.81$\\ & &$B^{(\mathrm{p})}_{\mathrm{f}}$ &$189$\\ & &$\eta_{\mathrm{0, s}}$ &$1.023$\\ & &$\eta_{\mathrm{1, s}}$ &$-0.568$\\ & &$\eta_{\mathrm{0, h}}$ &$1.527$\\ & &$\eta_{\mathrm{1, h}}$ &$-0.580$\\ \hline & &$\chi ^2/\mathrm{d.o.f.}$ &$1993.1/1795$\\ \hline \end{tabular}\\ \label{tab:spec_timing_parameter_values} \end{table} We come back to a spectral-timing analysis from a series of timing analyses above. We attempt to fit the energy-dependent timing properties from 2.6--48~keV along with the time-average energy spectrum at the corresponding energy range by minimizing \begin{equation} \begin{split} &\sum_{i} \left( \frac{S_{\mathrm{data}}(E_{i})-S_{\mathrm{model}}(E_{i})}{\Delta S_{\mathrm{data}}(E_{i})} \right) ^2 \\ + &\sum_{j, k} \left( \frac{P_{\mathrm{data}}(E_{j}, f_{k})-P_{\mathrm{model}}(E_{j}, f_{k})}{\Delta P_{\mathrm{data}}(E_{j}, f_{k})} \right) ^2 \\ + &\sum _{\substack{j, k\\(E_{j} \neq E_{\mathrm{r}})}} \left( \frac{\phi_{\mathrm{data}}(E_{\mathrm{r}}, E_{j}, f_{k})-\phi_{\mathrm{model}}(E_{\mathrm{r}}, E_{j}, f_{k})}{\Delta \phi_{\mathrm{data}}(E_{\mathrm{r}}, E_{\mathrm{j}}, f_{k})} \right) ^2 , \\ \end{split} \end{equation} where $S_{\mathrm{data}}(E)$, $\Delta S_{\mathrm{data}}(E)$ are the observed time-averaged spectrum and its one-sigma error, $S_{\mathrm{model}}(E)$ the modelled time-averaged spectrum, and $E_{i}$ each energy bin in the time-averaged spectrum. We remove clear calibration features seen in the ME spectrum for 20--24 keV (light green regions in Fig.~\ref{fig:observation_data} (top)) from the spectral modelling. To model the energy spectrum, we account for not only the soft and hard Comptonisation components but their disc reflection. We ignore emission from the turbulent disc due to its negligible contribution above the lowest energy of 2.6~keV (few \% at 2.6~keV in the spectral fit found in K22). We also ignore the negligible effect of galactic absorption. We use the {\tt XSPEC} model {\tt nthcomp} (\citealt{Zdziarski_1996, Zycki_1999}) for the Comptonisation components, and {\tt relxillCp} provided in {\tt relxill} version 2.0 (\citealt{Garcia_2014, Dauser_2022}) for the reflected components. Finally, we use \begin{equation} ({\tt nthcomp}+{\tt relxillCp}) + ({\tt nthcomp}+{\tt relxillCp}), \end{equation} where each bracket corresponds to the soft Comptonisation/reflection and hard Comptonisation/reflection, respectively. To connect the time-averaged spectrum and variability consistently, we take reverberation into account in our timing model. Its implementation is updated from that in K22 mainly due to the inclusion of spectral pivoting. We summarize how the reverberation behaves in our new model here, while the detailed formalism is described in Appendix~\ref{sec:app_reflection}. The illuminating Comptonisation spectrum changing its shape with time results in the reflected spectrum also changing its shape with time. As in the previous section, we consider a certain radius. Along with the mass accretion rate and direct emission, we account for the reflected emission associated with the direct emission. Defining the average and difference from it as $S^{(\mathrm{r})}_{0}(E)$ and $\Delta S^{(\mathrm{r})}(E, t)$, we assume $\Delta S^{(\mathrm{r})}(E, t)/S^{(\mathrm{r})}_{0}(E) = (\Delta S(E, t)/S_{0}(E))\otimes h(t)=\eta (E) (\Delta \dot{m}_{0}(t)/\dot{m}_{0})\otimes h(t)$. We use the upper script `${(\mathrm{r})}$' to stand for the reflected emission. The convolution in time is denoted by $\otimes$, and $h(t)$ is called the impulse response, which is the time evolution of reflected emission for a flash of illumination. All information as to the disc response, such as the delay for the direct emission due to an additional light crossing path and the duration due to the different delay times for different locations of reflection, are encoded in $h(t)$. The relation of spectral variation between the direct and reflected emission means that the reflected emission follows variations of the direct emission at the corresponding energy with some time delay, as long as the variability is slow enough not to be washed out via reprocessing, i.e., via the operation of the convolution. In the simple case of $h(t)=\delta (t-\tau)$, variations of the reflected emission exactly lag behind those of the direct emission with the time delay of $\tau$: $\Delta S^{(\mathrm{r})}(E, t)/S^{(\mathrm{r})}_{0}(E)=\Delta S(E, t-\tau )/S_{0}(E)$. Each reflected component has its own impulse response. We define the impulse response with a top-hat function: \begin{equation} h_{\mathrm{Y}} (t)= \begin{cases} 1/\Delta t_{0, \mathrm{Y}} & (|t-t_{0, \mathrm{Y}}|< \Delta t_{0, \mathrm{Y}}/2) ,\\ 0 & (\mathrm{othewise}), \end{cases} \end{equation} where $\mathrm{Y}={\mathrm{s, h}}$ are associated with the soft and hard Comptonisation components, respectively. The parameters $t_{0, \mathrm{Y}}$ and $\Delta t_{0, \mathrm{Y}}$ characterize the delay and duration, respectively. More realistic impulse responses are required, especially for low energy bands $E\lesssim 2\,\si{keV}$, where the quasi-thermal emission due to the reprocessing dominates high-frequency variability ($\gtrsim 1\,\si{Hz}$). However, the top-hat function appears to be a good approximation for high energies, where Comptonisation largely determines variability properties. For the consistency between the spectral modelling and variability modelling, we calculate the fractional time-averaged spectra required in our timing model, $S_{\mathrm{d}}(E)(=0)$, $S_{\mathrm{s}}(E)$, $S_{\mathrm{h}}(E)$, $S^{(\mathrm{r})}_{\mathrm{s}}(E)$, $S^{(\mathrm{r})}_{\mathrm{h}}(E)$, from the spectral models, {\tt nthcomp} and {\tt relxillCp}. We model the sensitivity parameter phenomenologically with \begin{equation} \eta _{\mathrm{Y}}(E)=\eta_{0, \mathrm{Y}} + \eta_{1, \mathrm{Y}} \log _{10} \left( E\,[\si{keV}] \right), \label{eq:eta_phenomenology} \end{equation} for $\mathrm{Y}={\mathrm{s, h}}$ (we do not need $\eta_{\mathrm{d}}(E)$ due to $S_{\mathrm{d}}(E)=0$). The sensitivity parameter is assumed to change with energy logarithmically. We note the difference in the model calculations between the timing fits (Section~\ref{sec:rev2}) and spectral-timing fits. In the timing fits, $\eta(E) S_{0}(E)$ is a model parameter, and it is impossible to disentangle this product. On the other hand, $S_{0}(E)$ and $\eta (E)$ are separately modelled in the spectral-timing fits. The former is calculated from spectral models, the latter is from equation (\ref{eq:eta_phenomenology}). In the joint spectral-timing fit, we fix the seed photon temperature of Comptonisation components to the typical disc temperature in this state, $kT_{\mathrm{seed, s}}=kT_{\mathrm{seed, h}}=0.2\,\si{keV}$ (\citealt{DeMarco_2021}, K22). Since the electron temperature is difficult to constrain from the energy band of interest, we fix it to $kT_{\mathrm{e, s}}=kT_{\mathrm{e, h}}=23\,\si{keV}$, as in K22. While we allow the inner radius of the reflector for the hard Comptonisation component to be free, we fix that for the soft Comptonisation component to $R_{\mathrm{in, s}}=45R_{\mathrm{g}}$ corresponding to the outer edge of the variable flow located at $r_{\mathrm{out}}=45$. Following K22, we fix the inclination angle to $i=66^{\circ}$ (\citealt{Torres_2020}) and Fe abundance to $Z_{\mathrm{Fe}}=1.1$. We also set the black hole spin to $a^{*}=0$, consistent with $r_{\mathrm{in}}=6$ in the timing model, and use the high electron density of $N_{\mathrm{e}}=10^{20}\,\si{cm^{-3}}$ (\citealt{Garcia_2016, Mastroserio_2021}). The delay and duration of the impulse response should be, in principle, calculated from the location and geometry of illuminating source and reflector. However, it is computationally expensive to get them in fitting and also difficult to constrain them. Thus, we simply fix $t_{0, \mathrm{s}}=t_{0, \mathrm{h}}=6\,\si{ms}$ and $\Delta t_{0, \mathrm{s}}=\Delta t_{0, \mathrm{h}}=10\,\si{ms}$. The top-hat impulse response with these values appears to be good approximations of more realistic ones (K22). The results of simultaneous modelling of the energy spectrum, six power spectra, and five phase-lag spectra are shown in Fig.~\ref{fig:spec_psd_phase_fit}. The comparison between the data and model is also plotted as the ratio for the energy spectrum and the difference divided by one-sigma errors for the variability. Model parameter values are found in Table~\ref{tab:spec_timing_parameter_values}. Overall, our new model successfully reproduces both time-averaged and variability properties, although the discrepancies are seen in the phase-lag spectrum between 35--48~keV and 2.6--4.8~keV (magenta), which is discussed in Section~\ref{sec:discussion_limitations}. This modelling is the first simultaneous fit to spectrum and variability using our model. The uncertainties of the derived parameter values are evaluated with an MCMC analysis in Appendix~\ref{sec:app_mcmc}. The spectral variation derived from the fit is shown in Fig.~\ref{fig:spectral_pivoting_fit_reflection}. The spectra for the mass accretion rate being its average and double the average are plotted with solid and dashed lines. For illustration purposes, we ignore all effects from the impulse response for reverberation, such as time delay, i.e., we assume $h(t)=\delta (t)$. This means that the Comptonisation and its associated reflection behave completely in the same way, $\Delta S(E, t)/S_{0}(E)=\Delta S^{(\mathrm{r})}(E, t)/S^{(\mathrm{r})}_{0}(E)$. Generally, all spectra are less sensitive for higher energies to mass accretion rate fluctuations, which results in a decrease in power spectrum with energy. We see the pivoting point at $\sim 50\,\si{keV}$ for the soft Comptonisation and its reflection, which roughly agrees with that at $\sim 35\,\si{keV}$ derived from the fit only to the timing properties in the previous section. \section{Discussion} \label{sec:discussion} \begin{figure} \includegraphics[width=\linewidth]{figure/fprop_comp.pdf} \caption{Propagation frequency (solid) as a function of radius derived from the spectral-timing fit. Red, green, and blue colours denote the variable disc, soft Comptonisation, and hard Comptonisation regions, respectively. The propagation frequency derived from our previous work and predicted from theoretical models are plotted with a dashed line and dash-dotted lines (see K22 for details). The black hole mass of $M_{\mathrm{BH}}=8M_{\odot}$ is assumed. The Keplerian frequency is also plotted with a grey dash-dotted line for reference. } \label{fig:fvisc_updates} \end{figure} \begin{figure} \includegraphics[width=\columnwidth]{figure/phase_extrapolation.pdf} \caption{Comparison of phase-lag spectra at high energies between the model extrapolation and observation. The reference band is 2.6--5.8~keV. The phase-lag spectrum for 35--48~keV, the highest energy band in the spectral-timing fitting, is plotted for reference. The observation shows a clear increase in phase lag with energy, while the phase lags calculated by our new model are rather insensitive to energy beyond 40~keV, pointing out the limitation of the model. } \label{fig:phase_extrapolation} \end{figure} \subsection{Generator time-scale and propagation time-scale} The characteristic time-scales on which the fluctuations are propagated at each radius are derived from the spectral-timing fit (Fig.~\ref{fig:spec_psd_phase_fit}). We compare the propagation frequency derived (solid) to those predicted by different hot flow models (dash-dotted) in Fig.~\ref{fig:fvisc_updates}. The theoretical propagation frequencies for the Advection Dominated Accretion Flow (ADAF; \citealt{Narayan_1997}), Standard And Normal Evolution accretion flow (SANE; \citealt{Narayan_2012}), Magnetically Arrested Disc (MAD; \citealt{Narayan_2012}), and Jet Emitting Disc (JAD; \citealt{Marcel_2018}) are calculated in a standard way by dividing the accretion speed by radius for the assumption of $M_{\mathrm{BH}}=8M_{\odot}$ (see K22 for details). The propagation frequency derived from our previous model from K22 is also plotted (dashed). In K22, we found fairly good agreement of the derived propagation time-scales with those in ADAF, SANE, and JED rather than MAD. The propagation time-scale derived from our new model is now not very similar to any theoretical predictions. Here, allowing the generator and propagation frequencies to be different makes their radial dependence steeper, i.e., from $f_{\mathrm{prop}}(r)\propto r^{-2.7}$ to $f_{\mathrm{prop}}(r)\propto r^{-4.22}$. This steep radial dependence is required to reproduce both the observed large phase-lags and high-frequency broad-band variability simultaneously. Indeed, the propagation time-scale derived is robust against uncertainties of the relationship between generator time-scale and propagation time-scale. Using our new model including the spectral pivoting, the assumption of $f_{\mathrm{gen}}(r)=f_{\mathrm{prop}}(r)$ also gives the similar propagation time-scale, although the fit is not as good as that obtained in the previous section with $f_{\mathrm{gen}}(r)\neq f_{\mathrm{prop}}(r)$. The key feature that our new model requires is a slow propagation speed enough to reproduce observed phase lags. The propagation time-scale of MAD (brown dash-dotted line) is too short to explain the observed phase lags. Thus, our results still prefer SANE rather than MAD, although MAXI~J1820+070 displays a powerful jet (\citealt{Bright_2020}). \subsection{Origin of QPOs} Our full spectral-timing modelling accounts for all the X-ray spectrum and rapid variability except for timing features of the QPOs. We account for the timing features of the QPOs by adding peaked Lorentzian components in the power spectra. No extra component is added to the phase-lag spectra simply because QPO features are not very clear across 2.6--48~keV. We did not add any other spectral components for the QPOs, assuming that the QPO is a modulation of the spectral components included in the model (multiplicative) rather than being associated with an additional spectral component (additive). Our successful modelling, in which there is no room for an additional emission component only related to the QPOs (Fig.~\ref{fig:spec_psd_phase_fit}), supports the assumption made. This result is consistent with a QPO produced predominantly from a global mode of the flow rather than an intrinsic change in intensity with QPO frequency. We specifically have in mind Lense-Thirring (vertical) precession of the entire hot flow, where the observed luminosity of the Compton component(s), including all their stochastic variability, are modulated by the changing projected area of the translucent hot flow as the viewing angle changes with QPO phase (\citealt{Fragile_2007, Ingram_2009, Ingram_2011, Ingram_2012}). This picture agrees with the new polarisation results for Cyg~X-1 in the low/hard state, which require the hot X-ray emitting plasma to be radially extended (\citealt{Krawczynski_2022}). The alternative precessing jet model suggested by \cite{Ma_2021} is, on the other hand, challenged by the above picture because it requires the hot X-ray emitting region to align with the jet. \subsection{Limitations of our new model} \label{sec:discussion_limitations} Our new model gives a poor fit to both the energy spectrum and phase lags beyond $\sim 40$--$50$~keV. From Fig.~\ref{fig:spec_psd_phase_fit} (left), the spectral model clearly underestimates the data above $\sim$ 100~keV. The hard Comptonisation spectrum rolls over too fast to match the observed data. The comparison of the phase-lag spectrum between the model and data at high energies are shown in Fig.~\ref{fig:phase_extrapolation}. The model phase lags increase with energy to $\sim 40\,\si{keV}$ (Fig.~\ref{fig:spec_psd_phase_fit} (right)) but then saturates to a constant value rather than continually increasing as in the data. The lag behaviour arises as the fraction of the total spectrum which is made of the hard (and long lagged) Comptonisation spectrum increases up to around $\sim 40\,\si{keV}$, but after this point, the hard Comptonisation dominates, leading to the saturation of lag. The spectral mismatch could be fixed if there is additional stratification of the energy spectrum of the hot flow, so the very innermost regions produce an even hotter/harder Comptonisation component. In many ways, this is quite natural. The two Comptonisation components used here for the spectral decomposition could be an approximation to a continuous flow with (presumably) continuous stratification. There should physically be two main regions. Close to the disc, seed photons for Comptonisation are predominantly from the disc. However, it is quite easy for this Comptonisation to become optically thick along the equatorial direction, shielding the inner regions from the disc photons so that seed photons are predominantly from cyclo-synchrotron (\citealt{Poutanen_2014}). Nonetheless, there could still be some radial temperature/spectral hardness gradients in this second region which could produce additional emissions at the hardest energies (\citealt{Poutanen_2014}). We note that the JED models (e.g. \citealt{Marcel_2018}) also predict a continuously increasing temperature/harder spectrum with radius in their hot flow. However, including the additional harder Comptonisation component probably does not fully solve the deviation of the phase lag. Although this will give a qualitatively increasing lag, the amount of the increased lag should be rather small as the propagation speed is already high. Yet the data show a large increase in lag between high energy bands. It seems more likely that there are other factors affecting lags at work. There is another feature which is lacking from our new model. It is a physical description of the spectral pivoting from the Comptonisation process. Currently, the model assumes that the spectra pivot in a synchronous way, i.e., the local spectrum at every energy responds to fluctuations of the local mass accretion rate simultaneously. Although the magnitude of the response can be different between different energies, as seen in Fig.~\ref{fig:spectral_pivoting_fit_reflection}, there is no causal connection between them. There can be 0 or $\pi$ of phase lags arising from the spectral pivoting itself (the phase lag of $\pi$ arises if one energy band is above the pivoting point and the other band is below it, i.e., $\eta(E_1)\eta(E_2)<0$). Thus, in our new model, the lags between different energies are still due to the propagating fluctuation process, as is the case for the model developed by \cite{Veledina_2016, Veledina_2018}. The spectral pivoting implemented in our new model can strongly affect power spectra but has only a relatively mild effect on phase-lag spectra. Indeed, our new model is able to reproduce energy-dependent power spectra fairly well up to $\sim 100\,\si{keV}$, even though it fails to match to the phase-lag spectra. However, the physical picture of Comptonisation described above yields a characteristic spectral pivoting pattern having a causal connection. A fluctuation from the edge of the truncated disc first gives a change in seed photons to the soft Comptonisation. Assuming an increase in seed photons, it increases the Compton cooling on the light travel time without any change in electron heating, so the spectrum softens. Then, after the accretion time-scale (propagation time-scale), the same fluctuation modulates the soft Comptonisation by increasing the electron density, increasing the heating rate, and causing the spectrum to harden. This gives an asynchronous rocking of the soft Comptonisation, where two mutually-correlated but lagged variability sources form its time-dependent behaviour. By contrast, in the hard Comptonisation region, the fluctuation gives a synchronous change in seed photons and electron heating as both are produced together around its outer edge. The synchronous pivoting implemented in our model is limiting the ability to properly model the data, as it is suppressing a real lag which occurs from the two time-scale propagation mechanism in the soft Comptonisation. It is worth noting that our implementation of the spectral pivoting is different from that in the {\tt RELTRANS} model (e.g. \citealt{Mastroserio_2018, Ingram_2019b}). \cite{Mastroserio_2018, Mastroserio_2019, Mastroserio_2021} consider the nonlinear effects in the time-varying continuum spectrum and have two variable terms in its expression to allow lags to arise from the spectral pivoting itself (\citealt{Kotov_2001}). Not specifying the underlying process causing the spectral pivoting may make {\tt RELTRANS} too flexible in producing the observed hard-lag data. On the other hand, our new model has only one variable term, i.e., the local mass accretion rate, in the expression of the local spectrum. The local spectrum varies linearly to this term, which does not produce lags except for $\pi$. As mentioned above, our model relies on the hard lags caused by the combination of the propagating fluctuations process and energy-dependent emission profile (\citealt{Veledina_2016, Veledina_2018}). We suspect that this lack of a physical spectral pivoting model, including the light crossing time spectral softening as well as the propagation time, is the major reason our new model fails to fit the phase-lag spectra from 50--150~keV. This more physical model for spectral pivoting is beyond the scope of this paper but will be considered in future work. \section{Conclusions} \label{sec:conclusions} We have studied X-ray spectral-timing properties of the black hole binary MAXI J1820+070 in the bright low/hard state using {\it Insight-HXMT} observation data. Particularly, we have focused on the energy-dependent broad-band variability on time-scales from milliseconds to seconds. We started with testing our previous model from K22, which included the propagating fluctuations process and reverberation and successfully explained soft X-ray timing properties ($< 10\,\si{keV}$), and found that it cannot be applied to higher energy bands. The key variability feature that our previous model missed was the decrease in fractional power spectrum with energy above $\sim 10\,\si{keV}$, which was difficult to explain with the simple propagating fluctuations picture but typically observed (\citealt{Nowak_1999, Malzac_2003, Axelsson_2018}). We have seen that additional effects proposed in the literature, such as the damping (\citealt{Rapisarda_2017a, Mahmoud_2018b}), are not very effective in reproducing both observed power spectra and phase-lag spectrum simultaneously. We have updated our model by implementing the physically natural picture, the spectral pivoting (\citealt{Malzac_2003, Gandhi_2008, Veledina_2016, Mastroserio_2018}), and shown that the new model succeeded in reproducing both power spectra and phase-lag spectra across the 2.6--48~keV band. The energy spectra being less sensitive to mass accretion rate fluctuations allow the model to suppress variability at high energies. In addition, we have successfully performed a joint spectral-timing fit, demonstrating that our timing model self-consistently combined with spectral models helps to break the spectral degeneracy. In modelling the time-averaged spectrum and broad-band variability, we have considered emission from the turbulent disc and soft and hard Comptonisation regions and their associated disc reflection, all of which constitute the accretion flow. Because the energy spectrum is filled up by the emission above, the emission of QPOs also present in the observation is likely to be produced from the considered spectral components, supporting that the QPOs originate from the accretion flow such as Lense-Thirring precession (\citealt{Ingram_2009}) rather than the jet (\citealt{Ma_2021}). However, our new model still has limitations in that there are clear discrepancies between the model and observation in phase-lag spectra above $\sim 40\,\si{keV}$. The model saturates the phase lag around $\sim 40\,\si{keV}$, while those observed keep increasing up to $\sim 150\,\si{keV}$. We suggest that additional harder spectral component(s) at the innermost region, the lags intrinsically arising from the spectral pivoting, or both could solve these discrepancies. \section*{Acknowledgements} This work was supported by JSPS KAKENHI Grant Numbers 18H05463 and 20H00153, and WPI MEXT. TK acknowledges support from JST SPRING, Grant Number JPMJFS2108. CD acknowledges support from the STFC consolidated grant ST/T000244/1. We thank Zi-Xu Yang, Liang Zhang, and Xiang Ma for help with {\it Insight-HXMT} data reduction. We dedicate this work to the memory of Magnus Axelsson, who died suddenly as we were finishing the paper. \section*{Data Availability} The observational data underlying this article are available at \url{http://hxmten.ihep.ac.cn/}. The model will be shared upon reasonable request to the corresponding author. \bibliographystyle{mnras}

7ff64a0958b7420a6e1c4cbcef27adf6d595e8d5

f080446304324098b74b9addc0a16960d33e1b84

\section{Introduction}\label{sec:intro} Nonconvex optimization plays a central role in machine learning because the training of many modern machine learning models, especially those from deep learning, requires optimization of nonconvex loss functions. Among algorithms for solving nonconvex optimization problems, stochastic gradient descent (SGD) and its variants, such as Adam~\cite{ADAM}, Adagrad~\cite{duchi2011adagrad}, etc., are widely used in practice. In theory, their provable guarantee has been studied from various perspectives. In this paper, we adopt the perspective of studying gradient descents via the analysis of their behavior in continuous-time limits as differential equations, following a recent line of work in~\cite{su2016differential,wibisono2016variational,jordan2018dynamical,shi2021understanding}. In particular, given a function $f\colon\mathbb{R}^{d}\to\mathbb{R}$, the SGD $x_{k+1} = x_k - s\nabla f(x_k) - s\xi_{k}$ with learning rate $s$ and the $k$th step noise $\xi_{k}$ can be approximated by a stochastic differential equation (SDE) as follows: \begin{align}\label{eq:SDE} \d x = -\nabla f(x)\d t + \sqrt{s}\d W, \end{align} where $W$ is a standard Brownian motion. Such approach enjoys clear intuition from physics. In particular, Eq.~\eq{SDE} is essentially a non-equilibrium thermodynamic process: gradient descent provides driving forces, the stochastic term serves as thermal motions, and a combination of these two ingredients enables convergence to the thermal distribution, also known as the Gibbs distribution. A systematic study of Eq.~\eq{SDE} was conducted in a recent work by~\cite{SSJ20}. See more details in \sec{classicalpre}. Nevertheless, algorithms based on gradient descents also have limitations because they only have access to local information about the function, which suffers from fundamental difficulties when facing landscapes with intricate local structures such as vanishing gradient~\cite{hochreiter1998vanishing}, nonsmoothness~\cite{KS21}, negative curvature~\cite{CB21}, etc. In terms of optimization, we are mostly interested in points with zero gradients, and they can be categorized as \emph{saddle points}, \emph{local optima}, and \emph{global optima}. It is known that variants of SGD can escape from saddle points \cite{ge2015escaping,jin2017escape,allen2018neon2,fang2018spider,fang2019sharp,jin2019stochastic,zhang2021escape}, but one of the most prominent issues in nonconvex optimization is to escape from local minima and reach global minima. Up to now, theoretical guarantee of escaping from local minima by SGD has only been known for some special nonconvex functions~\cite{kleinberg2018alternative}. In general, SGD has to climb through high barriers in landscapes to reach global minima, and this is typically intractable using only gradients that descend the function. In all, fundamentally different ideas, especially those that explores beyond local information, are expected to derive better algorithms for nonconvex optimization in general. This paper aims to study nonconvex optimization via dynamics from \emph{quantum mechanics}, which can leverage global information about a function $f\colon\mathbb{R}^{d}\to\mathbb{R}$. The fundamental rule in quantum mechanics is the \emph{Schr{\"o}dinger Equation:}\footnote{The standard Schr{\"o}dinger Equation in quantum mechanics is typically written as $i \hbar\frac{\partial}{\partial t} \Phi = \left(-\frac{\hbar^2}{2m}\Delta + f(x)\right)\Phi$. In this paper, we use the form in \eq{Schrodingereq} by setting the Planck constant $\hbar = 1$ and $h = \hbar/\sqrt{2m}$ which is a variable. See also \sec{quantumpre}.} \begin{align} i \frac{\partial}{\partial t} \Phi = \left(-h^2\Delta + f(x)\right) \Phi, \label{eq:Schrodingereq} \end{align} where $i$ is the imaginary unit, $h$ is defined as the \emph{quantum learning rate}, $\Delta=\sum_{j=1}^{d}\frac{\partial^2}{\partial x_{j}^2}$ is the Laplacian, and $\Phi(t,x)\colon\mathbb{R}\times\mathbb{R}^{d}\to\mathbb{C}$ is a quantum wave function satisfying $\int_{\mathbb{R}^{d}}|\Phi(t,x)|^2\d x=1$ for any $t$. Measuring the wave function at time $t$, $|\Phi(t,x)|^2$ is the probability density of finding the particle at position $x$. In Eq.~\eq{Schrodingereq}, the time evolution of wave functions is governed by the Hamiltonian\footnote{In this paper, we refer Hamiltonian to either the total energy of a system or the operator corresponding to the total energy of the system, depending on the context.} $H:=-h^2\Delta + f$, where $-h^2\Delta$ corresponds to the classical kinetic energy and $f$ the potential energy. In sharp contrast to classical particles, quantum wave functions can tunnel through high potential barriers with significant probability, and this is formally known as \emph{quantum tunneling.} Take a one-dimensional double-well potential $f\colon\mathbb{R}\to\mathbb{R}$ in \fig{1-dimexample} as an example, the goal is to move from the local minimum $x_{-}$ in the left region to the local minimum $x_{+}$ in the right region. Classically, the SDE in \eq{SDE} has to climb through the barrier with height $H_{f}$, and it can take $\exp(\Theta(H_{f}/s))$ time to reach $x_{+}$ (see Section 3.4 of \cite{SSJ20}). Quantumly, we denote $\Phi_-(x)$ and $\Phi_+(x)$ to be the ground state (i.e., the eigenstate corresponds to the smallest eigenvalue) of the left and the right region, respectively. These states $\Phi_\pm(x)$ are localized near $x_{\pm}$, respectively. We let the wave function be initialized at $\Phi_-(x)$, i.e., $\Phi(0,x)=\Phi_-(x)$. Under proper conditions, two eigenfunctions with eigenvalues $E_0$ and $E_1$ of $H=-h^2\Delta + f$ can be represented by superposition states \begin{align} \Phi_0(x):=(\Phi_+(x) + \Phi_-(x))/\sqrt{2},\\ \Phi_1(x):=(\Phi_+(x) - \Phi_-(x))/\sqrt{2}, \end{align} respectively. Note that $\Phi_0(x)$ and $\Phi_1(x)$ are \emph{not} localized because they have probability $1/2$ of reaching both $x_{+}$ and $x_{-}$. Specifically, given $\Phi(0,x) = \Phi_-(x) = (\Phi_0(x) + \Phi_1(x))/\sqrt{2}$, and because the dynamics of the Schr\"odinger equation \eq{Schrodingereq} is $\Phi(t,x) = e^{-iHt}\Phi(0,x)$, we have \begin{align} \Phi(t,x) = (e^{-iE_0t}\Phi_0(x) + e^{-iE_1t}\Phi_1(x))/\sqrt{2}. \end{align} As a result, after time $t$ where $|E_0-E_1|t =\pi$, we have $\Phi(t,x)\propto \Phi_+(x)$ localized near $x_{+}$. Intuitively, this can be viewed as global evolution and superposition of quantum states, which is capable of acquiring global information of the function $f$ and explains why for various choices of $f$, this quantum evolution time $t$ is much shorter than the classical counterpart by SDE which only takes gradients locally. \begin{figure} \centering \includegraphics[width=0.5\textwidth]{1-dimexample.pdf} \caption{An example of quantum tunneling in a double-well optimization function.} \label{fig:1-dimexample} \end{figure} It is a natural intuition to design quantum algorithms using quantum tunneling. Previously, \cite{FGS94,MAL16,CH16,BZ18} studied the phenomenon of quantum tunneling in quantum annealing algorithms \cite{FGS94,FGG01}. However, most of these results studied Boolean functions, which is essentially different from continuous optimization. In addition, quantum annealing focused on ground state preparation instead of the dynamics for quantum tunneling. Up to now, it is in general unclear when we can design quantum algorithms for optimization by adopting quantum tunneling. Therefore, we ask: \begin{question}\label{ques:configuration} On what kind of landscapes can we design algorithms efficiently using quantum tunneling? \end{question} To answer this question, we need to figure out specifications of the quantum algorithm, such as the initialization of the quantum wave packet, the landscape's parameters, the measurement strategy, etc. The next question is to understand the advantage of quantum algorithms based on quantum tunneling. A main reason of studying quantum computing is because it can solve various problems with significant speedup compared to classical state-of-the-art algorithms. In optimization, prior quantum algorithms have been devoted to semidefinite programs~\cite{brandao2016quantum,vanApeldoorn2017quantum,vanApeldoorn2018SDP,brandao2017SDP}, convex optimization~\cite{vanApeldoorn2020optimization,chakrabarti2020optimization}, escaping from saddle points~\cite{zhang2021quantum}, polynomial optimization~\cite{rebentrost2019quantum,li2021optimizing}, finding negative curvature directions~\cite{zhang2019quantum}, etc., but quantum algorithms for nonconvex optimization with provable guarantee in general is widely open as far as we know. Here we ask: \begin{question}\label{ques:accelaration} When do algorithms based on quantum tunneling give rise to quantum speedups? \end{question} \paragraph{Contributions.} We systematically study quantum algorithms based on quantum tunneling for a wide range of nonconvex optimization problems. Throughout the paper, we consider benign nonconvex landscapes where \emph{local minima are (approximately) global minima.} We point out that many common nonconvex optimization problems indeed yield objective functions satisfying such benign behaviors, such as tensor decomposition~\cite{ge2015escaping,GM20}, matrix completion~\cite{ge2016matrix,ma2018implicit}, and dictionary learning~\cite{qu2019analysis}, etc. In general, nonconvex problems with discrete symmetry satisfy this assumption, see the surveys by~\cite{Ma21,ZQW21}. In this paper, we demonstrate the power of quantum computing for the following main problem: \begin{named}{Main Problem}\label{prb:main} On a landscape whose local minima are (approximately) global minima, starting from one local minimum, find all local minima with similar function values or find a certain target minimum. \end{named} Such a problem is crucial for understanding the \emph{generalization} property of nonconvex landscapes, and in general it also sheds light on nonconvex optimization. First, local minima with similar function values can have dramatically different generalization performance (see Section 6.2.3 of \cite{Sun19}), and solving this \ref{prb:main} can be viewed as a subsequent step of optimization for finding the minimum which generalizes the best. Second, \ref{prb:main} implies the mode connectivity of landscapes, which has been applied to understanding the loss surfaces of various machine learning models including neural networks both empirically~\cite{draxler2018essentially,garipov2018loss} and theoretically~\cite{kuditipudi2019explaining,nguyen2019connected,shevchenko2020landscape}. Third, nonconvex landscapes where the \ref{prb:main} can be efficiently solved can also lead to efficient Monte Carlo sampling, which can be even faster than optimization~\cite{ma2019sampling,talwar2019computational}. Landscapes whose local minima are (approximately) global significantly facilitate quantum tunneling. Roughly speaking, since the total energy during our quantum evolution \eq{Schrodingereq} is conserved, quantum tunneling can only efficiently send a state from one minimum to another minimum with similar values. As a conclusion, if the quantum wave function is initialized near a local minimum, we can focus on quantum tunneling between the local ground state of each well, i.e., the tunneling of the particle from the bottom of a well to that of another well. To avoid complicated discussions on the value of the quantum learning rate $h$, we further restrict ourselves to functions whose local minima are global, which would not provide less intuition. Now, an answer to \ques{configuration} can be given as follows: \begin{theorem}[Quantum tunneling walks, informal]\label{thm:informalQTW} On landscapes whose local minima are global minima, we have an algorithm called quantum tunneling walks (QTW) which initiates the simulation of Eq.~\eq{Schrodingereq} from the local ground state at a minimum, and measures the position at a time which is chosen uniformly from $[0,\tau]$. To solve the \ref{prb:main} we can take \begin{align}\label{eq:tau-informal} \tau=O(\mathrm{poly}(N)/\Delta E), \end{align} where $N$ is the number of global minima and $\Delta E$ is the minimal spectral gap of the Hamiltonian restricted in a low-energy subspace. For sufficiently small $h$, we have \begin{align}\label{eq:QTW} \Delta E = \sqrt{h}(b + O(h))e^{-\frac{S_0}{h}}, \end{align} where $b, S_0>0$ are constants that depend only on $f$. \end{theorem} Formal description of the QTW can be found in \sec{QTW}. Here we highlight two important properties of QTW: Quantum mixing time and quantum hitting time. \\\\ \emph{Quantum mixing time (\lem{qtwmixingtime} in \sec{Qmixing}).} Since quantum evolutions are unitary, QTW never converges, a fundamental distinction from SGD. Therefore, to study the mixing properties of QTW, we follow quantum walk literature~\cite{CCD+03} by employing the measurement strategy, where we measure at $t$ uniformly chosen from $[0,\tau]$. The measured results obey a distribution which is a function of $\tau$, and when $\tau\to +\infty$, the distribution tends to its limit, $\mu_{\rm QTW}$. Quantum mixing time is the minimal $\tau$ enabling us to sample from $\mu_{\rm QTW}$ up to some small error. Alternatively speaking, the mixing time evaluates how fast the distribution yielded by QTW converges. We prove that $\mu_{\rm QTW}$ concentrates near minima, so that sampling from $\mu_{\rm QTW}$ repeatedly can give positions of all minima. In addition, $\mu_{\rm QTW}$ gives the upper bound on $\tau$ in \eq{tau-informal}. \\\\ \emph{Quantum hitting time (\lem{qtwhitting} in \sec{Qhitting})}. Hitting time is the duration it takes to hit a target region (usually a neighborhood of some minimum). Quantum hitting time is the minimum evolution time needed for hitting the region of interest once. Despite this straightforward intuition, the formal definition of quantum hitting time is very different from that of classical hitting time. Intuitively, repeatedly sampling from $\mu_{\rm QTW}$ can ensure the hitting of neighborhoods of particular minima, and thus we can use the mixing time to bound the hitting time. In short, to solve the \ref{prb:main}, we bound the quantum mixing and hitting time to obtain \thm{informalQTW}. \\\\ The minimal spectral gap $\Delta E$ in \thm{informalQTW} is calculated in \append{A.2.3}. The quantity $S_0$ is called the \emph{minimal Agmon distance} between different wells, formally defined in \defn{Agmon-distance}, which is related to both the height and width of potential barriers. The smaller $h$ is, the closer the measured results are to the minima (i.e., the more accurate QTW is), but the longer evolution time the Schr{\"o}dinger equation takes. As an application of \thm{informalQTW} and a justification of the practicability of QTW, we show how to use QTW to solve the orthogonal tensor decomposition problem. This problem asks to find all orthogonal components of a tensor. After transforming into a single optimization problem \cite{CLX+09,Hyv99}, the aim is to find all global minima. We present below a bound on the time cost of QTW on decomposing fourth-order tensors and details can be found in \sec{tendecom}. \begin{proposition}[Tensor decomposition, informal version of \prop{TenDecTtot}]\label{prop:TenDecTtot-main} Let $d$ be the dimension of the components of the fourth-order tensor $T \in \mathbb{R}^{d^4}$ satisfying \eq{4-tensor}, $\delta$ be the expected risk yielded by the limit distribution $\mu_{\rm QTW}$, and $\epsilon$ be the maximum error between $\mu_{\rm QTW}$ and the actual obtained distribution (quantified by $L^1$ norm). For sufficiently small $\epsilon$ and sufficiently small $\delta$, the total time $T_{\rm tot}$ for finding all orthogonal components of $T$ by QTW satisfies \begin{align} T_{\rm tot} = O(\mathrm{poly}(1/\delta, e^d, 1/\epsilon)) e^{\frac{(d-1) +o_{\delta}(1)}{2\delta}}. \end{align} \end{proposition} Next, we explore the advantages of the quantum tunneling mechanism comparing QTW with SGD and shown by describing landscapes where QTW outperforms SGD. The time cost for SGD to converge to global minima is loosely $O(1/\lambda_{s})$ and \begin{align}\label{eq:SGD} \lambda_{s} = (a + o(s))e^{-\frac{2H_f}{s}} \end{align} by~\cite{SSJ20}. Here, $s$ is the step size or learning rate of SGD. The constants $a>0$ and $H_f>0$ depend only on $f$. Interestingly, running time of QTW and that of SGD have similar form. In \eq{QTW} and \eq{SGD}, there are exponential terms $e^{S_0/h}$ and $e^{2H_f/s}$, respectively. Intuitively, the quantity $H_f$ is the characteristic height of potential barriers, and the quantity $S_0$ depends on not only the height but also the width of potential barriers. For the one-dimensional example in \fig{1-dimexample}, \begin{align} H_{f}=\max_{\xi\in[x_{-},x_{+}]}f(\xi),\qquad S_{0}=\int_{x_{-}}^{x_{+}} \sqrt{f(\xi)} \d\xi. \end{align} (Proof details are given in \sec{onedimexp}.) Other terms in the bounds, $\mathrm{poly}(N)/\Delta E$ and $1/\lambda_{s}$, are referred to as polynomial coefficients. We make the following comparisons: \begin{figure} \centerline{ \includegraphics[width=0.94\textwidth]{abstract1.pdf}} \caption{Flowchart of the \hyperref[slog:1]{Main Message}. } \label{fig:abstract1} \end{figure} \begin{itemize} \item Regarding the exponential terms $S_0$ and $H_f$, tall barriers means that $H_f$ is large, whereas if the barriers are thin enough, $S_0$ can still be small. This is consistent with the long-standing intuition that tall and thin barriers are easy for tunneling but difficult for climbing \cite{CH16}. \item Regarding the polynomial coefficients, they are mainly influenced by the distribution or relative positions of the wells. We observe that a symmetric distribution of wells, which can make (the local ground state in) any one well interacts with (the local ground states in) other wells, may reduce the running time of QTW but has no explicit impact on SGD. \item Flatness of wells is another important factor that influences the running time of both QTW and SGD. We propose standards for comparison (see \sec{standard}), which studies their running time when reaching the same accuracy $\delta$. Same to the effect of $h$, a smaller learning rate $s$ permits more accurate outputs but makes SGD more time consuming. For sufficiently flat minima, $h$ is larger than $s$, leading to a smaller running time for QTW. \end{itemize} In summary, we illustrate above observations in \fig{abstract1} and conclude the following: \begin{mainmessage}[Advantages of the quantum tunneling mechanism, a summary of \sec{standard} and \sec{illustration}]\label{slog:1} \emph{On landscapes whose local minima are global minima, QTW outperforms SGD on solving the \ref{prb:main} if barriers of the landscape $f$ is high but thin, wells are distributed symmetrically, and global minima are flat.} \end{mainmessage} \begin{remark} As is indicated above, we compare the costs of QTW and SGD under the same accuracy $\delta$. We introduce two definitions of accuracy in \sec{standard}: \stand{risk} concerns the expected risk, and \stand{distance} concerns the expected distance to some minima. Mathematically, \stand{risk} and \stand{distance} establish a relationship between the quantum and classical learning rates $h$ and $s$, respectively, enabling direct comparisons. \end{remark} Having introduced the general performance of the quantum tunneling walk, we further investigate \ques{accelaration} on some specific scenarios of the \ref{prb:main}. We focus on comparison between query complexities, namely the classical query complexity to local information and the quantum query complexity to the evaluation oracle\footnote{Query complexity of QTW is directly linked to the evolution time, in particular, evolving QTW for time $t$ needs $\tilde{O}(t)$ queries to $U_f$ (see details in \append{B.1}). As a result, it suffices to analyze the evolution time of QTW. Nevertheless, we state the query complexities for direct comparison.} \begin{align}\label{eqn:quantumquery} \hspace{-1.5mm}U_f (\ket{x}\otimes \ket{z}) = \ket{x}\otimes\ket{f(x)+z}\ \ \forall x\in\mathbb{R}^{d},z\in\mathbb{R}. \end{align} This is the standard assumption in existing literature on quantum optimization algorithms~\cite{vanApeldoorn2020optimization,chakrabarti2020optimization,zhang2021quantum}. Note that the $U_f$ here is a unitary transformation and it allows superposed input states, i.e., for a state $\sum_{j=1}^m c_j \ket{x_j}$ where $m\in \mathbb{N}$, $c_j \in \mathbb{C}$, and $\sum_{j=1}^m|c_j|^2=1$, \begin{align} U_f\Big(\sum_{j=1}^m c_j \ket{x_j} \otimes \ket{0}\Big) = \sum_{j=1}^m c_j \ket{x_j} \otimes \ket{f(x_j)}. \end{align} If we measure this output state, with probability $|c_{j}|^{2}$ we obtain $f(x_{j})$. The distribution can not only be sampled from a discrete set but also a continuous set. Different from classical queries that only learn local information of the landscape of $f$, quantum evaluation queries are essentially nonlocal as they can extract information of $f$ at different locations in superposition. Based on this fundamental difference, we are able to prove that QTW can solve a variant of the \ref{prb:main} with exponentially fewer queries than any classical counterparts: \begin{theorem}\label{thm:provableinformal} For any dimension $d$, there exists a landscape $f\colon \mathbb{R}^d \to \mathbb{R}$ such that its local minima are global minima, and on which, with high probability, QTW can hit the neighborhood of an unknown global minimum from the local ground state associated to a known minimum using queries polynomial in $d$, while no classical algorithm knowing the same minimum can hit the same target region with queries subexponential in $d$. \end{theorem} \begin{figure} \centerline{ \includegraphics[width=0.83\textwidth]{provable-acceleration3.pdf}} \caption{Sketch of the function in \thm{provableinformal}. The left figure explains the construction in the domain and the right figure plots the landscape of a two-dimensional example. } \label{fig:provable-acceleration} \end{figure} Details of \thm{provableinformal} are presented in \sec{separation}. Following similar idea to~\cite{JLG+18}, our construction relies on locally non-informative regions. Main structures of the constructed landscape are illustrated in \fig{provable-acceleration}, which has two global minima. $W_-$ and $W_+$ are two symmetric wells containing one global minimum respectively, $B_{\mathbf{v}}$ is a plateau connecting $W_-$ and $W_+$, and other places form a much higher plateau. The region $W_+$ is our target. We show that the landscape satisfies the following properties: \begin{itemize} \item $S_{\mathbf{v}}$, which is a band $\{\bf{x}\mid\bf{x}\in \mathbb{B}, |\bf{x}\cdot \bf{v}|\leq w\}$ with $w$ a constant and $\bf{v}$ a unit vector, occupies dominating measure in the ball $\mathbb{B}$. \item In $S_{\mathbf{v}}$, local queries (see \defn{localquery}) do not reveal information about the direction $\mathbf{v}$. \item Local queries outside $\mathbb{B}$ do not reveal information about the region inside $\mathbb{B}$. \end{itemize} Restricted in the ball $\mathbb{B}$, the first two properties make classical algorithms intractable to escape from $S_{\mathbf{v}}$ and thus cannot hit $W_+$ efficiently. The last property ensures that, without being restricted in $\mathbb{B}$, classical algorithms are still unable to hit $W_+$ efficiently. See \sec{clb} for details. Nevertheless, quantum tunneling can be efficient if we carefully design the function values and the parameter $h$. The design of the parameters should establish the following main conditions (See \sec{qub} for details): \begin{itemize} \item The wave function always concentrates in $W_-$ or $W_+$. \item The quantum learning rate $h$ is small such that our theory based on semi-classical analysis is valid. \item Quantum tunneling from $W_-$ to $W_+$ is always easy (can happen within time polynomial in $d$). \end{itemize} \paragraph{Organization.} \sec{prelim} introduces our assumptions and problem settings, both classical and quantum. In \sec{quantumal}, we explore QTW in details and state the formal version of \thm{informalQTW}. This includes a one-dimensional example, the formal definition of QTW, the mixing and hitting time of QTW, and the example on tensor decomposition (\prop{TenDecTtot-main}). \sec{comparison} covers detailed quantum-classical comparisons. First, we introduce fair criteria of the comparison. Second, we illustrate the advantages of quantum tunneling and give a detailed view of our \hyperref[slog:1]{Main Message}. Third, we prove \thm{provableinformal}. We corroborate our findings with numerical experiments in \sec{num}. At last, the paper is concluded with discussions in \sec{discussion}. \section{Preliminaries}\label{sec:prelim} \subsection{Notations} Throughout this paper, the space we consider is either $\mathbb{R}^d$ or a $d$-dimensional smooth compact Riemannian manifold denoted $M$. Bold lower-case letters $\mathbf{x}$, $\mathbf{y}$,\ldots, are used to denote vectors. If there is no ambiguity, we use normal lower-case letters, $x$, $y$,\ldots, to denote these vectors for simplicity. Depending on the context, $\d x$ may refer to either line differential or volume differential. We use $A_{jj'}$ to denote the element of the matrix $A$ at of row $j$ and column $j'$. Conversely, given all matrix elements $A_{jj'}$, we use the notation $(A_{jj'})$ to denote the matrix. Unless otherwise specified, $\| \cdot\|$ is used to denote the $\ell^2$ norm of vectors, spectral norm of matrices, and $L^2$ norm of functions. Similarly, $\| \cdot\|_1$ is used to denote the $\ell^1$ norm of vectors and $L^1$ norm of functions. For a function $f$, $\nabla f$ and $\nabla^2 f$ denote the gradient vector and Hessian matrix, respectively. $\Delta:=\sum_{j=1}^{d}\frac{\partial^2}{\partial x_{j}^2}$ is the Laplacian operator. $C^{\infty}(\mathbb{R}^d)$ is the set of all functions $f\colon\mathbb{R}^{d}\to\mathbb{R}$ that are continuous and differentiable up to any order. Notations about upper and lower bounds, $O(\cdot)$, $o(\cdot)$, $\Omega(\cdot)$, and $\Theta(\cdot)$, follow common definitions. We also write $f\ll g$ if $f=o(g)$, and $f\sim g$ if $f=\Theta(g)$. The $\tilde{O}$ notation omits poly-logarithmic terms, namely, $\tilde{O}(f):=O(f\mathrm{poly}(\log f))$ (in this paper, $\log$ denotes the logarithm with base $2$ and $\ln$ denotes the natural logarithm with base $e$). We write $f=O(g^{\infty})$ if \begin{align}\label{eqn:infty-power} \forall N>0,\quad f/g^{N} \to 0~(g\to 0). \end{align} Throughout the paper, we write $f\approx g$ if \begin{align}\label{eqn:approx-definition} f(x)=g+o(g) \end{align} when $g\neq 0$. When $g\to 0$, $f\to 0$ means $f=o(1)$ or, to stress the dependence, $f=o_g(1)$. For quantum mechanics, we use the \emph{Dirac notation} throughout the paper. Quantum states are vectors from a Hilbert space with unit norm. Let $\ket{\phi}$ denote a state vector, and $\bra{\phi}=(\ket{\phi})^{\dagger}$ denote the dual vector that equals to its conjugate transpose. The inner product of two states can be written as $\ip{\psi}{\phi}$. In the coordinate representation, for each $x\in\mathbb{R}^{d}$ we have the wave function $\phi(x):=\ip{x}{\phi}$, where $\ket{x}$ denotes the state localized at $x$. More basics on quantum mechanics and quantum computing can be found in standard textbooks, for instance~\cite{NC10}. \subsection{Classical preparations}\label{sec:classicalpre} Classical algorithms only have access to local information about the objective function at different sites, which is formalized as follows: \begin{definition}[Algorithms based on local queries]\label{defn:localquery} Denote a sequence of points and corresponding queries with size $T$ by $\{x_i, q(x_i)\}_{i=1}^T$, where each $q(x_i)$ can include the function value and arbitrary order derivatives (if exist). Algorithms based on local queries are those which determine the $j$th point $x_j$ by $\{x_i, q(x_i)\}_{i=1}^{j-1}$. \end{definition} As an example, the classical algorithm SGD can be mathematically described by \begin{definition}[Discrete model of SGD] Given a function $f(x)$, starting from an initial point $x_0$, SGD updates the iterates according to \begin{align} x_{k+1} = x_k - s\nabla f(x_k) - s\xi_{k}, \end{align} where $s$ is the \emph{learning rate} and $\xi_k$ is the noise term at the $k$th step. \end{definition} The local information in SGD is gradients. Since $s$ is small, define time $t_k = ks$, the points $\{x_k\}$ can be approximated by points on a smooth curve $\{X(t_k)\}$. The curve, which can be regarded as the continuous-time limit of discrete SGD, is determined by a learning-rate-dependent stochastic differential equation (lr-dependent SDE): \begin{definition}[SDE approximation of SGD] \begin{align} \d x = -\nabla f(x)\d t + \sqrt{s}\d W, \label{eq:lrsde} \end{align} where $W$ is a standard Brownian motion. \end{definition} The solution of \eq{lrsde}, $X(t)$, is a stochastic process whose probability density $\rho_{\rm SGD}(t,\cdot)$ evolves according to the Fokker–Planck–Smoluchowski equation \begin{align} \frac{\partial \rho_{\rm SGD}}{ \partial t} = \nabla\cdot (\rho_{\rm SGD} \nabla f) + \frac{s}{2} \Delta \rho_{\rm SGD}. \label{eq:FPS} \end{align} The validity of this SDE approximation has been discussed and verified in previous literature \cite{KY03,CS18,SSJ20,LMA21}. The results used in the present paper about SGD are based on analyses on \eq{lrsde}. For SGD, we consider an objective function $f$ in $\mathbb{R}^d$ and assume the following: \begin{assumption}[Confining condition~\cite{markowich1999trend,pavliotis2014stochastic}]\label{assum:confining} The objective function $f \in C^{\infty}(\mathbb{R}^d)$ should satisfy $\lim_{\|x\|\to +\infty } f(x) = +\infty$, and $\forall s>0,\exp(-2f/s)$ is integrable and \begin{align} \int_{\mathbb{R}^d} e^{-\frac{2f(x)}{s}} \d x < \infty. \end{align} \end{assumption} \begin{assumption}[Villani condition~\cite{villani2009hypocoercivity}]\label{assum:Villani} The following equation holds for all $s>0$: \begin{align} \|\nabla f(x) \|^2 - s\Delta f(x) \to \infty \quad(\|x\|\to \infty). \end{align} \end{assumption} \begin{assumption}[Morse function]\label{assum:morse} For any critical point $x$ of $f$ (i.e., $\nabla f(x) = 0$), the Hessian matrix $\nabla^2 f(x)$ is nondegenerate (i.e., all the eigenvalues of the Hessian are nonzero). \end{assumption} Under \assum{confining}, Eq.~\eq{FPS} admits a unique invariant Gibbs distribution \begin{align} \mu_{\rm SGD}(x) := \frac{e^{-\frac{2f(x)}{s}}}{\int_{\mathbb{R}^d}e^{-\frac{2f(x)}{s}} \d x}. \label{eq:Gibbs} \end{align} \begin{definition} A measurable function $g$ belongs to $L^2(\mu_{\rm SGD}^{-1})$, if \begin{align} \| g\|_{\mu_{\rm SGD}^{-1}} := \Big(\int_{\mathbb{R}^d} g^2 \mu_{\rm SGD}^{-1}(x) \d x\Big)^{\frac{1}{2}} < +\infty. \end{align} \end{definition} Under such measure, we have: \begin{lemma}[Lemma 2.2 and 5.2 of \cite{SSJ20}] Under \assum{confining}, if the initial distribution $\rho_{\rm SGD}(0,\cdot)\in L^2(\mu_{\rm SGD}^{-1})$, the lr-dependent SDE \eq{lrsde} admits a weak solution whose probability density \begin{align} \rho_{\rm SGD}(t,\cdot) \in C^1([0,+\infty), L^2(\mu_{\rm SGD}^{-1})), \end{align} is the unique solution to \eq{FPS} and $\rho_{\rm SGD}(t,\cdot) \to \mu_{\rm SGD}~(t \to \infty)$. \end{lemma} If we set $\psi_{\rm SGD}(t,\cdot) := \rho_{\rm SGD}(t,\cdot)/\sqrt{\mu_{\rm SGD}} \in L^2(\mathbb{R}^d)$, Eq.~\eq{FPS} is equivalent to \begin{align} s\frac{\partial \psi_{\rm SGD}}{ \partial t} = -\frac{\Delta_{f}^s}{2}\psi_{\rm SGD}, \label{eq:pseudo} \end{align} where $\Delta_{f}^s$ is called the \emph{Witten-Laplacian}, more specifically, \begin{align} \Delta_{f}^s:= -s^2 \Delta + \| \nabla f\|^2 - s\Delta f. \label{eq:defnWL} \end{align} Let $\delta_{s,1}$ be the smallest non-zero eigenvalue of $\Delta_{f}^s$, the following convergence guarantee for SGD holds: \begin{proposition}[Part of Theorem 2.8 in \cite{Mic19} and Lemma 5.5 in \cite{SSJ20}]\label{prop:mic19} Under \assum{confining}, \ref{assum:Villani}, and \ref{assum:morse}, for sufficiently small $s$, \begin{align} \|\rho_{\rm SGD} (t,\cdot) - \mu_{\rm SGD}\|_{\mu_{\rm SGD}^{-1}} \leq e^{-\frac{\delta_{s,1}}{2s}t}\|\rho_{\rm SGD} (0,\cdot) - \mu_{\rm SGD}\|_{\mu_{\rm SGD}^{-1}}, \end{align} where the smallest positive eigenvalues of the Witten-Laplacian $\Delta_{f}^s$ associated with $f$ satisfies \begin{align} \delta_{s,1} = s(\gamma_1 + o(s))e^{-\frac{2H_f}{s}}. \end{align} Here, $H_f$ and $\gamma_1$ are constants depending only on the function $f$. \end{proposition} \begin{corollary}\label{cor:sgdmixing} Assume the assumptions of \prop{mic19} are satisfied, for sufficiently small $s$ and any $\epsilon>0$, if \begin{align} t> \frac{2s}{\delta_{s,1}}\ln \frac{\|\rho_{\rm SGD} (0,\cdot) - \mu_{\rm SGD}\|_{\mu_{\rm SGD}^{-1}}}{\epsilon}, \end{align} then \begin{align} \|\rho_{\rm SGD} (t,\cdot) - \mu_{\rm SGD}\|_{\mu_{\rm SGD}^{-1}}<\epsilon. \end{align} \end{corollary} That is, the convergence time of SGD is loosely $O(s/\delta_{s,1})$ whose magnitude is largely related to $H_f$. The constant $H_f$ is called the \emph{Morse saddle barrier}, characterizing the largest height of barriers. Rigorous results about eigenvalues of the Witten-Laplacian are reviewed in \append{A.1}. \subsection{Quantum preparations}\label{sec:quantumpre} Our quantum algorithmic method essentially relies on the simulation of the Schr\"odinger equation: \begin{align} i \frac{\partial}{\partial t} \ket{\Phi(t)}= H \ket{\Phi(t)}, \label{eq:true} \end{align} where $i$ is the imaginary unit, $\hbar$ is the Planck constant, and $H$ is the Hamiltonian. Physically, we simulate a particle moving under a potential function $f$. Then, in the coordinate representation, the Schr\"odinger equation is specified: \begin{align} i \hbar\frac{\partial}{\partial t} \Phi = \Big(-\frac{\hbar^2}{2m}\Delta + f(x)\Big) \Phi. \end{align} Throughout this paper, we set $h = \hbar/\sqrt{2m}$. The spectrum of the Hamiltonian will highly depends on the variable $h$. More interestingly, by comparing \eq{pseudo} and \eq{true}, $h$ plays a similar role to that of the learning rate $s$. Thus we refer to $h$ as the quantum learning rate. In the reality, $\hbar$ is a fixed constant. However, since we are simulating quantum evolution by quantum computers, proper rescaling the simulation can equivalently be seen as varying the value of $\hbar$. The value of $\hbar$ affects the evolution time needed. However, rescaling $\hbar$ has no impact on quantum query complexity. Therefore, in this paper, $\hbar$ is an unimportant constant, i.e., $\hbar = 1$. As is introduced, we consider quantum tunneling from the bottom of a well to that of another well, in other words, tunneling between \emph{local ground states}. A local ground state of a well is the local eigenstate of the well with minimum eigenvalue. Technically, several kinds of local eigenstates are defined (see \blue{Definition A.7}, \blue{Definition A.8}, and \blue{Definition A.10} in \append{A.2.3}). Despite of the number of definitions, different kinds of local eigenstates are close to each other and share the same intuition: eigenstates of the Hamiltonian restricted in regions only contain one well. For convenience, if no otherwise specified, local eigenstates stand for orthonormalized eigenstates defined by \blue{Definition A.10}. Actually, there are also tunneling effects between local excited states. However, excited states are difficult to approximate accurately for general landscapes. Besides, due to interference, tunneling effects between different local excited states may cancel each other out. We restrict our attention to tunneling between local ground states in order to obtain explicit results along with a clear physical picture. Two local ground states can interact strongly with each other only if the difference between their energies is small relative to $h$ \cite{Hel88} (see also discussions after \blue{Proposition A.4} in \append{A.2.3}). In other words, this requires the function values between two local minima to be close and there is little resonance between the first (local) excited state in one well and the (local) ground state of the other \cite{Ras12,SCC91}. Therefore, our algorithms based on tunneling between local ground states are essentially restricted on landscapes where local minima are approximately global minima. Note that we can always find small enough $h$ to make two local ground states nonresonant, if the corresponding local minima are not exactly equal. As a result, to avoid more complicated restrictions on $h$, without loss of generality we assume that local minima are global minima, and they all have function value 0. More precisely: \begin{assumption}\label{assum:quantum1} The smooth objective function $f\colon \mathbb{R}^{d}\to \mathbb{R}$ satisfies \begin{equation} 0 = \min f < \lim_{\|x\|\to \infty} f. \end{equation} In addition, $f$ has finite number of local minima, and they can be decomposed as follows: \begin{equation} f^{-1}(0) = U_1 \cup U_2 \ldots \cup U_N, \end{equation} \begin{equation} U_j = \{x_j\}~\mathrm{is~a~point,}~\nabla f(x_j) = 0,~\mathrm{and}~\nabla^2 f(x_j) > 0~\mathrm{for}~j=1,\ldots,N. \end{equation} Each $U_j$ is called a \emph{well}. \end{assumption} This assumption will not affect the explicit forms of convergence time or present less physical insights. To further characterize the distance on such landscapes, an important geometric tool we use is the Agmon distance. \begin{definition}[Agmon distance]\label{defn:Agmon-distance} Under \assum{quantum1}, the Agmon distance $d(x,y)$ is defined as \begin{align} d(x,y) := \inf_{\gamma}\int_{\gamma} \sqrt{f(x)}\d x, \end{align} where $\gamma$ denotes pairwise $C^1$ paths connecting $x$ and $y$. For a set $U$, $d(x,U) = d(U,x) := \inf_{y\in U}d(x,y)$. And for two sets $U_1$ and $U_2$, $d(U_1,U_2) = \inf_{x\in U_1,y\in U_2}d(x,y)$. \end{definition} The minimal Agmon distance between wells are defined as \begin{align} S_0 :=\min_{j\neq k}d(U_j,U_k). \end{align} We only consider resonant wells by assuming the following for simplicity: \begin{assumption}[Informal]\label{assum:quantum2} The difference between any two local ground states are of the order $O(h^{\infty})$. In addition, for any well $U_j$, there exists another well $U_k$ ($k\neq j$) such that $d(U_j,U_k) = S_0$. \end{assumption} At last, to obtain explicit results we demand that \begin{assumption}\label{assum:quantum3} There are a finite number of paths of the Agmon length $S_0$ connecting $U_j$ and $U_k$ if $d(U_j,U_k) = S_0$. \end{assumption} \assum{quantum1}, \ref{assum:quantum2}, and \ref{assum:quantum3} are simplified from \blue{Assumptions A.5}, \blue{A.6}, and \blue{A.7} in \append{A.2} which presents details needed for analyzing quantum tunneling. Under \assum{quantum1}, \ref{assum:quantum2}, and \ref{assum:quantum3}, we state the main results of \append{A.2} as follows. For sufficiently small $h$, the orthonormalized local ground states $\ket{e_j},~j=1,\ldots,N$ almost localize near the wells $U_j,~j=1,\ldots,N$, respectively. The space $\mathcal{F}$ spanned by $\{\ket{e_j}:j=1,\ldots,N\}$ is exactly a low-energy invariant subspace of the Hamiltonian $H$. In other words, in the low-energy space $\mathcal{F}$, the particle walks between wells by quantum tunneling. The Hamiltonian restricted in $\mathcal{F}$, i.e., $H_{|\mathcal{F}}$, determines the strength of the quantum tunneling effect and is called the \emph{interaction matrix}. To explore $H_{|\mathcal{F}}$, we use the WKB method to estimate local ground states (\append{A.2.1}). Any local ground state function decays exponentially with respect to the Agmon distance to its corresponding well (\append{A.2.2}). Consequently, the tunneling effects would decay exponentially with respect to $S_0$. Having captured theses properties, \append{A.2.3} can give explicit estimations about $H_{|\mathcal{F}}$, namely, \blue{Proposition A.5}, \blue{Proposition A.6} (with $N^+ = N$), and \blue{Theorem A.1}. Finally, we restate a more formal version of our \ref{prb:main}: \begin{named}{Main Problem (restated)}\label{prb:mainformal} Given an objective function $f$ that satisfies \assum{quantum1}, \ref{assum:quantum2}, and \ref{assum:quantum3}, starting from one local minimum, find all local minima or find a certain target minimum. \end{named} We make the following remarks for clarification: \begin{remark} Assumptions in \sec{classicalpre} and \sec{quantumpre} are not contradictory. When considering SGD, we naturally add to the \ref{prb:mainformal} that the assumptions in \sec{classicalpre} should also be satisfied. \end{remark} \begin{remark} The assumption of starting from one local minimum enables quantum algorithms to prepare a local ground state, or more generally, a state largely in the aforementioned subspace $\mathcal{F}$. \end{remark} \begin{remark} Because finding a precise global minimum is impractical in general, it suffices to find points sufficiently close to the minima of interest. Later in \sec{standard}, we use two different measures of accuracy: 1. the function value difference; and 2. the distance to one of the minimum. \end{remark} \section{Quantum Tunneling Walks}\label{sec:quantumal} In this section, we present full details of the quantum tunneling walk (QTW). We begin with a one-dimensional example in \sec{onedimexp}, and then in \sec{QTW} we formally define QTW. In \sec{Qmixing} and \sec{Qhitting}, we study the mixing and hitting time of QTW, respectively. As an example, we give full details of applying QTW to tensor decomposition in \sec{tendecom}. \subsection{A one-dimensional example}\label{sec:onedimexp} We start the introduction of the quantum algorithm QTW with a one-dimensional example which quantifies the intuitions provided in \sec{intro}. \sec{onedimexp} also serves as a map connecting each step of the analysis to the needed mathematical tools. Later sections can be seen as generalizing results here for high dimensional and multi-well cases. General descriptions and results begin at \sec{QTW}. Consider the potential $f(x)$ in \fig{1-dimexample}, which has two global minima, $x_{\pm} = \pm a$. For simplicity, we take \begin{equation} f(x) = \frac{1}{2} \omega^2 (x + a)^2,\quad x \leq -\epsilon, \label{eq:eq01} \end{equation} where $\epsilon$ is a small number. In this way, the potential satisfies that $\min f =0$, and $f(x)$ is quadratic near minima. Besides, the symmetry of wells demands $f(x) = f(-x)$. The $f(x)$ for $x\in [-\epsilon,\epsilon]$ can always be made to be smooth. Near the two minima, $\pm a$, whose local harmonic frequencies are $\omega$, we can solve the Schr\"odinger equation locally and get two local ground states, $\Phi_{\pm}(x)$. For instance, around $-a$, if we set $y = x+ a$, the local ground state is determined by \begin{equation} H \Phi_{-}(x) = \varepsilon_0 \Phi_{-}(x),\quad H = - h^2 \frac{\d^2}{\d^2 y} + \frac{1}{2} \omega^2 y^2, \end{equation} where $\varepsilon_0 = h\omega/\sqrt{2}$. Physically, the demand of localization is equivalent to $\varepsilon_0 \ll f(0)$, indicating that the particle nearly cannot pass through the energy barrier. Concrete mathematical definitions and discussions on local ground states can be found in \append{A.2.2}. From a high-level perspective, the main idea of the present paper is to unite the interaction or tunneling between local ground states to realize algorithmic speedups. As we want to investigate the evolution of states, we need to determine the relationship between local ground states and the eigenstates of the Hamiltonian $H$. We set the eigenstates of $H$ as $\ket{n},~n=0,1,\ldots$ with energies $E_0 \leq E_1 \leq \cdots$, respectively (i.e., $\ket{0}$ is the global ground state, $\ket{1}$ is the first excited state, etc.). The overlap of states $\Phi_{\pm}$ is small (namely, $\ip{\Phi_{+}}{\Phi_{-}}\approx 0$), as they are local and separated by a high barrier. Denote the subspace spanned by $\Phi_{-}$ and $\Phi_{+}$ as $\mathcal{E}$, and that spanned by $\ket{0}$ and $\ket{1}$ as $\mathcal{F}$. Both $\mathcal{E}$ and $\mathcal{F}$ are 2-dimensional and contains states with low energies. It is intuitive that $\mathcal{E} \approx \mathcal{F}$ (which is guaranteed by \blue{Proposition A.3}). For the one-dimensional case, we just take $\mathcal{E} = \mathcal{F}$ for simplicity. In this way, we can represent $\ket{0}$ and $\ket{1}$ by $\ket{\Phi_{\pm}}$ in the following general way: \begin{align} \ket{0} &= \cos\theta \ket{\Phi_{-}} + \sin\theta \ket{\Phi_{+}},\\ \ket{1} &= \sin\theta \ket{\Phi_{-}} - \cos\theta \ket{\Phi_{+}}. \end{align} Restricted in the subspace $\mathcal{F}$, the two-level system Hamiltonian can be written as \begin{align} H|_{\mathcal{F}} &= \left( \begin{array}{cc} \varepsilon_{-} & -\nu \\ -\nu & \varepsilon_{+} \end{array} \right),\quad \mathrm{under~basis}~\{\ket{\Phi_{-}},\ket{\Phi_{+}}\},\\ H|_{\mathcal{F}} &= \left( \begin{array}{cc} E_0 & 0 \\ 0 & E_1 \end{array} \right),\quad\ \mathrm{under~basis}~\{\ket{0},\ket{1}\}, \end{align} where $\nu$ is called the tunneling amplitude, measuring the interaction between wells. Because the $f(x)$ we choose is symmetric, $\varepsilon_{-} = \varepsilon_{+} = \varepsilon_{0} = h\omega/\sqrt{2}$. Therefore, we have $\theta = \pi/4$ and the energy gap $\Delta E := E_1 - E_0 = 2\nu$. We will refer to this Hamiltonian restricted in subspace $\mathcal{F}$ as the \emph{interaction matrix}, indicating that $H_{|\mathcal{F}}$ characterizes the interaction between wells. In our setting, we can begin at a local minimum, where the local ground state is easy to prepare (see justifications in \append{B.2}). Without loss of generality, let us begin the quantum simulation at the state $\Phi_{-}$, namely, setting $\ket{\Phi(0)} = \ket{\Phi_{-}}$. After evolution of time $t$, the state becomes \begin{equation} \ket{\Phi(t)} = e^{-iHt}\ket{\Phi(0)} = \frac{1}{\sqrt{2}}(e^{-iE_0t}\ket{0} + e^{-iE_1t}\ket{1}) \propto \cos(\Delta E t/2)\ket{\Phi_{-}} + i\sin(\Delta E t/2)\ket{\Phi_{+}}. \end{equation} And the probabilities of finding the particle in the right and left wells are give by \begin{equation} P_{\pm}(t) = |\ip{\Phi_{\pm}}{\Phi(t)}|^2 = \frac{1\mp \cos(\Delta E t)}{2}. \end{equation} The energy gap $\Delta E$ is also called the Rabi oscillation frequency, suggesting that the particle oscillates between the two wells periodically. Our aim is to pass through the barrier and find other local minima (for the case here, is to find the other local minimum). Since for small $h$, the local state $\ket{\Phi_{+}}$ distributes in a very convex region near the right local minimum, it suffice to solve our problem by measuring the position of state $\ket{\Phi(t)}$ when $P_{+}(t)$ is large. However, we may not be able to know $\Delta E$ precisely in advance. So, we will apply the method of quantum walks: evolving system for time $t$ which is chosen randomly from $[0,\tau]$, and then measuring the position \cite{CCD+03}. The resulted distribution is \begin{align} \rho_{\rm QTW}(\tau,x) \approx |\ip{x}{\Phi_{-}}|^2\int_{0}^{\tau}P_{-}(t)\frac{\d t}{\tau} + |\ip{x}{\Phi_{+}}|^2\int_{0}^{\tau}P_{+}(t)\frac{\d t}{\tau}, \end{align} where QTW denotes quantum tunneling walk. Define $p_{-\to \pm}(\tau) = \int_{0}^{\tau}P_{\pm}(t)\frac{\d t}{\tau}$, which is the probabilities of finding the right and left local ground states, respectively. Since $|\ip{x}{\Phi_{-}}|^2$ is small for $x$ near $+a$, the probability of finding the particle near $+a$ is determined by \begin{align} \int_{\rm right~well} \rho_{\rm QTW}(\tau,x) \d x \approx \int_{\rm right~ well}\d x |\ip{x}{\Phi_{+}}|^2p_{-\to +} \approx p_{-\to +}. \end{align} Therefore, it suffice to study properties of $p_{-\to \pm}$. In the present case, \begin{align} p_{-\to \pm} = \frac{1}{2}\left(1 \mp \frac{\sin(\Delta E \tau)}{\Delta E \tau} \right), \end{align} which will converge when $\tau \to \infty$. This fact ensures that we can find the right local minimum $+a$ with a probability larger than some constant after evolving the system for enough long time. Starting from $\ket{\Phi_{-}}$, the hitting time for the right well is \begin{align} T_{\rm hit}(\Phi_{+}|\Phi_{-}) = \inf_{\tau > 0} \frac{\tau}{p_{-\to +}(\tau)}. \end{align} Since the probability for successful tunneling in one trial is $p_{-\to +}(\tau)$, we can repeat trials for $1/p_{-\to +}(\tau)$ times to secure one success and the total evolution time is $\tau/p_{-\to +}(\tau)$. For sufficiently small $\epsilon \ll 1$, if $\frac{1}{\Delta E \tau} \leq \epsilon$, we can get $p_{-\to +}(\tau) \geq \frac{1}{2}(1-\epsilon)$. Therefore, the hitting time can be bounded by $O(\frac{1}{\Delta E\epsilon})$. As is going to be shown later, if we want to find \emph{all} local minima, the mixing time would be a better quantifier. The limiting distribution is \begin{align} \mu_{\rm QTW}(x) := \lim_{\tau \to \infty}\rho_{\rm QTW}(\tau,x) \approx |\ip{x}{\Phi_{-}}|^2 p_{-\to -} + |\ip{x}{\Phi_{+}}|^2p_{-\to +}. \end{align} The mixing time measures how fast the distribution $\rho_{\rm QTW}(\tau,x)$ converges to $\mu_{\rm QTW}(x)$. We define $T_{\rm mix}$ as the $\epsilon$-close mixing time which satisfies \begin{align} T_{\rm mix} =\inf_{\|\rho_{\rm QTW}(\tau,\cdot) - \mu_{\rm QTW}(\cdot)\|_1 \leq \epsilon} \tau. \end{align} Because $\|\rho_{\rm QTW}(\tau,\cdot) - \mu_{\rm QTW}(\cdot)\|_1 \leq O(\frac{1}{\Delta E \tau})$, we have $\|\rho_{\rm QTW}(\tau,\cdot) - \mu_{\rm QTW}(\cdot)\|_1 \leq \epsilon$ if $\tau = \Omega(\frac{1}{\Delta E \epsilon})$. Therefore, $T_{\rm mix}$ could be bounded: $T_{\rm mix} = O(\frac{1}{\Delta E \epsilon})$. For simulating a time-independent Hamiltonian, the number of queries needed are roughly proportional to the total evolution time (as demonstrated in \sec{quantumpre} or see details in \append{B.1}). The major task left is to calculate the energy gap $\Delta E$, get different evolution times and compare them with classical results. As is specified in \append{A.2.3}, $0$ is the boundary of the two wells and the tunneling amplitude $\nu$ can be given by \begin{align} \nu = h^2 (\Phi_{-}(0)\Phi'_{+}(0) - \Phi'_{-}(0)\Phi_{+}(0)). \label{eq:eq15} \end{align} To obtain an explicit result, we need to use the WKB approximation of the local ground states (\append{A.2.1}): \begin{align} \Phi_{-}(x) \approx \frac{1}{ h^{1/4}} a_0(x)e^{-\frac{1}{h}\int_{-a}^{x}\sqrt{f(\xi)}\d \xi},\quad \Phi_{+}(x) = \Phi_{-}(-x), \label{eq:eq16} \end{align} where $a_0(x)$ is given by \begin{align} a_0(x) = \left( \frac{\omega}{\sqrt{2} \pi}\right)^{1/4}e^{-\frac{1}{2}\int_{-a}^{x}(\frac{f'(\xi)}{2f(\xi)} - \frac{\omega}{\sqrt{2f}})\d \xi }, \label{eq:eq17} \end{align} which is determined by the transport equation (Eq.~(\blue{S24}) in \append{A.2.1}) and the normalization condition. Substituting \eq{eq16} and \eq{eq17} to \eq{eq15}, we get \begin{align} \nu = 2\sqrt{\frac{h\omega f(0)}{\sqrt{2}\pi}}e^C[1+O(h)]e^{-\frac{S_0}{h}}, \end{align} where the constants $C$ and $S_0$ are given by \begin{align} C = \int_0^a \Big(\frac{\omega}{\sqrt{2f(\xi)}} - \frac{1}{a-\xi}\Big)\d \xi , \quad S_0 = \int_{-a}^a \sqrt{f(\xi)} \d\xi. \end{align} Note that if $f(x)$ is quadratic, the factor $C$ will be zero, indicating that $C$ measures the deviation of the landscape from being quadratic. The quantity $S_0$ is called the Agmon distance between two the local minima (see \append{A.2.2}). Since we assumed by \eq{eq01} that $f(x)$ is almost quadratic, we have $f(0)\approx \frac{1}{2}\omega^2a^2$ and \begin{align} \nu \approx h\omega\sqrt{\frac{\sqrt{2}\omega a^2}{h\pi}} e^{-\frac{S_0}{h}}. \end{align} As discussed above, the mixing time and hitting time could be bounded by the following characteristic time \begin{align} T_c = \frac{1}{\Delta E} = \frac{1}{2\nu } \approx \frac{\sqrt{\pi}}{2a\omega \sqrt{\sqrt{2}\omega h}}e^{\frac{1}{h}\int_{-a}^a \sqrt{2f(\xi)}\d \xi}. \end{align} Next, we need to find how long it takes for SGD to escape from the left local minimum. Discrete-time SGD with a small learning rate $s$ can be approximated by a learning-rate dependent stochastic differential equation (lr-dependent SDE) \cite{SSJ20}: \begin{align} \d X = - \nabla f(X) \d t + \sqrt{s}\d W, \end{align} where $W$ is a standard Brownian motion. Before hitting $0$, the SDE is almost an Ornstein–Uhlenbeck process: \begin{align} \d X = -\omega^2 (X+a) \d t + \sqrt{s}\d W. \end{align} The expected time for the Ornstein–Uhlenbeck process to first hit $0$ is \begin{align} \mathbb{E}T_0 \approx \frac{\sqrt{\pi s}}{a \omega^3}e^{\frac{2H_f}{s}}, \end{align} where $H_f$ is called the Morse saddle barrier and equals to $f(0) - f(-a) \approx \frac{1}{2}\omega^2 a^2$ in our case. Although $\mathbb{E}T_0$ is not a precise classical counterpart of either quantum mixing or hitting time, it is heuristic to compare $\mathbb{E}T_0$ with $T_c$ which can both reflect the time to escape from the left well. The forms of $\mathbb{E}T_0$ and $T_c$ are very similar. Two major differences can be observed: 1. the exponential term in $\mathbb{E}T_0$ is determined by the height of the barrier, while that in $T_c$ is related to an integral of $\sqrt{f}$; 2. $T_c \propto 1/\omega^{3/2}$ but $\mathbb{E}T_0 \propto 1/\omega^{3}$, indicating that the flatness of the wells affects differently on quantum and classical methods. We will show that for landscapes with multiple wells, the distribution of wells is also an important factor. In general, QTW could be faster than SGD if the barriers between local minima are high but thin, each well is close to many other wells, and wells are flat. The above comparison is intuitive but not rigorous. Two important technical details for comparison are needed for quantitative discussions. It is shown in \sec{quantumpre} that a super-polynomial separation between evolution time of QTW and SGD gives rise to a super-polynomial separation between quantum and classical queries for QTW and SGD, respectively. Therefore, it suffice to compare the evolution time, especially the exponential term $e^{S_0/h}$ and $e^{2H_f/s}$. The second problem is that $h$ and $s$ are not two constants but variables. The evolution times cannot be quantitatively compared if $h$ and $s$ are independent. In \sec{standard}, we develop two natural standards to make fair comparisons between QTW and SGD, which specifies $h$ and $s$. \subsection{Definition of quantum tunneling walks}\label{sec:QTW} We now formally describe the model of a quantum tunneling walk (QTW) on a general objective function $f(x)$ satisfying assumptions in \sec{quantumpre}. The wells are denoted by $U_j = \{x_j\}~(j=1,\ldots,N)$. Let $\ket{j}$ be the corresponding orthonormalized local ground state of $U_j$. It is ensured that $\{\ket{j}:j=1,\ldots,N\}$ spans a low energy subspace, $\mathcal{F}$, of the Hamiltonian $H = -h^2\Delta + f(x)$. If one has information about one well $U_j$ and its neighborhood, the construction of the local ground state should be easy which can be close to $\ket{j}$ or at least be almost in the subspace $\mathcal{F}$. We assume the initial state $\Phi(0)$ to be in $\mathcal{F}$. The evolution is determined by the Schr\"odinger equation, \begin{align} i \frac{\d}{\d t} \ip{x}{\Phi(t)} = \bra{x} H \ket{\Phi(t)}, \end{align} where $|\ip{x}{\Phi(t)}|^2$ is the probability distribution of finding the walker at $x$. The Schr\"odinger equation indicates the phenomenon of quantum tunneling since it can be rewritten as \begin{align} i \frac{\d}{\d t} \ip{j}{\Phi(t)} = \sum_{j'} \bra{j} H_{|\mathcal{F}} \ket{j'} \ip{j'}{\Phi(t)},~\mathrm{for~any}~j=1,\ldots,N, \label{eq:qtwsch} \end{align} given that $\ket{\Phi(0)}$ is in the subspace $\mathcal{F}$. Here, $(\langle j|H_{|\mathcal{F}}|j^{\prime} \rangle)$ is called as the \emph{interaction matrix} and is calculated by \blue{Proposition A.5} and \blue{A.6}. Once we get $\ip{j}{\Phi(t)}$ and $\ip{x}{j}$ for all $j$, we can obtain the probability distribution $|\ip{x}{\Phi(t)}|^2$. The overlap $\ip{x}{j}$ is invariant with respect to $t$. So, we may focus on \eq{qtwsch} to investigate the time evolution. As is shown by \eq{qtwsch}, restricted in the low energy subspace $\mathcal{F}$, the quantum evolution is similar to that of a quantum walk on a graph. The wells correspond to vertices of the graph, and the quantum tunneling effects between wells determine the graph connectivity. QTW walks among different wells by quantum tunneling, helping to find all other local minima. Finally, according to \blue{Lemma B.2} in \append{B.1}, the quantum query complexity of simulating the Schr\"odinger equation is directly linked to the evolution time $t$ and is bounded by \begin{align} O\left(\|f\|_{L^{\infty}(\Omega)} t \frac{\log (\|f\|_{L^{\infty}(\Omega)} t/\epsilon)}{\log \log (\|f\|_{L^{\infty}(\Omega)} t/\epsilon)}\right), \end{align} where $\Omega$ is a large region containing all minima of interest and $\epsilon$ is the precision quantified by the $L^2$ norm between the target and the obtained wave functions. Loosely speaking, we need $\tilde{O}(t)$ quantum queries if the evolution time is $t$. For SGD, the number of queries needed is at least $\Omega(t/s)$ for time $t$. Thus, as long as there is a super-polynomial separation between QTW and SGD evolution time, there is a super-polynomial separation between quantum queries and classical queries for QTW and SGD, respectively. Conclusions on speedups are essentially based on comparisons of query complexity. However, based on this relationship between evolution time and query complexity, we can focus on comparisons of time. To sum up, QTW is quantum simulation with the system Hamiltonian being $H = -h^2\Delta +f(x)$ and the initial state being in a low energy subspace of $H$, where $f(x)$ is the potential function of a type of benign landscapes (\assum{quantum1}, \ref{assum:quantum2}, and \ref{assum:quantum3}). QTW can be efficiently implemented on quantum computers. \subsection{Mixing time}\label{sec:Qmixing} For a given landscape $f(x)$, the complexity of Hamiltonian simulation mainly depends on the evolution time (see~\append{B.1}). In this section, we focus on the evolution time needed for fulfilling the tasks of finding all minima. Since quantum evolutions are unitary, different from SGD, QTW never converges. In this case, after running QTW for some time $t$, we measure the position of the walker. Similar to the quantum walks in \cite{CCD+03}, the evolution time $t$ can be chosen uniformly in $[0,\tau]$. Later, we will prove that under sufficiently large $\tau$, QTW can find other wells with probability larger than some constant. Note that there can be better strategies to determine the time for measurement $t$ than uniformly sampling from an interval $[0,\tau]$ \cite{AC21}. For simplicity, we only analyze the original strategy of \cite{CCD+03} in the present paper. As is mentioned earlier, we initialize at a state $\ket{\Phi(0)}\in \mathcal{F}$. In later sections, we may specify $\ket{\Phi(0)}$ to be one of the local ground states. Let the spectral decomposition of $H_{|\mathcal{F}}$ to be $H_{|\mathcal{F}} = \sum_{k=1}^N E_k \ket{E_k}\bra{E_k}$. Simulating the system for a time $t$ chosen uniformly in $[0,\tau]$, one can obtain the probability density of finding the walker at $x$\footnote{The distribution $\rho_{\rm QTW}(\tau,x)$ depends on the initial state $\ket{\Phi(0)}$. A more rigorous notation is $\rho_{\rm QTW}(\tau,x|\Phi(0))$. When there is no confusion, we omit the initial state for simplicity.} \begin{align} \rho_{\rm QTW}(\tau,x) :=& \frac{1}{\tau}\int_0^{\tau} \d t |\langle x|e^{-i H_{|\mathcal{F}} t}\ket{\Phi(0)}|^2 \nonumber \\ =& \sum_{E_k=E_{k^{\prime}}} \langle x|E_k\rangle \langle E_k |\Phi(0)\rangle \langle \Phi(0) |E_{k^{\prime}}\rangle \langle E_{k^{\prime}}| x\rangle \nonumber\\ &~+\sum_{E_k \neq E_{k^{\prime}}} \frac{1-e^{-i(E_k - E_{k^{\prime}})\tau}}{i(E_k - E_{k^{\prime}})\tau} \langle x|E_k\rangle \langle E_k |\Phi(0)\rangle \langle \Phi(0) |E_{k^{\prime}}\rangle \langle E_{k^{\prime}}| x\rangle. \end{align} The time-averaged probability density leads to a limiting distribution when $\tau \to \infty$: \begin{equation} \mu_{\rm QTW} := \sum_{E_k=E_{k^{\prime}}} \langle x|E_k\rangle \langle E_k |\Phi(0)\rangle \langle \Phi(0) |E_{k^{\prime}}\rangle \langle E_{k^{\prime}}| x\rangle. \end{equation} With the strategy of measuring at $t$ randomly chosen from $[0,\tau]$, QTW can output a distribution $\rho_{\rm QTW}(\tau,x)$ with limit. Such a process is regarded as mixing. Quantum mixing time evaluates how fast $\rho_{\rm QTW}(\tau,x)$ converges to $\mu_{\rm QTW}(x)$, and is rigorously defined as: \begin{definition}[Mixing time of QTW]\label{defn:mixingqtw} $T_{\rm mix}$ is called the $\epsilon$-close mixing time, iff for any $\tau\geq T_{\rm mix}$, we have \begin{align} \|\rho_{\rm QTW}(\tau,\cdot) - \mu_{\rm QTW}(\cdot)\|_1 \leq \epsilon. \label{eq:mixingqtw} \end{align} \end{definition} The following lemma provides a general bound for the QTW mixing time whose proof is postponed to \append{C.1.1}. \begin{lemma}[Upper bound for QTW mixing time]\label{lem:qtwmixingtime} The condition \eq{mixingqtw} can be satisfied if \begin{align} \tau \geq \frac{2}{\epsilon} \sum_{E_k\neq E_{k^{\prime}}}\frac{|\ip{E_k}{\Phi(0)} \ip{\Phi(0)}{E_{k'}}|}{|E_k - E_{k'}|}[1+(N-1)|O(h^{\infty})|], \end{align} and this implies \begin{align} \hspace{-2mm} T_{\rm mix} = O \bigg(\frac{1}{\epsilon} \sum_{E_k\neq E_{k^{\prime}}}\frac{|\ip{E_k}{\Phi(0)} \ip{\Phi(0)}{E_{k'}}|}{|E_k - E_{k'}|}[1+(N-1)|O(h^{\infty})|] \bigg) \leq O \bigg(\frac{N}{\epsilon \Delta E} [1+(N-1)|O(h^{\infty})|] \bigg), \label{eq:qtwmixing} \end{align} where $\Delta E := \min_{E_k\neq E_{k^{\prime}}}|E_k - E_{k'}|$ is referred to as the minimal gap of $H_{|\mathcal{F}}$. \end{lemma} The term $O(h^{\infty})$ in \eq{qtwmixing} originates from integrals $\int |\ip{x}{j} \ip{j'}{x}| \d x,~j\neq j'$. Intuitively, states $\ket{j}$ and $\ket{j'}$ localize in different wells, such that $\ip{x}{j} \ip{j'}{x}$ is exponentially small with respect to $h$ for any $x$. \lem{qtwmixingtime} highlights the dependence of the mixing time on the initial state $\ket{\Phi(0)}$ and the eigenvalue gaps of $H_{|\mathcal{F}}$. Concrete examples will be given in \sec{comparison}, where we further illustrate \eq{qtwmixing} and compare QTW with SGD. As is mentioned, \eq{qtwsch} indicates a quantum walk: a well $U_j$ can be seen as a vertex of a graph and $H_{|\mathcal{F}}$ implies graph connectivity (interaction between wells) similar to the graph Laplacian. The connection between QTW and quantum walks is helpful to simplify the physical picture of QTW. However, we also address the difference between quantum walks and QTW. For quantum walks, we only consider the probabilities of finding the walker at vertices, that is,\footnote{Similar to $\rho(\tau,x)$, the probability $p(\tau,j)$ depends on the initial state $\ket{\Phi(0)}$ and a more rigorous notation is $p(\tau,j|\Phi(0))$. When there is no confusion, we omit the initial state for simplicity.} \begin{align} p(\tau,j) &:= \frac{1}{\tau}\int_0^{\tau} \d t |\langle j|e^{-i H_{|\mathcal{F}} t}| \Phi(0) \rangle|^2 \nonumber \\ & = \sum_{E_k=E_{k^{\prime}}} \langle j|E_k\rangle \langle E_k |\Phi(0)\rangle \langle \Phi(0) |E_{k^{\prime}}\rangle \langle E_{k^{\prime}}| j\rangle \nonumber\\ &~~+\sum_{E_k \neq E_{k^{\prime}}} \frac{1-e^{-i(E_k - E_{k^{\prime}})\tau}}{i(E_k - E_{k^{\prime}})\tau} \langle j|E_k\rangle \langle E_k |\Phi(0)\rangle \langle \Phi(0) |E_{k^{\prime}}\rangle \langle E_{k^{\prime}}| j\rangle. \end{align} When $\tau \to \infty$, $p(\tau,j)$ also converges to a limit \begin{align} p(\infty,j) := \sum_{E_k=E_{k^{\prime}}} \langle j|E_k\rangle \langle E_k |\Phi(0)\rangle \langle \Phi(0) |E_{k^{\prime}}\rangle \langle E_{k^{\prime}}| j\rangle. \end{align} Following results, \lem{limitdis} and \lem{mixingqw}, show the connection and difference between QTW and quantum walks in a more quantitative way (detailed proofs can be found in \append{C.1.2} and \append{C.1.3}). \begin{lemma}[Limit distributions]\label{lem:limitdis} Limit distributions of the QTW and the quantum walk satisfy the following: \begin{align} \mu_{\rm QTW}(x) = \sum_{j} p(\infty, j) |\langle x|j\rangle|^2 + O(h^{\infty}). \end{align} \end{lemma} \begin{definition}[Mixing time of quantum walks \cite{CLR20}]\label{defn:mixingqw} $t_{\rm mix}$ is called the $\epsilon$-close mixing time of the quantum walk, iff for any $\tau \geq t_{\rm mix}$, \begin{align} \sum_{j=1}^N |p(\tau,j) - p(\infty, j)| \leq \epsilon. \label{eq:disofp} \end{align} \end{definition} \begin{lemma}[Upper bound for the mixing time of quantum walks]\label{lem:mixingqw} The condition \eq{disofp} is satisfied if \begin{align} \tau \geq \frac{2}{\epsilon} \sum_{E_k\neq E_{k^{\prime}}}\frac{|\ip{E_k}{\Phi(0)} \ip{\Phi(0)}{E_{k'}}|}{|E_k - E_{k'}|}, \end{align} and we have \begin{align} t_{\rm mix} = O \bigg(\frac{1}{\epsilon} \sum_{E_k\neq E_{k^{\prime}}}\frac{|\ip{E_k}{\Phi(0)} \ip{\Phi(0)}{E_{k'}}|}{|E_k - E_{k'}|} \bigg) \leq O \bigg(\frac{N}{\epsilon \Delta E} \bigg). \label{eq:qwmixing} \end{align} \end{lemma} By \lem{limitdis} and the comparison of \lem{qtwmixingtime} and \lem{mixingqw}, we know that for sufficiently small $h$, which indicates that local ground states localize sufficiently near their respective wells, QTW can be well characterized by a quantum walk. On a higher level of speaking, QTW generalizes quantum walks from walking on discrete graphs to propagating on continuous functions. And QTW may enable new phenomenons not shown in quantum walks when states $\ket{j}$ are poorly localized near $U_j$. On the other hand, QTW under proper conditions can be used to implement quantum walks. \subsection{Hitting time}\label{sec:Qhitting} If we aim at finding one particular well (the one with global minimum or the one with the best generalization properties), hitting time instead of mixing time should be of interest. Classically, the hitting time is the expected time required to find some target region or point. For quantum algorithms, we cannot output the position of the walker at all times and the system state would be destroyed by measuring its position. Thus, the definition of the hitting time for quantum algorithms is slightly different. We first see how previous literature defines the quantum walk hitting time: \begin{definition}[Hitting time for quantum walks \cite{AC21}] Consider a quantum walk governed by \eq{qtwsch}. Let the state $\ket{j}$ be the one of interest. Then, starting from the initial state $\ket{\Phi(0)}$, the hitting time of the quantum walk is defined as follows: \begin{align} t_{\rm hit}(j) := \inf_{\tau>0} \frac{\tau}{p(\tau,j)}. \end{align} \end{definition} To understand this definition, we first refer to the process, evolving the system for time $t$ uniformly chosen from $[0,\tau]$, as one trial. Using one trial, the probability of getting $\ket{j}$ is $p(\tau,j)$. So, repeating the trials for $1/p(\tau,j)$ times guarantees to hit $\ket{j}$ with high probability. In this case, the total evolution time needed is bounded by $\tau/p(\tau,j)$. In the same spirit, we can define and bound the QTW hitting time as follows: \begin{definition}[Hitting time of QTW] For QTW governed by \eq{qtwsch}, let the open and $C^2$-bounded region $\Omega$ be the region of interest. Then, starting from the initial state $\ket{\Phi(0)}$, the $\Omega$-hitting time of QTW is defined as follows: \begin{align} T_{\rm hit}(\Omega) := \inf_{\tau>0} \frac{\tau}{\int_{\Omega}\rho_{\rm QTW}(\tau,x)\d x}. \end{align} \end{definition} Basic results about the hitting time (\lem{qwhitting} and \lem{qtwhitting}) are present as follows. The proof of \lem{qtwhitting} is in \append{C.2.1} and that of \lem{qtwhitting} is in \append{C.2.2}. \begin{lemma}[Upper bound of the quantum walk hitting time]\label{lem:qwhitting} The probability of finding $\ket{j}$ can be bounded as follows \begin{align} p(\tau,j) \geq p(\infty,j) - \sum_{E_k \neq E_{k^{\prime}}} \frac{2}{|E_k - E_{k^{\prime}}|\tau} |\langle j|E_k\rangle \langle E_k |\Phi(0)\rangle \langle \Phi(0) |E_{k^{\prime}}\rangle \langle E_{k^{\prime}}| j\rangle| \geq p(\infty,j) - \frac{2}{\Delta E \tau}. \end{align} As a result, for any $\epsilon< p(\infty,j)$, setting $\tau_{\epsilon} = 2/\Delta E\epsilon$, we have \begin{align} t_{\rm hit}(j) \leq \frac{\tau_{\epsilon}}{p(\tau_{\epsilon},j)} \Rightarrow t_{\rm hit}(j) = O\bigg(\frac{1/\Delta E\epsilon}{p(\infty,j) - \epsilon}\bigg). \label{eq:qwhitting} \end{align} \end{lemma} If $\epsilon$ in \eq{qwhitting} is small enough, $\tau_{\epsilon} = 2/\Delta E\epsilon$ permits a good mixing and we may write \begin{align} t_{\rm hit}(j) = O\bigg(\frac{1}{p(\infty,j)\Delta E\epsilon} \bigg), \end{align} which suggests we are using the mixing time to bound the hitting time. \begin{lemma}[Upper bound of the QTW hitting time]\label{lem:qtwhitting} Consider an bounded open set $\Omega_{j}$ containing only one well $U_j$,\footnote{To be rigorous, satisfy (\blue{S85}) in \append{A.2.3}.} we have \begin{align} \hspace{-1mm} \int_{\Omega_j}\rho_{\rm QTW}(\tau,x)\d x \geq \int_{\Omega_j}\mu_{\rm QTW}(x)\d x - \frac{2}{\Delta E \tau}(1+|O(h^{\infty})|) = p(\infty, j) + O(h^{\infty}) - \frac{2}{\Delta E \tau}(1+|O(h^{\infty})|). \label{eq:qtwhitting} \end{align} For any $\epsilon< \int_{\Omega_j}\mu_{\rm QTW}(x)\d x$, let $\tau_{\epsilon} = 2 (1+|O(h^{\infty})|)/\Delta E\epsilon$, we have \begin{align} T_{\rm hit}(\Omega_j) \leq \frac{\tau_{\epsilon}}{\int_{\Omega_j}\rho_{\rm QTW}(\tau_{\epsilon},x)\d x} \Rightarrow T_{\rm hit}(\Omega_j) = O\bigg(\frac{1}{\Delta E\epsilon} \frac{1+|O(h^{\infty})|}{\int_{\Omega_j}\mu_{\rm QTW}(x)\d x - \epsilon}\bigg). \end{align} \end{lemma} The upper bounds we have obtained on mixing and hitting time are still not explicit, as $H_{|\mathcal{F}}$ is not given. Next, we establish relationships between an objective landscape, the corresponding interaction matrix $H_{|\mathcal{F}}$, and the time cost of the QTW algorithm. This is a main task of the present paper. In later sections, we figure out major geometric properties that affect $H_{|\mathcal{F}}$ on specific landscapes. \subsection{Application: Tensor decomposition}\label{sec:tendecom} After giving the definition of QTW and studying its mixing and hitting time, now we use QTW to solve a practical problem, orthogonal tensor decomposition, which is a central problem in learning many latent variable models \cite{AGH+14}. Specifically, we consider a fourth-order tensor $T\in \mathbb{R}^{d^4}$ that has orthogonal decomposition: \begin{equation} T = \sum_{i=1}^d a_j^{\otimes 4}, \label{eq:4-tensor} \end{equation} where the components $\{a_j \}$ form an orthonormal basis of a $d$-dimensional space ($a_j^{\top} a_j = \delta_{ij}$). The goal of orthogonal tensor decomposition is to find all components $\{a_j \}$. Following previous popular methods \cite{CLX+09,Hyv99}, we try to find all components by a single optimization problem. Concretely, we consider the following landscape \cite{FJK96}:\footnote{The original objective function used in previous papers including \cite{FJK96} is $T(u,u,u,u)$. The function in \eq{TenDecObj} is designed such that $\min f = 0$.} \begin{equation} f(u) = 1-T(u,u,u,u) = 1- \sum_{i=1}^d (u^{\top} a_j)^4,~\|u\|_2^2 = 1. \label{eq:TenDecObj} \end{equation} Without loss of generality, we work in the coordinate system specified by $\{a_j\}_{j=1}^{d}$. In particular, let $u = \sum_j^d x_j a_j$ and $x= (x_1,\ldots,x_d)$, we obtain $f(x) = 1-\sum_{i=1}^d x_i^4$. Later, we also use $a_j$ to denote the vector $(0,\ldots,1,\ldots,0)$ where the only nonzero coordinate with value 1 appears in the $j$th entry. $f(x)$ has $2d$ local minima $\pm a_1,\ldots\pm a_d$ uniformly distributed on the $d$-dimensional sphere $\mathbb{S}^{d-1}$.\footnote{Here, we need to consider quantum simulation on the manifold $\mathbb{S}^{d-1}$ which should cost the same quantum queries as quantum simulation on $\mathbb{R}^{d}$ under the same evolution time (see discussions in \append{B.1}).} Therefore, finding all the minima solves the orthogonal tensor decomposition problem. \begin{figure} \centerline{ \includegraphics[width=0.4\textwidth]{TenDeclabel.pdf}} \caption{The landscape given by \eq{TenDecObj} for dimension $d=3$: local minima $a_1$, $a_2$, and $-a_3$ are highlighted by red points ({\color{red}$\bullet$}) and corresponding labels $\alpha = (j(\alpha), \pi_{\alpha})$ are shown. } \label{fig:TenDeclabel} \end{figure} The problem of tensor decomposition has notable symmetries. As a result, the objective function $f$ \eq{TenDecObj} is nonconvex, and we can apply QTW to such a landscape. We can use the pair $\alpha = (j,\pi_{\alpha})$ to denote the local minima and corresponding wells, where $j = j(\alpha)$ (i.e. $j(\cdot)$ is a function) refers to $a_j$ and $\pi_{\alpha} \in \{\pm 1\}$ specifies whether it is $+ a_j$ or $- a_j$. \fig{TenDeclabel} shows the landscape $f$ for $d =3$, where some minima are labeled. The local ground state of the well $(j,\pi_{\alpha})$ is denoted by $|j,\pi_{\alpha}\rangle$. In the basis $\{|1,+\rangle,\ldots,|d,+\rangle,|1,-\rangle,\ldots,|d,-\rangle\}$ which spans a low-energy subspace $\mathcal{F}$, the interaction matrix modulus an exponential error has the form \begin{equation} H_{|\mathcal{F}} = \left( \begin{array}{cccc|cccc} \mu & w & \cdots & w & 0 & w & \cdots & w \\ w & \mu & \ddots & \vdots & w & 0 & \ddots & \vdots \\ \vdots &\ddots & \ddots& w & \vdots & \ddots &\ddots &w \\ w & \cdots& w & \mu & w & \cdots& w & 0 \\ \hline 0 & w & \cdots & w & \mu & w & \cdots & w\\ w & 0 &\ddots & \vdots & w & \mu & \ddots & \vdots\\ \vdots & \ddots&\ddots & w & \vdots & \ddots & \ddots& w\\ w & \cdots& w & 0 & w & \cdots& w & \mu \\ \end{array} \right). \label{eq:TenDec} \end{equation} Here the quantity $\mu$ stands for the energy of local ground states, and $w$ is the tunneling amplitude quantifying the interactions between wells. To understand \eq{TenDec}, for any $j$, imagine a sphere where $(j,+)$ is the north pole and $(j,-)$ the south pole, then for $j'\neq j$, $(j',\pm)$ are evenly distributed on the equator. The energy of all local ground states are the same because of the symmetry. So, diagonal elements of $H_{|\mathcal{F}}$ are all $\mu$. The interactions between $(j,+)$ and $(j',\pm)$ for all $j'\neq j$ should be the same as well. However, the interaction between $(j,+)$ and $(j,-)$ is exponentially weaker due to the longer distance between $(j,+)$ and $(j,-)$. As a result, we write $\bra{j,+}H_{|\mathcal{F}}\ket{j,-} = 0$ modulus an exponential error and all other off-diagonal elements as $w$. As is demonstrated in previous sections, the time cost of QTW highly depends on spectral gaps of $H_{|\mathcal{F}}$. The following lemma studies eigenstates and eigenvalues of $H_{|\mathcal{F}}$ (see proof in \append{C.3.1}). \begin{lemma}\label{lem:TenDecEig} The eigenstates and corresponding eigenvalues of $H_{|\mathcal{F}}$ in \eq{TenDec} are given by \begin{align} |E_k\rangle &= \frac{1}{\sqrt{2d}}\sum_j e^{i\frac{2\pi}{d}kj}(|j,+\rangle+ |j,-\rangle),~k=1,\ldots,d \\ |E_k\rangle &= \frac{1}{\sqrt{2}}(|k-d,+\rangle - |k-d,-\rangle),~k=d+1,\ldots,2d \end{align} where \begin{align} E_k &= \mu - 2w,~k=1,\ldots,d-1 \\ E_d &= \mu + 2w(d-1), \\ E_k &= \mu,~k=d+1,\ldots,2d. \end{align} \end{lemma} Evolving the system for at least the mixing time, the measured results would be subject to the limit distribution $\mu_{\rm QTW}$. Since $\mu_{\rm QTW}$ would concentrate near all minima, we are able to find all components. Combined with results of \sec{Qmixing}, we can get the mixing time of the QTW as follows (the proof is postponed to \append{C.3.2}): \begin{lemma}\label{lem:TenDecmix} For the landscape \eq{TenDecObj}, starting from a local ground state $|\alpha\rangle$, the distribution $\rho_{\rm QTW}(\tau,x)$ converges to the limiting distribution $\mu_{\rm QTW}$ obeying the following relation: \begin{align} \|\rho_{\rm QTW} - \mu_{\rm QTW}\|_{L^1(\mathbb{S}^{d-1})} \leq \frac{1}{|w| \tau}(\Theta(1/2)+ O(h^{\infty})). \label{eq:TenDecmix} \end{align} The $\epsilon$-close mixing time is subsequently bounded as \begin{align} T_{\rm mix} = O\bigg(\frac{1}{|w|\epsilon}(1+O(h^{\infty}))\bigg). \end{align} \end{lemma} To determine the total evolution time for finding all components $\{a_j\}$ (i.e., all global minima), we need to calculate $\int_{\Omega_{\beta}}\mu_{\rm QTW} \d x$, where $\Omega_{\beta}$ is an open set containing the minimum $\beta$. According to \lem{qtwhitting}, $\int_{\Omega_{\beta}}\mu_{\rm QTW} \d x$ is the probability of finding the particle in a neighborhood of $\pi_{\beta} a_{j(\beta)}$ and can be captured by the probability of finding the system at the state $\ket{\beta}$. Starting from a local state $|\alpha\rangle$, the probability of hitting $\ket{\beta}$ is given by the following lemma and the proof can be found in \append{C.3.3}. \begin{lemma}\label{lem:TenDecprobability} Initiating at a local state $|\alpha\rangle$ where $\alpha = (j(\alpha), \pi_{\alpha})$, after simulating for a time $t$ which is chosen uniformly from $[0,\tau],~\tau \to \infty$, the limiting distribution represented by the probability of tunneling to a local state $|\beta\rangle$ is given by \begin{equation} p(\infty,\beta | \alpha ) =\left\{\begin{array}{ll} \frac{1}{2d^2},~j(\alpha) \neq j(\beta),\\[3pt] \frac{1}{2}-\frac{(d-1)}{2d^2},~j(\alpha) = j(\beta). \end{array} \right. \end{equation} \end{lemma} Note that the two minima $\pm a_{j}$ are equivalent, representing one component. Thus, starting from $\ket{\alpha}$, we are able to find a component different from $\pm a_{j(\alpha)}$ if the measured result is in a well $\beta$ where $j(\beta) \neq j(\alpha)$. We can define the probability for a successful trial as $p_{\rm suc} = \sum_{j(\beta)\neq j(\alpha)} p(\infty,\beta | \alpha ) = \frac{d-1}{d^2}$. That is, evolving for time $T_{\rm mix}$ as described by \lem{TenDecmix}, we are able to approximately sample from the limiting distribution $\mu_{\rm QTW}$ and then get to another component with probability near $p_{\rm suc}$. The number of trials needed for finding another component is approximately $1/p_{\rm suc}$. And the time needed for finding another component from a known component is approximately $T_{\rm mix}/p_{\rm suc}$. Repeating the procedure of looking for one component that is different from a known one, we can obtain all orthogonal components with total time\footnote{The term $O(d\log d)$ appears because our procedure is equivalent to the Coupon Collector's Problem.} \begin{align} T_{\rm tot} = O(d\log d) T_{\rm mix}/p_{\rm suc} = \tilde{O}(d^2)\frac{1}{\epsilon|w|}. \label{eq:Ttot1} \end{align} To determine the time specifically, it remains to determine $1/|w|$ which depends exponentially on $d$ and $h$. We can obtain: \begin{lemma}\label{lem:TenDecw} For sufficiently small $h$, the tunneling amplitude $w$ in the interaction matrix \eq{TenDec} satisfies \begin{align} w = - \sqrt{h}(C_1 C_2^{d-1} + O(h))e^{-\frac{\sqrt{2}}{2h}}, \end{align} where $C_1$ and $C_2$ are constants depending only on the landscape and are independent of the dimension $d$. \end{lemma} The proof of \lem{TenDecw} is in \append{C.3.4}. It is intuitive to see from \lem{TenDecw} that the smaller $h$ is, the longer time it takes to find all components. However, small $h$ permits more accurate measurement results. A successful tunneling means we can find a point near a new component, but this point may not be the actual minimum. We add a constraint that the expected risk is $\delta$ (i.e., $\mathbb{E}_{x\sim \mu_{\rm QTW}}f(x) -\min f = \delta$). Subsequently, $h$ can be bounded using $\delta$ and we can have the following proposition: \begin{proposition}\label{prop:TenDecTtot} For sufficiently small $\epsilon$ (such that the measured positions nearly obey $\mu_{\rm QTW}$) and sufficiently small expected risk $\delta$ (such that $h$ can be estimated by $\delta$ and \lem{TenDecw} is valid), we have $h = \sqrt{2}\delta/(d-1 + o_{\delta}(1))$ and the total time for finding all orthogonal components of $T$ in \eq{4-tensor} by QTW satisfies \begin{align} T_{\rm tot} = O(\mathrm{poly}(1/\delta, e^d, 1/\epsilon)) e^{\frac{(d-1) +o_{\delta}(1)}{2\delta}}. \label{eq:TenDecTtot} \end{align} \end{proposition} \begin{remark} The strategy we adopt here, which is equivalent to repeating sampling from $\mu_{\rm QTW}$, is straightforward but may not be the optimal one under the framework of QTW. In other words, \prop{TenDecTtot} provides a general upper bound on the total evolution time needed. However, the term $e^{\frac{\sqrt{2}}{2h}}$ which gives the term $e^{\frac{d-1}{2\delta} + o_{\delta}(1)}$ in \eq{TenDecTtot} describes essential difficulty for tunneling through a barrier and would not disappear as long as we use quantum tunneling. \end{remark} To sum up, we provide a scenario that QTW can be used to solve orthogonal tensor decomposition problems. For a practical landscape, the spectrum of the interaction matrix and the mixing time of QTW is explicitly calculated. Running QTW for some time (bounded by the mixing time) repeatedly, we can sample points from a distribution near the limiting distribution and find all tensor components, and an upper bound on the total running time for QTW is derived. \section{Comparison Between Quantum Tunneling Walks and Classical Algorithms}\label{sec:comparison} In this section, we use comparisons between QTW and SGD to explain the advantages of quantum tunneling, resulting in our \hyperref[slog:1]{Main Message}. Because of distinctions between quantum and classical algorithms, preparations (i.e., standards for comparisons) in \sec{standard} are needed before specific comparisons in \sec{illustration}. Having such general understanding of QTW, in \sec{separation}, we further make use of the fact that quantum evolution is essentially global but classical algorithms rely on local queries, so that a hitting problem cannot be solved efficiently by classical algorithms can be tackled by QTW within polynomial queries when given reasonable initial states. \subsection{Criteria of fair comparison}\label{sec:standard} Through out \sec{comparison}, we adopt assumptions in both \sec{classicalpre} and \sec{quantumpre} for the objective landscape $f(x)$ of interest. We still use $U_j = \{x_j \}~(j=1,\ldots,N)$ to denote the wells and $\ket{j}$ the corresponding orthonormalized local ground states. The interaction matrix is $H_{|\mathcal{F}}$, where $H=-h^2\Delta + f(x)$ is the Hamiltonian and $\mathcal{F}$ the low energy subspace spanned by $\{ \ket{j}: j=1,\ldots,N\}$. As shown in \sec{onedimexp}, the hitting time of SGD is determined by the landscape and an adjustable learning rate $s$. Similarly, we can also adjust $h$ in Hamiltonian simulation. Therefore, we need to determine the relationship between $h$ and $s$ for the comparison between the time cost of QTW and SGD. Note that both QTW and SGD have limit distributions, namely, $\mu_{\rm QTW}$ and $\mu_{\rm SGD}$, respectively (see \sec{classicalpre} and \lem{limitdis} for details). If $h$ (or $s$) becomes smaller, $\mu_{\rm QTW}$ ($\mu_{\rm SGD}$) will concentrate more closely to global minima, giving more accurate outputs, whereas it would take more time for the QTW (SGD) to converge. Comparing the running time without specifying accuracy is not fair. In order to establish an relationship between $h$ and $s$, as well as to compare QTW and SGD fairly, we specify some kind of accuracy of the limit distributions. The two variables, $h$ and $s$, will be solved from the demand of accuracy. Hence, the time cost of different algorithms are only related to the accuracy, the dimension, and some geometric properties of the landscapes. There are different measures of accuracy we can choose depending on the tasks faced. Here, we introduce two kinds of measures along with the corresponding standards of comparison. \begin{standard}[Risk accuracy]\label{stand:risk} Let $\mu_{\rm QTW}$ be the limit distribution of QTW, and $\mu_{\rm SGD}$ the invariant Gibbs distribution of SGD. Two distributions are demanded to be $\delta$-risk-accurate: \begin{align} \mathbb{E}_{x\sim \mu_{\rm QTW}}f(x) - \min f = \mathbb{E}_{x\sim \mu_{\rm SGD}}f(x) - \min f = \delta. \label{eq:stand1} \end{align} \end{standard} \stand{risk} ensures that two limit distributions yield the same expected risk. Then, it is natural to compare how fast QTW and SGD would converge. The algorithm spending less time is more efficient on finding any one global minimum. Sometimes, the task is to find some target minima or one special minimum. In this case, using risk accuracy cannot emphasize the particularity of the minima of interest and we may need the following standard: \begin{standard}[Distance accuracy]\label{stand:distance} Let $\mu_{\rm QTW}$ be the limit distribution of QTW, and $\mu_{\rm SGD}$ be the invariant Gibbs distribution of SGD. The minima of interest are $x_{j_k},~k = 1,\ldots,m,~j_k \in \{1,\ldots,N\}$. Let $D(\cdot, \cdot)$ be any distance function. Two distributions are demanded to be $\delta$-distance-accurate with respect to $x_{j_k}$ and $D(\cdot, \cdot)$: \begin{align} \mathbb{E}_{x\sim \mu_{\rm QTW}}\sum_k D(x,x_{j_k}) = \mathbb{E}_{x\sim \mu_{\rm SGD}}\sum_k D(x,x_{j_k}) = \delta. \label{eq:stand2} \end{align} \end{standard} Conditions \eq{stand1} and \eq{stand2} can specify $h$ and $s$. To see this, we first study the expected risk for quadratic functions: \begin{lemma}\label{lem:stand1-quad} Assume the objective function $f\colon \mathbb{R}^d \to \mathbb{R}$ is quadratic and \begin{equation} f(0) = 0,~\nabla f(0) = 0,~\nabla^2 f(0) > 0, \end{equation} where the last inequality means the Hessian $\nabla^2 f(0)$ is positive definite. Then, we have \begin{equation} \mathbb{E}_{x\sim \mu_{\rm QTW}}f(x) = \frac{\sqrt{2}h}{4}\mathrm{tr} \sqrt{\nabla^2 f(0)}, \qquad \mathbb{E}_{x\sim \mu_{\rm SGD}}f(x) = \frac{sd}{4}. \end{equation} \end{lemma} \lem{stand1-quad} calculates the expected risks for a landscape with only one minimum whose proof is in \append{D.1.1}. For landscapes with multiple minima, the limit distributions concentrate near the global minima and the objective function in a small neighborhood of any minimum can be approximated by a quadratic function based on the assumptions. Hence, we can obtain the following general estimations (the proof is postponed to \append{D.1.2}). \begin{lemma}\label{lem:expectedrisk} If $\delta$ is sufficiently small and the objective function $f\colon \mathbb{R}^d \to \mathbb{R}$ satisfies satisfying assumptions in \sec{classicalpre} and \sec{quantumpre}, then \stand{risk} gives \begin{equation} h = \frac{\delta}{\frac{\sqrt{2}}{4}\sum_{j=1}^N p(\infty,j)\mathrm{tr}\big(\sqrt{\nabla^2f(x_j)}\big)+ o_{\delta}(1)}, \label{eq:hstand1} \end{equation} \begin{equation} s = \frac{\delta}{\frac{d}{4}(1+ o_{\delta}(1))}. \end{equation} \end{lemma} That is, we establish a relationship between $h$ and $s$ by \stand{risk}. Similarly, for \stand{distance}, we can have the following result: \begin{lemma}\label{lem:expecteddis} Assume the objective function $f\colon \mathbb{R}^d \to \mathbb{R}$ is quadratic and \begin{equation} f(0) = 0,~\nabla f(0) = 0,~\nabla^2 f(0) > 0, \end{equation} where the last inequality means the Hessian $\nabla^2 f(0)$ is positive definite. We define the distance function $D(x,y):=\|x-y\|^2_2,~\forall x,y\in \mathbb{R}^d$. Then, we have \begin{align} \mathbb{E}_{x\sim \mu_{\rm QTW}}D(x,0) &= \frac{\sqrt{2}h}{2}\mathrm{tr} (\nabla^2 f(0))^{-1/2}; \\ \mathbb{E}_{x\sim \mu_{\rm SGD}}D(x,0) &=\frac{s}{2}\mathrm{tr}(\nabla^2 f(0))^{-1}. \end{align} \end{lemma} The proof of \lem{expecteddis} is shown in \append{D.1.3}. Similar to the process from \lem{stand1-quad} to \lem{expectedrisk}, \lem{expecteddis} may be generalized to general landscapes. However, the generalization of \lem{expecteddis} is quite complicated as the distance function and the wells of interest are arbitrary. So, we stop at \lem{expecteddis}. Regardless of different standards, \lem{expectedrisk} and \lem{expecteddis} present some similar intuition: the dependence of $h$ on the flatness of wells are different from that of $s$, which is going to be shown in the following section as a source of quantum speedups.\footnote{Here, we use the Hessian matrix of $f$ at minima to quantify the concept ``flatness".} \subsection{Illustrating advantages of quantum tunneling}\label{sec:illustration} In this subsection, we compare QTW with SGD for several special landscapes. The goal is to explore geometric properties of the landscapes that affect relative efficiencies of QTW and SGD. Heuristically, the comparison reveals when quantum tunneling can be faster than thermal climbing (climbing over barriers between minima by stochastic motions), which are the two mechanisms behind QTW and many classical algorithms. For simplicity, we focus on the following kind of landscapes: \begin{definition}[One-dimensional partially periodic functions]\label{defn:funcforillus} A function $f\colon\mathbb{R}\to \mathbb{R}$ is partially periodic if it satisfies the assumptions in \sec{classicalpre} and \sec{quantumpre}, and all minima $\{x_j:j=1,\ldots,N\}$ are in a bounded interval which is a period of $f$. \end{definition} \begin{figure} \centerline{ \includegraphics[width=0.8\textwidth]{oneperiodic.pdf}} \caption{A one-dimensional partially periodic function. } \label{fig:oneperiodic} \end{figure} A sketch of functions in \defn{funcforillus} is shown in \fig{oneperiodic}. Neglect an exponentially small error and note the symmetry of the one-dimensional partially periodic function $f$, the interaction matrix under $\{ \ket{j}:j=1,\ldots,N\}$ should be given by \begin{equation} H_{|\mathcal{F}} = \left( \begin{array}{cccccc} \mu & w & & & & \\ w & \mu & w & & & \\ & w & \mu & w & & \\ & & \ddots& \ddots & \ddots & \\ & & & w & \mu & w\\ & & & & w & \mu\\ \end{array} \right), \label{eq:Hforline} \end{equation} where $\mu$ is the energy of one local ground state and $w$ quantifies the tunneling effect between two adjacent wells. Eigenstates and eigenvalues of $H_{|\mathcal{F}}$ can be given by the following lemma. \begin{lemma}\label{lem:IlluEig} The eigenstates and corresponding eigenvalues of the Hamiltonian \eq{Hforline} are given by \begin{align} |E_k\rangle &= \sqrt{\frac{2}{N+1}}\sum_{j=1}^{N}\sin\big(\frac{jk\pi}{N+1}\big)\ket{j},~k=1,2,\ldots,N; \\ E_k &= \mu + 2w \cos \frac{k\pi}{N+1},~k=1,2,\ldots,N. \end{align} \end{lemma} To describe $w$ in detail, as shown in \fig{oneperiodic}, we introduce new notations $\{x^{\bullet}_j: j=1,\ldots,N\}$ and $\{x^{\circ}_j: j=1,\ldots,N-1\}$ to denote minima and saddle points, respectively. A more general labeling of local minima and saddle points can be found in \append{A.1}. The Morse saddle barrier reflecting height of the barrier in the present case can be given by $H_f = f(x^{\circ}_1) - f(x^{\bullet}_1)$. Using results in \append{A.2.3}, we have: \begin{lemma}[Tunneling amplitude] The tunneling amplitude for the one-dimensional partially periodic function $f$ is given by \begin{align} w = -2\sqrt{\frac{h f''(x^{\bullet}_1) H_f}{\sqrt{2}\pi}} e^{\int_{x_1^{\bullet}}^{x_1^{\circ}} (\sqrt{\frac{ f''(x^{\bullet}_1)}{2f(\xi)}} - \frac{1}{\xi- x_1^{\bullet}})\d \xi } e^{-\frac{S_0}{h}},~\mathrm{where}~S_0 = \int_{x_1^{\bullet}}^{x_2^{\bullet}} \sqrt{f(\xi)} \d\xi. \end{align} \end{lemma} Now, we can obtain the spectrum of $H_{|\mathcal{F}}$ explicitly, and proceed by using \lem{qtwmixingtime} to get the quantum mixing time. \begin{lemma}[Quantum mixing time]\label{lem:illuqtwmixing} Staring from one local ground state of one minimum, the $\epsilon$-close mixing time of QTW is given by \begin{align} T_{\rm mix}^{\rm QTW} = O \bigg(\frac{N^3}{\epsilon |w|} [1+(N-1)|O(h^{\infty})|] \bigg) = O(\mathrm{poly}(N,1/h,1/\epsilon))e^{\frac{S_0}{h}}. \end{align} \end{lemma} Regarding SGD, we use the results introduced in \sec{classicalpre} to estimate the classical mixing time. First, \begin{lemma}[Exponential decay constant] In \prop{mic19}, let $\lambda = \delta_{s,1}/2s$ we have \begin{align} \lambda = \bigg(\frac{\sqrt{f''(x^{\circ})f''(x^{\bullet})}}{2\pi} + o(s)\bigg) e^{-\frac{2 H_f}{s}}. \end{align} \end{lemma} Then, by \cor{sgdmixing}, the following lemma holds. \begin{lemma}[Classical mixing time]\label{lem:illusgdmixing} Let $T_{\rm mix}^{\rm SGD}$ be the SGD $\epsilon$-close mixing time which is the minimum time enabling $\|\rho_{\rm SGD} (t,\cdot) - \mu_{\rm SGD}\|_{\mu_{\rm SGD}^{-1}}<\epsilon$, we have \begin{align} T_{\rm mix}^{\rm SGD} = O \bigg(\frac{1}{\lambda} \ln\frac{\|\rho(0,\cdot) - \mu_{\rm SGD}(\cdot) \|_{\mu^{-1}_{\rm SGD}}}{\epsilon} \bigg) = O(\mathrm{poly}(1/s,\ln(1/\epsilon)))e^{\frac{2H_f}{s}}. \end{align} \end{lemma} Later, we do not focus on the dependence of the mixing time on $\epsilon$, as the norms ($L^1$ norm for QTW and $L^2(\mu^{-1}_{\rm SGD})$ for SGD) used to capture convergence are different.\footnote{ In terms of $\epsilon$, the same argument in \cite{AC21} but with evolution time $t$ of QTW chosen as a sum of some random variables instead of chosen uniformly in an interval, $T_{\rm mix}^{\rm QTW}$ can also achieve $\ln(1/\epsilon)$ dependence instead of $1/\epsilon$.} The dominant terms affecting running time of QTW and SGD are $e^{\frac{S_0}{h}}$ and $e^{\frac{2H_f}{s}}$. \begin{lemma}[Comparison on one-dimensional periodic landscapes] Under \stand{risk}, let QTW and SGD be both $\delta$-accurate. For sufficiently small $\delta$, the QTW mixing time and SGD mixing time are dominated by \begin{align} LT_{\rm mix}^{\rm QTW}:=e^{\frac{\sqrt{2}S_0f''(x^{\bullet})}{4\delta}}~\mathrm{and}~LT_{\rm mix}^{\rm SGD}:=e^{\frac{H_f}{2\delta}}, \textrm{ respectively.} \end{align} \end{lemma} \begin{figure} \centerline{ \includegraphics[width=0.5\textwidth]{illuexamples.pdf}} \caption{The landscapes in \examp{critical}, \examp{flatness}, and \examp{sharp} (corresponding to Example 1, 2, and 3 in the figure, respectively) for illustrating the comparison between QTW and SGD. } \label{fig:illuexamples} \end{figure} As concrete examples, we present several specific functions to illustrate the advantages of quantum tunneling. Since the function in the region of our interest is periodic, we only need to specify the function value within one period to construct a concrete example. Without loss of generality, we set the interval $[-a,a+2b]$ to be one period, where $[-a,a]$ is called the well region and $[a,a+2b]$ the barrier region. The constructed landscape in $[-a,a+2b]$ is given by \begin{align} f(x) = \left\{ \begin{array}{ll} \frac{1}{2}k x^2\quad x\in [-a,a],\\[3pt] \frac{1}{2\pi \sigma^2}\exp\big(-\frac{(x-a-b)^2}{\sigma^2}\big) + \frac{1}{2}k a^2 - \varepsilon\quad x \in (a,a+2b]. \end{array} \right. \label{eq:illuVx} \end{align} Here, to reduce free parameters, we make $f(x)$ differentiable at $a$, the boundary of the well, and the barrier, such that \begin{align} b = \sigma \sqrt{\ln \frac{1}{2\pi \sigma^2 \varepsilon}}, \quad k = \frac{2\varepsilon}{a\sigma}\sqrt{\ln \frac{1}{2\pi \sigma^2 \varepsilon}}. \label{eq:illucondition} \end{align} \begin{remark} Note that the function in \eq{illuVx} is not smooth. We need to use the mollifier function $m_{r}$ (see detials in \append{D.3}) to smooth it such that assumptions in \sec{classicalpre} and \sec{quantumpre} are satisfied. Note that if $r\to 0$, the smoothed function will tend to be $f$, following results can be seen as arbitrarily accurate for a smooth function arbitrarily close to \eq{illuVx}. \end{remark} By giving specific $a$, $\sigma$, and $\varepsilon$ in \eq{illuVx}, we can design landscapes with different properties. Detailed variables, discussions and comparisons are given below. \begin{example}[Critical case]\label{examp:critical} For \eq{illuVx}, we set $a = 5.0$, $\sigma = 1.0$ and $\varepsilon=0.15$, and obtain $b\approx 0.243$ and $k\approx 0.0146$ by \eq{illucondition}. \end{example} \begin{example}[Flatness of minima]\label{examp:flatness} For \eq{illuVx}, we set $a = 5.0$, $\sigma = 1.0$ and $\varepsilon=0.009$, and obtain $b\approx 1.69$ and $k\approx 0.00610$ by \eq{illucondition}. \end{example} \begin{example}[Sharpness of barriers]\label{examp:sharp} For \eq{illuVx}, we set $a = 5.0$, $\sigma = 0.5$ and $\varepsilon=0.0088$, and obtain $b\approx 1.03$ and $k\approx 0.0146$ by \eq{illucondition}. \end{example} \fig{illuexamples} explicitly shows the shapes of above examples. The barrier region in \examp{critical} is small and most of the function in one period is quadratic, which is similar to the case introduced in \sec{onedimexp}. The Morse saddle barrier $H_f$ of \examp{flatness} is approximately equal to that of \examp{critical}, whereas, in \examp{flatness}, the well is more flat and the barrier is thicker. \examp{sharp} has almost the same well as \examp{critical} but is equipped with a much higher barrier. We call \examp{critical} as the critical case because QTW and SGD perform nearly the same on it in terms of the leading terms $LT_{\rm mix}^{\rm QTW}$ and $LT_{\rm mix}^{\rm SGD}$: \begin{lemma} \examp{critical} satisfies $b\ll a$ and $\frac{1}{2}ka^2 \approx H_f$. For such a landscape, we have \begin{align} \ln LT_{\rm mix}^{\rm QTW} &= \frac{H_f}{2\delta}\left[1+ 2b/a+ o(b/a) \right] + o(\delta),\\ \ln LT_{\rm mix}^{\rm SGD} &= \frac{H_f}{2\delta}(1+o(\delta)). \end{align} \end{lemma} QTW mixes faster on both \examp{flatness} and \examp{sharp} for sufficiently small $\delta$. Specifically, we have \begin{lemma} For \examp{flatness} and \examp{sharp}, the following holds \begin{align} \ln LT_{\rm mix}^{\rm QTW} &< \frac{k}{4\delta}a^2+ \frac{k}{2\delta}ab + \frac{\sqrt{2k}}{4\delta} + o(\delta),\\ \ln LT_{\rm mix}^{\rm SGD} &= \frac{1}{4\pi \sigma^2\delta}+\frac{k}{4\delta}a^2-\frac{\epsilon}{2\delta} + o(\delta). \end{align} Substituting the parameters, it is true for both \examp{flatness} and \examp{sharp} that \begin{align} \frac{k}{4}a^2+ \frac{k}{2}ab + \frac{\sqrt{2k}}{4} < \frac{1}{4\pi\sigma^2}+\frac{k}{4}a^2-\frac{\epsilon}{2}. \end{align} \end{lemma} Comparing to the critical case \examp{critical}, \examp{flatness} has a thicker barrier, which increases $S_0$ and causes difficulty for QTW. However, QTW can perform better in \examp{flatness}. This is mainly due to the more flat well of \examp{flatness}. Recall that by \stand{risk}, to ensure $\delta$-risk-accuracy, $h$ and $s$ should be \begin{align} h= \frac{\delta}{\frac{\sqrt{2k}}{4} + o_{\delta}(1)} \quad\mathrm{and}\quad s= \frac{\delta}{\frac{1}{4}(1+o_{\delta}(1))}, \end{align} respectively. That is, under the same risk accuracy, $h$ can be much larger than $s$ if the well is flat ($k$ is small), making tunneling easier. Note that there is a trade-off between accuracy and time cost: smaller $h$ (or $s$) ensures high accuracy but make tunneling effects (or thermal diffusion) weaker; conversely, larger $h$ (or $s$) permits faster tunneling (or diffusion) but yields inaccurate results. Discussions on quantum tunneling effects usually focus on properties of the barrier. In the present study, since we aim to find global minima, the precision of results obtained is one important concern. Therefore, the flatness of wells, which affects differently on the accuracy of QTW and SGD, is a crucial property determining the runtime of QTW and SGD. Loosely speaking, QTW is faster than SGD on landscapes with flat wells. \examp{sharp} adheres to the intuition that quantum tunneling is efficient on functions with tall and thin barriers. The wells of \examp{sharp} are almost the same as those of the critical case \examp{critical}. QTW can be faster in \examp{sharp} because we add a sharp barrier between wells. By \lem{illusgdmixing}, a high barrier (i.e., large $H_f$) would significantly hinder thermal climbing. However, the tall barrier is sufficiently thin, such that $S_0 = 2\int_{0}^{a+b}\sqrt{f(x)}\d x$ can still be small and by \lem{illuqtwmixing}, the tunneling effect would be strong. Moreover, in high dimensions, the distribution of wells can be very different from being on a line. As shown in \append{D.4}, distribution of wells can largely affect the dependence of time on $N$. However, such relation between the distribution of wells and running time is not explicitly shown for SGD. Therefore, the distribution of wells can also be a factor of quantum speedups. In summary, we can conclude our \hyperref[slog:1]{Main Message}. \subsection{Efficient quantum tunneling for solving a classically difficult hitting problem}\label{sec:separation} The above examples compare QTW driven by quantum tunneling with SGD. In this section, an exponential separation in terms of query complexity between QTW given initial states and classical algorithms knowing one well will be shown for a specific hitting problem on a constructed landscape. The landscape $f(\mathbf{x})$ we construct lives in $\mathbb{R}^d$. We use $\|\cdot\|$ to denote the $\ell_2$ norm of vectors, namely, $\|\mathbf{x}\| = \sqrt{\mathbf{x}\cdot \mathbf{x}}$. Let $\mathbb{B}(\mathbf{x},r)$ denote a $d$-dimensional ball centered at $\mathbf{x}$ with radius $r$. A special direction $\mathbf{v}$ is randomly chosen from the $d$-dimensional unit sphere. We define two regions $W_- = \mathbb{B}(\mathbf{0},a)$ and $W_+ = \mathbb{B}(2b\mathbf{v},a)$ with $b\geq a$. Let $R$ be sufficiently large s.t. $W_-, W_+ \subset \mathbb{B}(\mathbf{0},R)$. We denote the region $\{\mathbf{x}\mid\mathbf{x} \in \mathbb{B}(\mathbf{0},R), ~|\mathbf{x}\cdot \mathbf{v}| \leq w\}$ by $S_{\mathbf{v}}$, where $w$ will be chosen from $[\sqrt{3}a/2,0)$. We denote \begin{align} B_{\mathbf{v}}:=\{\mathbf{x}\mid w< \mathbf{x}\cdot \mathbf{v} < 2b-w,~\sqrt{\|\mathbf{x}\|^2 - (\mathbf{x}\cdot \mathbf{v})^2} < \sqrt{a^2-w^2}, \bf{x}\notin W_-\cup W_+\}. \end{align} \fig{provable-acceleration} illustrates positions of the newly defined regions. The constructed function $f$ is given by \begin{align} f(\mathbf{x}) = \left\{ \begin{array}{ll} \frac{1}{2}\omega^2 \|\mathbf{x}\|^2,~\mathbf{x} \in W_-, \\[3pt] \frac{1}{2}\omega^2 \|\mathbf{x}-2b\mathbf{v}\|^2,~\mathbf{x} \in W_+, \\[3pt] H_1,~\mathbf{x} \in B_{\mathbf{v}},\\[3pt] H_2,~\mathrm{otherwise}. \end{array} \right. \label{eq:hardinstance} \end{align} Here, we define $H_0 = \frac{1}{2}\omega^2a^2$ and demand that $0<H_0 \sim H_1 \ll H_2$. \begin{remark} The landscape $f$ in \eq{hardinstance} is not smooth and should be smoothed to be $F_r$ with the help of a mollifier function $m_r$ (see details in \append{D.3}) such that assumptions in \sec{quantumpre} can be satisfied. Because when $r\to 0$, $F_r\to f_r$, we can always find sufficiently small $r$ to make the following conclusions based on $f$ valid for $F_r$. \end{remark} There are two global minima, $\mathbf{0}$ and $2b\mathbf{v}$, of the function $f$. Given that we know $\mathbf{0}$ is a minimum, our goal is to find the other one. To avoid complicated justifications, we deal with a simpler problem: \begin{problem}\label{prb:provable} For the $f$ in \eq{hardinstance}, given that we only know $\mathbf{0}$ is a global minimum, find any point in $W_+$. \end{problem} \subsubsection{Classical lower bound}\label{sec:clb} Due to the concentration of measure, for any point $\mathbf{x} \in \mathbb{B}(\mathbf{0},R)$, the probability of $\mathbf{x}\in S_{\mathbf{v}}$ is given by \begin{align} P(\mathbf{x} \in S_{\mathbf{v}}) \geq 1 - O \big(e^{-\frac{dw^2}{2R^2}}\big). \label{eq:PxinSv} \end{align} Intuitively, restricted in $\mathbb{B}(\mathbf{0},R)$, any classical algorithm cannot escape from $S_{\mathbf{v}}$ efficiently. In $\mathbb{R}^d$, queries out of $\mathbb{B}(\mathbf{0},R)$ provide no information about the landscape inside $\mathbb{B}(\mathbf{0},R)$ and are unable to help to escape from $S_{\mathbf{v}}$. Therefore, classical algorithms cannot solve \prb{provable} efficiently with or without being constrained in $\mathbb{B}(\mathbf{0},R)$. To rigorously prove above intuitions, we first introduce a mathematical result indicating \eq{PxinSv}: \begin{lemma}[Measure concentration for the sphere]\label{lem:mconcentration} Let $\mathbb{S}^{d-1} = \{\mathbf{x}:\|\mathbf{x}\| =1 \}$ be the unit sphere in $\mathbb{R}^d$. Let $\mathrm{Cap}(\epsilon)$ denote the spherical cap of height $\epsilon$ above the origin (see the left part of \fig{capandcone}). We have \begin{align} \frac{\mathrm{Area}(\mathrm{Cap}(\epsilon))}{\mathrm{Area}(\mathbb{S}^{d-1})} \leq e^{-d\epsilon^2/2}. \end{align} \end{lemma} \begin{figure} \centerline{ \includegraphics[width=0.5\textwidth]{capandcone.pdf}} \caption{Estimating the area of a spherical cap. } \label{fig:capandcone} \end{figure} The estimation details are presented in \append{D.2.1}. Subsequently, it is readily to have (see details in \append{D.2.2}): \begin{lemma}\label{lem:provable-onep} For any randomly chosen point $\mathbf{x}\in \mathbb{B}(\mathbf{0},R)$, the probability of $\mathbf{x}\notin S_{\mathbf{v}}$ is $P(\mathbf{x}\notin S_{\mathbf{v}})\leq 2 e^{-\frac{dw^2}{2R^2}}$. \end{lemma} Recall that $w\in [a/2,a)$ and $R$ are independent of $d$, the measure of the region in $\mathbb{B}(\mathbf{0},R)$ and outside $S_{\mathbf{v}}$ is exponentially small with respect to the dimension $d$. By \defn{localquery}, classical algorithms depend on an adaptive sequence of points. we now need to demonstrate that it is difficult for the points to hit regions beyond $S_{\mathbf{v}}$. \begin{lemma}\label{lem:provable-manyp} For any classical algorithm (see \defn{localquery}), after running $T$ times, we get a sequence of points and corresponding queries $\{\mathbf{x}_i, q(\mathbf{x}_i)\}_{i=1}^T$. Restricted in $\mathbb{B}(\mathbf{0},R)$, as long as any $q(\mathbf{x}_i)~(\mathbf{x}\in S_{\mathbf{v}})$ is independent of $\mathbf{v}$, the probability $P(\exists t\leq T:\mathbf{x}_t\notin S_{\mathbf{v}})\leq 2T e^{-\frac{dw^2}{2R^2}}$. \end{lemma} We prove \lem{provable-manyp} in \append{D.2.3}. Now, we can prove that if the number of points and queries is small, with high probability, any classical algorithm cannot escape from $S_{\mathbf{v}}$. Rigorously, we have \begin{proposition}[Classical lower bound]\label{prop:provable-exp} Any classical algorithm (\defn{localquery}) will fail, with high probability, to solve \prb{provable} given only $o(e^{\frac{dw^2}{4R^2}})$ local queries with or without being restricted in $\mathbb{B}(\mathbf{0},R)$. \end{proposition} The proof sketch of \prop{provable-exp} goes as follows (see proof details in \append{D.2.4}). By \lem{provable-manyp}, it suffices to demonstrate that restricted in the ball $\mathbb{B}$, classical algorithms cannot escape from $W_-$ and hit $W_+$ efficiently. The left thing is to show that queries outside $\mathbb{B}(\mathbf{0},R)$ provide no information about $\mathbb{B}(\mathbf{0},R)$. And thus, without being restricted in $\mathbb{B}(\mathbf{0},R)$, classical algorithms still cannot hit $W_+$ by subexponential queries with high probability. \subsubsection{Quantum upper bound}\label{sec:qub} We now focus on the time needed for quantum tunneling to solve \prb{provable}. The landscape \eq{hardinstance} satisfies \blue{Assumption A.6} \begin{align} 0 = \min f < \lim_{\|\mathbf{x}\|\to \infty} f = H_2,\quad f^{-1}(0) = \{ \mathbf{0}\} \cup \{2b\mathbf{v}\}, \end{align} where $U_-:=\{ \mathbf{0}\}$ and $U_+ := \{2b\mathbf{v}\}$ are called as wells by definition. The neighborhoods of the two wells are quadratic, enabling the wells and corresponding local ground states to satisfy (\blue{S79}) and (\blue{S80}). Moreover, due to the symmetry of the function \eq{hardinstance}, the local ground states are also symmetric. Therefore, \blue{Assumption A.5} can be satisfied. To use \blue{Proposition A.6}, we only need to verify the conditions in \blue{Assumption A.7}, leading to the following lemma. \begin{lemma}\label{lem:provableS0} There exists a unique Agmon geodesic, denoted $\boldsymbol{\gamma}_{-+}: \mathbb{R} \to \mathbb{R}^d$, which links $U_-$ and $U_+$: \begin{equation} \boldsymbol{\gamma}_{-+}(s) = s \mathbf{v},\quad s\in [0,2b]. \label{eq:provable-geodesic} \end{equation} And the Agmon distance $S_0 := d(U_-, U_+)$ is \begin{equation} S_0 = \int_{-1/2}^{1/2} \sqrt{f(\boldsymbol{\gamma}_{-+}(s))} \d s = \frac{1}{\sqrt{2}}\omega a^2 + 2(b-a)\sqrt{H_1}. \label{eq:provable-S0} \end{equation} \end{lemma} The calculation details of \lem{provableS0} are presented in \append{D.2.5}. We are now ready to calculate the interaction matrix explicitly (see details in \append{D.2.6}): \begin{lemma}\label{lem:provablenu} Under the two orthonormalized local ground states, The interaction matrix is of the form \begin{equation} P = \left( \begin{array}{cc} \mu & \nu\\ \nu & \mu \end{array} \label{eq:provable-matrix} \right), \end{equation} and the next-to-leading order formula of $w$ is given by \begin{align} \nu = -\sqrt{\frac{2h}{\pi}} \sqrt{\frac{H_1(\sqrt{2}\omega)^d}{(4\sqrt{H_1}/b)^{d-1}}}\exp\left(-\frac{S_0}{h} + \frac{\omega d(b-a)}{\sqrt{2H_1}} - 2d \ln \frac{b}{a}\right). \label{eq:provable-w} \end{align} \end{lemma} Using the explicit tunneling amplitude, we can estimate the time needed for quantum tunneling. \begin{proposition}[Quantum upper bound]\label{prop:provable-poly} For any dimension $d$, we can always choose appropriate $h$, $\omega$, $a$, $b$, $H_1$, $H_2$, and $w$ satisfying previous restrictions, such that, given the local ground state associated to $W_-$ under the choosing $h$ as initial state, QTW can solve \prb{provable} with high probability $1-(1-C)^n$ using only $nO(\mathrm{poly}(d))$ queries, where $0<C<1$ is a constant independent of $d$. \end{proposition} \begin{remark} In \prop{provable-poly}, the constant $C$ can be understood as the probability of successful hitting in one trial and $n$ the number of trails. To reach a high probability of success, say $99$\%, the number of trials needed, $M$, enabling $1-(1-C)^M\geq 99$\%, is a constant independent of $d$. Since one trial needs only one initial state, only a constant number of copies (e.g., $M$ copies) of the local ground state are needed. \end{remark} The proof of \prop{provable-poly} is postponed to \append{D.2.7} which is explained briefly as follows. We take all the adjustable parameters as functions of $d$ and discuss the evolution time as a function of $d$. First, we have $h= \Theta(1/d)$ for sufficiently large $d$, which can eliminate the negative effects of measure concentration brought by increasing dimension, and on the other hand we prove that our theory on quantum tunneling walks is still valid. Thus, the quantum wave distributes near $W_-$ or $W_+$, and the limit distribution $\mu_{\rm QTW}$ permits a probability of finding the particle in $W_+$ larger than some constant independent of $d$. Then, based on the results of semi-classical analysis, we can tune the function values in $W_-$, $W_+$, and $B_{\mathbf{v}}$ such that the time needed for tunneling is a polynomial of $d$. As a result, the last three conditions at the end of \sec{intro} can be satisfied, and the first and third conditions suggest that with high probability, QTW can hit $W_+$ with queries polynomial in $d$. Finally, given the fact that we can use $\tilde{O}(t)$ quantum queries to evolve QTW for time $t$, with high probability QTW can hit $W_+$ with queries polynomial in $d$. Combining the results of \prop{provable-poly} and \prop{provable-exp}, we can obtain \thm{provableinformal}, which is restated in a more rigorous way as follows. \begin{theorem}\label{thm:thm2re} For any dimension $d$, there exists a landscape with the form \eq{hardinstance} such that with a high probability $1-(1-A)^n$, QTW can solve \prb{provable} with $nO(\mathrm{poly}(d))$ queries given the local ground state associated to $U_-$, but with a high probability $1-e^{-dB}$, no classical algorithm (\defn{localquery}) can solve the same problem for the same landscape $f$ with $o(e^{dB})$ queries, where $0<A<1$ and $B>0$ are two constants independent of $d$. \end{theorem} \subsubsection{The significance of proper initial states}\label{sec:no-exp-sep-init} The hardness of \prb{provable} can be abstracted as that of finding an exponentially small cone on a landscape which is isotropic outside the special cone.\footnote{Specifically, this cone is a region associated to the special direction $\bf{v}$, $\{\bf{x}:\bf{x}\cdot \bf{v}/\|\bf{x}\|<C \}$ for some constant $C$, which contains the parts $B_{\bf{v}}$ and $W_+$.} There can be exponentially many such cones disjoint with each other. Therefore, it can be proved that by solving \prb{provable} in $\mathbb{R}^d$, we can solve an unstructured search problem with a size exponential in $d$. To show this, we first introduce unstructured search. Say, we are given $N$ data points, only one of which is assigned the value $1$ and all other points are assigned a value $0$. The goal is to find the point assigned $1$ with an oracle outputting the assigned value of the input point. Intuitively, each data point can be mapped to a unique cone in $\mathbb{R}^d$ and the point assigned $1$ should correspond to the cone containing $B_{\bf{v}}$ and $W_+$. In this case, solving \prb{provable} can lead to the data point we want to find. Precisely speaking, if there is a quantum algorithm that can solve \prb{provable} with queries polynomial in $d$, it can solve an unstructured search whose size $N$ is exponential in $d$ within queries polynomial in $d$. That is, we can solve an $N$-size unstructured search within $O(\mathrm{poly}(\log N))$ queries with the help of the efficient algorithm for \prb{provable}. However, it is well known that quantum algorithms have a query complexity lower bound $\Omega(\sqrt{N})$ in solving unstructured search with $N$ data points \cite{BBBV97}. Therefore, we can conclude \begin{proposition}\label{prop:nospeedup} No quantum algorithm can solve \prb{provable} within queries polynomial in $d$. \end{proposition} We prove \prop{nospeedup} and related claims rigorously in \append{D.2.8}. It seems that \prop{nospeedup} contradicts with \prop{provable-poly}. But there is actually no paradox as in \prop{provable-poly} QTW does not solve \prb{provable} faithfully. The local ground state $\ket{\Phi_-}$ associated to $U_-$ under proper quantum learning rate $h$ is given to QTW as prior knowledge. To establish polynomial decay tunneling effect, the state $\ket{\Phi_-}$ has non-vanishing probability (maybe an inverse polynomial of $d$) in $W_+$. The state $\ket{\Phi_-}$ indicates a lot about the special direction $\bf{v}$ for QTW, such that what QTW does cannot be equivalent to unstructured search. Indeed, by the same spirit of \prop{nospeedup}, the state $\ket{\Phi_-}$ cannot be prepared within polynomial queries, or we can reach $W_+$ efficiently by measuring $\ket{\Phi_-}$ repeatedly. We admit that \thm{thm2re} requires the initial quantum state. Note that QTW only uses $M$ copies of the local ground state $\ket{\Phi_-}$ to hit $W_+$ with high probability in polynomial time, where $M$ is a number independent of $d$. If the possibility of learning about $\bf{v}$ from sampling tends to $0$ when $d\to \infty$, which is likely to be true, the expected queries needed by classical algorithms to hit $W_+$ cannot be subexponential in $d$. In this case, we have an exponential quantum-classical separation in evaluation queries even classical algorithms are given a constant number of samples from the initial distribution $|\ip{\bf{x}}{\Phi_-}|^2$. Essentially, this is because no classical evolution can make good use of the samples of the initial state. \section{Numerical Experiments}\label{sec:num} We conduct numerical experiments to examine our theory. All results and plots are obtained by simulations on a classical computer (Dual-Core Intel Core i5 Processor, 16GB memory) via MATLAB 2020b. Details of all numerical settings can be found in \append{E}. QTW is simulated by solving the Schr\"odinger equation by numerical methods, and SGD is performed with first-order queries and the noise of each step follows the standard Gaussian distribution. \paragraph{Quantum-classical comparisons.} To corroborate our \hyperref[slog:1]{Main Message}, we numerically study the performance of QTW and SGD on concrete examples (see details in \append{E.1}). The quantum learning rate $h$ and the classical learning rate $s$ are determined under \stand{risk} which equalizes expected risks yielded by QTW and SGD. The task is to hit a target neighborhood of one minimum beginning at another designated minimum. In \fig{histogramsExp123}, results on classical and quantum hitting time are shown. We examine QTW and SGD on three landscapes, Example 1, 2, and 3 in \fig{histogramsExp123} which correspond to concrete functions given by \examp{critical}, \examp{flatness}, and \examp{sharp} in \sec{illustration}, respectively. We conduct 1000 experiments for QTW and SGD on each example. For QTW, we use an \emph{experiment} to denote a process repeating trials until successfully hitting, where each \emph{trial} initiates QTW once and measures the position at $t$ randomly chosen from $[0,\tau]$. For SGD, an experiment begins at a designated minimum and stops until SGD hits the target region. The evolution time of an experiment is the sum of evolution time of the trials the experiment contains. We use $T^{\rm QTW}_{\rm hit}$ and $T^{\rm SGD}_{\rm hit}$ to denote the evolution time of one experiment for QTW and SGD, respectively. In \fig{histogramsExp123}, the histograms compare $T^{\rm QTW}_{\rm hit}$ with $T^{\rm SGD}_{\rm hit}/10$, and all presented examples demonstrate that QTW is faster. The number of quantum queries is approximately $\tilde{O}(\|f\|_{\infty} T^{\rm QTW}_{\rm hit})$ and the number of classical queries is $\Omega(T^{\rm SGD}_{\rm hit}/s)$. In addition, in the three examples $\|f\|_{\infty} \leq 0.85$ and $s<0.2$, and quantum advantage exists in terms of query complexity. This result matches our theory at large. For Example 1, we make direct comparison between the exponential terms $e^{S_0/h}$ and $e^{2H_f/s}$, and to remove the coefficients in front of them, we divide $T^{\rm SGD}_{\rm hit}$ by $10$ such that $T^{\rm SGD}_{\rm hit}/10$ has similar distribution t $T^{\rm QTW}_{\rm hit}$ for Example 1. In this way, we observe that whether $T^{\rm SGD}_{\rm hit}/10$ is relatively larger than $T^{\rm QTW}_{\rm hit}$ is determined only by $e^{S_0/h}$ and $e^{2H_f/s}$. For Example 2, $T^{\rm QTW}_{\rm hit}$ is not much smaller than $T^{\rm SGD}_{\rm hit}/10$, which is not completely coherent with our theory. This result can be explained as that for Example 2, the quantum learning rate $h$ is not small enough such that the initial state prepared does not well stay near a low energy subspace. Experiments on Example 2 suggest that higher energy may not be able to help quantum tunneling to run faster. For Example 3, the $h$ chosen is small enough (see details in \append{E.1}) such that a significant quantum speedup is achieved as expected. In \fig{histogramsExp123}, $T^{\rm SGD}_{\rm hit}/10$ is even several orders of magnitude larger than $T^{\rm QTW}_{\rm hit}$. \begin{figure} \centering \includegraphics[width=\linewidth]{histogramsExp123.pdf} \caption{Quantum-classical comparison between SGD and QTW on three landscapes. Example 1 is the critical case where the exponential terms in QTW and SGD evolution time are equal for sufficiently small accuracy $\delta$. Example 2 has flatter minima but similar barriers compared to Example 1, enabling QTW to be faster. Example 3 possesses the same flatness of minima as Example 1 but is equipped with sharp but thin barriers, enabling larger quantum speedups. We take $\tau=288, 800, 600$ in the three examples, respectively.} \label{fig:histogramsExp123} \end{figure} \paragraph{Dimension dependence.} Due to the limitations of solving the Schr\"odinger equation on classical computers, QTW is simulated only in low dimensions (i.e., $d=1$ and $d =2$). Here we examine \thm{provableinformal} by testing SGD and its the classical lower bound. The classical lower bound in \thm{provableinformal} ensures that for any $s$, SGD cannot cannot escape from $S_{\mathbf{v}}$ with subexponential queries with high probability. Based on the constructed landscape with parameters specified in \append{E.2}, we test SGD with different learning rates ($s\in [0.1,1]$) in various dimensions ($d \in [15,95]$). For each dimension and each $s$, 1000 experiments are conducted. The number of steps spent to escaping from $S_{\mathbf{v}}$ in one experiment is denoted as $Q_{\rm esc}$. Here, we present the relationship between average $Q_{\rm esc}$ and the dimension $d$ in \fig{dimdepenmean} (more details are deferred to \append{E.2}). For each fixed learning rate $s$, we observe that with the increase of $d$, the average $Q_{\rm esc}$ remains constant initially and then increase exponentially with respect to $d$. Increasing $s$ yields a smaller initial constant but larger exponential rate. Nevertheless, for all $s$, $Q_{\rm esc}$ eventually increases exponentially with respect to $d$ (the triangle in \fig{dimdepenmean} shows the slope $1/256$ corresponding to the exponential function $e^{d/256}$ which is a lower bound of the average $Q_{\rm esc}$), supporting our prediction. \begin{figure} \centering \begin{minipage}[t]{0.46\linewidth} \centering \includegraphics[width = \linewidth]{dimdepenmean.pdf} \caption{Relationship between the average $Q_{\rm esc}$ and the dimension $d$ under different $s$. The dashed line captures the lower bound of the average $Q_{\rm esc}$.} \label{fig:dimdepenmean} \end{minipage} \hspace{6mm} \begin{minipage}[t]{0.46\linewidth} \centering \includegraphics[width =\linewidth]{qlrdepend.pdf} \caption{$T_{\rm half}$ with respect to $h$ on a fixed landscape: theoretical prediction (red solid line) and time in experiments (blue circles).} \label{fig:qlrdepend} \end{minipage} \end{figure} \paragraph{Dependence on the quantum learning rate.} In QTW, the quantum learning rate $h$ is one of the most important variables. \thm{informalQTW} gives a general relationship between $h$ and the evolution time of QTW. We further test the relationship on the landscape constructed in \thm{provableinformal} (dimension $d=2$) with specified parameters given in \append{E.3}. Since the landscape has two symmetric wells, the time for tunneling from one well to the other, $T_{\rm half}$, is explicitly linked to $\Delta E$ (i.e., $T_{\rm half} = \pi/\Delta E$). On this concrete landscape, $\Delta E$ can be predicted, giving that \begin{align} \ln T_{\rm half} = \frac{S_0}{h} - \frac{1}{2}\ln h + C_f, \end{align} where $C_f$ is a constant depending on $f$ and can be explicitly calculated. Starting from one well, we stop when the probability of finding the other well exceeds 90\% and record the evolution time as $T_{\rm half}$. Experiments on $T_{\rm half}$ is shown in \fig{qlrdepend}. The results match our theory except a constant difference between the predicted and experimental $\ln T_{\rm half}$, indicating the correctness of $\frac{S_0}{h} - \frac{1}{2}\ln h$. The constant difference emerges because we stop evolution when the probability of tunneling exceeds 90\%, while theoretical $T_{\rm half}$ takes the time when the probability is nearly 100\%. To conclude, several aspects of the present theory are well supported by numerical experiments. \section{Discussion}\label{sec:discussion} In this paper, we explore quantum speedups for nonconvex optimization by quantum tunneling. In particular, we introduce the quantum tunneling walk (QTW) and apply it to nonconvex problems where local minima are approximately global minima. We show that QTW achieves quantum speedup over classical stochastic gradient descents (SGD) when the barriers between different local minima are high but thin and the minima are flat. Moreover, we construct a specific nonconvex landscape where QTW given proper initial states is exponentially faster than classical algorithms taking local queries for hitting the neighborhood of a target global minimum. Finally, we conduct numerical experiments to corroborate our theoretical results. We expect our results to have further impacts both classically and quantumly. In optimization theory, previous work has studied several physics-motivated optimization algorithms, including Nesterov's momentum method~\cite{su2016differential,wibisono2016variational,shi2021understanding}, stochastic gradient descents~\cite{SSJ20}, symplectic optimization~\cite{betancourt2018symplectic,jordan2018dynamical}, etc. We believe that our work can further inspire the design of optimization algorithms. From theory to practice, in this work we analyzed the performance of QTW on tensor decomposition, and we expect QTW to also have decent performance on other practical problems with benign landscapes. In quantum computing, on the one hand, previous work on continuous optimization only studies convex optimization~\cite{vanApeldoorn2020optimization,chakrabarti2020optimization} or local properties such as escaping from saddle points~\cite{zhang2021quantum}, and our work significantly extends the range of problems which quantum computers can efficiently solve to global problems in nonconvex optimization. On the other hand, we point out that QTW has the potential to be implemented on near-term quantum computers. In fact, current quantum computers have implemented both quantum simulation~\cite{arute2020hartree,ebadi2021quantum} and quantum walks~\cite{tang2018experimental,gong2021quantum} to decent scales. We deem QTW as a potential proposal for demonstrating quantum advantages in near term. Our paper also leaves several technical questions for future investigation: \begin{itemize} \item What is the performance of QTW on more general landscapes? For instance, a wide range of deep neural networks~\cite{KHK19} has some (but probably not all) local minima which are approximately global. Future work on weakening the assumptions on landscapes for QTW is preferred. \item Are there more examples with exponential quantum-classical separation? Our construction leverages a special kind of locally non-informative landscapes, and exponential quantum-classical separation can potentially be observed on other landscapes, such as nonsmooth landscapes and landscapes with negative curvature. \item QTW simulates the Schr\"odinger equation whose potential is set to be the optimization function, and this QTW can be efficiently simulated on quantum computers. In general, are there better PDEs which are more efficient for optimization and can still be efficiently simulated on quantum computers? \end{itemize} \section*{Acknowledgements} YL thanks Frédéric Hérau and Michael Hitrik for helping with understanding spectral theory, Rong Ge and Chenyi Zhang for inspiring discussions on tensor decomposition, and Jiaqi Leng for the guidance on numerical methods for quantum simulation. TL thanks Eric R. Anschuetz for general suggestions on a preliminary version of this paper, and Andrew M. Childs for helpful discussions that inspire \sec{no-exp-sep-init}. YL was funded by SURF from Tsien Excellence in Engineering Program, Tsinghua University. WJS was funded by an Alfred Sloan Research Fellowship and the Wharton Dean’s Research Fund. TL was funded by a startup fund from Peking University, and the Advanced Institute of Information Technology, Peking University. \bibliographystyle{alphaUrlePrint} \section{Introduction}\label{sec:intro} Nonconvex optimization plays a central role in machine learning because the training of many modern machine learning models, especially those from deep learning, requires optimization of nonconvex loss functions. Among algorithms for solving nonconvex optimization problems, stochastic gradient descent (SGD) and its variants, such as Adam~\cite{ADAM}, Adagrad~\cite{duchi2011adagrad}, etc., are widely used in practice. In theory, their provable guarantee has been studied from various perspectives. In this paper, we adopt the perspective of studying gradient descents via the analysis of their behavior in continuous-time limits as differential equations, following a recent line of work in~\cite{su2016differential,wibisono2016variational,jordan2018dynamical,shi2021understanding}. In particular, given a function $f\colon\mathbb{R}^{d}\to\mathbb{R}$, the SGD $x_{k+1} = x_k - s\nabla f(x_k) - s\xi_{k}$ with learning rate $s$ and the $k$th step noise $\xi_{k}$ can be approximated by a stochastic differential equation (SDE) as follows: \begin{align}\label{eq:SDE} \d x = -\nabla f(x)\d t + \sqrt{s}\d W, \end{align} where $W$ is a standard Brownian motion. Such approach enjoys clear intuition from physics. In particular, Eq.~\eq{SDE} is essentially a non-equilibrium thermodynamic process: gradient descent provides driving forces, the stochastic term serves as thermal motions, and a combination of these two ingredients enables convergence to the thermal distribution, also known as the Gibbs distribution. A systematic study of Eq.~\eq{SDE} was conducted in a recent work by~\cite{SSJ20}. See more details in \sec{classicalpre}. Nevertheless, algorithms based on gradient descents also have limitations because they only have access to local information about the function, which suffers from fundamental difficulties when facing landscapes with intricate local structures such as vanishing gradient~\cite{hochreiter1998vanishing}, nonsmoothness~\cite{KS21}, negative curvature~\cite{CB21}, etc. In terms of optimization, we are mostly interested in points with zero gradients, and they can be categorized as \emph{saddle points}, \emph{local optima}, and \emph{global optima}. It is known that variants of SGD can escape from saddle points \cite{ge2015escaping,jin2017escape,allen2018neon2,fang2018spider,fang2019sharp,jin2019stochastic,zhang2021escape}, but one of the most prominent issues in nonconvex optimization is to escape from local minima and reach global minima. Up to now, theoretical guarantee of escaping from local minima by SGD has only been known for some special nonconvex functions~\cite{kleinberg2018alternative}. In general, SGD has to climb through high barriers in landscapes to reach global minima, and this is typically intractable using only gradients that descend the function. In all, fundamentally different ideas, especially those that explores beyond local information, are expected to derive better algorithms for nonconvex optimization in general. This paper aims to study nonconvex optimization via dynamics from \emph{quantum mechanics}, which can leverage global information about a function $f\colon\mathbb{R}^{d}\to\mathbb{R}$. The fundamental rule in quantum mechanics is the \emph{Schr{\"o}dinger Equation:}\footnote{The standard Schr{\"o}dinger Equation in quantum mechanics is typically written as $i \hbar\frac{\partial}{\partial t} \Phi = \left(-\frac{\hbar^2}{2m}\Delta + f(x)\right)\Phi$. In this paper, we use the form in \eq{Schrodingereq} by setting the Planck constant $\hbar = 1$ and $h = \hbar/\sqrt{2m}$ which is a variable. See also \sec{quantumpre}.} \begin{align} i \frac{\partial}{\partial t} \Phi = \left(-h^2\Delta + f(x)\right) \Phi, \label{eq:Schrodingereq} \end{align} where $i$ is the imaginary unit, $h$ is defined as the \emph{quantum learning rate}, $\Delta=\sum_{j=1}^{d}\frac{\partial^2}{\partial x_{j}^2}$ is the Laplacian, and $\Phi(t,x)\colon\mathbb{R}\times\mathbb{R}^{d}\to\mathbb{C}$ is a quantum wave function satisfying $\int_{\mathbb{R}^{d}}|\Phi(t,x)|^2\d x=1$ for any $t$. Measuring the wave function at time $t$, $|\Phi(t,x)|^2$ is the probability density of finding the particle at position $x$. In Eq.~\eq{Schrodingereq}, the time evolution of wave functions is governed by the Hamiltonian\footnote{In this paper, we refer Hamiltonian to either the total energy of a system or the operator corresponding to the total energy of the system, depending on the context.} $H:=-h^2\Delta + f$, where $-h^2\Delta$ corresponds to the classical kinetic energy and $f$ the potential energy. In sharp contrast to classical particles, quantum wave functions can tunnel through high potential barriers with significant probability, and this is formally known as \emph{quantum tunneling.} Take a one-dimensional double-well potential $f\colon\mathbb{R}\to\mathbb{R}$ in \fig{1-dimexample} as an example, the goal is to move from the local minimum $x_{-}$ in the left region to the local minimum $x_{+}$ in the right region. Classically, the SDE in \eq{SDE} has to climb through the barrier with height $H_{f}$, and it can take $\exp(\Theta(H_{f}/s))$ time to reach $x_{+}$ (see Section 3.4 of \cite{SSJ20}). Quantumly, we denote $\Phi_-(x)$ and $\Phi_+(x)$ to be the ground state (i.e., the eigenstate corresponds to the smallest eigenvalue) of the left and the right region, respectively. These states $\Phi_\pm(x)$ are localized near $x_{\pm}$, respectively. We let the wave function be initialized at $\Phi_-(x)$, i.e., $\Phi(0,x)=\Phi_-(x)$. Under proper conditions, two eigenfunctions with eigenvalues $E_0$ and $E_1$ of $H=-h^2\Delta + f$ can be represented by superposition states \begin{align} \Phi_0(x):=(\Phi_+(x) + \Phi_-(x))/\sqrt{2},\\ \Phi_1(x):=(\Phi_+(x) - \Phi_-(x))/\sqrt{2}, \end{align} respectively. Note that $\Phi_0(x)$ and $\Phi_1(x)$ are \emph{not} localized because they have probability $1/2$ of reaching both $x_{+}$ and $x_{-}$. Specifically, given $\Phi(0,x) = \Phi_-(x) = (\Phi_0(x) + \Phi_1(x))/\sqrt{2}$, and because the dynamics of the Schr\"odinger equation \eq{Schrodingereq} is $\Phi(t,x) = e^{-iHt}\Phi(0,x)$, we have \begin{align} \Phi(t,x) = (e^{-iE_0t}\Phi_0(x) + e^{-iE_1t}\Phi_1(x))/\sqrt{2}. \end{align} As a result, after time $t$ where $|E_0-E_1|t =\pi$, we have $\Phi(t,x)\propto \Phi_+(x)$ localized near $x_{+}$. Intuitively, this can be viewed as global evolution and superposition of quantum states, which is capable of acquiring global information of the function $f$ and explains why for various choices of $f$, this quantum evolution time $t$ is much shorter than the classical counterpart by SDE which only takes gradients locally. \begin{figure} \centering \includegraphics[width=0.5\textwidth]{1-dimexample.pdf} \caption{An example of quantum tunneling in a double-well optimization function.} \label{fig:1-dimexample} \end{figure} It is a natural intuition to design quantum algorithms using quantum tunneling. Previously, \cite{FGS94,MAL16,CH16,BZ18} studied the phenomenon of quantum tunneling in quantum annealing algorithms \cite{FGS94,FGG01}. However, most of these results studied Boolean functions, which is essentially different from continuous optimization. In addition, quantum annealing focused on ground state preparation instead of the dynamics for quantum tunneling. Up to now, it is in general unclear when we can design quantum algorithms for optimization by adopting quantum tunneling. Therefore, we ask: \begin{question}\label{ques:configuration} On what kind of landscapes can we design algorithms efficiently using quantum tunneling? \end{question} To answer this question, we need to figure out specifications of the quantum algorithm, such as the initialization of the quantum wave packet, the landscape's parameters, the measurement strategy, etc. The next question is to understand the advantage of quantum algorithms based on quantum tunneling. A main reason of studying quantum computing is because it can solve various problems with significant speedup compared to classical state-of-the-art algorithms. In optimization, prior quantum algorithms have been devoted to semidefinite programs~\cite{brandao2016quantum,vanApeldoorn2017quantum,vanApeldoorn2018SDP,brandao2017SDP}, convex optimization~\cite{vanApeldoorn2020optimization,chakrabarti2020optimization}, escaping from saddle points~\cite{zhang2021quantum}, polynomial optimization~\cite{rebentrost2019quantum,li2021optimizing}, finding negative curvature directions~\cite{zhang2019quantum}, etc., but quantum algorithms for nonconvex optimization with provable guarantee in general is widely open as far as we know. Here we ask: \begin{question}\label{ques:accelaration} When do algorithms based on quantum tunneling give rise to quantum speedups? \end{question} \paragraph{Contributions.} We systematically study quantum algorithms based on quantum tunneling for a wide range of nonconvex optimization problems. Throughout the paper, we consider benign nonconvex landscapes where \emph{local minima are (approximately) global minima.} We point out that many common nonconvex optimization problems indeed yield objective functions satisfying such benign behaviors, such as tensor decomposition~\cite{ge2015escaping,GM20}, matrix completion~\cite{ge2016matrix,ma2018implicit}, and dictionary learning~\cite{qu2019analysis}, etc. In general, nonconvex problems with discrete symmetry satisfy this assumption, see the surveys by~\cite{Ma21,ZQW21}. In this paper, we demonstrate the power of quantum computing for the following main problem: \begin{named}{Main Problem}\label{prb:main} On a landscape whose local minima are (approximately) global minima, starting from one local minimum, find all local minima with similar function values or find a certain target minimum. \end{named} Such a problem is crucial for understanding the \emph{generalization} property of nonconvex landscapes, and in general it also sheds light on nonconvex optimization. First, local minima with similar function values can have dramatically different generalization performance (see Section 6.2.3 of \cite{Sun19}), and solving this \ref{prb:main} can be viewed as a subsequent step of optimization for finding the minimum which generalizes the best. Second, \ref{prb:main} implies the mode connectivity of landscapes, which has been applied to understanding the loss surfaces of various machine learning models including neural networks both empirically~\cite{draxler2018essentially,garipov2018loss} and theoretically~\cite{kuditipudi2019explaining,nguyen2019connected,shevchenko2020landscape}. Third, nonconvex landscapes where the \ref{prb:main} can be efficiently solved can also lead to efficient Monte Carlo sampling, which can be even faster than optimization~\cite{ma2019sampling,talwar2019computational}. Landscapes whose local minima are (approximately) global significantly facilitate quantum tunneling. Roughly speaking, since the total energy during our quantum evolution \eq{Schrodingereq} is conserved, quantum tunneling can only efficiently send a state from one minimum to another minimum with similar values. As a conclusion, if the quantum wave function is initialized near a local minimum, we can focus on quantum tunneling between the local ground state of each well, i.e., the tunneling of the particle from the bottom of a well to that of another well. To avoid complicated discussions on the value of the quantum learning rate $h$, we further restrict ourselves to functions whose local minima are global, which would not provide less intuition. Now, an answer to \ques{configuration} can be given as follows: \begin{theorem}[Quantum tunneling walks, informal]\label{thm:informalQTW} On landscapes whose local minima are global minima, we have an algorithm called quantum tunneling walks (QTW) which initiates the simulation of Eq.~\eq{Schrodingereq} from the local ground state at a minimum, and measures the position at a time which is chosen uniformly from $[0,\tau]$. To solve the \ref{prb:main} we can take \begin{align}\label{eq:tau-informal} \tau=O(\mathrm{poly}(N)/\Delta E), \end{align} where $N$ is the number of global minima and $\Delta E$ is the minimal spectral gap of the Hamiltonian restricted in a low-energy subspace. For sufficiently small $h$, we have \begin{align}\label{eq:QTW} \Delta E = \sqrt{h}(b + O(h))e^{-\frac{S_0}{h}}, \end{align} where $b, S_0>0$ are constants that depend only on $f$. \end{theorem} Formal description of the QTW can be found in \sec{QTW}. Here we highlight two important properties of QTW: Quantum mixing time and quantum hitting time. \\\\ \emph{Quantum mixing time (\lem{qtwmixingtime} in \sec{Qmixing}).} Since quantum evolutions are unitary, QTW never converges, a fundamental distinction from SGD. Therefore, to study the mixing properties of QTW, we follow quantum walk literature~\cite{CCD+03} by employing the measurement strategy, where we measure at $t$ uniformly chosen from $[0,\tau]$. The measured results obey a distribution which is a function of $\tau$, and when $\tau\to +\infty$, the distribution tends to its limit, $\mu_{\rm QTW}$. Quantum mixing time is the minimal $\tau$ enabling us to sample from $\mu_{\rm QTW}$ up to some small error. Alternatively speaking, the mixing time evaluates how fast the distribution yielded by QTW converges. We prove that $\mu_{\rm QTW}$ concentrates near minima, so that sampling from $\mu_{\rm QTW}$ repeatedly can give positions of all minima. In addition, $\mu_{\rm QTW}$ gives the upper bound on $\tau$ in \eq{tau-informal}. \\\\ \emph{Quantum hitting time (\lem{qtwhitting} in \sec{Qhitting})}. Hitting time is the duration it takes to hit a target region (usually a neighborhood of some minimum). Quantum hitting time is the minimum evolution time needed for hitting the region of interest once. Despite this straightforward intuition, the formal definition of quantum hitting time is very different from that of classical hitting time. Intuitively, repeatedly sampling from $\mu_{\rm QTW}$ can ensure the hitting of neighborhoods of particular minima, and thus we can use the mixing time to bound the hitting time. In short, to solve the \ref{prb:main}, we bound the quantum mixing and hitting time to obtain \thm{informalQTW}. \\\\ The minimal spectral gap $\Delta E$ in \thm{informalQTW} is calculated in \append{A.2.3}. The quantity $S_0$ is called the \emph{minimal Agmon distance} between different wells, formally defined in \defn{Agmon-distance}, which is related to both the height and width of potential barriers. The smaller $h$ is, the closer the measured results are to the minima (i.e., the more accurate QTW is), but the longer evolution time the Schr{\"o}dinger equation takes. As an application of \thm{informalQTW} and a justification of the practicability of QTW, we show how to use QTW to solve the orthogonal tensor decomposition problem. This problem asks to find all orthogonal components of a tensor. After transforming into a single optimization problem \cite{CLX+09,Hyv99}, the aim is to find all global minima. We present below a bound on the time cost of QTW on decomposing fourth-order tensors and details can be found in \sec{tendecom}. \begin{proposition}[Tensor decomposition, informal version of \prop{TenDecTtot}]\label{prop:TenDecTtot-main} Let $d$ be the dimension of the components of the fourth-order tensor $T \in \mathbb{R}^{d^4}$ satisfying \eq{4-tensor}, $\delta$ be the expected risk yielded by the limit distribution $\mu_{\rm QTW}$, and $\epsilon$ be the maximum error between $\mu_{\rm QTW}$ and the actual obtained distribution (quantified by $L^1$ norm). For sufficiently small $\epsilon$ and sufficiently small $\delta$, the total time $T_{\rm tot}$ for finding all orthogonal components of $T$ by QTW satisfies \begin{align} T_{\rm tot} = O(\mathrm{poly}(1/\delta, e^d, 1/\epsilon)) e^{\frac{(d-1) +o_{\delta}(1)}{2\delta}}. \end{align} \end{proposition} Next, we explore the advantages of the quantum tunneling mechanism comparing QTW with SGD and shown by describing landscapes where QTW outperforms SGD. The time cost for SGD to converge to global minima is loosely $O(1/\lambda_{s})$ and \begin{align}\label{eq:SGD} \lambda_{s} = (a + o(s))e^{-\frac{2H_f}{s}} \end{align} by~\cite{SSJ20}. Here, $s$ is the step size or learning rate of SGD. The constants $a>0$ and $H_f>0$ depend only on $f$. Interestingly, running time of QTW and that of SGD have similar form. In \eq{QTW} and \eq{SGD}, there are exponential terms $e^{S_0/h}$ and $e^{2H_f/s}$, respectively. Intuitively, the quantity $H_f$ is the characteristic height of potential barriers, and the quantity $S_0$ depends on not only the height but also the width of potential barriers. For the one-dimensional example in \fig{1-dimexample}, \begin{align} H_{f}=\max_{\xi\in[x_{-},x_{+}]}f(\xi),\qquad S_{0}=\int_{x_{-}}^{x_{+}} \sqrt{f(\xi)} \d\xi. \end{align} (Proof details are given in \sec{onedimexp}.) Other terms in the bounds, $\mathrm{poly}(N)/\Delta E$ and $1/\lambda_{s}$, are referred to as polynomial coefficients. We make the following comparisons: \begin{figure} \centerline{ \includegraphics[width=0.94\textwidth]{abstract1.pdf}} \caption{Flowchart of the \hyperref[slog:1]{Main Message}. } \label{fig:abstract1} \end{figure} \begin{itemize} \item Regarding the exponential terms $S_0$ and $H_f$, tall barriers means that $H_f$ is large, whereas if the barriers are thin enough, $S_0$ can still be small. This is consistent with the long-standing intuition that tall and thin barriers are easy for tunneling but difficult for climbing \cite{CH16}. \item Regarding the polynomial coefficients, they are mainly influenced by the distribution or relative positions of the wells. We observe that a symmetric distribution of wells, which can make (the local ground state in) any one well interacts with (the local ground states in) other wells, may reduce the running time of QTW but has no explicit impact on SGD. \item Flatness of wells is another important factor that influences the running time of both QTW and SGD. We propose standards for comparison (see \sec{standard}), which studies their running time when reaching the same accuracy $\delta$. Same to the effect of $h$, a smaller learning rate $s$ permits more accurate outputs but makes SGD more time consuming. For sufficiently flat minima, $h$ is larger than $s$, leading to a smaller running time for QTW. \end{itemize} In summary, we illustrate above observations in \fig{abstract1} and conclude the following: \begin{mainmessage}[Advantages of the quantum tunneling mechanism, a summary of \sec{standard} and \sec{illustration}]\label{slog:1} \emph{On landscapes whose local minima are global minima, QTW outperforms SGD on solving the \ref{prb:main} if barriers of the landscape $f$ is high but thin, wells are distributed symmetrically, and global minima are flat.} \end{mainmessage} \begin{remark} As is indicated above, we compare the costs of QTW and SGD under the same accuracy $\delta$. We introduce two definitions of accuracy in \sec{standard}: \stand{risk} concerns the expected risk, and \stand{distance} concerns the expected distance to some minima. Mathematically, \stand{risk} and \stand{distance} establish a relationship between the quantum and classical learning rates $h$ and $s$, respectively, enabling direct comparisons. \end{remark} Having introduced the general performance of the quantum tunneling walk, we further investigate \ques{accelaration} on some specific scenarios of the \ref{prb:main}. We focus on comparison between query complexities, namely the classical query complexity to local information and the quantum query complexity to the evaluation oracle\footnote{Query complexity of QTW is directly linked to the evolution time, in particular, evolving QTW for time $t$ needs $\tilde{O}(t)$ queries to $U_f$ (see details in \append{B.1}). As a result, it suffices to analyze the evolution time of QTW. Nevertheless, we state the query complexities for direct comparison.} \begin{align}\label{eqn:quantumquery} \hspace{-1.5mm}U_f (\ket{x}\otimes \ket{z}) = \ket{x}\otimes\ket{f(x)+z}\ \ \forall x\in\mathbb{R}^{d},z\in\mathbb{R}. \end{align} This is the standard assumption in existing literature on quantum optimization algorithms~\cite{vanApeldoorn2020optimization,chakrabarti2020optimization,zhang2021quantum}. Note that the $U_f$ here is a unitary transformation and it allows superposed input states, i.e., for a state $\sum_{j=1}^m c_j \ket{x_j}$ where $m\in \mathbb{N}$, $c_j \in \mathbb{C}$, and $\sum_{j=1}^m|c_j|^2=1$, \begin{align} U_f\Big(\sum_{j=1}^m c_j \ket{x_j} \otimes \ket{0}\Big) = \sum_{j=1}^m c_j \ket{x_j} \otimes \ket{f(x_j)}. \end{align} If we measure this output state, with probability $|c_{j}|^{2}$ we obtain $f(x_{j})$. The distribution can not only be sampled from a discrete set but also a continuous set. Different from classical queries that only learn local information of the landscape of $f$, quantum evaluation queries are essentially nonlocal as they can extract information of $f$ at different locations in superposition. Based on this fundamental difference, we are able to prove that QTW can solve a variant of the \ref{prb:main} with exponentially fewer queries than any classical counterparts: \begin{theorem}\label{thm:provableinformal} For any dimension $d$, there exists a landscape $f\colon \mathbb{R}^d \to \mathbb{R}$ such that its local minima are global minima, and on which, with high probability, QTW can hit the neighborhood of an unknown global minimum from the local ground state associated to a known minimum using queries polynomial in $d$, while no classical algorithm knowing the same minimum can hit the same target region with queries subexponential in $d$. \end{theorem} \begin{figure} \centerline{ \includegraphics[width=0.83\textwidth]{provable-acceleration3.pdf}} \caption{Sketch of the function in \thm{provableinformal}. The left figure explains the construction in the domain and the right figure plots the landscape of a two-dimensional example. } \label{fig:provable-acceleration} \end{figure} Details of \thm{provableinformal} are presented in \sec{separation}. Following similar idea to~\cite{JLG+18}, our construction relies on locally non-informative regions. Main structures of the constructed landscape are illustrated in \fig{provable-acceleration}, which has two global minima. $W_-$ and $W_+$ are two symmetric wells containing one global minimum respectively, $B_{\mathbf{v}}$ is a plateau connecting $W_-$ and $W_+$, and other places form a much higher plateau. The region $W_+$ is our target. We show that the landscape satisfies the following properties: \begin{itemize} \item $S_{\mathbf{v}}$, which is a band $\{\bf{x}\mid\bf{x}\in \mathbb{B}, |\bf{x}\cdot \bf{v}|\leq w\}$ with $w$ a constant and $\bf{v}$ a unit vector, occupies dominating measure in the ball $\mathbb{B}$. \item In $S_{\mathbf{v}}$, local queries (see \defn{localquery}) do not reveal information about the direction $\mathbf{v}$. \item Local queries outside $\mathbb{B}$ do not reveal information about the region inside $\mathbb{B}$. \end{itemize} Restricted in the ball $\mathbb{B}$, the first two properties make classical algorithms intractable to escape from $S_{\mathbf{v}}$ and thus cannot hit $W_+$ efficiently. The last property ensures that, without being restricted in $\mathbb{B}$, classical algorithms are still unable to hit $W_+$ efficiently. See \sec{clb} for details. Nevertheless, quantum tunneling can be efficient if we carefully design the function values and the parameter $h$. The design of the parameters should establish the following main conditions (See \sec{qub} for details): \begin{itemize} \item The wave function always concentrates in $W_-$ or $W_+$. \item The quantum learning rate $h$ is small such that our theory based on semi-classical analysis is valid. \item Quantum tunneling from $W_-$ to $W_+$ is always easy (can happen within time polynomial in $d$). \end{itemize} \paragraph{Organization.} \sec{prelim} introduces our assumptions and problem settings, both classical and quantum. In \sec{quantumal}, we explore QTW in details and state the formal version of \thm{informalQTW}. This includes a one-dimensional example, the formal definition of QTW, the mixing and hitting time of QTW, and the example on tensor decomposition (\prop{TenDecTtot-main}). \sec{comparison} covers detailed quantum-classical comparisons. First, we introduce fair criteria of the comparison. Second, we illustrate the advantages of quantum tunneling and give a detailed view of our \hyperref[slog:1]{Main Message}. Third, we prove \thm{provableinformal}. We corroborate our findings with numerical experiments in \sec{num}. At last, the paper is concluded with discussions in \sec{discussion}. \section{Preliminaries}\label{sec:prelim} \subsection{Notations} Throughout this paper, the space we consider is either $\mathbb{R}^d$ or a $d$-dimensional smooth compact Riemannian manifold denoted $M$. Bold lower-case letters $\mathbf{x}$, $\mathbf{y}$,\ldots, are used to denote vectors. If there is no ambiguity, we use normal lower-case letters, $x$, $y$,\ldots, to denote these vectors for simplicity. Depending on the context, $\d x$ may refer to either line differential or volume differential. We use $A_{jj'}$ to denote the element of the matrix $A$ at of row $j$ and column $j'$. Conversely, given all matrix elements $A_{jj'}$, we use the notation $(A_{jj'})$ to denote the matrix. Unless otherwise specified, $\| \cdot\|$ is used to denote the $\ell^2$ norm of vectors, spectral norm of matrices, and $L^2$ norm of functions. Similarly, $\| \cdot\|_1$ is used to denote the $\ell^1$ norm of vectors and $L^1$ norm of functions. For a function $f$, $\nabla f$ and $\nabla^2 f$ denote the gradient vector and Hessian matrix, respectively. $\Delta:=\sum_{j=1}^{d}\frac{\partial^2}{\partial x_{j}^2}$ is the Laplacian operator. $C^{\infty}(\mathbb{R}^d)$ is the set of all functions $f\colon\mathbb{R}^{d}\to\mathbb{R}$ that are continuous and differentiable up to any order. Notations about upper and lower bounds, $O(\cdot)$, $o(\cdot)$, $\Omega(\cdot)$, and $\Theta(\cdot)$, follow common definitions. We also write $f\ll g$ if $f=o(g)$, and $f\sim g$ if $f=\Theta(g)$. The $\tilde{O}$ notation omits poly-logarithmic terms, namely, $\tilde{O}(f):=O(f\mathrm{poly}(\log f))$ (in this paper, $\log$ denotes the logarithm with base $2$ and $\ln$ denotes the natural logarithm with base $e$). We write $f=O(g^{\infty})$ if \begin{align}\label{eqn:infty-power} \forall N>0,\quad f/g^{N} \to 0~(g\to 0). \end{align} Throughout the paper, we write $f\approx g$ if \begin{align}\label{eqn:approx-definition} f(x)=g+o(g) \end{align} when $g\neq 0$. When $g\to 0$, $f\to 0$ means $f=o(1)$ or, to stress the dependence, $f=o_g(1)$. For quantum mechanics, we use the \emph{Dirac notation} throughout the paper. Quantum states are vectors from a Hilbert space with unit norm. Let $\ket{\phi}$ denote a state vector, and $\bra{\phi}=(\ket{\phi})^{\dagger}$ denote the dual vector that equals to its conjugate transpose. The inner product of two states can be written as $\ip{\psi}{\phi}$. In the coordinate representation, for each $x\in\mathbb{R}^{d}$ we have the wave function $\phi(x):=\ip{x}{\phi}$, where $\ket{x}$ denotes the state localized at $x$. More basics on quantum mechanics and quantum computing can be found in standard textbooks, for instance~\cite{NC10}. \subsection{Classical preparations}\label{sec:classicalpre} Classical algorithms only have access to local information about the objective function at different sites, which is formalized as follows: \begin{definition}[Algorithms based on local queries]\label{defn:localquery} Denote a sequence of points and corresponding queries with size $T$ by $\{x_i, q(x_i)\}_{i=1}^T$, where each $q(x_i)$ can include the function value and arbitrary order derivatives (if exist). Algorithms based on local queries are those which determine the $j$th point $x_j$ by $\{x_i, q(x_i)\}_{i=1}^{j-1}$. \end{definition} As an example, the classical algorithm SGD can be mathematically described by \begin{definition}[Discrete model of SGD] Given a function $f(x)$, starting from an initial point $x_0$, SGD updates the iterates according to \begin{align} x_{k+1} = x_k - s\nabla f(x_k) - s\xi_{k}, \end{align} where $s$ is the \emph{learning rate} and $\xi_k$ is the noise term at the $k$th step. \end{definition} The local information in SGD is gradients. Since $s$ is small, define time $t_k = ks$, the points $\{x_k\}$ can be approximated by points on a smooth curve $\{X(t_k)\}$. The curve, which can be regarded as the continuous-time limit of discrete SGD, is determined by a learning-rate-dependent stochastic differential equation (lr-dependent SDE): \begin{definition}[SDE approximation of SGD] \begin{align} \d x = -\nabla f(x)\d t + \sqrt{s}\d W, \label{eq:lrsde} \end{align} where $W$ is a standard Brownian motion. \end{definition} The solution of \eq{lrsde}, $X(t)$, is a stochastic process whose probability density $\rho_{\rm SGD}(t,\cdot)$ evolves according to the Fokker–Planck–Smoluchowski equation \begin{align} \frac{\partial \rho_{\rm SGD}}{ \partial t} = \nabla\cdot (\rho_{\rm SGD} \nabla f) + \frac{s}{2} \Delta \rho_{\rm SGD}. \label{eq:FPS} \end{align} The validity of this SDE approximation has been discussed and verified in previous literature \cite{KY03,CS18,SSJ20,LMA21}. The results used in the present paper about SGD are based on analyses on \eq{lrsde}. For SGD, we consider an objective function $f$ in $\mathbb{R}^d$ and assume the following: \begin{assumption}[Confining condition~\cite{markowich1999trend,pavliotis2014stochastic}]\label{assum:confining} The objective function $f \in C^{\infty}(\mathbb{R}^d)$ should satisfy $\lim_{\|x\|\to +\infty } f(x) = +\infty$, and $\forall s>0,\exp(-2f/s)$ is integrable and \begin{align} \int_{\mathbb{R}^d} e^{-\frac{2f(x)}{s}} \d x < \infty. \end{align} \end{assumption} \begin{assumption}[Villani condition~\cite{villani2009hypocoercivity}]\label{assum:Villani} The following equation holds for all $s>0$: \begin{align} \|\nabla f(x) \|^2 - s\Delta f(x) \to \infty \quad(\|x\|\to \infty). \end{align} \end{assumption} \begin{assumption}[Morse function]\label{assum:morse} For any critical point $x$ of $f$ (i.e., $\nabla f(x) = 0$), the Hessian matrix $\nabla^2 f(x)$ is nondegenerate (i.e., all the eigenvalues of the Hessian are nonzero). \end{assumption} Under \assum{confining}, Eq.~\eq{FPS} admits a unique invariant Gibbs distribution \begin{align} \mu_{\rm SGD}(x) := \frac{e^{-\frac{2f(x)}{s}}}{\int_{\mathbb{R}^d}e^{-\frac{2f(x)}{s}} \d x}. \label{eq:Gibbs} \end{align} \begin{definition} A measurable function $g$ belongs to $L^2(\mu_{\rm SGD}^{-1})$, if \begin{align} \| g\|_{\mu_{\rm SGD}^{-1}} := \Big(\int_{\mathbb{R}^d} g^2 \mu_{\rm SGD}^{-1}(x) \d x\Big)^{\frac{1}{2}} < +\infty. \end{align} \end{definition} Under such measure, we have: \begin{lemma}[Lemma 2.2 and 5.2 of \cite{SSJ20}] Under \assum{confining}, if the initial distribution $\rho_{\rm SGD}(0,\cdot)\in L^2(\mu_{\rm SGD}^{-1})$, the lr-dependent SDE \eq{lrsde} admits a weak solution whose probability density \begin{align} \rho_{\rm SGD}(t,\cdot) \in C^1([0,+\infty), L^2(\mu_{\rm SGD}^{-1})), \end{align} is the unique solution to \eq{FPS} and $\rho_{\rm SGD}(t,\cdot) \to \mu_{\rm SGD}~(t \to \infty)$. \end{lemma} If we set $\psi_{\rm SGD}(t,\cdot) := \rho_{\rm SGD}(t,\cdot)/\sqrt{\mu_{\rm SGD}} \in L^2(\mathbb{R}^d)$, Eq.~\eq{FPS} is equivalent to \begin{align} s\frac{\partial \psi_{\rm SGD}}{ \partial t} = -\frac{\Delta_{f}^s}{2}\psi_{\rm SGD}, \label{eq:pseudo} \end{align} where $\Delta_{f}^s$ is called the \emph{Witten-Laplacian}, more specifically, \begin{align} \Delta_{f}^s:= -s^2 \Delta + \| \nabla f\|^2 - s\Delta f. \label{eq:defnWL} \end{align} Let $\delta_{s,1}$ be the smallest non-zero eigenvalue of $\Delta_{f}^s$, the following convergence guarantee for SGD holds: \begin{proposition}[Part of Theorem 2.8 in \cite{Mic19} and Lemma 5.5 in \cite{SSJ20}]\label{prop:mic19} Under \assum{confining}, \ref{assum:Villani}, and \ref{assum:morse}, for sufficiently small $s$, \begin{align} \|\rho_{\rm SGD} (t,\cdot) - \mu_{\rm SGD}\|_{\mu_{\rm SGD}^{-1}} \leq e^{-\frac{\delta_{s,1}}{2s}t}\|\rho_{\rm SGD} (0,\cdot) - \mu_{\rm SGD}\|_{\mu_{\rm SGD}^{-1}}, \end{align} where the smallest positive eigenvalues of the Witten-Laplacian $\Delta_{f}^s$ associated with $f$ satisfies \begin{align} \delta_{s,1} = s(\gamma_1 + o(s))e^{-\frac{2H_f}{s}}. \end{align} Here, $H_f$ and $\gamma_1$ are constants depending only on the function $f$. \end{proposition} \begin{corollary}\label{cor:sgdmixing} Assume the assumptions of \prop{mic19} are satisfied, for sufficiently small $s$ and any $\epsilon>0$, if \begin{align} t> \frac{2s}{\delta_{s,1}}\ln \frac{\|\rho_{\rm SGD} (0,\cdot) - \mu_{\rm SGD}\|_{\mu_{\rm SGD}^{-1}}}{\epsilon}, \end{align} then \begin{align} \|\rho_{\rm SGD} (t,\cdot) - \mu_{\rm SGD}\|_{\mu_{\rm SGD}^{-1}}<\epsilon. \end{align} \end{corollary} That is, the convergence time of SGD is loosely $O(s/\delta_{s,1})$ whose magnitude is largely related to $H_f$. The constant $H_f$ is called the \emph{Morse saddle barrier}, characterizing the largest height of barriers. Rigorous results about eigenvalues of the Witten-Laplacian are reviewed in \append{A.1}. \subsection{Quantum preparations}\label{sec:quantumpre} Our quantum algorithmic method essentially relies on the simulation of the Schr\"odinger equation: \begin{align} i \frac{\partial}{\partial t} \ket{\Phi(t)}= H \ket{\Phi(t)}, \label{eq:true} \end{align} where $i$ is the imaginary unit, $\hbar$ is the Planck constant, and $H$ is the Hamiltonian. Physically, we simulate a particle moving under a potential function $f$. Then, in the coordinate representation, the Schr\"odinger equation is specified: \begin{align} i \hbar\frac{\partial}{\partial t} \Phi = \Big(-\frac{\hbar^2}{2m}\Delta + f(x)\Big) \Phi. \end{align} Throughout this paper, we set $h = \hbar/\sqrt{2m}$. The spectrum of the Hamiltonian will highly depends on the variable $h$. More interestingly, by comparing \eq{pseudo} and \eq{true}, $h$ plays a similar role to that of the learning rate $s$. Thus we refer to $h$ as the quantum learning rate. In the reality, $\hbar$ is a fixed constant. However, since we are simulating quantum evolution by quantum computers, proper rescaling the simulation can equivalently be seen as varying the value of $\hbar$. The value of $\hbar$ affects the evolution time needed. However, rescaling $\hbar$ has no impact on quantum query complexity. Therefore, in this paper, $\hbar$ is an unimportant constant, i.e., $\hbar = 1$. As is introduced, we consider quantum tunneling from the bottom of a well to that of another well, in other words, tunneling between \emph{local ground states}. A local ground state of a well is the local eigenstate of the well with minimum eigenvalue. Technically, several kinds of local eigenstates are defined (see \blue{Definition A.7}, \blue{Definition A.8}, and \blue{Definition A.10} in \append{A.2.3}). Despite of the number of definitions, different kinds of local eigenstates are close to each other and share the same intuition: eigenstates of the Hamiltonian restricted in regions only contain one well. For convenience, if no otherwise specified, local eigenstates stand for orthonormalized eigenstates defined by \blue{Definition A.10}. Actually, there are also tunneling effects between local excited states. However, excited states are difficult to approximate accurately for general landscapes. Besides, due to interference, tunneling effects between different local excited states may cancel each other out. We restrict our attention to tunneling between local ground states in order to obtain explicit results along with a clear physical picture. Two local ground states can interact strongly with each other only if the difference between their energies is small relative to $h$ \cite{Hel88} (see also discussions after \blue{Proposition A.4} in \append{A.2.3}). In other words, this requires the function values between two local minima to be close and there is little resonance between the first (local) excited state in one well and the (local) ground state of the other \cite{Ras12,SCC91}. Therefore, our algorithms based on tunneling between local ground states are essentially restricted on landscapes where local minima are approximately global minima. Note that we can always find small enough $h$ to make two local ground states nonresonant, if the corresponding local minima are not exactly equal. As a result, to avoid more complicated restrictions on $h$, without loss of generality we assume that local minima are global minima, and they all have function value 0. More precisely: \begin{assumption}\label{assum:quantum1} The smooth objective function $f\colon \mathbb{R}^{d}\to \mathbb{R}$ satisfies \begin{equation} 0 = \min f < \lim_{\|x\|\to \infty} f. \end{equation} In addition, $f$ has finite number of local minima, and they can be decomposed as follows: \begin{equation} f^{-1}(0) = U_1 \cup U_2 \ldots \cup U_N, \end{equation} \begin{equation} U_j = \{x_j\}~\mathrm{is~a~point,}~\nabla f(x_j) = 0,~\mathrm{and}~\nabla^2 f(x_j) > 0~\mathrm{for}~j=1,\ldots,N. \end{equation} Each $U_j$ is called a \emph{well}. \end{assumption} This assumption will not affect the explicit forms of convergence time or present less physical insights. To further characterize the distance on such landscapes, an important geometric tool we use is the Agmon distance. \begin{definition}[Agmon distance]\label{defn:Agmon-distance} Under \assum{quantum1}, the Agmon distance $d(x,y)$ is defined as \begin{align} d(x,y) := \inf_{\gamma}\int_{\gamma} \sqrt{f(x)}\d x, \end{align} where $\gamma$ denotes pairwise $C^1$ paths connecting $x$ and $y$. For a set $U$, $d(x,U) = d(U,x) := \inf_{y\in U}d(x,y)$. And for two sets $U_1$ and $U_2$, $d(U_1,U_2) = \inf_{x\in U_1,y\in U_2}d(x,y)$. \end{definition} The minimal Agmon distance between wells are defined as \begin{align} S_0 :=\min_{j\neq k}d(U_j,U_k). \end{align} We only consider resonant wells by assuming the following for simplicity: \begin{assumption}[Informal]\label{assum:quantum2} The difference between any two local ground states are of the order $O(h^{\infty})$. In addition, for any well $U_j$, there exists another well $U_k$ ($k\neq j$) such that $d(U_j,U_k) = S_0$. \end{assumption} At last, to obtain explicit results we demand that \begin{assumption}\label{assum:quantum3} There are a finite number of paths of the Agmon length $S_0$ connecting $U_j$ and $U_k$ if $d(U_j,U_k) = S_0$. \end{assumption} \assum{quantum1}, \ref{assum:quantum2}, and \ref{assum:quantum3} are simplified from \blue{Assumptions A.5}, \blue{A.6}, and \blue{A.7} in \append{A.2} which presents details needed for analyzing quantum tunneling. Under \assum{quantum1}, \ref{assum:quantum2}, and \ref{assum:quantum3}, we state the main results of \append{A.2} as follows. For sufficiently small $h$, the orthonormalized local ground states $\ket{e_j},~j=1,\ldots,N$ almost localize near the wells $U_j,~j=1,\ldots,N$, respectively. The space $\mathcal{F}$ spanned by $\{\ket{e_j}:j=1,\ldots,N\}$ is exactly a low-energy invariant subspace of the Hamiltonian $H$. In other words, in the low-energy space $\mathcal{F}$, the particle walks between wells by quantum tunneling. The Hamiltonian restricted in $\mathcal{F}$, i.e., $H_{|\mathcal{F}}$, determines the strength of the quantum tunneling effect and is called the \emph{interaction matrix}. To explore $H_{|\mathcal{F}}$, we use the WKB method to estimate local ground states (\append{A.2.1}). Any local ground state function decays exponentially with respect to the Agmon distance to its corresponding well (\append{A.2.2}). Consequently, the tunneling effects would decay exponentially with respect to $S_0$. Having captured theses properties, \append{A.2.3} can give explicit estimations about $H_{|\mathcal{F}}$, namely, \blue{Proposition A.5}, \blue{Proposition A.6} (with $N^+ = N$), and \blue{Theorem A.1}. Finally, we restate a more formal version of our \ref{prb:main}: \begin{named}{Main Problem (restated)}\label{prb:mainformal} Given an objective function $f$ that satisfies \assum{quantum1}, \ref{assum:quantum2}, and \ref{assum:quantum3}, starting from one local minimum, find all local minima or find a certain target minimum. \end{named} We make the following remarks for clarification: \begin{remark} Assumptions in \sec{classicalpre} and \sec{quantumpre} are not contradictory. When considering SGD, we naturally add to the \ref{prb:mainformal} that the assumptions in \sec{classicalpre} should also be satisfied. \end{remark} \begin{remark} The assumption of starting from one local minimum enables quantum algorithms to prepare a local ground state, or more generally, a state largely in the aforementioned subspace $\mathcal{F}$. \end{remark} \begin{remark} Because finding a precise global minimum is impractical in general, it suffices to find points sufficiently close to the minima of interest. Later in \sec{standard}, we use two different measures of accuracy: 1. the function value difference; and 2. the distance to one of the minimum. \end{remark} \section{Quantum Tunneling Walks}\label{sec:quantumal} In this section, we present full details of the quantum tunneling walk (QTW). We begin with a one-dimensional example in \sec{onedimexp}, and then in \sec{QTW} we formally define QTW. In \sec{Qmixing} and \sec{Qhitting}, we study the mixing and hitting time of QTW, respectively. As an example, we give full details of applying QTW to tensor decomposition in \sec{tendecom}. \subsection{A one-dimensional example}\label{sec:onedimexp} We start the introduction of the quantum algorithm QTW with a one-dimensional example which quantifies the intuitions provided in \sec{intro}. \sec{onedimexp} also serves as a map connecting each step of the analysis to the needed mathematical tools. Later sections can be seen as generalizing results here for high dimensional and multi-well cases. General descriptions and results begin at \sec{QTW}. Consider the potential $f(x)$ in \fig{1-dimexample}, which has two global minima, $x_{\pm} = \pm a$. For simplicity, we take \begin{equation} f(x) = \frac{1}{2} \omega^2 (x + a)^2,\quad x \leq -\epsilon, \label{eq:eq01} \end{equation} where $\epsilon$ is a small number. In this way, the potential satisfies that $\min f =0$, and $f(x)$ is quadratic near minima. Besides, the symmetry of wells demands $f(x) = f(-x)$. The $f(x)$ for $x\in [-\epsilon,\epsilon]$ can always be made to be smooth. Near the two minima, $\pm a$, whose local harmonic frequencies are $\omega$, we can solve the Schr\"odinger equation locally and get two local ground states, $\Phi_{\pm}(x)$. For instance, around $-a$, if we set $y = x+ a$, the local ground state is determined by \begin{equation} H \Phi_{-}(x) = \varepsilon_0 \Phi_{-}(x),\quad H = - h^2 \frac{\d^2}{\d^2 y} + \frac{1}{2} \omega^2 y^2, \end{equation} where $\varepsilon_0 = h\omega/\sqrt{2}$. Physically, the demand of localization is equivalent to $\varepsilon_0 \ll f(0)$, indicating that the particle nearly cannot pass through the energy barrier. Concrete mathematical definitions and discussions on local ground states can be found in \append{A.2.2}. From a high-level perspective, the main idea of the present paper is to unite the interaction or tunneling between local ground states to realize algorithmic speedups. As we want to investigate the evolution of states, we need to determine the relationship between local ground states and the eigenstates of the Hamiltonian $H$. We set the eigenstates of $H$ as $\ket{n},~n=0,1,\ldots$ with energies $E_0 \leq E_1 \leq \cdots$, respectively (i.e., $\ket{0}$ is the global ground state, $\ket{1}$ is the first excited state, etc.). The overlap of states $\Phi_{\pm}$ is small (namely, $\ip{\Phi_{+}}{\Phi_{-}}\approx 0$), as they are local and separated by a high barrier. Denote the subspace spanned by $\Phi_{-}$ and $\Phi_{+}$ as $\mathcal{E}$, and that spanned by $\ket{0}$ and $\ket{1}$ as $\mathcal{F}$. Both $\mathcal{E}$ and $\mathcal{F}$ are 2-dimensional and contains states with low energies. It is intuitive that $\mathcal{E} \approx \mathcal{F}$ (which is guaranteed by \blue{Proposition A.3}). For the one-dimensional case, we just take $\mathcal{E} = \mathcal{F}$ for simplicity. In this way, we can represent $\ket{0}$ and $\ket{1}$ by $\ket{\Phi_{\pm}}$ in the following general way: \begin{align} \ket{0} &= \cos\theta \ket{\Phi_{-}} + \sin\theta \ket{\Phi_{+}},\\ \ket{1} &= \sin\theta \ket{\Phi_{-}} - \cos\theta \ket{\Phi_{+}}. \end{align} Restricted in the subspace $\mathcal{F}$, the two-level system Hamiltonian can be written as \begin{align} H|_{\mathcal{F}} &= \left( \begin{array}{cc} \varepsilon_{-} & -\nu \\ -\nu & \varepsilon_{+} \end{array} \right),\quad \mathrm{under~basis}~\{\ket{\Phi_{-}},\ket{\Phi_{+}}\},\\ H|_{\mathcal{F}} &= \left( \begin{array}{cc} E_0 & 0 \\ 0 & E_1 \end{array} \right),\quad\ \mathrm{under~basis}~\{\ket{0},\ket{1}\}, \end{align} where $\nu$ is called the tunneling amplitude, measuring the interaction between wells. Because the $f(x)$ we choose is symmetric, $\varepsilon_{-} = \varepsilon_{+} = \varepsilon_{0} = h\omega/\sqrt{2}$. Therefore, we have $\theta = \pi/4$ and the energy gap $\Delta E := E_1 - E_0 = 2\nu$. We will refer to this Hamiltonian restricted in subspace $\mathcal{F}$ as the \emph{interaction matrix}, indicating that $H_{|\mathcal{F}}$ characterizes the interaction between wells. In our setting, we can begin at a local minimum, where the local ground state is easy to prepare (see justifications in \append{B.2}). Without loss of generality, let us begin the quantum simulation at the state $\Phi_{-}$, namely, setting $\ket{\Phi(0)} = \ket{\Phi_{-}}$. After evolution of time $t$, the state becomes \begin{equation} \ket{\Phi(t)} = e^{-iHt}\ket{\Phi(0)} = \frac{1}{\sqrt{2}}(e^{-iE_0t}\ket{0} + e^{-iE_1t}\ket{1}) \propto \cos(\Delta E t/2)\ket{\Phi_{-}} + i\sin(\Delta E t/2)\ket{\Phi_{+}}. \end{equation} And the probabilities of finding the particle in the right and left wells are give by \begin{equation} P_{\pm}(t) = |\ip{\Phi_{\pm}}{\Phi(t)}|^2 = \frac{1\mp \cos(\Delta E t)}{2}. \end{equation} The energy gap $\Delta E$ is also called the Rabi oscillation frequency, suggesting that the particle oscillates between the two wells periodically. Our aim is to pass through the barrier and find other local minima (for the case here, is to find the other local minimum). Since for small $h$, the local state $\ket{\Phi_{+}}$ distributes in a very convex region near the right local minimum, it suffice to solve our problem by measuring the position of state $\ket{\Phi(t)}$ when $P_{+}(t)$ is large. However, we may not be able to know $\Delta E$ precisely in advance. So, we will apply the method of quantum walks: evolving system for time $t$ which is chosen randomly from $[0,\tau]$, and then measuring the position \cite{CCD+03}. The resulted distribution is \begin{align} \rho_{\rm QTW}(\tau,x) \approx |\ip{x}{\Phi_{-}}|^2\int_{0}^{\tau}P_{-}(t)\frac{\d t}{\tau} + |\ip{x}{\Phi_{+}}|^2\int_{0}^{\tau}P_{+}(t)\frac{\d t}{\tau}, \end{align} where QTW denotes quantum tunneling walk. Define $p_{-\to \pm}(\tau) = \int_{0}^{\tau}P_{\pm}(t)\frac{\d t}{\tau}$, which is the probabilities of finding the right and left local ground states, respectively. Since $|\ip{x}{\Phi_{-}}|^2$ is small for $x$ near $+a$, the probability of finding the particle near $+a$ is determined by \begin{align} \int_{\rm right~well} \rho_{\rm QTW}(\tau,x) \d x \approx \int_{\rm right~ well}\d x |\ip{x}{\Phi_{+}}|^2p_{-\to +} \approx p_{-\to +}. \end{align} Therefore, it suffice to study properties of $p_{-\to \pm}$. In the present case, \begin{align} p_{-\to \pm} = \frac{1}{2}\left(1 \mp \frac{\sin(\Delta E \tau)}{\Delta E \tau} \right), \end{align} which will converge when $\tau \to \infty$. This fact ensures that we can find the right local minimum $+a$ with a probability larger than some constant after evolving the system for enough long time. Starting from $\ket{\Phi_{-}}$, the hitting time for the right well is \begin{align} T_{\rm hit}(\Phi_{+}|\Phi_{-}) = \inf_{\tau > 0} \frac{\tau}{p_{-\to +}(\tau)}. \end{align} Since the probability for successful tunneling in one trial is $p_{-\to +}(\tau)$, we can repeat trials for $1/p_{-\to +}(\tau)$ times to secure one success and the total evolution time is $\tau/p_{-\to +}(\tau)$. For sufficiently small $\epsilon \ll 1$, if $\frac{1}{\Delta E \tau} \leq \epsilon$, we can get $p_{-\to +}(\tau) \geq \frac{1}{2}(1-\epsilon)$. Therefore, the hitting time can be bounded by $O(\frac{1}{\Delta E\epsilon})$. As is going to be shown later, if we want to find \emph{all} local minima, the mixing time would be a better quantifier. The limiting distribution is \begin{align} \mu_{\rm QTW}(x) := \lim_{\tau \to \infty}\rho_{\rm QTW}(\tau,x) \approx |\ip{x}{\Phi_{-}}|^2 p_{-\to -} + |\ip{x}{\Phi_{+}}|^2p_{-\to +}. \end{align} The mixing time measures how fast the distribution $\rho_{\rm QTW}(\tau,x)$ converges to $\mu_{\rm QTW}(x)$. We define $T_{\rm mix}$ as the $\epsilon$-close mixing time which satisfies \begin{align} T_{\rm mix} =\inf_{\|\rho_{\rm QTW}(\tau,\cdot) - \mu_{\rm QTW}(\cdot)\|_1 \leq \epsilon} \tau. \end{align} Because $\|\rho_{\rm QTW}(\tau,\cdot) - \mu_{\rm QTW}(\cdot)\|_1 \leq O(\frac{1}{\Delta E \tau})$, we have $\|\rho_{\rm QTW}(\tau,\cdot) - \mu_{\rm QTW}(\cdot)\|_1 \leq \epsilon$ if $\tau = \Omega(\frac{1}{\Delta E \epsilon})$. Therefore, $T_{\rm mix}$ could be bounded: $T_{\rm mix} = O(\frac{1}{\Delta E \epsilon})$. For simulating a time-independent Hamiltonian, the number of queries needed are roughly proportional to the total evolution time (as demonstrated in \sec{quantumpre} or see details in \append{B.1}). The major task left is to calculate the energy gap $\Delta E$, get different evolution times and compare them with classical results. As is specified in \append{A.2.3}, $0$ is the boundary of the two wells and the tunneling amplitude $\nu$ can be given by \begin{align} \nu = h^2 (\Phi_{-}(0)\Phi'_{+}(0) - \Phi'_{-}(0)\Phi_{+}(0)). \label{eq:eq15} \end{align} To obtain an explicit result, we need to use the WKB approximation of the local ground states (\append{A.2.1}): \begin{align} \Phi_{-}(x) \approx \frac{1}{ h^{1/4}} a_0(x)e^{-\frac{1}{h}\int_{-a}^{x}\sqrt{f(\xi)}\d \xi},\quad \Phi_{+}(x) = \Phi_{-}(-x), \label{eq:eq16} \end{align} where $a_0(x)$ is given by \begin{align} a_0(x) = \left( \frac{\omega}{\sqrt{2} \pi}\right)^{1/4}e^{-\frac{1}{2}\int_{-a}^{x}(\frac{f'(\xi)}{2f(\xi)} - \frac{\omega}{\sqrt{2f}})\d \xi }, \label{eq:eq17} \end{align} which is determined by the transport equation (Eq.~(\blue{S24}) in \append{A.2.1}) and the normalization condition. Substituting \eq{eq16} and \eq{eq17} to \eq{eq15}, we get \begin{align} \nu = 2\sqrt{\frac{h\omega f(0)}{\sqrt{2}\pi}}e^C[1+O(h)]e^{-\frac{S_0}{h}}, \end{align} where the constants $C$ and $S_0$ are given by \begin{align} C = \int_0^a \Big(\frac{\omega}{\sqrt{2f(\xi)}} - \frac{1}{a-\xi}\Big)\d \xi , \quad S_0 = \int_{-a}^a \sqrt{f(\xi)} \d\xi. \end{align} Note that if $f(x)$ is quadratic, the factor $C$ will be zero, indicating that $C$ measures the deviation of the landscape from being quadratic. The quantity $S_0$ is called the Agmon distance between two the local minima (see \append{A.2.2}). Since we assumed by \eq{eq01} that $f(x)$ is almost quadratic, we have $f(0)\approx \frac{1}{2}\omega^2a^2$ and \begin{align} \nu \approx h\omega\sqrt{\frac{\sqrt{2}\omega a^2}{h\pi}} e^{-\frac{S_0}{h}}. \end{align} As discussed above, the mixing time and hitting time could be bounded by the following characteristic time \begin{align} T_c = \frac{1}{\Delta E} = \frac{1}{2\nu } \approx \frac{\sqrt{\pi}}{2a\omega \sqrt{\sqrt{2}\omega h}}e^{\frac{1}{h}\int_{-a}^a \sqrt{2f(\xi)}\d \xi}. \end{align} Next, we need to find how long it takes for SGD to escape from the left local minimum. Discrete-time SGD with a small learning rate $s$ can be approximated by a learning-rate dependent stochastic differential equation (lr-dependent SDE) \cite{SSJ20}: \begin{align} \d X = - \nabla f(X) \d t + \sqrt{s}\d W, \end{align} where $W$ is a standard Brownian motion. Before hitting $0$, the SDE is almost an Ornstein–Uhlenbeck process: \begin{align} \d X = -\omega^2 (X+a) \d t + \sqrt{s}\d W. \end{align} The expected time for the Ornstein–Uhlenbeck process to first hit $0$ is \begin{align} \mathbb{E}T_0 \approx \frac{\sqrt{\pi s}}{a \omega^3}e^{\frac{2H_f}{s}}, \end{align} where $H_f$ is called the Morse saddle barrier and equals to $f(0) - f(-a) \approx \frac{1}{2}\omega^2 a^2$ in our case. Although $\mathbb{E}T_0$ is not a precise classical counterpart of either quantum mixing or hitting time, it is heuristic to compare $\mathbb{E}T_0$ with $T_c$ which can both reflect the time to escape from the left well. The forms of $\mathbb{E}T_0$ and $T_c$ are very similar. Two major differences can be observed: 1. the exponential term in $\mathbb{E}T_0$ is determined by the height of the barrier, while that in $T_c$ is related to an integral of $\sqrt{f}$; 2. $T_c \propto 1/\omega^{3/2}$ but $\mathbb{E}T_0 \propto 1/\omega^{3}$, indicating that the flatness of the wells affects differently on quantum and classical methods. We will show that for landscapes with multiple wells, the distribution of wells is also an important factor. In general, QTW could be faster than SGD if the barriers between local minima are high but thin, each well is close to many other wells, and wells are flat. The above comparison is intuitive but not rigorous. Two important technical details for comparison are needed for quantitative discussions. It is shown in \sec{quantumpre} that a super-polynomial separation between evolution time of QTW and SGD gives rise to a super-polynomial separation between quantum and classical queries for QTW and SGD, respectively. Therefore, it suffice to compare the evolution time, especially the exponential term $e^{S_0/h}$ and $e^{2H_f/s}$. The second problem is that $h$ and $s$ are not two constants but variables. The evolution times cannot be quantitatively compared if $h$ and $s$ are independent. In \sec{standard}, we develop two natural standards to make fair comparisons between QTW and SGD, which specifies $h$ and $s$. \subsection{Definition of quantum tunneling walks}\label{sec:QTW} We now formally describe the model of a quantum tunneling walk (QTW) on a general objective function $f(x)$ satisfying assumptions in \sec{quantumpre}. The wells are denoted by $U_j = \{x_j\}~(j=1,\ldots,N)$. Let $\ket{j}$ be the corresponding orthonormalized local ground state of $U_j$. It is ensured that $\{\ket{j}:j=1,\ldots,N\}$ spans a low energy subspace, $\mathcal{F}$, of the Hamiltonian $H = -h^2\Delta + f(x)$. If one has information about one well $U_j$ and its neighborhood, the construction of the local ground state should be easy which can be close to $\ket{j}$ or at least be almost in the subspace $\mathcal{F}$. We assume the initial state $\Phi(0)$ to be in $\mathcal{F}$. The evolution is determined by the Schr\"odinger equation, \begin{align} i \frac{\d}{\d t} \ip{x}{\Phi(t)} = \bra{x} H \ket{\Phi(t)}, \end{align} where $|\ip{x}{\Phi(t)}|^2$ is the probability distribution of finding the walker at $x$. The Schr\"odinger equation indicates the phenomenon of quantum tunneling since it can be rewritten as \begin{align} i \frac{\d}{\d t} \ip{j}{\Phi(t)} = \sum_{j'} \bra{j} H_{|\mathcal{F}} \ket{j'} \ip{j'}{\Phi(t)},~\mathrm{for~any}~j=1,\ldots,N, \label{eq:qtwsch} \end{align} given that $\ket{\Phi(0)}$ is in the subspace $\mathcal{F}$. Here, $(\langle j|H_{|\mathcal{F}}|j^{\prime} \rangle)$ is called as the \emph{interaction matrix} and is calculated by \blue{Proposition A.5} and \blue{A.6}. Once we get $\ip{j}{\Phi(t)}$ and $\ip{x}{j}$ for all $j$, we can obtain the probability distribution $|\ip{x}{\Phi(t)}|^2$. The overlap $\ip{x}{j}$ is invariant with respect to $t$. So, we may focus on \eq{qtwsch} to investigate the time evolution. As is shown by \eq{qtwsch}, restricted in the low energy subspace $\mathcal{F}$, the quantum evolution is similar to that of a quantum walk on a graph. The wells correspond to vertices of the graph, and the quantum tunneling effects between wells determine the graph connectivity. QTW walks among different wells by quantum tunneling, helping to find all other local minima. Finally, according to \blue{Lemma B.2} in \append{B.1}, the quantum query complexity of simulating the Schr\"odinger equation is directly linked to the evolution time $t$ and is bounded by \begin{align} O\left(\|f\|_{L^{\infty}(\Omega)} t \frac{\log (\|f\|_{L^{\infty}(\Omega)} t/\epsilon)}{\log \log (\|f\|_{L^{\infty}(\Omega)} t/\epsilon)}\right), \end{align} where $\Omega$ is a large region containing all minima of interest and $\epsilon$ is the precision quantified by the $L^2$ norm between the target and the obtained wave functions. Loosely speaking, we need $\tilde{O}(t)$ quantum queries if the evolution time is $t$. For SGD, the number of queries needed is at least $\Omega(t/s)$ for time $t$. Thus, as long as there is a super-polynomial separation between QTW and SGD evolution time, there is a super-polynomial separation between quantum queries and classical queries for QTW and SGD, respectively. Conclusions on speedups are essentially based on comparisons of query complexity. However, based on this relationship between evolution time and query complexity, we can focus on comparisons of time. To sum up, QTW is quantum simulation with the system Hamiltonian being $H = -h^2\Delta +f(x)$ and the initial state being in a low energy subspace of $H$, where $f(x)$ is the potential function of a type of benign landscapes (\assum{quantum1}, \ref{assum:quantum2}, and \ref{assum:quantum3}). QTW can be efficiently implemented on quantum computers. \subsection{Mixing time}\label{sec:Qmixing} For a given landscape $f(x)$, the complexity of Hamiltonian simulation mainly depends on the evolution time (see~\append{B.1}). In this section, we focus on the evolution time needed for fulfilling the tasks of finding all minima. Since quantum evolutions are unitary, different from SGD, QTW never converges. In this case, after running QTW for some time $t$, we measure the position of the walker. Similar to the quantum walks in \cite{CCD+03}, the evolution time $t$ can be chosen uniformly in $[0,\tau]$. Later, we will prove that under sufficiently large $\tau$, QTW can find other wells with probability larger than some constant. Note that there can be better strategies to determine the time for measurement $t$ than uniformly sampling from an interval $[0,\tau]$ \cite{AC21}. For simplicity, we only analyze the original strategy of \cite{CCD+03} in the present paper. As is mentioned earlier, we initialize at a state $\ket{\Phi(0)}\in \mathcal{F}$. In later sections, we may specify $\ket{\Phi(0)}$ to be one of the local ground states. Let the spectral decomposition of $H_{|\mathcal{F}}$ to be $H_{|\mathcal{F}} = \sum_{k=1}^N E_k \ket{E_k}\bra{E_k}$. Simulating the system for a time $t$ chosen uniformly in $[0,\tau]$, one can obtain the probability density of finding the walker at $x$\footnote{The distribution $\rho_{\rm QTW}(\tau,x)$ depends on the initial state $\ket{\Phi(0)}$. A more rigorous notation is $\rho_{\rm QTW}(\tau,x|\Phi(0))$. When there is no confusion, we omit the initial state for simplicity.} \begin{align} \rho_{\rm QTW}(\tau,x) :=& \frac{1}{\tau}\int_0^{\tau} \d t |\langle x|e^{-i H_{|\mathcal{F}} t}\ket{\Phi(0)}|^2 \nonumber \\ =& \sum_{E_k=E_{k^{\prime}}} \langle x|E_k\rangle \langle E_k |\Phi(0)\rangle \langle \Phi(0) |E_{k^{\prime}}\rangle \langle E_{k^{\prime}}| x\rangle \nonumber\\ &~+\sum_{E_k \neq E_{k^{\prime}}} \frac{1-e^{-i(E_k - E_{k^{\prime}})\tau}}{i(E_k - E_{k^{\prime}})\tau} \langle x|E_k\rangle \langle E_k |\Phi(0)\rangle \langle \Phi(0) |E_{k^{\prime}}\rangle \langle E_{k^{\prime}}| x\rangle. \end{align} The time-averaged probability density leads to a limiting distribution when $\tau \to \infty$: \begin{equation} \mu_{\rm QTW} := \sum_{E_k=E_{k^{\prime}}} \langle x|E_k\rangle \langle E_k |\Phi(0)\rangle \langle \Phi(0) |E_{k^{\prime}}\rangle \langle E_{k^{\prime}}| x\rangle. \end{equation} With the strategy of measuring at $t$ randomly chosen from $[0,\tau]$, QTW can output a distribution $\rho_{\rm QTW}(\tau,x)$ with limit. Such a process is regarded as mixing. Quantum mixing time evaluates how fast $\rho_{\rm QTW}(\tau,x)$ converges to $\mu_{\rm QTW}(x)$, and is rigorously defined as: \begin{definition}[Mixing time of QTW]\label{defn:mixingqtw} $T_{\rm mix}$ is called the $\epsilon$-close mixing time, iff for any $\tau\geq T_{\rm mix}$, we have \begin{align} \|\rho_{\rm QTW}(\tau,\cdot) - \mu_{\rm QTW}(\cdot)\|_1 \leq \epsilon. \label{eq:mixingqtw} \end{align} \end{definition} The following lemma provides a general bound for the QTW mixing time whose proof is postponed to \append{C.1.1}. \begin{lemma}[Upper bound for QTW mixing time]\label{lem:qtwmixingtime} The condition \eq{mixingqtw} can be satisfied if \begin{align} \tau \geq \frac{2}{\epsilon} \sum_{E_k\neq E_{k^{\prime}}}\frac{|\ip{E_k}{\Phi(0)} \ip{\Phi(0)}{E_{k'}}|}{|E_k - E_{k'}|}[1+(N-1)|O(h^{\infty})|], \end{align} and this implies \begin{align} \hspace{-2mm} T_{\rm mix} = O \bigg(\frac{1}{\epsilon} \sum_{E_k\neq E_{k^{\prime}}}\frac{|\ip{E_k}{\Phi(0)} \ip{\Phi(0)}{E_{k'}}|}{|E_k - E_{k'}|}[1+(N-1)|O(h^{\infty})|] \bigg) \leq O \bigg(\frac{N}{\epsilon \Delta E} [1+(N-1)|O(h^{\infty})|] \bigg), \label{eq:qtwmixing} \end{align} where $\Delta E := \min_{E_k\neq E_{k^{\prime}}}|E_k - E_{k'}|$ is referred to as the minimal gap of $H_{|\mathcal{F}}$. \end{lemma} The term $O(h^{\infty})$ in \eq{qtwmixing} originates from integrals $\int |\ip{x}{j} \ip{j'}{x}| \d x,~j\neq j'$. Intuitively, states $\ket{j}$ and $\ket{j'}$ localize in different wells, such that $\ip{x}{j} \ip{j'}{x}$ is exponentially small with respect to $h$ for any $x$. \lem{qtwmixingtime} highlights the dependence of the mixing time on the initial state $\ket{\Phi(0)}$ and the eigenvalue gaps of $H_{|\mathcal{F}}$. Concrete examples will be given in \sec{comparison}, where we further illustrate \eq{qtwmixing} and compare QTW with SGD. As is mentioned, \eq{qtwsch} indicates a quantum walk: a well $U_j$ can be seen as a vertex of a graph and $H_{|\mathcal{F}}$ implies graph connectivity (interaction between wells) similar to the graph Laplacian. The connection between QTW and quantum walks is helpful to simplify the physical picture of QTW. However, we also address the difference between quantum walks and QTW. For quantum walks, we only consider the probabilities of finding the walker at vertices, that is,\footnote{Similar to $\rho(\tau,x)$, the probability $p(\tau,j)$ depends on the initial state $\ket{\Phi(0)}$ and a more rigorous notation is $p(\tau,j|\Phi(0))$. When there is no confusion, we omit the initial state for simplicity.} \begin{align} p(\tau,j) &:= \frac{1}{\tau}\int_0^{\tau} \d t |\langle j|e^{-i H_{|\mathcal{F}} t}| \Phi(0) \rangle|^2 \nonumber \\ & = \sum_{E_k=E_{k^{\prime}}} \langle j|E_k\rangle \langle E_k |\Phi(0)\rangle \langle \Phi(0) |E_{k^{\prime}}\rangle \langle E_{k^{\prime}}| j\rangle \nonumber\\ &~~+\sum_{E_k \neq E_{k^{\prime}}} \frac{1-e^{-i(E_k - E_{k^{\prime}})\tau}}{i(E_k - E_{k^{\prime}})\tau} \langle j|E_k\rangle \langle E_k |\Phi(0)\rangle \langle \Phi(0) |E_{k^{\prime}}\rangle \langle E_{k^{\prime}}| j\rangle. \end{align} When $\tau \to \infty$, $p(\tau,j)$ also converges to a limit \begin{align} p(\infty,j) := \sum_{E_k=E_{k^{\prime}}} \langle j|E_k\rangle \langle E_k |\Phi(0)\rangle \langle \Phi(0) |E_{k^{\prime}}\rangle \langle E_{k^{\prime}}| j\rangle. \end{align} Following results, \lem{limitdis} and \lem{mixingqw}, show the connection and difference between QTW and quantum walks in a more quantitative way (detailed proofs can be found in \append{C.1.2} and \append{C.1.3}). \begin{lemma}[Limit distributions]\label{lem:limitdis} Limit distributions of the QTW and the quantum walk satisfy the following: \begin{align} \mu_{\rm QTW}(x) = \sum_{j} p(\infty, j) |\langle x|j\rangle|^2 + O(h^{\infty}). \end{align} \end{lemma} \begin{definition}[Mixing time of quantum walks \cite{CLR20}]\label{defn:mixingqw} $t_{\rm mix}$ is called the $\epsilon$-close mixing time of the quantum walk, iff for any $\tau \geq t_{\rm mix}$, \begin{align} \sum_{j=1}^N |p(\tau,j) - p(\infty, j)| \leq \epsilon. \label{eq:disofp} \end{align} \end{definition} \begin{lemma}[Upper bound for the mixing time of quantum walks]\label{lem:mixingqw} The condition \eq{disofp} is satisfied if \begin{align} \tau \geq \frac{2}{\epsilon} \sum_{E_k\neq E_{k^{\prime}}}\frac{|\ip{E_k}{\Phi(0)} \ip{\Phi(0)}{E_{k'}}|}{|E_k - E_{k'}|}, \end{align} and we have \begin{align} t_{\rm mix} = O \bigg(\frac{1}{\epsilon} \sum_{E_k\neq E_{k^{\prime}}}\frac{|\ip{E_k}{\Phi(0)} \ip{\Phi(0)}{E_{k'}}|}{|E_k - E_{k'}|} \bigg) \leq O \bigg(\frac{N}{\epsilon \Delta E} \bigg). \label{eq:qwmixing} \end{align} \end{lemma} By \lem{limitdis} and the comparison of \lem{qtwmixingtime} and \lem{mixingqw}, we know that for sufficiently small $h$, which indicates that local ground states localize sufficiently near their respective wells, QTW can be well characterized by a quantum walk. On a higher level of speaking, QTW generalizes quantum walks from walking on discrete graphs to propagating on continuous functions. And QTW may enable new phenomenons not shown in quantum walks when states $\ket{j}$ are poorly localized near $U_j$. On the other hand, QTW under proper conditions can be used to implement quantum walks. \subsection{Hitting time}\label{sec:Qhitting} If we aim at finding one particular well (the one with global minimum or the one with the best generalization properties), hitting time instead of mixing time should be of interest. Classically, the hitting time is the expected time required to find some target region or point. For quantum algorithms, we cannot output the position of the walker at all times and the system state would be destroyed by measuring its position. Thus, the definition of the hitting time for quantum algorithms is slightly different. We first see how previous literature defines the quantum walk hitting time: \begin{definition}[Hitting time for quantum walks \cite{AC21}] Consider a quantum walk governed by \eq{qtwsch}. Let the state $\ket{j}$ be the one of interest. Then, starting from the initial state $\ket{\Phi(0)}$, the hitting time of the quantum walk is defined as follows: \begin{align} t_{\rm hit}(j) := \inf_{\tau>0} \frac{\tau}{p(\tau,j)}. \end{align} \end{definition} To understand this definition, we first refer to the process, evolving the system for time $t$ uniformly chosen from $[0,\tau]$, as one trial. Using one trial, the probability of getting $\ket{j}$ is $p(\tau,j)$. So, repeating the trials for $1/p(\tau,j)$ times guarantees to hit $\ket{j}$ with high probability. In this case, the total evolution time needed is bounded by $\tau/p(\tau,j)$. In the same spirit, we can define and bound the QTW hitting time as follows: \begin{definition}[Hitting time of QTW] For QTW governed by \eq{qtwsch}, let the open and $C^2$-bounded region $\Omega$ be the region of interest. Then, starting from the initial state $\ket{\Phi(0)}$, the $\Omega$-hitting time of QTW is defined as follows: \begin{align} T_{\rm hit}(\Omega) := \inf_{\tau>0} \frac{\tau}{\int_{\Omega}\rho_{\rm QTW}(\tau,x)\d x}. \end{align} \end{definition} Basic results about the hitting time (\lem{qwhitting} and \lem{qtwhitting}) are present as follows. The proof of \lem{qtwhitting} is in \append{C.2.1} and that of \lem{qtwhitting} is in \append{C.2.2}. \begin{lemma}[Upper bound of the quantum walk hitting time]\label{lem:qwhitting} The probability of finding $\ket{j}$ can be bounded as follows \begin{align} p(\tau,j) \geq p(\infty,j) - \sum_{E_k \neq E_{k^{\prime}}} \frac{2}{|E_k - E_{k^{\prime}}|\tau} |\langle j|E_k\rangle \langle E_k |\Phi(0)\rangle \langle \Phi(0) |E_{k^{\prime}}\rangle \langle E_{k^{\prime}}| j\rangle| \geq p(\infty,j) - \frac{2}{\Delta E \tau}. \end{align} As a result, for any $\epsilon< p(\infty,j)$, setting $\tau_{\epsilon} = 2/\Delta E\epsilon$, we have \begin{align} t_{\rm hit}(j) \leq \frac{\tau_{\epsilon}}{p(\tau_{\epsilon},j)} \Rightarrow t_{\rm hit}(j) = O\bigg(\frac{1/\Delta E\epsilon}{p(\infty,j) - \epsilon}\bigg). \label{eq:qwhitting} \end{align} \end{lemma} If $\epsilon$ in \eq{qwhitting} is small enough, $\tau_{\epsilon} = 2/\Delta E\epsilon$ permits a good mixing and we may write \begin{align} t_{\rm hit}(j) = O\bigg(\frac{1}{p(\infty,j)\Delta E\epsilon} \bigg), \end{align} which suggests we are using the mixing time to bound the hitting time. \begin{lemma}[Upper bound of the QTW hitting time]\label{lem:qtwhitting} Consider an bounded open set $\Omega_{j}$ containing only one well $U_j$,\footnote{To be rigorous, satisfy (\blue{S85}) in \append{A.2.3}.} we have \begin{align} \hspace{-1mm} \int_{\Omega_j}\rho_{\rm QTW}(\tau,x)\d x \geq \int_{\Omega_j}\mu_{\rm QTW}(x)\d x - \frac{2}{\Delta E \tau}(1+|O(h^{\infty})|) = p(\infty, j) + O(h^{\infty}) - \frac{2}{\Delta E \tau}(1+|O(h^{\infty})|). \label{eq:qtwhitting} \end{align} For any $\epsilon< \int_{\Omega_j}\mu_{\rm QTW}(x)\d x$, let $\tau_{\epsilon} = 2 (1+|O(h^{\infty})|)/\Delta E\epsilon$, we have \begin{align} T_{\rm hit}(\Omega_j) \leq \frac{\tau_{\epsilon}}{\int_{\Omega_j}\rho_{\rm QTW}(\tau_{\epsilon},x)\d x} \Rightarrow T_{\rm hit}(\Omega_j) = O\bigg(\frac{1}{\Delta E\epsilon} \frac{1+|O(h^{\infty})|}{\int_{\Omega_j}\mu_{\rm QTW}(x)\d x - \epsilon}\bigg). \end{align} \end{lemma} The upper bounds we have obtained on mixing and hitting time are still not explicit, as $H_{|\mathcal{F}}$ is not given. Next, we establish relationships between an objective landscape, the corresponding interaction matrix $H_{|\mathcal{F}}$, and the time cost of the QTW algorithm. This is a main task of the present paper. In later sections, we figure out major geometric properties that affect $H_{|\mathcal{F}}$ on specific landscapes. \subsection{Application: Tensor decomposition}\label{sec:tendecom} After giving the definition of QTW and studying its mixing and hitting time, now we use QTW to solve a practical problem, orthogonal tensor decomposition, which is a central problem in learning many latent variable models \cite{AGH+14}. Specifically, we consider a fourth-order tensor $T\in \mathbb{R}^{d^4}$ that has orthogonal decomposition: \begin{equation} T = \sum_{i=1}^d a_j^{\otimes 4}, \label{eq:4-tensor} \end{equation} where the components $\{a_j \}$ form an orthonormal basis of a $d$-dimensional space ($a_j^{\top} a_j = \delta_{ij}$). The goal of orthogonal tensor decomposition is to find all components $\{a_j \}$. Following previous popular methods \cite{CLX+09,Hyv99}, we try to find all components by a single optimization problem. Concretely, we consider the following landscape \cite{FJK96}:\footnote{The original objective function used in previous papers including \cite{FJK96} is $T(u,u,u,u)$. The function in \eq{TenDecObj} is designed such that $\min f = 0$.} \begin{equation} f(u) = 1-T(u,u,u,u) = 1- \sum_{i=1}^d (u^{\top} a_j)^4,~\|u\|_2^2 = 1. \label{eq:TenDecObj} \end{equation} Without loss of generality, we work in the coordinate system specified by $\{a_j\}_{j=1}^{d}$. In particular, let $u = \sum_j^d x_j a_j$ and $x= (x_1,\ldots,x_d)$, we obtain $f(x) = 1-\sum_{i=1}^d x_i^4$. Later, we also use $a_j$ to denote the vector $(0,\ldots,1,\ldots,0)$ where the only nonzero coordinate with value 1 appears in the $j$th entry. $f(x)$ has $2d$ local minima $\pm a_1,\ldots\pm a_d$ uniformly distributed on the $d$-dimensional sphere $\mathbb{S}^{d-1}$.\footnote{Here, we need to consider quantum simulation on the manifold $\mathbb{S}^{d-1}$ which should cost the same quantum queries as quantum simulation on $\mathbb{R}^{d}$ under the same evolution time (see discussions in \append{B.1}).} Therefore, finding all the minima solves the orthogonal tensor decomposition problem. \begin{figure} \centerline{ \includegraphics[width=0.4\textwidth]{TenDeclabel.pdf}} \caption{The landscape given by \eq{TenDecObj} for dimension $d=3$: local minima $a_1$, $a_2$, and $-a_3$ are highlighted by red points ({\color{red}$\bullet$}) and corresponding labels $\alpha = (j(\alpha), \pi_{\alpha})$ are shown. } \label{fig:TenDeclabel} \end{figure} The problem of tensor decomposition has notable symmetries. As a result, the objective function $f$ \eq{TenDecObj} is nonconvex, and we can apply QTW to such a landscape. We can use the pair $\alpha = (j,\pi_{\alpha})$ to denote the local minima and corresponding wells, where $j = j(\alpha)$ (i.e. $j(\cdot)$ is a function) refers to $a_j$ and $\pi_{\alpha} \in \{\pm 1\}$ specifies whether it is $+ a_j$ or $- a_j$. \fig{TenDeclabel} shows the landscape $f$ for $d =3$, where some minima are labeled. The local ground state of the well $(j,\pi_{\alpha})$ is denoted by $|j,\pi_{\alpha}\rangle$. In the basis $\{|1,+\rangle,\ldots,|d,+\rangle,|1,-\rangle,\ldots,|d,-\rangle\}$ which spans a low-energy subspace $\mathcal{F}$, the interaction matrix modulus an exponential error has the form \begin{equation} H_{|\mathcal{F}} = \left( \begin{array}{cccc|cccc} \mu & w & \cdots & w & 0 & w & \cdots & w \\ w & \mu & \ddots & \vdots & w & 0 & \ddots & \vdots \\ \vdots &\ddots & \ddots& w & \vdots & \ddots &\ddots &w \\ w & \cdots& w & \mu & w & \cdots& w & 0 \\ \hline 0 & w & \cdots & w & \mu & w & \cdots & w\\ w & 0 &\ddots & \vdots & w & \mu & \ddots & \vdots\\ \vdots & \ddots&\ddots & w & \vdots & \ddots & \ddots& w\\ w & \cdots& w & 0 & w & \cdots& w & \mu \\ \end{array} \right). \label{eq:TenDec} \end{equation} Here the quantity $\mu$ stands for the energy of local ground states, and $w$ is the tunneling amplitude quantifying the interactions between wells. To understand \eq{TenDec}, for any $j$, imagine a sphere where $(j,+)$ is the north pole and $(j,-)$ the south pole, then for $j'\neq j$, $(j',\pm)$ are evenly distributed on the equator. The energy of all local ground states are the same because of the symmetry. So, diagonal elements of $H_{|\mathcal{F}}$ are all $\mu$. The interactions between $(j,+)$ and $(j',\pm)$ for all $j'\neq j$ should be the same as well. However, the interaction between $(j,+)$ and $(j,-)$ is exponentially weaker due to the longer distance between $(j,+)$ and $(j,-)$. As a result, we write $\bra{j,+}H_{|\mathcal{F}}\ket{j,-} = 0$ modulus an exponential error and all other off-diagonal elements as $w$. As is demonstrated in previous sections, the time cost of QTW highly depends on spectral gaps of $H_{|\mathcal{F}}$. The following lemma studies eigenstates and eigenvalues of $H_{|\mathcal{F}}$ (see proof in \append{C.3.1}). \begin{lemma}\label{lem:TenDecEig} The eigenstates and corresponding eigenvalues of $H_{|\mathcal{F}}$ in \eq{TenDec} are given by \begin{align} |E_k\rangle &= \frac{1}{\sqrt{2d}}\sum_j e^{i\frac{2\pi}{d}kj}(|j,+\rangle+ |j,-\rangle),~k=1,\ldots,d \\ |E_k\rangle &= \frac{1}{\sqrt{2}}(|k-d,+\rangle - |k-d,-\rangle),~k=d+1,\ldots,2d \end{align} where \begin{align} E_k &= \mu - 2w,~k=1,\ldots,d-1 \\ E_d &= \mu + 2w(d-1), \\ E_k &= \mu,~k=d+1,\ldots,2d. \end{align} \end{lemma} Evolving the system for at least the mixing time, the measured results would be subject to the limit distribution $\mu_{\rm QTW}$. Since $\mu_{\rm QTW}$ would concentrate near all minima, we are able to find all components. Combined with results of \sec{Qmixing}, we can get the mixing time of the QTW as follows (the proof is postponed to \append{C.3.2}): \begin{lemma}\label{lem:TenDecmix} For the landscape \eq{TenDecObj}, starting from a local ground state $|\alpha\rangle$, the distribution $\rho_{\rm QTW}(\tau,x)$ converges to the limiting distribution $\mu_{\rm QTW}$ obeying the following relation: \begin{align} \|\rho_{\rm QTW} - \mu_{\rm QTW}\|_{L^1(\mathbb{S}^{d-1})} \leq \frac{1}{|w| \tau}(\Theta(1/2)+ O(h^{\infty})). \label{eq:TenDecmix} \end{align} The $\epsilon$-close mixing time is subsequently bounded as \begin{align} T_{\rm mix} = O\bigg(\frac{1}{|w|\epsilon}(1+O(h^{\infty}))\bigg). \end{align} \end{lemma} To determine the total evolution time for finding all components $\{a_j\}$ (i.e., all global minima), we need to calculate $\int_{\Omega_{\beta}}\mu_{\rm QTW} \d x$, where $\Omega_{\beta}$ is an open set containing the minimum $\beta$. According to \lem{qtwhitting}, $\int_{\Omega_{\beta}}\mu_{\rm QTW} \d x$ is the probability of finding the particle in a neighborhood of $\pi_{\beta} a_{j(\beta)}$ and can be captured by the probability of finding the system at the state $\ket{\beta}$. Starting from a local state $|\alpha\rangle$, the probability of hitting $\ket{\beta}$ is given by the following lemma and the proof can be found in \append{C.3.3}. \begin{lemma}\label{lem:TenDecprobability} Initiating at a local state $|\alpha\rangle$ where $\alpha = (j(\alpha), \pi_{\alpha})$, after simulating for a time $t$ which is chosen uniformly from $[0,\tau],~\tau \to \infty$, the limiting distribution represented by the probability of tunneling to a local state $|\beta\rangle$ is given by \begin{equation} p(\infty,\beta | \alpha ) =\left\{\begin{array}{ll} \frac{1}{2d^2},~j(\alpha) \neq j(\beta),\\[3pt] \frac{1}{2}-\frac{(d-1)}{2d^2},~j(\alpha) = j(\beta). \end{array} \right. \end{equation} \end{lemma} Note that the two minima $\pm a_{j}$ are equivalent, representing one component. Thus, starting from $\ket{\alpha}$, we are able to find a component different from $\pm a_{j(\alpha)}$ if the measured result is in a well $\beta$ where $j(\beta) \neq j(\alpha)$. We can define the probability for a successful trial as $p_{\rm suc} = \sum_{j(\beta)\neq j(\alpha)} p(\infty,\beta | \alpha ) = \frac{d-1}{d^2}$. That is, evolving for time $T_{\rm mix}$ as described by \lem{TenDecmix}, we are able to approximately sample from the limiting distribution $\mu_{\rm QTW}$ and then get to another component with probability near $p_{\rm suc}$. The number of trials needed for finding another component is approximately $1/p_{\rm suc}$. And the time needed for finding another component from a known component is approximately $T_{\rm mix}/p_{\rm suc}$. Repeating the procedure of looking for one component that is different from a known one, we can obtain all orthogonal components with total time\footnote{The term $O(d\log d)$ appears because our procedure is equivalent to the Coupon Collector's Problem.} \begin{align} T_{\rm tot} = O(d\log d) T_{\rm mix}/p_{\rm suc} = \tilde{O}(d^2)\frac{1}{\epsilon|w|}. \label{eq:Ttot1} \end{align} To determine the time specifically, it remains to determine $1/|w|$ which depends exponentially on $d$ and $h$. We can obtain: \begin{lemma}\label{lem:TenDecw} For sufficiently small $h$, the tunneling amplitude $w$ in the interaction matrix \eq{TenDec} satisfies \begin{align} w = - \sqrt{h}(C_1 C_2^{d-1} + O(h))e^{-\frac{\sqrt{2}}{2h}}, \end{align} where $C_1$ and $C_2$ are constants depending only on the landscape and are independent of the dimension $d$. \end{lemma} The proof of \lem{TenDecw} is in \append{C.3.4}. It is intuitive to see from \lem{TenDecw} that the smaller $h$ is, the longer time it takes to find all components. However, small $h$ permits more accurate measurement results. A successful tunneling means we can find a point near a new component, but this point may not be the actual minimum. We add a constraint that the expected risk is $\delta$ (i.e., $\mathbb{E}_{x\sim \mu_{\rm QTW}}f(x) -\min f = \delta$). Subsequently, $h$ can be bounded using $\delta$ and we can have the following proposition: \begin{proposition}\label{prop:TenDecTtot} For sufficiently small $\epsilon$ (such that the measured positions nearly obey $\mu_{\rm QTW}$) and sufficiently small expected risk $\delta$ (such that $h$ can be estimated by $\delta$ and \lem{TenDecw} is valid), we have $h = \sqrt{2}\delta/(d-1 + o_{\delta}(1))$ and the total time for finding all orthogonal components of $T$ in \eq{4-tensor} by QTW satisfies \begin{align} T_{\rm tot} = O(\mathrm{poly}(1/\delta, e^d, 1/\epsilon)) e^{\frac{(d-1) +o_{\delta}(1)}{2\delta}}. \label{eq:TenDecTtot} \end{align} \end{proposition} \begin{remark} The strategy we adopt here, which is equivalent to repeating sampling from $\mu_{\rm QTW}$, is straightforward but may not be the optimal one under the framework of QTW. In other words, \prop{TenDecTtot} provides a general upper bound on the total evolution time needed. However, the term $e^{\frac{\sqrt{2}}{2h}}$ which gives the term $e^{\frac{d-1}{2\delta} + o_{\delta}(1)}$ in \eq{TenDecTtot} describes essential difficulty for tunneling through a barrier and would not disappear as long as we use quantum tunneling. \end{remark} To sum up, we provide a scenario that QTW can be used to solve orthogonal tensor decomposition problems. For a practical landscape, the spectrum of the interaction matrix and the mixing time of QTW is explicitly calculated. Running QTW for some time (bounded by the mixing time) repeatedly, we can sample points from a distribution near the limiting distribution and find all tensor components, and an upper bound on the total running time for QTW is derived. \section{Comparison Between Quantum Tunneling Walks and Classical Algorithms}\label{sec:comparison} In this section, we use comparisons between QTW and SGD to explain the advantages of quantum tunneling, resulting in our \hyperref[slog:1]{Main Message}. Because of distinctions between quantum and classical algorithms, preparations (i.e., standards for comparisons) in \sec{standard} are needed before specific comparisons in \sec{illustration}. Having such general understanding of QTW, in \sec{separation}, we further make use of the fact that quantum evolution is essentially global but classical algorithms rely on local queries, so that a hitting problem cannot be solved efficiently by classical algorithms can be tackled by QTW within polynomial queries when given reasonable initial states. \subsection{Criteria of fair comparison}\label{sec:standard} Through out \sec{comparison}, we adopt assumptions in both \sec{classicalpre} and \sec{quantumpre} for the objective landscape $f(x)$ of interest. We still use $U_j = \{x_j \}~(j=1,\ldots,N)$ to denote the wells and $\ket{j}$ the corresponding orthonormalized local ground states. The interaction matrix is $H_{|\mathcal{F}}$, where $H=-h^2\Delta + f(x)$ is the Hamiltonian and $\mathcal{F}$ the low energy subspace spanned by $\{ \ket{j}: j=1,\ldots,N\}$. As shown in \sec{onedimexp}, the hitting time of SGD is determined by the landscape and an adjustable learning rate $s$. Similarly, we can also adjust $h$ in Hamiltonian simulation. Therefore, we need to determine the relationship between $h$ and $s$ for the comparison between the time cost of QTW and SGD. Note that both QTW and SGD have limit distributions, namely, $\mu_{\rm QTW}$ and $\mu_{\rm SGD}$, respectively (see \sec{classicalpre} and \lem{limitdis} for details). If $h$ (or $s$) becomes smaller, $\mu_{\rm QTW}$ ($\mu_{\rm SGD}$) will concentrate more closely to global minima, giving more accurate outputs, whereas it would take more time for the QTW (SGD) to converge. Comparing the running time without specifying accuracy is not fair. In order to establish an relationship between $h$ and $s$, as well as to compare QTW and SGD fairly, we specify some kind of accuracy of the limit distributions. The two variables, $h$ and $s$, will be solved from the demand of accuracy. Hence, the time cost of different algorithms are only related to the accuracy, the dimension, and some geometric properties of the landscapes. There are different measures of accuracy we can choose depending on the tasks faced. Here, we introduce two kinds of measures along with the corresponding standards of comparison. \begin{standard}[Risk accuracy]\label{stand:risk} Let $\mu_{\rm QTW}$ be the limit distribution of QTW, and $\mu_{\rm SGD}$ the invariant Gibbs distribution of SGD. Two distributions are demanded to be $\delta$-risk-accurate: \begin{align} \mathbb{E}_{x\sim \mu_{\rm QTW}}f(x) - \min f = \mathbb{E}_{x\sim \mu_{\rm SGD}}f(x) - \min f = \delta. \label{eq:stand1} \end{align} \end{standard} \stand{risk} ensures that two limit distributions yield the same expected risk. Then, it is natural to compare how fast QTW and SGD would converge. The algorithm spending less time is more efficient on finding any one global minimum. Sometimes, the task is to find some target minima or one special minimum. In this case, using risk accuracy cannot emphasize the particularity of the minima of interest and we may need the following standard: \begin{standard}[Distance accuracy]\label{stand:distance} Let $\mu_{\rm QTW}$ be the limit distribution of QTW, and $\mu_{\rm SGD}$ be the invariant Gibbs distribution of SGD. The minima of interest are $x_{j_k},~k = 1,\ldots,m,~j_k \in \{1,\ldots,N\}$. Let $D(\cdot, \cdot)$ be any distance function. Two distributions are demanded to be $\delta$-distance-accurate with respect to $x_{j_k}$ and $D(\cdot, \cdot)$: \begin{align} \mathbb{E}_{x\sim \mu_{\rm QTW}}\sum_k D(x,x_{j_k}) = \mathbb{E}_{x\sim \mu_{\rm SGD}}\sum_k D(x,x_{j_k}) = \delta. \label{eq:stand2} \end{align} \end{standard} Conditions \eq{stand1} and \eq{stand2} can specify $h$ and $s$. To see this, we first study the expected risk for quadratic functions: \begin{lemma}\label{lem:stand1-quad} Assume the objective function $f\colon \mathbb{R}^d \to \mathbb{R}$ is quadratic and \begin{equation} f(0) = 0,~\nabla f(0) = 0,~\nabla^2 f(0) > 0, \end{equation} where the last inequality means the Hessian $\nabla^2 f(0)$ is positive definite. Then, we have \begin{equation} \mathbb{E}_{x\sim \mu_{\rm QTW}}f(x) = \frac{\sqrt{2}h}{4}\mathrm{tr} \sqrt{\nabla^2 f(0)}, \qquad \mathbb{E}_{x\sim \mu_{\rm SGD}}f(x) = \frac{sd}{4}. \end{equation} \end{lemma} \lem{stand1-quad} calculates the expected risks for a landscape with only one minimum whose proof is in \append{D.1.1}. For landscapes with multiple minima, the limit distributions concentrate near the global minima and the objective function in a small neighborhood of any minimum can be approximated by a quadratic function based on the assumptions. Hence, we can obtain the following general estimations (the proof is postponed to \append{D.1.2}). \begin{lemma}\label{lem:expectedrisk} If $\delta$ is sufficiently small and the objective function $f\colon \mathbb{R}^d \to \mathbb{R}$ satisfies satisfying assumptions in \sec{classicalpre} and \sec{quantumpre}, then \stand{risk} gives \begin{equation} h = \frac{\delta}{\frac{\sqrt{2}}{4}\sum_{j=1}^N p(\infty,j)\mathrm{tr}\big(\sqrt{\nabla^2f(x_j)}\big)+ o_{\delta}(1)}, \label{eq:hstand1} \end{equation} \begin{equation} s = \frac{\delta}{\frac{d}{4}(1+ o_{\delta}(1))}. \end{equation} \end{lemma} That is, we establish a relationship between $h$ and $s$ by \stand{risk}. Similarly, for \stand{distance}, we can have the following result: \begin{lemma}\label{lem:expecteddis} Assume the objective function $f\colon \mathbb{R}^d \to \mathbb{R}$ is quadratic and \begin{equation} f(0) = 0,~\nabla f(0) = 0,~\nabla^2 f(0) > 0, \end{equation} where the last inequality means the Hessian $\nabla^2 f(0)$ is positive definite. We define the distance function $D(x,y):=\|x-y\|^2_2,~\forall x,y\in \mathbb{R}^d$. Then, we have \begin{align} \mathbb{E}_{x\sim \mu_{\rm QTW}}D(x,0) &= \frac{\sqrt{2}h}{2}\mathrm{tr} (\nabla^2 f(0))^{-1/2}; \\ \mathbb{E}_{x\sim \mu_{\rm SGD}}D(x,0) &=\frac{s}{2}\mathrm{tr}(\nabla^2 f(0))^{-1}. \end{align} \end{lemma} The proof of \lem{expecteddis} is shown in \append{D.1.3}. Similar to the process from \lem{stand1-quad} to \lem{expectedrisk}, \lem{expecteddis} may be generalized to general landscapes. However, the generalization of \lem{expecteddis} is quite complicated as the distance function and the wells of interest are arbitrary. So, we stop at \lem{expecteddis}. Regardless of different standards, \lem{expectedrisk} and \lem{expecteddis} present some similar intuition: the dependence of $h$ on the flatness of wells are different from that of $s$, which is going to be shown in the following section as a source of quantum speedups.\footnote{Here, we use the Hessian matrix of $f$ at minima to quantify the concept ``flatness".} \subsection{Illustrating advantages of quantum tunneling}\label{sec:illustration} In this subsection, we compare QTW with SGD for several special landscapes. The goal is to explore geometric properties of the landscapes that affect relative efficiencies of QTW and SGD. Heuristically, the comparison reveals when quantum tunneling can be faster than thermal climbing (climbing over barriers between minima by stochastic motions), which are the two mechanisms behind QTW and many classical algorithms. For simplicity, we focus on the following kind of landscapes: \begin{definition}[One-dimensional partially periodic functions]\label{defn:funcforillus} A function $f\colon\mathbb{R}\to \mathbb{R}$ is partially periodic if it satisfies the assumptions in \sec{classicalpre} and \sec{quantumpre}, and all minima $\{x_j:j=1,\ldots,N\}$ are in a bounded interval which is a period of $f$. \end{definition} \begin{figure} \centerline{ \includegraphics[width=0.8\textwidth]{oneperiodic.pdf}} \caption{A one-dimensional partially periodic function. } \label{fig:oneperiodic} \end{figure} A sketch of functions in \defn{funcforillus} is shown in \fig{oneperiodic}. Neglect an exponentially small error and note the symmetry of the one-dimensional partially periodic function $f$, the interaction matrix under $\{ \ket{j}:j=1,\ldots,N\}$ should be given by \begin{equation} H_{|\mathcal{F}} = \left( \begin{array}{cccccc} \mu & w & & & & \\ w & \mu & w & & & \\ & w & \mu & w & & \\ & & \ddots& \ddots & \ddots & \\ & & & w & \mu & w\\ & & & & w & \mu\\ \end{array} \right), \label{eq:Hforline} \end{equation} where $\mu$ is the energy of one local ground state and $w$ quantifies the tunneling effect between two adjacent wells. Eigenstates and eigenvalues of $H_{|\mathcal{F}}$ can be given by the following lemma. \begin{lemma}\label{lem:IlluEig} The eigenstates and corresponding eigenvalues of the Hamiltonian \eq{Hforline} are given by \begin{align} |E_k\rangle &= \sqrt{\frac{2}{N+1}}\sum_{j=1}^{N}\sin\big(\frac{jk\pi}{N+1}\big)\ket{j},~k=1,2,\ldots,N; \\ E_k &= \mu + 2w \cos \frac{k\pi}{N+1},~k=1,2,\ldots,N. \end{align} \end{lemma} To describe $w$ in detail, as shown in \fig{oneperiodic}, we introduce new notations $\{x^{\bullet}_j: j=1,\ldots,N\}$ and $\{x^{\circ}_j: j=1,\ldots,N-1\}$ to denote minima and saddle points, respectively. A more general labeling of local minima and saddle points can be found in \append{A.1}. The Morse saddle barrier reflecting height of the barrier in the present case can be given by $H_f = f(x^{\circ}_1) - f(x^{\bullet}_1)$. Using results in \append{A.2.3}, we have: \begin{lemma}[Tunneling amplitude] The tunneling amplitude for the one-dimensional partially periodic function $f$ is given by \begin{align} w = -2\sqrt{\frac{h f''(x^{\bullet}_1) H_f}{\sqrt{2}\pi}} e^{\int_{x_1^{\bullet}}^{x_1^{\circ}} (\sqrt{\frac{ f''(x^{\bullet}_1)}{2f(\xi)}} - \frac{1}{\xi- x_1^{\bullet}})\d \xi } e^{-\frac{S_0}{h}},~\mathrm{where}~S_0 = \int_{x_1^{\bullet}}^{x_2^{\bullet}} \sqrt{f(\xi)} \d\xi. \end{align} \end{lemma} Now, we can obtain the spectrum of $H_{|\mathcal{F}}$ explicitly, and proceed by using \lem{qtwmixingtime} to get the quantum mixing time. \begin{lemma}[Quantum mixing time]\label{lem:illuqtwmixing} Staring from one local ground state of one minimum, the $\epsilon$-close mixing time of QTW is given by \begin{align} T_{\rm mix}^{\rm QTW} = O \bigg(\frac{N^3}{\epsilon |w|} [1+(N-1)|O(h^{\infty})|] \bigg) = O(\mathrm{poly}(N,1/h,1/\epsilon))e^{\frac{S_0}{h}}. \end{align} \end{lemma} Regarding SGD, we use the results introduced in \sec{classicalpre} to estimate the classical mixing time. First, \begin{lemma}[Exponential decay constant] In \prop{mic19}, let $\lambda = \delta_{s,1}/2s$ we have \begin{align} \lambda = \bigg(\frac{\sqrt{f''(x^{\circ})f''(x^{\bullet})}}{2\pi} + o(s)\bigg) e^{-\frac{2 H_f}{s}}. \end{align} \end{lemma} Then, by \cor{sgdmixing}, the following lemma holds. \begin{lemma}[Classical mixing time]\label{lem:illusgdmixing} Let $T_{\rm mix}^{\rm SGD}$ be the SGD $\epsilon$-close mixing time which is the minimum time enabling $\|\rho_{\rm SGD} (t,\cdot) - \mu_{\rm SGD}\|_{\mu_{\rm SGD}^{-1}}<\epsilon$, we have \begin{align} T_{\rm mix}^{\rm SGD} = O \bigg(\frac{1}{\lambda} \ln\frac{\|\rho(0,\cdot) - \mu_{\rm SGD}(\cdot) \|_{\mu^{-1}_{\rm SGD}}}{\epsilon} \bigg) = O(\mathrm{poly}(1/s,\ln(1/\epsilon)))e^{\frac{2H_f}{s}}. \end{align} \end{lemma} Later, we do not focus on the dependence of the mixing time on $\epsilon$, as the norms ($L^1$ norm for QTW and $L^2(\mu^{-1}_{\rm SGD})$ for SGD) used to capture convergence are different.\footnote{ In terms of $\epsilon$, the same argument in \cite{AC21} but with evolution time $t$ of QTW chosen as a sum of some random variables instead of chosen uniformly in an interval, $T_{\rm mix}^{\rm QTW}$ can also achieve $\ln(1/\epsilon)$ dependence instead of $1/\epsilon$.} The dominant terms affecting running time of QTW and SGD are $e^{\frac{S_0}{h}}$ and $e^{\frac{2H_f}{s}}$. \begin{lemma}[Comparison on one-dimensional periodic landscapes] Under \stand{risk}, let QTW and SGD be both $\delta$-accurate. For sufficiently small $\delta$, the QTW mixing time and SGD mixing time are dominated by \begin{align} LT_{\rm mix}^{\rm QTW}:=e^{\frac{\sqrt{2}S_0f''(x^{\bullet})}{4\delta}}~\mathrm{and}~LT_{\rm mix}^{\rm SGD}:=e^{\frac{H_f}{2\delta}}, \textrm{ respectively.} \end{align} \end{lemma} \begin{figure} \centerline{ \includegraphics[width=0.5\textwidth]{illuexamples.pdf}} \caption{The landscapes in \examp{critical}, \examp{flatness}, and \examp{sharp} (corresponding to Example 1, 2, and 3 in the figure, respectively) for illustrating the comparison between QTW and SGD. } \label{fig:illuexamples} \end{figure} As concrete examples, we present several specific functions to illustrate the advantages of quantum tunneling. Since the function in the region of our interest is periodic, we only need to specify the function value within one period to construct a concrete example. Without loss of generality, we set the interval $[-a,a+2b]$ to be one period, where $[-a,a]$ is called the well region and $[a,a+2b]$ the barrier region. The constructed landscape in $[-a,a+2b]$ is given by \begin{align} f(x) = \left\{ \begin{array}{ll} \frac{1}{2}k x^2\quad x\in [-a,a],\\[3pt] \frac{1}{2\pi \sigma^2}\exp\big(-\frac{(x-a-b)^2}{\sigma^2}\big) + \frac{1}{2}k a^2 - \varepsilon\quad x \in (a,a+2b]. \end{array} \right. \label{eq:illuVx} \end{align} Here, to reduce free parameters, we make $f(x)$ differentiable at $a$, the boundary of the well, and the barrier, such that \begin{align} b = \sigma \sqrt{\ln \frac{1}{2\pi \sigma^2 \varepsilon}}, \quad k = \frac{2\varepsilon}{a\sigma}\sqrt{\ln \frac{1}{2\pi \sigma^2 \varepsilon}}. \label{eq:illucondition} \end{align} \begin{remark} Note that the function in \eq{illuVx} is not smooth. We need to use the mollifier function $m_{r}$ (see detials in \append{D.3}) to smooth it such that assumptions in \sec{classicalpre} and \sec{quantumpre} are satisfied. Note that if $r\to 0$, the smoothed function will tend to be $f$, following results can be seen as arbitrarily accurate for a smooth function arbitrarily close to \eq{illuVx}. \end{remark} By giving specific $a$, $\sigma$, and $\varepsilon$ in \eq{illuVx}, we can design landscapes with different properties. Detailed variables, discussions and comparisons are given below. \begin{example}[Critical case]\label{examp:critical} For \eq{illuVx}, we set $a = 5.0$, $\sigma = 1.0$ and $\varepsilon=0.15$, and obtain $b\approx 0.243$ and $k\approx 0.0146$ by \eq{illucondition}. \end{example} \begin{example}[Flatness of minima]\label{examp:flatness} For \eq{illuVx}, we set $a = 5.0$, $\sigma = 1.0$ and $\varepsilon=0.009$, and obtain $b\approx 1.69$ and $k\approx 0.00610$ by \eq{illucondition}. \end{example} \begin{example}[Sharpness of barriers]\label{examp:sharp} For \eq{illuVx}, we set $a = 5.0$, $\sigma = 0.5$ and $\varepsilon=0.0088$, and obtain $b\approx 1.03$ and $k\approx 0.0146$ by \eq{illucondition}. \end{example} \fig{illuexamples} explicitly shows the shapes of above examples. The barrier region in \examp{critical} is small and most of the function in one period is quadratic, which is similar to the case introduced in \sec{onedimexp}. The Morse saddle barrier $H_f$ of \examp{flatness} is approximately equal to that of \examp{critical}, whereas, in \examp{flatness}, the well is more flat and the barrier is thicker. \examp{sharp} has almost the same well as \examp{critical} but is equipped with a much higher barrier. We call \examp{critical} as the critical case because QTW and SGD perform nearly the same on it in terms of the leading terms $LT_{\rm mix}^{\rm QTW}$ and $LT_{\rm mix}^{\rm SGD}$: \begin{lemma} \examp{critical} satisfies $b\ll a$ and $\frac{1}{2}ka^2 \approx H_f$. For such a landscape, we have \begin{align} \ln LT_{\rm mix}^{\rm QTW} &= \frac{H_f}{2\delta}\left[1+ 2b/a+ o(b/a) \right] + o(\delta),\\ \ln LT_{\rm mix}^{\rm SGD} &= \frac{H_f}{2\delta}(1+o(\delta)). \end{align} \end{lemma} QTW mixes faster on both \examp{flatness} and \examp{sharp} for sufficiently small $\delta$. Specifically, we have \begin{lemma} For \examp{flatness} and \examp{sharp}, the following holds \begin{align} \ln LT_{\rm mix}^{\rm QTW} &< \frac{k}{4\delta}a^2+ \frac{k}{2\delta}ab + \frac{\sqrt{2k}}{4\delta} + o(\delta),\\ \ln LT_{\rm mix}^{\rm SGD} &= \frac{1}{4\pi \sigma^2\delta}+\frac{k}{4\delta}a^2-\frac{\epsilon}{2\delta} + o(\delta). \end{align} Substituting the parameters, it is true for both \examp{flatness} and \examp{sharp} that \begin{align} \frac{k}{4}a^2+ \frac{k}{2}ab + \frac{\sqrt{2k}}{4} < \frac{1}{4\pi\sigma^2}+\frac{k}{4}a^2-\frac{\epsilon}{2}. \end{align} \end{lemma} Comparing to the critical case \examp{critical}, \examp{flatness} has a thicker barrier, which increases $S_0$ and causes difficulty for QTW. However, QTW can perform better in \examp{flatness}. This is mainly due to the more flat well of \examp{flatness}. Recall that by \stand{risk}, to ensure $\delta$-risk-accuracy, $h$ and $s$ should be \begin{align} h= \frac{\delta}{\frac{\sqrt{2k}}{4} + o_{\delta}(1)} \quad\mathrm{and}\quad s= \frac{\delta}{\frac{1}{4}(1+o_{\delta}(1))}, \end{align} respectively. That is, under the same risk accuracy, $h$ can be much larger than $s$ if the well is flat ($k$ is small), making tunneling easier. Note that there is a trade-off between accuracy and time cost: smaller $h$ (or $s$) ensures high accuracy but make tunneling effects (or thermal diffusion) weaker; conversely, larger $h$ (or $s$) permits faster tunneling (or diffusion) but yields inaccurate results. Discussions on quantum tunneling effects usually focus on properties of the barrier. In the present study, since we aim to find global minima, the precision of results obtained is one important concern. Therefore, the flatness of wells, which affects differently on the accuracy of QTW and SGD, is a crucial property determining the runtime of QTW and SGD. Loosely speaking, QTW is faster than SGD on landscapes with flat wells. \examp{sharp} adheres to the intuition that quantum tunneling is efficient on functions with tall and thin barriers. The wells of \examp{sharp} are almost the same as those of the critical case \examp{critical}. QTW can be faster in \examp{sharp} because we add a sharp barrier between wells. By \lem{illusgdmixing}, a high barrier (i.e., large $H_f$) would significantly hinder thermal climbing. However, the tall barrier is sufficiently thin, such that $S_0 = 2\int_{0}^{a+b}\sqrt{f(x)}\d x$ can still be small and by \lem{illuqtwmixing}, the tunneling effect would be strong. Moreover, in high dimensions, the distribution of wells can be very different from being on a line. As shown in \append{D.4}, distribution of wells can largely affect the dependence of time on $N$. However, such relation between the distribution of wells and running time is not explicitly shown for SGD. Therefore, the distribution of wells can also be a factor of quantum speedups. In summary, we can conclude our \hyperref[slog:1]{Main Message}. \subsection{Efficient quantum tunneling for solving a classically difficult hitting problem}\label{sec:separation} The above examples compare QTW driven by quantum tunneling with SGD. In this section, an exponential separation in terms of query complexity between QTW given initial states and classical algorithms knowing one well will be shown for a specific hitting problem on a constructed landscape. The landscape $f(\mathbf{x})$ we construct lives in $\mathbb{R}^d$. We use $\|\cdot\|$ to denote the $\ell_2$ norm of vectors, namely, $\|\mathbf{x}\| = \sqrt{\mathbf{x}\cdot \mathbf{x}}$. Let $\mathbb{B}(\mathbf{x},r)$ denote a $d$-dimensional ball centered at $\mathbf{x}$ with radius $r$. A special direction $\mathbf{v}$ is randomly chosen from the $d$-dimensional unit sphere. We define two regions $W_- = \mathbb{B}(\mathbf{0},a)$ and $W_+ = \mathbb{B}(2b\mathbf{v},a)$ with $b\geq a$. Let $R$ be sufficiently large s.t. $W_-, W_+ \subset \mathbb{B}(\mathbf{0},R)$. We denote the region $\{\mathbf{x}\mid\mathbf{x} \in \mathbb{B}(\mathbf{0},R), ~|\mathbf{x}\cdot \mathbf{v}| \leq w\}$ by $S_{\mathbf{v}}$, where $w$ will be chosen from $[\sqrt{3}a/2,0)$. We denote \begin{align} B_{\mathbf{v}}:=\{\mathbf{x}\mid w< \mathbf{x}\cdot \mathbf{v} < 2b-w,~\sqrt{\|\mathbf{x}\|^2 - (\mathbf{x}\cdot \mathbf{v})^2} < \sqrt{a^2-w^2}, \bf{x}\notin W_-\cup W_+\}. \end{align} \fig{provable-acceleration} illustrates positions of the newly defined regions. The constructed function $f$ is given by \begin{align} f(\mathbf{x}) = \left\{ \begin{array}{ll} \frac{1}{2}\omega^2 \|\mathbf{x}\|^2,~\mathbf{x} \in W_-, \\[3pt] \frac{1}{2}\omega^2 \|\mathbf{x}-2b\mathbf{v}\|^2,~\mathbf{x} \in W_+, \\[3pt] H_1,~\mathbf{x} \in B_{\mathbf{v}},\\[3pt] H_2,~\mathrm{otherwise}. \end{array} \right. \label{eq:hardinstance} \end{align} Here, we define $H_0 = \frac{1}{2}\omega^2a^2$ and demand that $0<H_0 \sim H_1 \ll H_2$. \begin{remark} The landscape $f$ in \eq{hardinstance} is not smooth and should be smoothed to be $F_r$ with the help of a mollifier function $m_r$ (see details in \append{D.3}) such that assumptions in \sec{quantumpre} can be satisfied. Because when $r\to 0$, $F_r\to f_r$, we can always find sufficiently small $r$ to make the following conclusions based on $f$ valid for $F_r$. \end{remark} There are two global minima, $\mathbf{0}$ and $2b\mathbf{v}$, of the function $f$. Given that we know $\mathbf{0}$ is a minimum, our goal is to find the other one. To avoid complicated justifications, we deal with a simpler problem: \begin{problem}\label{prb:provable} For the $f$ in \eq{hardinstance}, given that we only know $\mathbf{0}$ is a global minimum, find any point in $W_+$. \end{problem} \subsubsection{Classical lower bound}\label{sec:clb} Due to the concentration of measure, for any point $\mathbf{x} \in \mathbb{B}(\mathbf{0},R)$, the probability of $\mathbf{x}\in S_{\mathbf{v}}$ is given by \begin{align} P(\mathbf{x} \in S_{\mathbf{v}}) \geq 1 - O \big(e^{-\frac{dw^2}{2R^2}}\big). \label{eq:PxinSv} \end{align} Intuitively, restricted in $\mathbb{B}(\mathbf{0},R)$, any classical algorithm cannot escape from $S_{\mathbf{v}}$ efficiently. In $\mathbb{R}^d$, queries out of $\mathbb{B}(\mathbf{0},R)$ provide no information about the landscape inside $\mathbb{B}(\mathbf{0},R)$ and are unable to help to escape from $S_{\mathbf{v}}$. Therefore, classical algorithms cannot solve \prb{provable} efficiently with or without being constrained in $\mathbb{B}(\mathbf{0},R)$. To rigorously prove above intuitions, we first introduce a mathematical result indicating \eq{PxinSv}: \begin{lemma}[Measure concentration for the sphere]\label{lem:mconcentration} Let $\mathbb{S}^{d-1} = \{\mathbf{x}:\|\mathbf{x}\| =1 \}$ be the unit sphere in $\mathbb{R}^d$. Let $\mathrm{Cap}(\epsilon)$ denote the spherical cap of height $\epsilon$ above the origin (see the left part of \fig{capandcone}). We have \begin{align} \frac{\mathrm{Area}(\mathrm{Cap}(\epsilon))}{\mathrm{Area}(\mathbb{S}^{d-1})} \leq e^{-d\epsilon^2/2}. \end{align} \end{lemma} \begin{figure} \centerline{ \includegraphics[width=0.5\textwidth]{capandcone.pdf}} \caption{Estimating the area of a spherical cap. } \label{fig:capandcone} \end{figure} The estimation details are presented in \append{D.2.1}. Subsequently, it is readily to have (see details in \append{D.2.2}): \begin{lemma}\label{lem:provable-onep} For any randomly chosen point $\mathbf{x}\in \mathbb{B}(\mathbf{0},R)$, the probability of $\mathbf{x}\notin S_{\mathbf{v}}$ is $P(\mathbf{x}\notin S_{\mathbf{v}})\leq 2 e^{-\frac{dw^2}{2R^2}}$. \end{lemma} Recall that $w\in [a/2,a)$ and $R$ are independent of $d$, the measure of the region in $\mathbb{B}(\mathbf{0},R)$ and outside $S_{\mathbf{v}}$ is exponentially small with respect to the dimension $d$. By \defn{localquery}, classical algorithms depend on an adaptive sequence of points. we now need to demonstrate that it is difficult for the points to hit regions beyond $S_{\mathbf{v}}$. \begin{lemma}\label{lem:provable-manyp} For any classical algorithm (see \defn{localquery}), after running $T$ times, we get a sequence of points and corresponding queries $\{\mathbf{x}_i, q(\mathbf{x}_i)\}_{i=1}^T$. Restricted in $\mathbb{B}(\mathbf{0},R)$, as long as any $q(\mathbf{x}_i)~(\mathbf{x}\in S_{\mathbf{v}})$ is independent of $\mathbf{v}$, the probability $P(\exists t\leq T:\mathbf{x}_t\notin S_{\mathbf{v}})\leq 2T e^{-\frac{dw^2}{2R^2}}$. \end{lemma} We prove \lem{provable-manyp} in \append{D.2.3}. Now, we can prove that if the number of points and queries is small, with high probability, any classical algorithm cannot escape from $S_{\mathbf{v}}$. Rigorously, we have \begin{proposition}[Classical lower bound]\label{prop:provable-exp} Any classical algorithm (\defn{localquery}) will fail, with high probability, to solve \prb{provable} given only $o(e^{\frac{dw^2}{4R^2}})$ local queries with or without being restricted in $\mathbb{B}(\mathbf{0},R)$. \end{proposition} The proof sketch of \prop{provable-exp} goes as follows (see proof details in \append{D.2.4}). By \lem{provable-manyp}, it suffices to demonstrate that restricted in the ball $\mathbb{B}$, classical algorithms cannot escape from $W_-$ and hit $W_+$ efficiently. The left thing is to show that queries outside $\mathbb{B}(\mathbf{0},R)$ provide no information about $\mathbb{B}(\mathbf{0},R)$. And thus, without being restricted in $\mathbb{B}(\mathbf{0},R)$, classical algorithms still cannot hit $W_+$ by subexponential queries with high probability. \subsubsection{Quantum upper bound}\label{sec:qub} We now focus on the time needed for quantum tunneling to solve \prb{provable}. The landscape \eq{hardinstance} satisfies \blue{Assumption A.6} \begin{align} 0 = \min f < \lim_{\|\mathbf{x}\|\to \infty} f = H_2,\quad f^{-1}(0) = \{ \mathbf{0}\} \cup \{2b\mathbf{v}\}, \end{align} where $U_-:=\{ \mathbf{0}\}$ and $U_+ := \{2b\mathbf{v}\}$ are called as wells by definition. The neighborhoods of the two wells are quadratic, enabling the wells and corresponding local ground states to satisfy (\blue{S79}) and (\blue{S80}). Moreover, due to the symmetry of the function \eq{hardinstance}, the local ground states are also symmetric. Therefore, \blue{Assumption A.5} can be satisfied. To use \blue{Proposition A.6}, we only need to verify the conditions in \blue{Assumption A.7}, leading to the following lemma. \begin{lemma}\label{lem:provableS0} There exists a unique Agmon geodesic, denoted $\boldsymbol{\gamma}_{-+}: \mathbb{R} \to \mathbb{R}^d$, which links $U_-$ and $U_+$: \begin{equation} \boldsymbol{\gamma}_{-+}(s) = s \mathbf{v},\quad s\in [0,2b]. \label{eq:provable-geodesic} \end{equation} And the Agmon distance $S_0 := d(U_-, U_+)$ is \begin{equation} S_0 = \int_{-1/2}^{1/2} \sqrt{f(\boldsymbol{\gamma}_{-+}(s))} \d s = \frac{1}{\sqrt{2}}\omega a^2 + 2(b-a)\sqrt{H_1}. \label{eq:provable-S0} \end{equation} \end{lemma} The calculation details of \lem{provableS0} are presented in \append{D.2.5}. We are now ready to calculate the interaction matrix explicitly (see details in \append{D.2.6}): \begin{lemma}\label{lem:provablenu} Under the two orthonormalized local ground states, The interaction matrix is of the form \begin{equation} P = \left( \begin{array}{cc} \mu & \nu\\ \nu & \mu \end{array} \label{eq:provable-matrix} \right), \end{equation} and the next-to-leading order formula of $w$ is given by \begin{align} \nu = -\sqrt{\frac{2h}{\pi}} \sqrt{\frac{H_1(\sqrt{2}\omega)^d}{(4\sqrt{H_1}/b)^{d-1}}}\exp\left(-\frac{S_0}{h} + \frac{\omega d(b-a)}{\sqrt{2H_1}} - 2d \ln \frac{b}{a}\right). \label{eq:provable-w} \end{align} \end{lemma} Using the explicit tunneling amplitude, we can estimate the time needed for quantum tunneling. \begin{proposition}[Quantum upper bound]\label{prop:provable-poly} For any dimension $d$, we can always choose appropriate $h$, $\omega$, $a$, $b$, $H_1$, $H_2$, and $w$ satisfying previous restrictions, such that, given the local ground state associated to $W_-$ under the choosing $h$ as initial state, QTW can solve \prb{provable} with high probability $1-(1-C)^n$ using only $nO(\mathrm{poly}(d))$ queries, where $0<C<1$ is a constant independent of $d$. \end{proposition} \begin{remark} In \prop{provable-poly}, the constant $C$ can be understood as the probability of successful hitting in one trial and $n$ the number of trails. To reach a high probability of success, say $99$\%, the number of trials needed, $M$, enabling $1-(1-C)^M\geq 99$\%, is a constant independent of $d$. Since one trial needs only one initial state, only a constant number of copies (e.g., $M$ copies) of the local ground state are needed. \end{remark} The proof of \prop{provable-poly} is postponed to \append{D.2.7} which is explained briefly as follows. We take all the adjustable parameters as functions of $d$ and discuss the evolution time as a function of $d$. First, we have $h= \Theta(1/d)$ for sufficiently large $d$, which can eliminate the negative effects of measure concentration brought by increasing dimension, and on the other hand we prove that our theory on quantum tunneling walks is still valid. Thus, the quantum wave distributes near $W_-$ or $W_+$, and the limit distribution $\mu_{\rm QTW}$ permits a probability of finding the particle in $W_+$ larger than some constant independent of $d$. Then, based on the results of semi-classical analysis, we can tune the function values in $W_-$, $W_+$, and $B_{\mathbf{v}}$ such that the time needed for tunneling is a polynomial of $d$. As a result, the last three conditions at the end of \sec{intro} can be satisfied, and the first and third conditions suggest that with high probability, QTW can hit $W_+$ with queries polynomial in $d$. Finally, given the fact that we can use $\tilde{O}(t)$ quantum queries to evolve QTW for time $t$, with high probability QTW can hit $W_+$ with queries polynomial in $d$. Combining the results of \prop{provable-poly} and \prop{provable-exp}, we can obtain \thm{provableinformal}, which is restated in a more rigorous way as follows. \begin{theorem}\label{thm:thm2re} For any dimension $d$, there exists a landscape with the form \eq{hardinstance} such that with a high probability $1-(1-A)^n$, QTW can solve \prb{provable} with $nO(\mathrm{poly}(d))$ queries given the local ground state associated to $U_-$, but with a high probability $1-e^{-dB}$, no classical algorithm (\defn{localquery}) can solve the same problem for the same landscape $f$ with $o(e^{dB})$ queries, where $0<A<1$ and $B>0$ are two constants independent of $d$. \end{theorem} \subsubsection{The significance of proper initial states}\label{sec:no-exp-sep-init} The hardness of \prb{provable} can be abstracted as that of finding an exponentially small cone on a landscape which is isotropic outside the special cone.\footnote{Specifically, this cone is a region associated to the special direction $\bf{v}$, $\{\bf{x}:\bf{x}\cdot \bf{v}/\|\bf{x}\|<C \}$ for some constant $C$, which contains the parts $B_{\bf{v}}$ and $W_+$.} There can be exponentially many such cones disjoint with each other. Therefore, it can be proved that by solving \prb{provable} in $\mathbb{R}^d$, we can solve an unstructured search problem with a size exponential in $d$. To show this, we first introduce unstructured search. Say, we are given $N$ data points, only one of which is assigned the value $1$ and all other points are assigned a value $0$. The goal is to find the point assigned $1$ with an oracle outputting the assigned value of the input point. Intuitively, each data point can be mapped to a unique cone in $\mathbb{R}^d$ and the point assigned $1$ should correspond to the cone containing $B_{\bf{v}}$ and $W_+$. In this case, solving \prb{provable} can lead to the data point we want to find. Precisely speaking, if there is a quantum algorithm that can solve \prb{provable} with queries polynomial in $d$, it can solve an unstructured search whose size $N$ is exponential in $d$ within queries polynomial in $d$. That is, we can solve an $N$-size unstructured search within $O(\mathrm{poly}(\log N))$ queries with the help of the efficient algorithm for \prb{provable}. However, it is well known that quantum algorithms have a query complexity lower bound $\Omega(\sqrt{N})$ in solving unstructured search with $N$ data points \cite{BBBV97}. Therefore, we can conclude \begin{proposition}\label{prop:nospeedup} No quantum algorithm can solve \prb{provable} within queries polynomial in $d$. \end{proposition} We prove \prop{nospeedup} and related claims rigorously in \append{D.2.8}. It seems that \prop{nospeedup} contradicts with \prop{provable-poly}. But there is actually no paradox as in \prop{provable-poly} QTW does not solve \prb{provable} faithfully. The local ground state $\ket{\Phi_-}$ associated to $U_-$ under proper quantum learning rate $h$ is given to QTW as prior knowledge. To establish polynomial decay tunneling effect, the state $\ket{\Phi_-}$ has non-vanishing probability (maybe an inverse polynomial of $d$) in $W_+$. The state $\ket{\Phi_-}$ indicates a lot about the special direction $\bf{v}$ for QTW, such that what QTW does cannot be equivalent to unstructured search. Indeed, by the same spirit of \prop{nospeedup}, the state $\ket{\Phi_-}$ cannot be prepared within polynomial queries, or we can reach $W_+$ efficiently by measuring $\ket{\Phi_-}$ repeatedly. We admit that \thm{thm2re} requires the initial quantum state. Note that QTW only uses $M$ copies of the local ground state $\ket{\Phi_-}$ to hit $W_+$ with high probability in polynomial time, where $M$ is a number independent of $d$. If the possibility of learning about $\bf{v}$ from sampling tends to $0$ when $d\to \infty$, which is likely to be true, the expected queries needed by classical algorithms to hit $W_+$ cannot be subexponential in $d$. In this case, we have an exponential quantum-classical separation in evaluation queries even classical algorithms are given a constant number of samples from the initial distribution $|\ip{\bf{x}}{\Phi_-}|^2$. Essentially, this is because no classical evolution can make good use of the samples of the initial state. \section{Numerical Experiments}\label{sec:num} We conduct numerical experiments to examine our theory. All results and plots are obtained by simulations on a classical computer (Dual-Core Intel Core i5 Processor, 16GB memory) via MATLAB 2020b. Details of all numerical settings can be found in \append{E}. QTW is simulated by solving the Schr\"odinger equation by numerical methods, and SGD is performed with first-order queries and the noise of each step follows the standard Gaussian distribution. \paragraph{Quantum-classical comparisons.} To corroborate our \hyperref[slog:1]{Main Message}, we numerically study the performance of QTW and SGD on concrete examples (see details in \append{E.1}). The quantum learning rate $h$ and the classical learning rate $s$ are determined under \stand{risk} which equalizes expected risks yielded by QTW and SGD. The task is to hit a target neighborhood of one minimum beginning at another designated minimum. In \fig{histogramsExp123}, results on classical and quantum hitting time are shown. We examine QTW and SGD on three landscapes, Example 1, 2, and 3 in \fig{histogramsExp123} which correspond to concrete functions given by \examp{critical}, \examp{flatness}, and \examp{sharp} in \sec{illustration}, respectively. We conduct 1000 experiments for QTW and SGD on each example. For QTW, we use an \emph{experiment} to denote a process repeating trials until successfully hitting, where each \emph{trial} initiates QTW once and measures the position at $t$ randomly chosen from $[0,\tau]$. For SGD, an experiment begins at a designated minimum and stops until SGD hits the target region. The evolution time of an experiment is the sum of evolution time of the trials the experiment contains. We use $T^{\rm QTW}_{\rm hit}$ and $T^{\rm SGD}_{\rm hit}$ to denote the evolution time of one experiment for QTW and SGD, respectively. In \fig{histogramsExp123}, the histograms compare $T^{\rm QTW}_{\rm hit}$ with $T^{\rm SGD}_{\rm hit}/10$, and all presented examples demonstrate that QTW is faster. The number of quantum queries is approximately $\tilde{O}(\|f\|_{\infty} T^{\rm QTW}_{\rm hit})$ and the number of classical queries is $\Omega(T^{\rm SGD}_{\rm hit}/s)$. In addition, in the three examples $\|f\|_{\infty} \leq 0.85$ and $s<0.2$, and quantum advantage exists in terms of query complexity. This result matches our theory at large. For Example 1, we make direct comparison between the exponential terms $e^{S_0/h}$ and $e^{2H_f/s}$, and to remove the coefficients in front of them, we divide $T^{\rm SGD}_{\rm hit}$ by $10$ such that $T^{\rm SGD}_{\rm hit}/10$ has similar distribution t $T^{\rm QTW}_{\rm hit}$ for Example 1. In this way, we observe that whether $T^{\rm SGD}_{\rm hit}/10$ is relatively larger than $T^{\rm QTW}_{\rm hit}$ is determined only by $e^{S_0/h}$ and $e^{2H_f/s}$. For Example 2, $T^{\rm QTW}_{\rm hit}$ is not much smaller than $T^{\rm SGD}_{\rm hit}/10$, which is not completely coherent with our theory. This result can be explained as that for Example 2, the quantum learning rate $h$ is not small enough such that the initial state prepared does not well stay near a low energy subspace. Experiments on Example 2 suggest that higher energy may not be able to help quantum tunneling to run faster. For Example 3, the $h$ chosen is small enough (see details in \append{E.1}) such that a significant quantum speedup is achieved as expected. In \fig{histogramsExp123}, $T^{\rm SGD}_{\rm hit}/10$ is even several orders of magnitude larger than $T^{\rm QTW}_{\rm hit}$. \begin{figure} \centering \includegraphics[width=\linewidth]{histogramsExp123.pdf} \caption{Quantum-classical comparison between SGD and QTW on three landscapes. Example 1 is the critical case where the exponential terms in QTW and SGD evolution time are equal for sufficiently small accuracy $\delta$. Example 2 has flatter minima but similar barriers compared to Example 1, enabling QTW to be faster. Example 3 possesses the same flatness of minima as Example 1 but is equipped with sharp but thin barriers, enabling larger quantum speedups. We take $\tau=288, 800, 600$ in the three examples, respectively.} \label{fig:histogramsExp123} \end{figure} \paragraph{Dimension dependence.} Due to the limitations of solving the Schr\"odinger equation on classical computers, QTW is simulated only in low dimensions (i.e., $d=1$ and $d =2$). Here we examine \thm{provableinformal} by testing SGD and its the classical lower bound. The classical lower bound in \thm{provableinformal} ensures that for any $s$, SGD cannot cannot escape from $S_{\mathbf{v}}$ with subexponential queries with high probability. Based on the constructed landscape with parameters specified in \append{E.2}, we test SGD with different learning rates ($s\in [0.1,1]$) in various dimensions ($d \in [15,95]$). For each dimension and each $s$, 1000 experiments are conducted. The number of steps spent to escaping from $S_{\mathbf{v}}$ in one experiment is denoted as $Q_{\rm esc}$. Here, we present the relationship between average $Q_{\rm esc}$ and the dimension $d$ in \fig{dimdepenmean} (more details are deferred to \append{E.2}). For each fixed learning rate $s$, we observe that with the increase of $d$, the average $Q_{\rm esc}$ remains constant initially and then increase exponentially with respect to $d$. Increasing $s$ yields a smaller initial constant but larger exponential rate. Nevertheless, for all $s$, $Q_{\rm esc}$ eventually increases exponentially with respect to $d$ (the triangle in \fig{dimdepenmean} shows the slope $1/256$ corresponding to the exponential function $e^{d/256}$ which is a lower bound of the average $Q_{\rm esc}$), supporting our prediction. \begin{figure} \centering \begin{minipage}[t]{0.46\linewidth} \centering \includegraphics[width = \linewidth]{dimdepenmean.pdf} \caption{Relationship between the average $Q_{\rm esc}$ and the dimension $d$ under different $s$. The dashed line captures the lower bound of the average $Q_{\rm esc}$.} \label{fig:dimdepenmean} \end{minipage} \hspace{6mm} \begin{minipage}[t]{0.46\linewidth} \centering \includegraphics[width =\linewidth]{qlrdepend.pdf} \caption{$T_{\rm half}$ with respect to $h$ on a fixed landscape: theoretical prediction (red solid line) and time in experiments (blue circles).} \label{fig:qlrdepend} \end{minipage} \end{figure} \paragraph{Dependence on the quantum learning rate.} In QTW, the quantum learning rate $h$ is one of the most important variables. \thm{informalQTW} gives a general relationship between $h$ and the evolution time of QTW. We further test the relationship on the landscape constructed in \thm{provableinformal} (dimension $d=2$) with specified parameters given in \append{E.3}. Since the landscape has two symmetric wells, the time for tunneling from one well to the other, $T_{\rm half}$, is explicitly linked to $\Delta E$ (i.e., $T_{\rm half} = \pi/\Delta E$). On this concrete landscape, $\Delta E$ can be predicted, giving that \begin{align} \ln T_{\rm half} = \frac{S_0}{h} - \frac{1}{2}\ln h + C_f, \end{align} where $C_f$ is a constant depending on $f$ and can be explicitly calculated. Starting from one well, we stop when the probability of finding the other well exceeds 90\% and record the evolution time as $T_{\rm half}$. Experiments on $T_{\rm half}$ is shown in \fig{qlrdepend}. The results match our theory except a constant difference between the predicted and experimental $\ln T_{\rm half}$, indicating the correctness of $\frac{S_0}{h} - \frac{1}{2}\ln h$. The constant difference emerges because we stop evolution when the probability of tunneling exceeds 90\%, while theoretical $T_{\rm half}$ takes the time when the probability is nearly 100\%. To conclude, several aspects of the present theory are well supported by numerical experiments. \section{Discussion}\label{sec:discussion} In this paper, we explore quantum speedups for nonconvex optimization by quantum tunneling. In particular, we introduce the quantum tunneling walk (QTW) and apply it to nonconvex problems where local minima are approximately global minima. We show that QTW achieves quantum speedup over classical stochastic gradient descents (SGD) when the barriers between different local minima are high but thin and the minima are flat. Moreover, we construct a specific nonconvex landscape where QTW given proper initial states is exponentially faster than classical algorithms taking local queries for hitting the neighborhood of a target global minimum. Finally, we conduct numerical experiments to corroborate our theoretical results. We expect our results to have further impacts both classically and quantumly. In optimization theory, previous work has studied several physics-motivated optimization algorithms, including Nesterov's momentum method~\cite{su2016differential,wibisono2016variational,shi2021understanding}, stochastic gradient descents~\cite{SSJ20}, symplectic optimization~\cite{betancourt2018symplectic,jordan2018dynamical}, etc. We believe that our work can further inspire the design of optimization algorithms. From theory to practice, in this work we analyzed the performance of QTW on tensor decomposition, and we expect QTW to also have decent performance on other practical problems with benign landscapes. In quantum computing, on the one hand, previous work on continuous optimization only studies convex optimization~\cite{vanApeldoorn2020optimization,chakrabarti2020optimization} or local properties such as escaping from saddle points~\cite{zhang2021quantum}, and our work significantly extends the range of problems which quantum computers can efficiently solve to global problems in nonconvex optimization. On the other hand, we point out that QTW has the potential to be implemented on near-term quantum computers. In fact, current quantum computers have implemented both quantum simulation~\cite{arute2020hartree,ebadi2021quantum} and quantum walks~\cite{tang2018experimental,gong2021quantum} to decent scales. We deem QTW as a potential proposal for demonstrating quantum advantages in near term. Our paper also leaves several technical questions for future investigation: \begin{itemize} \item What is the performance of QTW on more general landscapes? For instance, a wide range of deep neural networks~\cite{KHK19} has some (but probably not all) local minima which are approximately global. Future work on weakening the assumptions on landscapes for QTW is preferred. \item Are there more examples with exponential quantum-classical separation? Our construction leverages a special kind of locally non-informative landscapes, and exponential quantum-classical separation can potentially be observed on other landscapes, such as nonsmooth landscapes and landscapes with negative curvature. \item QTW simulates the Schr\"odinger equation whose potential is set to be the optimization function, and this QTW can be efficiently simulated on quantum computers. In general, are there better PDEs which are more efficient for optimization and can still be efficiently simulated on quantum computers? \end{itemize} \section*{Acknowledgements} YL thanks Frédéric Hérau and Michael Hitrik for helping with understanding spectral theory, Rong Ge and Chenyi Zhang for inspiring discussions on tensor decomposition, and Jiaqi Leng for the guidance on numerical methods for quantum simulation. TL thanks Eric R. Anschuetz for general suggestions on a preliminary version of this paper, and Andrew M. Childs for helpful discussions that inspire \sec{no-exp-sep-init}. YL was funded by SURF from Tsien Excellence in Engineering Program, Tsinghua University. WJS was funded by an Alfred Sloan Research Fellowship and the Wharton Dean’s Research Fund. TL was funded by a startup fund from Peking University, and the Advanced Institute of Information Technology, Peking University. \bibliographystyle{alphaUrlePrint}

ac26ff1ecd29b5de7db2b96d8350d8ad183d89ac

\subsection{Adversarial Graph Signal Denoising Problem} Note that the second term in the GSD problem (Eq.~(\ref{graph signal denoising})) which controls the smoothness of the feature matrix over graphs, is related to both the graph Laplacian and the node features. Therefore, the slight changes in the graph Laplacian matrix could lead to an unstable denoising effect. Inspired by the recent studies in adversarial training~\citep{madry2018towards}, we formulate the adversarial graph signal denoising problem as a min-max optimization problem: \begin{equation} \label{adv gsd} \min_{\mathbf{F}} \left[\left\|\mathbf{F}-\mathbf{X}\right\|_{F}^{2}+\lambda \cdot \max_{\mathbf{L}^{\prime}} \ \operatorname{tr}\left(\mathbf{F}^{\top} \mathbf{L}^{\prime} \mathbf{F}\right)\right] \quad \operatorname{s.t.} \quad \left\|\mathbf{L}^{\prime}-\widetilde{\mathbf{L}}\right\|_{F} \leq \varepsilon. \end{equation} Intuitively, the inner maximization on the Laplacian $\mathbf{L}'$ generates perturbations on the graph structure\footnote{Here we do not need exact graph structure perturbations as in graph adversarial attacks~\citep{zugner2018adversarial,zugner2019adversarial} but a virtual perturbation that could lead to small changes in the Laplacian.}, and enlarges the distance between the node representations of connected neighbors. Such maximization finds the worst case perturbations on the graph Laplacian that hinders the global smoothness of $\mathbf{F}$ over the graph. Therefore, by training on those worse case Laplacian perturbations, one could obtain a robust graph signal denoising solution. Ideally, through solving Eq.~(\ref{adv gsd}), the smoothness of the node representations as well as the implicit denoising effect can be enhanced. \subsection{Minimization of the Optimization Problem} \label{Minimization of the Optimization Problem} The min-max formulation in Eq.~(\ref{adv gsd}) also makes the adversarial graph signal denoising problem much harder to solve. Fortunately, unlike adversarial training~\citep{madry2017towards} where we need to first adopt PGD to solve the inner maximization problem before we solve the outer minimization problem, here inner maximization problem is simple and has a closed form solution. In other words, we do not need to add random perturbations on the graph structure at each training epoch and can find the largest perturbation which maximizes the inner adversarial loss function. Denote the perturbation as $\bm{\delta}$, and $\mathbf{L}'=\widetilde{\mathbf{L}} + \bm{\delta}$. Directly solving\footnote{More details on how to solve the inner maximization problem can be found in Appendix~\ref{appendix:how to solve the optimization problem}.} the inner maximization problem, we get $\bm{\delta}=\varepsilon\nabla h(\bm{\delta})=\frac{\varepsilon\mathbf{F}\mathbf{F}^{\top}}{\left\|\mathbf{F}\mathbf{F}^{\top}\right\|_{F}}$. Plugging this solution into Eq.~(\ref{adv gsd}), we can rewrite the outer optimization problem as follows: \begin{equation} \rho(\mathbf{F})=\min_{\mathbf{F}} \left[\left\|\mathbf{F}-\mathbf{X}\right\|_{F}^{2}+\lambda \max \operatorname{tr}\left(\mathbf{F}^{\top} \widetilde{\mathbf{L}} \mathbf{F}\right)+\lambda\varepsilon\operatorname{tr}\frac{\mathbf{F}^{\top}\mathbf{F}\mathbf{F}^{\top}\mathbf{F}}{\left\|\mathbf{F}\mathbf{F}^{\top}\right\|_{F}}\right]. \end{equation} Taking the gradient of $\rho(\mathbf{F})$ to zero, we get the solution of the outer optimization problem as follows: \begin{equation}\label{eq:advF} \mathbf{F} = \left(\mathbf{I}+\lambda\widetilde{\mathbf{L}}+\lambda\varepsilon\frac{\mathbf{F}\mathbf{F}^{\top}}{\left\|\mathbf{F}\mathbf{F}^{\top}\right\|_{F}}\right)^{-1}\mathbf{X}. \end{equation} Both sides of Eq.~(\ref{eq:advF}) contains $\mathbf{F}$, directly computing the solution is difficult. Note that in Eq.~(\ref{adv gsd}) we also require $\mathbf{F}$ to be close to $\mathbf{X}$, we can approximate Eq.~(\ref{eq:advF}) by replacing the $\mathbf{F}$ with $\mathbf{X}$ in the inverse matrix on the right hand side. With the Neumann series expansion of the inverse matrix, we get the final approximate solution as \begin{equation} \label{RNGC filter} \mathbf{H} \approx \frac{1}{\lambda+1}\sum_{s=0}^{S}\left[\frac{\lambda}{\lambda+1}\left(\widetilde{\bm{\mathcal{A}}}-\frac{\varepsilon\mathbf{X}\mathbf{X}^{\top}}{\left\|\mathbf{X}\mathbf{X}^{\top}\right\|_{F}}\right)\right]^{s}\mathbf{X}\mathbf{W}. \end{equation} The difference between Eq.~(\ref{RNGC filter}) and Eq.~(\ref{NGC filter}) is that there is one more term in Eq.~(\ref{RNGC filter}) derived from solving the inner optimization problem of Eq.~(\ref{adv gsd}). Based on this, we proposed our robust Neumann graph convolution (RNGC). \paragraph{Scalability.}{Although RNGC introduces extra computational burdens for large graphs due to the $\mathbf{X} \mathbf{X}^{\top}$ term, if the feature matrix is sparse, the extra computational effort is minimal as the $\mathbf{X} \mathbf{X}^{\top}$ term can also be sparse. For the scalability of RNGC on large graphs with dense feature matrix, we only compute the inner product of feature vectors ($\mathbf{X}_i, \mathbf{X}_{j|j\in\mathcal{N}_{i}}$) between adjacent neighbors like masked attention in GAT. Compared with \name, the additional computation cost is $\mathcal{O}(|\mathcal{E}|)$.} \subsection{Denoising Effectiveness Comparison of Various GNN Models} In this section, we compare the denoising effectiveness of different GNN models through their test accuracy by training on the noisy feature matrix with Gaussian noise. \paragraph{Datasets.} In our experiments, we utilize three public citation network datasets Cora, Citeseer, and Pubmed~\citep{sen2008collective} which are homophily graphs for semi-supervised node classification. For the semi-supervised learning experimental setup, we follow the standard fixed splits employed in~\citep{yang2016revisiting}, with 20 nodes per class for training, 500 nodes for validation, and 1,000 nodes for testing. We also use four datasets: Cornell, Texas, Wisconsin, and Actor which are heterophily graphs for full-supervised node classification. For each dataset, we randomly split nodes into 60\%, 20\%, and 20\% for training, validation, and testing as suggested in \citep{pei2020geom}. Moreover, we utilize three large-scale graph datasets: Coauthor-CS, Coauthor-Phy~\citep{shchur2018pitfalls}, and ogbn-products~\citep{hu2020open} for evaluation. For Coauthor datasets, we split nodes into 60\%, 20\%, and 20\% for training, validation, and testing. For ogbn-products dataset, we follow the dataset split in OGB~\citep{hu2020open}. \paragraph{Baselines.} For the baselines, we consider graph neural networks derived from graph signal denoising, including GLP~\citep{li2019label}, S$^2$GC~\citep{zhu2021simple}, and IRLS~\citep{yang2021graph}; popular GNN architectures, such as GCN~\citep{kipf2017semi} and GAT~\citep{velivckovic2018graph}; and MLP which has no aggregation operation. \paragraph{Experimental Setup and Implementations.} We assume that the original feature matrix is clean and do not have noise and we synthesize the noise from the standard Gaussian distribution and add them on the original feature matrix. By default, we apply row normalization for data after adding the Gaussian noise\footnote{We also perform an analysis on the effect of row normalization in noisy feature matrix in Appendix~\ref{appendix:row norm}.}, and train all the models based on these noisy feature matrix. For the hyper-parameters of each model, we follow the setting that reported in their original papers. To eliminate the effect of randomness, we repeat such experiment for 100 or 10 times and report the mean accuracy. Note that in each repeated run, we add different Gaussian noises. While for the same run, we apply the same noisy feature matrix for training all the models. For our \name and R\name model, the hyper-parameter details can be found in Appendix~\ref{appendix:hyperparameter}. \input{heterophily_results} \input{large-scale_results} \paragraph{Results on Supervised Node Classification.} Figure~\ref{fig:noise} illustrates the comparison of classification accuracy against the various noise levels for semi-supervised node classification tasks. The noise level $\xi$ controls the magnitude of the Gaussian noise we add to the feature matrix: $\mathbf{X}+\xi\bm{\eta}$ where $\bm{\eta}$ is sampled from standard i.i.d., Gaussian distribution. For Cora and Citeseer, we test $\xi \in \{0.1, 0.2, 0.3, 0.4, 0.5\}$ and for Pubmed, we test $\xi \in \{0.01, 0.02, 0.03, 0.04, 0.05\}$. From Figure~\ref{fig:noise}, we can observe that the test accuracy of MLP is close to randomly guessing (RG) when the noise level is relatively large. This implies the weak denoising effect of MLP models. For shallow GNN models, such as GCN and GAT (which usually contain 2 layers), their denoising performance is limited especially on Pubmed since they do not aggregate information (features and noise) from higher-order neighbors. For models with deep layers{\footnote{We also perform an analysis on the denoising effect of depth in \name and R\name in Appendix~\ref{appendix:depth analysis}.}}, such as IRLS ($\geq 8$ layers), the denoising performance is much better compared to shallow models. Lastly, our \name and R\name model with 16 layers ($S=16$) achieve significantly better denoising performance compared with other baseline methods, which backup our theoretical analyses. In most cases, \name and R\name achieve very similar denoising performance but in general, R\name still slightly outperforms \name, suggesting that we indeed gain more benefits by solving the adversarial graph denoising problem. {Table~\ref{tab:heterophily} reports the comparison of classification accuracy against the various noise levels for full-supervised node classification tasks on heterophily graphs. The first- and second-highest accuracies are highlighted in bold. For these datasets, we test $\xi \in \{0.01, 1\}$. From Table~\ref{tab:heterophily}, we can observe that MLP is better than most GNN models in most cases due to the heterophily properties of these graphs. However, our proposed R\name achieves significantly better or matches denoising performance compared with other baseline methods, which demonstrates the superiority of our R\name.} {For ogbn-products, we only choose MLP, GCN, and S$^2$GC as baselines, since the results are sensitive concerning model size and various tricks from the OGB leaderboard. For fair comparison, the size of parameters for these baselines and R\name is the same. We also use full-batch training for the baselines and our model. Table~\ref{tab:coauthor} and \ref{tab:ogb} report the comparison of classification accuracy against the various noise levels for full-supervised node classification tasks on large-scale graphs. The first- and second-highest accuracies are highlighted in bold. For these datasets, we test $\xi \in \{0.1, 1\}$. Compared with the above small datasets, the node degree on these three datasets is larger, which means they have better connectivity. From Table~\ref{tab:coauthor} and \ref{tab:ogb}, we can observe that the test accuracy of MLP is far lower than GCN and R\name. This implies the weak denoising effect of MLP. The test accuracy of GCN is slightly smaller than R\name on these datasets since they are well-connected and have a large graph size and we can achieve a good denoising performance with shallow-layer GNN models. For the scalability of R\name on large graphs such as ogbn-products, we use the acceleration method mentioned in Sec.~\ref{Minimization of the Optimization Problem}.} \input{flip_defense_results} \subsection{Denoising Performance on Feature Flipping Perturbation} In this section, we compare the denoising effectiveness of different models through their test accuracy by training on the noisy feature matrix which is perturbated through flipping the individual feature with a small Bernoulli probability on three citation datasets. \paragraph{Setting and Results.} We flip the individual feature on three citation datasets: Cora, Citeseer, and Pubmed as the noise. And we compare the denoising performance of R\name with MLP and GCN. From Table~\ref{tab:flip}, we can observe that the denoising performance of R\name is much better than baselines when the flip probability is 0.4. In fact, the added perturbations by flipping the individual feature approximately follow a Bernoulli distribution, which is also a Sub-Gaussian distribution. The results verify our theoretical analysis further. \subsection{Defense Performance of R\name against Graph Structure Attack} Although we do not perform actual graph structure perturbations as in graph adversarial attacks~\citep{zugner2018adversarial,zugner2019adversarial} but a virtual perturbation in the Laplacian. Therefore, it's not clear how much perturbations on the Laplacian correspond to the actual perturbations on graph structure. Nevertheless, we still conduct the experiments of R\name against graph structure meta-attack where the ptb rate is 25\%. As shown in the Table~\ref{tab:attack}, our R\name model outperforms than GCN, GAT, RobustGCN \citep{zugner2019robustgcn}, GCN-Jaccard \citep{wu2019adversarial}, GCN-SVD \citep{entezari2020all}, and S$^2$GC on Cora, Citeseer, and Pubmed. \section{The details on how to solve the inner maximization problem in Sec.~\ref{Minimization of the Optimization Problem}} \label{appendix:how to solve the optimization problem} Different from the non-concave inner maximization problem in the adversarial attack, our inner maximization problem is indeed a convex optimization problem. Hence, we do not need to add random perturbations on the graph structure at each training epoch and can find the largest perturbation which maximizes the inner adversarial loss function. Denote the perturbation as $\bm{\delta}$, and $\mathbf{L}'=\widetilde{\mathbf{L}} + \bm{\delta}$. We can rewrite the inner maximization problem as \begin{equation} \max_{\mathbf{L}^{\prime}} \operatorname{tr}\left(\mathbf{F}^{\top} \mathbf{L}^{\prime} \mathbf{F}\right) =\langle\widetilde{\mathbf{L}}, \mathbf{F}^{\top}\mathbf{F}\rangle + \max_{\bm{\delta}}\langle\bm{\delta}, \mathbf{F}^{\top}\mathbf{F}\rangle \quad \operatorname{s.t.} \quad \left\|\bm{\delta}\right\|_{F} \leq \varepsilon. \end{equation} We denote $h(\bm{\delta})=\langle\bm{\delta}, \mathbf{F}^{\top}\mathbf{F}\rangle$. Obviously, $h(\bm{\delta})$ reaches the largest value when $\bm{\delta}$ has the same direction with the gradient of $h(\bm{\delta})$, e.g. $\bm{\delta}=\varepsilon\nabla h(\bm{\delta})=\frac{\varepsilon\mathbf{F}\mathbf{F}^{\top}}{\left\|\mathbf{F}\mathbf{F}^{\top}\right\|_{F}}$, which is illustrated in Fig.~\ref{figure:adv inner problem}. \begin{figure}[htbp] \begin{center} \includegraphics[width=.45\columnwidth]{images/adv_inner_problem.pdf} \end{center} \caption{The illustration of the inner maximization problem. The adversarial loss function reaches the largest value when the direction of $\bm{\delta}$ is the same with $\nabla h(\bm{\delta})$} \label{figure:adv inner problem} \end{figure} \section{Additional Details on the Neumann Series} \label{appendix:neumann series} We provide additional details and derivations on how to obtain the Neumann Series which leads to our Neumann Graph Convolution (NGC) method. Before we derive the Neumann Series, we first introduce the following lemmas which are crucial to the derivation of the Neumann Series. \begin{lemma} \rm {\textbf{(Gelfand formula)~\citep{bhatia2013matrix}} } \label{Gelfand formula} \emph{Given any matrix norm $\||\cdot|\|$, then $\rho(\mathbf{A})=\lim\limits_{k \rightarrow \infty}\||\mathbf{A}^{k}|\|^{1 / k}=\inf\limits_{k \geq 1}\||\mathbf{A}^k|\|^{1/k}\leq \||\mathbf{A}|\|$}. \end{lemma} Lemma~\ref{Gelfand formula} describes the relationship between the spectral radius of a matrix and its matrix norm, $i.e.$ $\rho(\mathbf{A})=\lim\limits_{k \rightarrow \infty}\||\mathbf{A}^{k}|\|^{1 / k}$. \begin{lemma} \label{convergence of neumann series} Let $\mathbf{A} \in \mathbb{C}^{n \times n}$, the spectral radius $\rho(\mathbf{A})=\max (\operatorname{abs}(\operatorname{spec}(\mathbf{A})))$, if $\rho(\mathbf{A})<1$, then $\sum_{k=0}^{\infty} \mathbf{A}^{k}$ converges to $(\mathbf{I}-\mathbf{A})^{-1}$. \end{lemma} \begin{proof} We first prove that $(\mathbf{I}-\mathbf{A})^{-1}$ exists as follows: Based on the definition of eigenvalues of $\mathbf{A}$, we have $|\lambda \mathbf{I} - \mathbf{A}| = 0$ and the solution is the eigenvalue of $\mathbf{A}$. Since $\rho(\mathbf{A}) < 1$, if $\lambda \geq 1$, then $|\lambda \mathbf{I} - \mathbf{A}| \neq 0$, so $|\mathbf{I} - \mathbf{A}| \neq 0$, which means $(\mathbf{I}-\mathbf{A})^{-1}$ exists. Since $\rho(\mathbf{A}) < 1$ and by Lemma~\ref{Gelfand formula}, we have $\lim\limits_{k \rightarrow \infty}\||\mathbf{A}^{k}|\|=\rho(\mathbf{A})^k=0$. Let $\mathbf{S}_k$ = $\mathbf{A}^0 + \mathbf{A}^1 + \cdots + \mathbf{A}^k$, then we have \begin{equation*} \begin{split} \lim_{k \rightarrow \infty}(\mathbf{S}^k-\mathbf{A}\mathbf{S}^k) &= \lim\limits_{k \rightarrow \infty}(\mathbf{I}-\mathbf{A})\mathbf{S}^k \\ &=\lim_{k \rightarrow \infty}(\mathbf{I}-\mathbf{A}^{k+1}) \\ &= \mathbf{I} \end{split} \end{equation*} Since $(\mathbf{I}-\mathbf{A})^{-1}$ exists, so we have $(\mathbf{I}-\mathbf{A})\lim\limits_{k \rightarrow \infty}\mathbf{S}^k=\mathbf{I}$, and $\lim\limits_{k \rightarrow \infty}\mathbf{S}^k=(\mathbf{I}-\mathbf{A})^{-1}$, which finishes the proof. \end{proof} Lemma~\ref{convergence of neumann series} describes the convergence of Neumann Series and the condition to get the convergence. \begin{lemma} \rm {\textbf{(Gerschgorin Disc)~\citep{bhatia2013matrix}}} \label{Gerschgorin Disc} \emph{Let $\mathbf{A} \in \mathbb{C}^{n \times n}$, with entries $a_{ij}$. For any eigenvalue $\lambda$, there exits $i$ and the corresponding Gerschgorin disc $D\left(a_{i i}, R_{i}\right) \subseteq \mathbb{C}$ such that $\lambda$ lies in this disc, i.e.} \begin{equation*} |\lambda-a_{ii}| \leq \sum_{j\neq i}^{n}|a_{ij}|. \end{equation*} \end{lemma} Lemma~\ref{Gerschgorin Disc} describes the estimated range of eigenvalues. Now we start to derive the Neumann Series expansion of the solution of GSD as follows. \begin{lemma} \label{Neumann derivation} Let $\mathbf{A} \in\{0,1\}^{n \times n}$ be the adjacency matrix of a graph and $\widetilde{\bm{\mathcal{A}}}=\widetilde{\mathbf{D}}^{-\frac{1}{2}} \widetilde{\mathbf{A}} \widetilde{\mathbf{D}}^{-\frac{1}{2}}$ or $\widetilde{\bm{\mathcal{A}}} = \widetilde{\mathbf{D}}^{-1} \widetilde{\mathbf{A}}$, then \begin{equation*} (\mathbf{I}-\frac{\lambda}{\lambda+1}\widetilde{\bm{\mathcal{A}}})^{-1}=\sum_{k=0}^{\infty} \left(\frac{\lambda}{\lambda+1}\widetilde{\bm{\mathcal{A}}}\right)^{k}. \end{equation*} \end{lemma} \begin{proof} We first prove that $\rho(\widetilde{\bm{\mathcal{A}}})\leq1$ where $\widetilde{\bm{\mathcal{A}}}=\widetilde{\mathbf{D}}^{-\frac{1}{2}} \widetilde{\mathbf{A}} \widetilde{\mathbf{D}}^{-\frac{1}{2}}$. Let $\lambda$ be the eigenvalue of $\widetilde{\bm{\mathcal{A}}}$, and $\mathbf{v}$ be the corresponding eigenvector. Then we have \begin{equation*} \begin{split} \left(\widetilde{\mathbf{D}}^{-\frac{1}{2}} \widetilde{\mathbf{A}} \widetilde{\mathbf{D}}^{-\frac{1}{2}}\right)\mathbf{v}=\lambda\mathbf{v} &\Longrightarrow \widetilde{\mathbf{D}}^{-\frac{1}{2}} \left(\widetilde{\mathbf{D}}^{-\frac{1}{2}} \widetilde{\mathbf{A}} \widetilde{\mathbf{D}}^{-\frac{1}{2}}\right)\mathbf{v}=\lambda\widetilde{\mathbf{D}}^{-\frac{1}{2}}\mathbf{v}\\ &\Longrightarrow \left(\widetilde{\mathbf{D}}^{-1} \widetilde{\mathbf{A}}\right) \widetilde{\mathbf{D}}^{-\frac{1}{2}}\mathbf{v} = \lambda\widetilde{\mathbf{D}}^{-\frac{1}{2}}\mathbf{v}, \end{split} \end{equation*} which means $(\lambda, \widetilde{\mathbf{D}}^{-\frac{1}{2}}\mathbf{v})$ is the eigen-pair of $\widetilde{\mathbf{D}}^{-1}\mathbf{A}$. By Lemma~\ref{Gerschgorin Disc}, there exists $i$, such that \begin{equation*} \begin{split} &\left|\lambda-\left(\widetilde{\mathbf{D}}^{-1}\widetilde{\mathbf{A}}\right)_{ii}\right|\leq \sum_{j\neq i}\left|\left(\widetilde{\mathbf{D}}^{-1}\widetilde{\mathbf{A}}\right)_{ij}\right|\\ &\Longrightarrow \left(\widetilde{\mathbf{D}}^{-1}\widetilde{\mathbf{A}}\right)_{ii} - \sum_{j\neq i}\left|\left(\widetilde{\mathbf{D}}^{-1}\widetilde{\mathbf{A}}\right)_{ij}\right| \leq \lambda \leq \left(\widetilde{\mathbf{D}}^{-1}\widetilde{\mathbf{A}}\right)_{ii} + \sum_{j\neq i}\left|\left(\widetilde{\mathbf{D}}^{-1}\widetilde{\mathbf{A}}\right)_{ij}\right|. \end{split} \end{equation*} Since $\left(\widetilde{\mathbf{D}}^{-1}\widetilde{\mathbf{A}}\right)_{ij}>0$ and $\sum_{j}\left|\left(\widetilde{\mathbf{D}}^{-1}\widetilde{\mathbf{A}}\right)_{ij}\right|=\sum_{j}\left(\widetilde{\mathbf{D}}^{-1}\widetilde{\mathbf{A}}\right)_{ij}=1$, obviously \begin{equation*} -1<\left(\widetilde{\mathbf{D}}^{-1}\widetilde{\mathbf{A}}\right)_{ii} - \sum_{j\neq i}\left|\left(\widetilde{\mathbf{D}}^{-1}\widetilde{\mathbf{A}}\right)_{ij}\right| \leq \lambda \leq \left(\widetilde{\mathbf{D}}^{-1}\widetilde{\mathbf{A}}\right)_{ii} + \sum_{j\neq i}\left|\left(\widetilde{\mathbf{D}}^{-1}\widetilde{\mathbf{A}}\right)_{ij}\right|=1. \end{equation*} So if $\widetilde{\bm{\mathcal{A}}}=\widetilde{\mathbf{D}}^{-\frac{1}{2}} \widetilde{\mathbf{A}} \widetilde{\mathbf{D}}^{-\frac{1}{2}}$, we have $\rho(\widetilde{\bm{\mathcal{A}}})\leq1$. When $\widetilde{\bm{\mathcal{A}}}=\widetilde{\mathbf{D}}^{-1}\widetilde{\mathbf{A}}$, we denote $(\lambda, \mathbf{v})$ as the eigen-pair of $\mathbf{\widetilde{\mathbf{D}}^{-1}\mathbf{A}}$. Similarly, by Lemma~\ref{Gerschgorin Disc}, there exists $i$, such that \begin{equation*} \begin{split} &\left|\lambda-\left(\widetilde{\mathbf{D}}^{-1}\widetilde{\mathbf{A}}\right)_{ii}\right|\leq \sum_{j\neq i}\left|\left(\widetilde{\mathbf{D}}^{-1}\widetilde{\mathbf{A}}\right)_{ij}\right|\\ &\Longrightarrow \left(\widetilde{\mathbf{D}}^{-1}\widetilde{\mathbf{A}}\right)_{ii} - \sum_{j\neq i}\left|\left(\widetilde{\mathbf{D}}^{-1}\widetilde{\mathbf{A}}\right)_{ij}\right| \leq \lambda \leq \left(\widetilde{\mathbf{D}}^{-1}\widetilde{\mathbf{A}}\right)_{ii} + \sum_{j\neq i}\left|\left(\widetilde{\mathbf{D}}^{-1}\widetilde{\mathbf{A}}\right)_{ij}\right|. \end{split} \end{equation*} Obviously, we can get the same conclusion for $\widetilde{\bm{\mathcal{A}}}=\widetilde{\mathbf{D}}^{-1}\widetilde{\mathbf{A}}$. So it is true for $\rho\left(\frac{\lambda}{\lambda+1}\widetilde{\bm{\mathcal{A}}}\right) \leq \frac{\lambda}{\lambda+1}<1$ By Lemma~\ref{convergence of neumann series}, we get the result $(\mathbf{I}-\frac{\lambda}{\lambda+1}\widetilde{\bm{\mathcal{A}}})^{-1}=\sum_{k=0}^{\infty} \left(\frac{\lambda}{\lambda+1}\widetilde{\bm{\mathcal{A}}}\right)^{k}$, which finishes the proof. \end{proof} By Lemma~\ref{Neumann derivation}, we approximate the inverse matrix $(\mathbf{I}+\lambda \tilde{\mathbf{L}})^{-1}$ up to $S$-th order with \begin{equation*} \left(\mathbf{I}+\lambda \widetilde{\mathbf{L}}\right)^{-1} =\frac{1}{\lambda+1}\left(\mathbf{I}-\frac{\lambda}{\lambda+1}\widetilde{\bm{\mathcal{A}}}\right)^{-1}\approx\frac{1}{\lambda+1}\sum_{s=0}^{S} \left(\frac{\lambda}{\lambda+1}\widetilde{\bm{\mathcal{A}}}\right)^{s}. \end{equation*} \section{The Row Summation of the Neumann Series} \label{appendix:row sum} We provide the derivations of the row sum of $\widetilde{\bm{\mathcal{A}}}_{S}$ in this section. Before we derive the row summation of $\widetilde{\bm{\mathcal{A}}}_{S}$, we first derive the row summation of $\widetilde{\bm{\mathcal{A}}}^{k}$. \begin{lemma} \label{transition matrix} Consider a probability matrix $\mathbf{P} \in \mathbb{R}^{n \times n}$, where $\mathbf{P}_{ij}\geq0$. Besides, for all $i$, we have $\sum_{j=1}^n\mathbf{P}_{ij}=1$. Then for any $s \in \mathbb{Z}_{+}$, we have $\sum_{j=1}^n\mathbf{P}^s_{ij}=1$, \end{lemma} \begin{proof} We give a proof by induction on $k$.\\ \textbf{Base case:} When $k=1$, the case is true.\\ \textbf{Inductive step:} Assume the induction hypothesis that for a particular $k$, the single case n = k holds, meaning $\mathbf{P}^k$ is true: \begin{equation*} \forall i, \sum_{j=1}^n \mathbf{P}_{ij}^k =1. \end{equation*} As $\mathbf{P}^{k+1}=\mathbf{P}^{k}\mathbf{P}$, so we have \begin{equation*} \sum_{j=1}^n\mathbf{P}^{k+1}_{ij}=\sum_{j=1}^n\sum_{k=1}^n\mathbf{P}^{k}_{ik}\mathbf{P}_{kj} = \sum_{k=1}^n\sum_{j=1}^n\mathbf{P}^{k}_{ik}\mathbf{P}_{kj} = \sum_{k=1}^n\mathbf{P}^{k}_{ik}\left(\sum_{j=1}^n\mathbf{P}_{kj}\right) = \sum_{k=1}^n\mathbf{P}^{k}_{ik} = 1, \end{equation*} which finishes the proof. \end{proof} Lemma~\ref{transition matrix} describes the row summation of $\widetilde{\bm{\mathcal{A}}}^{k}$ is 1. Now we can obtain the row summation for $\widetilde{\bm{\mathcal{A}}}_{S}$. Then for any $i$, we have \begin{equation} \begin{split} \sum_{j=1}^{n}\left[\widetilde{\bm{\mathcal{A}}}_{S}\right]_{ij} &=\frac{1}{\lambda+1}\sum_{s=0}^{S} \left(\frac{\lambda}{\lambda+1}\left[\widetilde{\bm{\mathcal{A}}}\right]_{ij}\right)^{s}\\ &=\frac{1}{\lambda+1}\sum_{s=0}^{S}\left(\frac{\lambda}{\lambda+1}\right)^{s} \\ &=1-\left(\frac{\lambda}{\lambda+1}\right)^{S+1}. \end{split} \end{equation} \section{Proof of Lemma 1} \label{appendix:upper bound with hoeffding} We provide the details of proof of Lemma 1. We first introduce the General Hoeffding Inequality~\citep{hoeffding1994probability}, which is essential for bounding $\left\|\widetilde{\bm{\mathcal{A}}}_S\bm{\eta}\right\|_{F}^{2}$. \begin{lemma} \label{hoeffding} \rm{\textbf{(General Hoeffding Inequality~\citep{hoeffding1994probability})}} \emph{Suppose that the variables $X_{1}, \cdots, X_{n}$ are independent, and $X_i$ has mean $\mu_{i}$ and sub-Gaussian parameter $\sigma_{i}$. Then for all $t\geq0$, we have} \begin{equation} \mathbb{P}\left[\sum_{i=1}^{n}\left(X_{i}-\mu_{i}\right) \geq t\right] \leq \exp \left\{-\frac{t^{2}}{2 \sum_{i=1}^{n} \sigma_{i}^{2}}\right\}. \end{equation} \end{lemma} Now let's prove Lemma~\ref{upper bound of aggregated noised matirx}. \begin{proof}[Proof of Lemma~\ref{upper bound of aggregated noised matirx}.] For any entry $\left[\widetilde{\bm{\mathcal{A}}}_S\bm{\eta}\right]_{ij}=\sum_{p=1}^{n}\left(\widetilde{\bm{\mathcal{A}}}_S\right)_{ip}\bm{\eta}_{pj}$, where $\bm{\eta}_{pj}$ is a sub-Gaussian variable with parameter $\sigma^{2}$. By the General Hoeffding inequality~\ref{hoeffding}, we have \begin{equation} \mathbb{P}\left(\left|\left[ \frac{1}{\lambda+1}\sum_{s=0}^{S}\left(\frac{\lambda}{\lambda+1}\widetilde{\bm{\mathcal{A}}}_{S}\right)^{s}\bm{\eta}\right]_{ij}\right|\geq t\right) \leq 2\exp \left\{-\frac{nt^{2}}{2\tau\left(1-\left(\frac{\lambda}{\lambda+1}\right)^{S+1}\right)^2\sigma^2}\right\}. \end{equation} where $\tau = \max_i \tau_i$ and $ \tau_i = {n\sum_{j=1}^{n}\left[\widetilde{\bm{\mathcal{A}}}_{S}\right]_{ij}^2}\Bigg/{\left(1-\left(\frac{\lambda}{\lambda+1}\right)^{S+1}\right)^2}$. Applying union bound~\citep{vershynin2010introduction} to all possible pairs of $i \in [n]$, $j \in [n]$, we get \begin{equation} \mathbb{P}\left(\left\|\widetilde{\bm{\mathcal{A}}}_S\bm{\eta}\right\|_{\infty, \infty}\geq t\right) \leq \sum_{i,j}\mathbb{P}\left(\left[\widetilde{\bm{\mathcal{A}}}_S\bm{\eta}\right]_{ij}\geq t\right) \leq 2n^2\exp \left\{-\frac{nt^{2}}{2\tau\left(1-\left(\frac{\lambda}{\lambda+1}\right)^{S+1}\right)^2\sigma^2}\right\}. \end{equation} Applying union bound again, we have \begin{equation} \mathbb{P}\left(\left\|\widetilde{\bm{\mathcal{A}}}_S\bm{\eta}\right\|_{F}^{2}\geq t\right) \leq \sum_{i,j}\mathbb{P}\left(\left\|\widetilde{\bm{\mathcal{A}}}_S\bm{\eta}\right\|_{\infty, \infty}\geq \sqrt{t}\right)\leq2n^4\exp \left\{-\frac{nt}{2\tau\left(1-\left(\frac{\lambda}{\lambda+1}\right)^{S+1}\right)^2\sigma^2}\right\}. \end{equation} Choose $t=2\tau\left(1-\left(\frac{\lambda}{\lambda+1}\right)^{S+1}\right)^2\left(4\log n+\log{2d}\right)/n$ and with probability $1-1/d$, we have \begin{equation} \left\|\widetilde{\bm{\mathcal{A}}}_S\bm{\eta}\right\|_{F}^{2}\leq \frac{2\tau\left(1-\left(\frac{\lambda}{\lambda+1}\right)^{S+1}\right)^2\sigma^2\left(4\log n+\log{2d}\right)}{n}, \end{equation} which finishes the proof. \end{proof} \section{Proof of the Main Theorem~\ref{theorem:main}} \label{appendix:upper bound with optimization} We provide the details of proof of main theorem~\ref{theorem:main}.\\ \input{optimization_upper_bound} \section{More Details on Equation~(\ref{graph signal denoising}).} We provide more details on how to obtain Equation~(\ref{graph signal denoising}). Note that if we set $\widetilde{\mathbf{L}}=\mathbf{I}-\widetilde{\mathbf{D}}^{-\frac{1}{2}}\widetilde{\mathbf{A}} \widetilde{\mathbf{D}}^{-\frac{1}{2}}$, we have $\operatorname{tr}\left(\mathbf{F}^{\top} \widetilde{\mathbf{L}} \mathbf{F}\right)=\operatorname{tr}\left(\mathbf{F}^{\top} (\mathbf{I}-\widetilde{\mathbf{D}}^{-\frac{1}{2}}\widetilde{\mathbf{A}} \widetilde{\mathbf{D}}^{-\frac{1}{2}}) \mathbf{F}\right)=\operatorname{tr}\left(\mathbf{F}^{\top}\mathbf{F}\right)-\operatorname{tr}\left(\mathbf{F}^{\top} \widetilde{\mathbf{D}}^{-\frac{1}{2}}\widetilde{\mathbf{A}} \widetilde{\mathbf{D}}^{-\frac{1}{2}} \mathbf{F}\right)=\operatorname{tr}\left(\mathbf{F}\mathbf{F}^{\top}\right)-\operatorname{tr}\left( \widetilde{\mathbf{D}}^{-\frac{1}{2}}\widetilde{\mathbf{A}} \widetilde{\mathbf{D}}^{-\frac{1}{2}} \mathbf{F}\mathbf{F}^{\top}\right)$. On the other hand, if we set $\widetilde{\mathbf{L}}=\mathbf{I}-\widetilde{\mathbf{D}}^{-1}\widetilde{\mathbf{A}}$, we have $\operatorname{tr}\left(\mathbf{F}^{\top} \widetilde{\mathbf{L}} \mathbf{F}\right)=\operatorname{tr}\left(\mathbf{F}^{\top} (\mathbf{I}-\widetilde{\mathbf{D}}^{-1}\widetilde{\mathbf{A}}) \mathbf{F}\right)=\operatorname{tr}\left(\mathbf{F}^{\top}\mathbf{F}\right)-\operatorname{tr}\left(\mathbf{F}^{\top} \widetilde{\mathbf{D}}^{-1}\widetilde{\mathbf{A}} \mathbf{F}\right)=\operatorname{tr}\left(\mathbf{F}\mathbf{F}^{\top}\right)-\operatorname{tr}\left( \widetilde{\mathbf{D}}^{-1}\widetilde{\mathbf{A}}\mathbf{F}\mathbf{F}^{\top}\right)$. We denote $\mathbf{F}=\left[\begin{array}{c} \mathbf{F}_{1} \\ \vdots \\ \mathbf{F}_{n} \\ \end{array}\right]$ and $\mathbf{F}^{\top}=\left[\mathbf{F}_{1}^{\top} \cdots \mathbf{F}_{n}^{\top}\right]$, where $\mathbf{F}_i=\left[\mathbf{F}_{i1} \cdots \mathbf{F}_{id}\right]$, then we have $\operatorname{tr}\left(\mathbf{F}\mathbf{F}^{\top}\right) = \sum_{i=1}^n \mathbf{F}_{i}\mathbf{F}^{\top}_{i}$. \\ When $\widetilde{\mathbf{L}}=\mathbf{I}-\widetilde{\mathbf{D}}^{-\frac{1}{2}}\widetilde{\mathbf{A}} \widetilde{\mathbf{D}}^{-\frac{1}{2}}$, we have \begin{equation*} \begin{split} &\quad\operatorname{tr}\left( \widetilde{\mathbf{D}}^{-\frac{1}{2}}\widetilde{\mathbf{A}} \widetilde{\mathbf{D}}^{-\frac{1}{2}} \mathbf{F}\mathbf{F}^{\top}\right)\\ &=\operatorname{tr}\left(\left[\begin{array}{cccc} \frac{\mathbf{A}_{11}}{\sqrt{d_{1}+1}\sqrt{d_{1}+1}} & \frac{\mathbf{A}_{12}}{\sqrt{d_{1}+1}\sqrt{d_{2}+1}} & \cdots & \frac{\mathbf{A}_{1n}}{\sqrt{d_{1}+1}\sqrt{d_{n}+1}} \\ \frac{\mathbf{A}_{21}}{\sqrt{d_{2}+1}\sqrt{d_{1}+1}} & \frac{\mathbf{A}_{22}}{\sqrt{d_{2}+1}\sqrt{d_{2}+1}} & \cdots & \frac{\mathbf{A}_{2n}}{\sqrt{d_{2}+1}\sqrt{d_{n}+1}} \\ \vdots & \ddots & \ddots & \vdots \\ \frac{\mathbf{A}_{n1}}{\sqrt{d_{n}+1}\sqrt{d_{1}+1}} & \frac{\mathbf{A}_{n2}}{\sqrt{d_{n}+1}\sqrt{d_{2}+1}} & \cdots & \frac{\mathbf{A}_{nn}}{\sqrt{d_{n}+1}\sqrt{d_{n}+1}} \end{array}\right] \left[\begin{array}{cccc} \mathbf{F}_{1}\mathbf{F}_{1}^{\top} & \mathbf{F}_{1}\mathbf{F}_{2}^{\top} & \cdots & \mathbf{F}_{1}\mathbf{F}_{n}^{\top} \\ \mathbf{F}_{2}\mathbf{F}_{1}^{\top} & \mathbf{F}_{2}\mathbf{F}_{2}^{\top} & \cdots & \mathbf{F}_{2}\mathbf{F}_{n}^{\top} \\ \vdots & \ddots & \ddots & \vdots \\ \mathbf{F}_{n}\mathbf{F}_{1}^{\top} & \mathbf{F}_{n}\mathbf{F}_{2}^{\top} & \cdots & \mathbf{F}_{n}\mathbf{F}_{n}^{\top} \end{array}\right]\right)\\ &=\sum_{i=1}^{n}\sum_{j=1}^{n}\frac{\mathbf{A}_{ij}}{\sqrt{d_{i}+1}\sqrt{d_{j}+1}}\mathbf{F}_{j}\mathbf{F}_{i}^{\top}. \end{split} \end{equation*} On the other hand, when $\widetilde{\mathbf{L}}=\mathbf{I}-\widetilde{\mathbf{D}}^{-1}\widetilde{\mathbf{A}}$, we have \begin{equation*} \begin{split} &\quad\operatorname{tr}\left( \widetilde{\mathbf{D}}^{-1}\widetilde{\mathbf{A}} \mathbf{F}\mathbf{F}^{\top}\right)\\ &=\operatorname{tr}\left(\left[\begin{array}{cccc} \frac{\mathbf{A}_{11}}{d_{1}+1} & \frac{\mathbf{A}_{12}}{d_{1}+1} & \cdots & \frac{\mathbf{A}_{1n}}{d_{1}+1} \\ \frac{\mathbf{A}_{21}}{d_{2}+1} & \frac{\mathbf{A}_{22}}{d_{2}+1} & \cdots & \frac{\mathbf{A}_{2n}}{d_{2}+1}\\ \vdots & \ddots & \ddots & \vdots \\ \frac{\mathbf{A}_{n1}}{d_{n}+1} & \frac{\mathbf{A}_{n2}}{d_{n}+1} & \cdots & \frac{\mathbf{A}_{nn}}{d_{n}+1} \end{array}\right] \left[\begin{array}{cccc} \mathbf{F}_{1}\mathbf{F}_{1}^{\top} & \mathbf{F}_{1}\mathbf{F}_{2}^{\top} & \cdots & \mathbf{F}_{1}\mathbf{F}_{n}^{\top} \\ \mathbf{F}_{2}\mathbf{F}_{1}^{\top} & \mathbf{F}_{2}\mathbf{F}_{2}^{\top} & \cdots & \mathbf{F}_{2}\mathbf{F}_{n}^{\top} \\ \vdots & \ddots & \ddots & \vdots \\ \mathbf{F}_{n}\mathbf{F}_{1}^{\top} & \mathbf{F}_{n}\mathbf{F}_{2}^{\top} & \cdots & \mathbf{F}_{n}\mathbf{F}_{n}^{\top} \end{array}\right]\right)\\ &=\sum_{i=1}^{n}\sum_{j=1}^{n}\frac{\mathbf{A}_{ij}}{d_{i}+1}\mathbf{F}_{j}\mathbf{F}_{i}^{\top}. \end{split} \end{equation*} So when $\widetilde{\mathbf{L}}=\mathbf{I}-\widetilde{\mathbf{D}}^{-\frac{1}{2}}\widetilde{\mathbf{A}} \widetilde{\mathbf{D}}^{-\frac{1}{2}}$, we have \begin{equation*} \begin{split} &\quad\operatorname{tr}\left(\mathbf{F}^{\top} \widetilde{\mathbf{L}} \mathbf{F}\right)\quad \left(\widetilde{\mathbf{L}}=\mathbf{I}-\widetilde{\mathbf{D}}^{-\frac{1}{2}}\widetilde{\mathbf{A}} \widetilde{\mathbf{D}}^{-\frac{1}{2}}\right)\\ &= \operatorname{tr}\left(\mathbf{F}^{\top} (\mathbf{I}-\widetilde{\mathbf{D}}^{-\frac{1}{2}}\widetilde{\mathbf{A}} \widetilde{\mathbf{D}}^{-\frac{1}{2}}) \mathbf{F}\right) \\ &=\operatorname{tr}\left(\mathbf{F}\mathbf{F}^{\top}\right)-\operatorname{tr}\left( \widetilde{\mathbf{D}}^{-\frac{1}{2}}\widetilde{\mathbf{A}} \widetilde{\mathbf{D}}^{-\frac{1}{2}} \mathbf{F}\mathbf{F}^{\top}\right) \\ &= \sum_{i=1}^n \mathbf{F}_{i}\mathbf{F}^{\top}_{i} -\sum_{i=1}^{n}\sum_{j=1}^{n}\frac{\mathbf{A}_{ij}}{\sqrt{d_{i}+1}\sqrt{d_{j}+1}}\mathbf{F}_{j}\mathbf{F}_{i}^{\top} \\ &= \frac{1}{2}\sum_{i=1}^n \mathbf{F}_{i}\mathbf{F}^{\top}_{i} + \frac{1}{2}\sum_{j=1}^n \mathbf{F}_{j}\mathbf{F}^{\top}_{j} -\sum_{i=1}^{n}\sum_{j=1}^{n}\frac{\mathbf{A}_{ij}}{\sqrt{d_{i}+1}\sqrt{d_{j}+1}}\mathbf{F}_{j}\mathbf{F}_{i}^{\top} \\ &=\frac{1}{2}\left(\sum_{i=1}^n \mathbf{F}_{i}\mathbf{F}^{\top}_{i} + \sum_{j=1}^n \mathbf{F}_{j}\mathbf{F}^{\top}_{j} -2\sum_{i=1}^{n}\sum_{j=1}^{n}\frac{\mathbf{A}_{ij}}{\sqrt{d_{i}+1}\sqrt{d_{j}+1}}\mathbf{F}_{j}\mathbf{F}_{i}^{\top}\right) \\ &= \frac{1}{2}\left( \sum_{i=1}^n \sum_{j=1}^n \frac{\mathbf{A}_{ij}\mathbf{F}_{i}\mathbf{F}^{\top}_{i}}{d_i+1} + \sum_{i=1}^n \sum_{j=1}^n \frac{\mathbf{A}_{ij}\mathbf{F}_{j}\mathbf{F}^{\top}_{j}}{d_j+1} -2\sum_{i=1}^{n}\sum_{j=1}^{n}\frac{\mathbf{A}_{ij}}{\sqrt{d_{i}+1}\sqrt{d_{j}+1}}\mathbf{F}_{j}\mathbf{F}_{i}^{\top}\right) \text{undirected graph}\\ &=\frac{1}{2}\left(\sum_{i=1}^{n}\sum_{j=1}^{n}\left(\frac{\mathbf{A}_{ij}\mathbf{F}_{i}\mathbf{F}^{\top}_{i}}{d_i+1} +\frac{\mathbf{A}_{ij}\mathbf{F}_{j}\mathbf{F}^{\top}_{j}}{d_j+1}-\frac{\mathbf{A}_{ij}}{\sqrt{d_{i}+1}\sqrt{d_{j}+1}}\mathbf{F}_{j}\mathbf{F}_{i}^{\top}-\frac{\mathbf{A}_{ij}}{\sqrt{d_{i}+1}\sqrt{d_{j}+1}}\mathbf{F}_{i}\mathbf{F}_{j}^{\top}\right)\right)\\ &=\frac{1}{2}\left(\sum_{i=1}^{n}\sum_{j=1}^{n}\mathbf{A}_{ij}\left(\frac{\mathbf{F}_{i}\mathbf{F}^{\top}_{i}}{d_i+1} +\frac{\mathbf{F}_{j}\mathbf{F}^{\top}_{j}}{d_j+1}-\frac{\mathbf{F}_{j}\mathbf{F}_{i}^{\top}}{\sqrt{d_{i}+1}\sqrt{d_{j}+1}}-\frac{\mathbf{F}_{i}\mathbf{F}_{j}^{\top}}{\sqrt{d_{i}+1}\sqrt{d_{j}+1}}\right)\right)\\ &=\frac{1}{2}\left(\sum_{i=1}^{n}\sum_{j=1}^{n}\mathbf{A}_{ij}\left(\frac{\mathbf{F}_i}{\sqrt{d_i+1}}-\frac{\mathbf{F}_j}{\sqrt{d_j+1}}\right)\left(\frac{\mathbf{F}_i^{\top}}{\sqrt{d_i+1}}-\frac{\mathbf{F}_j^{\top}}{\sqrt{d_j+1}}\right)\right)\\ &=\frac{1}{2}\left(\sum_{i=1}^{n}\sum_{j=1}^{n}\mathbf{A}_{ij}\left\|\frac{\mathbf{F}_{i}}{\sqrt{d_{i}+1}}-\frac{\mathbf{F}_{j}}{\sqrt{d_{j}+1}}\right\|_{2}^{2}\right) = \sum_{(i, j) \in \mathcal{E}} \mathbf{A}_{i j}\left\|\frac{\mathbf{F}_{i}}{\sqrt{d_{i}+1}}-\frac{\mathbf{F}_{j}}{\sqrt{d_{j}+1}}\right\|_{2}^{2}. \end{split} \end{equation*} On the other hand, when $\widetilde{\mathbf{L}}=\mathbf{I}-\widetilde{\mathbf{D}}^{-1}\widetilde{\mathbf{A}}$, we have \begin{equation*} \begin{split} &\operatorname{tr}\left(\mathbf{F}^{\top} \widetilde{\mathbf{L}} \mathbf{F}\right)\quad \left(\widetilde{\mathbf{L}}=\mathbf{I}-\widetilde{\mathbf{D}}^{-1}\widetilde{\mathbf{A}} \right)\\ &= \operatorname{tr}\left(\mathbf{F}^{\top} (\mathbf{I}-\widetilde{\mathbf{D}}^{-1}\widetilde{\mathbf{A}}) \mathbf{F}\right) \\ &= \sum_{i=1}^n \mathbf{F}_{i}\mathbf{F}^{\top}_{i} -\sum_{i=1}^{n}\sum_{j=1}^{n}\frac{\mathbf{A}_{ij}}{d_{i}+1}\mathbf{F}_{j}\mathbf{F}_{i}^{\top} \\ &= \frac{1}{2}\sum_{i=1}^n \mathbf{F}_{i}\mathbf{F}^{\top}_{i} + \frac{1}{2}\sum_{j=1}^n \mathbf{F}_{j}\mathbf{F}^{\top}_{j} -\sum_{i=1}^{n}\sum_{j=1}^{n}\frac{\mathbf{A}_{ij}}{d_{i}+1}\mathbf{F}_{j}\mathbf{F}_{i}^{\top} \\ &= \frac{1}{2}\left( \sum_{i=1}^n \sum_{j=1}^n \frac{\mathbf{A}_{ij}\mathbf{F}_{i}\mathbf{F}^{\top}_{i}}{d_i+1} + \sum_{i=1}^n \sum_{j=1}^n \frac{\mathbf{A}_{ij}\mathbf{F}_{j}\mathbf{F}^{\top}_{j}}{d_i+1} -2\sum_{i=1}^{n}\sum_{j=1}^{n}\frac{\mathbf{A}_{ij}}{\sqrt{d_{i}+1}\sqrt{d_{i}+1}}\mathbf{F}_{j}\mathbf{F}_{i}^{\top}\right) \text{undirected graph}\\ &=\frac{1}{2}\left(\sum_{i=1}^{n}\sum_{j=1}^{n}\mathbf{A}_{ij}\left(\frac{\mathbf{F}_i}{\sqrt{d_i+1}}-\frac{\mathbf{F}_j}{\sqrt{d_i+1}}\right)\left(\frac{\mathbf{F}_i^{\top}}{\sqrt{d_i+1}}-\frac{\mathbf{F}_j^{\top}}{\sqrt{d_i+1}}\right)\right)\\ &=\frac{1}{2}\left(\sum_{i=1}^{n}\sum_{j=1}^{n}\mathbf{A}_{ij}\left\|\frac{\mathbf{F}_{i}}{\sqrt{d_{i}+1}}-\frac{\mathbf{F}_{j}}{\sqrt{d_{i}+1}}\right\|_{2}^{2}\right) = \sum_{(i, j) \in \mathcal{E}} \mathbf{A}_{i j}\left\|\frac{\mathbf{F}_{i}}{\sqrt{d_{i}+1}}-\frac{\mathbf{F}_{j}}{\sqrt{d_{i}+1}}\right\|_{2}^{2}. \end{split} \end{equation*} \section{Datasets Details} Cora, Citeseer, and Pubmed are standard citation network benchmark datasets~\citep{sen2008collective}. Coauthor-CS and Coauthor-Phy are extracted from Microsoft Academic Graph~\citep{shchur2018pitfalls}. Cornell, Texas, Wisconsin, and Actor are constructed by \citet{pei2020geom}. ogbn-products is a large-scale product, constructed by \citet{hu2020open}. \begin{table}[!hbtp] \caption{Datasets statistics} \label{tab:stat} \setlength{\tabcolsep}{2mm} \begin{center} \begin{tabular}{lcccc} \toprule \multicolumn{1}{l}{\textbf{Dataset}} & \# Nodes & \# Edges & \# Features & \# Classes \\ \midrule Cora & 2708 & 5429 & 1433 & 7 \\ Citeseer & 3327 & 4732 & 3703 & 6 \\ Pubmed & 19717 & 44338 & 500 & 3 \\ Cornell & 183 & 295 & 1703 & 5 \\ Texas & 183 & 309 & 1703 & 5 \\ Wisconsin & 251 & 499 & 1703 & 5 \\ Actor & 7600 & 33544 & 931 & 5 \\ Coauthor-CS & 18333 & 81894 & 6805 & 15 \\ Coauthor-Phy & 34493 & 247962 & 8415 & 5 \\ ogbn-products & 2449029 & 61859140 & 100 & 42\\ \bottomrule \end{tabular} \end{center} \end{table} \section{Reproducibility} \subsection{Implementation Details} We use Pytorch~\citep{paszke2019pytorch} and PyG~\citep{fey2019fast} to implement \name and R\name. The codes of baselines are implemented referring to the implementation of MLP\footnote{https://github.com/tkipf/pygcn}\footnote{https://github.com/snap-stanford/ogb/blob/master/examples/nodeproppred/products/mlp.py}, GCN\footnote{https://github.com/tkipf/pygcn}\footnote{https://github.com/snap-stanford/ogb/blob/master/examples/nodeproppred/products/gnn.py}, GAT\footnote{https://github.com/pyg-team/pytorch\_geometric/blob/master/examples/gat.py}, GLP\footnote{https://github.com/liqimai/Efficient-SSL}, S$^2$GC\footnote{https://github.com/allenhaozhu/SSGC}, and IRLS\footnote{https://github.com/FFTYYY/TWIRLS}. All the experiments in this work are conducted on a single NVIDIA Tesla A100 with 80GB memory size. The software that we use for experiments are Python 3.6.8, pytorch 1.9.0, pytorch-scatter 2.0.9, pytorch-sparse 0.6.12, pyg 2.0.3, ogb 1.3.4, numpy 1.19.5, torchvision 0.10.0, and CUDA 11.1. \subsection{Hyperparameter Details} \label{appendix:hyperparameter} We provide details about hyparatemeters of \name and R\name in Table~\ref{tab:citation_hyper}, \ref{tab:heterophily_hyper}, \ref{tab:co-author_hyper}, \ref{tab:products_hyper}, and \ref{tab:citation_hyper_flip}. \begin{table}[t] \setlength{\tabcolsep}{0.6mm} \caption{The hyper-parameters for \name and R\name on three citation datasets.} \label{tab:citation_hyper} \vskip 0.15in \centering \begin{tabular}{l|c|c|c|c|c|c|c|c|c|c} \toprule Model & dataset & runs & lr & epochs & wight decay & hidden & dropout & $S$ & $\lambda$ & $\epsilon$\\ \midrule \name & Cora & 100 & 0.2 & 100 & 1e-5 & 0 & 0 & 16 & 32 & - \\ \name & Citeseer & 100 & 0.2 & 100 & 1e-5 & 0 & 0 & 16 & 32 & - \\ \name & Pubmed & 100 & 0.2 & 100 & 1e-5 & 0 & 0 & 16 & 32 & - \\ \hline R\name & Cora & 100 & 0.2 & 100 & 1e-5 & 0 & 0 & 16 & 32 & 1 \\ R\name & Citeseer & 100 & 0.2 & 100 & 1e-5 & 0 & 0 & 16 & 32 & 1 \\ R\name & Pubmed & 100 & 0.2 & 100 & 1e-5 & 0 & 0 & 16 & 32 & 1 \\ \bottomrule \end{tabular} \end{table} \begin{table}[t] \setlength{\tabcolsep}{0.6mm} \caption{The hyper-parameters for \name and R\name on four heterophily graphs.} \label{tab:heterophily_hyper} \vskip 0.15in \centering \begin{tabular}{l|c|c|c|c|c|c|c|c|c|c|c|c} \toprule Model & dataset & noise level & runs & lr & epochs & wight decay & hidden & dropout & $S$ & $\lambda$ & $\epsilon$ & +MLP\\ \midrule \name & Cornell & 0.01 & 10 & 0.2 & 200 & 5e-4 & 16 & 0.5 & 16 & 1 & - & y \\ \name & Cornell & 1 & 10 & 0.2 & 200 & 5e-4 & 16 & 0.5 & 16 & 1024 & - & y \\ \name & Texas & 0.01 & 10 & 0.2 & 200 & 5e-4 & 16 & 0.5 & 16 & 1 & - & y \\ \name & Texas & 1 & 10 & 0.2 & 200 & 5e-4 & 16 & 0.5 & 16 & 1024 & - & y \\ \name & Wisconsin & 0.01 & 10 & 0.2 & 1000 & 5e-4 & 16 & 0.5 & 2 & 1 & - & y \\ \name & Wisconsin & 1 & 10 & 0.2 & 1000 & 5e-4 & 16 & 0.5 & 2 & 1024 & - & y \\ \name & Actor & 0.01 & 10 & 0.2 & 1000 & 5e-4 & 16 & 0.5 & 2 & 1 & - & y \\ \name & Actor & 1 & 10 & 0.2 & 1000 & 5e-4 & 16 & 0.5 & 2 & 1024 & - & y \\ \hline R\name & Cornell & 0.01 & 10 & 0.2 & 200 & 5e-4 & 16 & 0.5 & 16 & 1 & 1 & y \\ R\name & Cornell & 1 & 10 & 0.2 & 200 & 5e-4 & 16 & 0.5 & 16 & 1024 & 1 & y \\ R\name & Texas & 0.01 & 10 & 0.2 & 200 & 5e-4 & 16 & 0.5 & 16 & 1 & 1 & y \\ R\name & Texas & 1 & 10 & 0.2 & 200 & 5e-4 & 16 & 0.5 & 16 & 1024 & 1 & y \\ R\name & Wisconsin & 0.01 & 10 & 0.2 & 1000 & 5e-4 & 16 & 0.5 & 2 & 1 & 1e-5 & y \\ R\name & Wisconsin & 1 & 10 & 0.2 & 1000 & 5e-4 & 16 & 0.5 & 2 & 1024 & 1e-5 & y \\ R\name & Actor & 0.01 & 10 & 0.2 & 1000 & 5e-4 & 16 & 0.5 & 2 & 1 & 1e-5 & y \\ R\name & Actor & 1 & 10 & 0.2 & 1000 & 5e-4 & 16 & 0.5 & 2 & 1024 & 1e-5 & y \\ \bottomrule \end{tabular} \end{table} \begin{table}[t] \setlength{\tabcolsep}{0.6mm} \caption{The hyper-parameters for \name and R\name on two co-author datasets.} \label{tab:co-author_hyper} \vskip 0.15in \centering \begin{tabular}{l|c|c|c|c|c|c|c|c|c|c|c} \toprule Model & dataset & noise level & runs & lr & epochs & wight decay & hidden & dropout & $S$ & $\lambda$ & $\epsilon$ \\ \midrule \name & Coauthor-CS & 0.1 & 10 & 0.2 & 1000 & 1e-7 & 0 & 0 & 16 & 1 & - \\ \name & Coauthor-CS & 1 & 10 & 0.2 & 1000 & 1e-7 & 0 & 0 & 16 & 128 & - \\ \name & Coauthor-Phy & 0.1 & 10 & 0.2 & 200 & 5e-4 & 16 & 0.5 & 16 & 1 & - \\ \name & Coauthor-Phy & 1 & 10 & 0.2 & 200 & 5e-4 & 16 & 0.5 & 16 & 1024 & - \\ \hline R\name & Coauthor-CS & 0.1 & 10 & 0.2 & 1000 & 1e-7 & 0 & 0 & 16 & 1 & 1 \\ R\name & Coauthor-CS & 1 & 10 & 0.2 & 1000 & 1e-7 & 0 & 0 & 16 & 128 & 1 \\ R\name & Coauthor-Phy & 0.1 & 10 & 0.2 & 200 & 5e-4 & 16 & 0.5 & 16 & 1 & 1 \\ R\name & Coauthor-Phy & 1 & 10 & 0.2 & 200 & 5e-4 & 16 & 0.5 & 16 & 1024 & 1 \\ \bottomrule \end{tabular} \end{table} \begin{table}[t] \setlength{\tabcolsep}{0.6mm} \caption{The hyper-parameters for \name and R\name on ogbn-products dataset.} \label{tab:products_hyper} \vskip 0.15in \centering \begin{tabular}{l|c|c|c|c|c|c|c|c|c|c|c} \toprule Model & noise level & runs & lr & epochs & hidden & dropout & $S$ & $\lambda$ & $\epsilon$ & layers & +MLP \\ \midrule \name & 0.1 & 10 & 0.01 & 300 & 256 & 0.5 & 128 & 32 & - & 3 & y \\ \name & 1 & 10 & 0.01 & 300 & 256 & 0.5 & 128 & 256 & - & 3 & y\\ \hline R\name & 0.1 & 10 & 0.01 & 300 & 256 & 0.5 & 128 & 32 & 1e-2 & 3 & y \\ R\name & 1 & 10 & 0.01 & 300 & 256 & 0.5 & 128 & 256 & 1e-2 & 3 & y\\ \bottomrule \end{tabular} \end{table} \begin{table}[t] \setlength{\tabcolsep}{0.6mm} \caption{The hyper-parameters for \name and R\name on three citation datasets of the flipping experiments.} \label{tab:citation_hyper_flip} \vskip 0.15in \centering \begin{tabular}{l|c|c|c|c|c|c|c|c|c|c|c} \toprule Model & dataset & flip probability & runs & lr & epochs & wight decay & hidden & dropout & $S$ & $\lambda$ & $\epsilon$\\ \midrule \name & Cora & 0.1 & 100 & 0.2 & 100 & 1e-5 & 0 & 0 & 32 & 64 & - \\ \name & Cora & 0.2 & 100 & 0.2 & 100 & 1e-5 & 0 & 0 & 16 & 32 & - \\ \name & Cora & 0.4 & 100 & 0.2 & 100 & 1e-5 & 0 & 0 & 16 & 32 & - \\ \name & Citeseer & 0.1 & 100 & 0.2 & 100 & 1e-5 & 0 & 0 & 16 & 32 & - \\ \name & Citeseer & 0.2 & 100 & 0.2 & 100 & 1e-5 & 0 & 0 & 16 & 32 & - \\ \name & Citeseer & 0.4 & 100 & 0.2 & 100 & 1e-5 & 0 & 0 & 16 & 32 & - \\ \name & Pubmed & 0.1 & 100 & 0.2 & 100 & 1e-5 & 0 & 0 & 16 & 32 & - \\ \name & Pubmed & 0.2 & 100 & 0.2 & 100 & 1e-5 & 0 & 0 & 16 & 32 & - \\ \name & Pubmed & 0.4 & 100 & 0.2 & 100 & 1e-5 & 0 & 0 & 16 & 32 & - \\ \hline R\name & Cora & 0.1 & 100 & 0.2 & 100 & 1e-5 & 0 & 0 & 32 & 64 & 1e-5 \\ R\name & Cora & 0.2 & 100 & 0.2 & 100 & 1e-5 & 0 & 0 & 16 & 32 & 1e-5 \\ R\name & Cora & 0.4 & 100 & 0.2 & 100 & 1e-5 & 0 & 0 & 16 & 32 & 1e-1 \\ R\name & Citeseer & 0.1 & 100 & 0.2 & 100 & 1e-5 & 0 & 0 & 16 & 32 & 1e-5 \\ R\name & Citeseer & 0.2 & 100 & 0.2 & 100 & 1e-5 & 0 & 0 & 16 & 32 & 1e-5 \\ R\name & Citeseer & 0.4 & 100 & 0.2 & 100 & 1e-5 & 0 & 0 & 16 & 32 & 1e-5 \\ R\name & Pubmed & 0.1 & 100 & 0.2 & 100 & 1e-5 & 0 & 0 & 16 & 32 & 1e-1 \\ R\name & Pubmed & 0.2 & 100 & 0.2 & 100 & 1e-5 & 0 & 0 & 16 & 32 & 1e-1 \\ R\name & Pubmed & 0.4 & 100 & 0.2 & 100 & 1e-5 & 0 & 0 & 16 & 32 & 1e-1 \\ \bottomrule \end{tabular} \end{table} \section{Additional Experiments} \subsection{Analysis on Row Normalization} \label{appendix:row norm} \begin{table}[htbp] \normalsize \caption{Summary of results of NGC w/o raw normalization on three datasets in terms of classification accuracy (\%)} \label{tab:row normalization} \setlength{\tabcolsep}{2.5mm} \begin{center} \begin{tabular}{cccccccccc} \toprule \multirow{2}{*}{Noise Level} & \multicolumn{3}{c}{Cora} & \multicolumn{3}{c}{Citeseer} & \multicolumn{3}{c}{Pubmed} \\ \cmidrule(r){2-4} \cmidrule(r){5-7} \cmidrule(r){8-10} & 1 & 10 & 100 & 1 & 10 & 100 & 1 & 10 & 100 \\ \midrule w/o RN &68.3 &59.7 &56.1 &43.5 &40.4 &37.6 &43.1 &38.8 &37.4 \\ w RN &66.1 &65.5 &66.2 &45.3 &45.1 &44.8 &62.3 &62.7 &62.1 \\ \bottomrule \end{tabular} \end{center} \end{table} In this section, we analyze the influence of row normalization on denoising performance. The noise level $\xi$ controls the magnitude of the Gaussian noise we add to the feature matrix: $\mathbf{X}+\xi\bm{\eta}$ where $\bm{\eta}$ is sampled from standard i.i.d., Gaussian distribution. For Cora, Citeseer, and Pubmed, we test $\xi \in \{1, 10, 100\}$. From Table~\ref{tab:row normalization}, we can observe that the denoising performance of w/ row normalization is better than w/o row normalization. Since row normalization can shrink the value of elements in $\bm{\eta}$, thus reducing the variance $\sigma$. In other words, row normalization make $\left\|\widetilde{\bm{\mathcal{A}}}_S\bm{\eta}\right\|_{F}^{2}$ converge to zero faster. \subsection{Analysis on the Depth of \name and R\name} \label{appendix:depth analysis} In this section, we analyze the influence of the depth of \name and R\name model on denoising performance by testing the classification accuracy on semi-supervised node classification tasks. We conduct two sets of experiments: with/without noise in feature matrix. For experiment with feature noise, we simple fix the noise level $\xi =1$. In each set of experiments, we evaluate the test accuracy with respect to \name and R\name model depth, which corresponding to the value of $S$ in $\widetilde{\mathcal{A}}_{S}$. From Figure~\ref{fig:depth_ngc} and \ref{fig:depth_rngc}, we can observe that the test accuracy barely changes with depth if the model is trained on the clean features on Cora and Pubmed but changes greatly if the model is trained on the clean feature on Citeseer. In this regard, the over-smoothing issue exists in R\name model on citeseer. However, the denoising performance of shallow R\name is not good as deeper R\name models, especially on the large graph like Pubmed. This suggests that we do need to increase the depth of GNN model to include more higher-order neighbors for better denoising performances. \begin{figure}[t] \begin{center} \includegraphics[width=0.325\textwidth]{images/cora_compar.pdf} \includegraphics[width=0.325\textwidth]{images/citeseer_compar.pdf} \includegraphics[width=0.325\textwidth]{images/pubmed_compar.pdf} \end{center} \caption{Comparison of classification accuracy v.s. \name model depth on semi-supervised node classification tasks. The experiments are conducted on clean and noisy features.} \label{fig:depth_ngc} \end{figure} \begin{figure}[t] \begin{center} \includegraphics[width=0.325\textwidth]{images/cora_compar_rngc.pdf} \includegraphics[width=0.325\textwidth]{images/citeseer_compar_rngc.pdf} \includegraphics[width=0.325\textwidth]{images/pubmed_compar_rngc.pdf} \end{center} \caption{Comparison of classification accuracy v.s. R\name model depth on semi-supervised node classification tasks. The experiments are conducted on clean and noisy features.} \label{fig:depth_rngc} \end{figure} \section{Introduction} \input{01introduction} \section{A Simple Unifying Framework: Neumann Graph Convolution} \label{sec:NGC} \input{03NGC} \section{Main Theory} \label{sec:main theory} \input{04theory} \section{Robust Neumann Graph Convolution} \label{sec:RNGC} \input{05adversarial} \section{Experiments} \label{sec:experiments} \input{06experiment} \section{Related Work} \input{07related} \section{Conclusion} \input{08conclusion} \newpage \normalem

81ba458b63b4ac3213f8254460aee132e301e2a2

\section{Introduction} Deep-inelastic scattering (DIS) of electrons off ions is at the forefront of experimental efforts to probe the internal structure of nucleons and nuclei and will be a primary focus of study at the Electron-Ion Collider. In semi-inclusive DIS, selected particles produced by the fragmentation of the struck quark are observed in coincidence with the scattered electron, \(e(k) + N(P) \rightarrow e(k') + h(p) + X\), resulting in observables which provide access to a convolution of parton distribution functions (PDFs), describing the momentum of partons within the nucleon, and fragmentation functions (FFs), describing the probability of producing a final state particle with some momentum from the struck quark in the factorization approach. \cite{bacchetta_dalesio_diehl_miller_2004} \begin{figure}[h] \centering \includegraphics[width=8cm]{trentoConvention.pdf} \caption{Definition of semi-inclusive kinematic variables in target rest frame, reprinted from \cite{bacchetta_dalesio_diehl_miller_2004}. } \label{fig:sidis} \end{figure} The kinematic variables describing the DIS process, with center of mass energy squared \(s=(P+k)^{2}\) can be defined in terms of the virtual photon four-momentum \(q\) as \cite{blumlein_2013}, \begin{equation} Q^{2} = -q^{2},\; y = \frac{P\cdot q }{P\cdot k}, \; x = \frac{Q^{2}}{sy} \end{equation} \section{SIDIS kinematic reconstruction} In semi-inclusive DIS, observables are extracted in the nucleon center of mass frame, with the SIDIS cross-section a function of the inclusive DIS variables as well as (\(p_{h\perp}, z, \phi_h\)). The relevant transverse momentum is defined with respect to the virtual photon axis, and the single-hadron azimuthal angle \(\phi_h\) is defined between the lepton scattering plane and hadron production plane (figure 1). \(z\) is defined as \( z = \frac{p_{h}\cdot P}{q\cdot P}\) The calculation of SIDIS kinematics therefore requires precise reconstruction of the four-momenta of the selected hadron and the exchanged virtual-photon. \subsection{Electron method} Extraction of SIDIS observables and multiplicities at the EIC presents a new challenge, as fully multi-dimensional SIDIS studies have so far only been carried out in lower energy fixed target experiments. In fixed target SIDIS studies, \(q\) has been determined using only the scattered electron, \(q = k - k'\). However, studies done for the EIC yellow report and EIC detector proposals have found that a significant contribution to uncertainty in SIDIS kinematics is poor reconstruction of the virtual photon four-momentum when using only the scattered electron. In particular, the electron method fails in such regions of kinematic phase space at the EIC such as at low y (\(y < 0.05\)), where the energy loss of the electron is small and not well resolved. This is a significant issue for the study of TMD effects at e-p colliders, as at low-\(Q^2\) and large-x spin-orbit correlations are expected to be most significant and higher twist effects are observable. Additionally, the low-y region will be critical for overlapping the phase space covered by the EIC and SIDIS studies carried out at other facilities such as Jefferson Lab. \begin{figure} \centering \includegraphics[width=10cm]{PIDcoverage_cross-25_pip_z02_flat.pdf} \caption{Momentum vs pseudorapidity in \(x-Q^{2}\) bins for positive pions with \(z > 0.2\) and requiring \(y > 0.05\), generated with Pythia-8. Different histogram colors represent different \(\sqrt{s}\) values. Red bands represent 3-\(\sigma\) PID coverage based on ATHENA proposal projections. } \label{fig:coverage} \end{figure} \subsection{Hadronic final state methods} Fast simulation studies for the EIC yellow report~\cite{yellow_report} and ATHENA (A Totally Hermetic Electron Nucleon Apparatus) proposal~\cite{athena_proposal} have demonstrated that DIS reconstruction methods developed at past e-p colliders~\cite{blumlein_2013} can be used to improve the reconstruction of inclusive DIS variables at the EIC. The DIS reconstruction methods developed at HERA utilized combinations of measured quantities from the scattered electron and the hadronic final state (HFS). The use of the HFS allowed these additional methods, such as the double angle (DA) and \(\Sigma\)-methods ~\cite{blumlein_2013}, to improve inclusive DIS kinematic reconstruction for various regions of the HERA kinematic space, as well as to make the reconstruction robust with respect to QED radiative effects ~\cite{Bassler:1994uq,RevModPhys.71.1275}. For the studies planned at the EIC, methods utilizing the HFS must be extended to the reconstruction of SIDIS kinematics. The authors of this contribution conducted first studies of SIDIS kinematic reconstruction for the EIC and demonstrated methods in which the hadronic final state can be used to improve the reconstruction of the virtual photon four momentum. This was carried out in the EIC yellow report and ATHENA proposal~\cite{yellow_report,athena_proposal} by first obtaining the transverse component of \(q\) from the recoil of the HFS transverse to the beamline through a sum of the momenta of HFS particles. Following the determination of this transverse recoil, the remaining two components of \(q\) can be constrained by the system of equations including \(q\) from the definitions of \(Q^2\) and \(y\), \begin{equation} q_x = \sum_{i}^{N_{HFS}} p_{x,i}, q_y = \sum_{i}^{N_{HFS}} p_{y,i}\end{equation} \begin{equation} q_z, q_t \leftarrow \begin{cases} Q^2 = -( q_x^2 + q_y^2 + q_z^2 - q_t^2 )\\ y = \frac{P_x q_x + P_y q_y + P_z q_z - P_t q_t}{P \cdot k} \end{cases} \end{equation} In the EIC yellow report and ATHENA proposal \cite{yellow_report,athena_proposal}, this procedure was carried out using various inclusive DIS reconstruction methods developed at HERA\cite{blumlein_2013}, in fast simulations showing improvements over the electron method in some regions of the DIS kinematic space. As methods such as the Jacquet-Blondel (JB) method \cite{blumlein_2013} use only the hadronic final state information, this also allows for the determination of \(q\) from the HFS alone. Results using this approach are shown in the next section compared to ML and electron methods, with resolution using this method expected to improve with further developed full simulations based on fast simulation results. \section{Machine learning kinematic reconstruction} \subsection{Network architecture} Multiple studies have been conducted demonstrating an improved resolution of inclusive DIS variables \(Q^2, y, x\) through deep learning approaches \cite{arratia_britzger_long_nachman_2022,diefenthaler_farhat_verbytskyi_xu_2021}, but these have not yet been extended to reconstruction of semi-inclusive DIS kinematics. In this study, we demonstrate that machine learning models which learn from the full HFS and scattered electron can be used to improve on current reconstruction methods to provide reliable reconstruction of the virtual photon axis across all of the DIS kinematic coverage at the EIC. This approach to semi-inclusive DIS reconstruction is centered on the use of deep neural networks to better leverage the full hadronic final state at the level of reconstructed tracks. While previous applications of deep learning to inclusive DIS reconstruction directly regressed the kinematic variables of interest~\cite{arratia_britzger_long_nachman_2022,diefenthaler_farhat_verbytskyi_xu_2021}, this study aims to improve kinematics by directly regressing the virtual photon four-momentum in the lab frame. Improvements to the HFS reconstruction are carried out through the use of Particle Flow Networks~\cite{Komiske:2018cqr}. Particle Flow Networks are an application of the deep sets neural network architecture, which learns a function of an unordered set of objects rather than from a fixed size input. The network consists of fully connected linear neural network layers which take as input the features of each particle individually, the outputs of which are summed over all particles to create a latent space representation of the event. The latent space variables and supplied global features of the event are then passed to another set of dense layers which produce the final output of the network ~\cite{Komiske:2018cqr}. Particle flow networks have seen particular success in tasks such as jet classification at the LHC. Particle flow networks implemented in Keras \cite{chollet2015keras} are included in the EnergyFlow python package. ~\cite{Komiske:2018cqr} \begin{figure} \centering \includegraphics[width=8cm]{particleFlow_SIDIS.drawio.pdf} \caption{Network diagram of Particle Flow Networks \cite{Komiske:2018cqr} with global event features. Features of each HFS particle supplied individually to layers \(\Phi\), then summed over to form a latent space representation. Latent space features and global features of event (green), including reconstructed inclusive DIS variables and DIS electron momentum, supplied to layer \(F\) which produces final output. } \label{fig:pfn_diagram} \end{figure} \subsection{Variables and dataset} \begin{figure*} \centering \includegraphics[width=8cm]{pTmean_vsy.pdf} \includegraphics[width=8cm]{pTrms_vsy.pdf} \caption{SIDIS \(p_{h\perp}\) resolution mean (left) and RMS (right) as a function of \(y_{true}\) for positive pions with \(z>0.2\), \(p_{h\perp} > 0.1 GeV\). HFS methods surpass electron method for very low \(y\), while PFN equals or outperforms electron method for all \(y\). } \label{fig:fullvsy_pt} \end{figure*} \begin{figure*} \centering \includegraphics[width=8cm]{phiHmean_vsy.pdf} \includegraphics[width=8cm]{phiHrms_vsy.pdf} \caption{SIDIS \(\phi_{h}\) resolution mean (left) and RMS (right) as a function of \(y_{true}\) for positive pions with \(z>0.2\), \(p_{h\perp} > 0.1 GeV\). } \label{fig:fullvsy_phi} \end{figure*} The features of the hadronic final state reconstructed particles provided to the particle flow network include the four-momentum of each particle, as well as the lab frame azimuthal angle and pseudorapidity to provide direct information on angular acceptance in addition to momentum in each direction. The global features used for training include the four-momentum of the scattered electron and the DIS variables \(x\) and \(Q^2\) from the electron, DA, and JB methods. By supplying the full electron four-momentum following the single-particle layers \(\Phi\), the model is intended to learn corrections to the electron method based on the hadronic final state latent space variables. When a greater amount of fully simulated EIC simulated data is available, the DIS methods could also be replaced by the output of the deep learning models for inclusive DIS variables described previously. The particle flow network was trained to predict the full four-momentum \(q\) in the lab frame. The particle flow network, implemented in Keras \cite{chollet2015keras} and available in the EnergyFlow python package, is used with per-particle dense layer units \(\phi = (500, 500, 500)\), \(l = 512\), and final dense layer units \(F = (200, 200, 200) \). Both the layers making up \(\phi\) and \(F\) employ a relu activation function, with the final output layer having linear activation. \begin{figure*} \centering \includegraphics[width=8cm]{pTmean_vspT.pdf} \includegraphics[width=8cm]{pTrms_vspT.pdf} \caption{SIDIS \(p_{h\perp}\) resolution mean (left) and RMS (right) as a function of \(p_{h\perp,true}\) for positive pions with \(z>0.2\). } \label{fig:fullvspT_largey_pT} \end{figure*} \begin{figure*} \centering \includegraphics[width=8cm]{phiHmean_vspT.pdf} \includegraphics[width=8cm]{phiHrms_vspT.pdf} \caption{SIDIS \(\phi_{h}\) resolution mean (left) and RMS (right) as a function of \(p_{h\perp,true}\) for positive pions with \(z>0.2\). } \label{fig:fullvspT_largey_phi} \end{figure*} The dataset used for the training and testing of this model was the ATHENA full simulation developed for the ATHENA detector proposal for the first interaction region at the EIC. ATHENA was developed with the objective of meeting the resolution and physics goals laid out in the EIC yellow report. The ATHENA full simulation was implemented in DD4hep, Geant4, and Juggler \cite{frank_markus_2018_1464634,ALLISON2016186,jugglerGit}. At the time of the detector proposal, PID algorithms were not fully implemented, meaning PID information was not included in this model. Additionally, the scattered electron was taken as always correctly identified by matching the scattered electron with the MC truth information. The simulated event sample used for model training and testing was a neutral current DIS sample generated using Pythia-8 \cite{pythia-manual-2022}, with additional beam smearing and crossing angle effects implemented. 3 million events with \(Q^2 > 1 \: \mathrm{GeV}^2\) and 2 million events with \(Q^2 > 10 \: \mathrm{GeV}^2\) were used for training with 1 million \(Q^2 > 1 \: \mathrm{GeV}^2\) set aside for model validation. \section{Results} As a function of \(y\) (figures 4 and 5), using the virtual photon four-momentum as predicted by the neural network model results in significantly improved reconstruction of \(p_{h\perp}\), \(\phi_h\) and \(z\) for low-y, when compared to both the electron method and methods utilizing information from the hadronic final state. The neural network reconstruction of \(q\) results in a distribution of the SIDIS variables which is both better centered around the true value, and with a significantly smaller RMS where the electron method begins to fail at low-y. At large-y, the neural network achieves performance only slightly surpassing that of the electron method, which is expected based on the projected energy and tracking resolution for the scattered electron with ATHENA. As a function of \(p_{h\perp,true}\), we also observe a significant improvement in kinematic reconstruction for both transverse momentum and for the semi-inclusive azimuthal angle. As the electron method begins to degrade for lower values of \(p_{h\perp}\), the neural network reconstruction of \(q\) results in stable performance to the lowest values of \(p_{h\perp}\) in the dataset. \section{Summary} The EIC will provide the first opportunity for semi-inclusive DIS measurements in an e-A collider context, giving access to new kinematic regions in which to precisely explore the 3-dimensional spin structure of nucleons. The development of reliable kinematic reconstruction methods will be critical to enabling precision extraction of SIDIS observables, especially at low-y. This can be achieved through the use of information from the hadronic final state alongside the scattered electron. As demonstrated in this contribution, machine learning, here using particle flow networks, can combine the information from the scattered electron and full HFS to provide reliable SIDIS kinematic reconstruction across the DIS variable space. Further steps in this work will include the consideration of QED radiative effects on SIDIS reconstruction, as well as possible extension to other neural network architectures exploiting correlations between particles. Additionally, this approach will continue to be studied and validated as more detailed full detector simulations are developed for the EIC. \section{Acknowledgements} We thank Markus Diefenthaler for helpful discussions throughout the development of the described machine learning approach and for comments on this submission. \clearpage

0a2b7507c6a77cb03a7cc7d63e3a215bf1aa3f38

\section{Introduction and Summary} Fundamental objects in relativistic theory are fields whose dynamics is derived from the action principle. The action is manifestly covariant object that is defined as the space-time integral of Lagrangian density. This manifest covariance is lost in the Hamiltonian formalism since its crucial point is an existence of one preferred coordinate which is the time coordinate. There is an alternative Hamiltonian formalism, known as covariant Hamiltonian theory or the Weyl-De Donder theory \cite{DeDonder, Weyl} that maintains manifest covariance. In the covariant Hamiltonian theory, the momenta are defined as derivatives with respect to all partial derivatives of coordinates so that there is no preferred direction and hence manifest covariance is preserved. This is very attractive idea that could be especially useful in manifest covariant theories as for example theory of gravity. In fact, the first covariant Hamiltonian theory of General Relativity was published by Ho\v{r}ava long time ago \cite{Horava:1990ba}. Recently, this work was further discussed examining its thermodynamic consequences in \cite{Parattu:2013gwa}. The obvious next step is to apply this approach to some generalized theories of gravity. The simplest one is the $F(R)$ gravity where the scalar curvature in the Lagrangian is replaced by a general function $F(R)$ of scalar curvature \footnote{For review of this theory see \cite{DeFelice:2010aj, Nojiri:2017ncd, Nojiri:2010wj}.}. The covariant Hamiltonian for $F(R)$ gravity in Jordan frame was found in \cite{Kluson:2020tzn}. It is well known that $F(R)$ gravity can also be formulated in Einstein frame, for detailed discussion, see for example \cite{DeFelice:2010aj}. The transformations between these two frames is based on the Weyl transformation of metric and corresponding Riemann and Ricci tensor. In more details, with appropriate chosen Weyl factor we can arrive to Einstein frame formulation of $F(R)$ gravity. $F(R)$ Lagrangian in Einstein frame is very similar to General Relativity Lagrangian minimally coupled to a scalar field. In 3+1 formalism, Hamiltonians for both frames were formulated and there was found that they are related by canonical transformation \cite{Deruelle:2009pu}. This leads to a question whether this is also true in the covariant Hamiltonian formalism. This question is answered with the presented paper. In more details, we firstly derive the covariant Hamiltonian for $F(R)$ gravity in Einstein frame. We find that covariant Hamiltonian for $F(R)$ gravity in Einstein frame has the same form as was found in \cite{Horava:1990ba} together with new additional scalar field contribution. On the other hand the covariant Hamiltonian formulation of $F(R)$ gravity in Jordan frame was performed in \cite{Kluson:2020tzn} and our goal is to show that these two formulations are related by canonical transformations. To do this we should firstly examine how canonical transformations are defined in covariant Hamiltonian formalism, following \cite{Struckmeier:2008zz,Struckmeier:2012de}. Then we study relationship between covariant Hamiltonians for $F(R)$ gravity in Jordan and Einstein frames and we find generating function resulting in proof of canonical transformation between those two Hamiltonians. This is really new and non-trivial result that shows that these two Hamiltonians are related by canonical transformations in the similar way how two Hamiltonians are related in $3+1$ canonical formalism \cite{Deruelle:2009pu}. On the other hand there is crucial difference between these two canonical transformations which is in the preservation of the Poisson brackets. In fact, it is not completely clear how to define Poisson brackets in covariant canonical formalism due to the fact that the conjugate momenta have additional vector index. Then it is natural to define Poisson bracket as in \cite{Struckmeier:2008zz} where the Poisson bracket is defined in the same way as canonical Poisson bracket so that it is now labeled by vector index. It was shown in \cite{Struckmeier:2008zz} that such Poisson brackets are not generally preserved under canonical transformations and we show that exactly this situation occurs in case of canonical transitions between Einstein and Jordan frame covariant Hamiltonians for $F(R)$-gravity. On the other hand we show that the form of Lagrangian brackets is preserved under canonical transformation. We mean that this is very interesting result that demonstrates nice application of the covariant canonical formalism for the study of $F(R)$-gravity. The organization of this paper is as follows. In the next section (\ref{second}) we formulate covariant Hamiltonian for $F(R)$ gravity in Einstein frame. The section (\ref{third}) is devoted to the canonical transformation between these two frames. At first the transformation's existence is proven using fundamental Lagrange brackets, then the explicit form of the generating function of this transformation is found, and at the end the Poisson brackets are also noted, they role is however a minor one since in the covariant Hamiltonian theory they do not serve as canonical invariant. The fourth section (\ref{fourth}) deals with the surface term of the Lagrangian \footnote{ This paper uses the East Coast convention with metric signature ($-,+,+,+$) and Latin indices running over $0...3$ interval while the Greek ones over $1...3$. The fundamental constants $c, G, \hbar, k_B$ are treated as equal to one. }. \section{$F(R)$-gravity in Einstein Frame and Its Covariant Hamiltonian}\label{second} We begin this section with the introduction of the Lagrangian for $F(R)$ gravity. $F(R)$ gravity is the simplest generalization of the Einstein-Hilbert action when we replace the linear dependence of the Lagrangian density on the scalar curvature~$R$ by more general function $F(R)$ \footnote{For review and extended list of references, see for example \cite{DeFelice:2010aj}.}. Explicitly, the Lagrangian density of $F(R)$ theory of gravity has the form \begin{equation}\label{lagFR} \mathcal{L} = \frac{\sqrt{-g}}{16\pi}F(R) \:. \end{equation} The presence of the function $F(R)$ implies that it is not straightforward procedure to find corresponding Hamiltonian. In order to overcome this issue it is convenient to introduce two scalar fields $A$ and $B$ and replace the Lagrangian density (\ref{lagFR}) by the following one \begin{equation}\label{FRextend} \mathcal{L} = \frac{\sqrt{-g}}{16\pi} \hzav{F(B) + A\kzav{R-B}}\ . \end{equation} In fact, equations of motion for $A$ and $B$ that follow from (\ref{FRextend}) have the form \begin{equation}\label{eqFRexed} R-B=0 \ , \quad F'(B)-A=0 \ , \quad F'(B)\equiv \frac{dF}{dB} \ . \end{equation} Then inserting the first equation in (\ref{eqFRexed}) into (\ref{FRextend}), we easily see the Lagrangian density (\ref{FRextend}) reduces into (\ref{lagFR}) that shows equivalence of these two actions. For our purposes, it is useful to use the second equation in (\ref{eqFRexed}) to solve $A$ as function of $B$ and hence the Lagrangian density in Jordan frame has the form \begin{equation} \mathcal{L}^J = \frac{\sqrt{-g}}{16\pi} \hzav{F(B) + F'(B)\kzav{R-B}}\:. \end{equation} As it is well known, we can formulate $F(R)$-gravity in the Einstein frame too, see for example \cite{DeFelice:2010aj}. Note that the Einstein frame is defined by requirement that the Lagrangian density is linear in the scalar curvature $R$. In the rest of this paper, the coordinates with a tilde are coordinates in Einstein frame while those without punctuation belong to Jordan frame. In order to find the transformation from Jordan frame to Einstein one we should perform Weyl transformation of the metric that is defined as \begin{equation}\label{trang} \widetilde{g}_{ij} = F' g_{ij}\ . \end{equation} As the next step we introduce connection with tilde \begin{equation} \widetilde{\Gamma}^i_{jk}=\frac{1}{2}\widetilde{g}^{il} (\partial_j \widetilde{g}_{lk}+\partial_k \widetilde{g}_{lj}-\partial_l \widetilde{g}_{jk}) \ . \end{equation} Then using (\ref{trang}) we find how it is related to $\Gamma^i_{ij}$ and to $F(B)$ \begin{equation} \label{tGamma} \widetilde{\Gamma}^i_{jk} = \Gamma^i_{jk} + \half{1} \delta^i_j \partial_k (\ln F') + \half{1} \delta^i_k \partial_j (\ln F') - \half{1} g_{jk} g^{km}\partial_m (\ln F')\:. \end{equation} With the help of this relation between connections, we can easily find relation between corresponding scalar curvatures and we get \begin{equation} \widetilde{R} = \widetilde{g}^{ab} \widetilde{R}_{ab} = \inv{F'}R - \frac{3}{F'} \Box \ln F' - \frac{3}{2F'} \partial_i \ln F' g^{ij} \partial_j \ln F' \:. \end{equation} For writing the Lagrangian in Einstein frame, we need to express $R$ in terms of $\widetilde{R}$, $\widetilde{g}_{ij}$ and corresponding derivatives. To do this, we use the following rules \begin{equation} \Box \ln F' = F' \widetilde{\Box} \ln F' - F' \kzav{\widetilde{\partial}\ln F'}^2 \:, \quad \kzav{\partial \ln F'}^2 = F'\kzav{\widetilde{\partial}\ln F'}^2 \ . \end{equation} Then we can easily express $R$ as \begin{equation} R = F' \hzav{ \widetilde{R} + 3 \widetilde{\Box} \ln F' - \frac{3}{2} \kzav{\widetilde{\partial} \ln F'}^2} \:. \end{equation} Using this expression together with the fact that $\sqrt{-\widetilde{g}} = F'^2 \sqrt{-g}$ we obtain that the Lagrangian density for $F(R)$ gravity takes the form \begin{equation} \mathcal{L} = \frac{\sqrt{-\widetilde{g}}}{16\pi} \hzav{ \frac{F - F'B}{F'^2} + \widetilde{R} + 3 \widetilde{\Box} \ln F' -\frac{3}{2} \kzav{\widetilde{\partial} \ln F'}^2 } \:, \end{equation} or equivalently \begin{equation} \mathcal{L} = \frac{\sqrt{-\widetilde{g}}}{16\pi} \hzav{ \frac{F - F'B}{F'^2} + \widetilde{R} -\frac{3}{2} \kzav{\widetilde{\partial} \ln F'}^2 } + \frac{3}{16\pi}\partial_i \kzav{\sqrt{-\widetilde{g}} \widetilde{\partial}^i \ln F'} \:. \end{equation} In order to get canonical form of the Lagrangian density, we introduce a new scalar variable $\widetilde{\phi} = \sqrt{\frac{3}{2}} \ln F'$ and a potential $V(B) = \frac{F'B - F}{F'^2}$ so that the Lagrangian density takes Einstein-Hilbert form \begin{equation} \mathcal{L}^E = \frac{\sqrt{-\widetilde{g}}}{16\pi} \hzav{ \widetilde{R} -\widetilde{g}^{ij}\partial_i\widetilde{\phi}\partial_j \widetilde{\phi} - V(\widetilde{\phi}) } + \frac{\sqrt{6}}{16\pi}\partial_i \kzav{\sqrt{-\widetilde{g}}\widetilde{g}^{ij} \partial_j \widetilde{\phi}} \:. \end{equation} This is the final form of Lagrangian density for $F(R)$ gravity in Einstein frame. To proceed to the covariant canonical formalism, it is necessary to separate Lagrangian into two parts \cite{Horava:1990ba}: the bulk term that contains only the first derivatives and the surface term which can be expressed as a total derivative. To construct Hamiltonian, solely the bulk term is needed, the surface Hamiltonian will be discussed in section (\ref{fourth}). Explicitly, we get \begin{eqnarray}\label{mLbulk} &&\mathcal{L}^E=\mathcal{L}^E_{bulk}+\mathcal{L}^E_{sur} \ , \nonumber \\ &&\mathcal{L}_{bulk} = \frac{\sqrt{-\widetilde{g}}}{16\pi} \hzav{ \widetilde{g}^{ab}\widetilde{\Gamma}^c_{bd}\widetilde{\Gamma}^d_{ac} - \widetilde{g}^{ac}\widetilde{\Gamma}^b_{bd}\widetilde{\Gamma}^d_{ac} - \widetilde{g}^{ij}\partial_i \widetilde{\phi} \partial_j \widetilde{\phi} - V(\widetilde{\phi}) } \ , \nonumber \\ &&\mathcal{L}_{sur} = \partial_i \hzav{\frac{\sqrt{-\widetilde{g}}}{16\pi} \kzav{\widetilde{g}^{ab}\widetilde{\Gamma}^i_{ab} - \widetilde{g}^{ai}\widetilde{\Gamma}^b_{ab} + \sqrt{6} \widetilde{g}^{ij}\partial_j \widetilde{\phi} } } \ . \nonumber \\ \label{Lagrangians} \end{eqnarray} We see that the bulk part reminds the bulk Lagrangian of General Relativity \cite{Parattu:2013gwa} together with added $\tilde{\phi}$-related terms. First step to transform the Lagrangian into covariant Hamiltonian is to find corresponding momenta. From (\ref{mLbulk}) we find that the momentum conjugated to $\widetilde{g}$ is equal to \begin{equation}\label{Mabc} \widetilde{M}^{prs} = \frac{\partial \mathcal{L}_{bulk}}{\partial \partial_p \widetilde{g}_{rs}} = \frac{\sqrt{-\widetilde{g}}}{16\pi} \hzav{ \widetilde{\Gamma}^p_{ab} \kzav{ \widetilde{g}^{ar} \widetilde{g}^{bs} - \half{1} \widetilde{g}^{ab} \widetilde{g}^{rs} } + \half{1} \widetilde{\Gamma}^a_{ak} \kzav{ \widetilde{g}^{rs} \widetilde{g}^{kp} - \widetilde{g}^{ps} \widetilde{g}^{kr} - \widetilde{g}^{pr} \widetilde{g}^{ks} } } \:. \end{equation} However it was shown in \cite{Horava:1990ba,Parattu:2013gwa,Kluson:2020tzn} that the covariant canonical formalism of general relativity is better formulated when we introduce coordinate $\widetilde{f^{ab}}$ that is related to $\widetilde{g}$ by following relation \begin{equation}\label{deffab} \widetilde{f}^{ab} = \sqrt{-\widetilde{g}} \widetilde{g}^{ab} \ , \end{equation} where, following \cite{Kluson:2020tzn}, we define $\tilde{f}_{ab}$ as inverse to $\tilde{f}^{ab}$ \footnote{This definition is different from the one used in \cite{Parattu:2013gwa}, where the new coordinate was defined so $f^{ab} f_{cb} = -f\delta^a_c$.} \begin{equation} \tilde{f}_{ab}\tilde{f}^{bc}=\delta_a^c \ . \end{equation} Note that from (\ref{deffab}) we also get useful result \begin{equation} \tilde{f} \equiv \det \tilde{f}^{ab}=\widetilde{g} \ . \end{equation} The momentum $\widetilde{N}^c_{ab}$ conjugated to $\tilde{f}^{ab}$ can be obtained directly when $\tilde{f}^{ab}$ is substituted into Lagrangian or by using formulae \cite{Kluson:2020tzn} \begin{equation} \widetilde{N}^c_{ab} = -\widetilde{M}^{cmn} \inv{\sqrt{-\widetilde{f}}} \widetilde{B}_{mn\:ab} \:, \qquad \widetilde{B}_{mn\:ab} = \half{1} \kzav{ \widetilde{g}_{ma} \widetilde{g}_{bn} + \widetilde{g}_{mb} \widetilde{g}_{an} - \widetilde{g}_{mn} \widetilde{g}_{ab} } \:. \end{equation} Using explicit form for $\widetilde{M}^{abc}$ given in (\ref{Mabc}) we obtain the well known result \begin{equation} \label{Ncab} \widetilde{N}^c_{ab} = \inv{16\pi} \hzav{ -\widetilde{\Gamma}^c_{ab} + \half{1} \kzav{ \widetilde{\Gamma}^k_{ak} \delta^c_b + \widetilde{\Gamma}^k_{kb} \delta^c_a } } \ . \end{equation} From (\ref{mLbulk}) we also find momentum conjugate to $\widetilde{\phi}$ \begin{equation} \widetilde{p}^{a} = \frac{\partial \mathcal{L}_{bulk}}{\partial \partial_a \widetilde{\phi}} = -\inv{8\pi} \widetilde{f}^{ab} \partial_b \widetilde{\phi} \:. \end{equation} The Hamiltonian is then computed using Legendre transformation as \begin{eqnarray} &&\mathcal{H}^E = \partial_c \widetilde{f}^{ab} \widetilde{N}^c_{ab} + \partial_a \widetilde{\phi} \widetilde{p}^a - \mathcal{L}_{bulk} =\nonumber \\ &&= \frac{1}{16\pi} \hzav{ \widetilde{f}^{ab} \kzav{ \widetilde{\Gamma}^c_{db}\widetilde{\Gamma}^d_{ca} - \widetilde{\Gamma}^c_{ab}\widetilde{\Gamma}^d_{dc} } + \sqrt{-\widetilde{f}} V -\widetilde{f}^{ab} \partial_a \widetilde{\phi} \partial_b \widetilde{\phi} } \ . \nonumber \\ \end{eqnarray} Finally we should express this Hamiltonian in terms of canonical variables which can be done using the relations \begin{equation} \label{inverseMomenta} \widetilde{\Gamma}^i_{jk} = 16\pi \hzav{ -\widetilde{N}^i_{jk} +\inv{3}\kzav{ \widetilde{N}^u_{ju} \delta^i_k +\widetilde{N}^u_{ku} \delta^i_j } } \:, \quad \partial_a \widetilde{\phi} = -8\pi \widetilde{f}_{ab} \widetilde{p}^b \:. \end{equation} Then the final form of covariant Hamiltonian for Einstein frame of $F(R)$ gravity, is expressed as \begin{equation} \label{EFHam} \mathcal{H}^E = 16\pi\widetilde{f}^{ab} \kzav{ \widetilde{N}^c_{bd} \widetilde{N}^d_{ac} -\inv{3} \widetilde{N}^c_{ac} \widetilde{N}^d_{bd} } + \frac{\sqrt{-\widetilde{f}}}{16\pi} V - 4\pi \widetilde{f}_{ab} \widetilde{p}^a \widetilde{p}^b \:. \end{equation} \section{Relationship with Jordan Frame Hamiltonian}\label{third} Now we proceed to the main part of this paper which is the relationship between covariant Hamiltonians in Einstein and Jordan frames respectively. Our work is motivated by an interesting paper \cite{Deruelle:2009pu} where the Hamiltonian for $F(R)$ gravity in $3+1$ formalism was analysed and it was shown there that they are related by canonical transformations. Then it is very interesting question whether such canonical transformation exists in the case of the covariant canonical formalism too. \subsection{Lagrange Brackets} If there is a canonical transformation between the two frames, the fundamental Lagrange brackets shall be preserved \cite{Struckmeier:2008zz}. Contrary to the conventional Hamiltonian theory, it is not good to use Poisson brackets for this purpose, because fundamental Poisson brackets are preserved only \cite{Struckmeier:2008zz} when the transformed momenta do not depend on original coordinates or on original momenta. This condition is quite strong and, as it will be presented in this section, does not hold for our system. The Poisson brackets will be noted shortly in section \ref{PoissonBrackets}. So the calculation of Lagrange brackets of Jordan frame variables in Einstein frame can reveal us the existence of a canonical transformation. In order to calculate the Lagrange brackets we need relations between Jordan frame variables and Einstein frame ones. We are already able to express the transformation rules for coordinates \begin{equation} \label{coordTransformation} \widetilde{f}^{ab} = F' f^{ab} \:, \qquad \widetilde{\phi} = \sqrt{\frac{3}{2}} \ln F' \:. \end{equation} In order to find transformation of momenta $\tilde{N}^c_{ab}$, we need to take \rf{Ncab} and \rf{tGamma} and compare it with the the Jordan frame momenta \cite{Kluson:2020tzn} \begin{equation} N^c_{ab} = \frac{F'}{16\pi}\hzav{ -\Gamma^c_{ab} + \half{1} \kzav{\Gamma^k_{ak}\delta^c_b + \Gamma^k_{bk} \delta^c_a} + \half{1} \kzav{ \delta^c_b \partial_a \ln F' + \delta^c_a \partial_b \ln F' + f^{gc} f_{ab}\partial_g \ln F' } } \:, \end{equation} resulting in transformation relation \begin{equation} \label{tNinJordan} \widetilde{N}^c_{ab} = \inv{F'} N^c_{ab} \:, \end{equation} that also implies following important relation \begin{equation} \label{fN=tftN} \tilde{f}^{ab}\tilde{N}^c_{ab} = f^{ab}N^c_{ab}\:. \end{equation} From the relation \rf{tNinJordan}, we see that the transformed momentum depends on original momentum as well as original coordinate, so the condition for canonical invariance of Poisson brackets is not met. As the next step, we proceed to the transformation of momentum $\tilde{p}^a$. First of all, we use the second relation in \rf{inverseMomenta} where we insert \rf{coordTransformation} so that we find relation between $\partial_a B$ and $\tilde{p}^a$. Further, we use the relation between $\partial_a B$ and $p^a$ that was derived in \cite{Kluson:2020tzn} \begin{equation} \partial_g B = \frac{16\pi}{3F''} f_{gc} \kzav{f^{ik} N^c_{ik} - \frac{F'}{F''} p^c} \ . \end{equation} If we combine these relations together we find desired relation \begin{equation} \label{tpinJordan} \widetilde{p}^a =\sqrt{\frac{2}{3}}\kzav{ \frac{F'}{F''}p^a - f^{bc} N^a_{bc} } \:. \end{equation} Having found the transformation relations, the brackets follow simply as \begin{align} &\left\{f^{ab},\:f^{cd}\right\}^j = \pd{\widetilde{f}^{ik}}{f^{ab}}\pd{\widetilde{N}^j_{ik}}{f^{cd}} -\pd{\widetilde{N}^j_{ik}}{f^{ab}}\pd{\widetilde{f}^{ik}}{f^{cd}} +\pd{\widetilde{\phi}}{f^{ab}}\pd{\widetilde{p^j}}{f^{cd}} -\pd{\widetilde{p^j}}{f^{ab}}\pd{\widetilde{\phi}}{f^{cd}} = 0 \:, \nonumber \\ &\left\{f^{ab},\:B\right\}^j = \pd{\widetilde{f}^{ik}}{f^{ab}}\pd{\widetilde{N}^j_{ik}}{B} -\pd{\widetilde{N}^j_{ik}}{f^{ab}}\pd{\widetilde{f}^{ik}}{B} +\pd{\widetilde{\phi}}{f^{ab}}\pd{\widetilde{p^j}}{B} -\pd{\widetilde{p^j}}{f^{ab}}\pd{\widetilde{\phi}}{B} = 0 \:,\nonumber \\ &\left\{B,\:B\right\}^j = 0 \text{ by definition} \:, \nonumber \\ &\left\{f^{ab},\:N^{c}_{de}\right\}^j = \pd{\widetilde{f}^{ik}}{f^{ab}}\pd{\widetilde{N}^j_{ik}}{N^{c}_{de}} -\pd{\widetilde{N}^j_{ik}}{f^{ab}}\pd{\widetilde{f}^{ik}}{N^{c}_{de}} +\pd{\widetilde{\phi}}{f^{ab}}\pd{\widetilde{p^j}}{N^{c}_{de}} -\pd{\widetilde{p^j}}{f^{ab}}\pd{\widetilde{\phi}}{N^{c}_{de}} = \delta^j_c \delta^{de}_{ab} \:, \nonumber \\ &\left\{f^{ab},\:p^{c}\right\}^j = \pd{\widetilde{f}^{ik}}{f^{ab}}\pd{\widetilde{N}^j_{ik}}{p^{c}} -\pd{\widetilde{N}^j_{ik}}{f^{ab}}\pd{\widetilde{f}^{ik}}{p^{c}} +\pd{\widetilde{\phi}}{f^{ab}}\pd{\widetilde{p^j}}{p^{c}} -\pd{\widetilde{p^j}}{f^{ab}}\pd{\widetilde{\phi}}{p^{c}} = 0 \:, \nonumber \\ &\left\{B,\:N^{c}_{de}\right\}^j = \pd{\widetilde{f}^{ik}}{B}\pd{\widetilde{N}^j_{ik}}{N^{c}_{de}} -\pd{\widetilde{N}^j_{ik}}{B}\pd{\widetilde{f}^{ik}}{N^{c}_{de}} +\pd{\widetilde{\phi}}{B}\pd{\widetilde{p^j}}{N^{c}_{de}} -\pd{\widetilde{p^j}}{B}\pd{\widetilde{\phi}}{N^{c}_{de}} = 0 \:, \nonumber \\ &\left\{B,\:p^{c}\right\}^j = \pd{\widetilde{f}^{ik}}{B}\pd{\widetilde{N}^j_{ik}}{p^{c}} -\pd{\widetilde{N}^j_{ik}}{B}\pd{\widetilde{f}^{ik}}{p^{c}} +\pd{\widetilde{\phi}}{B}\pd{\widetilde{p^j}}{p^{c}} -\pd{\widetilde{p^j}}{B}\pd{\widetilde{\phi}}{p^{c}} = \delta^j_c \:, \nonumber \\ &\left\{N^c_{ab},\:N^{d}_{ef}\right\}^j = \pd{\widetilde{f}^{ik}}{N^c_{ab}}\pd{\widetilde{N}^j_{ik}}{N^{d}_{ef}} -\pd{\widetilde{N}^j_{ik}}{N^c_{ab}}\pd{\widetilde{f}^{ik}}{N^{d}_{ef}} +\pd{\widetilde{\phi}}{N^c_{ab}}\pd{\widetilde{p^j}}{N^{d}_{ef}} -\pd{\widetilde{p^j}}{N^c_{ab}}\pd{\widetilde{\phi}}{N^{d}_{ef}} = 0 \:, \nonumber \\ &\left\{N^c_{ab},\:p^{d}\right\}^j = \pd{\widetilde{f}^{ik}}{N^c_{ab}}\pd{\widetilde{N}^j_{ik}}{p^{d}} -\pd{\widetilde{N}^j_{ik}}{N^c_{ab}}\pd{\widetilde{f}^{ik}}{p^{d}} +\pd{\widetilde{\phi}}{N^c_{ab}}\pd{\widetilde{p^j}}{p^{d}} -\pd{\widetilde{p^j}}{N^c_{ab}}\pd{\widetilde{\phi}}{p^{d}} = 0 \:, \nonumber \\ &\left\{p^{a},\:p^{b}\right\}^j = \pd{\widetilde{f}^{ik}}{p^{a}}\pd{\widetilde{N}^j_{ik}}{p^{b}} -\pd{\widetilde{N}^j_{ik}}{p^{a}}\pd{\widetilde{f}^{ik}}{p^{b}} +\pd{\widetilde{\phi}}{p^{a}}\pd{\widetilde{p^j}}{p^{b}} -\pd{\widetilde{p^j}}{p^{a}}\pd{\widetilde{\phi}}{p^{b}} = 0 \:, \nonumber \\ \end{align} where we have used abbreviation $\delta^{ab}_{cd} = \half{1} \kzav{\delta^a_c \delta^b_d + \delta^b_c \delta^a_d}$. From the equations above, it is easily visible that the fundamental Lagrange brackets are preserved, so there is a canonical transformation between Jordan and Einstein frames. \subsection{Generating Function of Canonical Transformation} Having proven existence of canonical transformation, its generating function is looked for. But first, let us review basic facts about canonical transformations in covariant formalism, following \cite{Struckmeier:2008zz, Struckmeier:2012de}. Let us consider covariant formulation of $F(R)$ gravity in Jordan frame with the canonical variables $f^{ab},N_{ab}^c,B,p^c_B$. On the other hand in case of covariant formulation of Einstein-frame $F(R)$-gravity the canonical variables are $\tilde{f}^{ab},\tilde{N}_{ab}^c,\tilde{\phi},\tilde{p}^a$. We demand that they give the same description of the physical systems so that we have a requirement \begin{eqnarray} \delta \int d^4x (N^c_{ab}\partial_c f^{ab}+p^c\partial_c B-\mathcal{H}^J)= \delta\int d^4x (\tilde{N}^c_{ab}\partial_c \tilde{f}^{ab}+\tilde{p}^c\partial_c \tilde{\phi}-\mathcal{H}^E) \ . \end{eqnarray} This result implies that the integrals can differ only by divergence of a vector function whose variation vanishes on the boundary $\partial R$ of the integration region $R$ \begin{equation} \label{CanonicalSurfaceCondition} \delta \int_R d^4x \partial_a G_1^a(f,B,\tilde{f},\tilde{\phi})= \delta \oint_{\partial R}G_1^a(f,B,\tilde{f},\tilde{\phi}) dS_a=0 \: . \end{equation} Using $G_1^a$, we can write \begin{equation} \label{eqLagrangians} N^c_{ab}\partial_c f^{ab}+p^c\partial_c B-\mathcal{H}^J= \tilde{N}^c_{ab}\partial_c \tilde{f}^{ab}+\tilde{p}^c\partial_c \tilde{\phi}-\mathcal{H}^E+ \partial_a G_1^a(f,B,\tilde{f},\tilde{\phi},x) \:. \end{equation} The divergence of $G_1^a$ can be expressed using coordinates as \begin{equation} \partial_a G^a_1= \frac{\partial G_1^a}{\partial f^{bc}}\partial_a f^{bc}+ \frac{\partial G_1^a}{\partial \tilde{f}^{bc}}\partial_a \tilde{f}^{bc}+ \frac{\partial G_1^a}{\partial B}\partial_a B+ \frac{\partial G_1^a}{\partial \tilde{\phi}}\partial_a \tilde{\phi}+ \frac{\partial G_1^a}{\partial x^a}|_{expl} \:. \end{equation} Putting this into \rf{eqLagrangians} and comparing terms proportional to $\partial_c f^{ab}$ and $\partial_c \tilde{f}^{ab}$ and $\partial_a B$ and $\partial_a \tilde{\phi}$ we obtain \begin{align} N^c_{ab} &=\frac{\partial G_1^c}{\partial f^{ab}} \ , \quad \quad p^c =\frac{\partial G_1^c}{\partial B} \ , \nonumber \\ \tilde{N}_{ab}^c &=-\frac{\partial G_1^c}{\partial \tilde{f}^{ab}} \ , \quad \: \tilde{p}^c =-\frac{\partial G_1^a}{\partial \tilde{\phi}} \ , \nonumber \\ \mathcal{H}^J&=\mathcal{H}^E-\frac{\partial G_1^a}{\partial x^a}|_{expl} \:. \end{align} On the other hand we can consider different type of generating function \begin{equation} \label{G2Def} G_1^a = G_2^a(f,B,\tilde{N},\tilde{p})-\tilde{f}^{ik}\tilde{N}_{ik}^a- \tilde{\phi} \tilde{p}^a \:, \end{equation} so that its total divergence is equal to \begin{eqnarray} && \partial_c (G^c_2 -\tilde{f}^{ab}\tilde{N}_{ab}^c- \tilde{\phi} \tilde{p}^c)=\nonumber \\ && \frac{\partial G^c_2}{\partial f^{ab}} \partial_c f^{ab}+ \frac{\partial G^c_2}{\partial B}\partial_c B+ \frac{\partial G^c_2}{\partial \tilde{N}_{ab}^d} \partial_c \tilde{N}_{ab}^d+ \frac{\partial G^c_2}{\partial \tilde{p}^a} \partial_ c\tilde{p}^a+ \frac{\partial G^c_2}{\partial x^c}|_{exp}-\nonumber \\ && -\tilde{N}_{ab}^c\partial_c \tilde{f}^{ab} -\tilde{f}^{ab}\partial_c \tilde{N}_{ab}^c-\tilde{p}^c\partial_c\tilde{\phi} - \tilde{\phi} \partial_c \tilde{p}^c \:.\nonumber \\ \end{eqnarray} Substituting the second type of generating function this into \rf{eqLagrangians} and then comparing the related terms yields a set of different equations \begin{eqnarray} \label{G2conditions} &&\mathcal{H}_J=\mathcal{H}_E-\frac{\partial G_2^c}{\partial x^c}|_{expl} \ , \nonumber \\ &&\frac{\partial G_2^c}{\partial \tilde{p}^a}=\delta^c_a \tilde{\phi} \ , \quad \frac{\partial G_2^c}{\partial \tilde{N}^d_{ab}}=\tilde{f}^{ab}\delta_d^c \ , \quad N^c_{ab}=\frac{\partial G_2^c}{f^{ab}} \ , \quad p^c=\frac{\partial G_2^c}{\partial B} \ . \nonumber \\ \end{eqnarray} After this brief discussion of the canonical formalism, we proceed to the analysis of the question how Einstein and Jordan frame formulations of $F(R)$ gravity are related by canonical transformations. Recall that the Covariant Hamiltonian in Jordan frame has the form \cite{Kluson:2020tzn} \begin{align} \label{HJ} \mathcal{H}^J =& \frac{16\pi}{F'} f^{ab} \kzav{ N^c_{bd} N^d_{ac} -\inv{3} N^c_{ac} N^d_{bd} } \nonumber \\ &-\frac{8\pi}{3F'} f_{ab} \kzav{\frac{F'}{F''}p^a - N^a_{cd} f^{cd}} \kzav{\frac{F'}{F''}p^b - N^b_{ef} f^{ef}} \nonumber \\ &-\frac{\sqrt{-f}}{16\pi}\kzav{F-F'B} \ . \end{align} Using the transformations \rf{coordTransformation}, \rf{tNinJordan}, \rf{tpinJordan}, we can express the Einstein frame Hamiltonian in Jordan frame coordinates \begin{align} \mathcal{H}^{E*} =& \frac{16\pi}{F'} f^{ab} \kzav{ N^c_{bd} N^d_{ac} -\inv{3} N^c_{ac} N^d_{bd} } \nonumber \\ &-\frac{8\pi}{3F'} f_{ab} \kzav{\frac{F'}{F''}p^a - N^a_{cd} f^{cd}} \kzav{\frac{F'}{F''}p^b - N^b_{ef} f^{ef}}\nonumber \\ &-\frac{\sqrt{-f}}{16\pi}\kzav{F-F'B} \:, \end{align} which agrees with \rf{HJ}. The fact that these two Hamiltonians are equal means that the generating function of canonical transformation is independent on coordinates. In fact, transformations \rf{coordTransformation}, \rf{tNinJordan} and \rf{tpinJordan} suggest that the generating function has the form of \begin{equation} \label{GeneratingFunctionG2} G_2^c=F'\tilde{N}^c_{ab}f^{ab}+\sqrt{\frac{3}{2}}\ln F' \tilde{p}^c \:. \end{equation} In order to verify that the suggested form of $G_2^c$ is really the generating function of canonical transformations let us calculate the following derivative \begin{eqnarray} \frac{\partial G_2^c}{\partial B}= F'' \tilde{N}_{ab}^c f^{ab} +\sqrt{\frac{3}{2}}\frac{F''}{F'}\tilde{p}^c = \frac{F''}{F'} N_{ab}^c f^{ab} +\frac{F''}{F'}\kzav{\frac{F'}{F''} p^c - N_{ab}^c f^{ab} } = p^c \end{eqnarray} that agrees with the the fourth equation in \rf{G2conditions}. Further, the derivative $G_2^c$ with respect to $f^{ab}$ gives \begin{equation} N^c_{ab}=\frac{\partial G_2^c}{\partial f^{ab}}= F'\widetilde{N}^c_{ab} \end{equation} that gives the relation (\ref{tNinJordan}). Finally, the derivative of $G_2^c$ with respect to $\tilde{p}^a$ and $\widetilde{N}^d_{ab}$ lead to \begin{eqnarray} &&\frac{\partial G_2^c}{\partial \tilde{p}^a}=\sqrt{\frac{3}{2}}\ln F'\delta^c_a=\delta^c_a\widetilde{\phi} \ , \nonumber \\ && \frac{\partial G_2^c}{\partial \widetilde{N}^d_{ab}}= F'\delta^c_d f^{ab}=\widetilde{f}^{ab}\delta^c_d \nonumber \\ \end{eqnarray} that lead to relations in (\ref{coordTransformation}). In summary, we have shown that \rf{GeneratingFunctionG2} is generating function of canonical transformation between Jordan and Einstein frame. This shows that these two frames are related by canonical transformations even in the case of covariant canonical formulations of these two theories which is new and non-trivial result. \subsection{Remark About Poisson Brackets}\label{PoissonBrackets} In Covariant Hamiltonian theory, the fundamental Poisson brackets are not always preserved under canonical transformation \cite{Struckmeier:2008zz}. Their purpose of canonical invariant is fulfilled with Lagrange brackets. This section summarizes Poisson brackets just for the reference. We will calculate Poisson brackets of Einstein frame variables in Jordan frame. Since no Einstein frame coordinate depends on any of Jordan frame momenta \rf{coordTransformation}, the Poisson brackets of coordinates are zero. For the mixed brackets, one obtains \begin{align} \hzav{\widetilde{f}^{ab} \:,\: \widetilde{N}^c_{de}}_j &= \frac{\partial \widetilde{f}^{ab}}{\partial f^{ik}}\frac{\partial \widetilde{N}^c_{de}}{\partial N^j_{ik}} + \frac{\partial \widetilde{f}^{ab}}{\partial B}\frac{\partial \widetilde{N}^c_{de}}{\partial p^j} - \frac{\partial \widetilde{N}^c_{de}}{\partial f^{ik}}\frac{\partial \widetilde{f}^{ab}}{\partial N^j_{ik}} - \frac{\partial \widetilde{N}^c_{de}}{\partial B }\frac{\partial \widetilde{f}^{ab}}{\partial p^j} = \delta^c_j \delta^{ab}_{cd} \:, \nonumber \\ \hzav{\widetilde{f}^{ab} \:,\: \widetilde{p}^c}_j &= \frac{\partial \widetilde{f}^{ab}}{\partial f^{ik}}\frac{\partial \widetilde{p}^c}{\partial N^j_{ik}} + \frac{\partial \widetilde{f}^{ab}}{\partial B}\frac{\partial \widetilde{p}^c}{\partial p^j} - \frac{\partial \widetilde{p}^c}{\partial f^{ik}}\frac{\partial \widetilde{f}^{ab}}{\partial N^j_{ik}} - \frac{\partial \widetilde{p}^c}{\partial B }\frac{\partial \widetilde{f}^{ab}}{\partial p^j} = 0 \:, \nonumber \\ \hzav{\widetilde{\phi} \:,\: \widetilde{N}^c_{de}}_j &= \frac{\partial \widetilde{\phi}}{\partial f^{ik}}\frac{\partial \widetilde{N}^c_{de}}{\partial N^j_{ik}} + \frac{\partial \widetilde{\phi}}{\partial B}\frac{\partial \widetilde{N}^c_{de}}{\partial p^j} - \frac{\partial \widetilde{N}^c_{de}}{\partial f^{ik}}\frac{\partial \widetilde{\phi}}{\partial N^j_{ik}} - \frac{\partial \widetilde{N}^c_{de}}{\partial B }\frac{\partial \widetilde{\phi}}{\partial p^j} = 0 \:, \nonumber \\ \hzav{\widetilde{\phi} \:,\: \widetilde{p}^c}_j &= \frac{\partial \widetilde{\phi}}{\partial f^{ik}}\frac{\partial \widetilde{p}^c}{\partial N^j_{ik}} + \frac{\partial \widetilde{\phi}}{\partial B}\frac{\partial \widetilde{p}^c}{\partial p^j} - \frac{\partial \widetilde{p}^c}{\partial f^{ik}}\frac{\partial \widetilde{\phi}}{\partial N^j_{ik}} - \frac{\partial \widetilde{p}^c}{\partial B }\frac{\partial \widetilde{\phi}}{\partial p^j} = \delta^c_j \:. \nonumber \\ \end{align} As we can see all mixed brackets yield the expected results. The brackets of momenta are those who break canonical invariance of Poisson brackets \begin{align} &\hzav{\widetilde{N}^f_{ab} \:,\: \widetilde{N}^c_{de}}_j = \frac{\partial \widetilde{N}^f_{ab}}{\partial f^{ik}}\frac{\partial \widetilde{N}^c_{de}}{\partial N^j_{ik}} + \frac{\partial \widetilde{N}^f_{ab}}{\partial B}\frac{\partial \widetilde{N}^c_{de}}{\partial p^j} - \frac{\partial \widetilde{N}^c_{de}}{\partial f^{ik}}\frac{\partial \widetilde{N}^f_{ab}}{\partial N^j_{ik}} - \frac{\partial \widetilde{N}^c_{de}}{\partial B }\frac{\partial \widetilde{N}^f_{ab}}{\partial p^j} = 0 \:, \nonumber \\ &\hzav{\widetilde{N}^d_{ab} \:,\: \widetilde{p}^c}_j = \frac{\partial \widetilde{N}^d_{ab}}{\partial f^{ik}}\frac{\partial \widetilde{p}^c}{\partial N^j_{ik}} + \frac{\partial \widetilde{N}^d_{ab}}{\partial B}\frac{\partial \widetilde{p}^c}{\partial p^j} - \frac{\partial \widetilde{p}^c}{\partial f^{ik}}\frac{\partial \widetilde{N}^d_{ab}}{\partial N^j_{ik}} - \frac{\partial \widetilde{p}^c}{\partial B }\frac{\partial \widetilde{N}^d_{ab}}{\partial p^j} = \nonumber \\ &\qquad = \sqrt{\frac{2}{3}}\inv{F'}\kzav{\delta^d_j N^c_{ab} - \delta^c_j N^d_{ab}} \:, \nonumber \\ &\hzav{\widetilde{p}^a \:,\: \widetilde{p}^c}_j = \frac{\partial \widetilde{p}^a}{\partial f^{ik}}\frac{\partial \widetilde{p}^c}{\partial N^j_{ik}} + \frac{\partial \widetilde{p}^a}{\partial B}\frac{\partial \widetilde{p}^c}{\partial p^j} - \frac{\partial \widetilde{p}^c}{\partial f^{ik}}\frac{\partial \widetilde{p}^a}{\partial N^j_{ik}} - \frac{\partial \widetilde{p}^c}{\partial B }\frac{\partial \widetilde{p}^a}{\partial p^j} = \nonumber \\ &\qquad= \frac{2}{3}\delta^c_j \kzav{ f^{ik} N^a_{ik} + \frac{F'(F''^2 - F'F''')}{F''^3} p^a } - \frac{2}{3}\delta^a_j \kzav{ f^{ik} N^c_{ik} + \frac{F'(F''^2 - F'F''')}{F''^3} p^c } \:. \nonumber \\ \end{align} Contrary to the conventional Hamiltonian theory, the canonical transformation does not preserve fundamental Poisson brackets for momenta. \section{Surface Lagrangian and Thermodynamic Properties}\label{fourth} It is also important to mention the so-called surface part of the Lagrangian. This is the part which can be expressed as a derivative, or alternatively said as a divergence of some vector potential, and usually is neglected since it does not contribute to the equations of motion. However, this is not true just for any surface and more importantly, this part of Lagrangian contains information about thermodynamic properties of the boundary region, typically a horizon. Previously, we have found the surface Lagrangian as \rf{Lagrangians} \begin{equation} \mathcal{L}_{sur} = \partial_i \hzav{\frac{\sqrt{-\widetilde{g}}}{16\pi} \kzav{\widetilde{g}^{ab}\widetilde{\Gamma}^i_{ab} - \widetilde{g}^{ai}\widetilde{\Gamma}^b_{ab} + \sqrt{6} \widetilde{\partial}^i \widetilde{\phi} } } \ . \end{equation} It is natural to express this surface term using canonical coordinates when we use the relations \rf{inverseMomenta} and we get \begin{equation} \label{L_sur_Einstein} \mathcal{L}_{sur} = -\partial_i \kzav{ \widetilde{f}^{ab} \widetilde{N}^i_{ab} + \sqrt{\half{3}} \widetilde{p}^i } \:. \end{equation} If we used coordinate transformation and put it back into the Jordan Frame, we would obtain \begin{equation} \label{L_sur_Jordan} \mathcal{L}_{sur} = - \partial_i \kzav{ \frac{F'}{F''} p^i } \:, \end{equation} which is the same result, as was derived for Jordan Frame formulation. Thus, the thermodynamic properties can be assumed to be the same as in Jordan case \cite{Matous:2021uqh}, which is the expected result. In General Relativity, there is a relation between the two Lagrangians (bulk and surface) \cite{Mukhopadhyay:2006vu} \begin{equation} \mathcal{L}_{sur} = - \partial_c \kzav{ g_{ab} \pd{\mathcal{L}_{bulk}}{\partial_c g_{ab}} } \:, \end{equation} which is better for our purposes to be written in $f-N$ formalism \cite{Parattu:2013gwa} \begin{equation}\label{mL_sur} \mathcal{L}_{sur} = - \partial_c \kzav{ f^{ab} N^c_{ab} } \:, \end{equation} as visible from surface Lagrangian of Jordan \rf{L_sur_Jordan} as well as Einstein \rf{L_sur_Einstein} frame, this relation does not hold for $F(R)$ gravity theory. One could argue that \rf{mL_sur} is a kind of first approximation of more general formula, however its form is non-trivial and so it remains an open question. {\bf Acknowledgement:} \\ The work of J. Kluso\v{n} is supported by the grant “Integrable Deformations” (GA20-04800S) from the Czech Science Foundation (GACR).

cf5dba109c268b667c924361ee9bc92a3a1fc699

\section{Introduction} For integers $s$, $u$, $t$, and $v$, with $u \ge s \ge 1$ and $v \ge t \ge 1$, a matroid~$M$ has the \emph{$(s,u,t,v)$-property} if every $s$-element subset of $E(M)$ is contained in a circuit of size~$u$, and every $t$-element subset of $E(M)$ is contained in a cocircuit of size~$v$. Matroids with this property appear regularly in the matroid theory literature: for example, wheels and whirls have the $(1,3,1,3)$-property, and (tipless) spikes have the $(2,4,2,4)$-property. Note that $M$ has the $(s,u,t,v)$-property if and only if $M^*$ has the $(t,v,s,u)$-property. Brettell, Campbell, Chun, Grace, and Whittle~\cite{bccgw2019} studied such matroids, and showed that if $u<2s$ or $v<2t$, then there are only finitely many matroids with the $(s,u,t,v)$-property~\cite[Theorem 3.3]{bccgw2019}. On the other hand, in the case that $s=t$ and $u=v=2t$, any sufficiently large matroid with the $(s,u,t,v)$-property is a member of a class of structured matroids referred to as \emph{$t$-spikes}. In particular, when $t=2$, this is the class typically known simply as \emph{(tipless) spikes}. Our focus in this paper is also on the case where $u=2s$ and $v=2t$, but we drop the requirement that $s=t$. For positive integers $s$ and $t$, an \emph{$(s,t)$-spike} is a matroid on at least $2\max\{s,t\}$ elements whose ground set has a partition $(S_1,S_2,\ldots,S_n)$ into pairs such that the union of every set of $s$ pairs is a circuit and the union of every set of $t$ pairs is a cocircuit. The following is our main result: \begin{theorem} \label{mainthm} There exists a function $f : \mathbb{N}^2 \rightarrow \mathbb{N}$ such that, if $M$ is a matroid with the $(s,2s,t,2t)$-property and $|E(M)| \ge f(s,t)$, then $M$ is an $(s,t)$-spike. \end{theorem} \noindent This proves the conjecture of Brettell et al.~\cite[Conjecture~1.2]{bccgw2019}. Our approach is essentially the same as in \cite{bccgw2019}, but some care is required to generalize the argument. We note also that \cref{modcut} corrects an erroneous lemma \cite[Lemma 6.6]{bccgw2019}. This paper is one in a developing series on matroids with the $(s,u,t,v)$-property. First, Miller~\cite{miller2014} studied matroids with the $(2,4,2,4)$-property, proving the specialization of \cref{mainthm} to the case where $s=t=2$. As previously mentioned, Brettell et al.~\cite{bccgw2019} considered the more general case where $s=t$ and $u=v=2t$, for any $t \ge 1$. Oxley, Pfeil, Semple, and Whittle considered the case where $s=2$, $u=4$, $t=1$, and $v \in \{3,4\}$, showing that a sufficiently large $v$-connected matroid with the $(2,4,1,v)$-property is isomorphic to $M(K_{v,n})$ for some $n$~\cite{pfeil}. A ``cyclic'' analogue of the $(s,u,t,v)$-property has also been considered, where a cyclic ordering $\sigma$ is imposed on $E(M)$, and only sets that appear consecutively with respect to $\sigma$ and have size~$s$ (or size~$t$) need appear in a circuit of size $u$ (or a cocircuit of size $v$, respectively). The case where $s = u-1$ and $t = v-1$ and $s=t$ was considered by Brettell, Chun, Fife, and Semple~\cite{bcfs2019}; whereas Brettell, Semple, and Toft dropped the requirement that $s=t$~\cite{bst2022}. This series of papers has been motivated by problems involving matroid connectivity. The well-known Wheels-and-Whirls Theorem of Tutte~\cite{tutte1966} states that wheels and whirls (which have the $(1,3,1,3)$-property) are the only $3$-connected matroids with no elements that can be either deleted or contracted to retain a $3$-connected matroid. Similarly, spikes (which have the $(2,4,2,4)$-property) are the only $3$-connected matroids on at least $13$ elements that have no triangles, no triads, and no pairs of elements that can be either deleted or contracted to preserve $3$-connectivity~\cite{williams2015}. The following conjecture was stated as \cite[Conjecture 1.3]{bccgw2019}. The case where $t=2$ was proved by Williams~\cite{williams2015}. \begin{conjecture} \label{conj:old} There exists a function $f : \mathbb{N} \rightarrow \mathbb{N}$ such that if $M$ is a $(2t-1)$-connected matroid with no circuits or cocircuits of size $2t-1$, and $|E(M)| \ge f(t)$, then either \begin{enumerate} \item there exists a $t$-element set $X \subseteq E(M)$ such that either $M/X$ or $M \backslash X$ is $(t+1)$-connected, or \item $M$ is a $(t,t)$-spike. \end{enumerate} \end{conjecture} Indeed, sufficiently large $(t,t)$-spikes are $(2t-1)$-connected matroids~\cite[Lemma~6.5]{bccgw2019}, they have no circuits or cocircuits of size $(2t-1)$~\cite[Lemma~6.3]{bccgw2019}, and for every $t$-element subset $X \subseteq E(M)$, neither $M/X$ nor $M \backslash X$ is $(t+1)$-connected. Optimistically, we offer the following generalization of \cref{conj:old}. \begin{conjecture} \label{conj:new} There exists a function $f : \mathbb{N}^2 \rightarrow \mathbb{N}$ such that if $M$ is a matroid with no circuits of size at most $2s-1$, no cocircuits of size at most $2t-1$, the matroid $M$ is $(2\min\{s,t\}-1)$-connected, and $|E(M)| \ge f(s,t)$, then either \begin{enumerate} \item there exists an $s$-element set $X \subseteq E(M)$ such that $M/X$ is $(s+1)$-connected, \item there exists a $t$-element set $X \subseteq E(M)$ such that $M \backslash X$ is $(t+1)$-connected, or \item $M$ is an $(s,t)$-spike. \end{enumerate} \end{conjecture} \cref{sec:Preliminaries} recalls some terminology and a Ramsey-theoretic result used later in the paper. In \cref{sec:echidnas}, we recall the definition of echidnas from~\cite{bccgw2019} and show that every matroid with the $(s,2s,t,2t)$-property and having a sufficiently large $s$-echidna is an $(s,t)$-spike. In \cref{sec:t2t}, we prove \cref{mainthm}. Finally, \cref{sec:tspikeprops} describes some properties of $(s,t)$-spikes, as well as a construction that allows us to build an $(s,t+1)$-spike from an $(s,t)$-spike. \section{Preliminaries} \label{sec:Preliminaries} Our notation and terminology follows Oxley~\cite{oxbook}. We refer to the fact that a circuit and a cocircuit cannot intersect in exactly one element as ``orthogonality''. A set $S_1$ \emph{meets} a set $S_2$ if $S_1 \cap S_2 \neq \emptyset$. We denote $\{1,2,\dotsc,n\}$ by $\seq{n}$, and, for positive integers $i < j$, we denote $\{i,i+1,\dotsc,j\}$ by $[i,j]$. We denote the set of positive integers by $\mathbb{N}$. In order to prove \cref{mainthm}, we will use some hypergraph Ramsey Theory~\cite{ramsey1930}. Recall that a hypergraph is \emph{$k$-uniform} if every hyperedge has size~$k$. \begin{theorem}[Ramsey's Theorem for $k$-uniform hypergraphs] \label{hyperramsey} For positive integers $k$ and $n$, there exists an integer $r_k(n)$ such that if $H$ is a $k$-uniform hypergraph on $r_k(n)$ vertices, then $H$ has either a clique on $n$ vertices, or a stable set on $n$ vertices. \end{theorem} \section{Echidnas and \texorpdfstring{$(s,t)$}{(s,t)}-spikes} \label{sec:echidnas} Recall that $M$ is an $(s,t)$-spike if there is a partition of $E(M)$ into pairs such that the union of any $s$ pairs is a circuit and the union of any $t$ pairs is a cocircuit. In this section, we prove a sufficient condition for $M$ to be an $(s,t)$-spike. Namely, we prove as \cref{lem:swamping} that if $M$ has the $(s,2s,t,2t)$-property, and a subset of $E(M)$ can be partitioned into $u$ pairs such that the union of any $t$ pairs is a circuit, then, when $u$ is sufficiently large, $M$ is an $(s,t)$-spike. Conforming with \cite{bccgw2019}, we call such a partition a $t$-echidna, as defined below. Let $M$ be a matroid. A $t$-\emph{echidna} of order $n$ is a partition $(S_1,\ldots, S_n)$ of a subset of $E(M)$ such that \begin{enumerate} \item $|S_i|=2$ for all $i \in \seq{n}$, and \item $\bigcup_{i \in I}S_i$ is a circuit for all $I \subseteq \seq{n}$ with $|I|=t$. \end{enumerate} For $i \in \seq{n}$, we say $S_i$ is a \emph{spine}. We say $(S_1,\ldots,S_n)$ is a \emph{$t$-coechidna} of $M$ if $(S_1,\ldots,S_n)$ is a $t$-echidna of $M^*$. Let $(S_1,\dotsc,S_n)$ be a $t$-echidna of a matroid $M$. If $(S_1,\dotsc,S_m)$ is a $t$-echidna of $M$, for some $m \ge n$, we say that $(S_1,\dotsc,S_n)$ \emph{extends} to $(S_1,\dotsc,S_m)$. We say that $\pi=(S_1,\dotsc,S_n)$ is \emph{maximal} if $\pi$ extends only to $\pi$. Note that a matroid~$M$ is an $(s,t)$-spike if there exists a partition $\pi=(A_1,\ldots,A_m)$ of $E(M)$ such that $\pi$ is an $s$-echidna and a $t$-coechidna, for some $m\geq\max\{s,t\}$. In this case, we say that the $(s,t)$-spike~$M$ has \emph{order~$m$}, we call $\pi$ the \emph{associated partition} of the $(s,t)$-spike~$M$, and we say that $A_i$ is an \emph{arm} of the $(s,t)$-spike for each $i \in \seq{m}$. An $(s,t)$-spike with $s=t$ is also called a \emph{$t$-spike}. Note that if $M$ is an $(s,t)$-spike, then $M^*$ is a $(t,s)$-spike. Throughout this section, we assume that $s$ and $t$ are positive integers. \begin{lemma} \label{lem:coechidna} Let $M$ be a matroid with the $(s,2s,t,2t)$-property. If $M$ has an $s$-echidna $(S_1,\ldots, S_n)$, where $n\geq s+2t-1$, then $(S_1,\ldots, S_n)$ is also a $t$-coechidna of $M$. \end{lemma} \begin{proof} Suppose $M$ has an $s$-echidna $(S_1,\ldots, S_n)$ with $n \ge s+2t-1$, and let $S_i=\{x_i,y_i\}$ for each $i \in [n]$. We show, for every $t$-element subset $J$ of $[n]$, that $\bigcup_{j \in J} S_j$ is a cocircuit. Without loss of generality, let $J=[t]$. By the $(s,2s,t,2t)$-property, $\{x_1,\ldots,x_{t}\}$ is contained in a $2t$-element cocircuit~$C^*$. Suppose for a contradiction that $C^*\neq\bigcup_{j \in J} S_j$. Then there is some $i \in [t]$ such that $y_i\notin C^*$. Without loss of generality, say $y_1\notin C^*$. Let $I$ be an $(s-1)$-element subset of $[t+1,n]$. For any such $I$, the set $S_1 \cup \bigcup_{i \in I} S_i$ is a circuit that meets $C^*$. By orthogonality, $\bigcup_{i \in I} S_i$ meets $C^*$. Thus, $C^*$ avoids at most $s-2$ of the $S_i$'s for $i \in [t+1,n]$. In fact, as $C^*$ meets each $S_i$ with $i \in [t]$, the cocircuit~$C^*$ avoids at most $s-2$ of the $S_i$'s for $i \in [n]$. Thus $|C^*| \ge n-(s-2) \ge (s+2t-1) -(s-2) =2t+1 > 2t$, a contradiction. Therefore, we conclude that $C^*=\bigcup_{j \in J} S_j$, and the result follows. \end{proof} \sloppy \begin{lemma} \label{lem:rep-orthog} Let $M$ be a matroid with the $(s,2s,t,2t)$-property, and let $(S_1,\ldots, S_n)$ be an $s$-echidna of $M$ with $n\geq\max\{s+2t,2s+t\}-1$. \begin{itemize} \item[(i)] Let $I$ be an $(s-1)$-subset of $[n]$. For $z\in E(M)-\bigcup_{i \in I}S_i$, there is a $2s$-element circuit containing $\{z\} \cup \bigcup_{i \in I}S_i$. \item[(ii)] Let $I$ be a $(t-1)$-subset of $[n]$. For $z\in E(M)-\bigcup_{i \in I}S_i$, there is a $2t$-element cocircuit containing $\{z\} \cup \bigcup_{i \in I}S_i$. \end{itemize} \end{lemma} \fussy \begin{proof} First we prove (i). For $i \in [n]$, let $S_i=\{x_i,y_i\}$. By the $(s,2s,t,2t)$-property, there is a $2s$-element circuit~$C$ containing $\{z\} \cup \{x_i : i \in I\}$. Let $J$ be a $(t-1)$-element subset of $[n]$ such that $C$ and $\bigcup_{j \in J}S_j$ are disjoint (such a set exists since $|C|=2s$ and $n \ge 2s+t-1$). For $i \in I$, let $C^*_i=S_i \cup \bigcup_{j \in J} S_j$, and observe that $x_i \in C^*_i \cap C$, and $C^*_i \cap C \subseteq S_i$. By \cref{lem:coechidna}, $(S_1,\dotsc,S_n)$ is a $t$-coechidna as well as an $s$-echidna; therefore, $C^*_i$ is a cocircuit. Now, for each $i \in I$, orthogonality implies that $|C^*_i \cap C| \ge 2$, and hence $y_i \in C$. So $C$ contains $\{z\} \cup \bigcup_{i \in I}S_i$, as required. Now, to prove (ii), recall that $(S_1,\dotsc,S_n)$ is a $t$-coechidna by Lemma \cref{lem:coechidna}. Therefore, (ii) follows by (i) and duality. \end{proof} \begin{lemma} \label{lem:swamping} Let $M$ be a matroid with the $(s,2s,t,2t)$-property. If $M$ has an $s$-echidna $\pi=(S_1,\ldots, S_n)$, where $n\geq\max\{s+2t-1,2s+t-1,3s+t-3\}$, then $(S_1,\ldots, S_n)$ extends to a partition of $E(M)$ that is both an $s$-echidna and a $t$-coechidna. \end{lemma} \begin{proof} Let $\pi'=(S_1, \dotsc, S_m)$ be a maximal $s$-echidna with $X=\bigcup_{i = 1}^{m} S_i\subseteq E(M)$. Suppose for a contradiction that $X\neq E(M)$. Since $\pi'$ is maximal, $m\geq n\geq s+2t-1$. Therefore, by Lemma \ref{lem:coechidna}, $\pi'$ is a $t$-coechidna. Let $z\in E(M)-X$. By Lemma \ref{lem:rep-orthog}, there is a $2s$-element circuit $C = (\bigcup_{i \in [s-1]} S_i)\cup \{z,z'\}$ for some $z'\in E(M)$. % We claim that $z'\notin X$. Towards a contradiction, suppose that $z'\in S_k$ for some $k\in [s,m]$. Let $J$ be a $t$-element subset of $[s,m]$ containing $k$. Then, since $(S_1,\dotsc,S_m)$ is a $t$-coechidna, $\bigcup_{j \in J}S_j$ is a cocircuit that contains $z'$. Now, this cocircuit intersects the circuit~$C$ in a single element $z'$, contradicting orthogonality. Thus, $z'\notin X$, as claimed. We next show that $(\{z,z'\}, S_{s}, S_{s+1}, \ldots, S_m)$ is a $t$-coechidna. Since $\pi'$ is a $t$-coechidna, it suffices to show that $\{z,z'\} \cup \bigcup_{i \in I}S_i$ is a cocircuit for each $(t-1)$-element subset~$I$ of $[s,m]$. Let $I$ be such a set. \Cref{lem:rep-orthog} implies that there is a $2t$-element cocircuit~$C^*$ of $M$ containing $\{z\} \cup \bigcup_{i\in I}S_i$. By orthogonality, $|C\cap C^*|>1$. Therefore, $z'\in C^*$. Thus, $(\{z,z'\}, S_{s}, S_{s+1}, \ldots, S_m)$ is a $t$-coechidna. Since this $t$-coechidna has order $1+m-(s-1)\geq n-s+2\geq2s+t-1$, the dual of \cref{lem:coechidna} implies that $(\{z,z'\}, S_{s}, S_{s+1}, \dotsc, S_m)$ is also an $s$-echidna. Next we show that $(\{z,z'\}, S_1, S_2, \dotsc, S_m)$ is a $t$-coechidna. Let $I$ be a $(t-1)$-element subset of $[m]$. We claim that $\{z,z'\} \cup \bigcup_{i \in I}S_i$ is a cocircuit. Let $J$ be an $(s-1)$-element subset of $[s,m]-I$. Then $C=\{z,z'\} \cup \bigcup_{j \in J}S_j$ is a circuit since $(\{z,z'\}, S_{s}, S_{s+1}, \dotsc, S_m)$ is an $s$-echidna. By \cref{lem:rep-orthog}, there is a $2t$-element cocircuit~$C^*$ containing $\{z\} \cup \bigcup_{i \in I}S_i$. By orthogonality between $C$ and $C^*$, we have $z'\in C^*$. Since $I$ was arbitrarily chosen, $(\{z,z'\}, S_1, S_2, \dotsc, S_m)$ is a $t$-coechidna. By the dual of \cref{lem:coechidna}, it is also an $s$-echidna, contradicting the maximality of $(S_1,\dotsc,S_m)$. \end{proof} \section{Matroids with the \texorpdfstring{$(s,2s,t,2t)$}{(s,2s,t,2t)}-property} \label{sec:t2t} In this section, we prove that every sufficiently large matroid with the $(s,2s,t,2t)$-property is an $(s,t)$-spike. We will show that a sufficiently large matroid with the $(s,2s,t,2t)$-property has a large $s$-echidna or $t$-coechidna; it then follows, by \cref{lem:swamping}, that the matroid is an $(s,t)$-spike. As in the previous section, we assume that $s$ and $t$ are positive integers. \begin{lemma} \label{lem:rank-t} Let $M$ be a matroid with the $(s,2s,t,2t)$-property, and let $X\subseteq E(M)$. \begin{enumerate} \item If $r(X)<s$, then $X$ is independent.\label{rt1} \item If $r(X)=s$, then $M|X\cong U_{s,|X|}$ and $|X|<s+2t$.\label{rt2} \end{enumerate} \end{lemma} \begin{proof} Every subset of $E(M)$ of size at most $s$ is independent since it is contained in a circuit of size $2s$. In particular, \ref{rt1} holds. Now let $r(X)=s$. Then every $(s+1)$-element subset of $X$ is a circuit, so $M|X\cong U_{s,|X|}$. Suppose for a contradiction that $|X|\geq s+2t$. Let $C^*$ be a $2t$-element cocircuit such that there is some $x\in X\cap C^*$. Then $X-C^*$ is contained in the hyperplane $E(M)-C^*$. Since $x\in X\cap C^*$, we have $r(X-C^*)<r(X)=s$. Therefore, $X-C^*$ is an independent set, so $|X-C^*|<s$. Since $|X|\geq s+2t$, we have $|C^*|>2t$, a contradiction. Thus, \ref{rt2} holds. \end{proof} \begin{lemma} \label{lemmaA} Let $M$ be a matroid with the $(s,2s,t,2t)$-property, and let $C_1^*,C_2^*,\dotsc,C_{s-1}^*$ be a collection of pairwise disjoint cocircuits of $M$. Let $Y = E(M)-\bigcup_{i \in [s-1]} C_i^*$. For all $y \in Y$, there is a $2s$-element circuit~$C_y$ containing $y$ such that either \begin{enumerate} \item $|C_y \cap C_i^*| = 2$ for all $i \in [s-1]$, or\label{A1} \item $|C_y \cap C_j^*| = 3$ for some $j \in [s-1]$, and $|C_y \cap C_i^*| = 2$ for all $i \in [s-1]-\{j\}$.\label{A2} \end{enumerate} Moreover, if $C_y$ satisfies \ref{A2}, then there are at most $s+2t-1$ elements $w \in Y$ such that $(C_y-y) \cup \{w\}$ is a circuit. \end{lemma} \begin{proof} Choose an element $c_i \in C_i^*$ for each $i \in [s-1]$. By the $(s,2s,t,2t)$-property, there is a $2s$-element circuit~$C_y$ containing $\{c_1,c_2,\dotsc,c_{s-1},y\}$, for each $y \in Y$. By orthogonality, $C_y$ satisfies \ref{A1} or \ref{A2}. Suppose $C_y$ satisfies \ref{A2}, and let $S =C_y-Y= C_y-\{y\}$. Let $W = \{w \in Y : S \cup \{w\} \textrm{ is a circuit}\}$. It remains to prove that $|W| < s+2t$. Observe that $W \subseteq \cl(S) \cap Y$, and, since $S$ contains $s-1$ elements in pairwise disjoint cocircuits that avoid $Y$, we have $r(\cl(S) \cup Y) \ge r(Y) + (s-1)$. Thus, \begin{align*} r(W) &\le r(\cl(S) \cap Y) \\ &\le r(\cl(S)) + r(Y) - r(\cl(S) \cup Y) \\ &\le (2s-1) + r(Y) - (r(Y)+ (s-1)) \\ &=s, \end{align*} using submodularity of the rank function at the second line. Now, by \cref{lem:rank-t}\ref{rt1}, if $r(W) < s$, then $W$ is independent, so $|W| = r(W) < s < s + 2t$. On the other hand, by \cref{lem:rank-t}\ref{rt2}, if $r(W)=s$, then $M|W \cong U_{t,|W|}$ and $|W|<s+2t$, as required. \end{proof} \begin{lemma} \label{lem:disjoint} There exists a function $h$ such that if $M$ is a matroid with at least $h(k,d,t)$ $k$-element circuits, and the property that every $t$-element set is contained in a $2t$-element cocircuit for some positive integer $t$, then $M$ has a collection of $d$ pairwise disjoint $2t$-element cocircuits. \end{lemma} \begin{proof} By \cite[Lemma 3.2]{bccgw2019}, there is a function $g$ such that if $M$ has at least $g(k,d)$ $k$-element circuits, then $M$ has a collection of $d$ pairwise disjoint circuits. We define $h(k,d,t) = g(k,dt)$, and claim that a matroid with at least $h(k,d,t)$ $k$-element circuits, and the property that every $t$-element set is contained in a $2t$-element cocircuit, has a collection of $d$ pairwise disjoint $2t$-element cocircuits. Let $M$ be such a matroid. Then $M$ has a collection of $dt$ pairwise disjoint circuits. We partition these into $d$ groups of size $t$: call this partition $(\mathcal{C}_1,\dotsc,\mathcal{C}_d)$. Since the $t$ circuits in any cell of this partition are pairwise disjoint, it now suffices to show that, for each $i \in [d]$, there is a $2t$-element cocircuit contained in the union of the members of $\mathcal{C}_i$. Let $\mathcal{C}_i = \{C_1,\dotsc,C_{t}\}$ for some $i \in [d]$. Pick some $c_j \in C_j$ for each $j \in [t]$. Then, since $\{c_1,c_2,\dotsc,c_{t}\}$ is a $t$-element set, it is contained in a $2t$-element cocircuit, which, by orthogonality, is contained in $\bigcup_{j \in [t]}C_j$. \end{proof} \begin{lemma} \label{setup} Let $M$ be a matroid with the $(s,2s,t,2t)$-property such that $r(M)\geq r^*(M)$. There exists a function $g$ such that, if $|E(M)| \ge g(s,t,q)$, then $M$ has $s-1$ pairwise disjoint $2t$-element cocircuits $C_1^*, C_2^*, \dotsc, C_{s-1}^*$, and there is some $Z \subseteq E(M)-\bigcup_{i \in [s-1]}C_i^*$ such that \begin{enumerate} \item $r_{M}(Z) \ge q$, and\label{ps1} \item for each $z \in Z$, there exists an element $z'\in Z-\{z\}$ such that $\{z,z'\}$ is contained in a $2s$-element circuit~$C$ with $|C \cap C_i^*|=2$ for each $i \in [s-1]$.\label{ps2} \end{enumerate} \end{lemma} \begin{proof} By \cref{lem:disjoint}, there is a function $h$ such that if $M$ has at least $h(k,d,t)$ $k$-element circuits, then $M$ has $d$ pairwise disjoint $2t$-element cocircuits. Suppose $|E(M)|\geq 2s\cdot h(2s,s-1,t)$. By the $(s,2s,t,2t)$-property, $M$ has at least $h(2s,s-1,t)$ distinct $2s$-element circuits. Therefore, by \cref{lem:disjoint}, $M$ has a collection of $s-1$ pairwise disjoint $2t$-element cocircuits $C_1^*,\dotsc, C_{s-1}^*$. Let $X = \bigcup_{i \in [s-1]}C_i^*$ and $Y=E(M)-X$. By \cref{lemmaA}, for each $y \in Y$ there is a $2s$-element circuit~$C_y$ containing $y$ such that $|C_y \cap C_j^*| = 3$ for at most one $j \in [s-1]$ and $|C_y \cap C_i^*| = 2$ otherwise. Let $W$ be the set of all $w \in Y$ such that $w$ is in a $2s$-element circuit~$C$ with $|C\cap C_j^*|=3$ for some $j \in [s-1]$, and $|C \cap C_i^*|=2$ for all $i \in [s-1]-\{j\}$. Now, letting $Z=Y-W$, we see that \ref{ps2} is satisfied. It remains to show that \ref{ps1} holds. Since each $C_i^*$ has size $2t$, there are $(s-1)\binom{2t}{3}\binom{2t}{2}^{s-2}$ sets $X'\subseteq X$ with $|X' \cap C_j^*|=3$ for some $j \in [s-1]$ and $|X' \cap C_i^*|=2$ for all $i \in [s-1]-\{j\}$. % It follows, by \cref{lemmaA}, that $|W| \le f(s,t)$ where \[f(s,t) = (s+2t-1)\left[(s-1)\binom{2t}{3}\binom{2t}{2}^{s-2}\right].\] We define \[g(s,t,q) = \max\left\{2s\cdot h(2s,s-1,t), 2\big(2t(s-1)+f(s,t)+q\big)\right\}.\] Suppose that $|E(M)| \ge g(s,t,q)$. Since $r(M)\geq r^*(M)$ and $|E(M)|\geq2(2t(s-1)+f(s,t)+q)$, we have $r(M) \ge 2t(s-1)+f(s,t)+q$. Then, \begin{align*} r_{M}(Z) &\ge r_{M}(Y) - |W| \\ &\ge \big(r(M)-2t(s-1)\big) - f(s,t) \\ &\ge q, \end{align*} so \ref{ps1} holds as well. \end{proof} \sloppy \begin{lemma} \label{lem:payoff} Let $M$ be a matroid with the $(s,2s,t,2t)$-property. Suppose $M$ has $s-1$ pairwise disjoint $2t$-element cocircuits $C_1^*, C_2^*, \dotsc, C_{s-1}^*$ and, for some positive integer~$p$, there is a set $Z \subseteq E(M)-\bigcup_{i \in [s-1]}C_i^*$ such that \begin{enumerate}[label=\rm(\alph*)] \item $r(Z) \ge \binom{2t}{2}^{s-1}(p + 2(s-1))$, and \item for each $z \in Z$, there exists an element $z'\in Z-\{z\}$ such that $\{z,z'\}$ is contained in a $2s$-element circuit $C$ of $M$ with $|C\cap C_i^*|=2$ for each $i\in [s-1]$. \end{enumerate} There exists a subset $Z' \subseteq Z$ and a partition $\pi=( Z_1', \dotsc, Z_p' )$ of $Z'$ into pairs such that \begin{enumerate} \item each circuit of $M|Z'$ is a union of pairs in $\pi$, and \item the union of any $s$ pairs in $\pi$ contains a circuit. \end{enumerate} \end{lemma} \fussy \begin{proof} We first prove the following: \begin{sublemma} \label{prelem:payoff} There exists a $(2s-2)$-element set $X$ such that $|X\cap C_i^*|=2$ for every $i\in[s-1]$ and a set $Z'\subseteq Z$ with a partition $\pi=\{ Z_1', \dotsc, Z_p' \}$ of $Z'$ into pairs such that \begin{enumerate}[label=\rm(\Roman*)] \item $X \cup Z_i'$ is a circuit, for each $i \in [p]$ and \label{ppo1} \item $\pi$ partitions the ground set of $(M/X)|Z'$ into parallel classes such that $r_{M/X}\big(\bigcup_{i \in [p]}Z_i'\big)=p$. \label{ppo2} \end{enumerate} \end{sublemma} \begin{subproof} By (b), for each $z \in Z$, there exists an element $z'\in Z-\{z\}$ and a set $X'$ such that $\{z,z'\} \cup X'$ is a circuit of $M$ and $X'$ is the union of pairs $Y_i$ for $i\in[s-1]$, with $Y_i\subseteq C_i^*$. Since $|C_i^*|=2t$ for each $i\in[s-1]$, there are $\binom{2t}{2}^{s-1}$ choices for $(Y_1,Y_2,\ldots,Y_{s-1})$. Therefore, for some $m\leq\binom{2t}{2}^{s-1}$, there are $(2s-2)$-element sets $X_1,X_2,\ldots,X_m$, and sets $Z_1,Z_2,\ldots,Z_m$ whose union is $Z$, such that each of $X_1,X_2,\ldots,X_m$ intersects $C_i^*$ in two elements for each $i\in[s-1]$, and such that, for each $j\in[m]$ and each $z_j\in Z_j$, there is an element $z_j'$ such that $\{z_j,z_j'\}\cup X_j$ is a circuit. Since $Z=\bigcup_{i \in [m]}Z_i$, we have $\sum_{i\in[m]}r(Z_i)\geq r(Z)$. Thus, the pigeonhole principle implies that there is some $j\in[m]$ such that \[r(Z_j) \ge \frac{r(Z)}{\binom{2t}{2}^{s-1}} \ge p+2(s-1),\] by (a). We define $Z' = Z_j$ and $X = X_j$. Observe that $X \cup \{z,z'\}$ is a circuit, for some pair $\{z,z'\} \subseteq Z'$, if and only if $\{z,z'\}$ is a parallel pair in $M/X$. Therefore, there is a partition of the ground set of $(M/X)|Z'$ into parallel classes, where every parallel class has size at least two. Let $\{\{z_1,z_1'\}, \dotsc,\{z_n,z_n'\}\}$ be a collection of pairs from each parallel class such that $\{z_1,z_2,\dotsc,z_n\}$ is an independent set in $(M/X)|Z'$. Note that $n\geq r_{M/X}(Z') = r(Z' \cup X) -r(X) \ge r(Z') - 2(s-1) \ge p$. For $i\in[p]$, let $Z_i'=\{z_i,z_i'\}$. Then $\pi=\{ Z_1', \dotsc, Z_p' \}$ satisfies \ref{prelem:payoff}. \end{subproof} Let $X$, $\pi$, and $Z'$ be as described in \ref{prelem:payoff}, and let $\mathcal{X} = \{X_1,\dotsc,X_{s-1}\}$, where $X_i = \{x_i,x_i'\} = X \cap C_i^*$. \begin{sublemma} \label{metamatroid} Each circuit of $M|(X \cup Z')$ is a union of pairs in $\mathcal{X} \cup \pi$. \end{sublemma} \begin{subproof} Let $C$ be a circuit of $M|(X \cup Z')$. If $x_i \in C$, for some $\{x_i,x_i'\} \in \mathcal{X}$, then orthogonality with $C_i^*$ implies that $x_i' \in C$. Assume for a contradiction that $\{z,z'\} \in \pi$ and $C \cap \{z,z'\} = \{z\}$. Let $W$ be the union of the pairs in $\pi$ containing elements of $(C-\{z\}) \cap Z'$. Then $z \in \cl(X \cup W)$. Hence $z \in \cl_{M/X}(W)$, contradicting \cref{prelem:payoff}\ref{ppo2}. \end{subproof} \begin{sublemma} \label{induct} Every union of $s$ pairs in $\mathcal{X} \cup \pi$ contains a circuit. \end{sublemma} \begin{subproof} Let $\mathcal{W}$ be a subset of $\mathcal{X} \cup \pi$ of size $s$. We proceed by induction on the number of pairs in $\mathcal{W} \cap \pi$. If there is only one pair in $\mathcal{W} \cap \pi$, then the union of the pairs in $\mathcal{W}$ contains a circuit (indeed, is a circuit) by \cref{prelem:payoff}\ref{ppo1}. Suppose the result holds for any subset containing $k$ pairs in $\pi$, and let $\mathcal{W}$ be a subset containing $k+1$ pairs in $\pi$. Let $\{x,x'\}$ be a pair in $\mathcal{X}-\mathcal{W}$, and let $W = \bigcup_{W' \in \mathcal{W}}W'$. Then $W \cup \{x,x'\}$ is the union of $s+1$ pairs of $\mathcal{X} \cup \pi$, of which $k+1$ are in $\pi$, so, by the induction hypothesis, $W \cup \{x,x'\}$ properly contains a circuit~$C_1$. If $\{x,x'\} \subseteq E(M)-C_1$, then $C_1 \subseteq W$, in which case the union of the pairs in $\mathcal{W}$ contains a circuit, as desired. Therefore, we may assume, by \cref{metamatroid}, that $\{x,x'\} \subseteq C_1$. Since $X$ is independent, there is a pair $\{z,z'\} \subseteq Z' \cap C_1$. By the induction hypothesis, there is a circuit~$C_2$ contained in $(W-\{z,z'\}) \cup \{x,x'\}$. Observe that $C_1$ and $C_2$ are distinct, and $\{x,x'\} \subseteq C_1 \cap C_2$. Circuit elimination on $C_1$ and $C_2$, and \cref{metamatroid}, imply that there is a circuit $C_3 \subseteq (C_1 \cup C_2) - \{x,x'\} \subseteq W$, as desired. The claim now follows by induction. \end{subproof} Now, \cref{induct} implies that the union of any $s$ pairs in $\pi$ contains a circuit, and the result follows. \end{proof} \begin{lemma} \label{lem:tis1} If $M$ is a matroid with the $(1,2,t,2t)$-property and at least $t$ elements, then $M$ is a $(1,t)$-spike. Dually, if $M$ is a matroid with the $(s,2s,1,2)$-property and at least $s$ elements, then $M$ is an $(s,1)$-spike. \end{lemma} \begin{proof} By duality, it suffices to consider the case where $M$ has the $(1,2,t,2t)$-property and at least $t$ elements. Since every element of $M$ is contained in a $2$-element circuit, there is a partition of $E(M)$ into parallel classes $P_1,P_2,\ldots,P_n$, where $|P_i|\geq2$ for each $i$. For each $P_i$, let $x_i\in P_i$. First, we consider the case where $n\geq t$. Let $X$ be a $t$-element subset of $\{x_1,\ldots,x_{n}\}$; for ease of notation, we assume $X=\{x_1,\ldots,x_{t}\}$. By the $(1,2,t,2t)$-property, $X\subseteq C^*$ for some $2t$-element cocircuit $C^*$. Since $P_i$ is a parallel class, $\{x_i,y_i\}$ is a circuit for each $y_i\in P_i-\{x_i\}$. By orthogonality, $y_i\in C^*$ for each such $y_i$, so $P_i\subseteq C^*$. Since $|C^*|=2t$, and $X$ is an arbitrary $t$-element subset of $\{x_1,\ldots,x_{n}\}$, it follows that $|P_i|=2$ for each $i\in[n]$, and that the union of any $t$ of the $P_i$'s is a cocircuit. Thus $M$ is a $(1,t)$-spike. It remains to consider the case where $n<t$. Since $M$ has at least $t$ elements, let $X$ be any $t$-element set containing $\{x_1,\ldots,x_n\}$. By the $(1,2,t,2t)$-property, there is a $2t$-element cocircuit $C^*$ containing $X$. For $i\in[n]$ and each $y_i\in P_i-\{x_i\}$, orthogonality implies $y_i\in C^*$. Thus, $E(M)=C^*$. It follows that $M\cong U_{1,2t}$, which is a $(1,t)$-spike. \end{proof} We now prove \cref{mainthm}, restated below. \begin{theorem} \label{mainthmtake2} There exists a function $f : \mathbb{N}^2 \rightarrow \mathbb{N}$ such that, if $M$ is a matroid with the $(s,2s,t,2t)$-property and $|E(M)| \ge f(s,t)$, then $M$ is an $(s,t)$-spike. \end{theorem} \begin{proof} If $s=1$ or $t=1$, then, by \cref{lem:tis1}, the theorem holds with $f(s,t) = \max\{s,t\}$. So we may assume that $\min\{s,t\} \ge 2$. A matroid is an $(s,t)$-spike if and only if its dual is a $(t,s)$-spike; moreover, a matroid has the $(s,2s,t,2t)$-property if and only if its dual has the $(t,2t,s,2s)$-property. Therefore, by duality, we may also assume that $r(M)\geq r^*(M)$. Let $r_k(n)$ be the Ramsey number described in \cref{hyperramsey}. For $k \in [s]$, we define the function $h_k : \mathbb{N}^2 \rightarrow \mathbb{N}$ such that \[h_{s}(s,t)=\max\{s+2t-1,2s+t-1,3s+t-3,s+3t-3\}\] and such that $h_k(s,t)=r_k(h_{k+1}(s,t))$ for $k\in[s-1]$. Note that $h_{k}(s,t) \ge h_{k+1}(s,t) \ge h_{s}(s,t)$, for each $k \in [s-1]$. Let $p = h_1(s,t)$ and let $q(s,t)=\binom{2t}{2}^{s-1}(p + 2(s-1))$. By \cref{setup}, there exists a function $g$ such that if $|E(M)| \ge g(s,t,q(s,t))$, then $M$ has $s-1$ pairwise disjoint $2t$-element cocircuits $C_1^*, C_2^*, \dotsc, C_{s-1}^*$, and there is some $Z \subseteq E(M)-\bigcup_{i \in [s-1]}C_i^*$ such that $r_M(Z) \ge q(s,t)$, and, for each $z \in Z$, there exists an element $z'\in Z'-\{z\}$ such that $\{z,z'\}$ is contained in a $2s$-element circuit~$C$ with $|C \cap C_i^*|=2$ for each $i \in [s-1]$. Let $f(s,t) = g(s,t,q(s,t))$, and suppose that $|E(M)| \ge f(s,t)$. Then, by \cref{lem:payoff}, there exists a subset $Z \subseteq Z'$ such that $Z$ has a partition into pairs $\pi = ( Z_1, \dotsc, Z_{p})$ such that \begin{enumerate}[label=\rm(\Roman*)] \item each circuit of $M|Z$ is a union of pairs in $\pi$, and \item the union of any $s$ pairs in $\pi$ contains a circuit.\label{rc2} \end{enumerate} Let $m=h_{s}(s,t)$. By \cref{lem:swamping} and its dual, it suffices to show that $M$ has either an $s$-echidna or a $t$-coechidna of order $m$. If the smallest circuit in $M|Z$ has size $2s$, then, by \ref{rc2}, $\pi$ is an $s$-echidna of order $p \ge m$. So we may assume that the smallest circuit in $M|Z$ has size $2j$ for some $j \in [s-1]$. \begin{sublemma} \label{iterramsey} If the smallest circuit in $M|Z$ has size $2j$, for $j \in [s-1]$, and $|\pi| \ge h_j(s,t)$, then either \begin{enumerate} \item $M$ has a $t$-coechidna of order $m$, or\label{ir1} \item there exists some $Z' \subseteq Z$ that is the union of $h_{j+1}(s,t)$ pairs in $\pi$ for which the smallest circuit in $M|Z'$ has size at least $2(j+1)$.\label{ir2} \end{enumerate} \end{sublemma} \begin{subproof} We define $H$ to be the $j$-uniform hypergraph with vertex set $\pi$ whose hyperedges are the $j$-subsets of $\pi$ that are partitions of circuits in $M|Z$. By \cref{hyperramsey}, and the definition of $h_k$, as $H$ has at least $h_j(s,t)$ vertices, it has either a clique or a stable set, on $h_{j+1}(s,t)$ vertices. If $H$ has a stable set~$\pi'$ on $h_{j+1}(s,t)$ vertices, then clearly \ref{ir2} holds, with $Z' = \bigcup_{P \in \pi'} P$. Therefore, we may assume that there are $h_{j+1}(s,t)$ pairs in $\pi$ such that the union of any $j$ of these pairs is a circuit. Let $Z''$ be the union of these $h_{j+1}(s,t)$ pairs. We claim that the union of any set of $t$ pairs contained in $Z''$ is a cocircuit. Let $T$ be a transversal of $t$ pairs in $\pi$ contained in $Z''$, and let $C^*$ be the $2t$-element cocircuit containing $T$. Suppose, for a contradiction, that there exists some pair $P \in \pi$ with $P \subseteq Z''$ such that $|C^* \cap P| = 1$. Select $j-1$ pairs $Z_1'',\dotsc,Z_{j-1}''$ in $\pi$ that are each contained in $Z''-C^*$ (these exist since $h_{j+1}(s,t) \ge s+2t-1 \ge 2t + j - 1$). Then $P \cup (\bigcup_{i \in [j-1]}Z_i'')$ is a circuit intersecting $C^*$ in a single element, contradicting orthogonality. We deduce that the union of any $t$ pairs in $\pi$ that are contained in $Z''$ is a cocircuit. Thus, $M$ has a $t$-coechidna of order $h_{j+1}(t) \ge m$, satisfying \ref{ir1}. \end{subproof} We now apply \cref{iterramsey} iteratively, for a maximum of $s-j$ iterations. If \ref{ir1} holds, at any iteration, then $M$ has a $t$-coechidna of order $m$, as required. Otherwise, we let $\pi'$ be the partition of $Z'$ induced by $\pi$; then, at the next iteration, we relabel $Z=Z'$ and $\pi=\pi'$. If \ref{ir2} holds for each of $s-j$ iterations, then we obtain a subset $Z'$ of $Z$ such that the smallest circuit in $M|Z'$ has size $2s$. Then, by \ref{rc2}, $M$ has an $s$-echidna of order $h_{s}(s,t)=m$, completing the proof. \end{proof} \section{Properties of \texorpdfstring{$(s,t)$}{(s,t)}-spikes} \label{sec:tspikeprops} In this section, we prove some properties of $(s,t)$-spikes. In particular, we show that an $(s,t)$-spike has order at least $s+t-1$; an $(s,t)$-spike of order~$m$ has $2m$ elements and rank~$m+s-t$; and the circuits of an $(s,t)$-spike that are not a union of $s$ arms meet all but at most $t-2$ of the arms. We also give some results about the connectivity of $(s,t)$-spikes of sufficiently large order. We also show that an appropriate concatenation of the associated partition of a $t$-spike is a $(2t-1)$-anemone, following the terminology of~\cite{ao2008}. Finally, we describe a construction that can be used to obtain an $(s,t+1)$-spike from an $(s,t)$-spike of sufficiently large order, and we show that every $(s,t+1)$-spike can be constructed from some $(s,t)$-spike in this way. We again assume that $s$ and $t$ are positive integers. \subsection*{Basic properties} \begin{lemma} \label{tspikeorder} Let $M$ be an $(s,t)$-spike with associated partition $(A_1,\ldots,A_m)$. Then $m \ge s+t-1$. \end{lemma} \begin{proof} By the definition of an $(s,t)$-spike, we have $m\geq\max\{s,t\}$. Let $Y = \bigcup_{j \in [t]}A_j$, and let $y\in Y$. Since $Y$ is a cocircuit, $Z=(E(M)-Y) \cup \{y\}$ spans $M$. Therefore, $r(M)\leq|Z|=2m-2t+1$. Similarly, by duality, $r^*(M)\leq2m-2s+1$. Therefore, \[2m = |E(M)| = r(M) + r^*(M) \le (2m-2t+1)+(2m-2s+1).\] The result follows. \end{proof} \begin{lemma} \label{lem:rank-matroid} Let $M$ be an $(s,t)$-spike of order~$m$. Then $r(M)=m+s-t$ and $r^*(M)=m-s+t$. \end{lemma} \begin{proof} Let $(A_1,\ldots,A_m)$ be the associated partition of $M$, and let $A_i = \{x_i,y_i\}$ for each $i \in [m]$. Choose $I \subseteq J \subseteq [m]$ such that $|I|=s-1$ and $|J| = m-t$. (This is possible by \cref{tspikeorder}.) Let $X = \{y_j : i \in I\} \cup \{x_j : j \in J\}$. Note that $\bigcup_{i \in I\cup J}A_i\subseteq\cl(X)$. Since $E(M)-\bigcup_{i \in I\cup J}A_i$ is a cocircuit, $\bigcup_{i \in I\cup J}A_i$ is a hyperplane. Therefore, $\bigcup_{i \in I\cup J}A_i=\cl(X)$, and we have $r(M)-1=r(X)\leq|X|=|I|+|J|=m+s-t-1$. Thus, $r(M)\leq m+s-t$. Similarly, by duality, $r^*(M)\leq m-s+t$. Therefore, we have \[2m=|E(M)|=r(M)+r^*(M)\leq(m+s-t)+(m-s+t)=2m.\] Thus, we must have equality, and the result holds. \end{proof} \sloppy \begin{lemma} \label{l:circuits} Let $M$ be an $(s,t)$-spike of order~$m$ with associated partition $(A_1,\ldots,A_m)$, and let $C$ be a circuit of $M$. \begin{enumerate} \item $C = \bigcup_{j \in J}A_j$ for some $s$-element set $J \subseteq [m]$, or\label{c1} \item $\left|\{i \in [m] : A_i \cap C \neq \emptyset\}\right| \ge m-(t-2)$ and $\left|\{i \in [m] : A_i \subseteq C\}\right| < s$.\label{c2} \end{enumerate} \end{lemma} \fussy \begin{proof} Let $S = \{i \in [m] : A_i \cap C \neq \emptyset\}$. Thus, $S$ is the minimal subset of $[m]$ such that $C \subseteq \bigcup_{i \in S}A_i$. We have $|S| \ge s$ since $C$ is independent otherwise. If $|S|=s$, then $C$ satisfies \ref{c1}. Therefore, we may assume $|S| > s$. We must have $\left|\{i \in [m] : A_i \subseteq C\}\right| < s$; otherwise $C$ properly contains a circuit. Thus, there is some $j \in S$ such that $A_j - C \neq \emptyset$. If $|S| \ge m-(t-2)$, then $C$ satisfies \ref{c2}. Therefore, we may assume $|S| \le m-(t-1)$. Let $T = ([m]-S) \cup \{j\}$. Then $|T|\ge t$, implying that $\bigcup_{i \in T}A_i$ contains a cocircuit intersecting $C$ in one element. This contradicts orthogonality. \end{proof} In the remainder of the paper, if $(A_1,\ldots,A_m)$ is the associated partition of an $(s,t)$-spike and $J\subseteq[m]$, then we define \[A_J=\bigcup_{j \in J} A_j.\] \begin{proposition} \label{pro:rank-func} Let $\pi=(A_1,\ldots,A_m)$ be the associated partition of an $(s,t)$-spike. If $J\subseteq[m]$, then \[r(A_J) = \begin{cases} 2|J| & \textrm{if $|J| < s$,}\\ s+|J|-1 & \textrm{if $s\leq|J| \leq m-t+1$,}\\ m+s-t & \textrm{if $|J| \ge m-t+1$.} \end{cases}\] \end{proposition} \begin{proof} If $|J|<s$, then $A_J$ is properly contained in a circuit and is therefore independent. Thus, $r(A_J)=|A_J|=2|J|$. We now prove that $r(A_J)=s+|J|-1$ if $s\leq|J| \leq m-t+1$. We proceed by induction on $|J|$. As a base case, if $|J|=s$, then $A_J$ is a circuit. Therefore, $r(A_J)=|A_J|-1=s+|J|-1$. Now, for the inductive step, let $s<|J|\leq m-t+1$, and let $J'\subseteq J$ with $|J'|=|J|-1$. By induction, $r(A_{J'})=s+|J|-2$. Let $\{x_i,y_i\}=A_J-A_{J'}$. By \cref{l:circuits}, since $|J|<m-t+2$, there is no circuit $C$ such that $x_i\in C\subseteq A_{J'}\cup\{x_i\}$. Therefore, $x_i\notin\cl(A_{J'})$, and $r(A_{J'}\cup\{x_i\})=r(A_{J'})+1$. On the other hand, since $|J|>s$, there is a circuit $C$ such that $y_i\in C\subseteq A_{J}$. Therefore, $y_i\in\cl(A_{J'}\cup\{x_i\})$, and $r(A_J)=r(A_{J'})+1=s+|J|-1$. Note that the preceding argument, along with \cref{lem:rank-matroid} implies that, if $|J|=m-t+1$, then $A_J$ is spanning. Thus, if $|J|\geq m-t+1$, then $r(A_J)=r(M)=m+s-t$. \end{proof} \subsection*{Connectivity} Let $M$ be a matroid with ground set $E$. Recall that the \emph{connectivity function} of $M$, denoted by $\lambda$, is defined as \begin{align*} \lambda(X) = r(X) + r(E - X) - r(M), \end{align*} for all subsets $X$ of $E$. In the case where $M$ is an $(s,t)$-spike of order $m$ and $X=A_J$ for some set $J\subseteq[m]$, this implies \begin{align*} \lambda(A_J) = r(A_J) + r(A_{[m]-J}) - r(M). \end{align*} Therefore, \cref{pro:rank-func} allows us to easily compute $\lambda(A_J)$. \begin{lemma} \label{lem:conn} Let $\pi=(A_1,\ldots,A_m)$ be the associated partition of an $(s,t)$-spike, and let $(J,K)$ be a partition of $[m]$ with $|J| \le |K|$. \begin{enumerate} \item If $|J|\leq t-1$, then $\lambda(A_J)=r(A_J)$. \item If $t-1\leq|J|\leq m-s$, then \[\lambda(A_J)= \begin{cases} t+|J|-1 & \textrm{if $|J| < s$,}\\ s+t-2 & \textrm{if $s\leq|J|\leq m-t+1$.} \end{cases}\] \item If $|J|> m-s$, then $\lambda(A_J)=m-s+t$. \end{enumerate} \end{lemma} \begin{proof} If $|J|\leq t-1$, then $|K|\geq m-t+1$. Therefore, $A_K$ is spanning, and $\lambda(A_J)=r(A_J)+r(A_K)-r(M)=r(A_J)$. Statement (i) follows. If $t-1\leq|J|\leq m-s$, then $s\leq|K|\leq m-t+1$. Therefore, $\lambda(A_J)=r(A_J)+r(A_K)-r(M)=r(A_J)+s+m-|J|-1-(m+s-t)$. Statement (ii) follows. (Note that we cannot have $|J|>m-t+1$ because otherwise $|K|<t-1\leq|J|$.) If $|J|> m-s$, then $s>|K|\geq|J|$. Therefore, $\lambda(A_J)=r(A_J)+r(A_K)-r(M)=2|J|+2(m-|J|)-(m+s-t)=m-s+t$. Statement (iii) follows. \end{proof} Using the terminology of~\cite{ao2008}, \cref{lem:conn} implies the following. \begin{proposition} \label{pro:anemone} Let $(A_1,\dotsc,A_m)$ be the associated partition of an $(s,t)$-spike~$M$, and suppose that $(P_1,\dotsc,P_k)$ is a partition of $E(M)$ such that, for each $i \in [k]$, $P_i = \bigcup_{i \in I}A_i$ for some subset $I$ of $[m]$, with $|I| \ge \max\{s-1,t-1\}$. Then $(P_1,\dotsc,P_k)$ is an $(s+t-1)$-anemone. \end{proposition} We now continue our study of the connectivity of $(s,t)$-spikes. \begin{lemma} \label{ind-and-coind} Let $M$ be an $(s,t)$-spike of order $m\geq3\max\{s,t\}-2$, and let $X\subseteq E(M)$ such that $|X|\leq2\min\{s,t\}-1$. Then $\lambda(X)=|X|$. \end{lemma} \begin{proof} By Lemma \ref{l:circuits}, if $X$ is dependent, then either $|X|=2s$ or $|X|\geq m-t+2\geq 3\max\{s,t\}-2-t+2=3\max\{s,t\}-t\geq2\max\{s,t\}\geq2s$. However, $|X|\leq2\min\{s,t\}-1<2s$. Therefore, $X$ is independent, which implies that $r(X)=|X|$. By a similar argument, using the dual of \cref{l:circuits}, $X$ is coindependent, implying that $r(E(M)-X)=r(M)$. Therefore, \begin{align*} \lambda(X)&=r(X)+r(E(M)-X)-r(M)\\ &=|X|+r(M)-r(M)\\ &=|X|, \end{align*} proving the lemma. \end{proof} \begin{theorem} Let $M$ be an $(s,t)$-spike of order \[m\geq\max\{3s+t,s+3t\}-4,\] where $\min\{s,t\}\geq2$. Then $M$ is $(2\min\{s,t\}-1)$-connected. \end{theorem} \begin{proof} Because $M^*$ is a $(t,s)$-spike and because $\lambda_{M^*}=\lambda_M$, we may assume without loss of generality that $t\leq s$. Note that $\max\{3s+t,s+3t\}=3\max\{s,t\}+\min\{s,t\}$. Therefore, $m\geq3s+t-4$, and we must show that $M$ is $(2t-1)$-connected. Now, suppose for a contradiction that $M$ is not $(2t-1)$-connected. Then there is a $k$-separation $(P,Q)$ of $M$, with $|P|\geq|Q|$, for some $k<2t-1$. Therefore, $\lambda(P)=\lambda(Q)<k\leq2t-2$. First, we consider the case where $A_I \subseteq P$, for some $(t-1)$-element set $I \subseteq [m]$. Let $U = \{u \in [m] : |P \cap A_u|= 1\}$. Then $A_j \subseteq \cl_{M^*}(P)$ for each $j \in U$. For such a $j$, it follows, by the definition of $\lambda_{M^*}$ (which is equal to $\lambda_M=\lambda$), that $\lambda(P \cup A_j) \le \lambda(P)$. We use this repeatedly below; in particular, we see that $\lambda(P\cup A_U)\leq\lambda(P)$. Let $P' = P\cup A_U$, and let $Q' = E(M)-P'$. Then there is a partition $(J,K)$ of $[m]$, with $|J|\leq|K|$, such that $Q'=A_J$ and $P'=A_K$. Moreover, $\lambda(Q')=\lambda(P')\leq\lambda(P)$. Suppose $|J|\geq t-1$. Note that $m\geq3s+t-4\geq2s$ since $\min\{s,t\}\geq2$. Therefore, $|J|\leq\frac{1}{2}m=m-\frac{1}{2}m\leq m-\frac{1}{2}(2s)=m-s$. Thus, to determine $\lambda(Q')$, we need only consider Lemma \ref{lem:conn}(ii). If $|J|\geq s$, then by Lemma \ref{lem:conn}(ii), \[\lambda(P)\geq\lambda(P')=\lambda(Q')=s+t-2\geq2t-2,\] a contradiction. Otherwise, $|J|<s$, implying by Lemma \ref{lem:conn}(ii) that \[\lambda(P)\geq\lambda(P')=\lambda(Q')=t+|J|-1\geq t+t-1-1=2t-2,\] another contradiction. Therefore, $|J|<t-1$. Let $U'\subseteq U$ such that $|U'|=|Q|-(2t-2)$. Then $\lambda(P) \ge \lambda\left(P \cup A_{U'}\right) = \lambda\left(Q- A_{U'}\right)$. Since $\left|Q- A_{U'}\right| = 2t-2$ and $m\geq3s+t-4\geq3s-2$, \cref{ind-and-coind} implies that $\lambda\left(Q-A_{U'}\right)=2t-2$, so $\lambda(P) \ge 2t-2$, a contradiction. Now we consider the case that $|\{i \in [m] : A_i \subseteq P\}| < t-1$. Since $|Q| \le |P|$, it follows that $|\{i \in [m] : A_i \subseteq Q\}| \le |\{i \in [m] : A_i \subseteq P\}| < t-1<s$. Now, since $|\{i \in [m] : A_i \subseteq P\}| < t-1$, we have $|\{i \in [m] : A_i \cap Q \neq \emptyset\}| > m-(t-1)$. Therefore, $r(Q) \ge m-(t-1)$ by \cref{l:circuits}. Similarly, $r(P) \ge m-(t-1)$. Thus, \begin{align*} \lambda(P) &= r(P) + r(Q) - r(M) \\ &\ge (m-(t-1)) + (m-(t-1)) - (m+s-t) \\ &=m-s-t+2 \\ &\ge 3s+t-4-s-t+2 \\ &= 2s-2\\ &\ge 2t-2, \end{align*} a contradiction. This completes the proof. \end{proof} \subsection*{Constructions} In \cite{bccgw2019}, a construction is described that, starting from a $(t,t)$-spike $M_0$, obtains a $(t+1,t+1)$-spike $M_1$. This construction consists of a certain elementary quotient $M_0'$ of $M_0$, followed by a certain elementary lift $M_1$ of $M_0'$. It is shown in \cite{bccgw2019} that $M_1$ is a $(t+1,t+1)$-spike as long as the order of $M_0$ is sufficiently large. In the process of constructing $M_1$ in this way, the intermediary matroid $M_0'$ is a $(t,t+1)$-spike. For the sake of completeness, we will review this construction in the more general case where $M_0$ is an $(s,t)$-spike, in which case $M_0'$ is an $(s,t+1)$-spike. To construct an $(s+1,t)$-spike, we perform the construction on $M^*$ and dualize. Since $(2,2)$-spikes (and indeed, $(1,1)$-spikes) are well known to exist, this means that $(s,t)$-spikes exist for all positive integers $s$ and $t$. It is also shown in \cite{bccgw2019} that all $(t,t)$-spikes can be constructed in this manner. We also extend this to the general case of $(s,t)$-spikes below. Recall that $M_1$ is an \emph{elementary quotient} of $M_0$ if there is a single-element extension $M^+_0$ of $M_0$ by an element~$e$ such that $M_1 = M^+_0 / e$. If $M_1$ is an elementary quotient of $M_0$, then $M_0$ is an \emph{elementary lift} of $M_1$. Also, note that if $M_1$ is an elementary lift of $M_0$, then $M_1^*$ is an elementary quotient of $M_0^*$. \begin{construction} \label{cons:quotient} Let $M$ be an $(s,t)$-spike of order~$m \ge s+t$, with associated partition $\pi$. Let $M+e$ be a single-element extension of $M$ by an element $e$ such that $e$ blocks each $2t$-element cocircuit that is a union of $t$ arms of $M$. Then let $M'=(M+e)/e$. \end{construction} In other words, $M+e$ has the property that $e\notin \cl_{M+e}(E(M)-C^*)$ for every $2t$-element cocircuit $C^*$ that is the union of $t$ arms. Note that one possibility is that $M+e$ is the free extension of $M$ by an element $e$. Since $m-t\geq s$, we have $e\notin\cl_{M+e}(C)$ for each $2s$-element circuit $C$. Thus, in $M'$, the union of any $s$ arms of the $(s,t)$-spike $M$ is still a circuit of $M'$. However, since $r(M') = r(M) - 1$, the union of any $t+1$ arms is a $2(t+1)$-element cocircuit. Therefore, $M'$ is an $(s,t+1)$-spike. Note that $M'$ is not unique; more than one $(s,t+1)$-spike can be constructed from a given $(s,t)$-spike $M$ using \cref{cons:quotient}. Given an $(s+1,t)$-spike~$M'$, we will describe how to obtain an $(s,t)$-spike~$M$ from $M'$ by a specific elementary quotient. This process reverses the dual of \cref{cons:quotient}. This will then imply that every $(s,t)$-spike can be constructed from a $(1,1)$-spike by repeated use of \cref{cons:quotient} and its dual. \cref{modcut} describes the single-element extension that gives rise to the elementary quotient we desire. Intuitively, the extension adds a ``tip'' to the $(s,t)$-spike. In the proof of this lemma, we assume knowledge of the theory of modular cuts (see \cite[Section~7.2]{oxbook}). The proof of \cref{modcut} will be very similar to the proof of \cite[Lemma 6.6]{bccgw2019}. However, we note that \cite[Lemma 6.6]{bccgw2019} is falsely stated; what is proven in \cite{bccgw2019} is essentially the specialisation of \cref{modcut}, below, in the case that $s=t$. The statement of \cite[Lemma 6.6]{bccgw2019} replaces the condition that $M$ is a $(t,t)$-spike with the weaker condition that $M$ has a $t$-echidna. To demonstrate that this is overly general, consider the rank-$3$ matroid consisting of two disjoint lines with four points. Let these lines be $\{a,b,c,d\}$ and $\{w,x,y,z\}$. Then $(\{a,b\},\{w,x\})$ is a $2$-echidna of order $2$. For \cite[Lemma 6.6]{bccgw2019} to be true, we would need a single-element extension $M^+$ by an element $e$ such that $e\in\cl_{M^+}(\{a,b\})$ but $e\notin\cl_{M^+}(\{c,d\})$. This is impossible since $\cl_M(\{a,b\})=\cl_M(\{c,d\})$. \begin{lemma} \label{modcut} Let $M$ be an $(s,t)$-spike. There is a single-element extension $M^+$ of $M$ by an element $e$ having the property that, for every $X \subseteq E(M)$, $e \in \cl_{M^+}(X)$ if and only if $X$ contains at least $s-1$ arms of $M$. \end{lemma} \begin{proof} Since $M$ is an $(s,t)$-spike, there is a partition $\pi=(S_1,\dotsc,S_m)$ of $E(M)$ that is both an $s$-echidna and a $t$-coechidna. Let $$\mathcal{F} = \left\{\bigcup_{i\in I}S_i : I \subseteq [m] \textrm{ and } |I|=s-1\right\}.$$ By the definition of an $s$-echidna, $\mathcal{F}$ is a collection of flats of $M$. Let $\mathcal{M}$ be the set of all flats of $M$ containing some flat $F \in \mathcal{F}$. We claim that $\mathcal{M}$ is a modular cut. Recall that, for distinct $F_1,F_2 \in \mathcal{M}$, the pair $(F_1,F_2)$ is \emph{modular} if $r(F_1) + r(F_2) = r(F_1 \cup F_2) + r(F_1 \cap F_2)$. To show that $\mathcal{M}$ is a modular cut, it suffices to prove that, for any $F_1,F_2 \in \mathcal{M}$ such that $(F_1,F_2)$ is a modular pair, $F_1 \cap F_2 \in \mathcal{M}$. For any $F \in \mathcal{M}$, since $F$ contains at least $s-1$ arms of $M$, and the union of any $s$ arms is a circuit, it follows that $F$ is a union of arms of $M$. Thus, let $F_1,F_2 \in \mathcal{M}$ be such that $F_1=\bigcup_{i\in I_1}S_i$ and $F_2=\bigcup_{i\in I_2}S_i$, where $I_1$ and $I_2$ are distinct subsets of $[m]$ with $u_1=|I_1| \ge s-1$ and $u_2=|I_2|\ge s-1$. Let $q=|I_1 \cap I_2|$. Then $F_1 \cup F_2$ is the union of $u_1 + u_2 - q \ge s-1$ arms, and $F_1\cap F_2$ is the union of $q$ arms. We show that if $q<s-1$, then $(F_1,F_2)$ is not a modular pair. We consider several cases. First, suppose $u_1,u_2\leq m-t+1$. By \cref{pro:rank-func}, \begin{align*} r(F_1) + r(F_2) &= (s + u_1 - 1) + (s + u_2 - 1) \\ &> (s-1 + u_1 + u_2 - q) +2q \\ &= s+|I_1\cup I_2|-1+2|I_1\cap I_2| \\ &\geq r(F_1 \cup F_2) + r(F_1 \cap F_2). \end{align*} Next, consider the case where $u_2\leq m-t+1<u_1$. (By symmetry, the argument is the same if $u_1$ and $u_2$ are swapped.) One can check that $u_1+u_2-q>m-t+1$. By \cref{pro:rank-func}, \begin{align*} r(F_1) + r(F_2) &= (m+s-t) + (s + u_2 - 1) \\ &> (m + s-t)+2q\\ &= r(F_1 \cup F_2) + r(F_1 \cap F_2). \end{align*} Finally, consider the case where $u_1,u_2>m-t-1$. We have \[r(F_1) + r(F_2) = 2m+2s -2t,\] which by \cref{tspikeorder}, is at least \begin{align*} m+3s-t-1 &> m+s-t+2q\\ &= r(F_1 \cup F_2) + r(F_1 \cap F_2). \end{align*} Thus, in all cases, $(F_1,F_2)$ is not a modular pair. Therefore, we have shown that $\mathcal{M}$ is a modular cut. Now, there is a single-element extension corresponding to the modular cut~$\mathcal{M}$, and this extension satisfies the requirements of the lemma (see, for example, \cite[Theorem~7.2.3]{oxbook}). \end{proof} \begin{theorem} Let $M$ be an $(s,t)$-spike of order $m\geq s+t$. Then $M$ can be constructed from a $(1,1)$-spike of order $m$ by applying \cref{cons:quotient} $t-1$ times, followed by the dual of \cref{cons:quotient} $s-1$ times. \end{theorem} \begin{proof} For $s=t=1$, the result is clear. Otherwise, by duality, we may assume without loss of generality that $t>1$. By induction and duality, it suffices to show that $M$ can be constructed from an $(s-1,t)$-spike of order $m$ by applying the dual of \cref{cons:quotient} once. Let $\pi=(A_1,\dotsc,A_m)$ be the associated partition of $M$. Let $M^+$ be the single-element extension of $M$ by an element~$e$ described in \cref{modcut}. Let $M'=M^+/e$. We claim that $\pi$ is an $(s-1)$-echidna and a $t$-coechidna that partitions the ground set of $M'$. Let $X$ be the union of any $s-1$ spines of $\pi$. Then $X$ is independent in $M$, and $X \cup \{e\}$ is a circuit in $M^+$, so $X$ is a circuit in $M'$. Thus, $\pi$ is an $(s-1)$-echidna of $M'$. Now let $C^*$ be the union of any $t$ spines of $\pi$, and let $H=E(M)-C^*$. Then $H$ is the union of at least $s-1$ spines, so $e \in \cl_{M^+}(H)$. Now $H \cup \{e\}$ is a hyperplane in $M^+$, so $C^*$ is a cocircuit in $M^+$ and therefore in $M'$. Hence $\pi$ is a $t$-coechidna of $M'$. Note that $M'$ is an elementary quotient of $M$, so $M$ is an elementary lift of $M'$ where none of the $2(s-1)$-element circuits of $M'$ are preserved in $M$. So the $(s,t)$-spike $M$ can be obtained from the $(s-1,t)$-spike $M'$ using the dual of \cref{cons:quotient}. \end{proof} \section*{Acknowledgements} Work for this project was begun during a visit, funded by the Heilbronn Institute for Mathematical Research, at the University of Bristol, United Kingdom. \sloppy

c2afd2fff6c82b26ccf0e57e0c56c472dcf41166

\section{Introduction} \label{sec:intro} The principal component analysis (PCA) model postulates that $N$ independent realisations of $K$-dimensional data ${\bf x}_i \in \mathbb{R}^K$ ($i=1,\ldots,N$) are described as \begin{equation} \label{eqn:mfa} {\bf x}_i = v_{i1} {\bf a}_1 + \cdots + v_{iJ} {\bf a}_J + \bm{\epsilon}_i = \left(\sum_{j=1}^J v_{ij} {\bf a}_j\right) + \bm{\epsilon}_i, \quad \bm{\epsilon} \sim {\rm N}_K ({\bf 0}_K, \sigma^2 {\bf I}_K), \end{equation} where $\{{\bf a}_1, \ldots, {\bf a}_J\}$ are the $J (< K)$ latent (unobserved) factor loadings with each factor loading ${\bf a}_j \in \mathbb{R}^K$, and ${\bf v}_j \sim {\rm N}({\bf 0}_N, {\bf I}_N)$ are the factor scores distributed as per the standard normal distribution. It is assumed that the residuals follow an isotropic zero mean normal distribution with the variance-covariance matrix $\sigma^2 {\bf I}_K$. Without loss of generality, we assume that the data has been centered to the mean. We can write this PCA model in matrix notation as follows \begin{equation} {\bf x}_i = {\bf A} {\bf v}_i + \bm{\varepsilon}, \quad {\bf A} \in \mathbb{R}^{K \times J}, \quad {\bf v}_i \in \mathbb{R}^J, \quad \bm{\varepsilon} \sim N_K({\bf 0}, \sigma^2 {\bf I}_K), \end{equation} where $i = (1,\ldots,N)$, ${\bf A} = ({\bf a}_1, \ldots, {\bf a}_J)$ and ${\bf V} = ({\bf v}_1, \ldots, {\bf v}_N) \in \mathbb{R}^{J \times N}$. Integrating out the factor scores yields the multivariate Gaussian marginal distribution of the data \begin{equation} \label{eqn:y:marginal} {\bf x}_i \sim N_K({\bf 0}_K, \bm{\Sigma}_0), \quad \bm{\Sigma} = {\bf A} {\bf A}^\prime + \sigma^2 {\bf I}_K . \end{equation} The probabilistic principal component model as described suffers from identifiability constraints~\citep{AndersonRubin56,LawleyMaxwell71}. A key reason for this is that the latent factors affect the likelihood function only through their outer product ${\bf A} {\bf A}^\prime$, which implies that an estimate of the factors can only be determined up to a rotation. To ensure the model is not over parametarised, the maximum number of latent factors to be estimated cannot exceed \begin{equation} J_\text{MAX} \leq K + \frac{1}{2} \left(1-\sqrt{8k + 1}\right) , \end{equation} see \citep{Beal03} (pp. 108) for details. For example, when $K = 4,5,6$ we have $J_\text{MAX} = 1, 2, 3$, respectively. \cite{TippingBishop99} showed how to interpret standard PCA model in a probabilistic framework and obtained maximum likelihood estimates of the latent factors and residual variance. This manuscript examines estimation of the probabilistic PCA model under the Bayesian minimum message length (MML) inductive inference framework. Although single and multiple factor analysis has been examined within the MML framework by \cite{WallaceFreeman92} and \cite{Wallace98} respectively, this manuscript departs from the earlier work in the following: \begin{itemize} \item We consider the marginal distribution of the data (\ref{eqn:y:marginal}) rather than the model (\ref{eqn:mfa}) analysed by \cite{Wallace98}. % \item Using polar decomposition, we write the factor load matrix ${\bf A}$ as a product of an orthogonal matrix and a diagonal matrix representing the direction and length of the loadings, respectively. Unlike the earlier MML approaches, we parameterize the orthogonal matrix via Givens rotations to explicitly capture orthogonality constraints. % \item We use matrix polar decomposition to develop a prior distribution for the latent factors ${\bf A}$ that is a product of a matrix variate Cauchy distribution and a uniform distribution over the corresponding Stiefel manifold. % \item We obtain analytic MML estimates of the parameters and find a polynomial whose roots yield the MML estimate of the residual variance. \end{itemize} \section{Maximum likelihood estimation} This section summarises the results of \cite{TippingBishop99}. The negative log-likelihood of the data under the probabilistic PCA model is \begin{equation} \ell (\bm{\theta} ) = \frac{N K}{2} \log (2\pi) + \frac{N}{2} \log |\bm{\Sigma} | + \frac{N}{2} {\rm tr} \left(\bm{\Sigma}^{-1} {\bf S}_x\right) \end{equation} where ${\bf S}_x = \frac{1}{N}\sum_i {\bf x}_i {\bf x}_i^\prime$ is the sample variance-covariance matrix. We have the observed data ${\bf X}$ and wish to estimate the number of latent factors $J$ and all parameters $\bm{\theta} = \left\{{\bf A},\sigma^2\right\}$. Differentiating the negative log-likelihood with respect to the factor loads \begin{align*} \partial \ell(\bm{\theta}) &= N {\rm tr} \, {\bf A}^\prime \bm{\Sigma}^{-1}(\partial {\bf A}) - N {\rm tr} \left( {\bf A}^\prime \bm{\Sigma}^{-1} {\bf S}_x \bm{\Sigma}^{-1} (\partial {\bf A}) \right) \end{align*} and setting the derivatives to zero we get \begin{align*} {\bf S}_x {\bf \Sigma}^{-1} {\bf A} &= {\bf A} \end{align*} Consider the singular value decomposition ${\bf A} = {\bf U} {\bf L} {\bf V}^\prime$, where ${\bf U} \in \mathbb{R}^{K\times J}$, ${\bf L} = {\rm diag}(l_1,\ldots,l_j)$ and ${\bf V} \in \mathbb{R}^{J\times J}$ is an orthogonal matrix. Noting that $\bm{\Sigma}^{-1} {\bf A} = {\bf U} {\bf L} ({\bf L}^2 + \sigma^2 {\bf I}_J )^{-1} {\bf V}^\prime$, we have \begin{align*} {\bf S}_x {\bf U} &= {\bf U} ({\bf L}^2 + \sigma^2 {\bf I}_J ) \\ {\bf S}_x {\bf u}_j &= (l_j^2 + \sigma^2 ) {\bf u}_j, \quad (j=1,\ldots,J), \end{align*} which is an example of the eigenvalue problem. That is, ${\bf U}$ is a $(K\times J)$ matrix whose columns are the top $J$ eigenvectors of the sample covariance matrix ${\bf S}_x$ corresponding to the $J$ largest eigenvalues \begin{eqnarray} \delta_j = \hat{l}^2_j + \sigma^2, \quad j=1,\ldots, J, \end{eqnarray} where $\hat{l}_j = (\delta_j - \sigma^2)^{\frac{1}{2}}$ is the $j$-th largest singular value of ${\bf A}$. Without loss of generality we assume that $\delta_1 > \delta_2 > \ldots > \delta_K > 0$ throughout the manuscript. This implies that the maximum likelihood estimate is \begin{eqnarray} \hat{\bf A}_\text{ML} = {\bf U} (\bm{\Delta} - \sigma^2 {\bf I}_J )^\frac{1}{2} {\bf O}, \quad \bm{\Delta} = {\rm diag}(\delta_1, \ldots, \delta_J) \end{eqnarray} where ${\bf O}$ is an arbitrary (orthogonal) rotation matrix and $\bm{\Delta}$ is a diagonal matrix with the $J$-th largest eigenvalues of ${\bf S}_x$. Substituting the maximum likelihood estimate of the factor loads into the negative log-likelihood we have \begin{align} \ell (\sigma, \hat{\bf A}_{\text{ML}} ) = \frac{N K}{2} \log (2\pi) + \frac{N}{2} \sum_{i=1}^J \log \delta_j + \frac{N (K-J)}{2} \log \sigma^2 + \frac{N J}{2} + \frac{N}{2 \sigma^2} \sum_{j=J+1}^K \delta_j . \end{align} The concentrated negative log-likelihood is minimised by \begin{align} \hat{\sigma}^2 = \frac{1}{K-J} \sum_{j=J+1}^K \delta_j \end{align} which is the sum of the (K-J) smallest eigenvalues of the sample variance-covariance matrix. \cite{TippingBishop99} show that these estimates minimise the negative log-likelihood and discuss other saddle points of the log-likelihood function. \section{Minimum message length analysis of the PCA model} Under suitable regularity conditions~\citep{Wallace05}) (pp. 226), the MML87 codelength approximation for data ${\bf x}$ is \begin{equation} \label{eqn:mml87:codelength} \mathcal{I}_{87}({\bf x}, \bm{\theta}) = \underbrace{-\log \pi(\bm{\theta}) + \frac{1}{2} \log \abs{{\bf J}_{\bm{\theta}}(\bm{\theta})} + \frac{P}{2} \log \kappa_P}_{\rm assertion} + \underbrace{\frac{P}{2} - \log p({\bf x}|\bm{\theta})}_{\rm detail} \end{equation} where $\pi_{\bm{\theta}}(\bm{\theta})$ is the prior distribution for the parameters $\bm{\theta}$, $\abs{{\bf J}_{\bm{\theta}}(\bm{\theta})}$ is the determinant of the expected Fisher information matrix, $p({\bf x}|\bm{\theta})$ is the likelihood function of the model and $\kappa_P$ is a quantization constant~\citep{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~\citep{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. \subsection{Orthogonality constraints} As discussed in Section~\ref{sec:intro}, it is well-known that the PCA model is not identifiable given the data. A key reason for this is that the latent vectors affect the likelihood only through their outer product $ {\bf A} {\bf A}^\prime = \sum_{j=1}^J {\bf a}_j {\bf a}^\prime_j. $ However, there are infinitely many sets of vectors that could generate the same matrix. To resolve this ambiguity, it is a convention to estimate the factor load vectors to be mutually orthogonal; that is, \begin{equation} {\bf A}^\prime {\bf A} = \bm{\alpha} = {\rm diag}(\alpha^2_1, \ldots, \alpha^2_J), \quad \alpha_j = ({\bf a}_j^\prime {\bf a}_j)^{\frac{1}{2}}, \quad (j=1,\ldots, J), \end{equation} where $\alpha_j$ denote the length of the $j$-th load vector. We enforce orthogonality constraints by parameterizing the matrix ${\bf A}$ in terms of Givens rotations ~\cite{PourzanjaniEtAl21}. Specifically, we can write ${\bf A}$ as \begin{align} {\bf A} &= \left[ R_{12}(\phi_{1,2}) \cdots R_{1,K}(\phi_{1,K}) R_{2,3}(\phi_{2,3}) \cdots R_{2,K}(\phi_{2,K}) \cdots R_{J,J+1}(\phi_{J,J+1}) \cdots R_{J,K}(\phi_{J,K}) {\bf I}_{K,J} \right] \bm{\alpha} \\ % &= {\bf R} \, \bm{\alpha} \end{align} where ${\bf I}_{K,J}$ is the first $J$ columns of a $K \times K$ identity matrix and $R_{i,j}(\phi_{i,j})$ is a $(K \times K)$ rotation matrix that is equal to the identity matrix except for the $(i, i)$ and $(j, j)$ positions which are replaced by $\cos (\phi_{i,j})$, and the $(i, j)$ and $(j, i)$ positions which are replaced by $-\sin (\phi_{i,j})$ and $\sin (\phi_{i,j})$ respectively. Thus ${\bf R} \in \mathbb{R}^{K \times J}$ and $\bm{\alpha} \in \mathbb{R}^{J \times J}$ denote the orientations and lengths of the factor load vectors, respectively. For example, when $K=J=2$, we have \begin{equation} {\bf A} = \left( \begin{array}{cc} \cos \left(\phi_{1,2}\right) & -\sin \left(\phi_{1,2}\right) \\ \sin \left(\phi_{1,2}\right) & \cos \left(\phi_{1,2}\right) \\ \end{array} \right) \left( \begin{array}{cc} 1 & 0 \\ 0 & 1 \\ \end{array} \right) \left( \begin{array}{c} \alpha _1 \\ \alpha _2 \\ \end{array} \right) . \end{equation} This parametarization explicitly takes into account that the estimated factor loads are pairwise orthogonal. The model parameters are now \begin{itemize} \item the lengths of the latent factors $\bm{\alpha}=(\alpha_1, \ldots, \alpha_J) \in \mathbb{R}_+^J$, \item the orientation of the factor load vectors as captured by the $D = J K- J(J+1)/2$ angles \begin{equation*} \bm{\phi} = (\phi_{1,2},\ldots,\phi_{1,K}, \phi_{2,3}, \ldots, \phi_{2,K},\ldots,\phi_{J,J+1},\ldots,\phi_{J,K}) \end{equation*} \item and the residual variance $\sigma^2 > 0$. \end{itemize} \subsection{Fisher information} Following lengthy and tedious algebra, the expected Fisher information matrix is seen to be block diagonal with determinant \begin{align} | {\bf J}(\bm{\alpha},\sigma,\bm{\phi}) | &= N^{P} | {\bf J}(\bm{\alpha},\sigma) | \, | {\bf J}(\bm{\phi}) | \\ | {\bf J}(\bm{\alpha},\sigma) | &= \frac{2^{J+1} (K-J)}{{\sigma^2}} \prod _{j=1}^J \frac{\alpha_j^2}{\left(\alpha_j^2+\sigma ^2\right)^2} \\ | {\bf J}(\bm{\phi}) | &= |J_{{\bf A} \to \bm{\phi}}|^2 \left(\prod _{i=1}^J \left(\frac{\alpha_i^4}{\sigma ^2}\right){}^{K-J} \frac{1}{\left(\alpha_i^2+\sigma ^2\right){}^{K-1}}\right) \prod _{j<k} \left(\alpha_j^2-\alpha_k^2\right)^2 \end{align} where $|J_{{\bf A} \to \bm{\phi}}|$ is the transformation of measure under the Givens representation~\citep{PourzanjaniEtAl21} \begin{equation*} |J_{{\bf A} \to \bm{\phi}}| = \prod_{i=1}^J \prod_{j=i+1}^K (\cos \phi_{i,j})^{j-i-1} , \end{equation*} and $P = (D + J + 1)$ is the total number of free parameters. Combining all the terms we have \begin{align} | {\bf J}(\bm{\alpha},\sigma,\bm{\phi}) | &= N^{P} \frac{\left(2^{J+1} (K-J)\right)}{{\sigma^2}} |J_{{\bf A} \to \bm{\phi}}|^2 \left(\prod _{i=1}^J \frac{\alpha_i^2}{\left(\alpha_i^2+\sigma ^2\right)^2} \left(\frac{\alpha_i^4}{\sigma ^2}\right){}^{K-J} \frac{1}{\left(\alpha_i^2+\sigma ^2\right){}^{K-1}}\right) \prod _{j<k} \left(\alpha_j^2-\alpha_k^2\right)^2 \nonumber \\ &= \frac{N^{P} 2^{J+1} (K-J) |J_{{\bf A} \to \bm{\phi}}|^2}{{\sigma^{2(J(K-J)+1)}}} \prod_{i=1}^J \frac{\alpha_i^{4(K-J)+2}}{\left(\alpha_i^2+\sigma ^2\right)^{K+1}} \prod _{j<k}^J \left(\alpha_j^2-\alpha_k^2\right)^2 \end{align} \subsection{Prior information} \label{sec:mfa:priors} The prior distribution for the residual standard deviation $\bm{\Sigma}^{\frac{1}{2}}$ is chosen to be the scale-invariant density \begin{equation} \label{eqn:prior:sigma} \pi_\sigma(\sigma) \propto \sigma^{-1}, \end{equation} defined over some suitable range. The prior distribution for the matrix of factor loads ${\bf A} \in \mathbb{R}^{K \times J}$ is not immediately obvious as the estimates of the factor loads are enforced to be mutually orthogonal. Ideally, we would like a prior distribution that is uniform over the direction of the $J$ factors, while the distribution of the lengths of these vectors should be heavy tailed to allow for a wide range of lengths. We follow a similar approach to \cite{Wallace98} and assume a prior distribution over the unknown latent vectors that is then transformed to account for the estimated factors being mutally orthogonal. Further, as in~\cite{Wallace98}, we shall consider a prior distribution for the scaled factors \begin{align*} {\bf b}_j = \left( \frac{{\bf a}_j}{\sigma} \right), \quad \beta_j = ({\bf b}_j^\prime {\bf b}_j)^{\frac{1}{2}}, \quad (j=1,\ldots,J) \end{align*} where the residual variance is used as a default scale. Let $\tilde{\bf B} \in \mathbb{R}^{K \times J}$ denote the matrix containing the $J$ true (unknown) scaled factors. We assume $\tilde{\bf B}$ to follow a matrix variate Cauchy distribution~\citep{BandekarDaya03} with probability density function \begin{equation} \pi_{\tilde{A}}(\tilde{\bf B}) = \frac{\Gamma_K((K+J)/2)}{\pi^{K J / 2} \Gamma_K(K/2)} {\rm det}({\bf I}_K + \tilde{\bf B} \tilde{\bf B}^\prime)^{-(K+J)/2}. \end{equation} This is a reasonable choice as the matrix variate Cauchy is spherically symmetric and has appropriately heavy tails. Further, our choice of the prior distribution implies that $\tilde{\bf B}^\prime \in \mathbb{R}^{J \times K}$ follows a matrix variate Cauchy distribution with density \begin{equation} \pi_{\tilde{B}^\prime}(\tilde{\bf B}^\prime) = \frac{\Gamma_J((K+J)/2)}{\pi^{K J / 2} \Gamma_J(J/2)} {\rm det}({\bf I}_J + \tilde{\bf B}^\prime \tilde{\bf B})^{-(K+J)/2} . \end{equation} Consider the unique matrix polar decomposition \begin{equation} \tilde{\bf B}^\prime = {\bf W}_B^{\frac{1}{2}} \, {\bf H}_B, \quad {\bf W}_B = \tilde{\bf B}^\prime \tilde{\bf B}, \quad {\bf H}_B = (\tilde{\bf B}^\prime \tilde{\bf B})^{-\frac{1}{2}} \tilde{\bf B}^\prime \end{equation} where ${\bf H}_B$ is defined over the Stiefel manifold $\mathcal{V}_J (\mathbb{R}^K )$ and ${\bf W}_B$ is a symmetric positive definite matrix. We may think of the matrix ${\bf H}_B$ as the orientation matrix, while the matrix ${\bf W}_B$ determines the squared lengths of the true scaled latent vectors. If $\tilde{\bf B}^\prime$ follows a matrix variate Cauchy distribution, it is known that ${\bf H}_B$ is distributed uniformly over the Stiefel manifold with density function~\citep{BandekarDaya03}: \begin{equation} \pi_H({\bf H}_B) = \frac{1}{{\rm Vol}( \mathcal{V}_{J} (\mathbb{R}^K ) )}, \quad {\rm Vol}(\mathcal{V}_J (\mathbb{R}^K ) ) = \frac{ 2^J \pi^{K J / 2} } { \Gamma_J (K/2)} , \end{equation} where $\Gamma_p( y )$ is the multivariate Gamma function \begin{align*} \Gamma_J(y) = \pi^{J(J-1)/4} \prod_{j=1}^J \Gamma(y + (1-j)/2) . \end{align*} Further, the random variable ${\bf W}_B$ representing the squared lengths of the true scaled factors is independent of ${\bf H}_B$ with probability density function~\cite{BandekarDaya03} \begin{eqnarray} \pi_W({\bf W}_B) &\propto& {\rm det}( {\bf W}_B )^{(K - J - 1)/2} {\rm det}({\bf I}_K + {\bf W}_B)^{-(K+J)/2} \end{eqnarray} which is a matrix variate beta type II distribution ${\bf W}_B \sim B_J^{II}(K/2, J/2)$ with parameters $(K/2, J/2)$ (see~\cite{GuptaNagar99}, pp. 166, for further details); this is also known as the matrix variate $F$ distribution. Recall that the estimated (scaled) factor load vectors obey \begin{equation} {\bf S}_B = \tilde{\bf B}\tilde{\bf B}^\prime = \sum_{j=1}^J \tilde{\bm{\beta}}_j \tilde{\bm{\beta}}_j^\prime = \sum_{j=1}^J \bm{\beta}_j \bm{\beta}_j^\prime = {\bf B} {\bf B}^\prime , \quad \bm{\beta}_j^\prime \bm{\beta}_{k \neq j} = 0 . \end{equation} where ${\bf S}_B$ is a $(K \times K)$ symmetric matrix of rank $J$. The distribution of the squared scaled lengths $\beta_j^2$ of the estimated latent vectors can then be taken as the joint distribution of the $J$ eigenvalues of ${\bf S}_B$ which is (see Appendix~F) \begin{align} \pi_{\bm{\beta}^2}(\beta_1^2,\ldots,\alpha_J^2) &= \frac{\pi^{J^2/2}}{\Gamma_J(J/2) \mathcal{B}_J(K/2,J/2)}\prod_{j=1}^J \beta_j^{(K-J-1)} (1 + \beta_j^2)^{-(K+J)/2} \prod_{j<k}^J | \beta_j^2 - \beta_k^2| , \nonumber \end{align} where $\mathcal{B}_p(a,b)$ denote the multivariate beta function \begin{align*} \mathcal{B}_J(a,b) = \frac{\Gamma_J(a) \Gamma_J(b)}{\Gamma_J(a+b)}. \end{align*} This implies that the prior distribution of the lengths of the scaled latent factors is \begin{align} \pi_{\bm{\beta}}(\beta_1,\ldots,\beta_J) &= \frac{2^J \pi^{J^2/2}}{\Gamma_J(J/2) \mathcal{B}_J(K/2,J/2)}\prod_{j=1}^J \beta_j^{(K-J)} (1 + \beta_j^2)^{-(K+J)/2} \prod_{j<k}^J | \beta_j^2 - \beta_k^2| . \end{align} Finally, the prior distribution for the lengths of the (unscaled) latent factors is \begin{align} \pi_{\bm{\alpha}}(\alpha_1,\ldots,\alpha_J) &= \frac{2^J \pi^{J^2/2} \sigma ^{J^2}}{\Gamma_J(J/2) \mathcal{B}_J(K/2,J/2) }\prod_{j=1}^J \alpha_j^{(K-J)} (\sigma^2 + \alpha_j^2)^{-(K+J)/2} \prod_{j<k}^J | \alpha_j^2 - \alpha_k^2| . \label{eqn:prior:len:b} \end{align} The complete prior distribution over all model parameters is \begin{align} \pi(\bm{\alpha},\sigma,\bm{\phi}) = \pi_\sigma(\bm{\Sigma}^{\frac{1}{2}}) \, \pi_{\bm{\alpha}}(\alpha_1,\ldots,\alpha_J) |J_{{\bf A} \to \bm{\phi}}| \, J! , \end{align} where the term $J!$ is included because the labelling of the latent factors is arbitrary and $|J_{{\bf A} \to \bm{\phi}}|$ is the transformation of measure from the matrix parametrization ${\bf A}$ to the orthogonality-preserving parameterization based on Givens rotations. \subsection{Codelength} Omitting constants, the \cite{WallaceFreeman87} codelength for the probabilistic PCA model is \begin{align*} \mathcal{I} &\propto \frac{N}{2} \log |\bm{\Sigma}| + \frac{N}{2} {\rm tr} \left( \bm{\Sigma}^{-1} {\bf S}_x\right) - K J \log (\sigma ) + \frac{1}{2} \sum_{j=1}^J \log \left[\alpha_j ^{2 (K-J+1)} \left(\alpha_j ^2+\sigma^2\right)^{(J-1)} \right] \end{align*} where ${\bf S}_x = \frac{1}{N}\sum_i {\bf x}_i {\bf x}_i^\prime$ is the sample variance-covariance matrix. To obtain MML estimates, we start with the Langrangian of the factor orientations \begin{align*} \psi({\bf R}) &= \log |\bm{\Sigma} | + {\rm tr} \left(\bm{\Sigma}^{-1} {\bf S}_x\right) - {\rm tr} {\bf L} ({\bf R}^\prime {\bf R} - I) , \end{align*} where ${\bf L}$ is a $J \times J$ symmetric matrix of Lagrange multipliers. Clearly, minimising $\psi({\bf R})$ is equivalent to minimising the codelength with respect to ${\bf R}$. The first differential of the Lagrangian is \begin{align*} \partial \psi({\bf R}) &= 2 \text{tr} \left[\bm{\alpha} {\bf A}^\prime \left( \bm{\Sigma}^{-1} - \bm{\Sigma}^{-1} {\bf S}_x \bm{\Sigma}^{-1}\right) (d{\bf R})\right] - 2 \text{tr} \left( {\bf L} {\bf R}^\prime (d{\bf R}) \right) , \end{align*} which implies the following first order conditions \begin{align} \bm{\alpha} {\bf A}^\prime \left( \bm{\Sigma}^{-1} - \bm{\Sigma}^{-1} {\bf S}_x \bm{\Sigma}^{-1}\right) &= {\bf 0} \label{eqn:lag:cond1}\\ {\bf L} {\bf R}^\prime &= {\bf 0} \label{eqn:lag:cond2} \\ {\bf R}^\prime {\bf R} &= {\bf I}_J \label{eqn:lag:cond3} \end{align} From (\ref{eqn:lag:cond2}) we have that ${\bf L} = {\bf 0}$ and from (\ref{eqn:lag:cond1}) \begin{align*} {\bf S}_x {\bf R} &= {\bf R} \, \text{diag} \left( \sigma^2 + \alpha_1^2, \ldots, \sigma^2 + \alpha_J^2\right) \\ {\bf S}_x {\bf r}_j &= {\bf r}_j (\sigma^2 + \alpha_j^2), \quad (j=1,\ldots,J) . \end{align*} We see that, at the codelength minimum, the MML estimate of the factor orientations is the matrix ${\bf R}$ whose columns are the top $J$ eigenvectors of the variance--covariance matrix ${\bf S}_x$ with eigenvalues $\delta_j = (\sigma^2 + \alpha_j^2)$, for $j = 1,\ldots J$. This is identical to the corresponding maximum likelihood estimate. The concentrated codelength, as a function of $\sigma^2$ is \begin{align} \mathcal{I}(\sigma) &\propto \frac{N}{2} \log \left( (\sigma^2)^{K-J} \prod_{j=1}^J (\alpha_j^2 + \sigma^2) \right) + \frac{N}{2 \sigma^2} \left(\sum_{j=1}^K \delta_j\right) - \frac{N}{2 \sigma^2} \sum_{j=1}^J \alpha_j^2 \nonumber \\ & \quad - K J \log (\sigma ) + \frac{1}{2} \sum_{j=1}^J \log \left[\alpha_j ^{2 (K-J+1)} \left(\alpha_j ^2+\sigma^2\right)^{(J-1)} \right] \nonumber \\ % &= \frac{N (K-J) - K J}{2} \log \left( \sigma^2 \right) + \frac{N}{2 \sigma^2} \left(\sum_{j=1}^K \delta_j\right) - \frac{N}{2 \sigma^2} \sum_{j=1}^J (\delta_j - \sigma^2) + \frac{(K-J+1)}{2} \sum_{j=1}^J \log \left(\delta_j - \sigma^2 \right) \label{eqn:msglen:conc} \end{align} The following theorem shows how to obtain the MML estimate of the residual variance from the concentrated message length. \begin{thm} \label{thm:mmlest} Let $\tau = \sigma^2$. The concentrated codelength (\ref{eqn:msglen:conc}) has $(J+1)$ stationary points whose location are the roots of the $n=(J+1)$-degree polynomial \begin{equation*} a_n \tau^n + a_{n-1} \tau^{n-1} + \cdots + a_1 \tau + a_0, \quad (0 < \tau < \delta_J) \end{equation*} with coefficients \begin{align*} a_0 &= - \hat{\tau}_{{\rm ML}} e_J, \\ a_j &= (-1)^{j+1} \left[ \hat{\tau}_{{\rm ML}} \, e_{J-j} + \left(1-\frac{(j-1) (J-1)}{N (K-J)}-\frac{K (J-j+1)}{N (K-J)}\right) e_{J-j+1} \right], \quad (1 \leq j \leq J) \\ % a_n &= (-1)^J \left[1 - \frac{J(J-1)}{N(K-J)}\right] \\ \end{align*} where the terms $e_t$ refer to elementary symmetric polynomials $e_t(\delta_1,\ldots,\delta_J)$ in $J$ variables $(\delta_1, \ldots, \delta_J)$. For example, for $J=3$, we have the following four elementary symmetric polynomials \begin{align*} e_0(\delta_1,\delta_2,\delta_3) &= 1 \\ e_1(\delta_1,\delta_2,\delta_3) &= \delta_1 + \delta_2 + \delta_3 \\ e_2(\delta_1,\delta_2,\delta_3) &= \delta_1 \delta_2 + \delta_1 \delta_3 + \delta_2 \delta_3 \\ e_3(\delta_1,\delta_2,\delta_3) &= \delta_1 \delta_2 \delta_3 . \end{align*} MML estimate of the residual variance is the stationary point in the domain $0 < \tau < \delta_J$ that yields the shortest codelength. MML estimates of the factor lengths can be obtained from $\hat{\alpha}_j = (\delta_j - \hat{\sigma}^2)^{\frac{1}{2}}$ for all $j = 1,\ldots,J$. % \end{thm} Importantly, all $(J+1)$ elementary symmetric polynomials can be efficiently computed in $O(J \log^2 J)$ time by Fast Fourier Transform (FFT) polynomial multiplication, following an algorithm widely attributed to Ben-Or. {\bf Example 1:} For a single latent factor ($J=1$), the stationary points of the concentrated codelength are the roots of the quadratic polynomial in $\tau$: \begin{align} - \delta_1 \hat{\tau}_{\text{ML}} + \left(\hat{\tau}_{\text{ML}} + c \delta_1 \right) \tau - \tau ^2=0, \quad c= 1-\frac{K}{N (K-1) }, \end{align} given by \begin{align*} \frac{1}{2}\left( \hat{\tau}_{\text{ML}} + c \, \delta_1 \pm \Delta^{\frac{1}{2}}\right), \quad \Delta = c^2 \delta _1^2+2 (c-2) \delta _1 \hat{\tau}_{\text{ML}}+\hat{\tau} _{\text{ML}}^2 . \end{align*} The quadratic polynomial has no real roots if: \begin{align} -\frac{c+2 \sqrt{1-c}-2}{c^2}<\frac{\delta_1}{\hat{\tau}_{\text{ML}}}<\frac{-c+2 \sqrt{1-c}+2}{c^2} \label{eqn:mml:1pc} \end{align} in which case the MML solution is the uncorrelated multivariate Gaussian model. For example, when $N = 25$ and $K = 4$, the quadratic will have no real roots if \begin{align*} 0.219 < \frac{\delta_1}{\delta_2 + \delta_3 + \delta_4} < 0.564 . \end{align*} In the limit as $N \to \infty$, we have \begin{align*} \lim_{N \to \infty} \pm \frac{c+2 \sqrt{1-c}-2}{c^2} = 1 . \end{align*} so that both the lower and upper bound in (\ref{eqn:mml:1pc}) approach 1. {\bf Example 2:} For a PCA model with two latent factors ($J=2$), the stationary points of the concentrated codelength are the roots of the cubic polynomial in $\tau$: \begin{align*} - \delta_1 \delta_2 \hat{\tau}_{\text{ML}} + \left( (\delta_1 + \delta_2) \hat{\tau}_{\text{ML}} + c_1 \delta_1 \delta_2 \right)\tau- \left(\hat{\tau}_{\text{ML}} + \left(c_0 + c_1\right) (\delta_1 + \delta_2)\right)\tau ^2 + (2 c_0+c_1) \tau ^3 \end{align*} where the constants \begin{align*} c_0 = \frac{K-1}{N(K-2) }, \quad c_1 = 1-\frac{2 K}{N(K-2) } , \end{align*} depend only on $N$ and $K$. \section{Experiments} \subsection{Parameter estimation} \label{sec:exp:pest} \begin{table*}[tbph] \scriptsize \begin{center} \begin{tabular}{ccccccccccccccc} \toprule $N$ & $K$ & $J$ & \multicolumn{2}{c}{$S_1$} & & \multicolumn{2}{c}{$S_2$} & & \multicolumn{2}{c}{KL Divergence} \\ & & & MLE & MML87 & ~ & MLE & MML87 & ~ & MLE & MML87 \\ \cmidrule{1-11} \multirow{9}{*}{25} & \multirow{3}{*}{5} & 1 & -0.072 & {\bf -0.023} & & 0.011 & {\bf 0.007} & & 0.167 & {\bf 0.130} \\ & & 2 & -0.093 & {\bf 0.001} & & 0.021 & {\bf 0.009} & & 0.279 & {\bf 0.179} \\ & \multirow{3}{*}{8} & 1 & -0.069 & {\bf -0.030} & & 0.008 & {\bf 0.004} & & 0.244 & {\bf 0.194} \\ & & 2 & -0.120 & {\bf -0.021} & & 0.018 & {\bf 0.005} & & 0.463 & {\bf 0.285} \\ & & 4 & -0.103 & {\bf 0.038} & & 0.020 & {\bf 0.007} & & 0.647 & {\bf 0.353} \\ & \multirow{3}{*}{16} & 1 & -0.066 & {\bf -0.034} & & 0.006 & {\bf 0.003} & & 0.407 & {\bf 0.332} \\ & & 2 & -0.111 & {\bf -0.040} & & 0.014 & {\bf 0.003} & & 0.818 & {\bf 0.557} \\ & & 4 & -0.207 & {\bf -0.012} & & 0.045 & {\bf 0.003} & & 1.809 & {\bf 0.764} \\ \vspace{-2mm} \\ \cmidrule{2-11} \vspace{-2mm} \\ \multirow{9}{*}{50} & \multirow{3}{*}{5} & 1 & -0.035 & {\bf -0.009} & & 0.004 & {\bf 0.003} & & 0.074 & {\bf 0.065} \\ & & 2 & -0.056 & {\bf 0.006} & & 0.008 & {\bf 0.004} & & 0.127 & {\bf 0.096} \\ & \multirow{3}{*}{8} & 1 & -0.034 & {\bf -0.012} & & 0.003 & {\bf 0.002} & & 0.113 & {\bf 0.100} \\ & & 2 & -0.060 & {\bf -0.008} & & 0.005 & {\bf 0.002} & & 0.208 & {\bf 0.159} \\ & & 4 & -0.069 & {\bf 0.034} & & 0.009 & {\bf 0.004} & & 0.324 & {\bf 0.209} \\ & \multirow{3}{*}{16} & 1 & -0.033 & {\bf -0.015} & & 0.002 & {\bf 0.001} & & 0.209 & {\bf 0.186} \\ & & 2 & -0.056 & {\bf -0.018} & & 0.004 & {\bf 0.001} & & 0.401 & {\bf 0.323} \\ & & 4 & -0.107 & {\bf -0.010} & & 0.012 & {\bf 0.001} & & 0.785 & {\bf 0.489} \\ \vspace{-2mm} \\ \cmidrule{2-11} \vspace{-2mm} \\ \multirow{9}{*}{100} & \multirow{3}{*}{5} & 1 & -0.017 & {\bf -0.004} & & 0.002 & {\bf 0.001} & & 0.033 & {\bf 0.031} \\ & & 2 & -0.031 & {\bf 0.004} & & 0.003 & {\bf 0.002} & & 0.058 & {\bf 0.049} \\ & \multirow{3}{*}{8} & 1 & -0.016 & {\bf -0.005} & & 0.001 & {\bf 0.001} & & 0.051 & {\bf 0.048} \\ & & 2 & -0.029 & {\bf -0.002} & & 0.002 & {\bf 0.001} & & 0.095 & {\bf 0.082} \\ & & 4 & -0.049 & {\bf 0.021} & & 0.004 & {\bf 0.002} & & 0.159 & {\bf 0.118} \\ & \multirow{3}{*}{16} & 1 & -0.016 & {\bf -0.006} & & 0.001 & {\bf 0.000} & & 0.100 & {\bf 0.094} \\ & & 2 & -0.027 & {\bf -0.007} & & 0.001 & {\bf 0.000} & & 0.193 & {\bf 0.171} \\ & & 4 & -0.054 & {\bf -0.004} & & 0.003 & {\bf 0.000} & & 0.367 & {\bf 0.283} \\ \vspace{-3mm} \\ \bottomrule \vspace{+1mm} \end{tabular} \caption{Performance metrics for maximum likelihood (MLE) and MML87 estimates of residual variance $\sigma^2$ computed over $10^5$ simulations. \label{tab:results:pest}} \end{center} \end{table*} This section compares the newly derived MML parameter estimates for the probabilistic PCA model to the standard approach based on the maximum likelihood estimator. Since MML and maximum likelihood estimates of the the factor lengths (for a given $\sigma^2$) and factor orientations are identical, the key difference between to two approaches is in the estimation of the residual variance. Our simulation experiments are loosely based on Section~6 in \citep{Wallace98}. We conducted $10^5$ simulations for each combination of the sample size $N \in \{25,50,100\}$, the dimensionality of the data $K \in \{5, 8, 16\}$ and the number of estimated latent factors $J \in \{1,2,4\}$. As both maximum likelihood and MML are scale invariant, the residual variance was set to $\sigma^2 = 1$ and the factors lengths were $\alpha_j = 1$ ($j=1,\ldots,J$) for each simulation run, without loss of generality. The factor directions were randomly sampled from a unit $K$-sphere. We used the three performance metrics discussed in \citep{Wallace98} to evaluate the estimators: \begin{align*} S_1 = \log \hat{\sigma}_i , \quad S_2 = (\log \hat{\sigma}_i)^2 , \end{align*} and the Kullback--Leibler (KL) divergence~\citep{KullbackLeibler51} between two multivariate Gaussian distributions \begin{equation*} \text{KL}(\bm{\Sigma}_0, \bm{\Sigma}_1) = \frac{1}{2} \left(\text{tr}\left( \bm{\Sigma}^{-1}_1 \bm{\Sigma}_0\right) + \log \left( \frac{ \bm{|\Sigma}_1|}{|\bm{\Sigma}_0|}\right) - K \right) . \end{equation*} The first metric $S_1$ is a measure of bias, while $S_2$ measures error in any direction. Both $S_1$ and $S_2$ are zero for exact estimates as the true residual variance was $\sigma^2 = 1$ in all experiments. The error measures were specifically chosen as they do not depend on the number of estimated latent vectors $J$. Simulation results averaged over $10^5$ iterations are shown in Table~\ref{tab:results:pest}. The MML estimate of the residual variance was found to be superior to the usual maximum likelihood estimate for all tested combinations of sample sizes, data dimensionality and the number of latent vectors. Maximum likelihood appeared to underestimate the residual variance more strongly compared to the minimum message length estimate. The differences in the performances of the two estimates were most pronounced when the sample size and data dimensionality was small ($N \leq 50$, $K = 5$). \subsection{Model selection} We have also compared the performance of MML model selection against the highly popular Bayesian information criterion (BIC) and Laplace's method for approximating the marginal distribution of the data~\citep{Minka00}, referred to as `Bayes' henceforth. Using numerical experiments, \cite{Minka00} demonstrated that approximating Bayesian evidence is superior to methods like cross validation. The simulation setup was identical to Section~\ref{sec:exp:pest} except the sample size was $N \in \{50,100\}$, the dimensionality of the data $K = 10$ and the number of estimated latent factors $J \in \{1,2,4\}$. Simulation results, averaged over $10^5$ iterations, are shown in Table~\ref{tab:results:msel}. As expected, both MML and the Bayes method have similar performance and both improve significantly over the popular BIC criterion. \begin{table*}[tbph] \scriptsize \begin{center} \begin{tabular}{cccccccc} \toprule $N$ & $J$ & Method & KL Divergence & & \multicolumn{3}{c}{Model Selection (\%)} \\ & & & &~ & $<J$ & $=J$ & $>J$ \\ \cmidrule{1-8} \multirow{9}{*}{50} & \multirow{3}{*}{1} & MML & 0.060 & & -- & 99.55 & 0.45\\ & & BIC & 0.063 & & -- & 100.00 & 0.00\\ & & Bayes & 0.066 & & -- & 97.19 & 2.81\\ & \multirow{3}{*}{2} & MML & 0.126 & & 70.40 & 28.32 & 1.27\\ & & BIC & 0.145 & & 96.40 & 3.60 & 0.00\\ & & Bayes & 0.129 & & 52.52 & 45.17 & 2.31\\ & \multirow{3}{*}{4} & MML & 0.198 & & 81.00 & 5.78 & 13.22\\ & & BIC & 0.257 & & 99.99 & 0.01 & 0.00\\ & & Bayes & 0.209 & & 91.09 & 7.92 & 0.98\\ \vspace{-2mm} \\ \cmidrule{2-8} \vspace{-2mm} \\ \multirow{9}{*}{100} & \multirow{3}{*}{1} & MML & 0.028 & & -- & 99.74 & 0.26\\ & & BIC & 0.029 & & -- & 100.00 & 0.00\\ & & Bayes & 0.030 & & -- & 97.86 & 2.14\\ & \multirow{3}{*}{2} & MML & 0.062 & & 30.47 & 68.49 & 1.05\\ & & BIC & 0.094 & & 74.75 & 25.25 & 0.00\\ & & Bayes & 0.060 & & 17.96 & 79.84 & 2.20\\ & \multirow{3}{*}{4} & MML & 0.096 & & 66.32 & 22.70 & 10.98\\ & & BIC & 0.166 & & 99.48 & 0.52 & 0.00\\ & & Bayes & 0.103 & & 72.71 & 25.93 & 1.36\\ \vspace{-3mm} \\ \bottomrule \vspace{+1mm} \end{tabular} \caption{Model selection simulation results for minimum message length (MML), Bayesian information criterion (BIC) and Laplace's method for estimating Bayesian evidence averaged over $10^5$ simulations. In all experiments, data dimensionality was $K=10$. \label{tab:results:msel}} \end{center} \end{table*}

c9344dbc056f182728ad69a78a4c3860808c41ba

\section{Leptoproduction of pseudoscalar mesons} In this paper, we analyze pseudoscalar meson electroproduction ($\pi^0$, $\eta$) on the basis of handbag approach. Its essential ingredients are the Generalized Parton Distributions (GPDs) that were proposed in Refs \cite{muell,ji1,rad1} and provide an extensive information on the hadron structure. GPDs are complicated nonperturbative objects which depend on $x_{B}$ -the momentum fraction of proton carried by parton, $\xi$- skewness and $t$- momentum transfer. GPDs are connected in the forward limit with Parton Distribution Functions (PDFs), they contain information about hadron form factors and the parton angular momentum \cite{ji2}. They give information on 3D structure of the hadrons, see e.g. \cite{3d}. More details on GPDs can be found e.g. in \cite{dvmp1,dvmp2,Diehl,Radyush}. GPDs were proposed to investigate exclusive reactions such as deeply virtual Compton scattering (DVCS) \cite{ji2,rad2,dvcs}, time-like Compton scattering (TCS) \cite{Berger:2001xd, Pire:2008ea, Mueller:2012sma} and deeply virtual meson production (DVMP) \cite{dvmp1,dvmp2}. Such processes at large photon vituality $Q^2$ can be factorized into the hard subprocess that can be calculated perturbatively and the GPDs \cite{ji2,rad2,dvcs}. Generally, this factorization was proved in the leading-twist amplitude with longitudinally polarized photon. This factorization formulae is valid up to power corrections of the order $1/Q$ to the leading twist results which are unknown. Study of exclusive meson electroproduction is one of the effective way to access GPDs. Experimental study of $\pi^0$ production was performed by CLAS \cite{clas} and COMPASS \cite{compass}. For $\eta$ production CLAS results are available at \cite{cleta}. These experimental data can be adopted to constrain the models of GPDs. On the other hand, Electron-Ion Colliders (EICs) are the next generation collider to study of nucleon structure. USA and China both design to build the EICs in future \cite{eic,eicc,Chen:2020ijn}. The GPDs property is one of the most important aims to investigate for the EICs \cite{Chavez:2021koz}. Theoretical investigation of DVMP in terms of GPDs is based on the handbag approach where, as mentioned before, the amplitude is factorized into the hard subprocess and GPDs \cite{ji1, rad1, ji2, rad2} see Fig.~1. This amplitude has an ingredient the non-perturbative meson Distribution Amplitudes, which probe the two-quark component of the meson wave functions. One of the popular way to construct GPDs is adopting so called Double Distribution (DD) \cite{mus99} which construct $\xi$ dependencies of GPDs and connect them with PDFs, modified by $t$- dependent term. The handbag approach with DD form of GPDs was successfully applied to the light vector mesons (VM) leptoproduction at high photon virtualities $Q^2$ \cite{gk06} and the pseudoscalar mesons (PM) leptoproduction \cite{gk09}. In this work, we continue our previous study of $\pi^0$ production \cite{gxc22} at the kinematics for EIC in China (EicC) based on the handbag approach. As it was shown in \cite{gk09} the leading twist longitudinal cross section $\sigma_L$ is rather small with respect to the predominant contribution determined by transversely polarized photons $\sigma_T$. This result was proved experimentally by JLab Hall A collaboration \cite{halla}. The transversity dominance $\sigma_T \gg \sigma_L$ is confirmed in \cite{gxc22} at all EicC energy ranges for $\pi^0$ production. This paper is organized as follows. In section 2, we discuss the contributions to the meson production amplitudes from the transversity GPDs $H_T$, $\bar E_T$ \cite{gk11}. Within the handbag approach, the transversity GPDs together with the twist-3 meson wave functions \cite{gk11} contribute to the amplitudes with transversely polarized photons which produce transverse cross sections $\sigma_T$. They give essential contribution to the cross sections that are consistent with experiment \cite{clas,compass}. In section 3, we consider two models for transversity GPDs that give results for the cross sections of the $\pi^0$ and $\eta$ leptoproduction that are consistent with experiment at CLAS and COMPASS energies \cite{clas,compass,cleta}. Predictions for $\eta$ cross section at EicC energies are done. Later on we extract information on the transversity GPDs contribution for these reactions. We discuss possibility to perform $u,d$ flavor separation for transversity GPDs $H_T$ and $\bar E_T$ using $\pi^0$ and $\eta$ cross sections \cite{kr-conv, kubar}. Finally, We give some discussion and conclusions in section 4. \section{Handbag approach. Properties of meson production amplitudes} The process amplitude in the handbag approach is depicted in Fig.~\ref{kt_h}. In handbag approach, the meson photoproduction amplitude is factorized into a hard subprocess amplitude ${\cal H}$ which is shown in the upper part of Fig.~\ref{kt_h} and GPDs $F$ which includes information on the hadron structure at sufficiently high $Q^2$. For the leading twist amplitude, with longitudinally polarized photons, its factorization has been proved \cite{ji1,rad1}. \begin{figure} \begin{center} \includegraphics[width=10cm]{GPDs-PM.eps} \end{center} \caption{The handbag diagram for the meson electroproduction off proton. Parton helicities for transversity GPDs contribution are shown.} \label{kt_h} \end{figure} In what follows, we consider the twist-3 contributions from transversity GPDs $H_T$ and $\bar {E}_T$ as well. Factorization for these twist-3 amplitudes is an assumption now. However factorization models give results which are consistent with experiment \cite{gk11}. In handbag method, the subprocess amplitude is calculated employing the modified perturbative approach (MPA) \cite{sterman}. The power $k_\perp^2/Q^2$ correction is considered in the propagators of the hard subprocess ${\cal H}$ together with the nonperturbative $\vk$-dependent meson wave functions \cite{koerner}. The gluonic corrections are regarded as the form of the Sudakov factors. Resummation of the Sudakov factor can be done in the impact parameter space \cite{sterman}. The unpolarized $e p\to e(\pi^0, \eta)p$ cross section can be decomposed into a number of partial cross sections which are expressed in terms of the $\gamma^* p\to (\pi^0, \eta) p$ helicity amplitudes. They have the following forms \begin{eqnarray}\label{ds} \frac{d\sigma_L}{dt} &=& \frac{1}{\kappa} \big(\vert{M}_{0+,0+}\vert^2 +\vert{M}_{0-,0+}\vert^2\big)\,,\nonumber\\ [0.3em] \frac{d\sigma_T}{dt} &=& \frac{1}{2 \kappa}(\vert{ M}_{0-,++}\vert^2 +2 \vert{M}_{0+,++}\vert^2)\,, \nonumber\\[0.3em] \frac{d\sigma_{LT}}{dt} &=& -\frac{1}{\sqrt{2} \kappa} {\rm Re}\Big[{M^*}_{0-,++}{M}_{0-,0+}\Big] \,,\nonumber\\ [0.3em] \frac{d\sigma_{TT}}{dt} &=& -\frac{1}{\kappa} \vert{M}_{0+,++}\vert^2\,. \end{eqnarray} Here $\kappa$ is the phase space factor, it reads \begin{equation}\label{kap} \kappa=16 \pi (W^2-m^2)\sqrt{\Lambda(W^2,-Q^2,m^2)}. \end{equation} $\Lambda(x, y, z)$ is expressed as $\Lambda(x, y, z) = (x^2 + y^2 + z^2) - 2xy - 2xz - 2 yz$. $\sigma_{LT}$ is the interference contributions of the longitudinal and transverse amplitudes and $\sigma_{TT}$ contains transverse amplitudes only. The leading twist amplitudes ${M}_{0-,0+} $ and ${M}_{0+,0+} $ are listed in our previous paper \cite{gxc22}. The transversity amplitudes that are essential in our study can be written in terms of convolutions as \begin{eqnarray}\label{conv} {M}_{0-,++}&=& \frac{e_0}{Q}\sqrt{1-\xi^2}\langle {H_T}\rangle,\nonumber\\ {M}_{0+,++}&=& -\frac{e_0}{Q}\frac{\sqrt{-t'}}{4m}\langle {\bar E_T}\rangle, \end{eqnarray} where $e_0 = \sqrt{4\pi \alpha}$ with $\alpha $ is the electronic-magnetic coupling. The other variables are defined as \begin{equation} \xi=\frac{x_B}{2-x_B}(1+\frac{m_P^2}{Q^2}),\;\; t'=t-t_0,\;\; t_0=-\frac{4 m^2\xi^2}{1-\xi^2}. \end{equation} $x_B$ is the Bjorken variable which is given as $x_B = Q^2/(W^2 + Q^2 - m^2)$. $m$ is the proton mass and $m_P$ is the pseudoscalar meson mass. The GPDs $F(x,\xi,t)$ are calculated as the integration of the double distributions function \cite{mus99} \begin{equation} F(x,\xi,t) = \int_{-1} ^{1}\, d\rho \int_{-1+|\rho|} ^{1-|\rho|}\, d\gamma \delta(\rho+ \xi \, \gamma - x) \, f(\rho,t)\,\upsilon(\rho,\gamma,t). \end{equation} For the valence quark double distributions read as \begin{equation}\label{ddf} \upsilon(\rho,\gamma,t)= \frac{3}{4}\, \frac{[(1-|\rho|)^2-\gamma^2]} {(1-|\rho|)^{3}}. \end{equation} The $t$- dependence in PDFs $f$ is expressed as the Regge form \begin{equation}\label{pdfpar} f(\rho,t)= N\,e^{(b-\alpha' \ln{\rho}) t}\rho^{-\alpha(0)}\,(1-\rho)^{\beta}, \end{equation} and $\alpha(t)=\alpha(0)+\alpha' t$ is the corresponding Regge trajectory factor. The parameters in Eq.~(\ref{pdfpar}) are fitted from the known information about CTEQ6 PDF \cite{CTEQ6} e.g, or from the nucleon form factor analysis \cite{pauli}. We consider $Q^2$ evolution of GPDs via evolution of PDF in Eq.~7, see \cite{gk06}. This form of evolution is proper near the forward limit. Generally in this work, the explicit form of GPDs evolution is not so important because we work at very limited $Q^2$ interval. It was found that for PM leptoproduction the contributions of the transversity GPDs $H_T$ and $\bar {E}_T=2 \tilde H_T+E_T$ are essential \cite{gk11}. It determines the amplitudes $M_{0-,++}$ and $M_{0+,++}$ respectively, see Eq.~(\ref{conv}). With the handbag approach the transversity GPDs are accompanied by a twist-3 PM wave functions in the hard amplitude ${\cal H}$ \cite{gk11} which is the same for both the ${M}_{0\pm,++}$ amplitudes in Eq.~(\ref{conv}). This property is demonstrated in Fig.~\ref{kt_h}, where the parton helicities of the subprocess amplitude ${\cal H}$ are presented. For corresponding transversity convolutions we have forms: \begin{equation}\label{ht} \langle H_T\rangle =\int_{-1}^1 dx {\cal H}_{0-,++}(x,...)\,H_T;\; \langle \bar E_T\rangle =\int_{-1}^1 dx {\cal H}_{0-,++}(x,...)\; \bar E_T. \end{equation} There is a parameter $\mu_P$ in twist-3 meson wave function that is large and enhanced by the chiral condensate. In our calculation, we use $\mu_P$ = 2 \gev at scale of 2 \gev. More details of leading twist polarized GPDs $\tilde H$ and $\tilde E$ which contribute to the leading twist amplitudes with longitudinally polarized virtual photons can be found in paper \cite{gk09,gk11}. These amplitudes contribute to longitudinal cross section $\sigma_L$ which is rather small with respect to transversity contribution $\sigma_T$ for $\pi^0$ and $\eta$ production. For additional information about transversity GPDs parameterization see \cite{gk11} and \cite{gxc22}. The $\pi^0$ estimations at EicC are presented in our previous paper \cite{gxc22}. Study of $\eta$ meson leptoproduction can be performed within the handbag approach too, for details see \cite{gk11}. \section{Model results for $\pi^0$ and $\eta$ leptoproduction and convolution extraction from the data} We consider the transversity effects described in Eq.~(\ref{ht}) and take into account the leading twist contribution in Eq.~(\ref{ds}). The amplitudes are transferred from program produced by PARTONS collaboration codes \cite{parton} which was changed into Fortran employing results of GK model for GPDs \cite{gk11}. In our previous paper \cite{gxc22}, two models for transversity GPDs were analyzed. Model-1 was applied in \cite{gk11} and described fine low energy CLAS data \cite{clas}, but gave results about two times larger with respect to COMPASS data \cite{compass}. It was the reason to change GPDs parameters, especially for $\bar E_T$ contribution that is important in $\sigma_{T}$ and $\sigma_{TT}$ cross sections. Some changes were done for $H_T$ as well. The parameters for new model labeled as Model-2 are exhibited at Table.~1 \cite{krollpr}. \begin{table*}[h] \renewcommand{\arraystretch}{1.4} \begin{center} \begin{tabular}{| c || c | c | c || c | c |} \hline GPD & $\alpha(0)$ & $\alpha^\prime [\gev^{-2}]$ & $b [\gev^{-2}]$ & $N^u$ & $N^d$ \\[0.2em] \hline $\bar{E}_T$& -0.1 & 0.45 & 0.67 & 29.23 & 21.61 \\[0.2em] $H_T$ & - & 0.45 & 0.04 & 0.68 & -0.186 \\[0.2em]\hline \end{tabular} \end{center} \caption{Regge parameters and normalizations of the GPDs at a scale of $2 \gev$ for Model-2.} \end{table*} \begin{figure}\label{comp} \epsfysize=65mm \centerline{\epsfbox{pi0-7-2-t1.eps}} \noindent\caption{Models results at COMPASS kinematics. Experimental data are taken from \cite{compass}, dashed line represents the results of Model-2 and solid curve indicates the prediction of Model-3} \end{figure} \begin{table*}[t] \renewcommand{\arraystretch}{1.4} \begin{center} \begin{tabular}{| c | c | c | c || c | c || c | c |} \hline GPD & $\alpha(0)$ & $\beta^u$& $\beta^d$& $\alpha^\prime [\gev^{-2}]$ & $b [\gev^{-2}]$ & $N^u$ & $N^d$ \\[0.2em] \hline $\bar{E}_T$& -0.1 &4 & 5& 0.45 & 0.77 & 20.91 & 15.46 \\[0.2em] $H_T$ & - & -& -& 0.45 & 0.3 & 1.1 & -0.3 \\[0.2em] \hline \end{tabular} \end{center} \caption{Regge parameters and normalizations of the GPDs at a scale of $2\,\gev$. Model-3.} \end{table*} Results of this model are shown at COMPASS energies in Fig.~2 by dashed lines. It can be seen that there are some discrepancy between Model-2 results and COMPASS data \cite{compass} at large $-t>0.3 \mbox{GeV}^2$. That was the reason to test in addition the new Model-3 results for $\pi^0$ and $\eta$ leptoproduction. The parameters for new Model-3 are listed at Table. 2 \cite{krollpr}. Note that in this model parameters are close to model I in \cite{gxc22}, only parameters of $\bar{E}_T$ was changed. It can be seen from the $N$ parameters that Model-2 have larger $\bar{E}_T$ and smaller $H_T$ values with respect to Model-3. In Model-3, we have smaller $\bar{E}_T$ and larger $H_T$. Both models describe well $\pi^0$ production at COMPASS. Model-3 gives better results for large $-t >0.3 \mbox{GeV}^2$, see Fig.~2. Model-2 and 3 results for $\pi^0$ production at CLAS energy are exhibited in Fig.~\ref{cl-pi0}. It can be seen that both models are in accordance with unseparated cross sections $\sigma=\sigma_T+\epsilon \sigma_L$, where $\sigma_T$ predominated. At the same time, Model-3 gives closer results for $\sigma_{TT}$ that is smaller with respect to Model-2. This confirms mentioned before smaller value of $\bar{E}_T$ in the Model-3. $\sigma_{LT}$ cross sections are shown as well. \begin{figure}[h!]\label{cl-pi0} \begin{center} \begin{tabular}{cc} \includegraphics[width=7.5cm]{pi0-2-56-2-21-2.eps} \includegraphics[width=7.5cm]{pi0-2-56-2-21-3.eps} \end{tabular} \end{center} \caption{Cross sections of $\pi^0$ production in the CLAS energy range together with the data \cite{clas}. Left graph is for $\sigma$ and $\sigma_{TT}$ while right graph is for $\sigma_{LT}$. } \end{figure} Calculation of the amplitudes of $\eta$ production is similar to the $\pi^0$ case and based on the \cite{gk11} results where the singlet-octet decomposition of $\eta$-state was used with redefined decay constants. The flavor factors for $\pi^0$ and $\eta$ production are appear in combinations \begin{eqnarray}\label{flf} F_{\pi^0} =\frac{1}{\sqrt{2}} (e^u F^u-e^d F^d)=\frac{1}{3 \sqrt{2}} (2 F^u+ F^d);\nonumber\\ F_{\eta} =\frac{1}{\sqrt{6}} (e^u F^u+e^d F^d)=\frac{1}{3 \sqrt{6}} (2 F^u- F^d). \end{eqnarray} The $\eta$ factor is written without strange sea contribution which is small and can be neglected. From Table 1 and 2, it can be seen that $\bar {E}_T$ has the same signs for $u$ and $d$ quarks but $H_T$ has the different signs, respectively. This means that for $\pi^0$ case $\bar E_T$ contributions for $u, d$ quarks are added but $H_T$ are subtracted. For $\eta$ production we have opposite case: $H_T$ contributions are added but $\bar E_T$ compensated. \begin{figure}[h!]\label{cl-eta} \begin{center} \begin{tabular}{cc} \includegraphics[width=7.5cm]{eta-cs-2-6-2.eps}& \includegraphics[width=7.5cm]{eta-cs-2-6-3.eps} \end{tabular} \end{center} \caption{Cross section of $\eta$ production in the CLAS energy range together with the data \cite{cleta}. Left graph is for $\sigma$ and $\sigma_{TT}$ while right graph is for $\sigma_{LT}$. } \end{figure} \begin{figure}[h!]\label{eta78} \begin{center} \begin{tabular}{cc} \includegraphics[width=7.5cm]{eta-cs-7-2.eps}& \includegraphics[width=7.5cm]{eta-cs-8-2.eps} \end{tabular} \end{center} \caption{Cross sections of $\eta$ production at EicC energy. Upper part of the figure presents $\sigma=\sigma_T+\epsilon\, \sigma_L$ and down part- $\sigma_{TT}$ as in Fig.~4. } \end{figure} \begin{figure}[h!]\label{eta1216} \begin{center} \begin{tabular}{cc} \includegraphics[width=7.5cm]{eta-cs-12-2.eps}& \includegraphics[width=7.5cm]{eta-cs-16-2.eps} \end{tabular} \end{center} \caption{Cross section of $\eta$ production at EicC energy. The labels are same as in Fig.~5. } \end{figure} Thus we have $\bar E_T$ enhancement for $\pi^0$ case. For $\eta$ production $H_T$ is increasing. Therefore, $\pi^0$ process is more sensitive to $\bar E_T$ effects but for $\eta$ production $H_T$ influences are more visible. Models results for $\eta$ production at CLAS energy \cite{cleta} are depicted in Fig.~4. It can be seen that Model-3 with larger $H_T$ contribution describes experimental data better at small momentum transfer. Model-2 with smaller $H_T$ produces essential dip in the cross section that is not observed at experiment. Cross sections $\sigma_{TT}$ and $\sigma_{LT}$ are described properly for both models. \begin{figure}[h!]\label{conv-cl} \begin{center} \begin{tabular}{cc} \includegraphics[width=7.5cm]{pi0-HET01.eps}& \includegraphics[width=7.5cm]{pi0-HET02.eps}\\ \includegraphics[width=7.5cm]{eta-HET01.eps}& \includegraphics[width=7.5cm]{eta-HET02.eps} \end{tabular} \end{center} \caption{Extracted from the cross section transversity convolutions $|\langle {H_T} \rangle|$ and $|\langle {\bar E_T} \rangle|$ for $\pi^0$ (upper part) and $\eta$ production (lower part) at CLAS energy range. } \end{figure} Model-2 predictions at EicC energies for $\pi^0$ production are presented at \cite{gxc22}. For Model-3 at these energies we have results similar to shown in Fig,~2. The $\pi^0$ cross sections for Model-3 don't have deep near $ \vert t^\prime\vert =$ 0 $\mbox{GeV}^2$ as we have for Model-2. Model-2 and 3 results are similar for $ \vert t^\prime \vert \sim$ 0.2 $\mbox{GeV}^2$ and cross section is a bit smaller for Model-3 with respect to Model-2 at $\vert t^\prime \vert >$ 0.3 $\mbox{GeV}^2$. Our results for EicC energies $W$= 7-16 $\mbox{GeV}$ for $\eta$ production are exhibited in the Figs.~5 and 6. It can be concluded that Model-3 results are higher for the cross section $\sigma$ with respect to Model-2 and for $\sigma_{TT}$ result is opposite- Model-2 gives higher results. This is caused by larger $H_T$ contribution in Model-3 and larger $\bar{E}_T$ effects in Model-2 that is important in $\sigma_{TT}$. These model results can be checked experimentally by EicC and determine what Model-2 or 3 is more adequate to experiment. \begin{figure}[h!]\label{conv-8} \begin{center} \begin{tabular}{cc} \includegraphics[width=7.5cm]{pi0-HET03.eps}& \includegraphics[width=7.5cm]{pi0-HET04.eps}\\ \includegraphics[width=7.5cm]{eta-HET03.eps}& \includegraphics[width=7.5cm]{eta-HET04.eps} \end{tabular} \end{center} \caption{Extracted from the cross section transversity convolutions for $\pi^0$ (upper part) and $\eta$ (lower part) production at EicC ($W$ = 8 $\mbox{GeV}$). } \end{figure} \begin{figure}[h!]\label{conv-12} \begin{center} \begin{tabular}{cc} \includegraphics[width=7.5cm]{pi0-HET05.eps}& \includegraphics[width=7.5cm]{pi0-HET06.eps}\\ \includegraphics[width=7.5cm]{eta-HET05.eps}& \includegraphics[width=7.5cm]{eta-HET06.eps} \end{tabular} \end{center} \caption{Extracted from the cross section transversity convolution functions for $\pi^0$ (upper part) and $\eta$ (lower part) production at EicC ($W$ = 12 $\mbox{GeV}$). } \end{figure} Now we shall discuss how we can get information about transversity convolutions $\bar E_T$ and $H_T$ from experimental data. From Eq.~(\ref{ds}), we can obtain \begin{eqnarray}\label{ampl} |M_{0+++}|&=&\sqrt{-\kappa \frac{d\sigma_{TT}}{dt}},\nonumber\\ |M_{0-++}|&=&\sqrt{2 \kappa (\frac{d\sigma_{T}}{dt}+\frac{d\sigma_{TT}}{dt})}, \end{eqnarray} we can determine the absolute values of the amplitudes. Employing normalization factor from Eq.~(\ref{conv}) we can determine $H_T$ and $\bar E_T$ convolutions. This procedure was adopted to extract tranvsersity convolutions from CLAS experimental data in Ref. \cite{kr-conv,kubar}. Now we don't have experimental data from China EicC. To demonstrate what can be done we shall use instead realistic experimental data, our model calculations for the cross sections $\frac{d\sigma_{T}}{dt}$ and $\frac{d\sigma_{TT}}{dt}$. Our results for $H_T$ and $\bar E_T$ convolutions for $\pi^0$ and $\eta$ production are depicted in Fig.~7. They are close to results found in \cite{kr-conv,kubar} at CLAS energies. As expected we find that $H_T$ convolution is larger for Model-3, at the same for Model-2 we get larger $\bar E_T$. Using these results, we can extract convolutions for $u$ and $d$ flavors under the help of: \begin{eqnarray}\label{flav} F^u=\frac{3}{4}(\sqrt{2} F^{\pi}+\sqrt{6} F^{\eta}),\nonumber\\ F^d=\frac{3}{2}(\sqrt{2} F^{\pi}-\sqrt{6} F^{\eta}), \end{eqnarray} which is a consequence of Eq.~(\ref{flf}). Here $F$ are corresponding transversity $H_T$ or $\bar{E}_T$ convolution functions. Such analyzes was performed at CLAS energies at \cite{kubar}. We will not do this here, because we have model results for flavor convolutions, but extraction of transversity convolution functions from future experimental data can be in important result in later experiments. Our predictions for $H_T$ or $\bar {E}_T$ convolution functions that were extracted from the cross sections at the energies $W=$ 8, 12 $ \mbox{GeV}$ which are typical at EicC energy range are exhibited in Figs.~8 and 9. Experimental analyses of these quantities can give information on the preferable models for transversity GPDs. In Fig.~10, we present our model predictions for energy dependencies of transversity convolution functions at fixed $Q^2$ and momentum transfer. Such analyses will be important to give a constraints on the $W$- dependence of $H_T$ or $\bar {E}_T$ GPDs from future experimental data. Note that the transversity dominance $\sigma_T\gg \sigma_L$ that was tested for $\pi^0$ production at high energies is valid for $\eta$ production at the energies $W = $2$\sim$ 15 $\mbox{GeV}$. This means that in experimental analyzes of transversity convolutions, unseparated cross section $\sigma = \sigma_T+ \epsilon \sigma_L$ can be applied instead $\sigma_T$ for both processes of $\pi^0$ and $\eta$ production. \begin{figure}[h!]\label{convW} \begin{center} \begin{tabular}{cc} \includegraphics[width=7.5cm]{pi0-HEW01.eps}& \includegraphics[width=7.5cm]{pi0-HEW02.eps}\\ \includegraphics[width=7.5cm]{eta-HEW01.eps}& \includegraphics[width=7.5cm]{eta-HEW02.eps} \end{tabular} \end{center} \caption{Energy dependencies of extracted trasnversity convolutions for $\pi^0$ (upper part) and $\eta$ (lower part) production at EicC ($Q^2$ = 2 $\mbox{GeV}^2$, $|t'|$ = 0.1 $\mbox{GeV}^2$). } \end{figure} \section{Conclusion} In this paper we investigate exclusive electroproduction of pseudoscalar $\pi^0$ and $\eta$ meson at China EicC energies. The process amplitudes are calculated in the model where amplitudes are factorized into subprocess amplitudes and GPDs. For the transversity twist-3 effects, the subprocess amplitude ${\cal H}_{0-,++}$ is the same in Eq.~(\ref{ht}) for both contributions that contains $H_T$ and $\bar{E}_T$ GPDs. We consider two GPDs parameterization Model-2 and Model-3. Both models describe properly $\pi^0$ and $\eta$ production at CLAS energies. It seems that Model-3 gives better results for $\pi^0$ production at COMPASS at large momentum transfer and gives better description of $\eta$ production at CLAS at momentum transfer $|t| < $0.5 $ \mbox{GeV}^2$. We perform predictions for unseparated $\sigma$ and $\sigma_{TT}$ cross sections for EicC kinematics for $\eta$ production with Model-2 and Model-3. We observe that transversity dominance $\sigma_T\gg \sigma_L$, found at low CLAS energies \cite{gk11} and confirmed at EicC energies in Ref.\cite{gxc22} for $\pi^0$ process is valid at all these energies for $\eta$ production too. Adopting combination of the cross sections, we extract GPDs $H_T$ and $\bar E_T$ convolutions determined in Eq.~(\ref{conv}) for the cases of $\pi^0$ and $\eta$ mesons. These results for EicC energies are quite different that give possibility to determine preferable model at future experiments. In addition, we analyze energy dependencies of transversity convolutions at fixed $t$ and $W$ that can give information about energy parameters of GPDs $H_T$ and $\bar {E}_T$ from the data. Note that reactions $\pi^0$ and $\eta$ production considered here have different flavor contributions to the amplitudes. This gives possibility to perform $u$ and $d$ flavor separation for transversity GPDs \cite{kubar}. Our results can be useful in future experiments at China EicC on the pseudoscalar mesons production and give more important knowledge on transversity influences at these energy ranges. \section*{Acknowledgment} S.G. expresses his gratitude to P.Kroll for collaboration on GPDs study and to V.Kubarovsky for important discussions. The work partially supported by is Strategic Priority Research Program of Chinese Academy of Sciences (Grant NO. XDB34030301) and the CAS president's international fellowship initiative (Grant No. 2021VMA0005).

e4b4e5fbef25f674528f73b0f3d04d95c46e2334

\section{Introduction} QMCPACK, is a modern high-performance open-source Quantum Monte Carlo (QMC)~\cite{becca_quantum_2017} simulation code~\cite{kim_qmcpack_2018,kent_qmcpack_2020}. Its main applications are electronic structure calculations of molecular, nanoscale and solid-state systems. Variational Monte Carlo (VMC), diffusion Monte Carlo (DMC) and several other advanced QMC methods are implemented with highly optimized algorithms. These algorithms sample the positions of electrons within the simulated system to accurately compute quantum mechanical properties. Their actual implementations in QMCPACK are called QMC drivers. When QMCPACK development started in the beginning of the first decade of this century, there were initially only multi-threaded CPU drivers. With over 15 years of development, QMCPACK has been well optimized to run on multicore CPU-only supercomputers like Cori at the National Energy Research Scientific Computing Center (NERSC) and Theta at the Argonne Leadership Computing Facility (ALCF). When NVIDIA GPUs emerged in the field of high-performance computing, GPU dedicated QMC drivers were introduced by using Compute Unified Device Architecture (CUDA). They perform extremely well on supercomputers with a CPU-GPU hybrid architecture like Summit at the Oak Ridge Leadership Computing Facility (OLCF), though at the cost of portability. The CUDA-based GPU drivers are completely incompatible with CPU-only drivers and call separate sets of subroutines of the wavefunction calculation. This is required due to the different data layouts and algorithms needed for high-performance on the GPUs\cite{esler2012}. For this reason, close to redundant feature implementations are needed to satisfy both architectures. If a feature implementation is missing in the GPU drivers, the whole simulation needs to run with a CPU-only build because mixing CPU and GPU features is not supported and the CPU/GPU selection is made at compile time. As supercomputers start to reach Exascale and architectural diversity has increased, and CUDA is no longer the only GPU programming model available. We must find approaches to efficiently address the different hardware\cite{ExascaleSkin2020}. The current Top500 rank-1 supercomputer, the AMD-based Frontier at OLCF, generally prefers Heterogeneous-Computing Interface for Portability (HIP) due to the installed AMD GPUs. To a large degree, CUDA source code can be treated as HIP code directly, and be easily compiled for AMD GPUs, making the portability issue not significant. However, on the Intel-based Aurora machine at ALCF, SYCL is the preferred programming model for Intel designed GPUs. The trick that treats CUDA as HIP won't work for SYCL considering the fundamental difference between CUDA and SYCL. The QMCPACK developers have already been struggling with maintaining both a CPU based and a CUDA based GPU implementation. Adding additional drivers for each preferred programming model is clearly not a sustainable direction. In this work, we introduce a new universal design of batched QMC drivers which may replace all the previous QMC drivers. The added flexibility in these drivers enables maximizing code performance on specific hardware once users match parallelism hierarchies properly to the actual software and hardware. The new design supports the necessary data movement to allow mixing CPU-only and GPU accelerated features to ensure a feature complete QMCPACK experience for the user regardless of the hardware being used. Code specialization for specific hardware remains possible for achieving potentially higher performance although this no longer needs to be at the driver level. This paper is organized as follows. Sec.~\ref{sec:drivers} analyzes the DMC algorithm and its implementation before the new batched drivers are added. Sec.~\ref{sec:newbatcheddrivers} introduces the details of the new drivers. Sec.~\ref{sec:results} shows how the code behaves in CUDA-based GPU drivers and the new drivers, and discusses application performance. Sec.~\ref{sec:conclusions} summarizes the hierarchical parallelism in QMCPACK. \section{QMC drivers without the batched design}\label{sec:drivers} \subsection{Basic QMC algorithm} Before analyzing all the three sets of drivers, let us first understand the characteristics of a DMC algorithm shown in Alg.~\ref{alg:DMC}. \begin{algorithm} \begin{algorithmic}[1] \FOR{$\text{MC generation}=1\cdots M$} \FOR{$\text{walker}=1\cdots N_w$} \STATE let ${\bf R} = \{{\bf r}_1 \ldots {\bf r}_{N}\}$ \FOR{$\text{particle}\ k=1 \cdots N$} \STATE set ${\bf r}_k^{\prime} \leftarrow {\bf r}_k+\nabla_k \Psi_T({\bf R)}+\delta$ \STATE let ${\bf R}^{\prime} = \{{\bf r}_1 \ldots {\bf r}_k^{\prime} \ldots{\bf r}_{N}\}$ \STATE \textbf{ratio} $\rho = \Psi_T ({\bf R}^{\prime})/\Psi_T ({\bf R})$ \STATE \textbf{derivatives} $\nabla_k \Psi_T({\bf R}^{\prime})$ \STATE Accept $\mathbf{r}_k \leftarrow \mathbf{r}_k^{\prime}$ or reject \ENDFOR \COMMENT{particle} \STATE \textbf{local energy} $E_L=\hat{H}\Psi_T({\bf R})/\Psi_T({\bf R})$ \ENDFOR \COMMENT{walker} \STATE reweight and branch walkers based on $E_L - E_T$ \STATE update $E_T$ and load balance via MPI. \ENDFOR \COMMENT{MC generation} \end{algorithmic} \caption{Pseudocode for diffusion Monte Carlo.\label{alg:DMC}} \end{algorithm} \begin{itemize} \item L1. The loop over generations is a sequential time-stepping loop for the DMC imaginary time evolution. \item L2. The walker evolution at each generation is independent of each other and thus this loop can be parallelized. On parallel computers, walkers are first parallelized over Message Passing Interface (MPI) and then parallelized within each MPI process. Due to the fact that MPI is only needed for aggregating results and handle walker count imbalance in L14, the parallel efficiency of QMC algorithms over MPI can be made nearly perfect~\cite{kim_qmcpack_2018}, even at a scale of thousands to millions of MPI processes. For the rest of this work, we restrict the discussion of parallelization schemes of walkers within an MPI process. \item L4. The loop of over particles (electrons) during random walking is also sequential. Each iteration is called a single particle move since only one particle of a walker is moved. This algorithm is referred to as particle-by-particle moves. \item L5-6. Proposing a new electron position requires relatively cheap computation. \item L7-8. When a single particle move gets proposed, heavy computational routines contain another vector loop over all the orbitals or particles. \item L9. The computational cost depends on whether a proposed move gets accepted or rejected. Upon accepting a move, additional computation is needed to update the internal data of a walker, including a determinant matrix inverse which contributes the leading term of algorithmic complexity. Rejecting a move doesn't need computation and results near zero cost. This line causes the major computational cost difference at each single particle move. In VMC simulations, the acceptance ratio is typically between 20--80\%. However, within the more costly DMC, the acceptance ratio is usually very high ($> 99\%$). \item L11. Energy evaluations are required for every walker after single particle moves. They are expensive. \end{itemize} \subsection{Multi-threaded CPU drivers} \begin{algorithm} \begin{algorithmic}[1] \FOR{$\text{MC generation}=1\cdots M$} \STATE \#pragma omp parallel for \FOR{$\text{walker}=1\cdots N_w$} \FOR{$\text{particle}\ k=1 \cdots N$} \STATE ... \ENDFOR \COMMENT{particle} \ENDFOR \COMMENT{walker} \ENDFOR \COMMENT{MC generation} \end{algorithmic} \caption{Pseudocode for the multi-threaded CPU implementation.\label{alg:CPU}} \end{algorithm} On multi-core CPUs, the multi-threaded CPU driver implementation distributes walkers over CPU cores via OpenMP threads as shown in Alg.~\ref{alg:CPU}. The needed code change is minimal. Although the cost of each single particle move depends on whether the proposed move is accepted or rejected, the overall cost of each walker is almost equal once the single particle move loop completes given the acceptance ratios across walkers. Thus, the load-balance of threads is also near perfect. In the CPU implementation, we also adopt OpenMP \texttt{simd} directives for the vector loop mentioned in Algo.~\ref{alg:DMC} to leverage the Single instruction, multiple data (SIMD) units on modern CPUs~\cite{sc17}. This parallelization strategy works extremely well on many-core wide vector CPUs including Intel Xeon, AMD EPYC and Fujitsu A64FX processors. \subsection{CUDA-based GPU drivers} The introduction of GPUs in HPC challenged the above parallelization strategy. The accelerator characteristics of GPUs require sufficiently heavy compute kernels to amortize kernel submission or synchronization cost in microseconds. In the multi-threaded CPU drivers, each compute routine only handles the work of a single walker. For large simulation problems ($>1000$ electrons), the workload of a single walker can keep GPUs busy. But many QMCPACK users run small to medium problem sizes ($<=1000$ electrons) in scientific production, so dispatching GPU computation from multi-threaded CPU drivers results in slow execution with most of the time being spent in GPU overhead. For this reason, a GPU friendly scheme~\cite{esler2012} was devised and implemented in the CUDA-based drivers in QMCPACK. \begin{algorithm} \begin{algorithmic}[1] \FOR{$\text{MC generation}=1\cdots M$} \FOR{$\text{particle}\ k=1 \cdots N$} \STATE Algorithm 1. Line 5,6,7,8,9 over all the $N_w$ walkers \ENDFOR \COMMENT{particle} \STATE \textbf{local energy} $E_L=\hat{H}\Psi_T({\bf R})/\Psi_T({\bf R})$ over $N_w$ \STATE reweight and branch walkers based on $E_L - E_T$ \STATE update $E_T$ and load balance via MPI. \ENDFOR \COMMENT{MC generation} \end{algorithmic} \caption{Pseudocode for the CUDA-based implementation.\label{alg:GPU}} \end{algorithm} In Alg.~\ref{alg:GPU} it appears that the loop over walkers in Alg.~\ref{alg:DMC} disappeared. Actually it is not removed but is added inside each of the computational routines previously serving only one walker at a time. All the computational routines on L3 now handle all the walkers in a batched operation. As a result, all the walkers advance in lock-step for each single particle move. All the compute kernels expose both vector computation and walker concurrency and fit extremely well the hierarchical design of GPUs with threads and thread blocks. For small simulated systems, GPUs have sufficient memory to enable batching over hundreds to thousands of walkers. This is sufficient to hide most of the GPU kernel overhead in practice. However, this scheme has a few limitations: (a) In the operation of accept/reject a single particle move, the number of walkers with their proposed moves accepted must be large enough to avoid leaving part of the compute hardware idle. (b) There is only one thread enqueuing kernels and handling synchronization, while all the other threads are idle. When the only working host thread is occupied with handling of pre-/post-kernel processing, the GPU is also left idle. (c) Most of the time, there is only one CUDA stream being used and there is limited overlap between kernel execution and data transfer or concurrent kernel execution. Using multiple CUDA streams can be added but requires significant code implementation. (d) Assigning one thread block per walker only works for small problem sizes. It doesn't allow leveraging more thread blocks per walker to further speed up the computation. Meanwhile, large problems cannot keep many walkers resident on GPU due to the device memory capacity limit and thus limits the full use of hardware resources. \subsection{Deficiency of multi-threaded and CUDA-based drivers} Constrained by their pre-determined parallelization schemes, either of the above driver designs only works efficiently in a limited parameter space as illustrated in Fig.~\ref{fig:working_space_legacy}. The lower triangle parameter space restriction comes from memory capacity limits. In addition, the multi-threaded CPU drivers only invoke single walker APIs while CUDA-based drivers only invoke ``batched'' or multi walker APIs. Due to the difference in data layout and assumptions of data locations, it is not possible to fallback from one implementation to the other and every feature must be implemented separately for both multi-threaded CPU and GPU drivers. i.e. The CPU and GPU codes were effectively internal forks of the codebase. This is clearly undesired due to the additional developer effort and added maintenance cost. Only compute-heavy features are worth porting to GPUs and running light computation on CPUs is usually sufficient. \begin{figure} \centering \includegraphics[width=0.95\columnwidth]{figs/legacydrivers.pdf} \caption{Parameter space of efficient runs with multi-threaded CPU (teal) or GPU (grey) driver.\label{fig:working_space_legacy}} \end{figure} With the above deficiency in mind, here we introduce a new high-performance design for the QMC drivers and overall application. It avoids diverging code paths at the driver level and works efficiently in the full possible parameter space as illustrated in Fig.~\ref{fig:working_space_pp}. \begin{figure} \centering \includegraphics[width=0.95\columnwidth]{figs/ppdrivers.pdf} \caption{Parameter space of efficient runs with batched drivers (brown).\label{fig:working_space_pp}} \end{figure} \section{Performance portable batched drivers}\label{sec:newbatcheddrivers} QMC drivers implemented in the new design are called batched drivers as walkers are handled by compute devices in batches. This applies even for CPU based machines, although a batch size of one is selectable, recovering close-to the old CPU-only algorithm. Before explaining the design detail, we first introduce a new concept ``crowd'', as a sub-organization of the walker population. A crowd is a subset of the walkers that are operated on as a single batch. Walkers within a crowd move through the operations of the QMC algorithm in lock-step. Walkers in different crowds remain fully asynchronous unless operations over the full population are needed. The batched DMC driver pseudocode is shown in Alg.~\ref{alg:batch}. Compared to the multi-threaded CPU implementation, the threaded loop over walkers has been replaced with a threaded loop over crowds and thus crowds are fully parallelized over host threads. The performance of multi-core CPUs can be easily maximized as long as the number of crowds is chosen equal to the core count. When the crowd size equal to 1, batched drivers behave exactly as multi-threaded CPU drivers. With the crowd size larger than 1, the throughput of generating statistical samples can potentially further increase due to improved data reuse or data locality. Compared to the CUDA-based implementation, the unchanged walker operation in batches ensures dispatching sufficiently heavy computation in each GPU invocation. The added multi-threaded crowds enable further improvements to GPU utilization. When the number of crowd is restricted to one, batched drivers behave exactly as CUDA-based drivers. With more than one crowd, crowds parallelized over host threads concurrently sending operations to GPUs; data transfer and kernel execution from different crowds may overlap if the underlying hardware allows. Thus, the first three limitations of the CUDA-based implementation are removed. Considering that the batched drivers do not mandate specific data layouts, the last limitation of the CUDA-based implementation can be removed by specializing compute kernels for extremely large problem sizes. With the added crowds, batched drivers have a flexible number of batches and batch sizes which can be tuned to maximize the performance of underlying hardware. In the new driver design, the old set of walker batched computational routines used by CUDA-based drivers are replaced with a new set which allow falling back to computation using the single walker APIs. Consequently, batched drivers allow mixing and matching CPU-only and GPU-accelerated features in a way that is neither feasible with the multi-threaded CPU implementation nor the CUDA-based GPU one. \begin{algorithm} \begin{algorithmic}[1] \FOR{$\text{MC generation}=1\cdots M$} \STATE \#pragma omp parallel for \FOR{$\text{crowd}=1\cdots C$} \FOR{$\text{particle}\ k=1 \cdots N$} \STATE Algorithm 1. Line 5,6,7,8,9 over all walkers with in this crowd \ENDFOR \COMMENT{particle} \STATE \textbf{local energy} $E_L=\hat{H}\Psi_T({\bf R})/\Psi_T({\bf R})$ over this crowd \STATE reweight and branch walkers based on $E_L - E_T$ \STATE update $E_T$ and load balance via MPI. \ENDFOR \COMMENT{crowd} CG \ENDFOR \COMMENT{MC generation} \end{algorithmic} \caption{Pseudocode for the batched DMC driver.\label{alg:batch}} \end{algorithm} \section{Results}\label{sec:results} \subsection{Demonstrating concurrent execution via GPU tracing} In order to verify that batched drivers behave as expected on GPUs with real simulations, we use NVIDIA Nsight Systems to trace GPU activities on an NVIDIA GPU when running a QMCPACK performance test, which is a NiO 8-atom supercell DMC simulation with 512 walkers. This is a ``small'' system where kernels are small and therefore hiding kernel latency and maximizing concurrency is critical to performance. Fig.~\ref{fig:qmcpack_legacy_gpu} shows the tracing of this test using the CUDA-based DMC driver. There is only 1 thread active even though there are 4 OpenMP threads available to the process. Kernel execution and data transfer are serialized. Fig.~\ref{fig:qmcpack_batch_gpu} shows the tracing of the same test using the new batched DMC driver. All the 4 OpenMP threads enqueue kernels and submit data transfers to the GPU dedicated to this process. The GPU keeps servicing requests from threads to maximize its utilization. Both concurrent kernel execution and overlapping kernel execution and data transfers are observed in the tracing. Higher efficiency is clearly obtained. \begin{figure} \centering \includegraphics[width=0.90\columnwidth]{figs/qmcpack_legacy_gpu.png} \caption{CUDA-based GPU driver GPU activity tracing. GPU API calls are made from a single thread and kernel execution and data transfers are all serialized.\label{fig:qmcpack_legacy_gpu}} \end{figure} \begin{figure} \centering \includegraphics[width=0.95\columnwidth]{figs/qmcpack_concurrent.png} \caption{Batched driver GPU activity tracing. GPU API calls are made from multiple threads. Both concurrent kernel execution and overlapping kernel execution and data transfers are observed.\label{fig:qmcpack_batch_gpu}} \end{figure} \subsection{Demonstrating gain in throughput on NVIDIA GPUs} When targeting GPUs, the batched drivers currently use OpenMP offload and vendor linear algebra libraries. Their performance are compared to the CUDA-based drivers. The comparison is not apples to apples. In general, the batched drivers have fewer pieces of features running on GPUs and additional data transfer can be necessary between the GPU and the host. Although all the features used in the NiO performance tests are accelerated in both the CUDA-based and batched drivers, the implementations still differ due to the fundamental design change in the batched drivers. And for the old CUDA-based drivers, all the kernels were handwritten and highly optimized in CUDA while the performance of batched drivers are affected by the quality of kernels generated by the OpenMP offload compilers. Here we can compare the performance of both drivers by the sampling throughput, namely the number of samples generated in a given time. The study was conducted on the Summit supercomputer at Oak Ridge National Laboratory. Each Summit node contains dual socket IBM Power 9 processors with 42 CPU cores in total and 6 NVIDIA V100 GPUs. The optimal way of running QMCPACK requires 1 MPI rank per GPU. Thus, on each node, we place 6 MPI ranks and each MPI process has its dedicated 7 CPU cores and 1 GPU. Both the CUDA-based GPU drivers and batched drivers require optimizing the walker count to maximize the throughput of a single GPU. Typically, the greatest number of walkers prior to exhausting GPU memory is optimal. In Fig.~\ref{fig:qmcpack_walker_scan}, the throughput of the runs with the CUDA-based driver increases rapidly as walker count increases. It quickly saturates at 1792 walkers once a single thread performance gets maxed out. With the batched DMC driver, throughput grows slower in small walker counts. When the total walker count per MPI rank is fixed, the walker batch size per thread is smaller in batched drivers and thus in total more GPU overhead gets exercised by all the threads. At larger walker counts when single thread maximal performance is reached, batched drivers have more potential to maximize the full GPU throughput by leveraging available threads. For this benchmark problem, the measured performance becomes higher than the CUDA-based GPU driver when the walker count exceeds 2000. \begin{figure} \centering \includegraphics[width=0.95\columnwidth]{figs/throughput_vs_nw_a32.png} \caption{Sampling throughput as a function of walker counts in the 32 atom cell NiO solid DMC simulation.\label{fig:qmcpack_walker_scan}} \end{figure} We also benchmark the code performance for a wide range of problem sizes. The benchmark uses DMC simulations of NiO solids with 16, 32, 64, 128 and 256 atoms in the simulation cell. In Fig.~\ref{fig:qmcpack_batch_gpu_throughput} the throughput of each problem size is rescaled by the throughput of CUDA-based GPU driver runs. When running CPU-only, the relative throughput is only 10\% which reflects the fact that most of compute power on Summit is from the NVIDIA V100 GPUs and the CUDA-based DMC driver is very well optimized. Our newly designed batched DMC driver shows 80\% to 115\% relative performance depending on the problem size. It is already suitable for scientific production simulations due to its competitive high-performance and feature complete nature. With further optimization, we expect that the batched drivers will exceed the CUDA-based driver in all the benchmark cases. As of August 2022, the compilers and libraries on non-NVIDIA platforms were not yet mature enough for benchmarking. We plan to evaluate them once compilers and runtime libraries are sufficiently mature. \begin{figure} \centering \includegraphics[width=0.95\columnwidth]{figs/throughput_clang15_gpu_track_cuda_base.png} \caption{Batched driver throughput compared with the CUDA-based GPU driver.\label{fig:qmcpack_batch_gpu_throughput}} \end{figure} \section{Conclusion}\label{sec:conclusions} Here we summarize the high-performance design of hierarchical parallelism in QMCPACK from the coarse level to the fine level as implemented in the batched drivers. \begin{enumerate} \item Fully MPI distributed walker population. Usually one MPI per CPU socket or GPU. Extremely good strong and weak scaling across thousands to millions of compute nodes. \item Multi-threaded crowds handle walkers within each MPI process. Each crowd does its independent time evolution. This is highly scalable on multi-core CPUs. In a CPU-GPU hybrid architecture, crowds may maximize the utilization of CPU cores before the shared GPUs are saturated by the workload. \item Batched computation of walkers within each crowd. On GPUs, their computations can be submitted to GPUs with minimal GPU API overhead. On CPUs, there remains the possibility of breaking them into smaller tasks which can run on additional CPU threads if they are available. \item Compute kernels of each walker operate on a set of orbitals, usually the same or more than the electron count, or all the electrons. They can be fully vectorized on single instruction, multiple threads (SIMT) and single instruction, multiple data (SIMD) hardware. For extremely large problem sizes, these vector loop can be broken up in order to leverage more threads on CPUs or thread blocks on GPUs. \end{enumerate} By matching these parallelism levels to appropriate software abstractions in a high-performance parallel computer, we believe that the maximal code performance can be achieved regardless of the underlying hardware and true performance portability can be achieved across CPU/GPU and even accelerators from any vendor, potentially also including FPGAs and ASICs. In future work we plan to demonstrate this portability. \section*{Acknowledgment} This research was supported by the Exascale Computing Project (17-SC-20-SC), a collaborative effort of the U.S. Department of Energy Office of Science and the National Nuclear Security Administration. This research used resources of the Oak Ridge Leadership Computing Facility, which is a DOE Office of Science User Facility supported under Contract DE-AC05-00OR22725. This research used resources of the Argonne Leadership Computing Facility, which is a DOE Office of Science User Facility supported under Contract DE-AC02-06CH11357. \newpage \bibliographystyle{IEEEtran} \input{Hipar22.bbl} \end{document}

569fb8c4408cb7c7a49b8955c18dfa7b136172de

\section{Introduction} High-entropy alloys (HEAs) receive much interest due to their rich functionality, such as high strength, corrosion resistance, energy storage, radiation protection, superconductivity, soft ferromagnetism, and biocompatibility\cite{Sathiyamoorthi:PMS2022,Chang:AM2020,Wang:JMCA2021,Marques:EES2021,Pickering:Entropy2021,Kitagawa:Metals2020,Chaudhary:MT2021,Castro:Metals2021}. The entropy state of an alloy is classified by the configurational entropy $\Delta S_\mathrm{mix}=-R\sum_{i=1}^{n}c_{i}\mathrm{ln}c_{i}$, where $n$ is the number of elements, $c_{i}$ is the atomic fraction, and $R$ is the gas constant. At the initial stage of HEAs research, HEA required $\Delta S_\mathrm{mix}$ larger than 1.62 $R$, which can be realized by the solid solution of more than five elements. The critical value defining HEA is now\cite{Yan:MMTA2021} 1.0 $R$. So the equiatomic quaternary alloy investigated in this study can be referred to as HEA. In HEAs, the microstructure often affects the physical properties, and, for example, the control of microstructure towards improved mechanical or magnetic properties is the central topic of HEAs\cite{Bhardwaj:TI2021,Rao:AFM2021}. The other current topic is the materials research on HEAs exhibiting novel phenomena \cite{Zherebtsov:Int2020,Kitagawa:APLMater2022} by utilizing the large compositional space of HEAs. Magnetic HEAs are attracting much attention because of their good soft ferromagnetism coexisted with high strength and/or high hardness\cite{Chaudhary:MT2021,Shivam:JALCOM2020}. The magnetic HEAs are a promising alternative to conventional soft magnetic materials with poor strength\cite{Chaudhary:MT2021}. Based on a face-centered cubic (fcc) FeCoNi with high saturation magnetization and low coercive field $H_\mathrm{c}$, many studies on the effect of adding alloying elements are carried out. For example, Al$_{x}$CoCrFeNi is well investigated HEA\cite{Kao:JALCOM2011}, which shows a structural change from fcc to body-centered cubic (bcc) structure with increasing $x$. The Curie temperature $T_\mathrm{C}$ of Al-free CoCrFeNi is 120 K, which can be enhanced above room temperature as $x$ increases\cite{Kao:JALCOM2011}. NiFeCoCrPd and NiFeCoCrMn are famous equiatomic HEAs with $T_\mathrm{C}$=440 K and 38 K, respectively\cite{Billington:PRB2020,Schneeweiss:PRB2017}. There are a few reports which aim at attaining a higher $H_\mathrm{c}$. FeCoNiAlCu$_{x}$Ti$_{x}$ shows a relatively high $H_\mathrm{c}$ of 955 Oe after heat treatment\cite{Na:AIPAdv2021}. The recent research on magnetic HEAs shows a growing interest in controlling magnetic properties through tailoring a microstructure\cite{Zuo:ActaMat2017,Rao:AFM2021}. For example, Fe$_{15}$Co$_{15}$Ni$_{20}$Mn$_{20}$Cu$_{30}$ exhibits a spinodal decomposition after a heat treatment\cite{Rao:AFM2021}, which leads to enhanced $T_\mathrm{C}$. \begin{table*} \centering \caption{\label{tab:table1}Lattice parameter $a$, atomic composition determined by EDX measurement, Curie temperature $T_\mathrm{C}$, saturation moment $M_\mathrm{s}$ at 50 K, and residual electrical resistivity $\rho_\mathrm{res}$ of each sample.} \begin{tabular}{cccccc} \hline sample & a (\AA) & atomic composition & $T_\mathrm{C}$ (K) & $M_\mathrm{s}$ ($\mu_\mathrm{B}$/f.u.) & $\rho_\mathrm{res}$ ($\mu\Omega$cm) \\ \hline FeCoNiPd (as-cast) & 3.683(2) & Fe$_{24.8(8)}$Co$_{26(1)}$Ni$_{25.2(5)}$Pd$_{24(1)}$ & 947 & 1.478 & 9.6 \\ FeCoNiPd (annealed) & 3.681(1) & Fe$_{25.2(9)}$Co$_{24(1)}$Ni$_{25.5(5)}$Pd$_{25.3(4)}$ & 955 & 1.458 & 8.1 \\ FeCoNiPt (as-cast) & 3.692(1) & Fe$_{25.4(6)}$Co$_{24(1)}$Ni$_{23.8(6)}$Pt$_{26.8(9)}$ & 848 & 1.464 & 26.1 \\ FeCoNiPt (annealed) & 3.686(1) & Fe$_{24(1)}$Co$_{25.0(5)}$Ni$_{26.0(7)}$Pt$_{25.0(7)}$ & 851 & 1.456 & 21.2 \\ \hline \end{tabular} \end{table*} We are focusing on HEAs with the combination of 3$d$ magnetic elements and noble metals (e.g., Rh, Ir, Pd, and Pt)\cite{Baba:Materials2021} because materials research on such an HEA is unexplored. Recently, Fukushima et al. reported the database ($T_\mathrm{C}$, spin moment, residual resistivity) of 147630 quaternary HEAs produced by a density functional theory calculation\cite{Fukushima:PRM2022}. This database contains fcc FeCoNiPd and fcc FeCoNiPt, which can be regarded as Pd- or Pt-added FeCoNi alloy, and the magnetic properties of these HEAs are not well investigated. Only the M\"{o}ssbauer effect of FeCoNiPd is reported\cite{Ciealak:JMMM2021}, and the fundamental magnetic properties such as $T_\mathrm{C}$ and saturation moment $M_\mathrm{s}$ are unknown. Therefore, this study's first purpose is to assess the fundamental magnetic properties of fcc FeCoNiPd and fcc FeCoNiPt and to compare the magnetic properties between these HEAs and FeCoNi. The Vickers microhardness is usually measured to evaluate the hardness of a material. In HEAs, it is reported that the Vickers microhardness empirically correlates with the valence electron count (VEC) per atom\cite{Tian:IM2015}. The hardness vs. VEC plot of HEAs with VEC ranging from 4.1 to 8.8 seems to form a broad peak at VEC$\sim$6.8, as mentioned below. The VEC of FeCoNiPd or FeCoNiPt is 9.25, and we are interested in the hardness to check the possible universal relationship between the hardness and the VEC. This examination would be useful for designing soft ferromagnetic HEAs with high hardness. The second purpose of this study is to investigate the Vickers microhardness of FeCoNiPd and FeCoNiPt. This paper reports the magnetic properties, electrical resistivity, and hardness of as-cast and annealed fcc FeCoNiPd and fcc FeCoNiPt. The electronic structure was calculated to elucidate the ferromagnetism in each HEA. We have found that the heat treatment does not significantly influence the physical properties of FeCoNiPd and FeCoNiPt. The soft ferromagnetic behaviors are observed in both HEAs. The comparisons between experimental and theoretical values are made for $T_\mathrm{C}$, $M_\mathrm{s}$, and the residual resistivity. The hardness of FeCoNiPd and FeCoNiPt do not deviate from the empirical relationship between the hardness and the VEC. \section{Materials and Methods} Polycrystalline samples of FeCoNiPd and FeCoNiPt were synthesized by a home-made arc furnace using constituent elements Fe (99.9 \%), Co (99.9 \%), Ni (99.9 \%), Pd (99.9 \%), and Pt (99.9 \%) under Ar atmosphere. The button-shaped samples were remelted several times on a water-cooled Cu hearth and flipped each time to ensure homogeneity. The as-cast samples were annealed in an evacuated quartz tube at 800 $^{\circ}$C for four days. Room temperature X-ray diffraction (XRD) patterns of samples were recorded using an X-ray diffractometer (XRD-7000L, Shimadzu) with Cu-K$\alpha$ radiation. We used thin slabs cut from the samples due to their high ductility. The scanning electron microscope (SEM) images were collected using a field emission scanning electron microscope (FE-SEM, JSM-7100F, JEOL). The chemical composition was also evaluated by an energy dispersive X-ray (EDX) spectrometer equipped with the FE-SEM. The temperature dependence of dc magnetization $\chi_\mathrm{dc}$ ($T$) between 50 and 300 K was measured using VersaLab (Quantum Design). The high-temperature $\chi_\mathrm{dc}$ ($T$) from 300 K to 1173 K was measured by a vibrating sample magnetometer (TM-VSM33483-HGC, Tamakawa) to estimate $T_\mathrm{C}$. The isothermal magnetization ($M$) curve at 50 K was taken using the VersaLab. The temperature dependence of electrical resistivity $\rho$ ($T$) between 3 K and 300 K was measured by a conventional dc four-probe method using a home-made sysytem in a GM refrigerator (UW404, Ulvac cryogenics). The Vickers microhardness was measured under the applied load of 0.49, 0.98, 1.96, 2.94, and 4.903 N, respectively, using a Shimadzu HMV-2T microhardness tester. The holding time under the diamond indenter is 10 s. We also performed the electronic structure calculation using the coherent potential approximation (CPA) approach because of no report of the density of states (DOS) in the previous study\cite{Fukushima:PRM2022}. We employed the Akai-KKR program package\cite{Akai:JPSJ1982}, which is based on the Korringa-Kohn-Rostoker (KKR) method with CPA. We used the generalized gradient approximation from Perdew-Burke-Ernzerhof (PBE) and treated the spin-polarization and the spin-orbit interaction. The spin-orbit interaction is not included in the previous study\cite{Fukushima:PRM2022}. \section{Results and Discussion} \begin{figure} \centering \includegraphics[width=0.9\linewidth]{fig1.eps \caption{\label{fig1} XRD patterns of FeCoNiPd and FeCoNiPt. The origin of each pattern is shifted by a value for clarity.} \end{figure} \begin{figure*} \centering \includegraphics[width=0.8\linewidth]{fig2.eps \caption{\label{fig2} SEM images of (a) FeCoNiPd (as-cast), (b) FeCoNiPd (annealed), (c) FeCoNiPt (as-cast), and (d) FeCoNiPt (annealed), respectively. The elemental mappings are also shown.} \end{figure*} \begin{figure*} \centering \includegraphics[width=0.85\linewidth]{fig3.eps \caption{\label{fig3} (a) Temperature dependences of $\chi_\mathrm{dc}$ of as-cast and annealed FeCoNiPd. The external field is 100 Oe. The inset is the temperature derivative of $\chi_\mathrm{dc}$ for each sample. (b) Temperature dependences of $\chi_\mathrm{dc}$ of as-cast and annealed FeCoNiPt. The external field is 100 Oe. The inset is the temperature derivative of $\chi_\mathrm{dc}$ for each sample. (c) Isothermal magnetization curves at 50 K of as-cast and annealed FeCoNiPd. The inset is the expanded view. (d) Isothermal magnetization curves at 50 K of as-cast and annealed FeCoNiPt. The inset is the expanded view.} \end{figure*} Figure \ref{fig1} shows the XRD patterns of as-cast and annealed FeCoNiPd and FeCoNiPt. All patterns can be indexed by fcc structure with the Miller indices denoted in the figure. The lattice parameters $a$ obtained by the least square method are listed in Table \ref{tab:table1}. SEM images and elemental mappings of all samples are displayed in Fig.\ref{fig2}. In each sample, no trace of impurity phase is detected, and the elemental mapping shows a homogeneous elemental distribution. The chemical compositions evaluated by EDX measurements are tabulated in Table \ref{tab:table1} and agree well with the ideal one with 25 at.\% for each element. HEAs often show composition segregation after heat treatment\cite{Vrtnik:JALCOM2017,Pacheco:InorgChem2019}. However, FeCoNiPd or FeCoNiPt forms a stable single-phase fcc against the heat treatment at 800 $^{\circ}$C. Figures \ref{fig3}(a) and (b) show $\chi_\mathrm{dc}$ ($T$) under the external field $H$ of 100 Oe for FeCoNiPd and FeCoNiPt, respectively. In each sample, a steep increase of $\chi_\mathrm{dc}$ is observed as the temperature is lowered, which indicates a ferromagnetic ordering. $T_\mathrm{C}$ is estimated by the temperature derivative of $\chi_\mathrm{dc}$ and defined by the minimum point described in the inset of Fig.\ref{fig3}(a) or (b). This is one of the effective ways to obtain $T_\mathrm{C}$ in transition metal-based ferromagnets\cite{Oikawa:APL2001,Yu:APL2003,Kitagawa:JMMM2018,Kitagawa:JSSC2020}. Thus obtained $T_\mathrm{C}$s are listed in Table \ref{tab:table1}, and the slight enhancement of $T_\mathrm{C}$ after the annealing is confirmed in each HEA. Fukushima et al. have provided $T_\mathrm{C}$ data obtained by the density functional theory calculation for 147630 quaternary HEAs and placed the data in a repository\cite{Fukushima:rep}. The predicted $T_\mathrm{C}$s of fcc FeCoNiPd and fcc FeCoNiPt are 1137 and 1085 K, respectively. Although the higher $T_\mathrm{C}$ of FeCoNiPd compared to FeCoNiPt is consistent with the experimental result, the theoretical value is approximately 200 K higher than the experimental one in each HEA. The mean-field approximation used for calculating $T_\mathrm{C}$ tends to overestimation, which is the reason for the relatively large difference between theoretical and experimental values. The isothermal $M$-$H$ curves of FeCoNiPd and FeCoNiPt measured at 50 K are exhibited in Figs.\ref{fig3}(c) and (d). With the increase of $H$ from 0 Oe in each HEA, $M$ increases steeply and soon saturates, which supports the ferromagnetic ground state. Thermal annealing seems not to affect $M$-$H$ curves. The values of $M_\mathrm{s}$ for HEAs investigated are summarized in Table \ref{tab:table1} and compared with the theoretical values below. The inset of each figure is the expanded view to show the hysteresis. The very weak hysteresis indicates the soft ferromagnetism in each HEA. While $H_\mathrm{c}$ of FeCoNiPt would be approximately 2 Oe, no hysteresis is observed for FeCoNiPd within the measurement accuracy. \begin{figure} \centering \includegraphics[width=0.7\linewidth]{fig4.eps \caption{\label{fig4} Temperature dependences of $\rho$ of FeCoNiPd and FeCoNiPt.} \end{figure} \begin{figure*} \centering \includegraphics[width=0.8\linewidth]{fig5.eps \caption{\label{fig5} Electronic density of states of fcc FeCoNiPd ((a) and (b)) and fcc FeCoNiPt ((c) and (d)). Each partial DOS ((b) and (d)) is drawn for only $d$-electrons due to the dominant contribution around the Fermi level. The Fermi energy is set to 0 Ry.} \end{figure*} Here we compare the magnetic properties of FeCoNiPd and FeCoNiPt with those of fcc FeCoNi. The lattice parameter, $T_\mathrm{C}$, and $M_\mathrm{s}$ of fcc FeCoNi are reported to be 3.599 \AA, 1000 K, and 163 emu/g (=1.687 $\mu_\mathrm{B}$/f.u.), respectively\cite{Zuo:JMMM2014,Lin:MCP2021}. The atomic radius of Pd(Pt) is 1.3754 \AA (1.387 \AA), which is larger than those of Fe, Co, and Ni (1.2412, 1.2510, and 1.2459 \AA, respectively)\cite{Miracle:ActMat2017}. Therefore, the lattice parameter of FeCoNiPd or FeCoNiPt expands compared to FeCoNi. As discussed later, Pd(Pt) carries the magnetic moment smaller than Fe, Co, and Ni moments, which causes the reduction of $M_\mathrm{s}$ in FeCoNiPd or FeCoNiPt. Going from FeCoNi, FeCoNiPd to FeCoNiPt, the lattice parameter increases, and $T_\mathrm{C}$ is systematically reduced. The expansion of lattice means the increase of interatomic distance between 3$d$ elements, which would lead to a weakened magnetic exchange interaction. In this case, the systematic reduction of $T_\mathrm{C}$ with increasing lattice parameter can be anticipated. We note that FeCoNiCr ($a$=3.580 \AA, $T_\mathrm{C}$=104 K)\cite{Chou:MSEB2009,Na:AIPAdv2018} and FeCoNiMn ($a$=3.6029 \AA, $T_\mathrm{C}$=332 K)\cite{Rao:PRM2020} significantly reduce $T_\mathrm{C}$, which is independent of the unit-cell volume change compared to FeCoNi. The alloying by antiferromagnetic elements of Cr or Mn severely decreases the averaged magnetic exchange interaction strength. FeCoNi possesses $H_\mathrm{c}$ of 1.5 Oe and this value is not largely enhanced in fcc FeCoNiAl$_{x}$ or fcc FeCoNiSi$_{x}$ ($H_\mathrm{c}$: 0.5$\sim$6 Oe)\cite{Chaudhary:MT2021}. The same trend is observed in FeCoNiPd and FeCoNiPt, and the soft ferromagnetism of fcc FeCoNi is robust against the addition of alloying elements. Next, $M_\mathrm{s}$ and $H_\mathrm{c}$ of FeCoNiPd and FeCoNiPt are compared with those of related systems. The values of $M_\mathrm{s}$ in the unit of emu/g for FeCoNiPd and FeCoNiPt after annealing are 116 emu/g and 88 emu/g, respectively. The corresponding values of FeCoNi-related systems\cite{Deng:MC2020,Dasari:Mat2020,Zuo:JMMM2014,Li:IM2017} are, for example, 130 emu/g in FeCoNiAl$_{x}$ ($x$=0.2 and 0.3), 126 emu/g in FeCoNiSi$_{0.25}$, and 101 emu/g in FeCoNiAl$_{0.25}$Mn$_{0.25}$. So, while $M_\mathrm{s}$ of FeCoNiPd is slightly reduced compared to FeCoNiAl$_{x}$ and FeCoNiSi$_{0.25}$, the relatively large reduction of $M_\mathrm{s}$ occurs in FeCoNiPt due to the heavy element Pt with a small magnetic moment. This comparison indicates that light elements are favorable for achieving high $M_\mathrm{s}$ soft ferromagnet. FeCoNiAl$_{x}$ ($x$=0.2 and 0.3), FeCoNiSi$_{0.25}$, and FeCoNiAl$_{0.25}$Mn$_{0.25}$ possess $H_\mathrm{c}$ ranging from 0.5 to 5 Oe\cite{Deng:MC2020,Dasari:Mat2020,Zuo:JMMM2014,Li:IM2017}. These values are comparable to $H_\mathrm{c}$ ($\sim$2 Oe) of FeCoNiPt. On the other hand, $H_\mathrm{c}$ of FeCoNiPd would be smaller compared to the related systems. $H_\mathrm{c}$ of HEAs and commercial soft ferromagnets are summarized in the review by Huang et al\cite{Huang:crystals2020}. According to the review, $H_\mathrm{c}$ of FeCoNiPd would locate in the range of commercial soft ferromagnets (Ni-Fe alloys with $H_\mathrm{c}$ of 0.004 Oe $\sim$ 0.1 Oe). Figure \ref{fig4} summarizes $\rho$ ($T$) of as-cast and annealed FeCoNiPd and FeCoNiPt. For each HEA, $\rho$ of the as-cast sample decreases after the heat treatment, which suggests a relaxation of the lattice distortion. Despite the existence of atomic disorders, $\rho$ smoothly decreases with cooling. The measured temperature range is well below $T_\mathrm{C}$, and relatively large temperature dependences mean the dominance of magnetic contribution to the electrical transport. Such a behavior is often observed in ferromagnetic metals with atomic disorders\cite{Kitagawa:Metals2020-2}. The database\cite{Fukushima:rep} by Fukushima et al. also includes the values of residual resistivity $\rho_\mathrm{res}$. According to the database, $\rho_\mathrm{res}$s of fcc FeCoNiPd and fcc FeCoNiPt are 5.1 and 21.6 $\mu\Omega$cm, respectively, which agree well with the experimental values of annealed HEAs. \begin{table*} \centering \caption{\label{tab:table2}Spin moment and orbital moment of each element in FeCoNiPd and FeCoNiPt with fcc structure obtained by electronic structure calculation. The total moments are 1.402 $\mu_\mathrm{B}$/f.u. and 1.421 $\mu_\mathrm{B}$/f.u. for FeCoNiPd and FeCoNiPt, respectively.} \begin{tabular}{cccccc} \hline \multicolumn{3}{c}{FeCoNiPd}&\multicolumn{3}{c}{FeCoNiPt}\\ atom & spin moment ($\mu_\mathrm{B}$) & orbital moment ($\mu_\mathrm{B}$) & atom & spin moment ($\mu_\mathrm{B}$) & orbital moment ($\mu_\mathrm{B}$)\\ \hline Fe & 2.854 & 0.0661 & Fe & 2.875 & 0.0657 \\ Co & 1.874 & 0.1015 & Co & 1.902 & 0.0891 \\ Ni & 0.779 & 0.0641 & Ni & 0.806 & 0.0505 \\ Pd & 0.233 & 0.0139 & Pt & 0.242 & 0.0413 \\ \hline \end{tabular} \end{table*} In the electronic structure calculations of fcc FeCoNiPd and fcc FeCoNiPt, the lattice parameters obtained for the annealed samples are used, and the perfect solid solution of constituent elements is assumed. Figures \ref{fig5} (a) and (c) exhibit the total DOSs of FeCoNiPd and FeCoNiPt, respectively. In each case, the difference in total DOS between spin-up and spin-down electrons supports the ferromagnetic ground state. Partial DOSs are displayed in Figs.\ref{fig5} (b) and (d) for FeCoNiPd and FeCoNiPt, respectively. Only $d$-electron DOS is drawn for each partial DOS due to the dominant contribution around the Fermi level. Reflecting the isoelectronic HEAs, the structures of partial DOS shown in Figs.\ref{fig5} (b) and (d) for each element are similar to each other. The partial DOSs indicate the presence of magnetic moment for all elements. The spin and orbital moment values calculated for all elements are summarized in Table \ref{tab:table2}. In each HEA, all moments align parallel, and Fe and Co spin moments are dominant. The spin moment values of Fe, Co, and Ni are slightly larger than those obtained by Fukushima et al\cite{Fukushima:rep}. Accordingly, the total moments of 1.402 $\mu_\mathrm{B}$/f.u. for FeCoNiPd and 1.421 $\mu_\mathrm{B}$/f.u. for FeCoNiPt are also respectively larger than 1.383 and 1.395 $\mu_\mathrm{B}$/f.u. in the database. The spin-orbit interaction considered in this study would be responsible for the slight increase of moment. The calculated total moments in this study or the database well explain the experimental $M_\mathrm{s}$s (see also Table \ref{tab:table1}), which means that the magnetic structure of FeCoNiPd or FeCoNiPt would be a simple one with all spins aligning parallel. Figure \ref{fig6} (a) depicts the Vickers microhardness vs. applied load profiles for all samples. The hardness in each sample gradually decreases as the load is increased, which is also reported in many HEAs and intermetallic compounds\cite{Zhu:JALCOM:2022,Ma:JSSC2021}. It is well known that the elastic recovery effect is responsible for the increase of Vickers microhardness with decreasing load\cite{Jindal:SC1988}. In FeCoNiPd, the annealed sample displays lower hardness than the as-cast sample. The annealing would cause the release of lattice distortion introduced in the rapid solidification process of arc melting, which leads to lower hardness. FeCoNiPt also shows similar behavior. It is proposed that the Vickers microhardness depends on the VEC. Figure \ref{fig6}(b) presents the VEC dependence of hardness. The solid curve is a guideline obtained for HEAs with the bcc structures (VEC: 6.0$\sim$7.55) and fcc structures (VEC: 7.8$\sim$8.8)\cite{Tian:IM2015}. The result of refractory bcc HEAs with VEC:4.14$\sim$5.65 is also shown\cite{Kitagawa:JALCOM2022,Han:IM2017,Li:Materials2019,Bhandari:JMRT2020,Ge:MSEA2020} and seems to be connected to the guideline. In addition, a deep learning study of the hardness of refractory HEAs with the VEC: 4$\sim$6 also supports the positive correlation between the hardness and VEC in that VEC range\cite{Bhandari:Crystals2021}. The VEC of FeCoNiPd or FeCoNiPt is 9.25, and the hardness of 188 HV obtained at 4.9030 N load is employed because the hardness at a higher load is usually employed. In addition we cannot obtain the hardness under a load higher than 4.9030 N due to the limitation of the microhardness tester used. Thus plotted data points in Figure \ref{fig6}(b) do not essentially deviate from the expected universal relation between the VEC and the hardness. We note that there are not enough data points from other HEAs at VEC larger than 9. Therefore, further study is required to prove the universal relation. This study implies a vital role of VEC in designing the hardness of magnetic HEAs. The VEC is also related to the phase stability of fcc and bcc HEAs\cite{Gao:book}: a single bcc phase for VEC between 5.0 and 6.87 and a single fcc phase for VEC larger than 8.0. Therefore, if we need a soft ferromagnetic HEA with high hardness within the limitation of a single fcc phase, a magnetic HEA based on FeCoNi with an alloying for tuning VEC=8 is desirable. \begin{figure} \centering \includegraphics[width=0.75\linewidth]{fig6.eps \caption{\label{fig6} (a) Load dependences of Vickers microhardness for FeCoNiPd and FeCoNiPt. The solid curves are guides for the eyes. (b) VEC dependence of Vickers microhardness of FeCoNiPd and FeCoNiPt with VEC=9.25. The refractory bcc HEAs reported in the literatures\cite{Kitagawa:JALCOM2022,Han:IM2017,Li:Materials2019,Bhandari:JMRT2020,Ge:MSEA2020} are also plotted (VEC:4.14$\sim$5.65). The solid curve represents the guideline obtained for HEAs with bcc structures (VEC: 6.0$\sim$7.55) and fcc structures (VEC: 7.8$\sim$8.8) taken from the literature\cite{Tian:IM2015}.} \end{figure} \section{Summary} We have investigated the fundamental magnetic properties, electrical resistivity, and Vickers microhardness of as-cast and annealed fcc FeCoNiPd and fcc FeCoNiPt. After the heat treatment at 800 $^{\circ}$C, both HEAs keep the single-phase fcc, and the annealing does not largely alter the physical properties. The values of $T_\mathrm{C}$ and $M_\mathrm{s}$ at 50 K are 955 K and 1.458 $\mu_\mathrm{B}$/f.u. for annealed FeCoNiPd, and 851 K and 1.456 $\mu_\mathrm{B}$/f.u. for annealed FeCoNiPt, respectively. The coercive fields of both HEAs are very small, and they are soft ferromagnets. In each HEA, the electrical resistivity shows the metallic temperature dependence. The electronic structure calculations of both HEAs were performed, and the ferromagnetic ground states were obtained. The total moments are close to those reported in the database made by the density functional theory calculation and agree with the experimental $M_\mathrm{s}$s. The theoretical values of $T_\mathrm{C}$ and $\rho_\mathrm{res}$ are also reported in the database. While the theoretical $T_\mathrm{C}$ values are relatively higher than the experimental ones in both HEAs, good agreement is confirmed between the theoretical and experimental $\rho_\mathrm{res}$ values. The database would be beneficial for the materials research on magnetic quarternary HEAs. In addition, the comparisons of magnetic properties of HEAs investigated and FeCoNi are made. The Vickers microhardness of FeCoNiPd or FeCoNiPt slightly decreases after the annealing. The hardness vs. VEC plot of these HEAs does not essentially deviate from the expected universal relation forming a broad peak at VEC$\sim$6.8. This study provides a possible material design of soft ferromagnetic HEA with high hardness; a magnetic HEA based on FeCoNi with an alloying for tuning VEC=8 is promising. \section*{ACKNOWLEDGMENTS} J.K. is grateful for the support provided by the Comprehensive Research Organization of Fukuoka Institute of Technology. \section*{REFERENCES}

23dd045555f240a1fa54ed36b00de84df880aae1

\section{Introduction}\label{s:0} \subsubsection*{Vertex algebras, chiral algebras and chiral operad} Nowadays vertex algebras appear in various branches of mathematics, mathematical physics and theoretical physics. They were introduced by Borcherds \cite{Bo} as an algebraic framework of the operator algebra structure appearing in the chiral part of two-dimensional conformal field theory. This framework fits very well with representation theory of infinite-dimensional algebras, and has been making vast progress since its appearance until today. On the other hand, it had been expected since at least mid 1980s that there should be a geometric framework of chiral conformal field theory. Along this expectation, together with the viewpoint of geometric Langlands program, Beilinson and Drinfeld started in the early 1990s to develop the algebro-geometric framework, which resulted in their book \cite{BD} of chiral algebras. At present, both vertex algebras and chiral algebras are used in the investigation of algebraic and geometric representation theory, and theoretical physics of conformal field theories and their cousins. Let us name \cite{FBZ} for the reference explaining both frameworks of vertex algebras and chiral algebras. The theory of chiral algebras is a flexible framework of chiral conformal field theory, and enables one to do geometric study. But it has one drawback: the definition is built on various complicated structures using the theory of $\mathcal{D}$-modules and that of operads, so that it is non-trivial for non-experts to check the relation \cite[0.15]{BD} between chiral algebras and Borcherds' axiom of vertex algebras. One feature of the theory is that a chiral algebra is defined to be a Lie algebra in a non-standard monoidal structure. Precisely speaking, on the category of $\mathcal{D}$-modules on a smooth algebraic curve, Beilinson and Drinfeld introduced the chiral tensor product and the \emph{chiral operad}. The latter object might be best called a sheaf operad, which is not an algebraic operad nor a topological operad appearing commonly. Then a (non-unital) chiral algebra is defined to be an operad morphism from the Lie operad to the chiral operad. Starting from such a definition, it is not a simple task to recover Borcherds' axiom of vertex algebras. In the late 2010s, Bakalov, De Sole, Heluani and Kac \cite{BDHK} started their investigation of the cohomology theory of vertex algebras, which continues up to present \cite{BDHK2,BDK20,BDK21,BDKV21}. As far as we understand, their study sits in the sequence of the investigation of Lie conformal algebras, vertex Poisson algebras and their cohomology theories started in the late 1990s (see \cite{BKV,DK13} for example). Among several results in \cite{BDHK}, they constructed an algebraic operad $P^{\textup{ch}}$ to which an operad morphism from the Lie operad corresponds bijectively to a vertex algebra. Thus, $P^{\textup{ch}}$ is a purely algebraic counterpart of the chiral operad of Beilinson and Drinfeld. The standard operadic cohomology theory then yields a natural cohomology theory of a vertex algebra, and one can investigate the relation between the obtained cohomology and those of vertex Poisson algebras. \subsubsection*{SUSY vertex algebras and SUSY chiral operad} This note is written in a simple motivation to give an analogue of the algebraic operad $P^{\textup{ch}}$ which will encode the structure of a \emph{SUSY vertex algebra}. The latter object is a super-field analogue of a vertex algebra, introduced by Heluani and Kac \cite{HK}. The framework of SUSY vertex algebras looks very natural in the viewpoint of superconformal field theory, and there are plenty examples, especially of geometric nature. However, as far as we understand, mathematical investigations of SUSY vertex algebras have not been pursued so much as ordinary vertex (super)algebras. There are two classes of SUSY vertex algebras: $N_W=N$ and $N_K=N$ SUSY vertex algebras, which originate in the classification of superconformal Lie algebras \cite{KL}. Accordingly, our argument is divided into two parts. We will introduce the algebraic operad $P^{\textup{ch}N_W}$ in \cref{s:W} and $P^{\textup{ch}N_K}$ in \cref{s:K}, which encode the structure of an $N_W=N$ and $N_K=N$ SUSY vertex algebra, respectively. These \cref{s:W} and \cref{s:K} are the main body of this note. Our argument basically follows the non-SUSY case given in \cite[\S5-\S7]{BDHK}. We first construct the operads $\Chom^{N_W}$ and $\Chom^{N_K}$ of SUSY Lie conformal algebras, and then construct the operads $P^{\textup{ch}N_W}$ and $P^{\textup{ch}N_K}$ of SUSY vertex algebras. Let us remark that, although the outline of our construction is quite the same as \cite{BDHK}, we need several technical modifications to treat the SUSY case appropriately. For example, in \cref{ss:W:VA} of the construction of the operad $P^{\textup{ch}N_W}$, we need to treat carefully the integral formulation of SUSY vertex algebras, and give detailed computations in \cref{dfn:W:indef}--\cref{prp:W:Jqas} and \cref{lem:W:circ}--\cref{lem:W:ResFn}, which we could not find in literature. The operads $P^{\textup{ch}N_W}$ and $P^{\textup{ch}N_K}$ yield the cohomology complexes of SUSY vertex algebras with coefficients in their modules, whose definition and analysis will be given in \cref{s:coh}. \subsubsection*{Summary of results} The main objects of this note are the $N_W=N$ SUSY chiral operad $P^{\textup{ch}N_W}$ introduced in \cref{dfn:W:Pch} and the $N_K=N$ SUSY chiral operad $P^{\textup{ch}N_K}$ introduced in \cref{dfn:K:Pch}. The main statements are \cref{thm:W:VA} and \cref{thm:K:VA} establishing the bijection between the odd Maurer-Cartan solutions of the operads $P^{\textup{ch}N_W},P^{\textup{ch}N_K}$ and the structures of non-unital $N_W=N$ and $N_K=N$ SUSY vertex algebras. Another block of main results is the cohomology theory given in \cref{s:coh}. The cohomology complex for an $N_W=N$ SUSY vertex algebra is given in \cref{dfn:W:VAmod}, and the interpretation of the low degree cohomology is given in \cref{thm:W:coh}. We have a straightforward analogue for the $N_K=N$ case, which will be briefly explained in \cref{ss:coh:K}. \subsubsection*{Organization} Let us explain the organization. \begin{itemize} \item \cref{s:op} is a preliminary section of super-objects, operads and operadic cohomology theory. In particular, we will explain in \cref{ss:1:Kos} the convolution Lie algebra $\mathfrak{g}(\oP,\oQ)$ for operads $\oP$ and $\oQ$ which encodes operad morphisms $\oP \to \oQ$, and explain in \cref{ss:1:LQ} that the Lie superalgebra $\mathfrak{g}(\oLie,\oQ)$ associated to an operad $\oQ$ introduced in \cite[\S 3]{BDHK} is nothing but the convolution Lie algebra. \item \cref{s:W} is devoted to study the algebraic operad encoding the structure of $N_W=N$ SUSY vertex algebras. We start with the preliminary of polynomial superalgebras in \cref{ss:W:poly}, which will be ``the base algebra'' of our operad. Then following the line of \cite[\S5, \S6]{BDHK}, we introduce an operad $\Chom^{N_W}$ encoding the structure of $N_W=N$ SUSY Lie conformal algebras in \cref{ss:W:LCA}, and an operad $P^{\textup{ch}N_W}$ encoding that of $N_W=N$ SUSY vertex algebras, which is one of the main objects of this note, in \cref{ss:W:VA}. In \cref{ss:W:CA}, we relate the operad $P^{\textup{ch}N_W}$ to the $N_W=N$ SUSY analogue of the Beilinson-Drinfeld chiral operad, following the line of the non-SUSY case \cite[Appendix A]{BDHK}. \item \cref{s:K} treas the $N_K=N$ SUSY vertex algebras, and the materials are delivered in the same order as in the previous \cref{s:W}. \item The operads $P^{\textup{ch}N_W}$ and $P^{\textup{ch}N_K}$ naturally yield the cohomology complexes of SUSY vertex algebras with coefficients in their modules, which will be studied in \cref{s:coh}. We start with the review of the operadic cohomology theory in \cref{ss:coh:CE}. Then we discuss the cohomology theory for $N_W=N$ SUSY vertex algebras in \cref{ss:coh:W}, and the theory of $N_K=N$ SUSY vertex algebras in \cref{ss:coh:K}. We close this note with concluding remarks in \cref{ss:cnc}. \end{itemize} \subsubsection*{Global notation} \begin{itemize} \item The symbol $\mathbb{N}$ denotes the set $\{0,1,2,\dotsc\}$ of non-negative integers. \item For a finite set $S$, we denote by $\# S$ the number of elements in $S$. \item For a positive integer $m$, the symbol $[m]$ denotes the set $\{1, 2,\dotsc, m\} \subset \mathbb{Z}$. \item The word `ring' or `algebra' means a unital associative one unless otherwise specified. \item The symbol $\partial_x$ denotes the partial differential $\frac{\partial}{\partial x}$ with respect to $x$. \end{itemize} \section{Lie algebra structures on operads}\label{s:op} This section gives an overview of \emph{the universal Lie superalgebra associated to an operad} introduced in \cite[\S3]{BDHK}. Although the definitions and arguments therein is clear, we feel that the reader would require extra explanation on the motivation and backgrounds, and this section is designed to give backgrounds on the theory of algebraic operads and to give some complementary comments on loc.\ cit. We refer to \cite{LV} for the comprehensive reference of algebraic operads, and to \cite{DK13} for the preceding study of the universal Lie superalgebra. In the starting \cref{ss:1:op}, we introduce notation and terminology on operads, and explain some fundamental notion. We will work in the super setting, following the presentation of \cite[\S2, \S3]{BDHK}. In the next \cref{ss:1:Kos}, we give a brief review of the theory of operadic Koszul duality and operadic deformation theory. In particular, for a binary quadratic operad $\oP$ and an operad $\oQ$, we introduce in \eqref{eq:1:gPQ} the convolution Lie algebra $\mathfrak{g}(\oP,\oQ)$, which concerns the $\oP$-algebra structure on $\oQ$ as \eqref{eq:1:MCQ}. In the last \cref{ss:1:LQ}, applying the arguments so far to the case $\oP=\oLie$ (the Lie operad), we recover the universal Lie superalgebra associated to $\oQ$ in the sense of \cite{BDHK}. See \cref{dfn:1:LP} and \cref{prp:1:LP} for the precise statement. In this section, all the objects (linear spaces, algebras and so on) are defined over the base field $\mathbb{K}$ of characteristic $0$ unless otherwise stated. The unit of $\mathbb{K}$ is denoted by $1_\mathbb{K}$. \subsection{Operads and algebras over operads}\label{ss:1:op} Here we collect from \cite{LV} and \cite{BDHK} some basic notions and symbols on algebraic operads in the super setting. Here is the list of our super terminology, which is more or less the standard one and can be found in \cite{DM} for example. \begin{itemize} \item We write $\mathbb{Z}_2 \ceq \mathbb{Z}/2\mathbb{Z}$, and denote by $\overline{i} \in \mathbb{Z}_2$ the parity of $i \in \mathbb{Z}$ . We also define $(-1)^{\overline{i}} \ceq (-1)^i$. \item A $\mathbb{Z}_2$-graded linear space $V$ is called a \emph{linear superspace}. The grading is expressed by $V=V_{\overline{0}}\oplus V_{\overline{1}}$. An element of $V_{\overline{0}} \mathbin{\setminus} \{0\}$, $V_{\overline{1}} \mathbin{\setminus} \{0\}$ and $V_{\textup{pure}} \ceq (V_{\overline{0}} \cup V_{\overline{1}}) \mathbin{\setminus} \{0\}$ is called \emph{even}, \emph{odd} and \emph{pure}, respectively. The parity $p(x)\in\mathbb{Z}_2$ of a pure element $x\in V_{\textup{pure}}$ is defined by \[ p(x) \ceq \begin{cases} \overline{0} & (x\in V_{\overline{0}}), \\ \overline{1} & (x\in V_{\overline{1}}). \end{cases} \] Whenever we write $p(x)$ in the text, we assume $x\in V_{\textup{pure}}$. \item For linear superspaces $V$ and $W$, we denote by $\Hom(V,W)$ the linear superspace of linear maps from $V$ to $W$. The $\mathbb{Z}_2$-grading is denoted by \[ \Hom(V,W)=\Hom(V,W)_{\overline{0}} \oplus \Hom(V,W)_{\overline{1}}, \] and an element of $\Hom(V,W)_{\overline{0}}$ (resp.\ $\Hom(V,W)_{\overline{1}}$) is called an even (resp.\ odd) linear map. Sometimes we denote $\Hom_{\mathbb{K}}(V,W) \ceq \Hom(V,W)$ to stress the base field $\mathbb{K}$. We also denote by $\End(V)=\End(V)_{\overline{0}} \oplus \End(V)_{\overline{1}}$ the linear superspace of linear endomorphisms of $V$. The $\mathbb{Z}_2$-graded $\mathbb{K}$-linear category of linear superspaces and linear maps is denoted by $\bbK\textup{-}\mathsf{Mod}$. \item We denote by $\otimes \ceq \otimes_{\mathbb{K}}$ the standard tensor product on $\bbK\textup{-}\mathsf{Mod}$. See \cite[(1.1)]{DM} for the sign rule of the tensor product. \item The \emph{parity change functor} is denoted by $\Pi$. Thus, for a linear superspace $V$, we have the linear superspace $\Pi V$ with $(\Pi V)_{\overline{0}} = V_{\overline{1}}$ and $(\Pi V)_{\overline{1}} = V_{\overline{0}}$. \item A $\mathbb{Z}_2$-graded algebra $A$ is called a \emph{superalgebra}, and a linear superspace equipped with a left (resp.\ right) $\mathbb{Z}_2$-graded left $A$-module structure is called a \emph{left} (resp.\ \emph{right}) \emph{$A$-supermodule}. A \emph{morphism $M \to N$ of left $A$-supermodules $M$ and $N$} means a linear map consistent with the supermodule structures of $M$ and $N$. The linear superspace of all morphisms of left $A$-supermodules is denote by \[ \Hom_A(M,N) = \Hom_A(M,N)_{\overline{0}} \oplus \Hom_A(M,N)_{\overline{1}}. \] \item A superalgebra $A$ is called commutative if the multiplication satisfies $x y=(-1)^{p(x) p(y)} y x$ for any pure $x,y \in A$. \end{itemize} Next we give some symbols for symmetric groups and their modules. \begin{itemize} \item For $n \in \mathbb{N}$, we denote by $\mathfrak{S}_n$ the $n$-th symmetric group with the convention $\mathfrak{S}_0 \ceq \{e\}$. We consider the group algebra $\mathbb{K}[\mathfrak{S}_n]$ as a purely even superalgebra. \item Let $V$ be a linear superspace. For $n \in \mathbb{N}$, the superspace $V^{\otimes n}$ carries a structure of a left $\mathfrak{S}_n$-supermodule by letting $\sigma \in \mathfrak{S}_n$ act on $v_1 \otimes \dotsb \otimes v_n \in V^{\otimes n}$ by \begin{align}\label{eq:1:Sn-Vn} \sigma(v_1\otimes \cdots\otimes v_n) \ceq \prod_{\substack{i<j \\ \sigma(i)>\sigma(j)}} (-1)^{p(v_i)p(v_j)} \cdot v_{\sigma^{-1}(1)}\otimes \cdots\otimes v_{\sigma^{-1}(n)}. \end{align} In what follows, we always consider $V^{\otimes n}$ with this structure of a left $\mathfrak{S}_n$-supermodule unless otherwise specified. \end{itemize} Now we introduce the notion of $\mathfrak{S}$-modules in the super setting. See \cite[\S5.1]{LV} for the detail in the non-SUSY case. \begin{itemize} \item An \emph{$\mathfrak{S}$-supermodule} $M=\bigl(M(n)\bigr)_{n \in \mathbb{N}}$ is a collection of right $\mathfrak{S}_n$-supermodules $M(n)$. The index $n$ is called the arity. A \emph{morphism $f\colon M \to N$ of $\mathfrak{S}$-supermodules} is a family of morphisms $f_n\colon M(n) \to N(n)$ of right $\mathfrak{S}_n$-modules. The resulting $\mathbb{K}$-linear category of $\mathfrak{S}$-supermodules is denoted by $\frS\textup{-}\mathsf{Mod}$. \item The category $\frS\textup{-}\mathsf{Mod}$ has a monoidal structure whose tensor product $M \circ N = \bigl((M \circ N)(n)\bigr)_{n \in \mathbb{N}}$ is given by \begin{align}\label{eq:pre:op:circ} (M \circ N)(n) \ceq \bigoplus_{k \in \mathbb{N}} M(k) \otimes_{\mathbb{K}[\mathfrak{S}_k]} \Bigl( \bigoplus_{n_1+\dotsb+n_k=n} \bigl(N(n_1) \otimes \dotsb \otimes N(n_k)\bigr) \otimes_{\mathbb{K}[\mathfrak{S}_{n_1} \times \dotsb \times \mathfrak{S}_{n_k}]} \mathbb{K}[\mathfrak{S}_n]\Bigr), \end{align} and whose unit object is $I \ceq (\delta_{n,1}\mathbb{K})_{n \in \mathbb{N}}=(0,\mathbb{K},0,0,\dotsc)$. \end{itemize} An operad is a monoid object in the monoidal category $(\frS\textup{-}\mathsf{Mod},\circ,I)$, i.e., it consists of an $\mathfrak{S}$-supermodule $\oP=\bigl(\oP(n)\bigr)_{n \in \mathbb{N}}$ and two morphisms $\gamma\colon \oP \circ \oP \to \oP$ and $\eta\colon I \to \oP$ of $\mathfrak{S}$-supermodules, which should satisfy the standard axioms of monoids. Unraveling the axioms, we obtain: \begin{dfn}[{\cite[\S3.1]{BDHK}}]\label{dfn:1:op} Consider the following data $(\oP,\gamma,1)$: \begin{itemize} \item An $\mathfrak{S}$-supermodule $\oP$, i.e, a map $n \mapsto \oP(n)$ associating a right $\mathfrak{S}_n$-supermodule to each $n\in\mathbb{N}$. \item A map $\gamma$ associating an even linear map \[ \gamma_{k, n_1, \ldots, n_k}\colon \oP(k)\otimes \oP(n_1)\otimes\cdots\otimes\oP(n_k)\to \oP(n_1+\cdots+n_k). \] to each $k, n_1, \ldots, n_k\in\mathbb{N}$. $\gamma$ is called the composition map. For simplicity, we denote \[ f(g_1\otimes \cdots \otimes g_k) \ceq \gamma_{k,n_1,\dotsc,n_k}(f \otimes g_1 \otimes \dotsb \otimes g_k) \] for $f\in\oP(k)$ and $g_i\in\oP(n_i)$ ($i=1,\dotsc,k$). \item An element $1\in\oP(1)$. $1$ is called the unit. \end{itemize} A triple $(\oP, \gamma, 1)$ is called an \emph{operad} if it satisfies the following conditions. Hereafter we use the symbol $[k] \ceq \{1,2,\dotsc,k\}$ for $k \in \mathbb{N}$ with the convention $[0] \ceq \emptyset$. \begin{clist} \item Let us given $k \in \mathbb{N}$, $m_1,\dotsc,m_k \in \mathbb{N}$ and $n_1,\dotsc,n_{m_1+\dotsb+m_k} \in \mathbb{N}$. For $f\in\oP(k)$, $g_i\in\oP(m_i)$ ($i\in[k]$) and $h_j\in\oP(n_j)$ ($j\in[m_1+\cdots+m_k]$), we have \begin{align*} f((g_1\otimes \cdots\otimes g_k)(h_1\otimes \cdots\otimes h_{m_1+\cdots+m_k})) =(f(g_1\otimes \cdots \otimes g_k))(h_1\otimes \cdots\otimes h_{m_1+\cdots+m_k}). \end{align*} \item For $f\in \oP(n)$, $f(1\otimes \cdots \otimes 1)=1(f)=f$. \item For $f\in\oP(k)$, $g_i\in\oP(n_i)$ ($i\in [k]$), $\sigma\in\mathfrak{S}_k$ and $\tau_i\in\mathfrak{S}_{n_i}$ ($i\in[k]$), we have \begin{align*} f^\sigma(g_1^{\tau_1}\otimes\cdots \otimes g_k^{\tau_k})=(f(\sigma(g_1\otimes \cdots \otimes g_k)))^{\sigma(\tau_1, \ldots, \tau_k). } \end{align*} Here $f^\sigma$ denotes the right action of $\sigma\in\mathfrak{S}_k$ on $f \in \oP(k)$, and $\sigma(\tau_1, \ldots, \tau_k)\in\mathfrak{S}_{n_1+\cdots+n_k}$ is given by \begin{align*} \sigma(\tau_1, \ldots, \tau_k)(l)\ceq n_{\sigma^{-1}(1)}+\cdots n_{\sigma^{-1}(\sigma(i)-1)}+\tau_i(n) \end{align*} for each $l \in [n_1+\dotsb+n_k]$ written uniquely as $l=n_1+\cdots+n_{i-1}+n$, $i \in [k]$ and $n \in [n_i]$. \end{clist} \end{dfn} Some complementary terminology is in order. \begin{itemize} \item Given an operad $(\oP, \gamma, 1)$, we will often say that $\oP$ is an operad, without mentioning $\gamma$ and $1$. \item An operad $\oP$ is called \emph{even} if the linear superspace $\oP(n)$ is even for each $n \in \mathbb{N}$. \item Let $\oP$ be an operad. For $m, n\in\mathbb{N}$ and $i\in[m]$, the map \begin{align}\label{eq:1:circ_i} \circ_i\colon \oP(m)\otimes \oP(n) \longrightarrow \oP(m+n-1), \quad f\circ_ig \ceq f(1\otimes \cdots \otimes 1 \otimes \overset{i}{\check{g}}\otimes 1\otimes \cdots \otimes 1) \end{align} is an even linear map. It is called the \emph{infinitesimal composition}. \end{itemize} Next we recall on the notion of algebras over an operad. \begin{itemize} \item A \emph{morphism $\alpha\colon \oP \to \oQ$ of operads} is a morphism of $\mathfrak{S}$-supermodules which is compatible with the monoid object structures. Operads and their morphisms form a subcategory $\mathsf{Op}$ of $\frS\textup{-}\mathsf{Mod}$, which is called the \emph{category of operads}. See \cite[\S5.2.1]{LV} for the detail. \item For a linear superspace $V$, we denote by $\oHom_V \in \mathsf{Op}$ the \emph{endomorphism operad} in the sense of \cite[\S5.2.11]{LV}. As an $\mathfrak{S}$-supermodule, it is given by \begin{align}\label{eq:1:HomV} \oHom_V(n) \ceq \Hom(V^{\otimes n},V) \end{align} with right $\mathfrak{S}_n$-module structure induced by the left $\mathfrak{S}_n$-action on $V^{\otimes n}$ by permuting tensor factors \eqref{eq:1:Sn-Vn}. \item For an operad $\oP$ and a linear superspace $V$, a \emph{$\oP$-algebra structure on $V$} is a morphism $\alpha\colon \oP \to \oHom_V$ of operads. See \cite[\S5.2.3, Proposition 5.2.2]{LV} for another equivalent definition. Such a pair $(V,\alpha)$ will be called a \emph{$\oP$-algebra}. \end{itemize} Let us recall the formulation of classical algebra structures in terms of the operad theory. \begin{itemize} \item Let $\oCom$ be the \emph{commutative operad} \cite[\S 5.2.10, \S 13.1]{LV}. We have $\oCom(n)=\mathbb{K}$, the trivial representation, for each $n \in \mathbb{N}$. A $\oCom$-algebra $(V,\alpha)$ is nothing but a commutative $\mathbb{K}$-algebra structure on $V$ whose multiplication $\mu\colon V^{\otimes 2}\to V$ is given by the image of $1_\mathbb{K} \in \mathbb{K}=\oCom(2)$ under $\alpha_2\colon \oCom(2) \to \oHom_V(2)=\Hom(V^{\otimes 2},V)$, i.e., $\mu=\alpha_2(1_\mathbb{K})$. \item Let $\oAss$ be the \emph{associative operad} \cite[\S 5.2.10, Chap.\ 9]{LV}. We have $\oAss(n)=\mathbb{K}[\mathfrak{S}_n]$, the regular representation of $\mathfrak{S}_n$. An $\oAss$-algebra structure $\alpha$ on $V$ is determined by the image of $e \in \oAss(2)=\mathbb{K}[\mathfrak{S}_2]=\mathbb{K} e+\mathbb{K}(1,2)$. The element $\mu=\alpha_2(e) \in \oHom_V(2)=\Hom(V^{\otimes 2},V)$ is nothing but the multiplication in the corresponding associative $\mathbb{K}$-algebra structure on $V$. \end{itemize} \subsection{Koszul operad theory and convolution Lie algebra}\label{ss:1:Kos} The operads $\oCom$ and $\oAss$ are \emph{quadratic operads}, for which one can develop cohomology theory and deformation theory using the operadic Koszul duality. Let us give a brief recollection. See \cite{GK} and \cite[Chap.\ 7]{LV} for the detail. \begin{itemize} \item For an $\mathfrak{S}$-supermodule $M$, we have the \emph{free operad} $\oT(M)$ which is an operad equipped with an $\mathfrak{S}$-supermodule morphism $\eta(M)\colon M \to \oT(M)$ having a universal property \cite[\S5.5.1]{LV}. The free operad $\oT(M)$ is endowed with the \emph{weight grading}, denoted as $\oT(M)=\bigoplus_{r \in \mathbb{N}} \oT(M)^{(r)}$ \cite[\S5.5.3]{LV}. \item An operadic quadratic data is a pair $(E,R)$ of an $\mathfrak{S}$-supermodule $E$ and a sub-$\mathfrak{S}$-supermodule $R \subset \oT(E)^{(2)}$. The \emph{quadratic operad} associated to an operadic quadratic data $(E,R)$ is defined to be $\oP(E,R) \ceq \oT(E)/(R)$, where $(R) \subset \oT(E)$ is the operadic ideal generated by $R$. It inherits the weight grading $\oP(E,R)=\bigoplus_{r \in \mathbb{N}}\oP(E,R)^{(r)}$. See \cite[\S7.1]{LV} for the detail. Below we always assume a quadratic operad to be even, i.e., the above $E$ is assumed to be an even linear space. \item A \emph{binary quadratic operad} is a quadratic operad $P(E,R)$ whose $\mathfrak{S}$-supermodule $E$ concentrates in the 2-arity. \end{itemize} The operads $\oCom$ and $\oAss$ are binary quadratic operads, and so is the operad $\oLie$ recalled below. \begin{itemize} \item Let $\oLie$ be the \emph{Lie operad} \cite[\S 13.2]{LV}. It is a binary quadratic operad $\oLie=\oT(\mathbb{K} c)/(R)$ with $R \ceq (c \circ_1 c)+(c \circ_1 c)^{(123)}+(c \circ_1 c)^{(132)} \subset \oT(\mathbb{K} c)^{(2)}$. Here we used the infinitesimal composition $\circ_1$ in \eqref{eq:1:circ_i}. given by \begin{align}\label{eq:circ_1} (X \circ_1 Y)(v_1 \otimes \dotsb \otimes v_{m+n-1}) \ceq X(Y(v_1 \otimes \dotsb \otimes v_n) \otimes v_{n+1} \otimes \dotsb \otimes v_{m+n-1}) \end{align} for $X \in \oHom_V(m)$ and $Y \in \oHom_V(n)$, $f^\sigma$ denotes the right action of $\sigma \in \mathfrak{S}_n$ on $f \in \oHom(V)(n)$, and $(R)$ denotes the operadic ideal generated by $R$. We also have $\oLie(2)=\mathbb{K} c$, and a $\oLie$-algebra $(V,\alpha)$ is nothing but a Lie algebra structure on $V$ over $\mathbb{K}$ whose Lie bracket is given by $\alpha_2(c) \in \oHom_V(2)=\Hom(V^{\otimes 2},V)$. \end{itemize} Next, we give an overview of the deformation theory of algebraic structures for quadratic operads. We refer to \cite[Chap.\ 12]{LV} for the detail. We begin with the notation for graded categories. \begin{itemize} \item In the following, we need $\mathbb{Z}$-graded objects \cite[\S6.2, \S6.3]{LV}, and everything in $\bbK\textup{-}\mathsf{Mod}$ which appeared so far is replaced by one in $\mathsf{grSuperLinSp}$, the ($\mathbb{Z}_2$,$\mathbb{Z}$)-bigraded category of $\mathbb{Z}$-graded linear superspaces $V=\bigoplus_{i \in \mathbb{Z}}V_i$. We call the additional $\mathbb{Z}$-grading $i$ the \emph{degree}. The morphism set $\Hom(V,W)$ for $V,W \in \mathsf{grSuperLinSp}$ is given by $\Hom(V,W)=\bigoplus_{i \in \mathbb{Z}}\Hom(V,W)_i$, where each $\Hom(V,W)_i$ is the linear superspace consisting of linear maps $f\colon V \to W$ satisfying $f(V_j) \subset W_{i+j}$ for each $j \in \mathbb{Z}$. \item We denote by $s$ the degree shift in $\mathsf{grSuperLinSp}$. So, for a graded linear superspace $V = \bigoplus_{i \in \mathbb{Z}} V_i$, we have $(s V)_i=V_{i-1}$. An element $f \in \Hom(V,W)_i$ is called a linear map of degree $i$. \item For example, a graded $\mathfrak{S}$-supermodule is a collection $M=(M(n))_{n \in \mathbb{N}}$ of $\mathbb{Z}$-graded right $\mathfrak{S}_n$-modules $M(n)=\bigoplus_{i \in \mathbb{Z}} M_i(n)$, and the index $i$ denotes the degree. We also have $(s M)_i(n)=M_{i-1}(n)$. \end{itemize} We now recall Koszul dual cooperads and convolution Lie algebras. \begin{itemize} \item We use the notion of \emph{cooperad} $\oC=(\oC,\Delta,\varepsilon)$ \cite[\S5.8]{LV}. In particular, for a ($\mathbb{Z}$-graded) $\mathfrak{S}$-supermodule $E$, we have the cofree cooperad $\oT^c(E)$ \cite[\S5.8.6]{LV}. \item Let $\oP=\oP(E,R)$ be a quadratic operad. Then we have the \emph{Koszul dual cooperad} $\oP^{\ash} \ceq \oC(s E,s^2R)$. It is a sub-cooperad of the cofree cooperad $\oT^c(s E)$ which is universal among the sub-cooperad $\oC$ such that the composite $\oC \to \oT^c(s E) \to \oT^c(s E)^{(2)}/(s^2 R)$ is zero, and also has the induced weight grading $\oP^{\ash}=\bigoplus_{r \in \mathbb{N}} (\oP^{\ash})^{(r)}$ from the weight grading of $\oP$. See \cite[\S7.1]{LV} for the detail of quadratic operads and Koszul dual cooperads. \item For a $\mathbb{Z}$-graded cooperad $\oC=(\oC,\Delta,\varepsilon)$ and a $\mathbb{Z}$-graded operad $\oP=(\oP,\gamma,\eta)$, we set \[ \Hom_{\mathfrak{S}}(\oC,\oP) \ceq \prod_{n \ge 0} \Hom_{\mathfrak{S}_n}\bigl(\oC(n),\oP(n)\bigr), \] where $\Hom_{\mathfrak{S}_n}\bigl(\oC(n),\oP(n)\bigr) \subset \Hom\bigl(\oC(n),\oP(n)\bigr)$ denotes the $\mathbb{Z}$-graded space of graded homomorphisms of $\mathbb{Z}$-graded $\mathfrak{S}_n$-supermodules, i.e., $\mathfrak{S}_n$-equivariant $\mathbb{Z}$-graded linear maps. Using the infinitesimal composition $\circ_{(1)}$ given in \eqref{eq:circ_1} and the infinitesimal operations $\gamma_{(1)}\colon \oP \circ_{(1)} \oP \to \oP$ and $\Delta_{(1)}\colon \oC \to \oC \circ_{(1)} \oC$ (see \cite[\S6.1]{LV} for the detail), we define the composite \begin{align}\label{eq:1:sq} f \square g \ceq \gamma_{(1)} \circ (f \circ_{(1)} g) \circ \Delta_{(1)} \end{align} for $f,g \in \Hom_{\mathfrak{S}_n}\bigl(\oC(n),\oP(n)\bigr)$, which is again in $\Hom_{\mathfrak{S}_n}\bigl(\oC(n),\oP(n)\bigr)$. Here we used the symbol $\square$ in \cite{BDHK}, instead of $\star$ in \cite{LV}. This operation $\square$ is a graded pre-Lie product \cite[Propositions 6.4.3, 6.4.5]{LV}, and hence we have a graded Lie superalgebra \begin{align}\label{eq:1:cvLie} \bigl(\Hom_{\mathfrak{S}}(\oC,\oP),[\cdot,\cdot]\bigr) \end{align} with $[f,g] \ceq f \square g-(-1)^{\abs{f}\abs{g}}g \square f$, where $\abs{f}$ denotes the $\mathbb{Z}$-grading of the morphism $f$. This graded Lie algebra is called the \emph{convolution Lie algebra}. \end{itemize} We have several application of the construction of convolution Lie algebra \eqref{eq:1:cvLie}. \begin{itemize} \item Let $V$ be a linear superspace and $\oP=\oP(E,R)$ be a quadratic operad which is \emph{homogeneous}, i.e., $R \subset \oT(E)^{(2)}$. Then we have the Koszul dual cooperad $\oP^{\ash}$ and the endomorphism operad $\oHom_V$. Applying the previous argument to the case $\oC=\oP^{\ash}$ and $\oP=\oHom_V$, we have the convolution Lie algebra. \begin{align}\label{eq:1:gPV} \mathfrak{g}_{\oP,V} \ceq \bigl(\Hom_{\mathfrak{S}}(\oP^{\ash},\oHom_V),[\cdot,\cdot]\bigr). \end{align} It has an $\mathbb{N}$-grading $\mathfrak{g} = \prod_{r \in \mathbb{N}} \mathfrak{g}^{r}$ induced by the weight grading on $\oP^{\ash}$. Explicitly, we have \[ \mathfrak{g}_{\oP,V}^{0} = \Hom_{\mathfrak{S}}(I,\oHom_V) \cong \End(V), \quad \mathfrak{g}_{\oP,V}^{r} = \Hom_{\mathfrak{S}}\bigl((\oP^{\ash})^{(r)},\oHom_V) \quad (r \ge 1). \] Then, by \cite[Proposition 10.1.4]{LV}, we have a bijection \begin{align}\label{eq:1:MCV} \begin{split} \{\text{$\oP$-algebra structures on $V$}\} &= \Hom_{\mathsf{Op}}(\oP,\oHom_V) \\ &\xrr{\sim} \MC\bigl(\mathfrak{g}_{\oP,V}\bigr) \ceq \{X \in \mathfrak{g}_{\oP,V}^1 \mid \tfrac{1}{2}[X,X]=0\}. \end{split} \end{align} The set $\MC(\mathfrak{g}_{\oP,V})$ is called the \emph{solutions of the Maurer-Cartan equation} of the (differential) graded Lie superalgebra $\mathfrak{g}_{\oP,V}$. \item Let us replace the endomorphism operad $\oHom_V$ in the previous item by an arbitrary operad $\oQ$. Then we obtain the convolution Lie algebra \begin{align}\label{eq:1:gPQ} \mathfrak{g}(\oP,\oQ)^{\bullet} \ceq \bigl(\Hom_{\mathfrak{S}}(\oP^{\ash},\oQ)^{\bullet},[\cdot,\cdot]\bigr), \end{align} so that the previous algebra is recovered as $\mathfrak{g}_{\oP,V}=\mathfrak{g}(\oP,\oHom_V)$. The argument in \cite[Proposition 10.1.4]{LV} works for general $\oQ$ as well, and we have a bijection \begin{align}\label{eq:1:MCPQ} \Hom_{\mathsf{Op}}(\oP,\oQ) \xrr{\sim} \MC\big(\mathfrak{g}(\oP,\oQ)\bigr) \ceq \{X \in \mathfrak{g}(\oP,\oQ)^1 \mid \tfrac{1}{2}[X,X]=0\}. \end{align} Let us call an operad morphism $\oP \to \oQ$ a \emph{$\oP$-algebra structure on $\oQ$}. Then the above bijection says that an $\oP$-algebra structure on $\oQ$ is in one-to-one correspondence with a Maurer-Cartan solution $X \in \mathfrak{g}(\oP,\oQ)^1$, $[X,X]=0$. \end{itemize} \subsection{Lie superalgebra associated to an operad}\label{ss:1:LQ} We are interested in the convolution Lie algebra \eqref{eq:1:gPQ} with $\oP=\oLie$, i.e., \begin{align}\label{eq:1:frg} \mathfrak{g}(\oLie,\oQ)^{\bullet} = \bigl(\Hom_{\mathfrak{S}}(\oLie^!,\oQ)^{\bullet},[\cdot,\cdot]\bigr), \quad [f,g] \ceq f \square g - (-1)^{\abs{f} \abs{g}}g \square f. \end{align} By \eqref{eq:1:MCPQ}, this graded Lie superalgebra concerns $\oLie$-algebra structures in the operad $\oQ$: \begin{align}\label{eq:1:MCQ} \{\text{$\oLie$-algebra structure on $\oQ$}\} \xrr{\sim} \MC\big(\mathfrak{g}(\oLie,\oQ)\bigr) = \{X \in \mathfrak{g}(\oLie,\oQ)^1 \mid \tfrac{1}{2}[X,X]=0\}. \end{align} Let us write down the graded component $\mathfrak{g}(\oLie,\oQ)^n$ for each $n \in \mathbb{N}$. Since the homogeneous quadratic operad $\oP=\oLie$ is binary. the argument in \cite[Proposition 12.1.1]{LV} works, and we have \[ \mathfrak{g}\bigl(\oLie,\oQ\bigr)^n \cong \Hom_{\mathfrak{S}_n}\bigl(\oLie^{\ash}(n+1),\oQ(n+1)\bigr) \] as linear superspaces. Next, by \cite[\S7,6]{LV}, for any homogeneous quadratic binary operad $\mathcal{P}$, each arity-graded component $\oP^{\ash}(n)$ of the Koszul dual cooperad $\oP^{\ash}$ satisfies \[ \oP^{\ash}(n) \cong \oP^!(n) \otimes_{\mathfrak{S}_n} \sgn_n \] as an $\mathfrak{S}_n$-module, where $\sgn_n$ denotes the signature representation of $\mathfrak{S}_n$, and $\oP^!$ is the \emph{Koszul dual operad} of the quadratic operad $\oP$. We also have $\oLie^! = \oCom$ by the Koszul duality \cite[\S 7.6]{LV}. Since $\oCom(n)$ is the trivial $\mathfrak{S}_n$-representation, we have \begin{align}\label{eq:1:frgn} \mathfrak{g}(\oLie,\oQ)^n \cong \bigl(\Pi \oQ(n+1)\bigr)^{\mathfrak{S}_{n+1}} \ceq \{f \in \oQ(n+1) \mid \forall \sigma \in \mathfrak{S}_{n+1}, \ f^{\sigma} = \sgn(\sigma) f\} \end{align} as linear superspaces, where $\Pi$ is the parity shift functor, and $f^\sigma$ denotes the right action of $\sigma \in \mathfrak{S}_{n+1}$ on $f \in \oQ(n+1)$. Next, we give an explicit expression of the pre-Lie product $\square$ in $\mathfrak{g}(\oLie,\oQ)$. Unraveling the definition \eqref{eq:1:sq}, we have for $f \in \mathfrak{g}(\oLie,\oQ)^n$ and $g \in \mathfrak{g}(\oLie,\oQ)^m$ that \begin{align}\label{eq:1:square} f \square g \ceq \sum_{\sigma \in \mathfrak{S}_{m+1,n}} (f \circ_1 g)^{\sigma^{-1}}, \end{align} where for $k,l \in \mathbb{N}$, the symbol $\mathfrak{S}_{k,l} \subset \mathfrak{S}_{k+l}$ denotes the subset of $(k,l)$-shuffles: \begin{align* \mathfrak{S}_{k,l} \ceq \{\sigma \in \mathfrak{S}_{k+l} \mid \sigma(1)<\dotsb<\sigma(k), \, \sigma(k+l)<\dotsb<\sigma(k+l)\}. \end{align*} The graded Lie superalgebra $\mathfrak{g}(\oLie,\oQ)$ is essentially the same as the \emph{universal Lie superalgebra associated to $\oQ$} in \cite[\S3.2]{BDHK}. The definition therein shifts the grading and compensates the parity shift. It is given as follows: \begin{dfn}\label{dfn:1:LP} For an operad $\oQ$, we define a $\mathbb{Z}_{\ge-1}$-graded linear superspace \begin{align*} L(\oQ) \ceq \bigoplus_{n \ge -1}L^n(\oQ), \quad L^n(\oQ) \ceq \oQ(n+1)^{\mathfrak{S}_{n+1}} = \{f\in\oQ(n+1) \mid \forall \sigma\in\mathfrak{S}_{n+1}, \, f^\sigma=f\} \end{align*} and a linear map $\square\colon L(\oQ)\otimes L(\oQ)\to L(\oQ)$ by \begin{align*} f\square g \ceq \sum_{\sigma \in \mathfrak{S}_{m+1, n}} (f\circ_1 g)^{\sigma^{-1}} \quad (f\in L^n(\oQ), \ g\in L^m(\oQ)). \end{align*} \end{dfn} Let us summarize the arguments so far in the following form. \begin{dfn}\label{dfn:1:Lie} For an operad $\oQ$, a \emph{Lie algebra structure on $\oQ$} means an operad morphism $\oLie \to \oQ$. \end{dfn} \begin{prp}[{\cite[Proposition 10.1.4]{LV}, \cite{DK13}, \cite[Theorem 3.4]{BDHK}}]\label{prp:1:LP} For an operad $\oQ$, the pair $(L(\oQ),\square)$ in \cref{dfn:1:LP} is a pre-Lie superalgebra (see \cite[\S1.4]{LV} for example). Hence we obtain a graded Lie superalgebra $(L(\oQ)^{\bullet},[\cdot,\cdot])$ with \begin{align*} [f, g]\ceq f\square g-(-1)^{p(f)p(g)}g\square f. \end{align*} Moreover, we have a bijection \[ \MC\bigl(L(\oQ)\bigr) \ceq \{X \in L^1(\oQ) \mid X \square X = 0\} \xrr{\sim} \{\text{Lie algebra structures on $\oQ$}\}. \] \end{prp} \section{\texorpdfstring{$N_W=N$}{NW=N} SUSY chiral operad}\label{s:W} We continue to work over the base field $\mathbb{K}$ of characteristic $0$ and to use the symbol $[k] \ceq \{1,2,\dotsc,k\}$ for $k \in \mathbb{N}$. We also fix a positive integer $N$. \subsection{Polynomial superalgebra}\label{ss:W:poly} Here we give a summary on the polynomial ring of supervariables. Most of the material is standard, and we refer to \cite[\S3.1]{HK} for a detailed description. \begin{dfn}\label{dfn:W:poly} Let $A$ be a set and $\Lambda_\alpha=(\lambda_\alpha, \theta_\alpha^1, \ldots, \theta_\alpha^N)$ be a sequence of sets for each $\alpha\in A$. We denote by $\mathbb{K}[\Lambda_\alpha]_{\alpha \in A}$ the free commutative $\mathbb{K}$-superalgebra generated by even $\lambda_\alpha$ $(\alpha\in A)$ and odd $\theta_\alpha^i$ $(\alpha\in A, i\in [N])$, i.e, the $\mathbb{K}$-superalgebra generated by these elements with relations \begin{align}\label{eq:W:poly} \lambda_\alpha\lambda_\beta-\lambda_\beta\lambda_\alpha=0, \quad \lambda_\alpha\theta_\beta^i-\theta_\beta^i\lambda_\alpha=0, \quad \theta_\alpha^i\theta_\beta^j+\theta_\beta^j\theta_\alpha^i=0 \quad (\alpha, \beta\in A, \ i, j\in[N]). \end{align} Each $\Lambda_\alpha$ for $\alpha\in A$ is called a \emph{$(1|N)_W$-supervariable}, and the $\mathbb{K}$-superalgebra $\mathbb{K}[\Lambda_\alpha]_{\alpha \in A}$ is called the \emph{$N_W=N$ polynomial superalgebra of the supervariables $(\Lambda_\alpha)_{\alpha\in A}$}. In the case $A=[n]$, we often denote the polynomial superalgebra by $\mathbb{K}[\Lambda_k]_{k=1}^n$ instead of $\mathbb{K}[\Lambda_k]_{k \in [n]}$. \end{dfn} For a $(1|N)_W$-supervariable $\Lambda=(\lambda, \theta^1, \ldots, \theta^N)$ and a subset $I=\{i_1<\cdots< i_r\}\subset [N]$, we denote $\theta^I\ceq \theta^{i_1}\cdots\theta^{i_r}$. Also, we define $\sigma(I, J)\in\{0, \pm 1\}$ for $I, J\subset [N]$ by the relation $\theta^I\theta^J=\sigma(I, J)\theta^{I\cup J}$, and set $\sigma(I)\ceq \sigma(I, [N] \mathbin{\setminus} I)$. For $\alpha \in A$, $m\in\mathbb{N}$ and $I\subset [N]$, we denote $\Lambda_\alpha^{m|I}\ceq \lambda_\alpha^m\theta_\alpha^I$. For a linear superspace $V$ and $(1|N)_W$-supervariables $\Lambda_\alpha$ ($\alpha \in A$), we denote \begin{align}\label{eq:W:VLam} V[\Lambda_\alpha]_{\alpha \in A} \ceq \mathbb{K}[\Lambda_\alpha]_{\alpha\in A} \otimes_{\mathbb{K}} V, \end{align} which is a $\mathbb{K}[\Lambda_\alpha]_{\alpha\in A}$-supermodule. As a preliminary of the following subsections, let us introduce: \begin{dfn}\label{dfn:W:clHW} Let $\mathcal{H}_W$ be the free commutative superalgebra generated by even $T$ and odd $S^i$ ($i\in[N]$), i.e., the superalgebra generated by these elements with relations \begin{align}\label{eq:W:TSi} TS^i-S^iT=0, \quad S^iS^j+S^jS^i=0\quad (i, j\in[N]). \end{align} For simplicity, we set \begin{align}\label{eq:W:nabla} \nabla \ceq (T, S^1, \ldots, S^N). \end{align} We also denote a linear superspace $V$ equipped with a left $\mathcal{H}_W$-supermodule structure as \[ (V,\nabla)=(V,T, S^1, \ldots, S^N), \] where $T$ is regarded as an even linear transformation on $V$ and $S^i$ as an odd linear transformation, satisfying the relations \eqref{eq:W:TSi}. \end{dfn} Note that $\mathcal{H}_W$ is isomorphic to $\mathbb{K}[\Lambda]$ as a superalgebra by the homomorphism defined by $T \mapsto -\lambda$, $S^i \mapsto -\theta^i$ ($i\in[N]$). Since $\mathcal{H}_W$ is a commutative superalgebra, we suppress the word `left' of an $\mathcal{H}_W$-supermodule hereafter. \subsection{Operad of $N_W=N$ SUSY Lie conformal algebras}\label{ss:W:LCA} In this subsection, we introduce a natural $N_W=N$ SUSY analogue of the operad $\Chom$ of Lie conformal algebra in \cite[\S5]{BDHK}. The main contents are \cref{dfn:W:Chom} and \cref{thm:W:LCA}. Let us fix a $(1|N)_W$-supervariable $\Lambda=(\lambda, \theta^1, \ldots, \theta^N)$. Recall the polynomial superalgebra $\mathbb{K}[\Lambda]$ in \cref{dfn:W:poly} and the $\mathbb{K}[\Lambda]$-supermodule $V[\Lambda] = \mathbb{K}[\Lambda] \otimes V$ in \eqref{eq:W:VLam}. \begin{dfn}[{\cite[Definition 3.2.2]{HK}}]\label{dfn:W:LCA} Let $(V, \nabla)=(V, T, S^1, \ldots, S^N)$ be an $\mathcal{H}_W$-supermodule and $[\cdot_\Lambda\cdot]\colon V\otimes V \to V[\Lambda]$ be a linear map of parity $\overline{N}$. A triple $(V,\nabla,[\cdot_\Lambda\cdot])$ is called an \emph{$N_W=N$ SUSY Lie conformal algebra} if it satisfies the following conditions: \begin{clist} \item (\emph{sesquilinearity}) For any $a, b\in V$, \begin{align}\label{eq:LCA:sesq} \begin{split} &[Ta_\Lambda b]=-\lambda[a_\Lambda b], \hspace{43pt} [a_\Lambda Tb]=(\lambda+T)[a_\Lambda b], \\ &[S^ia_\Lambda b]=-(-1)^{N}\theta^i[a_\Lambda b], \quad [a_\Lambda S^ib]=(-1)^{p(a)+\overline{N}}(\theta^i+S^i)[a_\Lambda b]\quad (i\in[N]). \end{split} \end{align} \item (\emph{skew-symmetry}) For any $a, b\in V$, \begin{align}\label{eq:LCA:ssym} [b_\Lambda a]=-(-1)^{p(a)p(b)+\overline{N}}[a_{-\Lambda-\nabla}b], \end{align} where we used $\nabla \ceq (T, S^1, \ldots, S^N)$ in \eqref{eq:W:nabla}. \item (\emph{Jacobi identity}) For any $a, b, c\in V$, \begin{align}\label{eq:LCA:Jac} [a_{\Lambda_1}[b_{\Lambda_2}c]] =(-1)^{(p(a)+\overline{N})\overline{N}}[[a_{\Lambda_1}b]_{\Lambda_1+\Lambda_2}c] +(-1)^{(p(a)+\overline{N})(p(b)+\overline{N})}[b_{\Lambda_2}[a_{\Lambda_1}c]], \end{align} where $\Lambda_1, \Lambda_2$ are $(1|N)_W$-supervariables. \end{clist} The linear map $[\cdot_\Lambda\cdot]$ is called the \emph{$\Lambda$-bracket} of the $\mathcal{H}_W$-supermodule $(V, \nabla)$. For simplicity, we say $(V, \nabla)$, or more simply $V$, is an $N_W=N$ SUSY Lie conformal algebra. \end{dfn} \begin{rmk} A notational remark is in order. The linear map $[\cdot_\Lambda \cdot]$ can be expressed as \[ [a_\Lambda b] = \sum_{m \in \mathbb{N}, \, I \subset [N]} \Lambda^{m|I} c_{m,I}, \quad c_{m,I} \in V. \] Then the term $[a_{-\Lambda-\nabla}b]$ in \eqref{eq:LCA:ssym} is defined to be \[ [a_{-\Lambda-\nabla}b] \ceq \sum_{m \in \mathbb{N}, \, I \subset [N]} (-\Lambda-\nabla)^{m|I} c_{m,I} \] with $(-\Lambda-\nabla)^{m|I} \ceq (-\lambda-T)^m(-\theta^{i_1}-S^{i_1})\dotsm(-\theta^{i_r}-S^{i_r})$ for $I=\{i_1<\dotsc< i_r\} \subset [N]$. \end{rmk} In the remaining of this section, let $(V, \nabla)$ be an $\mathcal{H}_W$-supermodule and $\Lambda_k=(\lambda_k, \theta_k^1, \ldots, \theta_k^N)$ be a $(1|N)_W$-supervariable for each $k\in\mathbb{Z}_{>0}$. Next, we turn to the definition of the operad $\Chom^{N_W}_V$. First, the polynomial superalgebra $\mathbb{K}[\Lambda_k]_{k=1}^n$ of $(1|N)_W$-supervariables $\Lambda_1,\dotsc,\Lambda_n$ carries a structure of a right $\mathcal{H}_W$-supermodule by letting \begin{align*} f(\Lambda_1, \ldots, \Lambda_n)\cdot T \ceq f(\Lambda_1, \ldots, \Lambda_n)\Bigl(-\sum_{k=1}^n\lambda_k\Bigr), \quad f(\Lambda_1, \ldots, \Lambda_n)\cdot S^i\ceq f(\Lambda_1, \ldots, \Lambda_n)\Bigl(-\sum_{k=1}^n\theta_k^i\Bigr) \end{align*} for $f(\Lambda_1, \ldots, \Lambda_n) \in \mathbb{K}[\Lambda_k]_{k=1}^n$. Thus, for an $\mathcal{H}_W$-supermodule $V=(V,\nabla)$, we obtain a linear superspace \begin{align}\label{eq:LCA:HWn-mod} V_{\nabla}[\Lambda_k]_{k=1}^n \ceq \mathbb{K}[\Lambda_k]_{k=1}^n \otimes _{\mathcal{H}_W} V. \end{align} By the definition of the right action of $\mathcal{H}_W$ on $\mathbb{K}[\Lambda_k]_{k=1}^n$, we have: \begin{lem}\label{lem:W:VnLn=VLn-1} The linear superspace $V_{\nabla}[\Lambda_k]_{k=1}^n$ is isomorphic to $V[\Lambda_k]_{k=1}^{n-1}$ by the linear map \begin{align*} f(\Lambda_1, \cdots, \Lambda_n)\otimes v \longmapsto f(\Lambda_1, \cdots, \Lambda_{n-1}, -\Lambda_1-\cdots-\Lambda_{n-1}-\nabla)v \quad (f\in\mathbb{K}[\Lambda]_{k=1}^n, v\in V). \end{align*} \end{lem} Let us explain that $V_{\nabla}[\Lambda_k]_{k=1}^n$ has a natural left $\mathcal{H}_W^{\otimes n}$-supermodule structure. Indeed, since any element of $\mathcal{H}_W$ is expressed uniquely as $P(\nabla)$ for some $P(\Lambda) \in \mathbb{K}[\Lambda]$, the space $\mathbb{K}[\Lambda_k]_{k=1}^n$ carries a structure of a left $\mathcal{H}_W^{\otimes n}$-supermodule by letting $P_1(\nabla)\otimes \cdots \otimes P_n(\nabla)\in\mathcal{H}_W^{\otimes n}$ act on $f(\Lambda_1, \ldots, \Lambda_n) \in \mathbb{K}[\Lambda_k]_{k=1}^n$ as \[ \bigl(P_1(\nabla)\otimes \cdots \otimes P_n(\nabla)\bigr)f(\Lambda_1 \ldots, \Lambda_n) \ceq P_1(-\Lambda_1)\cdots P_n(-\Lambda_n)f(\Lambda_1, \ldots, \Lambda_n). \] Thus, by the action on the first component of the tensor product \eqref{eq:LCA:HWn-mod}, we obtain a left $\mathcal{H}_W^{\otimes n}$-supermodule $V_{\nabla}[\Lambda]_{k=1}^n$. The tensor product $V^{\otimes n}$ is naturally a left $\mathcal{H}_W^{\otimes n}$-supermodule. Thus, the following definition makes sense, which is a SUSY analogue of the operad $\Chom$ in \cite[\S5.2]{BDHK}. \begin{dfn} For an $\mathcal{H}_W$-supermodule $V=(V,\nabla)$ and $n \in \mathbb{N}$, we denote \[ \Chom^{N_W}_V(n) \ceq \Hom_{\mathcal{H}_W^{\otimes n}}(V^{\otimes n}, V_{\nabla}[\Lambda_k]_{k=1}^n). \] In other words, $\Chom^{N_W}_{V}(n)$ is a linear sub-superspace of $\Hom_\mathbb{K}(V^{\otimes n}, V_{\nabla}[\Lambda_k]_{k=1}^n)$ spanned by elements $X$ such that \begin{align*} X(\varphi v) = (-1)^{p(\varphi)p(X)}\varphi X(v) \quad (\varphi \in \mathcal{H}_{W}^{\otimes n}, \ v \in V^{\otimes n}). \end{align*} To stress the variables $\Lambda_1,\ldots,\Lambda_n$, we express an element $X \in \Chom^{N_W}_{V}(n)$ as $X_{\Lambda_1,\ldots,\Lambda_n}$. \end{dfn} For $\sigma \in \mathfrak{S}_n$ and $X \in \Chom^{N_W}_V(n)$, we define a linear map $X^\sigma\colon V^{\otimes n} \to V_{\nabla}[\Lambda_k]_{k=1}^n$ by \begin{align*} X^\sigma(v_1\otimes \cdots \otimes v_n) \ceq X_{\sigma(\Lambda_1, \ldots, \Lambda_n)}\bigl(\sigma(v_1\otimes \cdots \otimes v_n)\bigr) \quad (v_1,\ldots,v_n \in V), \end{align*} where $\sigma(\Lambda_1, \ldots, \Lambda_n)\ceq (\Lambda_{\sigma^{-1}(1)}, \ldots, \Lambda_{\sigma^{-1}(n)})$ and $\sigma(v_1 \otimes \dotsb \otimes v_n) \ceq \pm v_{\sigma^{-1}(1)} \otimes \dotsb \otimes v_{\sigma^{-1}(n)}$ given in \eqref{eq:1:Sn-Vn}. Then one can check $\sigma(\varphi v)=(\sigma\varphi)(\sigma v)$ for $\varphi\in\mathcal{H}_W^{\otimes n}$ and $v\in V^{\otimes n}$, by which one has $X^\sigma \in \Chom^{N_W}_{V}(n)$. Hence, the linear superspace $\Chom^{N_W}_V(n)$ carries a structure of a right $\mathfrak{S}_n$-supermodule by this action. For $X \in \Chom^{N_W}_V(k)$ and $Y_i \in \Chom^{N_W}_V(n_i)$ ($i\in[k]$), we define a linear map \begin{align*} X(Y_1\otimes \cdots\otimes Y_k)\colon V^{\otimes (n_1+\cdots +n_k)} \longrightarrow V_{\nabla}[\Lambda_l]_{l=1}^{n_1+\cdots+n_k} \end{align*} by letting \begin{align*} &X(Y_1\otimes \cdots\otimes Y_k)(v_1\otimes \cdots \otimes v_{n_1+\cdots+n_k}) \ceq \\ &X_{\Lambda_1', \ldots, \Lambda_k'}\bigl( ((Y_1)_{\Lambda_1, \ldots, \Lambda_{N_1}}\otimes \cdots \otimes (Y_k)_{\Lambda_{N_{k-1}+1}, \ldots, \Lambda_{N_k}}) ((v_1\otimes \cdots \otimes v_{N_1}) \otimes \cdots \otimes (v_{N_{k-1}+1}\otimes \cdots \otimes v_{N_k}))\bigr) \end{align*} for $v_1, \ldots, v_{n_1+\cdots+n_k}\in V$, where we used $N_i \ceq n_1+\cdots+n_i$ and $\Lambda_i'\ceq \Lambda_{N_{i-1}+1}+\cdots+\Lambda_{N_i}$. Then a direct calculation shows that $X(Y_1 \otimes \cdots \otimes Y_k) \in \Chom^{N_W}_V(n_1+\cdots+n_k)$. \begin{rmk} Let us explain the meaning of the above definition of $X(Y_1 \otimes \cdots \otimes Y_k)$. First, notice that we can identify $V_{\nabla}[\Lambda_l]_{l=1}^{N_1} \otimes \cdots \otimes V_{\nabla}[\Lambda_l]_{l=N_{k-1}+1}^{N_k}$ with $\mathbb{K}[\Lambda_l]_{l=1}^{N_k} \otimes_{\mathcal{H}_W^{\otimes k}} V^{\otimes k}$ as a linear superspace by the isomorphism defined by \begin{align*} (f_1 v_1)\otimes\cdots\otimes(f_k v_k) \longmapsto \prod_{1 \le i<j \le k} (-1)^{p(v_i)p(f_j)} \cdot (f_1\cdots f_k) \otimes (v_1\otimes \cdots \otimes v_k) \end{align*} for $f_i \in \mathbb{K}[\Lambda_l]_{l=N_{i-1}+1}^{N_i}$ and $v_i\in V$. Thus we can consider $(Y_1)_{\Lambda_1, \ldots, \Lambda_{N_1}}\otimes \cdots\otimes (Y_k)_{\Lambda_{N_{k-1}+1}, \ldots, \Lambda_{N_k}}$ in the definition of $X(Y_1\otimes \cdots\otimes Y_k)$ as a map \begin{align*} (Y_1)_{\Lambda_1, \ldots, \Lambda_{N_1}}\otimes\cdots\otimes (Y_k)_{\Lambda_{N_{k-1}+1}, \ldots, \Lambda_{N_k}}\colon V^{\otimes N_k} \longrightarrow \mathbb{K}[\Lambda_l]_{l=1}^{N_k} \otimes_{\mathcal{H}_W^{\otimes k}} V^{\otimes k}. \end{align*} Second, $X_{\Lambda_1', \ldots, \Lambda_k'}$ in the definition of $X(Y_1\otimes \cdots\otimes Y_k)$ can be considered as a map \begin{align*} (\mu\otimes \id_V)\circ(\id\otimes X_{\Lambda_1', \ldots, \Lambda_k'})\colon \mathbb{K}[\Lambda_l]_{l=1}^{N_k}\otimes_{\mathcal{H}^{\otimes k}} V^{\otimes k} \longrightarrow V_{\nabla}[\Lambda_l]_{l=1}^{N_k}, \end{align*} where $\id$ is the identity map of $\mathbb{K}[\Lambda_l]_{l=1}^{N_k}$ and \begin{align*} \mu\colon \mathbb{K}[\Lambda_l]_{l=1}^{N_k} \otimes_{\mathcal{H}^{\otimes k}} \mathbb{K}[\Lambda_l]_{l=1}^{N_k} \longrightarrow \mathbb{K}[\Lambda_l]_{l=1}^{N_k} \end{align*} is the linear map induced by the multiplication over $\mathbb{K}[\Lambda_l]_{l=1}^{N_k}$. \end{rmk} Now recall \cref{dfn:1:op} of operads. The $\mathfrak{S}_n$-supermodules $\Chom^{N_W}_V(n)$ give rise to: \begin{lem} For an $\mathcal{H}_W$-supermodule $V=(V,\nabla)$, the $\mathfrak{S}$-supermodule \[ \Chom^{N_W}_V \ceq \bigl(\Chom^{N_W}_{V}(n)\bigr)_{n\in\mathbb{N}} \] is an operad by letting $X \otimes Y_1 \otimes \cdots \otimes Y_k \mapsto X(Y_1 \otimes \cdots \otimes Y_k)$ be the composition map and $\id_V$ be the unit. \end{lem} \begin{proof} The proof in the non-SUSY case \cite[\S5.2]{BDHK} works with minor modification. We omit the detail. \end{proof} \begin{dfn}\label{dfn:W:Chom} The operad $\Chom^{N_W}_V$ is called the \emph{$N_W=N$ SUSY Lie conformal operad} on the $\mathcal{H}_W$-supermodule $V=(V, \nabla)$. \end{dfn} For an $\mathcal{H}_W$-supermodule $V=(V,\nabla)$ and $k \in \mathbb{N}$, we denote by $\Pi^k V=(\Pi^kV,\nabla)$ the linear superspace $\Pi^k V$ equipped with the supermodule structure $\nabla=(T,S^1,\dotsc,S^N)$ over the commutative superalgebra $\mathcal{H}_W$. Also recall the set $\MC\bigl(L(\oQ)\bigr)$ of Maurer-Cartan solutions $X \square X = 0$, $X \in L^1(\oQ)$ in the graded Lie superalgebra $L(\oQ)$ associated to an operad $\oQ$. Now we can state the main claim of this \cref{ss:W:LCA}, which is a natural $N_W=N$ SUSY analogue of \cite[Proposition 5.1]{BDHK}. \begin{thm}\label{thm:W:LCA} Let $V=(V,\nabla)$ be an $\mathcal{H}_W$-supermodule, i.e., a linear superspace $V$ endowed with $T \in \End_{\mathbb{K}}(V)_{\overline{0}}$ and $S^i \in \End_{\mathbb{K}}(V)_{\overline{1}}$ ($i\in[N]$) satisfying \eqref{eq:W:TSi}. \begin{enumerate} \item For an odd element $X\in L^1\bigl(\Chom^{N_W}_{\Pi^{N+1}V}\bigr)$ such that $X\square X=0$, define a linear map $[\cdot_\Lambda\cdot]_X\colon V\otimes V\to V[\Lambda]$ by \begin{align}\label{eq:LCAstr} [a_\Lambda b]_X \ceq (-1)^{p(a)(\overline{N}+\overline{1})}X_{\Lambda, -\Lambda-\nabla}(a \otimes b) \quad (a,b \in V). \end{align} Then $(V, \nabla, [\cdot_\Lambda\cdot]_X)$ is a $N_W=N$ SUSY Lie conformal algebra. \item The map $X \mapsto [\cdot_\Lambda\cdot]_X$ gives a bijection \[ \MC\bigl(L\bigl(\Chom^{N_W}_{\Pi^{N+1}V}\bigr)\bigr)_{\overline{1}} \xrr{\sim} \{\text{$N_W=N$ SUSY Lie conformal algebra structures on $(V,\nabla)$}\}. \] \end{enumerate} \end{thm} \begin{proof} We denote $\widetilde{V} \ceq \Pi^{N+1}V=(\Pi^{N+1}V,\nabla)$ and $\Chom_{\widetilde{V}} \ceq \Chom^{N_W}_{\Pi^{N+1}V}$ for simplicity, and let $p,\widetilde{p}$ be the parity of $V,\widetilde{V}$ respectively. Consider the linear map $[\cdot_\Lambda\cdot]_X\colon V\otimes V \to V[\Lambda]$ defined by \cref{eq:LCAstr} for an odd element $X\in\Chom_{\widetilde{V}}(2)$. The linear map $[\cdot_\Lambda\cdot]_X$ has parity $\overline{N}$: Since $X$ is an odd linear map $X\colon \widetilde{V}\otimes \widetilde{V} \to \widetilde{V}_{\nabla}[\Lambda_k]_{k=1, 2}$, \begin{align*} p(X(a\otimes b))=\widetilde{p}(X(a\otimes b))+\overline{N}+\overline{1}=\widetilde{p}(a)+\widetilde{p}(b)+\overline{N}=p(a)+p(b)+\overline{N} \end{align*} holds for $a, b\in V$. The linear map $[\cdot_\Lambda\cdot]_X$ satisfies (i) in \cref{dfn:W:LCA}: This can be checked by direct calculation using \begin{align*} X(\varphi v)=(-1)^{\widetilde{p}(\varphi)}\varphi X(v) \quad (\varphi\in\mathcal{H}_{W}^{\otimes 2}, \, v \in \widetilde{V}^{\otimes 2}). \end{align*} For instance, we can calculate \begin{align*} [a^i_\Lambda S^i b] &=(-1)^{p(a)(\overline{N}+\overline{1})}X_{\Lambda, -\Lambda-\nabla}(a\otimes S^ib) =(-1)^{p(a)(\overline{N}+\overline{1})+\widetilde{p}(a)} X_{\Lambda, -\Lambda-\nabla}\bigl((\id_V \otimes S^i)(a\otimes b)\bigr) \\ &=(-1)^{p(a)(\overline{N}+\overline{1})+\widetilde{p}(a)+1}(\theta^i+S^i)X_{\Lambda,-\Lambda-\nabla}(a\otimes b) \\ &=(-1)^{p(a)+\overline{N}}(\theta^i+S^i)[a_\Lambda b]_X \quad (a, b\in V). \end{align*} Note now that the map $X \mapsto [\cdot_\Lambda\cdot]_X$ gives a bijective correspondence between the set of all odd elements $X\in \Chom_{\widetilde{V}}(2)$ and the set of all linear maps $[\cdot_\Lambda\cdot]: V\otimes V \to V[\Lambda]$ of parity $\overline{N}$ satisfying (i) in \cref{dfn:W:LCA}. The skew-symmetry \eqref{eq:LCA:ssym} $\mathbin{\Leftrightarrow}$ $X^\sigma=X$ for any $\sigma\in\mathfrak{S}_2$: The element $X$ satisfies $X^\sigma=X\ (\sigma\in\mathfrak{S}_2)$ if and only if \begin{align*} (-1)^{\widetilde{p}(a)\widetilde{p}(b)}X_{\Lambda_2, \Lambda_1}(b\otimes a) =X_{\Lambda_1, \Lambda_2}(a\otimes b)\quad(a, b\in V). \end{align*} Since \begin{align*} [b_\Lambda a]=(-1)^{p(b)(\overline{N}+\overline{1})}X_{\Lambda, -\Lambda-\nabla}(b\otimes a), \quad [a_{-\Lambda-\nabla} b] =(-1)^{p(a)(\overline{N}+\overline{1})}X_{-\Lambda-\nabla, \Lambda}(a\otimes b) \quad (a, b\in V), \end{align*} the skew-symmetry is equivalent to $X^\sigma=X$ $(\sigma\in\mathfrak{S}_2)$. In what follows, let $X$ be an odd element $X\in L^1\bigl(\Chom_{\widetilde{V}}\bigr)$. Thus $[\cdot_\Lambda\cdot]_X\colon V\otimes V \to V_{\nabla}[\Lambda]$ is a linear map of parity $\overline{N}$ satisfying (i), (ii) in \cref{dfn:W:LCA}. The Jacobi identity \eqref{eq:LCA:Jac}$\mathbin{\Leftrightarrow}$ $X\square X=0$: For $a, b, c\in V$, we have \begin{align*} &(X\square X)(a\otimes b\otimes c)\\ &=(X\circ_1X)(a\otimes b\otimes c)+(X\circ_2X)(a\otimes b\otimes c)+(X\circ_2X)^{(1, 2)}(a\otimes b\otimes c)\\ &=X_{\Lambda_1+\Lambda_2, \Lambda_3}(X_{\Lambda_1, \Lambda_2}(a\otimes b)\otimes c) +(-1)^{\widetilde{p}(a)}X_{\Lambda_1, \Lambda_2+\Lambda_3}(a\otimes X_{\Lambda_2, \Lambda_3}(b\otimes c)) \\ &\qquad +(-1)^{\widetilde{p}(a)\widetilde{p}(b)+\widetilde{p}(b)}X_{\Lambda_2, \Lambda_1+\Lambda_3}(b\otimes X_{\Lambda_1, \Lambda_3}(a\otimes c))\\ &=(-1)^{p(a)\overline{N}+(p(b)+\overline{1})(\overline{N}+\overline{1})}\bigl([a_{\Lambda_1}[b_{\Lambda_2}c]] -(-1)^{(p(a)+\overline{N})\overline{N}}[[a_{\Lambda_1}b]_{\Lambda_1+\Lambda_2}c] -(-1)^{(p(a)+\overline{N})(p(b)+\overline{N})}[b_{\Lambda_2}[a_{\Lambda_1}c]]\bigr). \end{align*} Thus the Jacobi identity is equivalent to $X\square X=0$. \end{proof} \begin{rmk} Combined with \cref{prp:1:LP}, the theorem says that a Lie conformal algebra structure on $(V,\nabla)$ is an odd Lie algebra structure on the operad $\Chom_{\Pi^{N+1}V}^{N_W}$. The condition being odd comes from \cref{dfn:W:LCA} of $\Lambda$-bracket to have parity $\overline{N}$. \end{rmk} \subsection{\texorpdfstring{$N_W=N$}{NW=N} SUSY chiral operad}\label{ss:W:VA} Here is the main part of this \cref{s:W}, where we will introduce a natural $N_W=N$ SUSY analogue of the operad of vertex algebra in \cite[\S6]{BDHK}. Let us fix a $(1|N)_W$-supervariable $\Lambda=(\lambda, \theta^1, \ldots, \theta^N)$. For even linear transformations $F$ and $G$ on a linear superspace $V$, we define a linear map $\int_F^G d\Lambda\colon V[\Lambda] \to V$ by \begin{align}\label{eq:W:intLv} \int_F^G d\Lambda\ \Lambda^{m|I}v\ceq\frac{\delta_{I, [N]}}{m+1}(G^{m+1}v-F^{m+1}v) \quad (m\in\mathbb{N}, \, I\subset [N], \, v\in V). \end{align} The linear map $\int_F^G d\Lambda$ has the parity $\overline{N}$. Also, if $V$ is a superalgebra (not necessarily unital nor associative), we define a linear map $\int_F^G d\Lambda \, a\colon V[\Lambda] \to V$ for $a\in V$ by \begin{align*} \Bigl(\int_F^G d\Lambda \, a\Bigr)\Lambda^{m|I}v \ceq \Bigl(\int_F^G d\Lambda\ \Lambda^{m|I}a\Bigr)v, \end{align*} where the term $\int_F^G d\Lambda\ \Lambda^{m|I}a$ in the right hand side is given by \eqref{eq:W:intLv}. Then the linear map $\int_F^G d\Lambda \,a$ has the parity $p(a)$. Using this integral, we introduce: \begin{dfn}[{\cite[Definition 3.3.15]{HK}}]\label{dfn:NWVA} Let $(V,\nabla,[\cdot_\Lambda\cdot])=(V,T,S^1,\dotsc,S^N,[\cdot_\Lambda\cdot])$ be an $N_W=N$ SUSY Lie conformal algebra (\cref{dfn:W:LCA}) and $\mu\colon V \otimes V \to V$ be an even linear map. We denote $a b\ceq\mu(a\otimes b)$ for $a, b\in V$. A tuple $(V,\nabla,[\cdot_\Lambda\cdot],\mu)$ is called a \emph{non-unital $N_W=N$ SUSY vertex algebra} if it satisfies the following conditions: \begin{clist} \item For any $a, b\in V$, \begin{align}\label{eq:VA:der} T(ab)=(Ta)b+a(Tb), \quad S^i(ab)=(S^ia)b+(-1)^{p(a)}a(S^ib)\quad(i\in[N]). \end{align} \item (\emph{quasi-commutativity}) For any $a, b\in V$, \begin{align}\label{eq:VA:qcom} ab-(-1)^{p(a)p(b)}ba=\int_{-T}^0d\Lambda\ [a_\Lambda b]. \end{align} \item (\emph{quasi-associativity}) For any $a, b, c\in V$, \begin{align}\label{eq:VA:qass} (ab)c-a(bc) = \left(\int_0^Td\Lambda a\right)[b_\Lambda c] +(-1)^{p(a)p(b)}\left(\int_0^Td\Lambda b\right)[a_\Lambda c]. \end{align} \item (\emph{Wick formula}) For any $a, b, c\in V$, \begin{align}\label{eq:VA:Wick} [a_\Lambda bc]=[a_\Lambda b]c+(-1)^{(p(a)+\overline{N})p(b)}b[a_\Lambda c]+\int_0^\lambda d\Gamma [[a_\Lambda b]_\Gamma c], \end{align} where $\Gamma$ is an additional $(1|N)_W$-supervariable. \end{clist} For simplicity, we say $(V, \nabla)$, or more simply $V$, is a non-unital $N_W=N$ SUSY vertex algebra. The map $\mu$ is called the \emph{multiplication} of $V$. \end{dfn} \begin{dfn} A non-unital $N_W=N$ SUSY vertex algebra $V$ is called an \emph{$N_W=N$ SUSY vertex algebra} if there exists an even element $\vac \in V$ such that $a\vac=\vac a=a$ for all $a\in V$. \end{dfn} In the remaining of this \cref{ss:W:VA}, let $(V,\nabla)$ be an $\mathcal{H}_W$-supermodule and $\Lambda_k=(\lambda_k, \theta_k^1, \ldots, \theta_k^N)$ be a $(1|N)_W$-supervariable for each $k\in\mathbb{N}$. Now we give several preliminaries to introduce the $N_W=N$ SUSY chiral operad (see \cref{dfn:W:Pch}). \cref{lem:rWick} and \cref{lem:lsym} below are SUSY analogues of \cite[(1.38), (1.41)]{DK}. \begin{lem}\label{lem:rWick} Let $[\cdot_\Lambda\cdot]: V\otimes V \to V[\Lambda]$ be a linear map of parity $\overline{N}$ satisfying the sesquilinearity \eqref{eq:LCA:sesq} and $\mu: V\otimes V\to V$ be an even linear map satisfying \eqref{eq:VA:der}. In addition, we assume that $[\cdot_\Lambda\cdot]$ and $\mu$ satisfy the skew symmetry \eqref{eq:LCA:ssym} and the quasi-commutativity \eqref{eq:VA:qcom}. Then the Wick formula \eqref{eq:VA:Wick} is equivalent to the following identity which is called the \emph{right Wick formula}: \begin{align*} [ab_\Lambda c] =(-1)^{p(a)\overline{N}}(e^{\nabla\cdot\partial_{\Lambda}}a)[b_\Lambda c] +(-1)^{(p(a)+\overline{N})p(b)}(e^{\nabla\cdot\partial_{\Lambda}}b)[a_\Lambda c] +(-1)^{(p(a)+\overline{N})p(b)}\int_0^\lambda d\Gamma[b_\Gamma[a_{\Lambda-\Gamma}c]]. \end{align*} Here we denoted $\partial_\Lambda\ceq(\partial_\lambda, \partial_{\theta^1}, \ldots, \partial_{\theta^N})$, $\nabla\cdot\partial_\Lambda\ceq T\partial_\lambda+\sum_{i=1}^NS^i\partial_{\theta^i}$, and for $a\in V$, we defined a linear map $e^{\nabla\cdot\partial_{\Lambda}}a\colon V[\Lambda] \to V[\Lambda]$ by \begin{align*} \bigl(e^{\nabla\cdot\partial_{\Lambda}}a\bigr)\Lambda^{m|I}v \ceq (-1)^{p(a) \cdot \overline{\#I}}\bigl(e^{\nabla\cdot\partial_{\Lambda}}\Lambda^{m|I}a\bigr)v \quad (m \in \mathbb{N}, \, I \subset [N], \, v \in V). \end{align*} \end{lem} \begin{proof} If the Wick formula holds, then \begin{align*} [ab_\Lambda c] &=-(-1)^{(p(a)+p(b))p(c)+\overline{N}}[c_{-\Lambda-\nabla}ab]\\ &=-(-1)^{(p(a)+p(b))p(c)+\overline{N}}e^{\nabla\cdot\partial_{\Lambda}}[c_{-\Lambda}ab]\\ &=-(-1)^{(p(a)+p(b))p(c)+\overline{N}}e^{\nabla\cdot\partial_{\Lambda}}\Bigl([c_{-\Lambda}a]b+(-1)^{(p(c)+\overline{N})p(a)}a[c_{-\Lambda}b]+\int_0^{-\lambda}d\Gamma[[c_{-\Lambda}a]_{\Gamma}b]\Bigr)\\ &=-(-1)^{p(a)p(c)+\overline{N}+(p(a)+\overline{N})p(b)}e^{\nabla\cdot\partial_\Lambda}b[c_{-\Lambda}a] -(-1)^{p(b)p(c)+\overline{N}+p(a)\overline{N}}e^{\nabla\cdot\partial_\Lambda}a[c_{-\Lambda}b]\\ &\hspace{193pt} -(-1)^{(p(a)+p(b))p(c)+\overline{N}}e^{\nabla\cdot\partial_\Lambda}\int_{-T}^{-\lambda}[[c_{-\Lambda}a]_\Gamma b]. \end{align*} Here we used the skew-symmetry in the first equality, the Wick formula in the third equality, and the quasi-commutativity in the last equality. A direct calculation shows \begin{align*} &-(-1)^{p(a)p(c)+\overline{N}}e^{\nabla\cdot\partial_\Lambda}b[c_{-\Lambda}a] =\bigl(e^{\nabla\cdot\partial_\Lambda}b\bigr)[a_\Lambda c], \\ &-(-1)^{(p(a)+p(b))p(c)+\overline{N}}e^{\nabla\cdot\partial_\Lambda} \int_{-T}^{-\lambda}[[c_{-\Lambda}a]_\Gamma b] =(-1)^{(p(a)+\overline{N})p(b)}\int_0^\lambda d\Gamma[b_\Gamma[a_{\Lambda-\Gamma}c]], \end{align*} by which the right Wick formula holds. We can similarly prove the converse. \end{proof} \begin{lem}\label{lem:lsym} Let $[\cdot_\Lambda\cdot]\colon V\otimes V \to V[\Lambda]$ and $\mu\colon V\otimes V\to V$ be the same as in \cref{lem:rWick}. In addition, we assume that $[\cdot_\Lambda\cdot]$ and $\mu$ satisfy the skew-symmetry \eqref{eq:LCA:ssym}, the quasi-commutativity \eqref{eq:VA:qcom} and the Wick formula \eqref{eq:VA:Wick}. Then the quasi-associativity \eqref{eq:VA:qass} is equivalent to the following identity: \begin{align}\label{eq:lsym} a(bc)-(-1)^{p(a)p(b)}b(ac)=(ab-(-1)^{p(a)p(b)}ba)c \end{align} \end{lem} \begin{proof} Since the right-hand side of the quasi-associativity is symmetric with respect to $a$ and $b$, it is clear that the quasi-associativity implies the identity \eqref{eq:lsym}. Conversely, we assume the identity \eqref{eq:lsym}. By the quasi-commutativity and the identity \eqref{eq:lsym}, we have \begin{align*} (ab)c-a(bc) &=(-1)^{(p(a)+p(b))p(c)}c(ab)-(-1)^{p(b)p(c)}a(cb) +\int_{-T}^0d\Lambda[ab_\Lambda c]-a\int_{-T}^0d\Lambda [b_\Lambda c]\\ &=-(-1)^{p(b)p(c)}\Bigl(\int_{-T}^0d\Lambda[a_\Lambda c]\Bigr)b +\int_{-T}^0d\Lambda[ab_\Lambda c]-a\int_{-T}^0d\Lambda [b_\Lambda c]\\ &=(-1)^{p(a)p(b)}\int_{-T}^0d\Gamma\Bigl[b_{\Gamma}\Bigl(\int_{-T}^0d\Lambda [a_\Lambda c]\Bigr)\Bigr]\\ &\quad-(-1)^{p(a)p(b)}b\int_{-T}^0d\Lambda [a_\Lambda c] +\int_{-T}^0d\Lambda[ab_\Lambda c]-a\int_{-T}^0d\Lambda [b_\Lambda c]. \end{align*} We can calculate the double integral as \begin{align*} \int_{-T}^0d\Lambda_2\Bigl[b_{\Lambda_2}\Bigl(\int_{-T}^0d\Lambda_1 [a_{\Lambda_1} c]\Bigr)\Bigr] &=(-1)^{(p(b)+\overline{N})\overline{N}}\int_{-T}^0d\Lambda_2\int_{-\lambda_2-T}^0d\Lambda_1[b_{\Lambda_2}[a_{\Lambda_1} c]]\\ &=(-1)^{(p(b)+\overline{N})\overline{N}}\int_{-T}^0d\Lambda_2\int_{-T}^{\lambda_2}d\Lambda_1[b_{\Lambda_2}[a_{\Lambda_1-\Lambda_2} c]]\\ &=(-1)^{p(b)N}\int_{-T}^0d\Lambda_1\int_{\lambda_1}^0d\Lambda_2[b_{\Lambda_2}[a_{\Lambda_1-\Lambda_2}c]], \end{align*} so by \cref{lem:rWick}, we obtain \begin{align*} &(-1)^{p(a)p(b)}\int_{-T}^0d\Gamma\Bigl[b_{\Gamma}\Bigl(\int_{-T}^0d\Lambda [a_{\Lambda} c]\Bigr)\Bigr]\\ &=-\int_{-T}^0d\Lambda[ab_\Lambda c] +(-1)^{p(a)\overline{N}}\int_{-T}^0d\Lambda(e^{\nabla\cdot\partial_\Lambda}a)[b_\Lambda c] +(-1)^{(p(a)+\overline{N})p(b)}\int_{-T}^0d\Lambda(e^{\nabla\cdot\partial_\Lambda}b)[a_\Lambda c]. \end{align*} Thus \begin{align*} (ab)c-a(bc) &=(-1)^{p(a)\overline{N}}\int_{-T}^0d\Lambda(e^{\nabla\cdot\partial_\Lambda}a)[b_\Lambda c] -a\int_{-T}^0d\Lambda [b_\Lambda c]\\ &\quad+(-1)^{p(a)p(b)}\Bigl((-1)^{p(b)\overline{N}}\int_{-T}^0d\Lambda(e^{\nabla\cdot\partial_\Lambda}b)[a_\Lambda c]-b\int_{-T}^0d\Lambda [a_\Lambda c]\Bigr). \end{align*} Since a direct calculation shows \begin{align*} (-1)^{p(a)\overline{N}}\int_{-T}^0d\Lambda(e^{\nabla\cdot\partial_\Lambda}a)[b_\Lambda c] -a\int_{-T}^0d\Lambda [b_\Lambda c] =\Bigl(\int_0^Td\Lambda a\Bigr)[b_\Lambda c], \end{align*} the identity \eqref{eq:lsym} implies the quasi-associativity. \end{proof} Next we introduce the indefinite integral of the $\Lambda$-bracket (\cref{dfn:W:indef}), and translate the definition of $N_W=N$ SUSY vertex algebra using the integral (\cref{prp:W:skecom,prp:W:Jqas}). As a preliminary, let us show: \begin{lem}\label{lem:intbra} For a linear map $[\cdot_\Lambda\cdot]\colon V \otimes V \to V[\Lambda]$ of parity $\overline{N}$ satisfying the sesquilinearity \eqref{eq:LCA:sesq} and an even linear map $\mu\colon V\otimes V\to V$ satisfying \eqref{eq:VA:der}, there exists a unique linear map $F\colon V\otimes V \to V[\Lambda]$ of parity $\overline{N}$ satisfying the following conditions for every $a, b\in V$: \begin{align*} &\Res_\Lambda(\lambda^{-1}F(S^i a \otimes b))= -(-1)^N \Res_\Lambda(\lambda^{-1}\theta^i F(a \otimes b)) \quad (i \in [N]), \\ &\partial_\lambda F(a \otimes b) = [a_\Lambda b], \quad \Res _\Lambda\left(\lambda^{-1}F(a \otimes b)\right) =ab. \end{align*} \end{lem} \begin{proof} Let us define a linear map $F\colon V \otimes V \to V[\Lambda]$ by \begin{align*} F(a \otimes b)\ceq \sum_{I \subset [N]}(-1)^{\# I \cdot (N+1)}\sigma(I)\theta^{[N] \mathbin{\setminus} I}(S^Ia)b +\int_0^\lambda d\lambda \,[a_\Lambda b] \quad (a,b \in V). \end{align*} Then $F$ has the parity $\overline{N}$ and satisfies the above three conditions. Also, if linear maps $F, G\colon V\otimes V \to V[\Lambda]$ satisfy the above three conditions, then we have the identities \begin{align*} \partial_\lambda F(a\otimes b)=\partial_\lambda G(a\otimes b), \quad \Res_\Lambda(\lambda^{-1}\theta^IF(a\otimes b)) =\Res_\Lambda(\lambda^{-1}\theta^IG(a\otimes b)) \quad (I\subset [N]), \end{align*} which yield $F=G$. \end{proof} \begin{dfn}\label{dfn:W:indef} Given a linear map $[\cdot_\Lambda\cdot]\colon V \otimes V \to V[\Lambda]$ of parity $\overline{N}$ satisfying the sesquilinearity \eqref{eq:LCA:sesq} and an even linear map $\mu\colon V\otimes V\to V$ satisfying \eqref{eq:VA:der}, we denote by \[ \int^\Lambda d\Gamma[\cdot_\Gamma \cdot]\colon V \otimes V \longrightarrow V[\Lambda] \] the linear map $F$ in the \cref{lem:intbra}, and call it the integral of $[\cdot_\Lambda \cdot]$ (with respect to $\mu$). \end{dfn} \cref{prp:W:skecom,prp:W:Jqas} below are SUSY analogues of \cite[Definition 1.24]{DK}. \begin{prp}\label{prp:W:skecom} Let $[\cdot_\Lambda\cdot]\colon V \otimes V \to V[\Lambda]$ be a linear map of parity $\overline{N}$ satisfying the sesquilinearity \eqref{eq:LCA:sesq}, and $\mu\colon V \otimes V\to V$ be an even linear map satisfying \eqref{eq:VA:der}. Then the skew-symmetry \eqref{eq:LCA:ssym} and the quasi-commutativity \eqref{eq:VA:qcom} are equivalent to the following identity: \begin{align}\label{eq:W:skecom} \int^\Lambda d\Gamma [b_\Gamma a] = (-1)^{p(a)p(b)+\overline{N}}\int^{-\Lambda-\nabla}d\Gamma [a_\Gamma b]. \end{align} \end{prp} \begin{proof} By a direct calculation, we get \begin{align*} &\Res_\Lambda\Bigl(\lambda^{-1}\int^{-\Lambda-\nabla}d\Gamma[a_\Gamma S^ib]\Bigr) =(-1)^{p(a)}\Res_\Lambda\Bigl(\lambda^{-1}\theta^i\int^{-\Lambda-\nabla}d\Gamma[a_\Gamma b]\Bigr), \\ &\partial_\lambda\int^{-\Lambda-\nabla}d\Gamma[a_\Gamma b]=-[a_{-\Lambda-\nabla} b], \quad \Res_{\Lambda}\Bigl(\lambda^{-1}\int^{-\Lambda-\nabla}d\Gamma[a_\Gamma b]\Bigr) =(-1)^N\Bigl(ab-\int_{-T}^0d\Lambda[a_\Lambda b]\Bigr). \end{align*} Thus the skew-symmetry and the quasi-commutativity are equivalent to the identity \eqref{eq:W:skecom}. \end{proof} \begin{prp}\label{prp:W:Jqas} Let $[\cdot_\Lambda\cdot]\colon V \otimes V \to V[\Lambda]$ and $\mu\colon V \otimes V \to V$ be the same in \cref{prp:W:skecom}. In addition, we assume that $[\cdot_\Lambda\cdot]$ and $\mu$ satisfy the skew-symmetry and the quasi-commutativity. Then the Jacobi identity \eqref{eq:LCA:Jac}, the quasi-associativity \eqref{eq:VA:qass} and the Wick formula \eqref{eq:VA:Wick} are equivalent to the following identity: \begin{align}\label{eq:W:Jqas} \begin{split} &\int^{\Lambda_1}d\Gamma_1\left[a_{\Gamma_1}\left(\int^{\Lambda_2}d\Gamma_2[b_{\Gamma_2}c]\right)\right] -(-1)^{(p(a)+\overline{N})(p(b)+\overline{N})}\int^{\Lambda_2}d\Gamma_2\left[b_{\Gamma_2}\left(\int^{\Lambda_1}d\Gamma_1[a_{\Gamma_1}c]\right)\right]\\ &=(-1)^{(p(a)+\overline{N})\overline{N}}\int^{\Lambda_1+\Lambda_2}d\Gamma\left[\left(\int^{\Lambda_1}d\Gamma_1[a_{\Gamma_1}b]-\int^{-\Lambda_2-\nabla}d\Gamma_2[a_{\Gamma_2}b]\right)_{\Gamma}c\right]. \end{split} \end{align} \end{prp} \begin{proof} Let $F(a\otimes b\otimes c)$ (resp.\ $G(a\otimes b\otimes c)$) denote the left hand side (resp.\ the right hand side) of \cref{eq:W:Jqas}. The identity $\partial_{\lambda_1}\partial_{\lambda_2}F(a\otimes b\otimes c) =\partial_{\lambda_1}\partial_{\lambda_2}G(a\otimes b\otimes c)$ is equivalent to the Jacobi identity: In fact, a direct calculation yields \begin{align*} &\partial_{\lambda_1}\partial_{\lambda_2}F(a\otimes b\otimes c) =[a_{\Lambda_1}[b_{\Lambda_2}c]]-(-1)^{(p(a)+\overline{N})(p(b)+\overline{N})}[b_{\Lambda_2}[a_{\Lambda_1}b]], \\ &\partial_{\lambda_1}\partial_{\lambda_2}G(a\otimes b\otimes c) =(-1)^{(p(a)+\overline{N})\overline{N}}[[a_{\Lambda_1}b]_{\Lambda_1+\Lambda_2}c]. \end{align*} The identity $\partial_{\lambda_1}\Res_{\Lambda_2}(\lambda_2^{-1}F(a\otimes b\otimes c)) =\partial_{\lambda_1}\Res_{\Lambda_2}(\lambda_2^{-1}G(a\otimes b\otimes c))$ is equivalent to the Wick formula: In fact, one can get by direct calculation that \begin{align*} &\partial_{\lambda_1}\Res_{\Lambda_2}(\lambda_2^{-1}F(a\otimes b\otimes c)) =(-1)^{(p(a)+\overline{N})\overline{N}}[a_{\Lambda_1}bc] -(-1)^{(p(a)+\overline{N})(p(b)+\overline{N})}b[a_{\Lambda_1}c],\\ &\partial_{\lambda_1}\Res_{\Lambda_2}(\lambda_2^{-1}G(a\otimes b\otimes c)) =(-1)^{(p(a)+\overline{N})\overline{N}} \left([a_{\Lambda_1}b]c+\int_0^{\lambda_1}d\Gamma [[a_{\Lambda_1}b]_\Gamma c]\right). \end{align*} The identity $\Res_{\Lambda_1}\Res_{\Lambda_2}(\lambda_1^{-1}\lambda_2^{-1}F(a\otimes b\otimes c)) =\Res_{\Lambda_1}\Res_{\Lambda_2}(\lambda_1^{-1}\lambda_2^{-1}G(a\otimes b\otimes c))$ is equivalent to the identity \eqref{eq:lsym}: In fact, we have \begin{align*} &\Res_{\Lambda_1}\Res_{\Lambda_2}(\lambda_1^{-1}\lambda_2^{-1}F(a\otimes b\otimes c)) =(-1)^{(p(a)+\overline{N})\overline{N}}(a(bc)-(-1)^{p(a)p(b)}b(ac)),\\ &\Res_{\Lambda_1}\Res_{\Lambda_2}(\lambda_1^{-1}\lambda_2^{-1}G(a\otimes b\otimes c)) =(-1)^{(p(a)+\overline{N})\overline{N}}(ab-(-1)^{p(a)p(b)}ba)c. \end{align*} Since we have \begin{align*} \Res_{\Lambda_1}(\lambda_1^{-1}\theta_1^IF(a\otimes b\otimes c)) &=(-1)^{\#I(N+1)}\Res_{\Lambda_1}(\lambda_1^{-1}F(S^Ia\otimes b\otimes c)), \\ \Res_{\Lambda_1}(\lambda_1^{-1}\theta_1^IG(a\otimes b\otimes c)) &=(-1)^{\#I(N+1)}\Res_{\Lambda_1}(\lambda_1^{-1}G(S^Ia\otimes b\otimes c)), \end{align*} and \begin{align*} \Res_{\Lambda_2}(\lambda_2^{-1}\theta_2^IF(a\otimes b\otimes c)) &=(-1)^{\overline{\#I}(p(a)+\overline{1})}\Res_{\Lambda_2}(\lambda_2^{-1}F(a\otimes S^Ib\otimes c)), \\ \Res_{\Lambda_2}(\lambda_2^{-1}\theta_2^IG(a\otimes b\otimes c)) &=(-1)^{\overline{\#I}(p(a)+\overline{1})}\Res_{\Lambda_2}(\lambda_2^{-1}G(a\otimes S^Ib\otimes c)), \end{align*} for any $I\subset [N]$, the identity $F(a\otimes b\otimes c)=G(a\otimes b\otimes c)$ holds if and only if the following identities hold: \begin{gather*} \partial_{\lambda_1}\partial_{\lambda_2}F(a\otimes b\otimes c) =\partial_{\lambda_1}\partial_{\lambda_2}G(a\otimes b\otimes c), \\ \partial_{\lambda_1}\Res_{\Lambda_2}(\lambda_2^{-1}F(a\otimes b\otimes c)) =\partial_{\lambda_1}\Res_{\Lambda_2}(\lambda_2^{-1}G(a\otimes b\otimes c)), \\ \Res_{\Lambda_1}\Res_{\Lambda_2}(\lambda_1^{-1}\lambda_2^{-1}F(a\otimes b\otimes c)) =\Res_{\Lambda_1}\Res_{\Lambda_2}(\lambda_1^{-1}\lambda_2^{-1}G(a\otimes b\otimes c)). \end{gather*} Hence, by \cref{lem:lsym}, the Jacobi identity, the quasi-associativity and the Wick formula are equivalent to the identity \eqref{eq:W:Jqas}. \end{proof} Now we introduce several superalgebras which are $N_W=N$ SUSY analogue of those in \cite[\S6.3]{BDHK}. \begin{dfn}\label{dfn:W:O} Let $n \in \mathbb{N}$, and $Z_k=(z_k, \zeta_k^1, \ldots, \zeta_k^N)$ be a $(1|N)_{W}$-supervariable for each $k \in [n]$. We set \begin{align}\label{eq:W:zkl} z_{k,l}\ceq z_k-z_l, \quad \zeta_{k,l}^{i} \ceq \zeta_k^i-\zeta_l^i \end{align} for $k,l \in [n]$ and $i \in [N]$. Also, we set $Z_{k,l} \ceq (z_{k,l},\zeta_{k,l}^1,\dotsc,\zeta_{k,l}^N)$ for simplicity. \begin{enumerate} \item We denote by \[ \mathcal{O}_n^{\star} \ceq \mathbb{K}[Z_k]_{k=1}^n[z_{k, l}^{-1}]_{1 \le k < l \le n} \] the localization of $\mathbb{K}[Z_k]_{k=1}^n$ by the multiplicatively closed set generated by $\{z_{k,l} \mid 1 \le k<l \le n\}$. \item Let $\mathcal{O}_n^{\star \mathrm{T}}$ denote the sub-superalgebra of $\mathcal{O}_n^{\star}$ generated by the subset $\{z_{k, l}^{\pm 1}\mid 1\le k<l\le n\} \cup \{\zeta_{k,l}^i \mid i \in [N], \, 1 \le k<l \le n\}$. Thus, we have \[ \mathcal{O}_n^{\star \mathrm{T}} = \mathbb{K}[z_{k,l}^{\pm1},\zeta_{k,l}^i \mid i \in [N], \ 1 \le k<l \le n]. \] \item Let $\mathcal{D}_n$ denote the superalgebra of regular differential operators of $Z_k$, i.e, the sub-superalgebra of $\End_{\mathbb{K}}(\mathbb{K}[Z_k]_{k=1}^n)$ generated by $Z_k=(z_k,\zeta_k^1,\dotsc,\zeta_k^N)$ and $\partial_{Z_k}=(\partial_{z_k}, \partial_{\zeta_k^1}, \ldots, \partial_{\zeta_k^N})$ for $k \in [n]$. As a superspace, we have \begin{align*} \mathcal{D}_n = \mathbb{K}[Z_k]_{k=1}^n [\partial_{Z_k}]_{k=1}^n = \mathbb{K}[z_k,\zeta_k^1,\dotsc,\zeta_k^N]_{k=1}^n [\partial_{z_k},\partial_{\zeta_k^1},\dotsc,\partial_{\zeta_k^N}]_{k=1}^n. \end{align*} Here $\mathbb{K}[Z_k]_{k=1}^n[\partial_{Z_k}]_{k=1}^n$ is the free commutative superalgebra over $\mathbb{K}[Z_k]_{k=1}^n$ generated by even $\partial_{z_k}$ and odd $\partial_{\zeta_k^i}$ for $i \in [N]$ and $k \in [n]$. \item Let $\mathcal{D}_n^{\mathrm{T}}$ denote the sub-superalgebra of $\mathcal{D}_n$ generated by the subset $\{z_{k,l}, \zeta_{k, l}^i\mid i\in[N], 1\le k<l\le n\}\cup \{\partial_{z_k}, \partial_{\zeta_k^i}\mid i\in[N], k\in[n]\}$. Thus we have \begin{align*} \mathcal{D}_n^{\mathrm{T}}=\mathbb{K}[Z_{k, l}]_{1\le k<l\le n}[\partial_{Z_k}]_{k=1}^n. \end{align*} \end{enumerate} \end{dfn} As in the non-SUSY case \cite[\S6.2]{BDHK}, the superalgebra $\mathcal{O}_n^{\star\mathrm{T}}$ is the translation invariant subalgebra of $\mathcal{O}_n^{\star}$. In other words, we have \begin{align}\label{eq:W:Ker-delta} \mathcal{O}_n^{\star\mathrm{T}} = \Ker(\delta^0) \cap \Ker(\delta^1) \cap \dotsb \cap \Ker(\delta^N) \subset \mathcal{O}_n^{\star} \end{align} with $\delta^0 \ceq \sum_{k=1}^n \partial_{z_k}$ and $\delta^i \ceq \sum_{k=1}^n \partial_{\zeta_k^i}$ for $i \in [N]$. The superspace $V^{\otimes n}\otimes \mathcal{O}_n^{\star\mathrm{T}}$ carries a structure of a right $\mathcal{D}_n$-supermodule by letting $Z_k=(z_k, \zeta_k^1, \ldots, \zeta_k^N)$ act as \begin{align*} (v\otimes f)\cdot z_k\ceq v\otimes fz_k, \quad (v\otimes f)\cdot \zeta_k^i \ceq v\otimes f\zeta_k^i, \end{align*} and $\partial_{Z_k}=\bigl(\partial_{z_k}, \partial_{\zeta_k^1}, \ldots, \partial_{\zeta_k^N}\bigr)$ act as \begin{align*} (v\otimes f)\cdot \partial_{z_k}\ceq T^{(k)}v\otimes f-v\otimes \partial_{z_k}f, \quad (v\otimes f)\cdot\partial_{\zeta_k^i}\ceq (-1)^{p(v)+p(f)}(S^i)^{(k)}v\otimes f+(-1)^{p(f)}v\otimes \partial_{\zeta_k^i}f \end{align*} for $v \in V^{\otimes n}$, $f \in \mathcal{O}_n^{\star\mathrm{T}}$. Here, for a linear transformation $\varphi \in \End V$, the symbol $\varphi^{(k)}$ denotes the linear transformation on $V^{\otimes n}$ defined by $\varphi^{(k)} \ceq \id_V \otimes \dotsb \otimes \overset{k}{\varphi} \otimes \dotsb \otimes \id_V$. Now let us recall the superspace $V_\nabla[\Lambda_k]_{k=1}^n$ in \eqref{eq:LCA:HWn-mod}. It also has a structure of right $\mathcal{D}_n$-supermodule by letting \begin{align*} a(\Lambda_1,\dotsc,\Lambda_n) \cdot z_k \ceq -\partial_{\lambda_k} a(\Lambda_1,\dotsc,\Lambda_n), \quad a(\Lambda_1,\dotsc,\Lambda_n) \cdot \zeta_k^i \ceq -(-1)^{p(a)}\partial_{\theta_k^i} a(\Lambda_1,\dotsc,\Lambda_n) \end{align*} and \begin{align*} a(\Lambda_1, \ldots, \Lambda_n)\cdot \partial_{z_k}\ceq -a(\Lambda_1, \ldots, \Lambda_n)\lambda_k, \quad a(\Lambda_1, \ldots, \Lambda_n)\cdot \partial_{\zeta_k^i}\ceq -a(\Lambda_1, \ldots, \Lambda_n)\theta_k^i \end{align*} for $a(\Lambda_1, \ldots, \Lambda_n)\in V_\nabla[\Lambda_k]_{k=1}^n$. Thus, by restriction, we have right $\mathcal{D}_n^\mathrm{T}$-supermodules $V^{\otimes n}\otimes \mathcal{O}_n^{\star\mathrm{T}}$ and $V_\nabla[\Lambda_k]_{k=1}^n$. Now we introduce the main object of this \cref{ss:W:VA}, which is an $N_W=N$ SUSY analogue of the operad $P^{\textup{ch}}$ in \cite[\S6.3]{BDHK}. \begin{dfn} For an $\mathcal{H}_W$-supermodule $V=(V,\nabla)$ and $n \in \mathbb{N}$, we denote \begin{align*} P^{\textup{ch}N_W}_V(n) \ceq \Hom_{\mathcal{D}_n^\mathrm{T}} \bigl(V^{\otimes n} \otimes \mathcal{O}_n^{\star \mathrm{T}},V_{\nabla}[\Lambda_k]_{k=1}^n\bigr). \end{align*} To stress the variables $\Lambda_1, \ldots, \Lambda_n$, we express an element $X \in P^{\textup{ch}N_W}_V(n)$ as $X_{\Lambda_1, \ldots, \Lambda_n}$. \end{dfn} In other words, $P^{\textup{ch}N_W}_V(n)$ is a linear sub-superspace of $\Hom_\mathbb{K}\bigl(V^{\otimes n} \otimes \mathcal{O}_n^{\star \mathrm{T}},V_{\nabla}[\Lambda_k]_{k=1}^n\bigr)$ spanned by elements $X$ such that \begin{align*} X((v \otimes f) \cdot \varphi) = X(v \otimes f) \cdot \varphi \quad (v \in V^{\otimes n}, \, f \in \mathcal{O}_n^{\star\mathrm{T}}, \, \varphi \in \mathcal{D}_n^{\mathrm{T}}). \end{align*} For $\sigma \in \mathfrak{S}_n$ and $X \in P^{\textup{ch}N_W}_V(n)$, we define a linear map $X^\sigma\colon V^{\otimes n} \otimes \mathcal{O}_n^{\star \mathrm{T}} \to V_{\nabla}[\Lambda_k]_{k=1}^n$ by \begin{align*} X^\sigma(v_1\otimes \cdots \otimes v_n \otimes f) \ceq X_{\sigma(\Lambda_1,\ldots,\Lambda_n)}(\sigma(v_1\otimes \cdots \otimes v_N)\otimes \sigma f) \end{align*} for $v_1, \ldots, v_n\in V$ and $f \in \mathcal{O}_n^{\star \mathrm{T}}$, where $\sigma(\Lambda_1, \ldots, \Lambda_n) \ceq (\Lambda_{\sigma^{-1}(1)},\ldots,\Lambda_{\sigma^{-1}(n)})$ as before and \begin{align*} (\sigma f)(Z_{k, l}) \ceq f(Z_{\sigma^{-1}(k), \sigma^{-1}(l)}). \end{align*} Then, one can easily verify $X^\sigma \in P^{\textup{ch}N_W}_V(n)$. The linear superspace $P^{\textup{ch}N_W}_V(n)$ carries a structure of a right $\mathfrak{S}_n$-supermodule by this action. For $X \in P^{\textup{ch}N_W}_V(k)$ and $Y_i\in P^{\textup{ch}N_W}_V(n_i)$ with $i \in [k]$, we define a linear map \begin{align*} X(Y_1\otimes \cdots\otimes Y_k)\colon V^{\otimes (n_1+\cdots+n_k)} \otimes \mathcal{O}_{n_1+\cdots+n_k}^{\star \mathrm{T}} \longrightarrow V_{\nabla}[\Lambda_l]_{l=1}^{n_1+\cdots+n_k} \end{align*} by letting \begin{align*} &X(Y_1\otimes \cdots \otimes Y_k)(v_1\otimes \cdots\otimes v_{n_1+\cdots+n_k}\otimes f) \ceq \\ &X_{\Lambda_1', \ldots, \Lambda_k'}\Bigl( \bigotimes_{i=1}^k (Y_i)_{\Lambda_{N_{i-1}+1}-\partial_{Z_{N_{i-1}+1}}, \ldots, \Lambda_{N_i}-\partial_{Z_{N_i}}} \Bigl(\bigotimes_{i=1}^k(v_{N_{i-1}+1}\otimes \cdots\otimes v_{N_i}\otimes f_i)\Bigr) \otimes f_0|_{Z_j=Z_{N_i}(N_{i-1}+1\le j\le N_i)} \Bigr) \end{align*} for $v_1,\dotsc,v_{n_1+\cdots+n_k} \in V$ and $f \in \mathcal{O}_{n_1+\cdots+n_k}^{\star \mathrm{T}}$. Here we used $N_i\ceq n_1+\cdots+n_i$, $\Lambda_i'\ceq \Lambda_{N_{i-1}+1}+\cdots+\Lambda_{N_i}$, and $f$ is written in the form \begin{align*} f(Z_1, \ldots, Z_{N_k})=f_0(Z_1, \ldots, Z_{N_k})\prod_{i=1}^kf_i(Z_{N_{i-1}+1}, \ldots Z_{N_i}) \end{align*} so that $f_0$ has no poles at $z_i=z_j$ for $N_{i-1}+1\le i<j\le N_i$. The symbol $f_0|_{Z_j=Z_{N_i}(N_{i-1}+1\le j\le N_i)}$ in the right hand side means that we apply the derivatives $\partial_{Z_i}$ to $f_0$ before setting $Z_j=Z_{N_i}$. \begin{lem} The $\mathfrak{S}$-supermodule \[ P^{\textup{ch}N_W}_V \ceq \bigl(P^{\textup{ch}N_W}_V(n)\bigr)_{n \in \mathbb{N}} \] is an operad by letting $X \otimes Y_1 \otimes \cdots \otimes Y_k \mapsto X(Y_1\otimes \cdots \otimes Y_k)$ be the composition map and $\id_V$ be the identity. \end{lem} \begin{proof} The proof of non-SUSY case in \cite [Proposition 6.7]{BDHK} works with minor modification. We omit the detail. \end{proof} Summarizing the discussion so far, we name the obtained operad as follows. Recall the commutative superalgebra $\mathcal{H}_W$ in \cref{dfn:W:clHW}. \begin{dfn}\label{dfn:W:Pch} For an $\mathcal{H}_W$-supermodule $V=(V,\nabla)$, we call the operad $P^{\textup{ch}N_W}_V$ the \emph{$N_W=N$ SUSY chiral operad}. \end{dfn} Recall the symbol $\Pi^{k}V = (\Pi^{k}V,\nabla)$, which is explained right after \cref{dfn:W:Chom}. Also recall the set $\MC\bigl(L(\oQ)\bigr)$ of Maurer-Cartan solutions $X \square X = 0$, $X \in L(\oQ)^1$ in the graded Lie superalgebra $L(\oQ)$ associated to an operad $\oQ$, explained in \cref{ss:1:LQ}. The following is the main statement of this \cref{ss:W:VA}, which is a natural $N_W=N$ SUSY analogue of \cite[Theorem 6.12]{BDHK}. \begin{thm}\label{thm:W:VA} Let $(V,\nabla)=(V,T,S^1,\dotsc,S^N)$ be an $\mathcal{H}_W$-supermodule. \begin{enumerate} \item For an odd element $X \in L^1\bigl(P_{\Pi^{N+1}V}^{\textup{ch}N_W}\bigr)$ satisfying $X\square X=0$, define linear maps $[\cdot_\Lambda\cdot]_X\colon V \otimes V \to V[\Lambda]$ and $\mu_X\colon V\otimes V \to V$ by \begin{align} \label{eq:VAlbra} [a_\Lambda b]_X &\ceq (-1)^{p(a)(\overline{N}+\overline{1})}X_{\Lambda, -\Lambda-\nabla}(a\otimes b\otimes 1_\mathbb{K}), \\ \label{eq:VAprod} \mu_X(a\otimes b) &\ceq (-1)^{p(a)(\overline{N}+\overline{1})+\overline{1}} \Res_\Lambda (\lambda^{-1}X_{\Lambda, -\Lambda-\nabla}(a\otimes b\otimes z_{1, 2}^{-1})) \end{align} for any $a, b \in V$. Then $(V,\nabla,[\cdot_\Lambda \cdot]_X,\mu_X)$ is a non-unital $N_W=N$ SUSY vertex algebra. \item The map $X \mapsto ([\cdot_\Lambda \cdot]_X,\mu_X)$ gives a bijection \[ \MC\bigl(L\bigl(\oP_{\Pi^{N+1}V}^{\textup{ch}N_W}\bigr)\bigr)_{\overline{1}} \xrr{\sim} \{\text{non-unital $N_W=N$ SUSY vertex algebra structures on $(V,\nabla)$}\}. \] \end{enumerate} \end{thm} In the rest of this section, we prove \cref{thm:W:VA}. We denote $\widetilde{V} \ceq \Pi^{N+1}V$ and $\oP^{\textup{ch}}_{\widetilde{V}} \ceq \oP^{\textup{ch}N_W}_{\Pi^{N+1}V}$ for simplicity, and let $p$, $\widetilde{p}$ be the parity of $V$, $\widetilde{V}$ respectively. Consider the linear maps $[\cdot_\Lambda\cdot]_X\colon V\otimes V \to V[\Lambda]$ and $\mu_X\colon V \otimes V \to V$ defined by \cref{eq:VAlbra} and \cref{eq:VAprod} for an odd element $X \in \oP_{\widetilde{V}}^{\textup{ch}}(2)$. The linear map $[\cdot_\Lambda\cdot]_X\colon V\otimes V \to V[\Lambda]$ has the parity $\overline{N}$: In fact, since $X$ is an odd linear map $X\colon \widetilde{V}^{\otimes 2} \otimes \mathcal{O}^{\star \mathrm{T}}_2 \to \widetilde{V}_\nabla[\Lambda_k]_{k=1, 2}$, \begin{align*} p(X(a\otimes b\otimes 1_\mathbb{K}))=\widetilde{p}(X(a\otimes b\otimes 1_\mathbb{K}))+\overline{N}+\overline{1} =\widetilde{p}(a)+\widetilde{p}(b)+\overline{N}=p(a)+p(b)+\overline{N} \end{align*} holds for $a, b\in V$. The linear map $\mu_X\colon V \otimes V \to V$ is even: This is clear since the residue map $\Res_\Lambda$ has the parity $\overline{N}$. The linear map $[\cdot_\Lambda\cdot]_X$ satisfies (1) in \cref{dfn:W:LCA}: This can be checked by a direct calculation similar to the proof of \cref{thm:W:LCA}. The linear map $\mu_X$ satisfies (1) in \cref{dfn:NWVA}: Since \begin{align*} &X_{\Lambda, -\Lambda-\nabla}(Ta\otimes b\otimes z_{1, 2}^{-1}) =-\lambda X_{\Lambda, -\Lambda-\nabla}(a\otimes b\otimes z_{1, 2}^{-1}) -X_{\Lambda, -\Lambda-\nabla}(a\otimes b\otimes z_{1, 2}^{-2}), \\ &X_{\Lambda, -\Lambda-\nabla}(a\otimes Tb\otimes z_{1, 2}^{-1}) =(\lambda+T)X_{\Lambda, -\Lambda-\nabla}(a\otimes b\otimes z_{1, 2}^{-1}) +X_{\Lambda, -\Lambda-\nabla}(a\otimes b\otimes z_{1, 2}^{-2}) \end{align*} hold for $a, b\in V$, we have $T(ab)=(Ta)b+a(Tb)$. Similarly, we can prove $S^i(ab)=(S^ia)b+a(S^ib)$. Note now that there exists the integral of $[\cdot_\Lambda\cdot]_X$. By \cref{lem:intbra}, we have \begin{align*} \int^\Lambda d\Gamma[a_\Gamma b]_X =(-1)^{p(a)(\overline{N}+\overline{1})+\overline{1}}X_{\Lambda, -\Lambda-\nabla}(a\otimes b\otimes z_{1, 2}^{-1}) \quad (a, b\in V). \end{align*} Thus we find that the map $X \mapsto ([\cdot_\Lambda\cdot]_X, \mu_X)$ gives a bijective correspondence between the set of all odd elements $X\in \oP_{\widetilde{V}}^{\textup{ch}}(2)$ and the set of all pairs $([\cdot_\Lambda\cdot], \mu)$ of a linear map $[\cdot_\Lambda\cdot]\colon V\otimes V \to V[\Lambda]$ of parity $\overline{N}$ satisfying (i) in \cref{dfn:W:LCA} and an even linear map $\mu\colon V \otimes V \to V$ satisfying (i) in \cref{dfn:NWVA}. The skew-symmetry \eqref{eq:LCA:ssym} and the quasi-commutativity $\eqref{eq:VA:qcom}$ are equivalent to $X^\sigma=X$ for any $\sigma\in\mathfrak{S}_2$: The element $X$ satisfies $X^\sigma=X$ ($\sigma\in\mathfrak{S}_2$) if and only if \begin{align*} (-1)^{\widetilde{p}(a)\widetilde{p}(b)}X_{\Lambda_2, \Lambda_1}(b\otimes a\otimes z_{2, 1}^{-1}) =X_{\Lambda_1, \Lambda_2}(a\otimes b\otimes z_{1, 2}^{-1}). \end{align*} Since we have, for $a, b\in V$, \begin{align*} \int^\Lambda d\Gamma[b_\Gamma a] &= (-1)^{p(b)(\overline{N}+\overline{1})}X_{\Lambda, -\Lambda-\nabla}(b\otimes a\otimes z_{1, 2}^{-1}), \\ \int^{-\Lambda-\nabla} d\Gamma[a_\Gamma b] &= (-1)^{p(a)(\overline{N}+\overline{1})}X_{-\Lambda-\nabla, \Lambda}(a\otimes b\otimes z_{1, 2}^{-1}), \end{align*} by \cref{prp:W:skecom}, the skew-symmetry and the quasi-commutativity are equivalent to $X^\sigma=X$ ($\sigma\in\mathfrak{S}_2$). In what follows, let $X \in L^1\bigl(\oP^\textup{ch}_{\widetilde{V}}\bigr)_{\overline{1}}$, so that $[\cdot_\Lambda\cdot]_X\colon V \otimes V \to V[\Lambda]$ and $\mu_X\colon V \otimes V \to V$ are linear maps satisfying (i), (ii) in \cref{dfn:W:LCA} and (i), (ii) in \cref{dfn:NWVA}. The Jacobi identity \eqref{eq:LCA:Jac}, the quasi-associativity \eqref{eq:VA:qass} and the Wick formula \eqref{eq:VA:Wick} are equivalent to $X\square X=0$: First, we have the following lemmas. \begin{lem}\label{lem:W:circ} For an odd element $X\in P_{\Pi^{N+1}V}^{\textup{ch}N_W}(2)$, we have \begin{align*} &(X\circ_1X)(a\otimes b\otimes c\otimes z_{1, 2}^{-1}z_{1, 3}^{-1}z_{2, 3}^{-1}) =(-1)^{p(b)(\overline{N}+\overline{1})}\int^{\Lambda_1+\Lambda_2}d\Gamma\Bigl[\Bigl(\int^{\lambda_1}_{-\lambda_2-T}d\lambda_1\int^{\Lambda_1}d\Gamma_1[a_{\Gamma_1}b]\Bigr)_\Gamma c\Bigr] \\ &(X\circ_2X)(a\otimes b\otimes c\otimes z_{1, 2}^{-1}z_{1, 3}^{-1}z_{2, 3}^{-1}) =(-1)^{(p(a)+\overline{N})\overline{N}+p(b)(\overline{N}+1)+1} \\ & \hspace{190pt} \times \int^{\Lambda_1}d\Gamma_1\Bigl[\Bigl(\int^{\lambda_1+\lambda_2+T}_{\lambda_2}d\lambda_2a\Bigr)_{\Gamma_1}\Bigl(\int^{\Lambda_2}d\Gamma_2[b_{\Gamma_2}c]\Bigr)\Bigr] \\ &(X\circ_2X)^{(1, 2)}(a\otimes b\otimes c\otimes z_{1, 2}^{-1}z_{1, 3}^{-1}z_{2, 3}^{-1}) =(-1)^{p(a)p(b)+p(b)+1}\\ & \hspace{190pt} \times\int^{\Lambda_2}d\Gamma_2\Bigl[\Bigl(\int_{\lambda_1}^{\lambda_1+\lambda_2+T}d\lambda_1b\Bigr)_{\Gamma_2}\Bigl(\int^{\Lambda_1}d\Gamma_1[a_{\Gamma_1}c]\Bigr)\Bigr] \end{align*} \end{lem} \begin{proof} First, we get \begin{align*} (X\circ_1X)(a\otimes b\otimes c\otimes z_{1, 2}^{-1}z_{1, 3}^{-1}z_{2, 3}^{-1}) &=X_{\Lambda_1+\Lambda_2, \Lambda_3}(X_{\Lambda-\partial_{Z_1}, \Lambda_2-\partial_{Z_2}}(a\otimes b\otimes z_{1, 2}^{-1})\otimes c\otimes z_{1, 3}^{-1}z_{2, 3}^{-1}|_{z_1=z_2})\\ &=(-1)^{p(a)(\overline{N}+\overline{1})+1}X_{\Lambda_1+\Lambda_2, \Lambda_3}\Bigl(\int^{\Lambda_1-\partial_{Z_1}}d\Gamma_1[a_{\Gamma_1}b]\otimes c\otimes z_{1, 3}^{-1}z_{2, 3}^{-1}|_{z_1=z_2}\Bigr). \end{align*} For $F(\Lambda) \in V[\Lambda]$, we have \begin{align*} F(\Lambda_1-\partial_{Z_1})\otimes c\otimes z_{1, 3}^{-1}z_{2, 3}^{-1} &=e^{-\partial_{Z_1}\partial_{\Lambda_1}}F(\Lambda_1)\otimes c\otimes z_{1, 3}^{-1}z_{2, 3}^{-1}\\ &=\sum_{n\ge 0}\frac{(-1)^n}{n!}\partial_{\lambda_1}^nF(\Lambda_1)\otimes c\otimes (\partial_{z_1}^nz_{1, 3}^{-1}z_{2, 3}^{-1})|_{z_1=z_2}\\ &=\sum_{n\ge 0}\frac{(-1)^n}{n!}\partial_{\lambda_1}^nF(\Lambda_1)\otimes c\otimes \frac{-1}{n+1}\partial_{z_1}^{n+1}z_{1, 3}^{-1}\\ &=-\int_0^{\partial_{z_1}}d\Lambda\,\theta^{[N]}e^{-\Lambda\cdot\partial_{\Lambda_1}}F(\Lambda_1)\otimes c\otimes z_{1, 3}^{-1}\\ &=-\int_0^{\partial_{z_1}}d\Lambda\, \theta^{[N]}F(\Lambda_1-\Lambda)\otimes c\otimes z_{1, 3}^{-1} \end{align*} Thus we obtain \begin{align*} &(X\circ_1X)(a\otimes b\otimes c\otimes z_{1, 2}^{-1}z_{1, 3}^{-1}z_{2, 3}^{-1})\\ &=(-1)^{p(a)(\overline{N}+\overline{1})}X_{\Lambda_1+\Lambda_2, \Lambda_3}\Bigl(\int_0^{\partial_{z_1}}d\Lambda\,\theta^{[N]}\int^{\Lambda_1-\Lambda}d\Gamma_1[a_{\Gamma_1}b]\otimes c\otimes z_{1, 3}^{-1}\Bigr)\\ &=(-1)^{p(a)(\overline{N}+\overline{1})}X_{\Lambda_1+\Lambda_2, \Lambda_3}\Bigl(\Bigl(\int_0^{\lambda_1+\lambda_2+T}d\Lambda\,\theta^{[N]}\int^{\Lambda_1-\Lambda}d\Gamma_1[a_{\Gamma_1}b]\Bigr)\otimes c\otimes z_{1, 3}^{-1}\Bigr)\\ &=(-1)^{p(a)(\overline{N}+\overline{1})+\overline{1}}X_{\Lambda_1+\Lambda_2, \Lambda_3}\Bigl(\Bigl(\int_{\lambda_1}^{-\lambda_2-T}d\lambda_1\int^{\Lambda_1}d\Gamma_1[a_{\Gamma_1}b]\Bigr)\otimes c\otimes z_{1, 3}^{-1}\Bigr)\\ &=(-1)^{p(b)(\overline{N}+\overline{1})}\int^{\Lambda_1+\Lambda_2}d\Gamma\Bigl[\Bigl(\int^{\lambda_1}_{-\lambda_2-T}d\lambda_1\int^{\Lambda_1}d\Gamma_1[a_{\Gamma_1}b]\Bigr)_\Gamma c\Bigr]. \end{align*} Similarly, we can prove the rest of two identities. \end{proof} \begin{lem}\label{lem:boxJqW} For an odd element $X\in L^1(P_{\Pi^{N+1}{V}}^{\textup{ch}N_W})$, we have \begin{align*} &(-1)^{(p(a)+\overline{N})\overline{N}+p(b)(\overline{N}+\overline{1})+\overline{1}}(X\square X)(a\otimes b\otimes c\otimes z_{1, 3}^{-1}z_{2, 3}^{-1})\\ &=\int^{\Lambda_1}d\Gamma_1\left[a_{\Gamma_1}\left(\int^{\Lambda_2}d\Gamma_2[b_{\Gamma_2}c]\right)\right] -(-1)^{(p(a)+\overline{N})(p(b)+\overline{N})}\int^{\Lambda_2}d\Gamma_2\left[b_{\Gamma_2}\left(\int^{\Lambda_1}d\Gamma_1[a_{\Gamma_1}c]\right)\right]\\ &\quad -(-1)^{(p(a)+N)N}\int^{\Lambda_1+\Lambda_2}d\Gamma\left[\left(\int^{\Lambda_1}d\Gamma_1[a_{\Gamma_1}b]-\int^{-\Lambda_2-\nabla}d\Gamma_2[a_{\Gamma_2}b]\right)_{\Gamma}c\right]. \end{align*} \end{lem} \begin{proof} This is immediate by \cref{lem:W:circ}. \end{proof} By \cref{prp:W:Jqas} and \cref{lem:boxJqW}, it is clear that $X\square X=0$ implies the Jacobi identity , the quasi-associativity and the Wick formula. Conversely, assume the Jacobi identity, the quasi-associativity and the Wick formula. To prove $X\square X=0$, it is sufficient to show that \begin{align}\label{eq:W:ResFn} \Res_{\Lambda_1}\Res_{\Lambda_2}\bigl(\lambda_1^{-1} \lambda_2^{-1}\partial_{\lambda_1}^n(X\square X)(a\otimes b\otimes c\otimes z_{1, 2}^{-1}z_{1, 3}^{-1}z_{2, 3}^{-1})\bigr)=0 \end{align} for all $n \in \mathbb{N}$ and $a,b,c \in V$, because we have \begin{align*} &(X\square X)(S^ia\otimes b\otimes c\otimes z_{1, 2}^{-1}z_{1, 3}^{-1}z_{2, 3}^{-1}) =-\theta_1^i(X\square X)(a\otimes b\otimes c\otimes z_{1, 2}^{-1}z_{1, 3}^{-1}z_{2, 3}^{-1}), \\ &(X\square X)(a\otimes S^ib\otimes c\otimes z_{1, 2}^{-1}z_{1, 3}^{-1}z_{2, 3}^{-1}) =-(-1)^{\widetilde{p}(a)}\theta_2^i(X\square X)(a\otimes b\otimes c\otimes z_{1, 2}^{-1}z_{1, 3}^{-1}z_{2, 3}^{-1}) \end{align*} and \begin{align*} (\partial_{\lambda_2}-\partial_{\lambda_1})(X\square X)(a\otimes b\otimes c\otimes z_{1, 2}^{-1}z_{1, 3}^{-1}z_{2, 3}^{-1}) =(X\square X)(a\otimes b\otimes c\otimes z_{1, 3}^{-1}z_{2, 3}^{-1})=0. \end{align*} For $n\in\mathbb{N}$, let us denote \begin{align}\label{eq:W:Fn} F_n(a\otimes b\otimes c) \ceq (-1)^{p(b)(\overline{N}+\overline{1})}\partial_{\lambda_1}^n\Res_{\Lambda_2}(\lambda_2^{-1}(X\square X)(a\otimes b\otimes c\otimes z_{1, 2}^{-1}z_{1, 3}^{-1}z_{2, 3}^{-1})). \end{align} By \cref{lem:W:circ}, we get \begin{align}\label{eq:W:F0} \begin{split} F_0(a\otimes b\otimes c) &=\Bigl(\int_{-T}^{\lambda_1}d\lambda_1\int^{\Lambda_1}d\Gamma_1[a_{\Gamma_1}b]\Bigr)c +\int_0^{\lambda_1}d\Lambda\Bigl[ \Bigl(\int_{-T}^{\lambda_1}\int^{\Lambda_1}d\Gamma_1[a_{\Gamma_1}b]\Bigr)_\Lambda c\Bigr]\\ &\quad -\int^{\Lambda_1}d\Gamma_1[(\lambda_1+T)a_{\Gamma_1}bc] -\int^{\Lambda_1}d\Gamma_1\Bigl[\Bigl(\int_0^{\lambda_1+T}d\Lambda a\Bigr)_{\Gamma_1}\Bigl(\int_0^{\lambda_2}d\lambda[b_\Lambda c]\Bigr)\Bigr]\\ &\quad -(-1)^{(p(a)+\overline{N})p(b)}\Bigl(\int_{\lambda_1}^{\lambda_1+T}d\lambda_1b\Bigr)\int^{\Lambda_1}d\Gamma_1[a_{\Gamma_1}c]. \end{split} \end{align} \begin{lem}\label{lem:ResF01} For $n=0,1$ and $a,b,c \in V$, we have \begin{align*} \Res_{\Lambda_1}\Res_{\Lambda_2}\bigl(\lambda_1^{-1} \lambda_2^{-1}\partial_{\lambda_1}^n(X\square X)(a\otimes b\otimes c\otimes z_{1, 2}^{-1}z_{1, 3}^{-1}z_{2, 3}^{-1})\bigr)=0 \end{align*} \end{lem} \begin{proof} Since we get \begin{align*} &\Res_{\Lambda_1}(\lambda_1^{-1}F_0(a\otimes b\otimes c))\\ &=\Bigl(((Ta)b)c-(Ta)(bc)-\Bigl(\int_0^Td\Lambda Ta\Bigr)[b_\Lambda c]-(-1)^{p(a)p(b)}\Bigl(\int_0^Td\Lambda b\Bigr)[Ta_\Lambda c]\Bigr) \\ &\quad +(-1)^{p(a)p(b)}\Bigl(((Tb)a)c-(Tb)(ac)-\Bigl(\int_0^Td\Lambda Tb\Bigr)[a_\Lambda c]-(-1)^{p(a)p(b)}\Bigl(\int_0^Td\Lambda a\Bigr)[Tb_\Lambda c]\Bigr), \end{align*} the equation \eqref{eq:W:ResFn} for $n=0$ follows from the quasi-associativity. Also, by \eqref{eq:W:F0} \begin{align*} F_1(a\otimes b\otimes c) &=\Bigl(\int^{\Lambda_1}d\Gamma_1[a_{\Gamma_1}b]\Bigr)c +\int_0^{\lambda_1}d\Lambda\Bigl[\Bigl(\int^{\Lambda_1}d\Gamma_1[a_{\Gamma_1}b]\Bigr)_\Lambda c\Bigr]\\ &\quad -\int^{\Lambda_1}d\Gamma_1[a_{\Gamma_1}bc] -\int^{\Lambda_1}d\Gamma_1\Bigl[\Big(\int_0^{\lambda_1+T}d\Lambda a\Bigr)_{\Gamma_1}[b_{\Lambda}c]\Bigr]\\ &\quad -(-1)^{(p(a)+\overline{N})p(b)}\Bigl(\int_{\lambda_1}^{\lambda_1+T}d\lambda_1 b\Bigr)[a_{\Lambda_1}c], \end{align*} thus we get \begin{align*} \Res_{\Lambda_1}(\lambda_1^{-1}F_1(a\otimes b\otimes c)) =(ab)c-a(bc)-\Bigl(\int_0^Td\Lambda a\Bigr)[b_\Lambda c]-(-1)^{p(a)p(b)}\Bigl(\int_0^Td\Lambda b\Bigr)[a_\Lambda c]. \end{align*} Hence the equation \eqref{eq:W:ResFn} for $n=1$ follows from the quasi-associativity. \end{proof} To show \eqref{eq:W:ResFn} for $n \ge 2$, we expand $[\cdot_\Lambda \cdot]$ in the form \begin{align*} [a_\Lambda b]=\sum_{m \in \mathbb{N}, \, I \subset [N]} \Lambda^{m|I} a_{m|I}b \quad (a, b\in V). \end{align*} We also denote $a_{m|N}b \ceq a_{m|[N]}b$. Using induction on $n$, we get for any $n\in\mathbb{N}$ that \begin{align* \begin{split} F_{n+2}(a\otimes b\otimes c) &=\sum_{m\ge n}n!\binom{m}{n}\lambda_1^{m-n} \Bigl(\int^{\Lambda_1}d\Gamma_1[a_{\Gamma_1}b]\Bigr)_{m|N}c +\sum_{k=0}^{n-1}k!\binom{m}{k}\binom{n}{k} \lambda_1^{m-k}\partial_{\lambda_1}^{n-k-1}[a_{\Lambda_1}b]_{m|N}c\\ &\quad -\sum_{m\ge n}n!\binom{m}{n}\int^{\Lambda_1}d\Gamma_1[(\lambda_1+T)^{m-n}a_{\Gamma_1}(b_{m|N}c)] -\sum_{k=0}^{n-1}k!\partial_{\lambda_1}^{n-k-1}[a_{\Lambda_1}(b_{k|N}c)] \\ &\quad -(-1)^{(p(a)+\overline{N})p(b)}\sum_{m \ge n, \, I\subset [N]} (-1)^{p(b)\overline{\# I}}n!\binom{m}{n}((\lambda_1+T)^{m-n}\theta_1^Ib)(a_{m|I}c). \end{split} \end{align*} \begin{lem}\label{lem:W:ResFn} For $n \ge 2$ and $a,b,c \in V$, we have \begin{align*} \Res_{\Lambda_1}\Res_{\Lambda_2}\bigl(\lambda_1^{-1} \lambda_2^{-1}\partial_{\lambda_1}^n(X\square X)(a\otimes b\otimes c\otimes z_{1, 2}^{-1}z_{1, 3}^{-1}z_{2, 3}^{-1})\bigr)=0 \end{align*} \end{lem} \begin{proof} Let us use the notation $F_n$ in \eqref{eq:W:Fn}. First, we have \begin{align*} &\Res_{\Lambda_1}(\lambda_1^{-1}F_{n+2}(a\otimes b\otimes c))\\ &=n!\Res_{\Lambda}\bigl(\lambda^{-n-1} ([ab_{\Lambda}c] -(-1)^{p(a)\overline{N}}(e^{\nabla\cdot\partial_{\Lambda}}a)[b_\Lambda c] -(-1)^{(p(a)+\overline{N})p(b)}(e^{\nabla\cdot\partial_{\Lambda}}b)[a_\Lambda c])\bigr) \\ &\quad -\sum_{k=0}^{n-1}k!(n-k-1)!\Bigl(a_{n-k-1|N}(b_{k|N}c)-\binom{n}{k}(a_{n-k-1|N}b)_{k|N}c\Bigr). \end{align*} Since the Jacobi identity yields \begin{align*} [a_{\Lambda_1-\Lambda_2}[b_{\Lambda_2}c]] =(-1)^{(p(a)+\overline{N})\overline{N}}[[a_{\Lambda_1-\Lambda_2}b]_{\Lambda_1}c]+(-1)^{(p(a)+\overline{N})(p(b)+\overline{N})}[b_{\Lambda_2}[a_{\Lambda_1-\Lambda_2}c]], \end{align*} we obtain \begin{align*} & (-1)^{(p(a)+\overline{N})p(b)} \Res_{\Lambda_1} \Bigl(\lambda_1^{-n-1} \int_0^{\lambda_1}d\Lambda_2[b_{\Lambda_2}[a_{\Lambda_1-\Lambda_2}c]\Bigr) \\ &=(-1)^{(p(a)+\overline{N})\overline{N}} \Res_{\Lambda_1} \Bigl(\lambda_1^{-n-1} \int_0^{\lambda_1}d\Lambda_2[a_{\Lambda_1-\Lambda_2}[b_{\Lambda_2}c]]\Bigr) -\Res_{\Lambda_1} \Bigl(\lambda_1^{-n-1} \int_0^{\lambda_1}d\Lambda_2[[a_{\Lambda_1-\Lambda_2}b]_{\Lambda_1}c]\Bigr) \\ &=\sum_{k=0}^{n-1} \frac{k!(n-k-1)!}{n!} \Bigl(a_{n-k-1|N}(b_{k|N}c)-\binom{n}{k}(a_{n-k-1|N}b)_{k|N}c\Bigr). \end{align*} Thus \begin{align*} \Res_{\Lambda_1}\bigl(\lambda_1^{-1}F_{n+2}(a\otimes b\otimes c)\bigr) =n!\Res_{\Lambda}\Bigl(\lambda^{-n-1} \bigl([ab_{\Lambda}c] &-(-1)^{p(a)\overline{N}}(e^{\nabla\cdot\partial_{\Lambda}}a)[b_\Lambda c] \\ -(-1)^{(p(a)+\overline{N})p(b)}(e^{\nabla\cdot\partial_{\Lambda}}b)[a_\Lambda c] &-(-1)^{(p(a)+\overline{N})p(b)}\int_0^{\lambda}d\Gamma[b_{\Gamma}[a_{\Lambda-\Gamma}c]\bigr)\Bigr). \end{align*} Hence the equation \eqref{eq:W:ResFn} for $n\ge 2$ follows from the right Wick formula. \end{proof} As was mentioned around \eqref{eq:W:ResFn}, $X\square X=0$ follows from \cref{lem:ResF01} and \cref{lem:W:ResFn}. The proof of \cref{thm:W:VA} is now complete. \subsection{Relation to $N_W=N$ SUSY chiral algebras}\label{ss:W:CA} In \cite[Appendix A]{BDHK}, the authors discussed the relation between their operadic theory of vertex algebras and the theory of chiral algebra \cite{BD}, which is an algebro-geometric reformulation of vertex algebras. Here we give a $N_W=N$ SUSY analogue of the argument in \cite[Appendix A]{BDHK}. A SUSY analogue of chiral algebra was introduced by Heluani in \cite[\S4]{H}, which will be used in our discussion. We continue to work over a field $\mathbb{K}$ of characteristic $0$, and take a positive integer $N$. We will also extensively use the theory of superscheme. See \cite{DM}, \cite[\S2.2]{H} and \cite[\S1]{KV} for the detail. An ordinary scheme is called an even scheme for distinction. \subsubsection{$N_W=N$ SUSY chiral algebras}\label{sss:W:CA} Let $X$ be an irreducible smooth projective $(1|N)$-dimensional supercurve over $\mathbb{K}$ in the sense of \cite[p.1085]{KV}. More explicitly, it is a pair $(X,\mathscr{O}_X)$ of a topological space $X$ and a sheaf $\mathscr{O}_X$ of commutative $\mathbb{K}$-superalgebras over $X$ such that \begin{clist} \item $(X,\mathscr{O}_X^{\textup{red}})$ is an irreducible smooth projective curve, where $\mathscr{O}_X^{\textup{red}}$ is the reduced part of $\mathscr{O}_X$. The sheaf $\mathscr{O}_X$ is called the structure sheaf of $X$. \item For every $x \in X(\mathbb{K})$, there is an open subsets $U \subset X$ and some linearly independent odd elements $\theta^i \in \mathscr{O}_X(U)$ ($i \in [N]$) such that $\mathscr{O}_X(U)=\mathscr{O}_X^{\textup{red}}(U) \otimes \mathbb{K}[\theta^1,\dotsc,\theta^N]$. Such $U$ is called a coordinate neighborhood of $x$. \end{clist} Hereafter we call such $X$ a smooth $(1|N)$-supercurve for simplicity. On each coordinate neighborhood $U \subset X$ of $x$, we can take a local coordinate $z$ of the even curve $(X,\mathscr{O}_X^{\textup{red}})$, and the tuple $Z=(z,\zeta^1,\dotsc,\zeta^N)$ satisfies the relation \eqref{eq:W:poly}. We call $Z$ a local coordinate of $U$. On a smooth $(1|N)$-supercurve $X$, we have the sheaf $\mathscr{D}_X$ of differential operators, whose sections $\mathscr{D}_X(U)$ on a coordinate neighborhood $U \subset X$ is the superalgebra \[ \mathcal{D} \ceq \mathbb{K}[Z][\partial_Z] = \mathbb{K}[z,\zeta^1,\dotsc,\zeta^n][\partial_z,\partial_{\zeta^1},\dotsc,\partial_{\zeta^N}] \] of regular differential operators generated by $\partial_z \ceq \frac{\partial}{\partial z}$ and $\partial_{\zeta^k} \ceq \frac{\partial}{\partial \zeta^k}$ ($i \in [N]$). We denote by $\crMod \mathscr{D}_{X}$ the category of quasi-coherent sheaves of right $\mathscr{D}_{X}$-modules on $X$. For each $n \in \mathbb{Z}_{>0}$, we denote by $\Delta\colon X \hookrightarrow X^n$ the embedding of the big diagonal in $X^n$, i.e., the union of the hypersurfaces \begin{align}\label{eq:W:bigdiag} Z_i = Z_j \end{align} for $i \ne j \in [n]$, using the local coordinates $Z_i=(z_i,\zeta_i^1,\dotsc,\zeta_i^N)$. Hereafter we denote $\Delta \ceq \Delta(X) \subset X^n$ for simplicity. As commented in the beginning of \cite[\S4]{H}, the big diagonal has relative codimension $1|N$, so that it is not a divisor (of codimension $1|0$) in genuine sense, and there is an ambiguity of the pushforward functor $\Delta_*$. We follow \cite[\S4.1.4]{H} to resolve this problem: Let us consider the embedding \[ j\colon X^n \mathbin{\setminus} \Delta \lhook\joinrel\longrightarrow X^n, \] where $X^n \mathbin{\setminus} \Delta$ is the superscheme with underlying topological space $U \ceq \abs{X}^n \mathbin{\setminus} \abs{\Delta}$ and structure sheaf $\rst{\mathscr{O}_{X^n}}{U}$. Then we define the pushforward $\Delta_* \mathscr{A}$ of a right $\mathscr{D}_X$-module $\mathscr{A}$ to be $\Delta_* \mathscr{A} \ceq j_* j^*(\mathscr{O}_{X^{n-1}} \boxtimes \mathscr{A})/(\mathscr{O}_{X^{n-1}} \boxtimes \mathscr{A})$. Now, for a right $\mathscr{D}_X$-module $\mathscr{A}$, we set \[ \oP^{\textup{ch}N_W}_{\mathscr{A}}(n) \ceq \Hom_{\crMod \mathscr{D}_{X^n}}\bigl(j_* j^* \mathscr{A}^{\boxtimes n},\Delta_*\mathscr{A}\bigr). \] By the argument in the even case \cite{BD}, $\oP^{\textup{ch}N_W}_{\mathscr{A}}=\bigl(\oP^{\textup{ch}N_W}_{\mathscr{A}}(n)\bigr)_{n \in \mathbb{N}}$ has a structure of an operad. The image of the generator of $\oLie(2)$ under $\varphi$ gives a binary operation $\mu \in \oP^{\textup{ch}N_W}_{\mathscr{A}}(2)$, which is a map $\mu_{\mathscr{A}}\colon j_* j^* \mathscr{A} \boxtimes \mathscr{A} \to \Delta_* \mathscr{A}$ satisfying skew-symmetry and Jacobi identity. We call $\mu_{\mathscr{A}}$ the chiral operation of $\mathscr{A}$, following the terminology in \cite[\S3.1]{BD}. A \emph{non-unital $N_W=N$ SUSY chiral algebra on $X$} is a pair $(\mathscr{A},\varphi)$ of a right $\mathscr{D}_X$-module $\mathscr{A}$ and an operad morphism $\varphi\colon \oLie \to \oP^{\textup{ch}N_W}_{\mathscr{A}}$. In the language of \cref{dfn:1:Lie}, such a morphism $\varphi$ is nothing but a Lie algebra structure on the operad $\oP^{\textup{ch}N_W}_{\mathscr{A}}$. Let us denote by $\omega_X$ the Berezinian bundle \cite[2.2.7]{H} of the smooth $1|X$-dimensional curve $X$, which can be regarded as a SUSY analogue of the canonical bundle of an even curve. It is a right $\mathscr{D}_X$-module in the standard way, and moreover carries a non-unital $N_W=N$ SUSY chiral algebra structure, established by Heluani in \cite[4.1.7]{H}. We denote by $\mu_{\omega} \in \oP^{\textup{ch}N_W}_{\omega_X}(2)$ the corresponding chiral operation. Now we define an \emph{$N_W=N$ SUSY chiral algebra $\mathscr{A}$ on $X$} to be a non-unital one equipped with a morphism $\iota\colon \omega_X \to \mathscr{A}$ of right $\mathscr{D}_X$-modules such that the composition $\mu_{\iota \otimes \id_{\mathscr{A}}}$ (the restriction of $\mu_{\mathscr{A}}$ to $j_* j^* \omega_X \boxtimes \mathscr{A} \to j_* j^* \mathscr{A} \boxtimes \mathscr{A}$) coincides with $j_* j^* \omega_X \boxtimes \mathscr{A} \to (j_* j^* \omega_X \boxtimes \mathscr{A})/\omega_X \boxtimes \mathscr{A} \xr{\sim} \Delta_*\mathscr{A}$ (the latter in the even case is called the unit operation \cite[3.1.12, 3.3.3]{BD}). \subsubsection{SUSY chiral algebras on the superline} We write down $N_W=N$ SUSY chiral algebras over the affine superline $\mathbb{A}^{1|N}$, following the argument for the even case \cite[\S A.2]{BDHK}. Let $Z=(z,\zeta^1,\dotsc,\zeta^N)$ be a $(1|N)_W$-supervariable and $\mathbb{K}[Z]$ be the corresponding polynomial superring in the sense of \cref{ss:W:poly}. In the setting of the previous \cref{sss:W:CA}, consider the case when the smooth $(1|N)$-supercurve $X$ is taken to be \[ X = \mathbb{A}^{1|N} \ceq \Spec(\mathbb{K}[Z]), \] i.e., the affine superspace of dimension $1|N$ in the sense of \cite[Example 1.1.6]{KV}. A right $\mathscr{D}_X$-module $\mathscr{A}$ is determined by the right module $A \ceq \Gamma(X,\mathscr{A})$ over the superalgebra $\mathcal{D}=\Gamma(X,\mathscr{D}_{X})$ of regular differential operators on the supervariable $Z$. Let $n \in \mathbb{Z}_{>0}$, and $Z_k=(z_k,\zeta_k^1,\dotsc,\zeta_k^N)$ be a $(1|N)_W$-supervariable for $k \in [n]$, regarded as the local coordinate of the $k$-th component of the $n$-th product supervariety $X^n = (\mathbb{A}^{1|N})^n$. Then we denote the superalgebra $\Gamma(X^n,\mathscr{O}_{X^n})$ of regular functions on $X^n$ by \[ \mathcal{O}_n \ceq \mathbb{K}[Z_1,\dotsc,Z_n]. \] Also recall the superalgebra $\mathcal{O}_n^{\star}=\mathbb{K}[Z_k]_{k=1}^n[z_{k,l}^{-1}]_{1 \le k<l \le n}$ in \cref{dfn:W:O}. Then, by the argument in \cite[\S A.2]{BDHK}, we have \begin{align}\label{eq:W:A2} \Gamma(X,j_* j^* \mathscr{A}^{\boxtimes n}) \cong \mathcal{O}_n^{\star} \otimes_{\mathcal{O}_n} A^{\otimes n}, \end{align} which is a right module over the superalgebra \[ \mathcal{D}_n \ceq \mathbb{K}[Z_1,\dotsc,Z_n][\partial_{Z_1},\dotsc,\partial_{Z_n}] \] of regular differential operators on $X^n$. The module structure is given by the action on the right factor in \eqref{eq:W:A2}. Let $I \subset \mathcal{D}_n$ be the left ideal generated by $z_1-z_k$ for $k \in \{2,\dotsc,n\}$. Then the quotient $I \mathbin{\setminus} \mathcal{D}_n$ is a $\mathcal{D}_1$-$\mathcal{D}_n$-bimodule, where the $\mathcal{D}_n$-action is the right multiplication, and the left action of $\mathcal{D}_1 = \mathbb{K}[Z][\partial_Z]$-action is given by letting $z,\zeta^i$ act from left by $z_1,\zeta_1^i$ and $\partial_z,\partial_{\zeta^i}$ act as multiplication on the left by $\sum_{k=1}^n \partial_{z_k}, \sum_{k=1}^n \partial_{\zeta^i_k}$. Then, by the argument in \cite[\S A.2]{BDHK}, we have \begin{align}\label{eq:W:A3} \Gamma(X^n,\Delta_*\mathscr{A}) \cong A \otimes_{\mathcal{D}_1} (I \mathbin{\setminus} \mathcal{D}_n), \end{align} where the right $\mathcal{D}_n$-module structure is given by right multiplication on the right factor. Combining \eqref{eq:W:A2} and \eqref{eq:W:A3}, we obtain an $\mathfrak{S}_n$-supermodule isomorphism \begin{align}\label{eq:W:A4} \mathcal{P}^{\textup{ch}N_W}_{\mathscr{A}}(n) \cong \Hom_{\crMod \mathcal{D}_n}\bigl( \mathcal{O}^{\star}_n \otimes_{\mathcal{O}_n} A^{\otimes n}, A \otimes_{\mathcal{D}_1}(I \mathbin{\setminus} \mathcal{D}_n)\bigr), \end{align} where $\crMod \mathcal{D}_n$ denotes the category of right $\mathcal{D}_n$-supermodules. \subsubsection{Equivariant SUSY chiral algebras on the superline} Next, we restrict the above description to the translation-equivariant part. We use the terminology on equivariant $\mathscr{D}$-modules in \cite[\S A.3]{BDHK}, which we extend to superschemes in an obvious way. In particular, we will use the notion of \emph{weakly equivariant $\mathcal{D}$-modules}. The affine superline $X=\mathbb{A}^{1|N}$ has a natural action of the group superscheme $\mathrm{Aff}^{1|N}$ of affine transformations. Consider the even subgroup scheme $\mathrm{T} \ceq \mathbb{G}_a \times \mathbb{G}_a^N \subset \mathrm{Aff}^{1|N}$ consisting of translations. We have an equivalence of categories between $\mathrm{T}$-equivariant quasi-coherent $\mathscr{O}_X$-modules and linear superspaces by taking the stalk at the origin $0 \in X$. The inverse functor sends a linear superspace $V$ to the sheaf associated to the $\mathbb{K}[Z]$-supermodule $V[Z]=\mathbb{K}[Z] \otimes_\mathbb{K} V$ equipped with the action of $(u,\upsilon^1,\dotsc,\upsilon^N) \in \mathrm{T}$ by $v(z,\zeta^1,\dotsc,\zeta^N) \mapsto v(z+u,\zeta^1+\upsilon^1,\dotsc,\zeta^N+\upsilon^N)$. As for $\mathrm{T}$-equivariant $\mathcal{D}$-modules on $X=\mathbb{A}^{1|N}$, following the even case \cite[\S A.4]{BDHK}, we have an equivalence of categories \begin{align}\label{eq:W:TDmod=Hmod} \begin{split} \bigl(\text{weakly $\mathrm{T}$-equivariant quasi-coherent $\mathscr{D}_X$-modules $\mathscr{A}$ on $X$}) \\ \xrr{\sim} \bigl(\text{$\mathcal{H}_W$-supermodules $(V,\nabla) = (V,T,S^1,\dotsc,S^N)$}\bigr). \end{split} \end{align} For the latter, see \cref{dfn:W:clHW}. Indeed, given a $\mathscr{D}_X$-module $\mathscr{A}$, taking the stalk of $\mathscr{A}$ at the origin $0 \in X$, we obtain $V$ as in the last paragraph, together with the endomorphisms $\nabla=(T,S^1,\dotsc,S^N)$ on $V$ induced by the differential operators $\partial_Z=(\partial_z,\partial_{\zeta^1},\dotsc,\partial_{\zeta^N})$ which satisfy the relations \eqref{eq:W:TSi}. Conversely, given a pair $(V,\nabla)$, we have a $\mathcal{D}_1$-module structure on $V[Z]=\mathbb{K}[z,\zeta^1,\dotsc,\zeta^i] \otimes_{\mathbb{K}} V$ by letting $Z$ act by multiplication by $Z$, and $\partial_z$ (resp.\ $\partial_{\zeta^i}$) act by $T-\frac{\partial}{\partial z}$ (resp.\ $S^i-\frac{\partial}{\partial \zeta^i}$). Now, following the argument in \cite[\S A.6]{BDHK}, let us consider a weakly $\mathrm{T}$-equivariant $\mathcal{D}$-module $\mathscr{A}$ on $X=\mathbb{A}^{1|N}$ corresponding to the $\mathcal{H}_W$-supermodule $(V,\nabla)$. Then the associated $\mathcal{D}_n$-module \eqref{eq:W:A3} is given by \[ V \otimes_{\mathbb{K}[\nabla]}\mathbb{K}[Z][\nabla_1,\dotsc,\nabla_n] \] with $Z=(z,\zeta^1,\dotsc,\zeta^N)$ and $\nabla_k=(T_k,S_k^1,\dotsc,S_k^N)$ satisfying the relations \eqref{eq:W:TSi} for each $k \in [n]$. The commutative superalgebra $\mathbb{K}[Z][\nabla_1,\dotsc,\nabla_n]$ is regarded as a $\mathbb{K}[\partial]$-$\mathcal{D}_n$-bimodule in the following way: \begin{itemize} \item The left action of $\nabla=(T,S^1,\dotsc,S^N)$ is given by $(\sum_{k=1}^n T_k,\sum_{k=1}^n S_k^1,\dotsc,\sum_{k=1}^n S_k^N)$. \item The right action of $\partial_{Z_k}=(\partial_{z_k},\partial_{\zeta_k^1},\dotsc,\partial_{\zeta_k^N})$ is the right multiplication by $\nabla_k$, and the right action of $Z_k=(z_k,\zeta_k^1,\dotsc,\zeta_k^N)$ on $f=f(Z,\nabla_1,\dotsc,\nabla_n) \in \mathbb{K}[Z][\nabla_1,\dotsc,\nabla_n]$ is given by \begin{align*} f \cdot z_k \ceq z f+\frac{\partial}{\partial T_k}f, \quad f \cdot \zeta_k^i \ceq (-1)^{p(f)}\bigl(\zeta^i f+\frac{\partial}{\partial S_k^i}f\bigr), \end{align*} where $p(f)$ denotes the parity of $f$. \end{itemize} Hence, $\oP^{\textup{ch}N_W}_{\mathscr{A}}(n)$ in \eqref{eq:W:A4} is equal to \begin{align}\label{eq:W:A8} \oP^{\textup{ch}N_W}_{\mathscr{A}}(n) = \Hom_{\crMod \mathcal{D}_n} \bigl(\mathcal{O}^{\star}_n \otimes V^{\otimes n}, V \otimes_{\mathbb{K}[\nabla]}\mathbb{K}[Z][\nabla_1,\dotsc,\nabla_n]\bigr). \end{align} The translation group $\mathrm{T} = \mathbb{G}_a \times \mathbb{G}_a^N$ acts on $\oP^{\textup{ch}N_W}_{\mathscr{A}}(n)$ by letting $U=(u,\upsilon^1,\dotsc,\upsilon^N) \in \mathrm{T}$ act on $\varphi \in \oP^{\textup{ch}N_W}_{\mathscr{A}}(n)$ to give $\varphi^U \in \oP^{\textup{ch}N_W}_{\mathscr{A}}(n)$ with \begin{align}\label{eq:W:A10} \varphi^U \bigl(f(Z_0,\dotsc,Z_n) \otimes v_1 \otimes \dotsb \otimes v_n) \ceq \rst{\varphi(f(Z_1-U,\dotsc,Z_n-U) \otimes v_1 \otimes \dotsb \otimes v_n)}{Z=Z+U} \end{align} for $f=f(Z_1,\dotsc,Z_n) \in \mathcal{O}^{\star}_n$ and $v_k \in V$, $k \in [n]$. Then the operad $\oP^{\textup{ch}N_W}_{\mathscr{A}}$ is $\mathrm{T}$-equivariant in the sense of \cite[\S A.5]{BDHK}, and the $\mathrm{T}$-invariants $\oP^{\textup{ch}N_W,\mathrm{T}}_{\mathscr{A}}(n) \subset \oP^{\textup{ch}N_W}_{\mathscr{A}}(n)$ form a sub-operad \[ \oP^{\textup{ch}N_W,\mathrm{T}}_{\mathscr{A}} \subset \oP^{\textup{ch}N_W}_{\mathscr{A}}. \] Now, we have the main statement of this \cref{ss:W:CA}. \begin{prp} Let $V=(V,\nabla)$ be an $\mathcal{H}_W$-supermodule (\cref{dfn:W:clHW}), and $\mathscr{A}$ be the corresponding weakly $\mathrm{T}$-equivariant $\mathcal{D}$-module on $\mathbb{A}^{1|N}$ under the equivalence \eqref{eq:W:TDmod=Hmod}. Let $\oP^{\textup{ch}N_W}_V$ be the $N_W=N$ SUSY chiral operad in \cref{dfn:W:Pch}. Then there is an isomorphism of operads \[ \oP^{\textup{ch}N_W,\mathrm{T}}_{\mathscr{A}} \cong \oP^{\textup{ch}N_W}_V. \] \end{prp} \begin{proof} The even case argument \cite[Lemma A.1]{BDHK} works, so we explain the outline only. By \eqref{eq:W:A8}, we want to determine the $\mathrm{T}$-invariant parts of $\mathcal{O}^{\star}_n \otimes V^{\otimes n}$ and $V \otimes_{\mathbb{K}[\nabla]}\mathbb{K}[Z][\nabla_1,\dotsc,\nabla_n]$. As for the former, by \eqref{eq:W:Ker-delta}, we have an isomorphism \[ (\mathcal{O}^{\star}_n \otimes V^{\otimes n})^{\mathrm{T}} = (\mathcal{O}^{\star}_n)^{\mathrm{T}} \otimes V^{\otimes n} = \mathcal{O}^{\star \mathrm{T}}_n \otimes V^{\otimes n} \] as $\mathcal{D}_n^{\mathrm{T}}$-supermodules, where \[ \mathcal{D}_n^{\mathrm{T}} \ceq \Ker(\ad \delta^0) \cap \Ker(\ad \delta^1) \cap \dotsb \cap \Ker(\ad \delta^N) \bigr) \subset \mathcal{D}_n \] with $\delta^i$ given after \eqref{eq:W:Ker-delta}. As for the latter $V \otimes_{\mathbb{K}[\nabla]}\mathbb{K}[Z][\nabla_1,\dotsc,\nabla_n]$, by the argument of \cref{lem:W:VnLn=VLn-1}, we have a $\mathcal{D}_n^{\mathrm{T}}$-supermodule isomorphism \begin{align*} (V \otimes_{\mathbb{K}[\nabla]}\mathbb{K}[Z][\nabla_1,\dotsc,\nabla_n])^{\mathrm{T}} = V \otimes_{\mathbb{K}[\nabla]} \mathbb{K}[\nabla_1,\dotsc,\nabla_n] &\xrr{\sim} V[\Lambda_k]_{k=1}^n/\Img(\nabla+\Lambda_1+\dotsb+\Lambda_n), \\ v \otimes f(\nabla_1,\dotsc,\nabla_n) &\longmapsto f(-\Lambda_1,\dotsc,-\Lambda_n)v. \end{align*} Now, any $n$-operation $\varphi \in \oP^{\textup{ch}N_W,\mathrm{T}}_{\mathscr{A}}(n)$ on $f \otimes v_1 \otimes \dotsb \otimes v_n \in \mathcal{O}_n^{\star} \otimes V^{\otimes n}$ is independent of $Z$ by \eqref{eq:W:A10}, so that it defines an element of $V[\Lambda_k]_{k=1}^n/\Img(\nabla+\Lambda_1+\dotsb+\Lambda_n)$. Thus we obtain an element of $\oP^{\textup{ch}N_W}_V(n)$ from $\varphi$. An inverse construction is given by the Taylor-expansion technique in the proof of \cite[Lemma A.1]{BDHK}. \end{proof} \section{\texorpdfstring{$N_K=N$}{NK=N} SUSY chiral operad}\label{s:K} In this section, we introduce $N_K=N$ SUSY analogue of the operad $\Chom$ and $P^{\textup{ch}}$ in \cite{BDHK}. Most results and proofs in \cref{s:W} for the $N_W=N$ case carry over to the $N_K=N$ case with minor modifications, so we omit several details. \subsection{Polynomial superalgebra}\label{ss:K:poly} Here we give a summary on the polynomial ring of supervariables in the $N_K=N$ setting. \begin{dfn} Let $A$ be a set and $\Lambda_\alpha=(\lambda_\alpha,\theta_\alpha^1,\ldots,\theta_\alpha^N)$ be a sequence of sets for each $\alpha\in A$. We denote by $\mathbb{K}[\Lambda_\alpha]_{\alpha\in A}$ the $\mathbb{K}$-superalgebra defined by even generators $\lambda_\alpha$ ($\alpha \in A$) and odd generators $\theta_\alpha^i$ ($\alpha \in A$, $i \in [N]$) with relations \begin{align*} \lambda_\alpha \lambda_\beta -\lambda_\beta \lambda_\alpha = 0, \quad \lambda_\alpha \theta_\beta^i-\theta_\beta^i \lambda_\alpha = 0, \quad \theta_\alpha^i \theta_\beta^j+\theta_\beta^j \theta_\alpha^i = -2\delta_{\alpha,\beta}\delta_{i,j} \lambda_\alpha \quad (\alpha,\beta \in A, \, i,j \in [N]). \end{align*} Each $\Lambda_\alpha$ for $\alpha \in A$ is called a \emph{$(1|N)_K$-supervariable}, and $\mathbb{K}$-superalgebra $\mathbb{K}[\Lambda_\alpha]_{\alpha \in A}$ is called the \emph{$N_K=N$ polynomial superalgebra of the supervariables $(\Lambda_\alpha)_{\alpha \in A}$}. \end{dfn} For a linear superspace $V$ and $(1|N)_K$-supervariable $\Lambda_\alpha$ ($\alpha\in A$) , we denote \begin{align*} V[\Lambda_\alpha]_{\alpha \in A} \ceq \mathbb{K}[\Lambda_\alpha]_{\alpha \in A} \otimes_\mathbb{K} V, \end{align*} which is a $\mathbb{K}[\Lambda_\alpha]_{\alpha\in A}$-supermodule. \begin{dfn} Let $\mathcal{H}_K$ be the $\mathbb{K}$-superalgebra defined by an even generator $T$ and odd generators $S^i$ $(i\in[N])$ with relations \begin{align*} TS^i-S^iT=0, \quad S^iS^j+S^jS^i=2\delta_{i, j}T\quad (i, j\in[N]). \end{align*} For simplicity, we set $\nabla\ceq(T, S^1, \ldots, S^N)$. \end{dfn} Note that $\mathcal{H}_K$ is isomorphic to $\mathbb{K}[\Lambda]$ as a superalgebra by the homomorphism defined by $T \mapsto -\lambda$, $S^i \mapsto -\theta^i$ ($i\in[N]$). \subsection{Operad of $N_K=N$ SUSY Lie conformal algebras}\label{ss:K:LCA} In this subsection, we introduce an $N_K=N$ SUSY analogue of the operad $\Chom$ in \cite{BDHK}. All results and proofs in $N_W=N$ case \cref{ss:W:LCA} are valid for $N_K=N$ setting with modifications below: \begin{itemize} \item Replace the superalgebra $\mathcal{H}_W$ by $\mathcal{H}_K$. \item Replace all $(1|N)_W$-supervariables by $(1|N)_K$-supervariables. \end{itemize} Therefore we only refer to the results omitting their proofs. \medskip Let us fix a $(1|N)_K$-supervariable $\Lambda=(\lambda, \theta^1, \ldots, \theta^N)$. \begin{dfn}[{\cite[Definition 4.10]{HK}}]\label{dfn:K:LCA} Let $V$ be a left $\mathcal{H}_K$-supermodule and $[\cdot_\Lambda\cdot]\colon V\otimes V\to V[\Lambda]$ be a linear map of parity $\overline{N}$. A pair $(V, [\cdot_\Lambda\cdot])$ is called an \emph{$N_K=N$ SUSY Lie conformal algebra} if it satisfies the conditions (i)--(iv) in \cref{dfn:W:LCA} replacing $(1|N)_W$-supervariables $\Lambda, \Lambda_1, \Lambda_2$ by $(1|N)_K$-supervariables. The linear map $[\cdot_\Lambda\cdot]$ is called the \emph{$\Lambda$-bracket} of the left $\mathcal{H}_K$-supermodule $V$. For simplicity, we say $V$ is an $N_K=N$ SUSY Lie conformal algebra. \end{dfn} Hereafter we fix a left $\mathcal{H}_K$-supermodule $V=(V,\nabla)$. The polynomial superalgebra $\mathbb{K}[\Lambda_k]_{k=1}^n$ of $(1|N)_K$-supervariables $(\Lambda_k)_{k=1}^n$ carries a structure of a right $\mathcal{H}_K$-supermodule by letting \begin{align*} f(\Lambda_1, \ldots, \Lambda_n)\cdot T \ceq f(\Lambda_1, \ldots, \Lambda_n)\Bigl(-\sum_{k=1}^n\lambda_k\Bigr), \quad f(\Lambda_1, \ldots, \Lambda_n)\cdot S^i \ceq f(\Lambda_1, \ldots, \Lambda_n)\Bigl(-\sum_{k=1}^n\theta_k\Bigr) \end{align*} for $f(\Lambda_1, \ldots, \Lambda_n) \in \mathbb{K}[\Lambda_k]_{k=1}^n$. Thus, we obtain a linear superspace \begin{align}\label{eq:LCA:HKn-mod} V_\nabla[\Lambda_k]_{k=1}^n \ceq \mathbb{K}[\Lambda_k]_{k=1}^n \otimes_{\mathcal{H}_K} V. \end{align} As in the $N_W=N$ case (\cref{lem:W:VnLn=VLn-1}), we have: \begin{lem}\label{lem:K:VnLn=VLn-1} The linear superspace $V_\nabla[\Lambda_k]_{k=1}^n$ is isomorphic to $V[\Lambda_k]_{k=1}^{n-1}$ by the linear map \begin{align*} f(\Lambda_1, \cdots, \Lambda_n)\otimes v \longmapsto f(\Lambda_1, \cdots, \Lambda_{n-1}, -\Lambda_1-\cdots-\Lambda_{n-1}-\nabla)v \quad (f\in\mathbb{K}[\Lambda]_{k=1}^n, v\in V). \end{align*} \end{lem} The linear space $V_\nabla[\Lambda_k]_{k=1}^n$ has a natural structure of a left $\mathcal{H}_K^{\otimes n}$-supermodule. Indeed, since any element of $\mathcal{H}_K$ is expressed uniquely as $P(\nabla)$ for some $P(\Lambda)\in\mathbb{K}[\Lambda]$, the space $\mathbb{K}[\Lambda_k]_{k=1}^n$ carries a structure of a left $\mathcal{H}_K^{\otimes n}$-supermodule by letting $P_1(\nabla) \otimes \dotsb \otimes P_n(\nabla) \in \mathcal{H}_K^{\otimes n}$ act on $f(\Lambda_1, \ldots, \Lambda_n)\in\mathbb{K}[\Lambda_k]_{k=1}^n$ as \begin{align*} (P_1(\nabla)\otimes \cdots \otimes P_n(\nabla))f(\Lambda_1, \ldots, \Lambda_n) \ceq P_1(-\Lambda_1)\cdots P_n(-\Lambda_n)f(\Lambda_1, \ldots, \Lambda_n). \end{align*} Thus we obtain a left $\mathcal{H}_K^{\otimes n}$-supermodule structure of $V_\nabla[\Lambda_k]_{k=1}^n$. Since the tensor product $V^{\otimes n}$ is naturally a left $\mathcal{H}_K^{\otimes n}$-supermodule, the following definition makes sense: \begin{dfn} For a left $\mathcal{H}_K$-supermodule $V=(V,\nabla)$ and $n \in \mathbb{N}$, we denote \begin{align*} \Chom_V^{N_K}(n) \ceq \Hom_{\mathcal{H}_K^{\otimes n}}(V^{\otimes n}, V_\nabla[\Lambda_k]_{k=1}^n). \end{align*} In other words, $\Chom_V^{N_K}(n)$ is the linear sub-superspace of $\Hom_{\mathbb{K}}(V^{\otimes n}, V_\nabla[\Lambda_k]_{k=1}^n)$ spanned by elements $X$ such that \begin{align*} X(\varphi v) = (-1)^{p(\varphi)p(X)}\varphi X(v) \quad (\varphi \in \mathcal{H}_K^{\otimes n}, \ v \in V^{\otimes n}). \end{align*} To stress the variable $\Lambda_1, \ldots, \Lambda_n$, we express an element $X\in\Chom_V^{N_K}(n)$ as $X_{\Lambda_1, \ldots, \Lambda_n}$. \end{dfn} As in \cref{ss:W:LCA} for the $N_W=N$ case, define the action of $\mathfrak{S}_n$ on $\Chom_V^{N_K}$ and the composition map, then we have an operad $\Chom_V^{N_K}\ceq(\Chom_V^{N_K}(n))_{n\in\mathbb{N}}$. \begin{dfn}\label{dfn:K:Chom} The operad $\Chom_V^{N_K}$ is called the \emph{$N_K=N$ SUSY Lie conformal operad} on the left $\mathcal{H}_K$-supermodule $V$. \end{dfn} The following is the main claim of this \cref{ss:K:LCA}, which is a natural $N_K=N$ SUSY analogue of \cite[Proposition 5.1]{BDHK}. This can be proved in the same way as in the $N_W=N$ case \cref{thm:W:LCA} with minor modification. \begin{thm}\label{thm:K:LCA} Let $V=(V, \nabla)$ be a left $\mathcal{H}_K$-supermodule. \begin{enumerate} \item For an odd element $X\in L^1\bigl(\Chom_{\Pi^{N+1}V}^{N_K}\bigr)$ such that $X\square X=0$, define a linear map $[\cdot_\Lambda\cdot]_X\colon V \otimes V\to V[\Lambda]$ by \begin{align*} [a_\Lambda b]\ceq (-1)^{p(a)(\overline{N}+\overline{1})}X_{\Lambda, -\Lambda-\nabla}(a\otimes b) \quad (a, b\in V). \end{align*} Then $(V, [\cdot_\Lambda\cdot]_X)$ is an $N_K=N$ SUSY Lie conformal algebra. \item The map $X \mapsto [\cdot_\Lambda\cdot]_X$ gives a bijection \begin{align*} \MC\bigl(L\bigl(\Chom_{\Pi^{N+1}V}^{N_K}\bigr)\bigr)_{\overline{1}}\xrr{\sim} \{\text{$N_K=N$ Lie conformal algebra structures on $(V, \nabla)$}\}. \end{align*} \end{enumerate} \end{thm} \subsection{$N_K=N$ SUSY chiral operad}\label{ss:K:VA} In this subsection, we introduce an $N_K=N$ SUSY analogue of the operad $P^{\textup{ch}}$ in \cite{BDHK}. Most results and proofs in \cref{ss:W:VA} for $N_W=N$ case are valid for $N_K=N$ setting with modifications below: \begin{itemize} \item Replace the superalgebra $\mathcal{H}_W$ by $\mathcal{H}_K$. \item Replace $(1|N)_W$-supervariables $\Lambda$, $\Lambda_k$ $(k\in\mathbb{N})$ and $\Gamma$ by $(1|N)_K$-supervariables, while do not replace $(1|N)_W$-supervariables $Z_k$ $(k\in\mathbb{N})$ by $(1|N)_K$-supervariables. \item For $(1|N)_W$-supervariables $Z_k=(z_k, \zeta^1_k, \ldots, \zeta^N_k)$, $k\in\mathbb{N}$, replace \begin{align}\label{eq:K:Z-W} z_{k,l}=z_k-z_l \text{ by } z_k-z_l-\sum_{i=1}^N \zeta^i_k\zeta^i_l, \quad \partial_{\zeta^i_k} \text{ by } D_{\zeta^i_k}\ceq \partial_{\zeta^i_k}+\zeta^i_k\partial_{z_k} \end{align} \end{itemize} \medskip Let us fix a $\Lambda=(\lambda, \theta^1, \ldots, \theta^N)$ be a $(1|N)_K$-supervariable. For even linear transformations $F, G$ on linear superspace $V$, we can define a linear map $\int_F^Gd\Lambda\colon V[\Lambda]\to V$ of parity $\overline{N}$ as in \cref{ss:W:VA}. Also, if $V$ is a superalgebra (not necessarily unital nor associative), we can define a linear map $\int_F^Gd\Lambda\, a$ ($a\in V$) of parity $p(a)$. Using this integral, we introduce: \begin{dfn}[{\cite[Definition 4.19]{HK}}]\label{dfn:NKVA} Let $(V,\nabla,[\cdot_\Lambda\cdot])=(V,T, S^1,\ldots,S^N,[\cdot_\Lambda\cdot])$ be an $N_K=N$ SUSY Lie conformal algebra and $\mu\colon V\otimes V\to V$ be an even linear map. We denote $ab\ceq \mu(a\otimes b)$ for $a, b\in V$. A tuple $(V,\nabla,[\cdot_\Lambda\cdot],\mu)$ is called a \emph{non-unital $N_K=N$ SUSY vertex algebra} if it satisfies the conditions (i)--(iv) in \cref{dfn:NWVA} replacing $(1|N)_W$-supervariables $\Lambda, \Gamma$ by $(1|N)_K$-supervariables. For simplicity, we say $(V, \nabla)$, or more simply $V$, is a non-unital $N_K=N$ SUSY vertex algebra. The linear map $\mu$ is called the \emph{multiplication} of $V$. \end{dfn} \begin{dfn} A non-unital $N_K=N$ SUSY vertex algebra $V$ is called an \emph{$N_K=N$ SUSY vertex algebra} if there exists an even element $\vac\in V$ such that $a\vac=\vac a=a$ for all $a\in V$. \end{dfn} In the remaining of this \cref{ss:K:VA}, we fix a left $\mathcal{H}_K$-supermodule $(V, \nabla)$. \cref{lem:rWick}, \cref{lem:lsym} and their proofs are valid for $N_K=N$ setting with modifications replacing all $(1|N)_W$-supervariables by $(1|N)_K$-supervariables. Also \cref{lem:intbra} and its proof are valid replacing supervariables, thus we have the integral of $\Lambda$-bracket $\int^\Lambda d\Gamma\colon V\otimes V\to V[\Lambda]$ for $(1|N)_K$-supervariable $\Lambda$. Using this integral, one can similarly prove \cref{prp:W:skecom} and \cref{prp:W:Jqas} with $(1|N)_W$-supervariables replaced by $(1|N)_K$-supervariables. \begin{dfn} Let $n\in\mathbb{N}$ and $Z_k=(z_k, \zeta_k^1, \ldots, \zeta_k^N)$ be a $(1|N)_W$-supervariable for each $k\in[n]$. We set \begin{align*} z_{k, l}\ceq z_k-z_l-\sum_{i=1}^N\zeta_k^i\zeta_l^i, \quad \zeta_{k, l}^i\ceq \zeta_k^i-\zeta_l^i \end{align*} for $k, l\in[n]$ and $i\in[N]$. Also, we set $Z_{k, l}\ceq (z_{k, l}, \zeta_{k, l}^1, \ldots, \zeta_{k, l}^N)$ for simplicity. \begin{enumerate} \item We denote by \begin{align*} \mathcal{O}_n^{\star}\ceq \mathbb{K}[Z_k]_{k=1}^n[z_{k, l}^{-1}]_{1\le k<l\le n} \end{align*} the localization of $\mathbb{K}[Z_k]_{k=1}^n$ by the multiplicatively closed set generated by $\{z_{k, l}\mid 1\le k<l\le n\}$. \item Let $\mathcal{O}_n^{\star\mathrm{T}}$ denote the sub-superalgebra of $\mathcal{O}_n^{\star}$ generated by the subset $\{z_{k, l}^{\pm 1}\mid 1\le k<l\le n\}\cup \{\zeta^i_{k, l}\mid i\in[N], 1\le k< l\le n\}$. Thus, we have \begin{align*} \mathcal{O}_n^{\star\mathrm{T}}=\mathbb{K}[z_{k, l}^{\pm1}, \zeta_{k, l}^i\mid i\in[N], 1\le k<l\le n]. \end{align*} \item Let $\mathcal{D}_n$ denote the superalgebra of regular differential operators of $Z_k$, i.e, the sub-superalgebra of $\End_\mathbb{K}\bigl(\mathbb{K}[Z_k]_{k=1}^n\bigr)$ generated by $Z_k=(z_k, \zeta_k^1, \ldots, \zeta_k^N)$ and $D_{Z_k}=(\partial_{z_k}, D_{\zeta_k^1}, \ldots, D_{\zeta_k^N})$ for $k\in[n]$. As a superspace, we have \begin{align*} \mathcal{D}_n=\mathbb{K}[Z_k]_{k=1}^n[D_{Z_k}]_{k=1}^n =\mathbb{K}[z_k, \zeta_k^1, \ldots, \zeta_k^N]_{k=1}^n[\partial_{z_k}, D_{\zeta_k^1}, \ldots, D_{\zeta_k^N}]_{k=1}^n. \end{align*} \item Let $\mathcal{D}_n^\mathrm{T}$ denote the sub-superalgebra of $\mathcal{D}_n$ generated by the subset $\{z_{k, l}, \zeta_{k, l}^i\mid i\in[N], 1\le k<l\le n\}\cup \{\partial_{z_k}, D_{\zeta_k^i}\mid i\in[N], k\in[n]\}$. Thus we have \begin{align*} \mathcal{D}_n^\mathrm{T}=\mathbb{K}[Z_{k, l}]_{1\le k<l\le n}[D_{Z_k}]_{k=1}^n. \end{align*} \end{enumerate} \end{dfn} The superspace $V^{\otimes n}\otimes \mathcal{O}_n^{\star\mathrm{T}}$ carries a structure of a right $\mathcal{D}_n$-supermodule by letting $Z_k=(z_k, \zeta_k^1, \ldots, \zeta_k^N)$ act as \begin{align*} (v\otimes f)\cdot z_k\ceq v\otimes fz_k, \quad (v\otimes f)\cdot \zeta_k^i \ceq v\otimes f\zeta_k^i, \end{align*} and $D_{Z_k}=\bigl(\partial_{z_k}, D_{\zeta_k^1}, \ldots, D_{\zeta_k^N}\bigr)$ act as \begin{align*} (v\otimes f)\cdot \partial_{z_k}\ceq T^{(k)}v\otimes f-v\otimes \partial_{z_k}f, \quad (v\otimes f)\cdot D_{\zeta_k^i}\ceq (-1)^{p(v)+p(f)}(S^i)^{(k)}v\otimes f+(-1)^{p(f)}v\otimes \partial_{\zeta_k^i}f \end{align*} for $v \in V^{\otimes n}$, $f \in \mathcal{O}_n^{\star\mathrm{T}}$. Here, for a linear transformation $\varphi \in \End V$, the symbol $\varphi^{(k)}$ denotes the linear transformation on $V^{\otimes n}$ defined by $\varphi^{(k)} \ceq \id_V \otimes \dotsb \otimes \overset{k}{\varphi} \otimes \dotsb \otimes \id_V$. Now let us recall the superspace $V_\nabla[\Lambda_k]_{k=1}^n$ in \eqref{eq:LCA:HKn-mod}. It also has a structure of right $\mathcal{D}_n$-supermodule by letting \begin{align*} a(\Lambda_1,\dotsc,\Lambda_n) \cdot z_k \ceq -\partial_{\lambda_k} a(\Lambda_1,\dotsc,\Lambda_n), \quad a(\Lambda_1,\dotsc,\Lambda_n) \cdot \zeta_k^i \ceq -(-1)^{p(a)}\partial_{\theta_k^i} a(\Lambda_1,\dotsc,\Lambda_n) \end{align*} and \begin{align*} a(\Lambda_1, \ldots, \Lambda_n)\cdot \partial_{z_k}\ceq -a(\Lambda_1, \ldots, \Lambda_n)\lambda_k, \quad a(\Lambda_1, \ldots, \Lambda_n)\cdot D_{\zeta_k^i}\ceq -a(\Lambda_1, \ldots, \Lambda_n)\theta_k^i \end{align*} for $a(\Lambda_1, \ldots, \Lambda_n)\in V_\nabla[\Lambda_k]_{k=1}^n$. Thus, by restriction, we have right $\mathcal{D}_n^\mathrm{T}$-supermodules $V^{\otimes n}\otimes \mathcal{O}_n^{\star\mathrm{T}}$ and $V_\nabla[\Lambda_k]_{k=1}^n$. \begin{dfn} For a left $\mathcal{H}_K$-supermodule $V=(V,\nabla)$ and $n \in \mathbb{N}$, we denote \begin{align*} P^{\textup{ch}N_K}_V(n) \ceq \Hom_{\mathcal{D}_n^\mathrm{T}} \bigl(V^{\otimes n} \otimes \mathcal{O}_n^{\star \mathrm{T}},V_{\nabla}[\Lambda_k]_{k=1}^n\bigr). \end{align*} To stress the variables $\Lambda_1, \ldots, \Lambda_n$, we express an element $X \in P^{\textup{ch}N_K}_V(n)$ as $X_{\Lambda_1, \ldots, \Lambda_n}$. \end{dfn} In other words, $P^{\textup{ch}N_K}_V(n)$ is a linear sub-superspace of $\Hom_\mathbb{K}\bigl(V^{\otimes n} \otimes \mathcal{O}_n^{\star \mathrm{T}},V_{\nabla}[\Lambda_k]_{k=1}^n\bigr)$ spanned by elements $X$ such that \begin{align*} X((v \otimes f) \cdot \varphi) = X(v \otimes f) \cdot \varphi \quad (v \in V^{\otimes n}, \, f \in \mathcal{O}_n^{\star\mathrm{T}}, \, \varphi \in \mathcal{D}_n^{\mathrm{T}}). \end{align*} As in \cref{ss:W:VA} for the $N_W=N$ case, define the action of $\mathfrak{S}_n$ and the composition map, then we have an operad $P^{\textup{ch}N_K}_V\ceq(P^{\textup{ch}N_K}_V(n))_{n\in\mathbb{N}}$. \begin{dfn}\label{dfn:K:Pch} For a left $\mathcal{H}_K$-supermodule $V=(V, \nabla)$, we call the operad $P^{\textup{ch}N_K}_V$ the \emph{$N_K=N$ SUSY chiral operad}. \end{dfn} The following is the main statement of this \cref{ss:K:VA}, which is a natural $N_K=N$ SUSY analogue of \cite[Theorem 6.12]{BDHK}. \begin{thm}\label{thm:K:VA} Let $(V, \nabla)=(V, T, S^1, \ldots, S^N)$ be a left $\mathcal{H}_K$-supermodule. \begin{enumerate} \item For an odd element $X\in L^1\bigl(P^{\textup{ch}N_K}_{\Pi^{N+1}V}\bigr)$ satisfying $X\square X=0$, define linear maps $[\cdot_\Lambda\cdot]_X\colon V\otimes V\to V[\Lambda]$ and $\mu_X\colon V\otimes V\to V$ by \begin{align} \label{eq:K:VAlbra} [a_\Lambda b]_X &\ceq (-1)^{p(a)(\overline{N}+\overline{1})}X_{\Lambda, -\Lambda-\nabla}(a\otimes b\otimes 1_{\mathbb{K}}), \\ \label{eq:K:VAprod} \mu_X(a\otimes b) &\ceq (-1)^{p(a)(\overline{N}+\overline{1})+\overline{1}}\Res_\Lambda(\lambda^{-1}X_{\Lambda, -\Lambda-\nabla}(a\otimes b\otimes z_{1, 2}^{-1})) \end{align} for $a, b\in V$. Then $(V, \nabla, [\cdot_\Lambda\cdot]_X, \mu_X)$ is a non-unital $N_K=N$ SUSY vertex algebra. \item The map $X \mapsto ([\cdot_\Lambda\cdot]_X, \mu_X)$ gives a bijection \begin{align*} \MC\bigl(L\bigl(P^{\textup{ch}N_K}_{\Pi^{N+1}V}\bigr)\bigr)_{\overline{1}} \xrr{\sim} \{\text{non-unital $N_K=N$ SUSY vertex algebra structures on $(V,\nabla)$}\}. \end{align*} \end{enumerate} \end{thm} In the rest of this section, we prove \cref{thm:K:VA}. We denote $\widetilde{V}\ceq \Pi^{N+1}V$ and $P^{\textup{ch}}_{\widetilde{V}}\ceq P^{\textup{ch}N_K}_{\Pi^{N+1}V}$ for simplicity, and let $p, \widetilde{p}$ be the parity of $V, \widetilde{V}$ respectively. Consider the linear maps $[\cdot_\Lambda\cdot]_X\colon V\otimes V\to V[\Lambda]$ and $\mu_X\colon V\otimes V\to V$ defined by \eqref{eq:K:VAlbra} and \eqref{eq:K:VAprod} for an odd element $X\in P^{\textup{ch}}_{\widetilde{V}}(2)$. Similarly to the proof of \cref{thm:W:VA}, one can show that the map $X \mapsto ([\cdot_\Lambda\cdot]_X, \mu_X)$ gives a bijective correspondence between the set of all odd elements $X\in P^{\textup{ch}}_{\widetilde{V}}(2)$ and the set of all pairs $([\cdot_\Lambda\cdot], \mu)$ of a linear map $[\cdot_\Lambda\cdot]\colon V\otimes V\to V[\Lambda]$ of parity $\overline{N}$ satisfying (i) in \cref{dfn:K:LCA} and an even linear map $\mu\colon V\otimes V\to V$ satisfying (i) in \cref{dfn:NKVA}. Also, one can prove that the skew-symmetry and the quasi-commutativity are equivalent to $X^\sigma=X$ for any $\sigma\in\mathfrak{S}_2$. In what follows, let $X\in L^1\bigl(P^{\textup{ch}}_{\widetilde{V}}\bigr)_{\overline{1}}$, so that $[\cdot_\Lambda\cdot]_X\colon V\otimes V\to V[\Lambda]$ and $\mu_X\colon V\otimes V\to V$ are linear maps satisfying (i), (ii) in \cref{dfn:K:LCA} and (i), (ii) in \cref{dfn:NKVA}. We need to prove that the Jacobi identity, the quasi-associativity and the Wick formula are equivalent to $X\square X=0$. First, we have the following lemmas. \begin{lem}\label{lem:K:circ} For an odd element $X\in P^{\textup{ch}N_K}_{\Pi^{N+1}V}(2)$, we have \begin{align*} &\Res_{\Lambda_2}(\lambda_2^{-1}(X\circ_1X)(a\otimes b\otimes c\otimes z_{1, 2}^{-1}z_{1, 3}^{-1}z_{2, 3}^{-1}))\\ &=(-1)^{p(b)(\overline{N}+\overline{1})}\Bigl(\Bigl(\int_{-T}^{\lambda_1}d\lambda_1\int^{\Lambda_1}d\Gamma_1[a_{\Gamma_1}b]\Bigr)c +\int_0^{\lambda_1}d\Lambda\Bigl[ \Bigl(\int_{-T}^{\lambda_1}\int^{\Lambda_1}d\Gamma_1[a_{\Gamma_1}b]\Bigr)_\Lambda c\Bigr]\Bigr), \\ &\Res_{\Lambda_2}(\lambda_2^{-1}(X\circ_2X)(a\otimes b\otimes c\otimes z_{1, 2}^{-1}z_{1, 3}^{-1}z_{2, 3}^{-1}))\\ &=-(-1)^{p(b)(\overline{N}+\overline{1})}\Bigl(\int^{\Lambda_1}d\Gamma_1[(\lambda_1+T)a_{\Gamma_1}bc] +\int^{\Lambda_1}d\Gamma_1\Bigl[\Bigl(\int_0^{\lambda_1+T}d\Lambda a\Bigr)_{\Gamma_1}\Bigl(\int_0^{\lambda}d\lambda[b_\Lambda c]\Bigr)\Bigr]\Bigr), \\ &=\Res_{\Lambda_2}(\lambda_2^{-1}(X\circ_2X)^{(1, 2)}(a\otimes b\otimes c\otimes z_{1, 2}^{-1}z_{1, 3}^{-1}z_{2, 3}^{-1}))\\ &=-(-1)^{p(b)(\overline{N}+\overline{1})}(-1)^{(p(a)+\overline{N})p(b)} \Bigl(\int_{\lambda_1}^{\lambda_1+T}d\lambda_1b\Bigr)\int^{\Lambda_1}d\Gamma_1[a_{\Gamma_1}c]. \end{align*} \end{lem} \begin{proof} One can calculate the first identity in the same way as in the proof of \cref{lem:W:circ}. Similarly, we get the third identity. Also, we have \begin{align*} &\Res_{\Lambda_2}(\lambda_2^{-1}(X \circ_2 X) (a \otimes b \otimes c \otimes z_{1, 2}^{-1}z_{1, 3}^{-1}z_{2, 3}^{-1}))\\ &=(-1)^{p(a)(\overline{N}+\overline{1})}(-1)^{p(b)(\overline{N}+\overline{1})} X_{\Lambda_1, -\Lambda_1-\nabla}\Bigl(a \otimes \Res_{\Lambda_2}\Bigl(\lambda_2^{-1}\int^{\Lambda_2-D_{Z_2}}d\Gamma_2 [b_{\Gamma_2}c]\Bigr) \otimes z_{1, 2}^{-1}z_{1, 3}^{-1}|_{z_2=z_3}\Bigr) \\ &=-(-1)^{p(b)(\overline{N}+\overline{1})} \Bigl(\int^{\Lambda_1} d\Gamma_1 [(\lambda_1+T)a_{\Gamma_1}bc] +\int^{\Lambda_1}d\Gamma_1 \Bigl[\Bigl(\int_0^{\lambda_1+T}d\Lambda a\Bigr)_{\Gamma_1} \Bigl(\int_0^{\lambda}d\lambda[b_\Lambda c]\Bigr)\Bigr]\Bigr). \end{align*} \end{proof} \begin{lem}\label{lem:K:boxJqW} For an odd element $X\in L^1\bigl(P^{\textup{ch}N_K}_{\Pi^{N+1}V}\bigr)$, we have \begin{align*} &(-1)^{(p(a)+\overline{N})\overline{N}+p(b)(\overline{N}+\overline{1})+\overline{1}}(X\square X) (a\otimes b\otimes c\otimes z_{1, 3}^{-1}z_{2, 3}^{-1}) \\ &=\int^{\Lambda_1} d\Gamma_1 \left[a_{\Gamma_1}\left(\int^{\Lambda_2}d\Gamma_2[b_{\Gamma_2}c]\right)\right] -(-1)^{(p(a)+\overline{N})(p(b)+\overline{N})} \int^{\Lambda_2}d\Gamma_2 \left[b_{\Gamma_2}\left(\int^{\Lambda_1}d\Gamma_1[a_{\Gamma_1}c]\right)\right]\\ &\quad -(-1)^{(p(a)+N)N} \int^{\Lambda_1+\Lambda_2}d\Gamma \left[\left(\int^{\Lambda_1}d\Gamma_1 [a_{\Gamma_1}b] -\int^{-\Lambda_2-\nabla}d\Gamma_2 [a_{\Gamma_2}b]\right)_{\Gamma}c\right]. \end{align*} \end{lem} \begin{proof} First, a direct calculation shows that \begin{align*} &(X \circ_2 X)(a \otimes b \otimes c\otimes z_{1, 3}^{-1}z_{2, 3}^{-1}) =(-1)^{(p(a)+\overline{N})\overline{N}+p(b)(\overline{N}+\overline{1})+\overline{1}} \int^{\Lambda_1}d\Gamma_1 \Bigl[a_{\Gamma_1}\Bigl(\int^{\Lambda_2}d\Gamma_2[b_{\Gamma_2}c]\Bigr)\Bigr], \\ &(X \circ_2 X)^{(1, 2)}(a \otimes b\otimes c\otimes z_{1, 3}^{-1}z_{2, 3}^{-1}) =(-1)^{p(a)p(b)+p(b)} \int^{\Lambda_2} d\Gamma_2 \Bigl[b_{\Gamma_2}\Bigl(\int^{\Lambda_1} d\Gamma_1 [a_{\Gamma_1}c]\Bigr)\Bigr]. \end{align*} Also, we have \begin{align*} (X \circ_1 X)(a \otimes b \otimes c \otimes z_{1, 3}^{-1}z_{2, 3}^{-1}) &=(\partial_{\lambda_2}-\partial_{\lambda_1})(X \circ_1X) (a \otimes b \otimes c \otimes z_{1, 2}^{-1}z_{1, 3}^{-1}z_{2, 3}^{-1})\\ &=(X \circ_1 X)(a \otimes b \otimes c \otimes z_{1, 2}^{-1}z_{2, 3}^{-1}) -(X \circ_1 X)(a \otimes b \otimes c \otimes z_{1, 2}^{-1}z_{1, 3}^{-1}). \end{align*} Then, a direct calculation yields \begin{align*} &(X \circ_1X)(a\otimes b\otimes c\otimes z_{1, 2}^{-1}z_{2, 3}^{-1}) =(-1)^{p(b)(\overline{N}+\overline{1})} \int^{\Lambda_1+\Lambda_2} d\Gamma \Bigl[\Bigl(\int^{\Lambda_1}d\Gamma_1[a_{\Gamma_1}b]\Bigr)_\Gamma c\Bigr], \\ &(X \circ_1X)(a\otimes b\otimes c\otimes z_{1, 2}^{-1}z_{1, 3}^{-1}) =(-1)^{p(b)(\overline{N}+\overline{1})} \int^{\Lambda_1+\Lambda_2} d\Gamma \Bigl[\Bigl(\int^{-\Lambda_2-\nabla}d\Gamma_2[a_{\Gamma_2}b]\Bigr)c\Bigr], \end{align*} and using $X^\sigma=X$ $(\sigma\in\mathfrak{S}_2)$ for the second identity, we get \begin{align*} (X \circ_1 X)(a \otimes b \otimes c \otimes z_{1, 3}^{-1}z_{2, 3}^{-1}) =(-1)^{p(b)(\overline{N}+\overline{1})} \int^{\Lambda_1+\Lambda_2} d\Gamma \left[\left(\int^{\Lambda_1} d\Gamma_1 [a_{\Gamma_1}b] -\int^{-\Lambda_2-\nabla} d\Gamma_2 [a_{\Gamma_2}b]\right)_{\Gamma}c\right]. \end{align*} The claim of this lemma is now clear. \end{proof} By \cref{lem:K:boxJqW}, it is clear that $X\square X=0$ implies the Jacobi identity, the quasi-associativity and the Wick formula. Conversely, assume the Jacobi identity, the quasi-associativity and the Wick formula. To prove $X\square X=0$, it is sufficient to show that \begin{align}\label{eq:K:ResFn} \Res_{\Lambda_1}\Res_{\Lambda_2}\bigl(\lambda_1^{-1} \lambda_2^{-1}\partial_{\lambda_1}^n(X\square X)(a\otimes b\otimes c\otimes z_{1, 2}^{-1}z_{1, 3}^{-1}z_{2, 3}^{-1})\bigr)=0 \end{align} for all $n \in \mathbb{N}$ and $a,b,c \in V$, because we have \begin{align*} & \Res_{\Lambda_1}\Res_{\Lambda_2}\bigl(\lambda_1^{-1} \lambda_2^{-1}\partial_{\lambda_1}^n\theta_1^I\theta_2^J(X\square X)(a\otimes b\otimes c\otimes z_{1, 2}^{-1}z_{1, 3}^{-1}z_{2, 3}^{-1})\bigr)\\ & =(-1)^{\#I}(-1)^{(\widetilde{p}(a)+\overline{1})\#J} \Res_{\Lambda_1}\Res_{\Lambda_2}\bigl(\lambda_1^{-1} \lambda_2^{-1}\partial_{\lambda_1}^n(X\square X)(S^Ia\otimes S^Jb\otimes c\otimes z_{1, 2}^{-1}z_{1, 3}^{-1}z_{2, 3}^{-1})\bigr) \end{align*} for any $I, J\subset [N]$ and \begin{align*} (\partial_{\lambda_2}-\partial_{\lambda_1})(X\square X)(a\otimes b\otimes c\otimes z_{1, 2}^{-1}z_{1, 3}^{-1}z_{2, 3}^{-1}) =(X\square X)(a\otimes b\otimes c\otimes z_{1, 3}^{-1}z_{2, 3}^{-1})=0. \end{align*} The identity \eqref{eq:K:ResFn} can be proved as in the proof of \cref{thm:W:VA} using \cref{lem:K:circ}. \subsection{Relation to $N_K=N$ SUSY chiral algebras}\label{ss:K:CA} We can relate our operad $P^{\textup{ch}N_K}$ to the $N_K=N$ SUSY chiral algebras, as we did for the $N_W=N$ case in \cref{ss:W:CA}. We just comment the necessary modification, and suppress the full presentation. Instead of a $(1|N)$-supercurve, we consider an oriented superconformal $(1|N)$-supercurve $X$ \cite[2.2.11]{H}. It is defined to be a $(1|N)$-supercurve equipped with the differential form $\omega$ which is locally give by \[ \omega = d z + \sum_{k=1}^N \zeta^i d \zeta^i \] with respect to some local coordinate $Z_i=(z,\zeta^1,\dotsc,\zeta^N)$, and is well-defined up to multiplication by a function. Then, recalling \eqref{eq:K:Z-W}, we replace the big diagonal $\Delta\colon X \subset X^n$ in \eqref{eq:W:bigdiag} by the \emph{super-diagonal} $\Delta^s\colon X \subset X^n$. Using local coordinates $Z_i=(z_i,\zeta_i^1,\dotsc,\zeta_i^N)$, it is the union of the divisors (sub-superscheme of $1|0$-codimension) in $X^n$ defined by \[ 0 = \zeta_i-\zeta_j-\sum_{k=1}^N \zeta_i^k \zeta_j^k \quad (i \ne j \in [n]). \] Then we have the standard pushforward functor $\Delta^s_*$ of $\mathcal{D}$-modules. Together with the open embedding $j\colon U \hookrightarrow X^n$, we define \[ \oP^{\textup{ch}N_K}_{\mathscr{A}}(n) \ceq \Hom_{\crMod \mathscr{D}_{X^n}}\bigl(j_* j^* \mathscr{A}^{\boxtimes n},\Delta^s_*\mathscr{A}\bigr) \] for a right $\mathscr{D}_X$-module $\mathscr{A}$. It gives rise to an operad $\oP^{\textup{ch}N_K}_{\mathscr{A}} = \bigl(\oP^{\textup{ch}N_K}_{\mathscr{A}}(n)\bigr)_{n \in \mathbb{N}}$. The translation group $\mathrm{T} = \mathbb{G}_a \times \mathbb{G}_a^N$ acts on $\oP^{\textup{ch}N_K}_{\mathscr{A}}(n)$ by the formula \eqref{eq:W:A10} with $Z_i-U$ understood as \eqref{eq:K:Z-W}. Then the operad $\oP^{\textup{ch}N_K}_{\mathscr{A}}$ is $\mathrm{T}$-equivariant, and the $\mathrm{T}$-invariants $\oP^{\textup{ch}N_K,\mathrm{T}}_{\mathscr{A}}(n) \subset \oP^{\textup{ch}N_K}_{\mathscr{A}}(n)$ form a sub-operad \[ \oP^{\textup{ch}N_K,\mathrm{T}}_{\mathscr{A}} \subset \oP^{\textup{ch}N_K}_{\mathscr{A}}. \] The remaining part goes similarly as the $N_W=N$ case in \cref{ss:W:CA}, and we have: \begin{prp} Let $V=(V,\nabla)$ be an $\mathcal{H}_K$-supermodule, and $\mathscr{A}$ be the corresponding weakly $\mathrm{T}$-equivariant $\mathcal{D}$-module on $\mathbb{A}^{1|N}$. Let $\oP^{\textup{ch}N_K}_V$ be the $N_K=N$ SUSY chiral operad. Then there is an isomorphism of operads \[ \oP^{\textup{ch}N_K,\mathrm{T}}_{\mathscr{A}} \cong \oP^{\textup{ch}N_K}_V. \] \end{prp} \section{Cohomology of SUSY vertex algebras}\label{s:coh} \subsection{Recollection on the Lie algebra cohomology}\label{ss:coh:CE} Here we give a brief recollection on the Chevalley-Eilenberg cohomology of a Lie algebra from the viewpoint of the operad theory. For the detail, we refer to \cite[Chap.\ 12, \S 13.2]{LV}. We use the language of operads reviewed in \cref{s:op}. Let $V$ be a linear space. Recall the convolution Lie algebra \[ \mathfrak{g}_{\oLie,V} = \bigl(\Hom_{\mathfrak{S}}(\oLie^{\ash},\oHom_V),[\cdot,\cdot]\bigr) \] in \eqref{eq:1:gPV} and the set of solutions of the Maurer-Cartan equation \[ \MC(\mathfrak{g}_{\oLie,V}) = \{X \in \mathfrak{g}_{\oLie,V}^1 \mid \tfrac{1}{2}[X,X]=0\}, \] both being specialized for the Lie operad $\oP = \oLie$. A Maurer-Cartan solution $X \in \MC(\mathfrak{g}_{\oLie,V})$ corresponds bijectively to a Lie algebra structure $[\cdot,\cdot]_X$ on $V$ . Given a Maurer-Cartan solution $X \in \MC(\mathfrak{g}_{\oLie,V})$, we have a differential \[ d_X \ceq [X,-]\colon \mathfrak{g}_{\oLie,V}^{\bullet} \longrightarrow \mathfrak{g}_{\oLie,V}^{\bullet+1}, \quad d_X^2=0. \] The obtained cochain complex $(\mathfrak{g}_{\oLie,V}^{\bullet},d_X)$ coincides with the Chevalley-Eilenberg cochain complex $C^\bullet(L,L)$ of the Lie algebra $L=(V,[\cdot,\cdot]_X)$ with coefficients in itself as an adjoint module, up to a degree shift. It is known that the cochain complex $(\mathfrak{g}_{\oLie,V}^{\bullet},d_X)$ has a Lie bracket, giving a dg Lie algebra structure, and that the Lie bracket is essentially equal to the Nijenhuis-Richardson bracket on the Chevalley-Eilenberg cochain complex. We can extend the construction of the cochain complex $(\mathfrak{g}_{\oLie,V}^{\bullet},d_X)$ to the version with coefficients in a module over the Lie algebra $(V,[\cdot,\cdot]_X)$ along the line of the general argument for a quadratic operad in \cite[\S12.4]{LV}. Let us give a brief recollection. \begin{itemize} \item First, for an operad $\oP$ and a linear space $V$, we have an isomorphism \begin{align}\label{eq:coh:PV} \Hom_{\mathfrak{S}}(\oP,\oHom_V) \cong C_{\oP}(V,V) \ceq \Hom(\oP(V),V) \end{align} by the hom-tensor adjunction and \eqref{eq:1:HomV}, where \[ \oP(V) \ceq \bigoplus_{n \ge 0}\oP(V)(n) \otimes_{\mathfrak{S}_n} V^{\otimes n} \] is the image of the Schur functor of the operad $\oP$ \cite[\S 5.1.2]{LV}. Thus, a $\oP$-algebra structure $\varphi\colon \oP \to \oHom_V$ can be encoded by $C_{\oP}(V,V)$. The linear space $C_{\oP}(V,V)$ has an $\mathbb{N}$-grading with $C_{\oP}^n(V,V) \ceq \Hom(\oP(V)(n) \otimes_{\mathfrak{S}_n} V^{\otimes n},V)$. \item Next, recall from \cite[\S 12.3.1]{LV} that a module $M$ over a $\oP$-algebra $(V,\varphi)$ is defined to be a linear space equipped with two morphisms $\gamma_M\colon \oP \circ (V;M) \to M$ and $\eta_M\colon M \to \oP \circ (V;M)$ satisfying some axioms. Here we used the composite product \cite[\S6.1.1]{LV} \[ \oP \circ (V;M) \ceq \bigoplus_{n \ge 0}\oP(n) \otimes_{\mathfrak{S}_n} \Bigl( \bigoplus_{1 \le i \le n} V^{\otimes i} \otimes M \otimes V^{\otimes n-i}\Bigr). \] As it implies, the notion of modules over a $\oP$-algebra is a natural operadic analogue of bimodules over the algebra structure considered. \item Following \cite[\S 12.4.1]{LV}, let $\oP$ be a homogeneous quadratic operad, $V$ be a linear space equipped with a $\oP$-algebra structure, and $M$ be a module over the $\oP$-algebra $V$. Let us consider the linear space \begin{align}\label{eq:coh:CVM} C_{\oP}(V,M) \ceq \Hom(\oP^{\ash}(V),M) \end{align} which has the natural $\mathbb{N}$-grading induced by that on $\oP^{\ash}(V)$. In the case $M=V$, this graded space coincides with $C_{\oP}^{\bullet}(V,V)$ in \eqref{eq:coh:PV}. According to \cite[\S 12.4.1]{LV}, the graded space $C_{\oP}^{\bullet}(V,M)$ has a natural differential $d$ coming from the $\oP$-algebra structure on $V$ and the $V$-module structure on $M$, and we obtain a cochain complex $(C_{\oP}^{\bullet}(V,M),d)$ called the \emph{operadic cochain complex}. \item The cohomology of the operadic cochain complex $(C_{\oP}^{\bullet}(V,M),d)$ is denoted by $H^{\bullet}_{\oP}(A,M)$. The low degree cohomology have analogous interpretation as the classical cohomology. For example, we have $H^0_{\oP}(A,M) \cong \Der_A(A,M)$, the space of derivations. We refer to \cite[\S 12.4]{LV} for the detail. \item In the case $\oP=\oLie$, the operadic cochain complex $(C_{\oP}(V,M),d)$ coincides, up to a degree shift, with the Chevalley-Eilenberg cochain complex of the Lie algebra $V$ with coefficients in the $V$-module $M$. Similarly, we can recover the Harrison cochain complex in the case $\oP=\oCom$, and recover the Hochschild cochain complex in the case $\oP=\oAss$. \end{itemize} In \cite[\S7]{BDHK}, the authors introduced a natural analogue of the Chevalley-Eilenberg complex for vertex algebras, based on their operad $P^{\textup{ch}}_V$. As explained in \cref{s:op}, in order to have a vertex algebra analogue in the operad theory, we should keep $\oLie$ as it is, and replace the endomorphism operad $\oHom_V$ by the chiral operad $\oP^{\textup{ch}}_V$. The construction in loc.\ cit.\ can be understood as such a replacement in \eqref{eq:coh:PV} and \eqref{eq:coh:CVM}. We will not give the detail of the construction in \cite[\S7]{BDHK}. Instead, we explain a straightforward modification for SUSY vertex algebras in the following \cref{ss:coh:W} and \cref{ss:coh:K}. \subsection{Cohomology of $N_W=N$ SUSY vertex algebras}\label{ss:coh:W} In this subsection, we introduce a cohomology of an $N_W=N$ SUSY vertex algebra with coefficients in its module, and investigate the low degree cohomology. First, we introduce modules over an $N_W=N$ SUSY vertex algebra. \begin{dfn}\label{dfn:W:LCAmod} Let $(V, \nabla, [\cdot_\Lambda\cdot])$ be an $N_W=N$ SUSY Lie conformal algebra, $(M,\nabla)$ be an $\mathcal{H}_W$-supermodule and $\rho_\Lambda\colon V \otimes M \to M[\Lambda]$ be a linear map of parity $\overline{N}$. We denote \begin{align}\label{eq:coh:aLx} a_\Lambda x \ceq \rho_\Lambda(a \otimes x) \end{align} for $a \in V$ and $x \in M$. A triple $(M,\nabla,\rho_\Lambda)$ is called a \emph{module over $(V,\nabla,[\cdot_\Lambda\cdot])$} if the following conditions are satisfied. \begin{clist} \item For any $a \in V$ and $x \in M$, \begin{align}\label{eq:Lmodsesq} \begin{aligned} & (Ta)_\Lambda x = - \lambda a_\Lambda x, & & a_\Lambda (Tx) = (\lambda+T)a_\Lambda x,\\ &(S^ia)_\Lambda x = -(-1)^N \theta^i a_\Lambda x, & & a_\Lambda (S^ix) = (-1)^{p(a)+\overline{N}}(\theta^i+S^i)a_\Lambda x \quad (\forall i \in [N]). \end{aligned} \end{align} \item For any $a, b\in V$ and $x\in M$, \begin{align}\label{eq:LmodJ} a_{\Lambda_1}(b_{\Lambda_2}x) =(-1)^{(p(a)+\overline{N})\overline{N}}[a_{\Lambda_1}b]_{\Lambda_1+\Lambda_2}x +(-1)^{(p(a)+\overline{N})(p(b)+\overline{N})}b_{\Lambda_2}(a_{\Lambda_1}x), \end{align} where $\Lambda_1, \Lambda_2$ are $(1|N)_W$-supervariables. \end{clist} For simplicity, we say $(M, \nabla)$, or more simply $M$, is a module over $V$. The map $\rho_\Lambda$ is called \emph{the} (\emph{left}) \emph{$\Lambda$-action on $M$}. \end{dfn} \begin{dfn}\label{dfn:W:VAmod} Let $(V, \nabla, [\cdot_\Lambda\cdot],\mu)$ be a non-unital $N_W=N$ SUSY vertex algebra, $(M,\nabla,\rho_\Lambda)$ be a module over $N_W=N$ Lie conformal algebra $(V,\nabla,[\cdot_\Lambda \cdot])$ and $\rho\colon V \otimes M \to M$ be an even linear map. We denote \begin{align}\label{eq:coh:adx} a \cdot x \ceq \rho(a \otimes x) \end{align} for $a \in V$ and $x \in M$. Then, a tuple $(M,\nabla,\rho_\Lambda,\rho)$ is called a \emph{module over $(V,\nabla,[\cdot_\Lambda\cdot],\mu)$} if it satisfies the following conditions: \begin{clist} \item For any $a \in V$ and $x \in M$, \begin{align}\label{eq:VAmodder} T(a \cdot x) = (Ta) \cdot x + a \cdot (Tx), \quad S^i(a \cdot x) = (S^ia) \cdot x + (-1)^{p(a)}a \cdot (S^ix) \quad (i \in [N]). \end{align} \item For any $a, b\in V$ and $x\in M$, \begin{align*} a\cdot(b\cdot x) =(ab) \cdot x + \Bigl(\int_0^T d\Lambda a\Bigr)\cdot b_\Lambda x +(-1)^{p(a)p(b)}\Bigl(\int_0^T d\Lambda b\Bigr)\cdot a_\Lambda x. \end{align*} \item For any $a, b\in V$ and $x\in M$, \begin{align*} a_\Lambda(b \cdot x) =[a_\Lambda b] \cdot x + (-1)^{(p(a)+\overline{N})p(b)}b \cdot (a_\Lambda x) + \int_0^\lambda d\Gamma[a_\Lambda b]_\Gamma x. \end{align*} \end{clist} For simplicity, we say $(M, \nabla)$, or more simply $M$, is a module over $V$. \end{dfn} \begin{rmk} For a module $M$ over $V$, we define a right action of $V$ by \begin{align}\label{eq:ract} x_\Lambda a \ceq -(-1)^{p(a)p(x)+\overline{N}} a_{-\Lambda-\nabla}x, \quad x \cdot a \ceq (-1)^{p(a)p(x)} a \cdot x \quad (a \in V, \, x \in M). \end{align} Then $M$ is a ``right'' module over $V$. \end{rmk} In the remaining of this \cref{ss:coh:W}, the symbol $V$ denotes a non-unital $N_W=N$ SUSY vertex algebra, and $M$ denotes a module over $V$. Let us introduce the underlying graded superspace of the cohomology complex. \begin{dfn} For a module $M$ over an $N_W=N$ SUSY vertex algebra $V$ and $n \in \mathbb{N}$, we define a linear superspace $C^n(V,M)$ by \begin{align*} C^n(V, M)\ceq \Hom_{\mathcal{D}_n^\mathrm{T}}\bigl((\Pi^{N+1} V)^{\otimes n}\otimes \mathcal{O}_n^{\star \mathrm{T}}, (\Pi^{N+1}M)_\nabla[\Lambda_k]_{k=1}^n\bigr)^{\mathfrak{S}_n}. \end{align*} \end{dfn} Note that we can naturally consider $C^n(V, M)\subset L^{n-1}\bigl(P^{\textup{ch} N_W}_{\Pi^{N+1}(V\oplus M)}\bigr)$ for $n\in\mathbb{N}$ (see \cref{dfn:1:LP} for the symbol $L^{n-1}$). Next, we want to define the differential on $C^{\bullet}(V,M)$. For that, we define linear maps $[\cdot_\Lambda\cdot]\colon V\oplus M\to (V\oplus M)[\Lambda]$ and $\mu\colon V\oplus M\to V\oplus M$ by \begin{align*} [(a+x)_\Lambda (b+y)] \ceq [a_\Lambda b]+a_\Lambda y+x_\Lambda b, \quad \mu((a+x)\otimes(b+y)) \ceq ab + a \cdot y + x \cdot b \end{align*} for $a,b \in V$ and $x,y \in M$, where we used the right action of $V$ defined by \eqref{eq:ract}. These yield a non-unital $N_W=N$ SUSY vertex algebra $(V\oplus M, \nabla, [\cdot_\Lambda\cdot], \mu)$, which is called the trivial structure of non-unital $N_W=N$ SUSY vertex algebra over $V \oplus M$. Then, by \cref{thm:W:VA}, we have the corresponding Maurer-Cartan solution $X \in \MC\bigl(L\bigl(P^{\textup{ch} N_W}_{\Pi^{N+1}(V\oplus M)}\bigr)\bigr)_{\overline{1}}$. Since $X$ is an odd element of $ L^1\bigl(P^{\textup{ch} N_W}_{\Pi^{N+1}(V\oplus M)}\bigr)$, we have \begin{align*} [X, Y] = X \square Y-(-1)^{\widetilde{p}(Y)}Y\square X \in L^n\bigl(P^{\textup{ch} N_W}_{\Pi^{N+1}(V \oplus M)}\bigr) \end{align*} for $Y \in C^n(V,M)$, where $\widetilde{p}$ denotes the parity in $\Pi^{N+1}(V\oplus M)$. A direct calculation shows: \begin{lem} For any $Y\in C^n(V,M)$, we have $\rst{[X,Y]}{(\Pi^{N+1} V)^{\otimes (n+1)} \otimes \mathcal{O}_{n+1}^{\star \mathrm{T}}}\in C^{n+1}(V,M)$. \end{lem} Now we have the main object in this \cref{ss:coh:W}. It is an $N_W=N$ SUSY analogue of \cite[Definition 7.3]{BDHK}. \begin{dfn} Let $V$ be a non-unital $N_W=N$ SUSY vertex algebra, and $M$ be a module over $V$. Define a linear map $\partial^n\colon C^n(V,M) \to C^{n+1}(V,M)$ by \begin{align*} \partial^n Y \ceq \rst{[X,Y]}{(\Pi^{N+1} V)^{\otimes (n+1)} \otimes \mathcal{O}_{n+1}^{\star \mathrm{T}}}. \end{align*} Since $X$ satisfies $X\square X=0$, we have $\partial^{n+1}\circ \partial^n=0$. Thus we get a cochain complex \begin{align}\label{eq:W:comp} C^\bullet(V, M)\ceq(C^n(V,M), \partial^n)_{n\in\mathbb{Z}} \end{align} with convention $C^n(V,M) \ceq 0$ and $\partial^n \ceq 0$ for $n<0$, which is called \emph{the cochain complex of $V$ with coefficients in M}. The $n$-th cohomology is denoted by \begin{align*} H^n_{\textup{ch}}(V,M) \ceq H^n(C^\bullet(V,M)) = \Ker\partial^{n}/\Img\partial^{n-1}. \end{align*} \end{dfn} Similarly as the Lie algebra cohomology $H^n_{\oLie}(L,M)$ recalled in \cref{ss:coh:CE}, and as the non-SUSY case \cite[Theorem 7.6]{BDHK}, the lower degree SUSY vertex algebra cohomology $H^n_{\textup{ch}}(V, M)$, $n=0,1,2$, have explicit description. \begin{thm}\label{thm:W:coh} Let $M$ be a module over a non-unital $N_W=N$ SUSY vertex algebra $V$. \begin{enumerate} \item For the zeroth cohomology, we have a linear isomorphism \[ H_\textup{ch}^0(V, M) \cong \Pi^{N+1}\Cas(V, M), \] where $\Cas(V, M)$ is the linear superspace of Casimir elements (\cref{dfn:Cas}). \item For the first cohomology, we have a linear isomorphism \begin{align*} H_\textup{ch}^1(V, M)\cong \Der(V, M)/\Ind(V, M), \end{align*} where $\Der(V, M)$ is the linear superspace of derivations and $\Ind(V, M)$ is the linear superspace of inner derivations (\cref{dfn:der}). \item The odd part of the second cohomology $H_\textup{ch}^2(V, M)_{\overline{1}}$ is isomorphic to the linear space of isomorphism class of $\mathcal{H}_W$-split extensions of $V$ by module $M$ (\cref{dfn:NWext}). \end{enumerate} \end{thm} The proof will be given in the last part of this \cref{ss:coh:W}. Let us explain the undefined terminology in the statement. We start with: \begin{dfn}\label{dfn:Cas} A \emph{Casimir element of $M$} is an element $\int x\in M/\nabla M$ such that $a_{-\nabla}x=0$ for all $a\in V$. Here we set $\nabla M\ceq \Img(T+S^1+\dotsb+S^N)\subset M$, and denote $\int\colon M\to M/\nabla M$ the canonical projection. We denote by $\Cas(V,M)$ the sub-superspace of $M/\nabla M$ consisting of all Casimir elements of $M$. \end{dfn} Next, we want to introduce derivations on $V$ with values in $M$. We need a few preparations. \begin{lem}\label{lem:intact} Let $M$ be an $\mathcal{H}_W$-supermodule. For a linear map $\rho_\Lambda\colon V\otimes M\to M[\Lambda]$ of parity $\overline{N}$ satisfying \eqref{eq:Lmodsesq} and an even linear map $\rho\colon V\otimes M\to M$, there exists a unique linear map $F\colon V\otimes M\to M[\Lambda]$ of parity $\overline{N}$ such that \begin{align*} &\Res(\lambda^{-1}F(S^ia \otimes x)) = -(-1)^N\Res_\Lambda(\lambda^{-1}\theta^iF(a \otimes x)) \quad (i \in [N]), \\ &\partial_\lambda F(a \otimes x) = a_\Lambda x, \quad \Res(\lambda^{-1}F(a \otimes x)) = a \cdot x \end{align*} for every $a \in V$ and $x \in M$. Here we used the notations \eqref{eq:coh:aLx} and \eqref{eq:coh:adx}. \end{lem} \begin{proof} Define a linear map $F\colon V \otimes M \to M[\Lambda]$ by \begin{align*} F(a \otimes x) \ceq \sum_{I \subset [N]} (-1)^{\#I(N+1)} \sigma(I) \theta^{[N] \mathbin{\setminus} I}(S^Ia)\cdot x + \int_0^\lambda d\lambda (a_\Lambda x). \end{align*} Then, similarly as in the proof of \cref{lem:intbra}, we can show that $F$ is the desired map. \end{proof} The linear map $F$ in \cref{lem:intact} is denoted by \begin{align}\label{eq:coh:intact} \int^{\Lambda}d\Gamma\, (\cdot_\Gamma \cdot)\colon V \otimes M \longrightarrow M[\Lambda], \end{align} and called \emph{the integral of the} (\emph{left}) \emph{$\Lambda$-action $\rho_\Lambda$} (\emph{with respect to $\rho$}). For a module $M$ over non-unital $N_W=N$ SUSY vertex algebra $V$, we define \emph{the integral of the right $\Lambda$-action} \eqref{eq:ract} by \begin{align}\label{eq:intract} \int^{\Lambda}d\Gamma (x_\Gamma a) \ceq (-1)^{p(a)p(x)+\overline{N}} \int^{-\Lambda-\nabla}d\Gamma (a_\Gamma x) \quad (a \in V, \, x \in M). \end{align} \begin{dfn}\label{dfn:der} Let $V$ and $M$ be as above. \begin{enumerate} \item An $\mathcal{H}_W$-supermodule homomorphism $D\colon V \to M$ is called a \emph{derivation of $V$ with values in $M$} if it satisfies \begin{align*} (-1)^{p(D)\overline{N}} D\Bigl(\int^\Lambda d\Gamma[a_\Gamma b]\Bigr) = \int^\Lambda d\Gamma(Da_\Gamma b) + (-1)^{p(D)p(a)} \int^\Lambda d\Gamma (a_\Gamma Db) \end{align*} for any $a,b \in V$. Here we used the integrals of the $\Lambda$-action \eqref{eq:coh:intact} and of the right action \eqref{eq:intract}. We denote by $\Der(V, M)$ the sub-superspace of $\Hom_{\mathcal{H}_W}(V, M)$ consisting of all derivations of $V$ with values in $M$. \item For $x \in M$, the map $V \to M$ given by \begin{align*} a \longmapsto \rst{x_{\Lambda}a}{\Lambda=0}=-(-1)^{p(a)p(x)+\overline{N}}a_{-\nabla} x \end{align*} is a derivation of $V$ with values in $M$. We call it an \emph{inner derivation}, and denote by $\Ind(V,M)$ the linear superspace of inner derivations. \end{enumerate} \end{dfn} To define the notion of extension (\cref{dfn:NWext}), we need to introduce morphisms of non-unital $N_W=N$ SUSY vertex algebras and those of their modules. For non-unital $N_W=N$ SUSY vertex algebras $V$ and $W$, a \emph{morphism $V \to W$} is an even $\mathcal{H}_W$-supermodule homomorphism $\varphi\colon V\to W$ satisfying \begin{align*} \varphi\Bigl(\int^\Lambda d\Gamma[a_\Gamma b]\Bigr) = \int^\Lambda d\Gamma[\varphi(a)_\Gamma \varphi(b)] \end{align*} for any $a,b \in V$. We see that non-unital $N_W=N$ SUSY vertex algebras and their morphisms form an abelian category. An exact sequence therein means a diagram of non-unital $N_W=N$ SUSY vertex algebras \[ \begin{tikzcd} 0 \arrow[r] & U \arrow[r, "\iota"] & V \arrow[r, "\pi"] & W \arrow[r] & 0 \end{tikzcd} \] with $\iota$ injective, $\pi$ surjective, and $\Ker \pi = \Img \iota$. Also, for modules $M$ and $N$ over a non-unital $N_W=N$ SUSY vertex algebra $V$, a \emph{morphism $M \to N$} is an even $\mathcal{H}_W$-supermodule homomorphism $\varphi\colon V\to W$ satisfying \begin{align*} \varphi\Bigl(\int^\Lambda d\Gamma(a_\Gamma x)\Bigr) =\int^\Lambda d\Gamma(a_\Gamma \varphi(x)) \end{align*} for any $a \in V$ and $x \in M$. \begin{dfn}\label{dfn:NWext} Let $V$ and $M$ be as above. \begin{enumerate} \item An \emph{extension $E$ of $V$ by $M$} is a non-unital $N_W=N$ SUSY vertex algebra $E$ satisfying the following conditions. \begin{clist} \item $E$ sits in an exact sequence of non-unital $N_W=N$ SUSY vertex algebras \begin{equation}\label{eq:coh:NWext} \begin{tikzcd} 0 \arrow[r] & M \arrow[r, "\iota"] & E \arrow[r, "\pi"] & V \arrow[r] & 0 \end{tikzcd} \end{equation} Here we regard $M$ as a non-unital $N_W=N$ SUSY vertex algebra with trivial operations. \item $\iota$ is a morphism of modules over the non-unital $N_W=N$ SUSY vertex algebra $E$, i.e., $\iota$ satisfies \begin{align*} \iota\Bigl(\int^\Lambda d\Gamma(\pi(a)_\Gamma x)\Bigr) = \int^\Lambda d\Gamma[a_\Gamma \iota(x)] \end{align*} for any $a \in E$ and $x \in M$. \end{clist} \item An extension of $V$ by module $M$ is called \emph{$\mathcal{H}_W$-split} if the diagram \eqref{eq:coh:NWext} is a split exact sequence of $\mathcal{H}_W$-supermodules. \end{enumerate} \end{dfn} As a preliminary of the proof of \cref{thm:W:coh}, we have: \begin{lem}\label{lem:NWext} For $V$ and $M$ as above, there exists a surjection \begin{align}\label{eq:extsrj} \begin{split} &\left\{Y \in \MC\bigl(L\bigl(\oP_{\Pi^{N+1}(V\oplus M)}^{\textup{ch} N_W}\bigr)\bigr)_{\overline{1}} \;\middle|\; \begin{array}{l} (X-Y)(a\otimes b\otimes z_{1, 2}^{-1})\in M_\nabla[\Lambda_1, \Lambda_2] \\ (X-Y)(a\otimes x\otimes z_{1, 2}^{-1})=0 \\ Y(x\otimes y\otimes z_{1, 2}^{-1})=0 \ (a,b \in V, \, x,y \in M) \end{array} \right\}\\ &\longrightarrow \{\text{isomorphism classes of $\mathcal{H}_W$-split extexsions of $V$ by $M$}\}. \end{split} \end{align} Here an isomorphism of $\mathcal{H}_W$-split extensions of $V$ by $M$ is defined to be an isomorphism of exact sequences of non-unital $N_W=N$ SUSY vertex algebras. \end{lem} \begin{proof} For $Y$ in the left hand side of \eqref{eq:extsrj}, let $E_Y$ denote the $\mathcal{H}_W$-supermodule $V \oplus M$ with the non-unital $N_W=N$ SUSY vertex algebra structure corresponding to $Y$. Then we have an $\mathcal{H}_W$-split extension \[ \begin{tikzcd} 0 \arrow[r] & M \arrow[r, "\iota"] & E_Y \arrow[r, "\pi"] & V \arrow[r] & 0, \end{tikzcd} \] where $\iota\colon M \to V \oplus M$ is the canonical inclusion and $\pi\colon V \oplus M \to V$ is the canonical projection. The map $Y\mapsto E_Y$ gives rise to the surjection \eqref{eq:extsrj}. \end{proof} Now let us start: \begin{proof}[{Proof of \cref{thm:W:coh}}] We denote by $p$ the parity of $V$ or $M$, and by $\widetilde{p}$ the parity of $\Pi^{N+1}V$ or $\Pi^{N+1}M$. \begin{enumerate} \item On the zeroth cohomology, we have \begin{align*} C^0(V, M) = \Hom_\mathbb{K}(\mathbb{K}, \Pi^{N+1}(M/\nabla M)), \quad H_\textup{ch}^0(V, M) = \Ker\bigl(\partial^0\colon C^0(V,M) \to C^1(V,M)\bigr). \end{align*} Define a map $F'_0\colon C^0(V, M) \to M/\nabla M$ by $F'_0(Y) \ceq Y(1_\mathbb{K})$ for $Y\in C^0(V, M)$, and denote by $F_0\colon C^0(V, M)\to \Pi^{N+1}(M/\nabla M)$ the composition of $F'_0$ with parity shift. Then it is clear that $F_0$ is a linear isomorphism. For $Y \in C^0(V,M)$, we can represent $Y(1_\mathbb{K})=\int x$ with $x \in M$. Then we have \begin{align*} (\partial^0Y)(a\otimes 1_\mathbb{K})&=(X\circ_1 Y)(a\otimes 1_\mathbb{K}) =X_{0, \Lambda}(x\otimes a\otimes 1_\mathbb{K})\\ &=(-1)^{\widetilde{p}(a)\widetilde{p}(x)}X_{\Lambda, 0}(a\otimes x\otimes 1_\mathbb{K}) \\ &=(-1)^{p(x)(\overline{N}+\overline{1})}(-1)^{p(a)p(x)+\overline{N}+\overline{1}}a_{-\nabla}x \end{align*} for any $a \in V$. Thus, the restriction of $F_0$ to $H_\textup{ch}^0(V, M)$ gives $H_\textup{ch}^0(V, M)\cong \Pi^{N+1}\Cas(V, M)$. \item Define a map $F_1\colon C^1(V, M)\to \Hom_{\mathcal{H}_W}(V, M)$ by $F_1(Y)\ceq (-1)^{\widetilde{p}(Y)(\overline{N}+\overline{1})}Y_{-\nabla}(\cdot\otimes 1_\mathbb{K})$ for $Y\in C^1(V, M)$, then $F_1$ is a linear isomorphism. For $Y\in C^1(V, M)$ and $a, b\in V$, we have \begin{align*} &(\partial^1Y)(a \otimes b \otimes z_{1,2}^{-1}) \\ &=(X\circ_1Y)(a \otimes b \otimes z_{1,2}^{-1}) +(X\circ_1Y)^{(1,2)}(a \otimes b \otimes z_{1,2}^{-1}) -(-1)^{\widetilde{p}(Y)}(Y\circ_1X)(a \otimes b \otimes z_{1,2}^{-1}) \\ &=X_{\Lambda_1, \Lambda_2}(Y_{\Lambda_1}(a \otimes 1_\mathbb{K})\otimes z_{1,2}^{-1}) +(-1)^{\widetilde{p}(a)\widetilde{p}(b)}X_{\Lambda_2,\Lambda_1}( Y_{\Lambda_2}(b \otimes 1_\mathbb{K})\otimes a \otimes z_{2,1}^{-1}) \\ &\quad -(-1)^{\widetilde{p}(Y)} Y_{\Lambda_1+\Lambda_2}( X_{\Lambda_1,\Lambda_2}(a \otimes b \otimes z_{1,2}^{-1})\otimes 1_\mathbb{K}) \\ &=(-1)^{p(a)(\overline{N}+\overline{1})+\overline{1}} \Bigl( \int^\Lambda d\Gamma (F_1(Y)a_\Gamma b) +(-1)^{p(F_1(Y))p(a)}\int^\Lambda d\Gamma(a_\Gamma F_1(Y)b) \\ &\hspace{9em} -(-1)^{p(F_1(Y))\overline{N}}F_1(Y)\Bigl(\int^\Lambda d\Gamma [a_\Gamma b]\Bigr)\Bigr). \end{align*} Thus, $F_1(Y)$ belongs to $\Der(V,M)$ if and only if $Y\in\Ker \partial^1$. Also, by the proof of (1), we have \begin{align*} F_1(\partial^0 Y)(a) = -(-1)^{p(a)p(x)+\overline{N}}a_{-\nabla}x \quad (a \in V) \end{align*} for $Y\in C^0(V,M)$, where we represented $Y(1_\mathbb{K}) = \int x \in M/\nabla M$. Thus, for $Y \in C^1(V,M)$, $F_1(Y)$ is an inner derivation if and only in $Y\in\Img\partial^0$. Hence $F_1$ induces a linear isomorphism $H_\textup{ch}^1(V,M) \cong \Der(V,M)/\Ind(V,M)$. \item Define a map $F_2'\colon C^2(V,M)_{\overline{1}} \to L^1\bigl(\oP_{\Pi^{N+1}(V\oplus M)}^{\textup{ch} N_W}\bigr)_{\overline{1}}$ by $F_2'(Y) \ceq X+Y$. If $Y \in (\Ker\partial^2)_{\overline{1}}$, then $F_2'(Y)$ satisfies $F_2'(Y) \square F_2'(Y)=0$, so $F_2'$ gives a surjection from $(\Ker \partial^2)_{\overline{1}}$ to \begin{align*} \left\{Y \in \MC\bigl(L\bigl(\oP_{\Pi^{N+1}(V\oplus M)}^{\textup{ch} N_W}\bigr)\bigr)_{\overline{1}} \;\middle|\; \begin{array}{l} (X-Y)(a\otimes b\otimes z_{1, 2}^{-1})\in M_\nabla[\Lambda_1,\Lambda_2] \\ (X-Y)(a\otimes x\otimes z_{1, 2}^{-1})=0\\ Y(x\otimes y\otimes z_{1, 2}^{-1})=0 \ (a,b \in V, \, x,y \in M) \end{array} \right\}. \end{align*} Thus, the composition with the surjection in \cref{lem:NWext} gives a surjection \begin{align*} F_2\colon (\Ker\partial^2)_{\overline{1}} \longrightarrow \{\text{isomorphism classes of $\mathcal{H}_W$-split extexsions of $V$ by module $M$}\}. \end{align*} Hence it is enough to show that $F_2(Y)=F_2(Y')$ is equivalent to $Y-Y'\in (\Img \partial^1)_{\overline{1}}$ for $Y,Y' \in (\Ker\partial^2)_{\overline{1}}$. Note that, for $Y, Y'\in(\Ker\partial^2)_{\overline{1}}$, we have $F_2(Y)=F_2(Y')$ if and only if there exists a morphism of non-unital $N_W=N$ SUSY vertex algebras $f\colon E\to E'$ satisfying \begin{align}\label{eq:exthom} f(a)-a \in M, \quad f(x)=x \quad (a \in V, \, x \in M). \end{align} Here the symbol $E$ (resp.\ $E'$) denotes the $\mathcal{H}_W$-supermodule $V\oplus M$ with the non-unital $N_W=N$ SUSY vertex algebra structure corresponding to $F_2'(Y)$ (resp.\ $F_2'(Y')$). For an even $\mathcal{H}_W$-supermodule homomorphism $f\colon E\to E'$ satisfying \eqref{eq:exthom}, define an even linear map $\Phi_f\colon V\to M$ by $\Phi_f(a)\ceq f(a)-a$ for $a\in V$. Then the mapping $f \mapsto \Phi_f$ gives a bijection \begin{align*} \Phi\colon\{f\in\Hom_{\mathcal{H}_W}(E, E')_{\overline{0}}\ \text{satisfying \eqref{eq:exthom}}\} \xrr{\sim} \Hom_{\mathcal{H}_W}(V, M)_{\overline{0}}. \end{align*} For an even $\mathcal{H}_W$-supermodule homomorphism $f\colon E\to E'$ satisfying \eqref{eq:exthom} and $a, b\in V$, we have \begin{align*} f\Bigl(\int^\Lambda d\Gamma [a_\Gamma b]_E\Bigr) &=\Phi_f\Bigl(\int^\Lambda d\Gamma [a_\Gamma b]_E\Bigr)+\int^\Lambda d\Gamma [a_\Gamma b]_E, \\ \int^\Lambda d\Gamma [f(a)_\Gamma f(b)] &=\int^\Lambda d\Gamma (\Phi_f(a)_\Gamma b)+\int^\Lambda d\Gamma (a_\Gamma \Phi_f(b)) +\int^\Lambda d\Gamma [a_\Gamma b]_{E'}. \end{align*} Thus, $f\colon E \to E'$ is a morphism of non-unital $N_W=N$ SUSY vertex algebras if and only if \begin{align*} \int^\Lambda d\Gamma [a_\Gamma b]_E-\int^\Lambda d\Gamma [a_\Gamma b]_{E'} =\int^\Lambda d\Gamma (\Phi_f(a)_\Gamma b)+\int^\Lambda d\Gamma (a_\Gamma \Phi_f(b)) -\Phi_f\Bigl(\int^\Lambda d\Gamma [a_\Gamma b]_E\Bigr) \end{align*} for any $a, b\in V$, which is equivalent to $Y-Y'=\partial^1(F_1^{-1}(\Phi_f))$. Here $F_1$ is the isomorphism $C^1(V, M) \xr{\sim} \Hom_{\mathcal{H}_W}(V, M)$ defined in the proof of (2). Therefore we find that $F_2(Y)=F_2(Y')$ is equivalent to $Y-Y'\in (\Img\partial^1)_{\overline{1}}$ for $Y,Y'\in(\Ker\partial^2)_{\overline{1}}$. \end{enumerate} \end{proof} \subsection{Cohomology of $N_K=N$ SUSY vertex algebras}\label{ss:coh:K} The theory of $N_W=N$ SUSY vertex algebra cohomology in \cref{ss:coh:W} carries over to the $N_K=N$ setting with minor modifications. So we only give the main definitions and statements. \begin{dfn} Let $(V,\nabla,[\cdot_\Lambda\cdot])$ be an $N_K=N$ SUSY Lie conformal algebra, $(M,\nabla)$ be a left $\mathcal{H}_K$-supermodule and $\rho_\Lambda\colon V\otimes M\to M[\Lambda]$ be a linear map of parity $\overline{N}$. We denote $a_\Lambda x \ceq \rho_\Lambda(a\otimes x)$ for $a\in V$, $x\in M$. A triple $(M,\nabla,\rho_\Lambda)$ is called a \emph{module over $(V,\nabla,[\cdot_\Lambda \cdot])$} if it satisfies the conditions (i) and (ii) in \cref{dfn:W:LCAmod} replacing the $(1|N)_W$-supervariables $\Lambda,\Lambda_1,\Lambda_2$ by the corresponding $(1|N)_K$-supervariables. \end{dfn} \begin{dfn} Let $(V,\nabla,[\cdot_\Lambda\cdot],\mu)$ be a non-unital $N_K=N$ SUSY vertex algebra, $(M,\nabla,\rho_\Lambda)$ be a module over the $N_K=N$ Lie conformal algebra $(V,\nabla,[\cdot_\Lambda\cdot])$ and $\rho\colon V \otimes M \to M$ be an even linear map. We denote $a \cdot x \ceq \rho(a\otimes x)$. A tuple $(M,\nabla,\rho_\Lambda, \rho)$ is called a \emph{module over $(V,\nabla,[\cdot_\Lambda\cdot],\mu)$} if it satisfies the conditions (i)--(iii) in \cref{dfn:W:VAmod} replacing the $(1|N)_W$-supervariables $\Lambda,\Gamma$ by the corresponding $(1|N)_K$-supervariables. \end{dfn} For a module $M$ over a non-unital $N_K=N$ SUSY vertex algebra $V$, one can define the cochain complex of $V$ with coefficients in $M$ as in \eqref{eq:W:comp}. Also, the notions of Casimir elements, derivations and extensions are defined in the similar way. Then we find that \cref{thm:W:coh} is valid for a non-unital $N_K=N$ SUSY vertex algebra $V$ and its module $M$. \subsection{Concluding remarks}\label{ss:cnc} As mentioned in \cref{s:0}, the paper \cite{BDHK} is continued to the development of the cohomology theory of vertex algebras in \cite{BDHK2,BDK20,BDK21,BDKV21}. One of the main topics of these works is the relation between the cohomology theory of vertex algebras and that of vertex Poisson algebras. As far as we understand, the status of the investigation of this relation is still complicated, although several fundamental results are established. Thus, it is natural to consider that the next step of the study of our cohomology theory for SUSY vertex algebras is the investigation of the operad theory and the cohomology theory of SUSY vertex Poisson algebras. We leave it to a future work.

b2a4aff7901b5380b7f4c82cd8412e3b51aa0d4c

\section{Introduction} We are interested in the construction of solutions of the following generalized parabolic Cahn-Hilliard equation \begin{equation}\label{eq1.1} u_t=-\Delta\left[\Delta u -W'(u)\right]+W''(u)\left[\Delta u -W'(u)\right], \qquad \forall\, (t, x)\in \widetilde{\mathbb R}\times \R^n, \end{equation} where $n=2$ or $n\geq4$, $\widetilde{\mathbb R}$ is given by $$ \widetilde{\mathbb R}=\left\{ \begin{array}{rl} (0, \infty), &\quad \mbox{if } n=2, \\[2mm] (-\infty, 0), & \quad \mbox{if } n\geq 4, \end{array} \right. $$ and the functions $W'(s)$ and $W''(s)$ denote the derivatives of first and second orders of $W$ respectively. The potential $W(s)$ is a smooth function, which satisfies the following assumptions \begin{equation}\label{eq1.4} \left\{ \begin{array}{l} W(s)>W(-1)=W(1)\qquad \text{in}\ (-1, 1), \\[1mm]W(s)=W(-s),\qquad \text{for all}\ s\in\R, \\[1mm] W'(-1)=W'(1)=0, \\[1mm] W''(-1)=W''(1)>0. \end{array} \right. \end{equation} The potential $W(u)$ has two non-degenerate local minimum points $u=+1$ and $u=-1$, which are stable equilibria of \eqref{eq1.1}. In particular, the function $W(s)=\frac{1}{4}(1-s^2)^2$ obviously satisfies the above conditions in (\ref{eq1.4}). Up to a scaling, the function $W(s)=\cos(s)$ also satisfies \eqref{eq1.4}. \subsection{Backgrounds} Generally speaking, the Cahn-Hilliard equation means the following equation \begin{equation}\label{eq1.2} \left\{ \begin{array}{l} u_t=-\Delta\left[\Delta u-W'(u)\right], \qquad \forall\, (t, x)\in (-\infty, +\infty)\times\Omega, \\[2mm] u(0, x)=u_0(x), \hspace{2cm} \forall\, x\in\Omega, \end{array} \right. \end{equation} where $\Omega$ denotes a smooth bounded domain in $\R^n$ or the whole space $\R^n$($n\geq1$), which describes phase separation processes of binary alloys in \cite{ch}. Various kinds of problems of this class equation have been extensively studied in recent thirty years. By applying a priori estimate and continuity argument, the existence and asymptotic behaviors of global smooth solutions of Cauchy problem \eqref{eq1.2} have been proved in \cite{lwz} when the initial value $u_0$ is close to stable equilibria $\bar{u}$ ($W(\bar{u})=0$) in the $L^\infty\cap L^1(\R^n)$ space. Using Fourier transform and estimates on the kernel of a linear parabolic operator, a uniform $L^\infty$ bound estimate for solutions of perturbed Cauchy problem \eqref{eq1.2} with additional nonlinear terms, was established by Caffarelli and Muler in \cite{cm}. In a bounded domain, the global well-posedness and long-time behavior of solutions to problem \eqref{eq1.2} with several types of dynamic boundary conditions were studied in \cite{ cgw, cgm, gms, lw, rz}. Based on the works of De Giorgi in \cite{d, de}, numerous authors \cite{bm, rs, t} studied the diffuse approximation of the Willmore functional: \begin{equation}\label{wf} \mathcal{W}(S, \Omega)=\frac{1}{2}\int_{\partial S\cap\Omega}\abs{H_{\partial S}(x)}^2 {\mathrm d}\mu^{n-1}, \end{equation} where $\Omega$ is a given open set in $\R^n$, the set $S\subset\R^n$ with smooth boundary $\partial S\subset \Omega$, and $H_{\partial S}(x)$ is the mean curvature of surface $\partial S$ at point $x\in \partial S$. The approximating functional is defined by \begin{equation} \mathbb{W}_\varepsilon(u)=\left\{ \begin{array}{l} \frac{1}{2\varepsilon}\int_{\Omega}\Big(\varepsilon\Delta u -\frac{W'(u)}{\varepsilon}\Big)^2{\mathrm d}x,\qquad \text{if}\ u\in L^1(\Omega)\cap W^{2, 2}(\Omega), \\[2mm] +\infty,\hspace{3.9cm} \text{otherwise in}\ L^1(\Omega), \end{array} \right. \label{CHEnergy} \end{equation} where the function $W$ satisfies \eqref{eq1.4}. The essential and more challenging work is to rigourously prove that the approximating functional $\mathbb{W}_\varepsilon(u)$ $\Gamma$-converges to the Willmore functional $ \mathcal{W}(S, \Omega)$ as $\varepsilon$ goes to $0$. Bellettini and Paolini in \cite{bp} proved the $\Gamma$-lim sup inequality for smooth Willmore hypersurfaces. However, the $\Gamma$-lim inf inequality is more hard to prove. Up to now, it has been proved in $\R^n$ with $n=2, 3$ in \cite{rs} or $n=2$ in \cite{nt}. This problem is still open in $\R^n$ with $n\geq4$. The relation of the critical points of \eqref{wf} and \eqref{CHEnergy} was exemplified by M. Rizzi \cite{r} and also A. Malchiodi, R. Mandel, M. Rizzi \cite{MalchiodiMandelRizzi}. On the other hand, for the parabolic Cahn-Hilliard equation \eqref{eq1.2}, the gamma convergence results have obtained by Le in \cite{l} under suitable conditions. And see \cite{bb} for the case of \eqref{eq1.2} on the one-dimensional torus. The $L^2$ gradient flow of the approximating energy $\mathbb{W}_\varepsilon (u)$ is equivalent to the evolution equation: \begin{equation}\label{eq1} \partial_tu_\varepsilon=-\Delta\Big[\Delta u_\varepsilon -\frac{1}{\varepsilon^2}W'(u_\varepsilon)\Big] +\frac{1}{\varepsilon^2}W''(u_\varepsilon)\Big[\Delta u_\varepsilon -\frac{1}{\varepsilon^2}W'(u_\varepsilon)\Big] \qquad \text{in}\ (-\infty, +\infty)\times \Omega, \end{equation} which was introduced in \cite{dlw} to describe the deformation of a vesicle membrane under the elastic bending energy, with prescribed bulk volume and surface area. The well-posedness of the phase field model \eqref{eq1} with fixed $\varepsilon$ has been proved in \cite{cl} providing a volume constraint for the average of $u$, or in \cite{cl1} with both volume and area constraints. By applying formal method of matched asymptotic expansions, Loreti and March in \cite{lm} (or Wang in \cite{w}) showed that if $\Gamma(t)\subset \R^n$ with $n=2$ or $3$, is a family of compact closed smooth interfaces and evolves by Willmore flow, it can be approximated by nodal set of the solution $u_\varepsilon$ to the phase field (\ref{eq1}) when $\varepsilon$ goes to $0$. The Willmore flow equation is given by \begin{equation}\label{willmore} V(t)=\Delta_{\Sigma(t)}H-\frac{1}{2}H^3+H\|A\|^2, \end{equation} which is the $L^2$ gradient flow for \eqref{wf} with $\partial S(t)=\Sigma(t)$, where $V(t)$ denotes the outer normal velocity at $x \in \Sigma(t)$, $\Delta_{\Sigma(t)}$ is the Laplace-Beltrami operator on surface $\Sigma(t)$, $H$ and $A$ are the mean curvature and the second fundamental form of $\Sigma(t)$ respectively, and $\|A\|^2$ is the sum of squared coefficients of $A$. Fei and Liu \cite{fl} proved that for given a solution $\Gamma_0(t)$ of \eqref{willmore} in $\R^n(n=2, 3)$, there exists a solution $u_\varepsilon$ of equation \eqref{eq1} with Neumann boundary condition such that its level set convergence to $\Gamma_0(t)$ as the parameter $\varepsilon$ goes to zero. Moreover, a variety of problems for the Willmore flow \eqref{willmore} has been investigated by Kuwert and Sch\"{a}tzle in several papers (for example see \cite{ks, ks1, ks2}). However, the study of connections between equation \eqref{eq1} and the Willmore flow \eqref{willmore} is a challenging work when $n\geq4$. The present paper is one of the first attempts in this direction. \subsection{Main results} In this paper, we want to find solutions of (\ref{eq1.1}) whose values lie at all times in $[-1, 1]$, and approach either $+1$ or $-1$ in the most of the space $\R^n$. This type of solution corresponds to a continuous realization of a material, in which the two states ($u=- 1$ and $u=+1$) coexist. The main difficult point of the study of this type solution of (\ref{eq1.1}), is to derive qualitative information on the interface region(the walls separating the two phases). It is easy to find that $u(t, x)$ is a solution of \eqref{eq1.1} if and only if $u_\eps(t, x):=u(\eps^{-4}t, \eps^{-1} x)$ satisfies equation \eqref{eq1}. Basing on the results in \cite{lm} or \cite{w} with $n=2$ and $3$, we know that the nodal set of $u_\varepsilon(t, x)$ approximates to the solution of Willmore flow \eqref{willmore}. Let us consider the sphere $\Gamma_n(t)$ evolving by the Willmore flow in \eqref{willmore}. Then we have that the radius $\gamma_n(t)$ of $\Gamma_n(t)$ satisfies the equation \begin{equation}\label{willmoreflowradial} \gamma_n'(t)=-\frac{1}{2}\Big(\frac{n-1}{\gamma_n(t)}\Big)^3+\frac{\big(n-1\big)^2}{\big(\gamma_n(t)\big)^3}, \end{equation} which has a solution \begin{equation}\label{sphere eq} \gamma_n(t):=\sqrt[4]{-2(n-3)(n-1)^2t}, \end{equation} where $t\leq0$ when $n\geq 3$ and $t\geq0$ when $n=2$. Due to the self-similarity, there holds that the sphere $\abs{x}=\gamma_n(t)$ is also the transition layer (nodal set) for $u(t, x)$, which is a solution of \eqref{eq1.1}. Our aim is to construct solutions to equation \eqref{eq1.1} with one transition layer closes to the sphere $\abs{x}=\gamma_n(t)$. In which, $t\leq0$ when $n\geq 4$, these solutions are called ancient solutions. And $t\geq0$ for $n=2$, they are long time solutions. We shall mention that the ancient radially symmetric solutions for the parabolic Allen-Cahn equation \begin{equation}\label{Allen} u_t=\Delta u+u-u^3 \qquad \text{in} \ (-\infty, 0]\times\R^n, \end{equation} have been obtained by del Pino and Gkikas in \cite{dG1} for $n=1$ and \cite{dG2} with $n\geq2$. Let us introduce a layer function by considering the following problem of semilinear elliptic equation \begin{equation}\label{eqq} \omega''(y)-W'\big(\omega(y)\big)=0, \quad \omega'(y)>0, \quad y\in\R, \quad \omega(0)=0\quad \text{and} \quad \lim\limits_{y\rightarrow\pm\infty}\omega(y)=\pm 1, \end{equation} which has a unique smooth solution $\omega(y)$ obtained in \cite{ac}, where $W$ satisfies the conditions in (\ref{eq1.4}). In particular if we choose $W'(s)=s^3-s$, then (\ref{eqq}) is the elliptic Allen-Cahn equation, and its solution is written as $$ \omega(y)=\tanh(\frac{y}{\sqrt{2}}). $$ For general functions $W$ satisfying \eqref{eq1.4}, the solution $\omega$ of (\ref{eqq}) has no explicit expression. In \cite{ac}, the inverse function of $\omega$ is given by \begin{equation}\label{eqq2} \lambda(s):=\int^s_{0}\frac{1}{\sqrt{2(W(\tau)-W(1))}}\mathrm{d}\tau, \qquad s\in(-1, 1). \end{equation} Thanks to (\ref{eq1.4}), the function $\lambda(s)$ is well-defined. Taking $\omega$ as a basic layer (the name layer is motivated by the fact that $\omega$ approaches the limits $1$ and $-1$ at $\pm \infty$), we will construct a solution of (\ref{eq1.1}) with a `transition layer' which is symmetric about the sphere $\abs{x}=\gamma_n(t)$ in \eqref{sphere eq}. More precisely, we want to find a solution of equation \eqref{eq1.1}, which has the following asymptotical behavior \begin{equation}\label{eq1.12} u(t, x)\approx \omega\big(\abs{x}-\rho(t)\big), \end{equation} where $\omega$ is given by (\ref{eqq}). The function $\rho(t)$ satisfies \begin{equation}\label{eq1.13} \rho(t)=\gamma_n(t)+h(t), \qquad \text{with}\ h(t)=O\left(\frac{1}{\log\abs{t}}\right), \end{equation} where $h(t)$ is an $C^1$ function with respect of $t$ and the function $\gamma_n(t)$ is defined by \eqref{sphere eq}. In fact, $\rho(t)$ can be chosen by solving the following ODE \begin{equation*} \rho'(t)+\frac{(n-3)(n-1)^2}{2\rho^3(t)}=Q\big(\rho(t), \rho'(t)\big), \qquad \text{with}\ Q\big(\rho(t), \rho'(t)\big)=O\left(\frac{1}{\abs{t}\big[\log\abs{t}\big]^{2(p-1)}}\right), \end{equation*} for all $t\leq0$ when $n\geq4$ or $t\geq0$ when $n=2$, where $p\in(n, n+1]$, see Section \ref{sec:tc}. Our main results can be stated as follows. \begin{thm}\label{thm1} When $n=2$, there exists a radial solution $u(t, \abs{x})$ of equation $(\ref{eq1.1})$ with $t\geq0$, which has the form $u(t, x)\approx \omega\big(\abs{x}-\rho(t)\big)+\phi(t, \abs{x})$, with \begin{equation}\label{1.4} \rho(t)=\sqrt[4]{2t}+h(t). \end{equation} Moreover, $h(t)$ is a $C^1$ function with the decay of order $O(\frac{1}{\log\abs{t}})$ as $t\rightarrow+\infty$ and $\lim\limits_{t\rightarrow +\infty}\phi(t, \abs{x})=0$ uniformly in $x\in \R^2$. \qed \end{thm} \begin{thm}\label{thm4} When $n\geq4$, there exists an ancient solution $u(t, x)$ of equation with $t\leq0$, which has the form $u(t, x)\approx \omega\big(\abs{x}-\rho(t)\big)+\phi(t, \abs{x})$ with \begin{equation}\label{1.7} \rho(t)=\sqrt[4]{-2(n-3)(n-1)^2t}+h(t). \end{equation} Moreover, $h(t)$ is a $C^1$ function with the decay of order $O(\frac{1}{\log\abs{t}})$ as $t\rightarrow-\infty$ and $\lim\limits_{t\rightarrow -\infty}\phi(t, \abs{x})=0$ uniformly in $x\in \R^n$. \qed \end{thm} \begin{rem} In this paper, we mainly prove that the results of Theorem \ref{thm4} hold. For Theorem \ref{thm1}, its proof is similar and we just now notice that $t>0$ in this case $n=2$. When $n=1$ and $3$, it is easily to check that problem \eqref{eq1.1} has a radical solution of the form \begin{equation*} u(t, \abs{x})=\omega(\abs{x}-c), \end{equation*} with any constant $c\in\R$, where $\omega(y)$ is the solution of \eqref{eqq}. \qed \end{rem} \begin{rem} We prove Theorem \ref{thm4} by using Lyapunov-Schmidt reduction method. This method has been widely applied to solving the existence problem of solutions to various kinds of equations \cite{cdm, CDD, dd1, dd2, dG2}. Comparing to the previous works, we are facing three difficulties during the process of dealing with problem \eqref{eq1.1} as the following: \noindent {\textbf{(1)}}. Firstly, the main term $\omega\big(\abs{x}-\rho(t)\big)$ in \eqref{eq1.13} is not a good approximate solution since it does not satisfy the boundary conditions in \eqref{eq2.2} and does not provide enough decay of the error. To modify it, we introduce a cut-off function around the origin and a correction function, see \eqref{dz} in Section \ref{sec:a}. This main idea comes from \cite{r}. \noindent {\textbf{(2)}}. Secondly, equation \eqref{eq1.1} is a parabolic equation of fourth order, which does not satisfy the Maximum Principle since the heat kernel of the biharmonic parabolic operator is sign-changing, see \cite{FG}. To overcome it, we employ the blow-up technique together with the representation of parabolic kernel in Section \ref{sec:lp}. \noindent {\textbf{(3)}}. Lastly, the correlative heat kernel of the linearization operator of the equation \eqref{eq1.1} is more intricate, which leads to hard task to get suitable a priori estimates of linearized problem of equation \eqref{eq1.1}. To solve it, we modify the estimate of the error term $E(t, r)$ in \eqref{Error} with an polynomial decay in Section \ref{sec:a}, and perform more delicate calculations in Section \ref{sec:lp}. \qed \end{rem} The paper is organized as follows. $\clubsuit$ In the first part of Section \ref{sec:a}, we deduce some estimates of decay for $\omega$ in (\ref{eqq}) and its derivatives. After that, an approximate solution, say $z(t,\abs{x})$ with a parameter $\rho(t)$ in \eqref{dz} and \eqref{drho1}, will be defined. By the perturbation of $z(t,\abs{x})+\phi(t,\abs{x})$, the setting-up of a projected form of \eqref{eq1.1} will be derived, see (\ref{eq2.10})-(\ref{eq2.11}) with the Lagrange multiplier $c(t)$. With the introduction of a suitable norm, the estimates of the error will be provided in the last part of Section \ref{sec:a}. $\clubsuit$ Section \ref{sec:lp} is devoting to the collection of some results of linear parabolic equations with a biharmonic operator and then obtain the solvability of a linear projected problem in (\ref{eq3.1}). $\clubsuit$ In Section \ref{sec:nl}, we solve the nonlinear problem (\ref{eq2.10})-(\ref{eq2.11}) by an argument of the fixed-point theorem. $\clubsuit$ In Section \ref{sec:tc}, in order to obtain a radial solution to \eqref{eq1.1} we choose a suitable parameter $h(t)$ (in other words, adjusting the parameter $\rho$ given by \eqref{drho1}) such that $c(t)$ equals to zero, in problem (\ref{eq2.10})-\eqref{eq2.11}. \section{The setting-up: ansatz, the nonlinear projected problem for perturbation term} \label{sec:a} \subsection{Some estimates of the basic layer} Before proving the main theorems, we first derive the decay estimates of the basic layer $\omega$, which is the solution to equation (\ref{eqq}). \begin{lem}\label{lem1}Let $\omega$ be the solution of problem \eqref{eqq}, then we have \begin{equation}\label{eq2.0} \begin{aligned} \left\{ \begin{array}{l} \lim\limits_{y\rightarrow+\infty}\frac{\omega(y)-1+\beta e^{-\sqrt{W''(1)}y}}{\beta^2e^{-2\sqrt{W''(1)}y}}=\frac{W^{(3)}(1)}{6W''(1)}; \\[3mm] \lim\limits_{y\rightarrow-\infty}\frac{\omega(y)+1-\beta e^{\sqrt{W''(-1)}y}}{\beta^2e^{2\sqrt{W''(-1)}y}}=-\frac{W^{(3)}(-1)}{6W''(-1)}; \\[3mm] \lim\limits_{y\rightarrow+\infty}\frac{\omega'(y)}{e^{-\sqrt{W''(1)}y}}=\beta\sqrt{W''(1)} \quad \text{and}\quad \lim\limits_{y\rightarrow-\infty}\frac{\omega'(y)}{e^{\sqrt{W''(-1)}y}}=\beta\sqrt{W''(-1)}, \end{array} \right. \end{aligned} \end{equation} where $W''(s)$ and $W^{(3)}(s)$ denote the derivatives of second and third orders of the function $W(s)$ respectively, and \begin{equation}\label{beta} \beta:=\exp\Bigg\{ \sqrt{W''(1)}\int^1_0 \Bigg[\frac{1}{\sqrt{2\big(W(s)-W(1)\big)}}-\frac{1}{\sqrt{W''(1)}\,(1-s)}\Bigg]\mathrm{d}s\Bigg\}. \end{equation} \end{lem} \begin{proof} Since the function $W$ satisfies the conditions in $(\ref{eq1.4})$, especially $W''(1)=W'(-1)>0$, the above limits in $(\ref{eq2.0})$ are well-defined and $\beta$ is finite. Recall that the inverse function of $\omega$ is the function $\lambda(s)$ given by (\ref{eqq2}). Using L'Hospital's rule, Taylor's formula, the function $W\in C^{3, 1}_{\text{loc}}(\R)$ satisfies (\ref{eq1.4}) and (\ref{eqq2}), we find that $$\lim_{s\rightarrow 1^-}\frac{1}{1-s} \Bigg\{ \lambda(s)+\frac{\ln(1-s)}{\sqrt{W''(1)}}-\int^1_0\frac{1} {\sqrt{2(W(s)-W(1))}}-\frac{1}{\sqrt{W''(1)}(1-s)}\mathrm{d}s\Bigg\} =-\frac{W^{(3)}(1)}{6(W''(1))^{3/2}} $$ and $$ \lim_{s\rightarrow (-1)^+}\frac{1}{1+s} \Bigg\{ \lambda(s)- \frac{\ln(1+s)}{\sqrt{W''(-1)}}-\int^0_{-1}\frac{1} {\sqrt{2(W(s)-W(1))}}-\frac{1}{\sqrt{W''(1)}(1+s)}\mathrm{d}s\Bigg\} =\frac{W^{(3)}(-1)}{6(W''(-1))^{3/2}}. $$ According to the fact that if $x=\lambda(s)$, then $s=\omega(y)$, using Taylor's formula again and the evenness of $W$ in $(-1, 1)$, we easily deduce that the first two limits in (\ref{eq2.1}) hold. So we have $$ \lim\limits_{y\rightarrow+\infty}\frac{1-\omega(y)}{e^{-\sqrt{W''(1)}y}}=\beta\ \ \text{and}\ \ \lim\limits_{y\rightarrow-\infty}\frac{1+\omega(y)}{e^{\sqrt{-W''(1)}y}}=\beta.$$ Using L'Hospital's rule again, the condition (\ref{eq1.4}) and the equation (\ref{eqq}), we have \begin{equation*} \begin{aligned} \lim\limits_{y\rightarrow+\infty}\frac{\omega'(y)}{e^{-\sqrt{W''(1)}y}} = \lim\limits_{y\rightarrow+\infty}\Bigg\{\frac{\omega(y)-1}{-\sqrt{W''(1)}e^{-\sqrt{W''(1)}y}} \times\frac{W'(\omega)-W'(1)}{\omega(y)-1}\Bigg\} =\beta\sqrt{W''(1)} \end{aligned} \end{equation*} and $$\lim\limits_{y\rightarrow-\infty}\frac{\omega'(y)}{e^{\sqrt{W''(-1)}y}} = \lim\limits_{y\rightarrow-\infty}\Bigg\{\frac{\omega(y)+1}{\sqrt{W''(-1)}e^{\sqrt{W''(-1)}y}} \times\frac{W'(\omega)-W'(-1)}{\omega(y)+1}\Bigg\} =\beta\sqrt{W''(-1)}.$$ \end{proof} Next we consider the kernel of a fourth order linear operator. The main result is stated as the following. \begin{lem} Let $W(s)$ be a smooth function satisfying the conditions in \eqref{eq1.4}. Then we have that any solution of the homogeneous problem \begin{equation*} \big[\,\partial_{xx}-W'(\omega(x))\,\big]^2\varphi=0, \quad \abs{\varphi}\leq 1, \quad \abs{\varphi_{xx}}\leq 1 \qquad \text{in}\ \R, \end{equation*} has the form \begin{equation*} \varphi(x)=c\omega'(x), \end{equation*} with some constant $c\in\R$. \end{lem} For the proof of this lemma, see Lemma \ref{lem2} which is a more general result. \qed \subsection{The setting-up of the problem}\label{section2.2} We will prove that Theorem \ref{thm4} holds. Thus, we always assume that $n\geq 4$ and $t<0$ in the rest of the present paper. Let $\widetilde{u}(t, \abs{x})$ be a solution of \eqref{eq1.1}, by a translation $u(t, \abs{x})=\widetilde{u}(t-T, \abs{x})$ with some abuse of notation, then we have that $u(t, r)$ satisfies the following problem \begin{equation}\label{eq2.2} \left\{ \begin{array}{l} u_t=-u_{rrrr}-\frac{2(n-1)}{r}u_{rrr}+\Big(2W''(u)-\frac{(n-1)(n-3)}{r^2}\Big)u_{rr} +\Big(\frac{2(n-1)W''(u)}{r}+\frac{(n-1)(n-3)}{r^3}\Big)u_{r} \\[3mm] \hspace{0.8cm}+W'''(u)u^2_r-W'(u)W''(u), \qquad \ \ \forall\, (t, r)\in (-\infty, -T]\times(0, +\infty), \\[3mm] u_{rrr}(t, 0)=u_r(t, 0)=0,\hspace{2.35cm}\text{for all}\ t\in (-\infty, -T], \end{array} \right. \end{equation} where $r=\abs{x}$ and $T$ is a large positive number whose value can be adjusted at different steps. For convenience, we denote the right hand side of the first equation in (\ref{eq2.2}) by $F(u)$, that is \begin{equation}\label{defF} \begin{aligned} F(u):=&-u_{rrrr}-\frac{2(n-1)}{r}u_{rrr}+\Big(2W''(u)-\frac{(n-1)(n-3)}{r^2}\Big)u_{rr} \\[2mm] &+\Big(\frac{2(n-1)W''(u)}{r}+\frac{(n-1)(n-3)}{r^3}\Big)u_{r}+W'''(u)u^2_r-W'(u)W''(u). \end{aligned} \end{equation} Our purpose is to find a solution of \eqref{eq2.2} with the property \begin{equation*} u(t, r)\approx \omega\big(r-\rho(t)\big), \end{equation*} where $\omega(y)$ is the solution of problem \eqref{eqq}. Firstly, we notice that $\omega\big(r-\rho(t)\big)$ does not satisfies the boundary conditions in \eqref{eq2.2}. A smooth cut-off function $\chi(r)$ can be defined in the form \begin{equation}\label{dcut-off} \chi(r)=0, \qquad \text{for}\ r\leq\frac{\delta_0}{2}\qquad \text{and}\qquad \chi(r)=1, \qquad \text{for}\ r\geq\delta_0, \end{equation} for some small fixed positive number $\delta_0$. We define the first approximate solution of (\ref{eq2.2}) as the following \begin{equation}\label{eq2.3} \widehat{\omega}(t, r)=\omega\big(r-\rho(t)\big)\chi(r)+\chi(r)-1. \end{equation} Here we assume that the function $\rho(t)$ has the from \begin{equation} \rho(t)= \gamma_n(t)+h(t), \label{drho1} \end{equation} where the function $h(t)=O((\log|t|)^{-1})$ as $t\rightarrow-\infty$ and the function $\gamma_n(t)$ is defined in \eqref{willmoreflowradial}-\eqref{sphere eq}, i.e. \begin{equation}\label{dgamma-n} \gamma_n(t):=\sqrt[4]{-2(n-1)^2(n-3)t}, \qquad t<0, \end{equation} which is a radial solution of Willmore flow equation \eqref{willmore} with $n\geq4$. More precisely, we assume that $h(t)$ satisfies the following constraint \begin{equation}\label{assumeh} \sup_{t\leq-1}\abs{h(t)}+\sup_{t\leq-1}\left\{\frac{\abs{t}}{\log\abs{t}}\abs{h'(t)}\right\}\leq 1. \end{equation} Secondly, by using \eqref{eqq}, we have that \begin{equation} \begin{aligned}\label{eq2.4} -\partial_t\widehat{\omega}(t, r)+F(\widehat{\omega}(t, r))&=\rho'(t)\omega'\big(r-\rho(t)\big)-\frac{(n-3)(n-1)}{r^2}\omega''\big(r-\rho(t)\big) \\[2mm] &\quad+\frac{(n-1)(n-3)}{r^3}\omega'\big(r-\rho(t)\big), \qquad \text{for}\ r>\delta_0, \end{aligned} \end{equation} where the operator $F(u)$ is defined by \eqref{defF}. By \eqref{dgamma-n}, we find that the second term in the right hand side of equality \eqref{eq2.4}, that is \begin{equation*} -\frac{(n-1)(n-3)}{r^2}\omega''\big(r-\rho(t)\big), \end{equation*} has a slow decay of order $O\left(\abs{t}^{-\frac{1}{2}}\right)$ as $t$ goes to negative infinity. However, it is not enough to solve equation \eqref{eq2.4} since this term is much bigger than other terms in \eqref{eq2.4}. To cancel it and improve the approximate solution, inspired by \cite{r}, we define a correction function \begin{equation}\label{w1} \widetilde{\omega}(y) :=-\omega'(y)\int_0^y \Bigg[ (\omega'(\hat{y}))^{-2}\int_{-\infty}^{\hat{y}}\frac{s(\omega'(s))^2}{2}\mathrm{d}s \Bigg]{\mathrm d}\hat{y}, \end{equation} Then we have that \begin{equation}\label{eq2.1}\begin{aligned} L^*(\widetilde{\omega}):=\big[-\partial_{yy}+W''(\omega(y))\big]\widetilde{\omega}(y)=\frac{1}{2}y\omega'(y), \qquad (L^*)^2\big[\widetilde{\omega}(y)\big]=-\omega''(y), \qquad \forall\, y\in\R, \end{aligned}\end{equation} and $\widetilde{\omega}(y)$ is an odd function with exponential decay such that \begin{equation}\label{dwtd} \int_{\R}\omega'(y)\widetilde{\omega}(y){\mathrm d}y=0\quad \text{and}\quad \abs{\widetilde{\omega}(y)}\leq Ce^{-\frac{3\alpha}{4}\abs{y}},\qquad \text{for}\ y\in\R. \end{equation} At last, we define an approximate solution of problem \eqref{eq2.2} as the following \begin{equation}\label{dz} z(t, r):=\omega\big(r-\rho(t)\big)\chi(r)+\chi(r)-1+\frac{(n-1)(n-3)}{r^2} \widetilde{\omega}\big(r-\rho(t)\big)\chi(r), \end{equation} where the cut-off function $\chi(r)$ and the function $\rho(t)$ are given by \eqref{dcut-off} and \eqref{drho1}. We will look for a solution of equation (\ref{eq2.2}) of the form \begin{equation}\label{eq2.30000} u(t, r)=z(t, r)+\phi(t, r), \end{equation} where $\phi$ is a small perturbation term. This can be done by using the Lyapunov-Schmidt reduction method in two steps. \noindent {\bf (1).} The first step (see Sections \ref{sec:lp}-\ref{sec:nl}) is solving the following projected version of problem (\ref{eq2.2}) in terms of $\phi(t, r)$: \begin{equation}\label{eq2.10} \begin{aligned} \phi_t=L[\phi]+E(t, r)+N(\phi)-c(t)\partial_r\widehat{\omega}(t, r)\qquad \text{in}\ (-\infty, -T]\times (0, \infty), \end{aligned} \end{equation} and \begin{equation}\label{eq2.11} \int_{0}^\infty \phi(t, r)\omega'\big(r-\rho(t)\big) r^{n-1}{\mathrm d}r=0, \qquad \text{for all}\ t<-T, \end{equation} where the function $\widehat{\omega}(t, r)$ is defined by \eqref{eq2.3}, the error term $E(t, r)$ and nonlinear term $N(\phi)$ are defined respectively by \begin{equation}\label{Error} E(t, r):=F\big(z(t, r)\big)-\frac{\partial z(t, r)}{\partial t} \end{equation} and \begin{equation}\label{nonlinearterm} N(\phi):=F\big(z(t, r)+\phi(t, r)\big)-F\big(z(t, r)\big)-F'\big(z(t, r)\big)[\phi]. \end{equation} In the above, $F(u)$ is defined by \eqref{defF} and the linear operator $L[\phi]:=F'\big(z(t, r)\big)[\phi]$ is defined as follows \begin{equation}\label{eq2.13} \begin{aligned} F'\big(z(t, r)\big)[\phi]:=&-\phi_{rrrr}-\frac{2(n-1)}{r}\phi_{rrr} +\Bigg[2W''\big(z(t, r)\big) -\frac{(n-1)(n-3)}{r^2}\Bigg]\phi_{rr} \\[2mm] &-\big(W''\big(z(t, r)\big)\big)^2\phi +\Bigg[\frac{2(n-1)W''\big(z(t, r)\big)}{r}-\frac{(3-n)(n-1)}{r^3}\Bigg]\phi_{r} \\[2mm] &+2W'''\big(z(t, r)\big)\phi_rz_r-W'''\big(z(t, r)\big)W'\big(z(t, r)\big)\phi+W^{(4)}\big(z(t, r)\big)\abs{z_r}^2\phi \\[2mm] &+2W'''\big(z(t, r)\big)z_{rr}\phi+2\frac{n-1}{r}W'''\big(z(t, r)\big)z_r\phi. \end{aligned} \end{equation} \noindent {\bf (2).} The second step is to choose the function $c(t)$ in such a way that $\phi$ satisfies the orthogonality condition (\ref{eq2.11}), namely the following equality holds: \begin{align} &c(t)\int_{0}^\infty\partial_r\widehat{\omega}(t, r)\omega'\big(r-\rho(t)\big)r^{n-1}{\mathrm d}r \nonumber\\[2mm] &=\int_{0}^\infty \Big[\omega'''\big(r-\rho(t)\big)+\frac{n-1}{r}\omega''\big(r-\rho(t)\big) -W''\big(z(t, r)\big)\omega'\big(r-\rho(t)\big)\Big] \nonumber\\[2mm] &\qquad\qquad\times\left(-\phi_{rr}-\frac{n-1}{r}\phi_r+W''\big(z(t, r)\big)\phi\right) r^{n-1}{\mathrm d}r \nonumber\\[2mm] &\quad +\int_{0}^\infty\left[\partial_{rr}z(t, r) +\frac{n-1}{r}\partial_{r}z(t, r)-W'\big(z(t, r)\big)\right]\phi \omega'\big(r-\rho(t)\big)r^{n-1}{\mathrm d}r \nonumber\\[2mm] &\quad +\int_{0}^\infty \phi(t, r)\partial_t\big[\omega'\big(r-\rho(t)\big)\big]r^{n-1}{\mathrm d}r \ +\ \int_{0}^\infty\big(E(t, r)+N(\phi)\big)\omega'\big(r-\rho(t)\big)r^{n-1}{\mathrm d}r, \label{eq2.14} \end{align} for all $t<-T$. Later on, in Section \ref{sec:tc}, we will choose $h(t)$ such that $c(t)=0$. This means that the function $u$ in \eqref{eq2.30000} will exactly solve \eqref{eq2.2}. \subsection {Estimates of the error terms} We will establish some estimates for the error term $E(t, r)$ in \eqref{Error}. By Taylor's formula, the definitions in \eqref{defF} and \eqref{dz}, we have that \begin{align*} E(t, r)=&F\big(z(t, r)\big)-\frac{\partial z(t, r)}{\partial t} \\[2mm] =&F\big(\widehat{\omega}(t, r)+\widetilde{z}(t, r)\big) -\frac{\partial\big[ \widehat{\omega}(t, r)+\widetilde{z}(t, r)\big]}{\partial t} \\[2mm] =&F\big(\widehat{\omega}(t, r)\big) +F'\big(\widehat{\omega}(t, r)\big)\big[\widetilde{z}(t, r)\big] +F''\big(\widehat{\omega}(t, r)+\theta\widetilde{z}(t, r)\big)\big[\widetilde{z}(t, r), \widetilde{z}(t, r)\big] \\[2mm] &-\frac{\partial \widetilde{z}(t, r)}{\partial t}-\frac{\partial \widehat{\omega}(t, r)}{\partial t} \\[2mm] :=&E_1(t, r)+E_2(t, r), \end{align*} where $\theta\in(0, 1)$. In the above, the operators $F(u)$ and $F'(u)[v]$ are given by \eqref{defF} and \eqref{eq2.13} respectively, and the function $\widehat{\omega}(t, r)$ is given by \eqref{eq2.3} and $\widetilde{z}(t, r)$ is defined by \begin{equation}\label{dzt}\begin{aligned} \widetilde{z}(t, r):=\frac{(n-1)(n-3)}{r^2} \widetilde{\omega}\big(r-\rho(t)\big)\chi(r), \end{aligned}\end{equation} with $\rho(t)$ and $\chi(r)$ given by \eqref{drho1} and \eqref{dcut-off} respectively. The operator $F''(u)[v_1, v_2]$ is defined as the following \begin{equation}\label{dFs}\begin{aligned} F''(u)[v_1, v_2]:=&\Delta\big[W'''(u)v_1v_2\big] \ -\ \left\{W'''(u)W''(u) \ +\ W^{(4)}(u)\big[-\Delta u+W'(u)\big]\right\}v_1v_2 \\[2mm] &\ +\ W'''(u)\Big\{\big[\Delta v_1-W''(u)v_1\big]v_2 \ +\ \big[\Delta v_2-W''(u)v_2\big]v_1\Big\}. \end{aligned} \end{equation} The terms $E_1(t, r)$ and $E_2(t, r)$ have the following explicit forms \begin{equation}\label{de1} E_1(t, r):=F\big(\widehat{\omega}(t, r)\big)-\frac{\partial \widehat{\omega}(t, r)}{\partial t}+\frac{(n-1)(n-3)}{r^2}\partial_{rr}\widehat{\omega}(t, r)+\frac{(n-1)(n-3)^2}{2r^3}\partial_r\widehat{\omega}(t, r), \end{equation} and \begin{align} E_2(t, r):=&-\left[\frac{(n-1)(n-3)}{r^2}\partial_{rr}\widehat{\omega}(t, r)+\frac{(n-1)(n-3)^2}{2r^3}\partial_r\widehat{\omega}(t, r)\right] +F'\big(\widehat{\omega}(t, r)\big)\big[\widetilde{z}(t, r)\big] \nonumber \\[2mm] & +F''\big(\widehat{\omega}(t, r)+\theta\widetilde{z}(t, r)\big)\big[\widetilde{z}(t, r), \widetilde{z}(t, r)\big] -\frac{\partial \widetilde{z}(t, r)}{\partial t}. \label{de2}\end{align} The main result is given by the following lemma. \begin{lem}\label{lem10} Let $\alpha:=\sqrt{W''(1)}$, $p\in(n, n+1]$ and $T>1$, we set \begin{equation}\label{dpsi} \Phi(t, r):=\left\{ \begin{array}{l} \frac{\log\abs{t}}{\abs{t}^{\frac{1}{2}}}\frac{1}{\big(1+\abs{r-\gamma_n(t)-\frac{1}{4\alpha}\log\abs{t}}\big)^{p}}, \qquad \text{if} \ r\in\left[\delta_0, +\infty\right); \\[4mm] \frac{\log\abs{t}}{\abs{t}^{\frac{1}{2}}}\overline{\chi}_{\left\{\frac{\delta_0}{2}\leq r<\delta_0\right\}}, \hspace{2.33cm}\text{if}\ r\in\left(0, \delta_0\right); \end{array} \right. \end{equation} where the function $\gamma_n(t)$ is defined by \eqref{dgamma-n} and $\delta_0$ is a small positive number given in \eqref{dcut-off}. Here $\overline{\chi}_{A}$ is the characteristic function of the set $A$. Then there exists constant $C>0$ which depends only on $\delta_0$, $\alpha$, and $n$, such that $$|E(t, r)|\leq C\frac{\Phi(t, r)}{\log\abs{t}}, $$ for all $(t, r)\in (-\infty, -T]\times(0, +\infty)$. \end{lem} \begin{proof} We first estimate the term $E_1(t, r)$ in \eqref{de1}. By the definitions of $F(u)$ and $\widehat{\omega}(t, r)$ in \eqref{defF} and \eqref{eq2.3}, lemma \ref{lem1}, we have that \begin{equation*}\begin{aligned} E_1(t, r)=&\left[\rho'(t)+\frac{(n-1)^2(n-3)}{2r^3}\right]\omega'\big(r-\rho(t)\big)\overline{\chi}_{\left\{ r\geq\delta_0\right\}}+O\left(\frac{1}{[\gamma_n(t)]^2}\right)\overline{\chi}_{\left\{ \frac{\delta_0}{2}<r<\delta_0\right\}}, \end{aligned}\end{equation*} where $\overline{\chi}_{A}$ is the characteristic function of the set $A$. Furthermore, using the assumption of $\rho(t)$ in \eqref{drho1}-\eqref{assumeh}, the definition of $\gamma_n(t)$ in \eqref{dgamma-n} and Lemma \ref{lem1}, we have that \begin{align} \nonumber \abs{E_1(t, r)} \leq&C\Bigg[\frac{1}{\abs{t}^{\frac{3}{4}}}\overline{\chi}_{\left\{ r\geq\delta_0\right\}}+\frac{1}{r^3}\overline{\chi}_{\left\{ r\geq\frac{\gamma_{n}(t)}{2}\right\}}+\frac{1}{r^3}\overline{\chi}_{\left\{ \frac{\gamma_{n}(t)}{2}\geq r\geq\delta_0\right\}}\Bigg]e^{-\alpha\abs{r-\gamma_n(t)}} \,+\, \frac{C}{\abs{t}^{\frac{1}{2}}}\overline{\chi}_{\left\{ \frac{\delta_0}{2}<r<\delta_0\right\}} \nonumber \\[3mm] \leq &C\frac{e^{-\frac{3\alpha}{4}\abs{r-\gamma_n(t)-\frac{1}{3\alpha}\log\abs{t}}}}{\abs{t}^{\frac{1}{2}}}\overline{\chi}_{\left\{ r\geq\delta_0\right\}} +\frac{C}{r^3}e^{-\frac{3\alpha}{4}\abs{r-\gamma_n(t)}}\overline{\chi}_{\left\{ \frac{\gamma_{n}(t)}{2}\geq r\geq\delta_0\right\}} +\frac{C}{\abs{t}^{\frac{1}{2}}}\overline{\chi}_{\left\{ \frac{\delta_0}{2}<r<\delta_0\right\}} \nonumber \\[3mm] \leq& C\frac{\Phi(t, r)}{\log\abs{t}},\qquad \text{ for all}\ (t, r)\in (-\infty, -T]\times(0, +\infty), \label{eE1} \end{align} where $C$ is a positive constant only depending on $\delta_0$, $\alpha$ and $n$. Here we used the fact that there exits a positive constant $C>0$ only depending on $\delta_0$, $\alpha$ and $n$ such that \begin{equation}\label{dafact} \frac{1}{r^3}e^{-\frac{3\alpha}{4}\abs{r-\gamma_n(t)}}\overline{\chi}_{\left\{ \frac{\gamma_{n}(t)}{2}\geq r\geq\delta_0\right\}}\leq Ce^{-\frac{\alpha}{4}\gamma_n(t)}\quad \text{and} \quad e^{-\frac{3\alpha}{4}\abs{x}}\leq \frac{C}{\big(1+\abs{x}\big)^{p}}, \end{equation} for all $ x\in\R$ and $p\in(n, n+1]$, where $C>0$ does not depend on $x$ and $t$. Next we consider the term $E_2(t, r)$ in \eqref{de2}. Using the definitions of linear operators $F'(u)[v]$ and $F''(u)[v, v]$ in \eqref{eq2.13} and \eqref{dFs}, Lemma \ref{lem1} and estimate in \eqref{dwtd}, we derive that \begin{equation*} \begin{aligned} E_2(t, r)=&\Bigg\{F'\big(\omega\big(r-\rho(t)\big)\big)\big[\widetilde{z}(t, r)\big]-\frac{(n-1)(n-3)}{r^2}\omega''\big(r-\rho(t)\big) +\frac{(n-1)(n-3)^2}{2r^3}\omega'\big(r-\rho(t)\big) \\[2mm] & +F''\big(\omega\big(r-\rho(t)\big)+\theta\widetilde{z}(t, r)\big)\big[\widetilde{z}(t, r), \widetilde{z}(t, r)\big] +\rho'(t)\omega'\big(r-\rho(t)\big)\frac{(n-1)(n-3)}{r^2}\Bigg\}\overline{\chi}_{\left\{r\geq\delta_0\right\}}\\&+O\left(\frac{1}{[\gamma_n(t)]^2}\right)\overline{\chi}_{\left\{ \frac{\delta_0}{2}<r<\delta_0\right\}} \\[2mm] =&\Bigg\{\left[-\Big(\partial_{rr}-W''\big(\omega\big(r-\rho(t)\big)\big)\Big)^2\widetilde{\omega}\big(r-\rho(t)\big)-\omega''\big(r-\rho(t)\big)\right]\frac{(n-1)(n-3)}{r^2} \\[2mm] & +O\left(\frac{1}{r^3}+\frac{\rho'(t)}{r^2}\right)e^{-\frac{3\alpha}{4}\abs{r- \rho(t)}} \Bigg\}\overline{\chi}_{\left\{r\geq\delta_0\right\}} \ +\ O\left(\frac{1}{[\gamma_n(t)]^2}\right)\overline{\chi}_{\left\{ \frac{\delta_0}{2}<r<\delta_0\right\}} \\[2mm] =&O\left(\frac{1}{r^3}+\frac{\rho'(t)}{r^2}\right)e^{-\frac{3\alpha}{4}\abs{r- \rho(t)}}\overline{\chi}_{\left\{r\geq\delta_0\right\}} \ +\ O\left(\frac{1}{[\gamma_n(t)]^2}\right)\overline{\chi}_{\left\{ \frac{\delta_0}{2}<r<\delta_0\right\}}, \end{aligned} \end{equation*} where we used the equalities in \eqref{eq2.1} and the fact that the function $\widetilde{\omega}(t, r)$ in \eqref{w1} and its derivatives are all exponentially decaying. By the same argument in \eqref{eE1} and the equality in \eqref{dafact}, we can get that \begin{equation*} |E_2(t, r)|\leq C\frac{\Phi(t, r)}{\log\abs{t}},\qquad \text{ for all}\ (t, r)\in (-\infty, -T]\times(0, +\infty), \end{equation*} where $C$ is a positive constant only depending on $\alpha$ and $n$. Eventually, combining the above estimates of the terms $E_1(t,r)$ and $E_2(t,r)$, we can obtain the desired results. \end{proof} \begin{rem} According to the above proof of Lemma \ref{lem10}, it is easy to find that the error term $E(t, r)$ has an exponentially decaying in space variable \begin{equation}\label{deee} \abs{E(t, r)}\leq \frac{C}{\abs{t}^{\frac{1}{2}}e^{\frac{\alpha}{2}\abs{r-\gamma_n(t)-\frac{1}{4\alpha}\log\abs{t}}}}, \end{equation} for all $r\geq\delta_0$, where $C>0$ does only depend on $n$ and $\alpha$. Our goal is to solve nonlinear problem \eqref{eq2.10}-\eqref{eq2.11}. Hence, according to \eqref{deee}, we may consider the following linear parabolic problem of fourth order \begin{equation}\label{deeee} -\partial_t\varphi+F'(z(t, r))[\varphi]=f(t, r), \qquad (t, r)\in(-\infty, -T)\times(0, +\infty), \end{equation} with the function $f(t, r)$ satisfying \begin{equation*} \sup_{(t, r)\in(-\infty, -T)\times(0, +\infty)} \frac{\abs{f(t, r)}}{\abs{t}^{\frac{1}{2}}e^{\frac{\alpha}{2}\abs{r-\gamma_n(t)-\frac{1}{4\alpha}\log\abs{t}}}}<+\infty, \end{equation*} where the linear operator $F'(z(t, r))$ is defined in \eqref{eq2.13}. However, the heat kernel of linear parabolic operator $-\partial_t+F'(z(t, r))$ would not has the exponential decay in \eqref{deee}. As far as we known, it can satisfy the polynomial decay. Hence, we provide an estimate of the error term $E(t,r)$ with the polynomial decay in the above lemma. \qed \end{rem} \section{The Linear Problem}\label{sec:lp} In this section, firstly, we will obtain the solvability of a class of semilinear biharmonic parabolic equations by applying some properties of biharmonic heat kernel and fixed-point arguments, which is given in Proposition \ref{prop8}. Secondly, we will prove that the linear projected problem \eqref{eq3.1} is solvable by using Proposition \ref{prop8} and a priori estimate in Lemma \ref{lem5}. The main result is given by Proposition \ref{prop2}. \subsection{A few results of linear parabolic equations with a biharmonic operator} We first collect some known results for homogeneous biharmonic parabolic equation, from \cite{ FG, GG1, G}. Its solution can be represented by the convolution of a biharmonic heat kernel and initial function. \begin{prop}\label{prop6}We consider the following Cauchy problem for the biharmonic heat equation: \begin{equation}\label{heateq} \left\{ \begin{array}{l} u_t+(-\Delta)^2u=0\qquad \text{in}\ \R^{n+1}_+:=(0, +\infty)\times\R^n, \\[2mm] u(0, x)=u_0(x)\hspace{0.3cm} \qquad \text{in}\ \R^n, \end{array} \right. \end{equation} where $n\geq 1$ and $u_0\in C^1(\R^n)\cap L^\infty(\R^n)$. Then \eqref{heateq} admits a unique global in time solution explicitly given by $$ u(t, x)=\int_{\R^n}u_0(y)p_n(t, x-y){\mathrm d}y, \qquad p_n(t, x):=\bar{\alpha}_nt^{-n/4}f_n\Big(\frac{\abs{x}}{t^{1/4}}\Big), \qquad \forall\,(t, x)\in \R^{n+1}_+. $$ Here $p_n(t, x)$ is called the biharmonic heat kernel and $\bar{\alpha}_n$ denotes a suitable positive normalization number which depends on $n$ and satisfies $$ \bar{\alpha}_nt^{-n/4}\int_{\R^n}f_n\Big(\frac{\abs{y}}{t^{1/4}}\Big){\mathrm d}y=1, \qquad \text{for all}\ t>0. $$ Moreover, the function $f_n$ is given by $$ f_n(s):=s^{1-n}\int_{0}^{+\infty}e^{-\varrho^4}(s\varrho)^{\frac{n}{2}}J_{\frac{n-2}{2}}(s\varrho) {\mathrm d}\varrho, \qquad \text{for}\ s>0, $$ where $J_\nu$ denotes the $\nu$-th Bessel function of the first kind. There exist two constants $K_n$ and $\mu_n$ depending on $n$ such that \begin{equation}\label{esf} \abs{f_n(s)}\leq K_n\exp(-\mu_n s^{4/3}), \qquad \text{for all}\ s\geq 0. \end{equation} For the derivative of $f_n$, the following formula holds \begin{equation}\label{fdf} f'_n(s)=-sf_{n+2}(s). \end{equation} \qed \end{prop} Next we will use the above results to study the following inhomogeneous problem: \begin{equation}\label{noneq} \left\{ \begin{array}{l} u_t+(-\Delta)^2u=f(t, x)\qquad\text{in}\ (t_0, t_1)\times\R^{n}, \\[2mm] u(t_0, x)=u_0(x)\hspace{1.7cm} \text{in}\ \R^n, \end{array} \right. \end{equation} where $f(t, x)\in L^\infty((t_0, t_1)\times\R^{n})$ and $u_0\in C^1(\R^n)\cap L^\infty(\R^n)$. \begin{defn}\label{demild} Assume that $t_1\in (t_0, +\infty)$. We say that $u\in L^\infty((t_0, t_1)\times\R^n)$ is a mild solution of \eqref{noneq} if \begin{equation}\label{demil} u(t, x)=\Gamma_{t-t_0}[u_0](x)+\int^t_{t_0}\Gamma_{t-\tau}[f(\tau, \cdot)](x){\mathrm d}\tau,\qquad\text{for}\ (t, x)\in (t_0, t_1)\times\R^n. \end{equation} Here $\Gamma_{t}$ is a linear operator defined by $$ \Gamma_{t}[u](x):=\Big(p_n(t, \cdot)\ast u\Big)(x)=\int_{\R^n}p_n(t, x-y)u(y){\mathrm d}y, $$ where the biharmonic heat kernel $p_n(t, x)$ is given by Proposition \ref{prop6}. In addition, when $t_1=+\infty$, a mild solution also can be defined by \eqref{demil} provided $u(t, x)\in L^\infty_{loc}(t_0, +\infty)\times L^\infty(\R^n)$. \end{defn} Notice that it is not hard to verify that $u$ is a mild solution of \eqref{noneq} if and only if $u$ solves it in the pointwise sense. We give some regularity estimates for mild solutions. \begin{prop}\label{prop7}Let $u$ be a mild solution of problem \eqref{noneq}. Assume that $t_1<+\infty$. Then the following estimates are valid: \begin{itemize} \item If $u_0\in L^{\infty}(\R^n)\cap C^1(\R^n)$, then for every $t_*\in(t_0, t_1)$, $t_*>0$, $\vartheta\in (0, 4)$ and $\theta\in (0, 1)$, it holds that \begin{equation}\label{es1} \begin{aligned} \sup_{t\in(t_*, t_1)}\|u(t, \cdot)\|_{C^{\vartheta}(\R^n)} +\sup_{x\in\R^n}\|u(\cdot, x)\|_{C^{\theta}(t_*, t_1)} \leq C_1\Big(\|f\|_{L^{\infty}((t_*, t_1)\times\R^n)}+\|u_0\|_{L^\infty(\R^n)}\Big), \end{aligned} \end{equation} for some positive constant $C_1$ depending on $n, \theta, t_1-t_0, t_1-t_*, \vartheta$ and $t_*$. \item If $u_0\in C^\vartheta(\R^n)$ for some $\vartheta\in (0, 4)$ and $t_0>0$, then it holds \begin{equation}\label{es2} \begin{aligned} \sup_{t\in(t_0, t_1)}\|u(t, \cdot)\|_{C^{\vartheta}(\R^n)}+\sup_{x\in\R^n}\|u(\cdot, x)\|_{C^{\frac{\vartheta}{4}}(t_0, t_1)} \leq C_{2}\Big(\|f\|_{L^{\infty}((t_0, t_1)\times\R^n)}+\|u_0\|_{C^\vartheta(\R^n)}\Big), \end{aligned} \end{equation} for some positive constant $C_2$ depending on $t_1-t_0, n$ and $\vartheta$. \end{itemize} \end{prop} \begin{proof} According to the formula \eqref{fdf} and some direct computations, we can derive some estimates for the derivatives of biharmonic heat kernel $p_n(t, x)$ as follows: \begin{itemize}\label{1} \item $\abs{\partial_tp_n(t, x)}\leq C_n\Big(t^{-1}\abs{p_n(t, x)}+t^{-1}\abs{x}^2\abs{p_{n+2}(t, x)}\Big)$; \medskip \item $\abs{\nabla_x p_n(t, x)}\leq C_n\abs{x}\abs{p_{n+2}(t, x)}$; \medskip \item $\abs{\nabla^2_x p_n(t, x)}\leq C_n\Big(\abs{p_{n+2}(t, x)}+\abs{x}^2\abs{p_{n+4}(t, x)}\Big)$; \medskip \item $\abs{\nabla^3_x p_n(t, x)}\leq C_n\Big(\abs{x}\abs{p_{n+4}(t, x)}+\abs{x}^3\abs{p_{n+6}(t, x)}\Big)$; \medskip \item $\abs{\nabla^4_x p_n(t, x)}\leq C_n\Big(\abs{p_{n+4}(t, x)}+\abs{x}^2\abs{p_{n+6}(t, x)}+\abs{x}^4\abs{p_{n+8}(t, x)}\Big)$, \end{itemize} where $C_n>0$ only depends on $n$. We decompose the mild solution in (\ref{demil}) as follows \begin{equation*} u(t, x)=U_1(t, x)+U_2(t, x), \end{equation*} where $$ U_1(t, x)=\Gamma_{t-t_0}[u_0](x), \quad U_2(t, x)=\int^t_{t_0}\Gamma_{t-\tau}[f(\tau, \cdot)](x){\mathrm d}\tau. $$ The analysis will begin with the first term $U_1(t, x)$. For any $t>t_0$ and $x\in\R^n$, using the above estimates and \eqref{esf}, we derive that \begin{align*} \abs{\partial_tU_1(t, x)}=\abs{\int_{\R^n}\partial_tp_n(t-t_0, x-y)u_0(y){\mathrm d}y} \leq \frac{C_n}{t-t_0}\|u_0\|_{L^\infty(\R^n)}. \end{align*} Moreover, using (\ref{esf}), we have that $$ \abs{U_1(t, x)}\leq\int_{\R^n}\abs{p_n(t-t_0, x-y)u_0(y)}{\mathrm d}y\leq C_n\|u_0\|_{L^\infty(\R^n)}. $$ For any $\vartheta\in(0, 4)$ and $t, t^1\in(t_0, t_1)$, using the fact $p_n(t, x)=t^{-n/4}p_n(\frac{x}{t^{1/4}}, 1)$, \begin{equation}\label{esq}\begin{aligned} \frac{\abs{U_1(t, x)-U_1(t^1, x)}}{\abs{t-t^1}^{\frac{\vartheta}{4}}}&\leq\abs{\int_{\R^n} \frac{u_0\big((t-t_0)^{1/4}y\big) \,-\, u_0\big((t^1-t_0)^{1/4}y\big)} {\abs{t-t^1}^{\frac{\vartheta}{4}}}p_n(1, y){\mathrm d}y} \\[2mm] &\leq\|u_0\|_{C^\vartheta(\R^n)}\int_{\R^n}\abs{p_n(1, y)}\abs{y}^\vartheta {\mathrm d}y\leq C_n\|u_0\|_{C^\vartheta(\R^n)}. \end{aligned}\end{equation} Here we have used the inequality $a^{\vartheta/4}-b^{\vartheta/4}\leq (a-b)^{\vartheta/4}$ for any $a\geq b\geq0$ and $\vartheta\in(0, 4)$ in the second inequality, $C_n>0$ only depends on $n$. Moreover, for the derivatives of $U_1$ with respect to $x$, by (\ref{esf}), we have that \begin{align*} \abs{\nabla_xU_1(t, x)}&=\abs{\int_{\R^n}\nabla_x p_n(t-t_0, x-y)u_0(y){\mathrm d}y} \\[2mm] &\leq C_n\|u_0\|_{L^\infty(\R^n)}\int_{\R^n}\abs{y}\abs{p_{n+2}(y, t-t_0)}{\mathrm d}y \\[2mm] &\leq \frac{C_n}{(t-t_0)^{1/4}}\|u_0\|_{L^\infty(\R^n)}, \end{align*} where $C_n$ does only depend on $n$. Similarly, we can get that \begin{align*} \abs{\nabla^2_xU_1(t, x)}\leq&\frac{C_n}{(t-t_0)^{1/2}}\|u_0\|_{L^\infty(\R^n)}, \qquad \abs{\nabla^3_xU_1(t, x)}\leq \frac{C_n}{(t-t_0)^{3/4}}\|u_0\|_{L^\infty(\R^n)}, \end{align*} and $$\abs{\nabla^4_xU_1(t, x)}\leq \frac{C_n}{(t-t_0)}\|u_0\|_{L^\infty(\R^n)}.$$ Thus for any $x, y\in\R^n$ and $\theta\in(0, 1)$, we have $$[U_1]_\theta(t):=\frac{\abs{U_1(t, x)-U_1(t, y)}}{\abs{x-y}^\theta}\leq\left\{ \begin{array}{l} 2\|U_1(t, \cdot)\|_{L^\infty(\R^n)},\ \quad \ \quad \text{if}\ \abs{x-y}\geq1, \\[2mm] \|\nabla _x U_1(t, \cdot)\|_{L^\infty(\R^n)},\qquad \text{if}\ \abs{x-y}\leq1. \end{array} \right.$$ Hence, for any $t_*>t_0$, we have $$\sup_{t\in(t_*, t_1)}[U_1]_\theta(t)\leq \frac{C_n}{(t_*-t_0)^{1/4}}\|u_0\|_{L^\infty(\R^n)}.$$ By similar arguments, we can obtain the following estimates $$\sup_{t\in(t_*, t_1)}[\nabla U_1]_\theta(t)\leq \frac{C_n}{(t_*-t_0)^{1/2}}\|u_0\|_{L^\infty(\R^n)}, \qquad \sup_{t\in(t_*, t_1)}[\nabla^2 U_1]_\theta(t)\leq \frac{C_n}{(t_*-t_0)^{3/4}}\|u_0\|_{L^\infty(\R^n)}, $$ and $$ \sup_{t\in(t_*, t_1)}[\nabla^3 U_1]_\theta(t)\leq \frac{C_n}{(t_*-t_0)}\|u_0\|_{L^\infty(\R^n)}. $$ Combining the above estimates, we have that for any $\vartheta\in(0, 4)$, $t_*>t_0$ and $\theta\in(0, 1)$, \begin{equation}\label{esU1} \sup_{t\in(t_*, t_1)}\|U_1(t, \cdot)\|_{C^{\vartheta}(\R^n)}+\sup_{x\in\R^n}\|U_1(\cdot, x)\|_{C^{\theta}(t_*, t_1)} \leq C_{n}\big[1+\varrho(t_*-t_0)\big]\|u_0\|_{L^\infty(\R^n)} \end{equation} where $\varrho(t_*-t_0)=\frac{1}{(t_*-t_0)}+\frac{1}{(t_*-t_0)^{1/4}}+\frac{1}{(t_*-t_0)^\theta}$. Using the formula $$ U_1(t, x)=\int_{\R^n}p_n(t-t_0, y)u_0(x-y){\mathrm d}y $$ and \eqref{esq}, for $\vartheta\in (0, 4)$, we have \begin{equation}\label{esU2} \sup_{t\in(t_0, t_1)}\|U_1(t, \cdot)\|_{C^{\vartheta}(\R^n)}+\sup_{x\in\R^n}\|U_1(\cdot, x)\|_{C^{\frac{\vartheta}{4}}(t_0, t_1)} \leq C_{n}\|u_0\|_{C^\vartheta(\R^n)}, \end{equation} where $C_{n}$ is a positive constant which depends on $n, \vartheta$. Next we will estimate $U_2(t, x)$. First we find that \begin{align*} \abs{U_2(t, x)}\leq& C_n\int_{t_0}^t\int_{\R^n}\abs{p_n(t-\tau, x-y)f(\tau, y)}{\mathrm d}y{\mathrm d}\tau \\[2mm] \leq& C_n(t_1-t_0)\|f\|_{L^{\infty}}((t_1, t)\times\R^n). \end{align*} For any $t, t^1\in(t_0, t_1)$, by the previous estimate of $\partial_tp_n(t, x)$, we have $$ \abs{p_n(t, x)-p_n(t^1, x)}\leq C_n\abs{\log t- \log t^1}\Big(\abs{p_n\big((1-\theta_1)t+\theta_1 t^1, x\big)}+\abs{x}^2\abs{p_{n+2}\big((1-\theta_1)t+\theta_1 t^1, x\big)}\Big), $$ for some $\theta_1\in(0, 1)$. Using the above inequality and \eqref{esf}, for $\theta\in (0, 1)$, we have \begin{align*} \frac{\abs{U_2(t, x)-U_2(t^1, x)}}{\abs{t-t^1}^\theta}\leq& \frac{\abs{\int_{t_0}^t\int_{\R^n}\Big[p_n(t-\tau, x-y)-p_n(t^1-\tau, x-y)\Big]f(y, \tau){\mathrm d}y{\mathrm d}\tau}}{\abs{t-t^1}^\theta} \\[2mm] &+\frac{\abs{\int^t_{t^1}\int_{\R^n}p_n(t^1-\tau, x-y)f(\tau, y){\mathrm d}y{\mathrm d}\tau}}{\abs{t-t^1}^\theta} \\[2mm] \leq &C_n\Bigg(1+\log\frac{t_1-t_0}{\abs{t-t^1}}\Bigg)\abs{t-t^1}^{1-\theta}\|f\|_{L^{\infty}((t_1, t)\times\R^n)}, \end{align*} where $C_n$ is a positive constant which only depends on $n$. Second, we study the derivatives of $U_2(t, x)$ with respect to $x$. By some direct computations and (\ref{esf}), we have \begin{align*} \abs{\nabla_xU_2(t, x)}&=\abs{\int_{t_0}^{t}\int_{\R^n}\nabla_x p_n(t-\tau, x-y)f(\tau, y){\mathrm d}y{\mathrm d}\tau} \\[2mm] &\leq C_n \|f\|_{L^{\infty}((t_0, t)\times\R^n)}\int_{t_0}^t\int_{\R^n}\abs{y}\abs{p_{n+2}(t-\tau, y)}{\mathrm d}y{\mathrm d}\tau \\[2mm] &\leq C_n \|f\|_{L^{\infty}((t_0, t)\times\R^n)}\int_{t_0}^t(t-\tau)^{-1/4}{\mathrm d}\tau \\[2mm] &\leq C_n (t-t_0)^{3/4} \|f\|_{L^{\infty}((t_0, t)\times\R^n)}. \end{align*} By the same arguments as above, we can derive that $$ \abs{\nabla^2_xU_2(t, x)}\leq C_n (t-t_0)^{1/2} \|f\|_{L^{\infty}((t_0, t)\times\R^n)}, \qquad \abs{\nabla^3_xU_2(t, x)}\leq C_n (t-t_0)^{1/4} \|f\|_{L^{\infty}((t_0, t)\times\R^n)}, $$ and for any $\theta\in(0, 1)$, \begin{align*} &\frac{\abs{\nabla^3_xU_2(t, x_1)-\nabla^3_xU_2(t, x_2)}}{\abs{x_1-x_2}^\theta} \\[2mm] & =\alpha_n\frac{\abs{\int_{t_0}^t\int_{\R^n}(t-\tau)^{-\frac{n+3}{4}}\Big[\nabla^3_xf_n\Big(\frac{\abs{x_1-y}}{(t-\tau)^{1/4}}\Big)- \nabla^3_xf_n\Big(\frac{\abs{x_2-y}}{(t-\tau)^{1/4}}\Big)\Big]f(y, \tau){\mathrm d}y{\mathrm d}\tau}}{\abs{x_1-x_2}^\theta} \\[2mm] &\leq C_n \frac{\abs{\int_{t_0}^t\int_{\R^n}(t-\tau)^{-\frac{3}{4}}\Big[\nabla^3_xf_n\Big(\abs{y-\frac{x_1}{(t-\tau)^{1/4}}}\Big)- \nabla^3_xf_n\Big(\abs{y-\frac{x_2}{(t-\tau)^{1/4}}}\Big)\Big]f\Big((t-\tau)^{1/4}y, \tau\Big){\mathrm d}y{\mathrm d}\tau}}{\abs{x_1-x_2}^\theta} \\[2mm] &\leq C_n(t-t_0)^{(1-\theta)/4}\|f\|_{L^{\infty}((t_0, t)\times\R^n)}. \end{align*} Combing the above estimates of $U_2$, for any $\vartheta\in(0, 4)$ and $\theta\in(0, 1)$, we have that \begin{equation}\label{esU21} \sup_{t\in(t_0, t_1)}\|U_2(t, \cdot)\|_{C^{\vartheta}(\R^n)}+\sup_{x\in\R^n}\|U_2(\cdot, x)\|_{C^{\theta}(t_0, t_1)} \leq C_{n}\Big[\upsilon(t_1-t_0)+(t_1-t_0)^{1-\theta}\Big]\|f\|_{L^{\infty}((t_0, t)\times\R^n)}. \end{equation} On the other hand, for $\vartheta\in(0, 4)$, there holds \begin{equation}\label{esU22} \sup_{t\in(t_0, t_1)}\|U_2(t, \cdot)\|_{C^{\vartheta}(\R^n)}+\sup_{x\in\R^n}\|U_2(\cdot, x)\|_{C^{\frac{\vartheta}{4}}(t_0, t_1)} \leq C_{n}\upsilon(t_1-t_0)\|f\|_{L^{\infty}((t_0, t)\times\R^n)}, \end{equation} where $\upsilon(t-t_0)=(t-t_0)+(t-t_0)^{1-\frac{\vartheta}{4}}$ and $C_{n}$ is a positive constant which depends on $n$. The Proposition \ref{prop7} follows from \eqref{esU1}, \eqref{esU2}, \eqref{esU21} and \eqref{esU22}. \end{proof} We will apply the above proposition to study the solvability of a class of semilinear parabolic equations. Let $u_0\in C^1(\R^n)\cap L^\infty(\R^n)$, and we consider the initial value problem \begin{equation}\label{heatnonlineareq} \left\{ \begin{array}{l} u_t+(-\Delta)^2u=G[\Delta u, \nabla u, u, t, x]\qquad \text{in}\ (t_0, t_1)\times\R^n, \\[2mm] u(t_0, x)=u_0(x)\hspace{3.35cm} \text{in}\ \R^n, \end{array} \right. \end{equation} where $G[p, \vec{q}, s, t, x]:\R\times\R^n\times\R\times[t_0, +\infty)\times\R^n\rightarrow \R$ is a measurable function satisfying \begin{enumerate} \item For every $M>0$ and $T>t_0$, there exists $C_{T, M}>0$ such that $\Big{|}G[p, \vec{q}, s, t, x]\Big{|}\leq C_{T, M}$ for all $x\in\R^n$, $t\in[t_0, T]$ and $p, \abs{\vec{q}}, s \in[-M, M]$. \item There is a constant $\sigma>0$ in such a way that \begin{equation}\label{delpf} \Big{|}G[p_1, \vec{q}_1, s_1, t, x]-G[p_2, \vec{q}_2, s_2, t, x]\Big{|}\leq \sigma\Big(\abs{p_1-p_2}+\abs{\vec{q}_1-\vec{q}_2}+\abs{s_1-s_2}\Big), \end{equation} for all $ x\in\R^n, t\geq t_0$, $p_1, p_2 \in\R$, $\vec{q}_1, \vec{q}_2\in\R^n$ and $s_1, s_2\in\R$. \end{enumerate} In particular, we can take \begin{equation}\label{defG} G[\Delta u, \nabla u, u, t, x]=a(t, x)\Delta u+\sum_{i=1}^nb^i(t, x)\nabla_{x_i} u+c(t, x)u+g(t, x), \end{equation} where those functions $a(t, x)$, $b^1(t, x), \cdots, b^n(t, x)$, $c(t, x)$ are all belonging to $L^\infty ([t_0, +\infty), C^1(\R^n))$ and $g\in L^\infty([t_0, +\infty)\times\R^n)$. Assume that $u(t, x)$ satisfies that $$ \Delta u, \ \abs{\nabla u}, \ u\in L^{\infty}((t_0, t_1)\times\R^n), $$ then by \eqref{demil}, $u$ is a mild solution of \eqref{heatnonlineareq} if and only if \begin{equation}\label{demils} u(t, x)=\Gamma_{t-t_0}[u_0](x)+\int^t_{t_0}\Gamma_{t-\tau}\Big[G[\Delta u, \nabla u, u, \tau, \cdot](\tau, \cdot)\Big](x){\mathrm d}\tau\qquad\text{for}\ (t, x)\in (t_0, t_1)\times\R^n. \end{equation} Notice that $G[\Delta u, \nabla u, u, t, x]\in L^{\infty}((t_0, t_1)\times\R^n)$ by the assumptions on $u$ and $G$. Thus (\ref{demils}) is well defined by Proposition \ref{prop7}. Now we devote to the study of the solvability of problem \eqref{heatnonlineareq}. For convenience, we define a map $\mathcal{N}_{G, u_0}$ from $L^\infty[(t_0, t_1); C^{2}(\R^{n})]$ to itself as follows $$\mathcal{N}_{G, u_0}[u](t, x):=\Gamma_{t-t_0}[u_0](x)+\int^t_{t_0}\Gamma_{t-\tau}\Big[G[\Delta u, \nabla u, u](\cdot, \tau)\Big](x){\mathrm d}\tau\qquad\text{for}\ (x, t)\in \R^n\times(t_0, t_1), $$ for $u_0\in C^1(\R^n)\cap L^\infty(\R^n)$ and $u\in L^\infty[(t_0, t_1); C^{2}(\R^{n})]$. The main result is the following: \begin{prop}\label{prop8} Given any $u_0\in C^1(\R^n)\cap L^\infty(\R^n)$, there exists a unique mild solution $u$ to problem \eqref{heatnonlineareq} with $t_1=+\infty$. Moreover, if $u_0\in C^2(\R^n)$, then the solution $u$ has the property: for $s>t_0$, there exists $C$ independent of $t$ such that \begin{equation}\label{ess} \sup_{t\in(t_0, s)}\|u(t, \cdot)\|_{C^{2}(\R^n)} \leq C\Big(\Big[(s-t_0)+(s-t_0)^{1/2}\Big]\|G(0, 0, 0, x, t)\|_{L^{\infty}((t_0, s)\times\R^n)}+\|u_0\|_{C^2(\R^n)}\Big). \end{equation} \end{prop} \begin{proof} Notice that this proposition is equivalent to that problem \eqref{heatnonlineareq} has a unique solution $u$ for any $t_1\in (t_0, +\infty)$. According to the definition in \eqref{demil}, $u$ is a mild solution of \eqref{heatnonlineareq} if and only if $u$ is a fixed point of the map $\mathcal{N}_{G, u_0}$. Thus we will study the existence and uniqueness of fixed point for the map $\mathcal{N}_{G, u_0}$. Using Proposition \ref{prop7} and the assumptions on $G$, we have that $\mathcal{N}_{G, u_0}$ is an operator from the space $L^\infty[(t_0, t_1); C^{2}(\R^{n})]$ to itself. We claim that the map $\mathcal{N}_{G, u_0}$ is a contraction when $T_1:=t_1-t_0$ is small enough. Indeed, by Proposition \ref{prop6}, the assumptions on $G$ in (\ref{delpf}) and some direct computations, we have that for any $u, w\in L^\infty[(t_0, t_1); C^{2}(\R^{n})]$, \begin{align*} &\big|\mathcal{N}_{G, u_0}[u](t, x)-\mathcal{N}_{G, u_0}[w](t, x)\big| \\[2mm] &\leq \int^t_{t_0}\int_{\R^n}\abs{p_n(t-\tau, x-y)}\cdot \Big{|}G[\Delta u, \nabla u, u](\tau, y)-G[\Delta w, \nabla w, w](\tau, y)\Big{|}{\mathrm d}y{\mathrm d}\tau \\[2mm] &\leq \sigma T_1C_n\Big(\|\Delta u-\Delta w\|_{L^{\infty}((t_0, t_1)\times\R^n) }+\|\nabla u-\nabla w\|_{L^{\infty}((t_0, t_1)\times\R^n )}+\|u-w\|_{L^{\infty}((t_0, t_1)\times \R^n)}\Big), \end{align*} for a.e. $x\in\R^n$, $t\in (t_0, t_1)$, where $C_n$ is a positive number which only depends on $n$. Hence, we can choose $T_1\leq\frac{1}{2\sigma C_n}$ such that \begin{align*} &\big\|\mathcal{N}_{G, u_0}[u](t, x)-\mathcal{N}_{G, u_0}[w](t, x)\big\|_{L^{\infty}((t_0, t_1)\times\R^n )} \\[2mm] & \leq \frac{1}{2}\Big(\|\Delta u-\Delta w\|_{L^{\infty}((t_0, t_1)\times\R^n )}+\|\nabla u-\nabla w\|_{L^{\infty}((t_0, t_1)\times\R^n )}+\|u-w\|_{L^{\infty}((t_0, t_1)\times\R^n )}\Big). \end{align*} Thus, the map $\mathcal{N}_{G, u_0}$ is a contraction. Next we consider the following iterative sequence $$u_m(t, x)=(\mathcal{N}_{G, u_0})^m[0](t, x), $$ by Banach Fixed Theorem, then there is $u\in L^\infty[(t_0, t_1); C^{2}(\R^{n})]$ such that $u_m(x, t)$ converges to $ u(x, t)$ in the space $L^\infty[(t_0, t_1); C^{2}(\R^{n})]$ and $\mathcal{N}_{G, u_0} (u)=u$, which is the unique solution of problem \eqref{heatnonlineareq} with $t_1=t_0+\frac{1}{2\sigma C_n}$. Let $u^1(t, x)$ denote the mild solution of \eqref{heatnonlineareq}, which can be obtained by the above arguments. We consider the following initial problem: \begin{equation*} \left\{ \begin{array}{l} u_t+(-\Delta)^2u=G[\Delta u, \nabla u, u, x, t] \hspace{1cm}\text{in}\ \big(t_0+\frac{1}{2\sigma C_n}, t_0+\frac{1}{\sigma C_n}\big)\times\R^n, \\[2mm] u\big(t_0+\frac{1}{2\sigma C_n}, x\big)=u^1\big(t_0+\frac{1}{2\sigma C_n}, x\big)\qquad \text{in}\ \R^n. \end{array} \right. \end{equation*} Repeating the previous arguments, we can get that the above equation has a unique solution $u^2(t, x)$. Thus we can extend the time interval of existence to $\big(t_0, t_0+\frac{l}{2\sigma C_n}\big)$ with any positive integer $l\geq 1$ step by step. In this process, we used the fact that the initial function is belonging to $C^1(\R^n)\cap L^\infty(\R^n)$ on each step. Indeed, the solutions are well-defined bounded continuous functions of $t$ and $x$, and are continuously differentiable in $x$, due to the estimate \eqref{es2} with $\vartheta=1$ in Proposition \ref{prop7}. The estimate \eqref{ess} comes from Proposition \ref{prop7}. \end{proof} \subsection{A priori estimate of a linear problem involving the operator $L$} Recall the functions $\rho$, $\gamma_n$ and $h$ in \eqref{drho1}-\eqref{assumeh}. By defining $$ \|\psi\|_{C_{\Phi}\left((t_1, t_2)\times(0, \infty)\right)}:=\left\| \frac{\psi}{\Phi}\right\| _{L^\infty((t_1, t_2)\times(0, \infty))}, $$ we choose a set consisting of continuous functions as the following \begin{align} & C_{\Phi}((t_1, t_2)\times(0, \infty)) \nonumber\\[2mm] &:=\left\{ \psi \ :\ \psi(t, r)=0, \ \forall\, (t, r)\in(t_1, t_2)\times\left(0, \frac{\delta_0}{2}\right), \ \text{and}\ \|\psi\|_{C_{\Phi}((t_1, t_2)\times(0, \infty))}<+\infty \right\}, \label{eq3.3} \end{align} where $t_1<t_2<0$, $\delta_0>0$ is a small fixed number, the functions $\Phi(t, r)$ and is given by (\ref{dpsi}). Let us consider the following Cauchy problem: \begin{equation}\label{eq3.5} \left\{ \begin{array}{l} \psi_t=L[\psi]+g(t, r)\qquad \ \text{in}\ (s, -T]\times(0, \infty), \\[2mm] \psi(s, r)=0,\hspace{1.85cm} \forall\, r\in (0, +\infty), \\[2mm] \partial^j_r\psi(t, 0)=0, \hspace{1.55cm} \forall\, t\in (s, -T]\ \text{and}\ j=1, 3, \end{array} \right. \end{equation} where $g\in C_{\Phi}\left((-\infty, -T]\times(0, \infty)\right)$, $T>0$ and $s+1<-T$. Notice that problem \eqref{eq3.5} has a unique solution $\psi^s(t, r)$. Indeed, by the definition of the operator $L[\psi]$ in \eqref{eq2.13}, we can rewrite $L[\psi]$ as follows \begin{align*} L[\psi]=&-(\Delta)^2\psi+2W''\big(z(t, r)\big)\Delta \psi+2W'''\big(z(t, r)\big)\Delta z(t, r)\psi +W^{(4)}\big(z(t, r)\big)\abs{\nabla z(t, r)}^2\psi \\[2mm] &+2W'''\big(z(t, r)\big)\nabla z(t, r)\cdot\nabla \psi -\Big[\big|W''\big(z(t, r)\big)\big|^2-W'\big(z(t, r)\big)W'''\big(z(t, r)\big)\Big]\psi, \end{align*} where we recall that approximate solution $$ z(t, r)=\left(\omega\big(r-\rho(t)\big)+\frac{(n-1)(n-3)}{r^2} \widetilde{\omega}\big(r-\rho(t)\big)\right)\chi\big(r\big)+\chi\big(r\big)-1, $$ and the functions $\chi(r)$ and $\rho(t)$ are given by \eqref{dcut-off} and \eqref{drho1} respectively. According to Lemma \ref{lem1} and the assumptions on $W(s)$ in \eqref{eq1.4}, those coefficients in front of $\Delta \psi$, $\nabla\psi$ and $\psi$ are all smooth and bounded for $x\in\R^n$ and $t<-2$. Thus we know that problem \eqref{eq3.5} is uniquely solvable by applying Proposition \ref{prop8}, and denote this solution by $\psi^s(t, r)$. Moreover, we have that $\psi^s(t, r)=0$ when $r\in(0,\frac{\delta_0}{2})$. Indeed, we consider \begin{equation*} \psi^{e}(t,r):=\left\{ \begin{array}{l} \psi^s(t,r),\qquad \text{if}\ r\in(0,\frac{\delta_0}{2}]; \\[2mm] 0,\hspace{1.65cm} \text{if}\ r\in(\frac{\delta_0}{2},+\infty), \end{array} \right. \end{equation*} which satisfies that \begin{equation*} \left\{ \begin{array}{l} \psi_t=-(\Delta)^2\psi+4\Delta\psi-4\psi\qquad \ \text{in}\ (s, -T]\times(0,+\infty), \\[2mm] \psi(s, r)=0,\hspace{3.1cm} \forall\, r\in (0, +\infty), \\[2mm] \partial^j_r\psi(t, 0)=0, \hspace{2.83cm} \forall\, t\in (s, -T]\ \text{and}\ j=1, 3, \end{array} \right. \end{equation*} where $n\geq4$ and we have used the fact that $z(t,r)=-1$ when $r\in(0,\frac{\delta_0}{2}]$. According to the proposition \ref{prop8}, the above equation has a unique mild solution $\psi(t,r)=0$. Hence $\psi^e(t,r)=0$, that is $\psi^s(t, r)=0$ when $r\in(0,\frac{\delta_0}{2}]$. We next establish a priori estimate for the solutions of problem (\ref{eq3.5}). \begin{lem}\label{lem5} Let $g\in C_{\Phi}((s, -T)\times(0, \infty))$ and $\psi^s$ be a solution of problem $(\ref{eq3.5})$ which satisfies the orthogonality condition \begin{equation}\label{eq3.6} \int_0^\infty \psi^s(t, r)\omega'\big(r-\rho(t)\big)r^{n-1}{\mathrm d}r=0, \qquad s<t<-T. \end{equation} Then there exists a uniform constant $T_0>T$ such that for any $t\in (s, -T_0]$, the following estimate is valid \begin{equation}\label{eq3.7} \sum_{l=0}^3\|\partial_r^l\psi^s\|_{C_{\Phi}((s, t)\times(0, \infty))}\leq C\|g\|_{C_{\Phi}((s, t)\times(0, \infty))}, \end{equation} where $C>0$ is a uniform positive constant independent of $t$ and $s$. \end{lem} The proofs of this lemma consist of tedious analysis, which will be given in the sequel. \subsubsection{Proof of Lemma \ref{lem5}} Here are the details. \begin{proof} We prove the above lemma by contradiction. Assume that there exist two sequences $\{s_i\}$ and $\{t_i\}$ such that $s_i+1<t_i< 0$ and $s_i\rightarrow -\infty$, $t_i\rightarrow -\infty$ when the sub-index $i$ goes to infinity. We assume that: for each given $i$, there exists $g^{i}\in C_{\Phi}((s_i, t_i)\times(0, \infty))$, such that \begin{equation}\label{eq3.8} \begin{aligned} \sum_{l=0}^3\big\|\partial_r^l\psi^i\big\|_{C_{\Phi}((s_i, t_i)\times(0, \infty))}=1, \end{aligned} \end{equation} and \begin{equation}\label{eq8} \|g^{i}\|_{C_{\Phi}((s_i, t_i)\times(0, \infty))}\rightarrow 0, \qquad \text{as}\ i\rightarrow+\infty, \end{equation} where $\psi^i$ is the solution of problem (\ref{eq3.5})-(\ref{eq3.6}) with $g=g^i$, $s=s_i$ and $-T=t_i$. \textbf{Assertion}: \emph{ For any $R>0$, we have that \begin{equation}\label{eq3.9} \lim_{i\rightarrow\infty}\sum_{l=0}^3\left\|\frac{\partial^l_r\psi^i}{\Phi}\right\|_{L^\infty\big(A^{(s_i, t_i)}_{R}\big)}=0, \end{equation} where $\Phi(t, r)$ is given by $(\ref{eq2.13})$ and $A^{(s_i, t_i)}_{R}$ is defined by $$ A^{(s_i, t_i)}_{R}:=\Big\{(t, r)\in (s_i, t_i)\times (0, \infty): \ \abs{r-\gamma_n(t)}<R+1\Big\} $$ and the function $\gamma_n(t)$ is defined by \eqref{dgamma-n}. } The proof of (\ref{eq3.9}) will be provided in Section \ref{proofAssertion}. We first accept the validity of (\ref{eq3.9}). According to \eqref{eq3.8}, we have that there exists $l\in\{0, 1, 2, 3\}$ such that $$\big\|\partial_r^l\psi^i\big\|_{C_{\Phi}((s_i, t_i)\times(0, \infty))}\geq\frac{1}{4}.$$ Without loss of generality, we assume that $\|\psi^i\|_{C_{\Phi}((s_i, t_i)\times(0, \infty))}\geq\frac{1}{4}$. For the other situations, the following arguments can be adapted in a simple way. Recall the definition of $C_{\Phi}((s_i, t_i)\times(0, \infty))$ in \eqref{eq3.3}, then we derive that \begin{equation*} \|\psi^i\|_{C_{\Phi}((s_i, t_i)\times(0, \infty))}=\sup_{(t, r)\in(s_i, t_i)\times(0, \infty)}\frac{\abs{\psi^i(t, r)}}{\Phi(t, r)}\geq\frac{1}{4}. \end{equation*} Thus, there exists $(\bar{t}_i, \bar{r}_i)\in (s_i, t_i)\times(0, \infty)$ such that \begin{equation*} \frac{\abs{\psi^i(\bar{t}_i, \bar{r}_i)}}{\Phi(\bar{t}_i, \bar{r}_i)}\geq\frac{1}{5}. \end{equation*} Furthermore, by (\ref{eq3.9}), we have \begin{equation}\label{j1} \lim_{i\rightarrow +\infty}\abs{\bar{r}_i-\gamma_{n}(\bar{t}_i)}=+\infty. \end{equation} Let us define \begin{equation*} \phi^i(\nu, y):=\frac{\psi^i\big(\nu+\bar{t}_i, y+\bar{y}_i+\gamma_n(\nu+\bar{t}_i)\big)} {\Phi\big(\bar{t}_i, \bar{y}_i+\gamma_{n}(\bar{t}_i)\big)}, \qquad\text{where}\ \bar{y}_i=\bar{r}_i-\gamma_{n}(\bar{t}_i), \end{equation*} where the functions $\gamma_n(t)$ and $\Phi(t,r)$ are given by \eqref{dgamma-n} and \eqref{dpsi} respectively. Thus $\phi^i(\nu, y)$ satisfies the following problem \begin{align} \partial_\nu\phi^i =& -\phi_{yyyy}^{i} -\frac{2(n-1)}{y+\bar{y}_i+\gamma_{n}(\nu+\bar{t}_i)}\phi_{yyy}^i +\Bigg[2W''\big(\bar{z}(\nu, y)\big)-\frac{(n-3)(n-1)}{ \big(y+\bar{y}_i+\gamma_{n}(\nu+\bar{t}_i)\big)^2}\Bigg]\phi^i_{yy} \nonumber\\[3mm] &-\Big(W''\big(\bar{z}(\nu, y)\big)\Big)^2\phi^i+\Bigg[\frac{2(n-1)W''\big(\bar{z}(\nu, y)\big)}{y+\bar{y}_i+\gamma_{n}(\nu+\bar{t}_i)} -\frac{(3-n)(n-1)}{\big(y+\bar{y}_i+\gamma_{n}(\nu+\bar{t}_i)\big)^3}\Bigg]\phi^i_{y} \nonumber\\[3mm] &+2W'''\big(\bar{z}(\nu, y)\big)\bar{z}_y\phi^i_y-W'''\big(\bar{z}(\nu, y)\big)W'\big(\bar{z}(\nu, y)\big)\phi+W^{(4)}\big(\bar{z}(\nu, y)\big)\abs{\bar{z}_y}^2\phi^i \nonumber\\[3mm] &+2W'''\big(\bar{z}(\nu, y)\big)\bar{z}_{yy}\phi^i +\frac{2(n-1)}{y+\bar{y}_i+\gamma_{n}(\nu+\bar{t}_i)}W'''\big(\bar{z}(\nu, y)\big)\bar{z}_y\phi^i +\phi^i_y\partial_\nu\gamma_{n}(\nu+\bar{t}_i) \nonumber\\[3mm] &+\frac{g^i\big(\nu+\bar{t}_i, y+\bar{y}_i+\gamma_{n}(\nu+\bar{t}_i)\big)} {\Phi\big(\bar{t}_i, \bar{y}_i+\gamma_{n}(\bar{t}_i)\big)} \qquad \text{in}\ B^{(s_i, t_i)}_i, \label{eqj.2} \end{align} with the conditions \begin{align} \abs{\phi^i(0, 0)}\geq\frac{1}{5}, \qquad \phi^i(s_i-\bar{t}_i, y)=0,\qquad \text{for}\ y\in \big(-\bar{y}_i-\gamma_{n}(\nu+\bar{t}_i), +\infty\big), \end{align} where $$ \bar{z}(\nu, y):=z\big(\nu+\bar{t}_i, y+\bar{y}_i+\gamma_{n}(\nu+\bar{t}_i)\big) $$ and the set $B^{(s_i, t_i)}_i$ is defined by $$B^{(s_i, t_i)}_i:=(s_i-\bar{t}_i, 0]\times(-\bar{y}_i-\gamma_{n}(\nu+\bar{t}_i), +\infty).$$ Since $\lim\limits_{i\rightarrow +\infty} \abs{\bar{y}_i}=+\infty$ due to \eqref{eq3.9}, we have that $\lim\limits_{i\rightarrow +\infty} \bar{y}_i=+\infty$ or $-\infty$ up to a subsequence. Moreover, notice that there exist two cases: \noindent\textbf{Case $1$}, $\lim\limits_{i\rightarrow\infty}s_i-\bar{t}_i=-\infty$. \noindent\textbf{Case $2$}, $\lim\limits_{i\rightarrow\infty}s_i-\bar{t}_i=-\varsigma$ for some positive constant $\varsigma>0$. \\ Next we will prove that Lemma \ref{lem5} holds, and the process will be divided into the following two steps. \textbf{Step one.} We first consider \textbf{Case $1$}, and then have that $\phi^i\rightarrow\phi$ locally uniformly. Moreover, $\abs{\phi(0, 0)}>\frac{1}{5}$ and $\phi$ satisfies \begin{equation}\label{limeq} \phi_{\nu}=-\phi_{yyyy}+2W''(1)\phi_{yy}-\big[W''(1)\big]^2\phi\qquad \text{in}\ (-\infty, 0]\times\R. \end{equation} By the assumption on $\psi^i$ in \eqref{eq3.8}, we have that for all $(\nu, y)\in B^j_i$, \begin{equation}\label{esp0} \abs{\phi^i(\nu, y)}=\abs{\frac{\psi^i\big(\nu+\bar{t}_i, y+\bar{y}_i+\gamma_{n}(\nu+\bar{t}_i)\big)} {\Phi\big(\bar{t}_i, \bar{y}_i+\gamma_{n}(\bar{t}_i)\big)}} \leq \abs{\frac{\Phi\big(\nu+\bar{t}_i, y+\bar{y}_i+\gamma_{n}(\nu+\bar{t}_i)\big)} {\Phi\big(\bar{t}_i, \bar{y}_i+\gamma_{n}(\bar{t}_i)\big)}}, \end{equation} where the sets $B_i^j$ with $j=0, 1$, are defined by \begin{equation}\begin{aligned}\label{dB} B_i^0:&=(s_i-\bar{t}_i, 0]\times\big(-\bar{y}_i-\gamma_{n}(\nu+\bar{t}_i), \ \delta_0-\gamma_n(\nu+\bar{t}_i)-\bar{y}_i\big], \\[2mm] B_i^1:&=(s_i-\bar{t}_i, 0]\times\big(\delta_0-\gamma_n(\nu+\bar{t}_i)-\bar{y}_i, \ +\infty\big). \end{aligned}\end{equation} Next we estimate the right hand side of the above inequality \eqref{esp0}. There exist the following different situations: \\ \noindent{\bf (1).} When $j=1$ and $\bar{r}_i=\bar{y}_i+\gamma_{n}(\bar{t}_i)>\delta_0$, by the definitions of $\gamma_n(t)$ and $\Phi(t, r)$ in \eqref{dgamma-n} and \eqref{dpsi}, we have that \begin{align*} \abs{\frac{\Phi\big(\nu+\bar{t}_i, y+\bar{y}_i+\gamma_{n}(\nu+\bar{t}_i)\big)} {\Phi\big(\bar{t}_i, \bar{y}_i+\gamma_{n}(\bar{t}_i)\big)}} &\leq \frac{\hbar(\nu+\bar{t}_i)}{\hbar(\bar{t}_i)} \left(\frac{1+\abs{\bar{y}_i-\frac{1}{3\alpha}\log\abs{\bar{t}_i}}} {1+\abs{y+\bar{y}_i-\frac{1}{3\alpha}\log\abs{\nu+\bar{t}_i}}}\right)^p \\[2mm] &\leq C \frac{\hbar(\nu+\bar{t}_i)}{\hbar(\bar{t}_i)} \Big(1+\abs{y}+\log\Big[\frac{\abs{\nu+\bar{t}_i}}{\abs{\bar{t}_i}}\Big]\Big)^p \\[2mm] &\leq C\left(\abs{y}+1\right)^p, \end{align*} where the function $\hbar(t)$ is defined by \begin{equation}\label{dehat} \hbar(t):=\abs{t}^{-\frac{1}{2}}\log\abs{t}, \end{equation} and $C$ is a positive constant which depends on $n$ and $\alpha$. \noindent{\bf (2).} When $j=1$ and $\frac{\delta_0}{2}<\bar{r}_i\leq\delta_0$, by the same argument as above, we have that \begin{equation*} \begin{aligned} &\abs{\frac{\Phi\big(\nu+\bar{t}_i, y+\bar{y}_i+\gamma_n(\nu+\bar{t}_i)\big)} {\Phi\big(\bar{t}_i, \bar{y}_i+\gamma_n(\bar{t}_i)\big)}}\leq C. \end{aligned} \end{equation*} \\ \noindent{\bf (3).} When $j=0$, by the definitions of $\gamma_n(t)$ and $\Phi(t, r)$ in \eqref{dgamma-n} and \eqref{dpsi}, we have that \begin{equation*} \begin{aligned} \abs{\frac{\Phi\big(\nu+\bar{t}_i, y+\bar{y}_i+\gamma_{n}(\nu+\bar{t}_i)\big)} {\Phi\big(\bar{t}_i, \bar{y}_i+\gamma_{n}(\bar{t}_i)\big)}} &\leq \frac{\hbar(\nu+\bar{t}_i)}{\hbar(\bar{t}_i)} \left(1+\abs{\bar{y}_i-\frac{\alpha}{3}\log\abs{\bar{t}_i}}\right)^p \\[2mm] &\leq C(1+\abs{y})^p, \end{aligned} \end{equation*} where we have used the fact that $\abs{\bar{y}_i-\frac{\alpha}{4}\log\abs{t_i}}\leq \abs{y}$ since $(\nu, y)\in B_i^0$ given by \eqref{dB}. Combining above estimates, we derive that \begin{equation}\label{esp} \abs{\phi^i(\nu, y)}\leq C(\abs{y}+1)^p, \end{equation} for all $(\nu, y)\in B_i^j$ with $j=0, 1$, where $p\in(n, n+1)$ and $C$ is a positive constant only depending on $\alpha$ and $n$. Similarly, we have that \begin{equation}\label{esp1} \abs{\phi^i_y(\nu, y)}\leq C(\abs{y}+1)^p, \quad \abs{\phi^i_{yy}(\nu, y)}\leq C(\abs{y}+1)^p\quad \text{and}\quad \abs{\phi^i_{yyy}(\nu, y)}\leq C(\abs{y}+1)^p, \end{equation} for all $(\nu, y)\in B_i^j$ and $C$ is a positive constant which only depends on $\alpha$ and $n$. Notice that $$ \cup_{j=0}^1B_i^j=\big(s_i-\bar{t}_i, 0\big]\times\big(-\bar{y}_i-\gamma_n(\nu+\bar{t}_i),\, +\infty\big) $$ due to \eqref{dB}. Owing to \eqref{esp} and \eqref{esp1}, we have that $$ \phi^i\rightarrow \phi\qquad \mbox{in}\ L_{\text{loc}}^2\big[(-\bar{\varsigma}, 0); H^2_{\text{loc}}(\R)\big], $$ where $\bar{\varsigma}=\varsigma$ or $\infty$. Moreover, we note that there holds that $$ \lim\limits_{i\rightarrow+\infty}-\gamma_n(\nu+\bar{t}_i)-\bar{y}_i=M, $$ up to a subsequence, where $M=-\infty$ or $M\in(-\infty, 0]$. For the second case, that is $M\in(-\infty, 0]$, we can get that $\phi$ satisfies that \begin{equation*} \phi_{\nu}=-\phi_{yyyy}+2W''(1)\phi_{yy}-\big[W''(1)\big]^2\phi\qquad \text{in}\ (-\infty, 0]\times[0, +\infty). \end{equation*} Then, we can consider the function \begin{equation*} \widetilde{\phi}(\nu, x):=\phi(\nu, x)\quad \text{when}\ x\geq0, \quad \text{and} \quad \widetilde{\phi}(\nu, x):=\phi(\nu, -x)\quad \text{when}\ x<0, \end{equation*} which is the even extension of $\phi$ with respect to $x$. It is easy to check that $ \widetilde{\phi}$ satisfies equation \eqref{limeq}. For \textbf{Case $2$}: $\lim\limits_{i\rightarrow\infty}s_i-\bar{t}_i=-\varsigma$ for some constant $\varsigma>0$, we have $\phi^i\rightarrow\phi$ locally uniformly, and $\phi$ satisfies \begin{equation*} \left\{ \begin{array}{l} \phi_{\nu}=-\phi_{yyyy}+2W''(1)\phi_{yy}-\big[W''(1)\big]^2\phi\qquad \text{in}\ (-\varsigma, 0]\times\R, \\[3mm] \phi(-\varsigma, y)=0, \qquad \text{for all}\ y\in\R. \end{array} \right. \end{equation*} By Proposition \ref{prop8}, the above equation has a unique solution $\phi\equiv0$. However, we have that $$ \abs{\phi(0, 0)}=\lim\limits_{i\rightarrow\infty}\abs{\phi^i(0, 0)}\geq\frac{1}{5} $$ in \eqref{eqj.2}, which leads to a contradiction. Hence the \textbf{Case $2$} cannot happen. \textbf{Step two.} We claim that \begin{equation}\label{des} \phi(\nu, y)\equiv0, \end{equation} for all $(\nu, y)\in(-\infty, 0]\times\R$. This result contradicts with $\abs{\phi(0, 0)}>\frac{1}{5}$, thus we can derive that Lemma \ref{lem5} holds. In the rest part of this proof, the job is to show that the above conclusion in \eqref{des} holds. Recall that $\alpha=\sqrt{W''(1)}>0$, we then consider the following parabolic equation \begin{equation*} u_\nu=-u_{yyyy}+2\alpha^2 u_{yy}, \qquad \forall\, (\nu, y)\in (0, +\infty)\times\R. \end{equation*} Using the Fourier transformation for $y$, the above equation has a formula solution \begin{equation}\label{11} \mathbf{Q}(\nu, y)=\frac{1}{2\pi}\int_\R \exp\Big\{-\nu\big(\abs{\xi}^4+2\alpha^2\abs{\xi}^2\big)\Big\}e^{i\xi y}{\mathrm d}\xi. \end{equation} Hence, for any $T>0$ and $f\in L^\infty((-T, 0)\times\R)$, the following initial value problem \begin{equation*} u_\nu=-u_{yyyy}+2\alpha^2 u_{yy}-\alpha^4u+f(\nu, y), \qquad (\nu, y)\in (-T, 0)\times\R, \qquad u(-T, y)=0, \qquad \forall\, y\in\R, \end{equation*} has a solution of the form \begin{equation}\label{f1} u(\nu, y)=\frac{1}{2\pi}\int_{0}^{\nu-T_1}\int_{\R}e^{-\alpha^4\tau}\mathbf{Q}(\tau, x)f(\nu-\tau, y-x){\mathrm d}x{\mathrm d}\tau, \qquad \forall\, (\nu, y)\in(-T, 0)\times\R. \end{equation} According to formula \eqref{f1} and equation \eqref{eqj.2}, we have that \begin{equation}\label{dfs} \phi^i(\nu, y)=\int_{0}^{\nu-s_i+\bar{t}_i}\int_{\R}e^{-\alpha^4\tau}\mathbf{Q}(\tau, x)f_i\big(\phi^i(\nu-\tau, y-x),\, \nu-\tau, y-x\big){\mathrm d}x{\mathrm d}\tau, \end{equation} for all $(\nu, y)\in(s_i-\bar{t}_i, 0)\times\R$, where the function $f_i(\phi^i, \nu, y)$ is defined as follows \begin{align*} f_i(\phi^i, \nu, y):=&-\frac{2(n-1)}{y+\bar{y}_i+\gamma_{n}(\nu+\bar{t}_i)}\phi_{yyy}^i-\frac{(n-3)(n-1)}{ \big(y+\bar{y}_i+\gamma_{n}(\nu+\bar{t}_i)\big)^2}\phi^i_{yy} \\[2mm] &+\Bigg[\frac{2(n-1)W''\big(\bar{z}(\nu, y)\big)}{y+\bar{y}_i+\gamma_{n}(\nu+\bar{t}_i)} -\frac{(3-n)(n-1)}{\big(y+\bar{y}_i+\gamma_{n}(\nu+\bar{t}_i)\big)^3}\Bigg]\phi^i_{y} \\[2mm] &+2W'''\big(\bar{z}(\nu, y)\big)\bar{z}_y\phi^i_y-W'''\big(\bar{z}(\nu, y)\big)W'\big(\bar{z}(\nu, y)\big)\phi+W^{(4)}\big(\bar{z}(\nu, y)\big)\abs{\bar{z}_y}^2\phi^i \\[2mm] &+2W'''\big(\bar{z}(\nu, y)\big)\bar{z}_{yy}\phi^i+2\frac{n-1}{y+\bar{y}_i+\gamma_{n}(\nu+\bar{t}_i)}W'''\big(\bar{z}(\nu, y)\big)\bar{z}_y\phi^i +\phi^i_y\partial_\nu\gamma_{n}(\nu+\bar{t}_i) \\[2mm] &+\frac{g^i\big(\nu+\bar{t}_i, y+\bar{y}_i+\gamma_{n}(\nu+\bar{t}_i)\big)} {\Phi\big(\bar{t}_i, \bar{y}_i+\gamma_{n}(\bar{t}_i)\big)} \\[2mm] &+2\Big[W''\big(\bar{z}(\nu, y)\big)-W''(1)\Big]\phi^i_{yy} -\Big(\big[W''\big(\bar{z}(\nu, y)\big)\big]^2-\big[W''(1)\big]^2\Big)\phi^i, \end{align*} where $\bar{z}(\nu, y)=z\big(\nu+\bar{t}_i, y+\bar{y}_i+\gamma_{n}(\nu+\bar{t}_i)\big)$, the function $z(\nu, r)$ is given by \eqref{dz}. Using \eqref{eq3.3}, \eqref{eq8}, \eqref{esp}, \eqref{esp1}, the same arguments in proof of Lemma \ref{lem2}, we have that \begin{align*} \abs{f_i\big(\phi^i(\nu, y), \nu, y\big)}\leq& C(1+\abs{y})^p\Bigg[\exp\big\{-\alpha\abs{y+\bar{y}_i}\big\} +\|g^{i}\|_{C_{\Phi}((s_i, t_i)\times(0,+\infty))} \\[2mm] &\qquad\qquad+\frac{1}{\left[-(\nu+\bar{t}_i)\right]^{3/4}} +\sum_{l=1}^3\big(y+\bar{y}_i+\gamma_{n}(\nu+\bar{t}_i)\big)^{-l}\Bigg] \overline{\chi}_{\left\{y>\frac{\delta_0}{2}-\gamma_n(\nu+\bar{t}_i)-\bar{y}_i\right\}}, \end{align*} for all $(\nu, y)\in (s_i-\bar{t}_i, 0]\times(-\bar{y}_i-\gamma_n(\nu+\bar{t}_i), +\infty)$, where $C$ depends on $\alpha$ and $n$, and $\overline{\chi}_{A}$ denotes the characteristic function of the set $A$. Moreover a straightforward shift of contour gives that there exists a positive constant $C$ such that $$\abs{\mathbf{Q}(\nu, y)}\leq C \nu^{-\frac{1}{4}}\exp\left\{-\frac{\abs{y}}{\nu^{1/4}}\right\}, $$ or see $(3.2)$ in \cite{BKT}. Thus, using \eqref{dfs} and the above inequality, we have that \begin{align}\label{13} \abs{\phi^i(\nu, x)}&\leq C\int_0^{\nu-s_i+\bar{t}_i}\int^{+\infty}_{\frac{\delta_0}{2}-\bar{y}_i-\gamma_{n}(\nu+\bar{t}_i)}\tau^{-\frac{1}{4}} \exp\Bigg\{-\frac{\abs{x-y}}{\tau^{1/4}}-\alpha^4\tau\Bigg\} \Bigg[\exp\big\{-\alpha\abs{y+\bar{y}_i}\big\} \\[2mm] &\qquad\quad+\frac{1}{\left[-(\nu+\bar{t}_i-\tau)\right]^{3/4}} +\|g^{i}\|_{C_{\Phi}((s_i, t_i)\times\R)} +\sum_{l=1}^3\big[y+\bar{y}_i+\gamma_{n}(\nu+\bar{t}_i)\big]^{-l}\Bigg]\big(1+\abs{y}\big)^p {\mathrm d}y {\mathrm d}\tau. \nonumber \end{align} Thus, by Lebesgue dominated convergence theorem, the facts $$ \abs{\bar{y}_i} \rightarrow \infty, \qquad \bar{t}_i\rightarrow-\infty \qquad\mbox{and}\qquad \|g^{i}\|_{C_{\Phi}((s_i, t_i)\times(0,+\infty))}\rightarrow 0, $$ when $i$ goes to infinity in \eqref{eq8}, we can obtain \eqref{des}. \end{proof} \subsubsection{Proof of \textbf{Assertion}}\label{proofAssertion} We will prove (\ref{eq3.9}) by contradiction arguments in the following five steps. We assume that (\ref{eq3.9}) is not valid. Then there exists a sequence $\{i_m\in \mathbb{N}\}$ with $\lim\limits_{m\rightarrow \infty} i_m=\infty$ and $R>0$ such that there exists $l\in\{0, 1, 2, 3\}$ satisfying \begin{equation*} \left\|\frac{\partial^l_r\psi^{i_m}}{\Phi}\right\|_{L^\infty\left(A^{\left(s_{i_m}, t_{i_m}\right)}_{R}\right)}>0. \end{equation*} Without loss of the generality, we assume that $$ \Big\|\frac{\psi^{i_m}}{\Phi}\Big\|_{L^\infty\left( A^{\left(s_{i_m}, t_{i_m}\right)}_{R}\right)}>0. $$ For the other situations, the following arguments are similar. Let $(\hat{t}_{i_m}, r_{i_m})\in A^{(s_{i_m}, t_{i_m})}_{R}=\big\{(t, r)\in (s_{i_m}, t_{i_m})\times \R: \abs{r-\gamma_{n}(t)}<R+1\big\}$ such that there exist $\delta>0$ and \begin{equation}\label{eq3.12} \abs{\frac{\psi^{i_m}(\hat{t}_{i_m}, r_{i_m})}{\Phi(\hat{t}_{i_m}, r_{i_m})}}>\delta>0. \end{equation} Let us introduce the following change of variables \begin{equation*} \nu=t-\hat{t}_{i_m},\quad y=r-y_{i_m}-\rho(t+\hat{t}_{i_m})\quad \text{and}\quad y_{i_m}=r_{i_m}-\gamma_{n}(\hat{t}_{i_m}), \end{equation*} and set \begin{equation}\label{eq3.14} \widehat{\psi}^{i_m}(\nu, y):=\frac{\psi_{i_m}\big(\nu+\hat{t}_{i_m},\, y+y_{i_m}+\rho(\nu+\hat{t}_{i_m})\big)\,} {\Phi\big(\hat{t}_{i_m}, y_{i_m} +\rho(\hat{t}_{i_m})\big)}, \end{equation} where we recall that $\rho(t)=\gamma_n(t)+h(t)$ defined in \eqref{drho1}. Hence, there exists $y_0$ such that $$ \lim\limits_{m\rightarrow\infty}y_{i_m}=y_0 \quad \mbox{and}\quad \abs{y_0}<R+2. $$ According to (\ref{eq3.5}), we have that $\widehat{\psi}^{i_m}$ satisfies the following problem \begin{align} \widehat{\psi}^{i_m}_\nu =&-\widehat{\psi}^{i_m}_{yyyy} -\frac{2(n-1)}{y+y_{i_m}+\rho(\nu+\hat{t}_{i_m})}\widehat{\psi}^{i_m}_{yyy} +\Bigg[ 2W''\big(\hat{z}(\nu, y)\big)-\frac{(n-3)(n-1)}{ \big(y+y_{i_m}+\rho(\nu+\hat{t}_{i_m})\big)^2}\Bigg] \widehat{\psi}^{i_m}_{yy} \nonumber\\[2mm] &-\Big(W''\big(\hat{z}(\nu, y)\big)\Big)^2\widehat{\psi}^{i_m} +\Bigg[\frac{2(n-1)W''\big(\hat{z}(\nu, y)\big)}{y+y_{i_m}+\rho(\nu+\hat{t}_{i_m})}-\frac{(3-n)(n-1)}{\big(y+y_{i_m}+\rho(\nu+\hat{t}_{i_m})\big)^3}\Bigg]\widehat{\psi}^{i_m}_{y} \nonumber\\[2mm] &+2W'''\big(\hat{z}(\nu, y)\big)\hat{z}_{yy}\widehat{\psi}^{i_m} +2\frac{n-1}{y+y_{i_m}+\rho(\nu+\hat{t}_{i_m})}W'''\big(\hat{z}(\nu, y)\big)\hat{z}_y\widehat{\psi}^{i_m} \nonumber\\[2mm] &+2W'''\big(\hat{z}(\nu, y)\big)\hat{z}_y\widehat{\psi}^{i_m}_y -W'''\big(\hat{z}(\nu, y)\big)W'\big(\hat{z}(\nu, y)\big)\widehat{\psi}^{i_m} +W^{(4)}\big(\hat{z}(\nu, y)\big)\abs{\hat{z}_y}^2 \widehat{\psi}^{i_m} \nonumber\\[2mm] &+\frac{g^{i_m}\big(\nu+\hat{t}_{i_m}, \, y+y_{i_m}+\rho(\nu+\hat{t}_{i_m})\big)} {\Phi\big(\hat{t}_{i_m}, y_{i_m}+\rho(\hat{t}_{i_m})\big)} -\partial_\nu \rho(\nu+\hat{t}_{i_m})\widehat{\psi}^{i_m}_{y} \qquad \text{in}\ \Gamma^{s_{i_m}, t_{i_m}}, \label{eq3.10} \end{align} with the conditions \begin{align} \abs{\widehat{\psi}^{i_m}(0, 0)}\geq\delta, \quad \widehat{\psi}^{i_m}(s_{i_m}-\hat{t}_{i_m}, y)=0, \quad y\in\big(-y_{i_m}-\rho(\nu+\hat{t}_{i_m}), +\infty\big), \label{eq3.10-111} \end{align} where $$ \hat{z}(\nu, y)=z\big(\nu+\hat{t}_{i_m}, y+y_{i_m}+\rho(\nu+\hat{t}_{i_m})\big), $$ and $$ \Gamma^{s_{i_m}, t_{i_m}}:=\Big\{(\nu, y)\in \big(s_{i_m}-\hat{t}_{i_m},\, 0\big]\times \big(-y_{i_m}-\rho(\nu+\hat{t}_{i_m}),\, +\infty\big)\Big\}. $$ As the same as previous part, $\lim\limits_{m\rightarrow\infty}(s_{i_m}-\hat{t}_{i_m})=\hat{\varsigma}$, where $\hat{\varsigma}$ equal to $-\infty$ or a negative number. \textbf{Step} $\mathbf{1}$. We first consider the case: $\lim\limits_{m\rightarrow\infty}(s_{i_m}-\hat{t}_{i_m})=-\infty$. We have that $\widehat{\psi}^{i_m}\rightarrow\widehat{\psi}$ locally uniformly, $\abs{\widehat{\psi}(0, 0)}>\delta>0$, which $\widehat{\psi}$ satisfies the following equation \begin{equation}\label{eq3.13} \begin{aligned} \widehat{\psi}_\nu=&-\widehat{\psi}_{yyyy}+2W''\big(\omega(y+y_0)\big)\widehat{\psi}_{yy} +2W'''\big(\omega(y+y_0)\big)\omega'(y+y_0)\widehat{\psi}_y -\Big[W''\big(\omega(y+y_0)\big)\Big]^2\widehat{\psi} \\[3mm] &+\Big[W''''\big(\omega(y+y_0)\big)\big(\omega'(y+y_0)\big)^2 +W'''\big(\omega(y+y_0)\big)\omega''(y+y_0)\Big]\widehat{\psi}, \qquad \text{in} \ (-\infty, 0]\times\R. \end{aligned} \end{equation} According to (\ref{dpsi}), (\ref{eq3.8}) and (\ref{eq3.14}), by the definition of $\rho$ in \eqref{drho1}, similar as \eqref{esp}, we have \begin{align}\label{eq3.15} \nonumber\abs{\widehat{\psi}^{i_m}(\nu, y)}=&\abs{\frac{\psi_{i_m}\big(\nu+\hat{t}_{i_m}, y+y_{i_m}+\rho(\nu+\hat{t}_{i_m})\big)} {\Phi\big(\hat{t}_{i_m}, y_{i_m}+\rho(\hat{t}_{i_m})\big)}} \\[2mm]\nonumber \leq&\abs{\frac{\Phi\big(\nu+\hat{t}_{i_m}, y+y_{i_m}+\rho(\nu+\hat{t}_{i_m})\big)} {\Phi\big(\hat{t}_{i_m}, y_{i_m}+\rho(\hat{t}_{i_m})\big)}} \\[2mm] \leq &C\big(\alpha, n, R, \|h\|_{L^\infty}\big) \big(1+\abs{y}\big)^p, \qquad \forall\, (\nu, y)\in B_{\hat{t}_{i_m}, j}, \end{align} for all $j=0, 1$, where $\rho(t)=\gamma_n(t)+h(t)$ in \eqref{drho1} and the sets $B_{\hat{t}_{i_m}, j}$ are defined as the following \begin{align*} B_{\hat{t}_{i_m}, 0}&:=\Big\{(\nu, y)\in (s_{i_m}-\hat{t}_{i_m}, 0]\times\R: 0 < y+y_{i_m}+\rho(\nu+\hat{t}_{i_m})\leq \delta_0\Big\}, \\[2mm] B_{\hat{t}_{i_m}, 1}&:=\Big\{(\nu, y)\in (s_{i_m}-\hat{t}_{i_m}, 0]\times\R: \delta_0 \leq y+y_{i_m}+\rho(\nu+\hat{t}_{i_m})<+\infty\Big\}. \end{align*} We notice that \begin{equation}\label{eq3.17} \Gamma^{s_{i_m}, t_{i_m}}=\cup_{l=0}^{1}B_{\hat{t}_{i_m}, j} =\big(s_{i_m}-\hat{t}_{i_m}, 0\big]\times\big(-y_{i_m}-\rho(\nu+\hat{t}_{i_m}),\, +\infty\big). \end{equation} Similarly, we have that \begin{equation}\label{f4} \sum_{l=1}^3\abs{\partial_y^l\widehat{\psi}^{i_m}(\nu, y)}\leq C (1+\abs{y})^p, \end{equation} for all $(\nu, y)\in \Gamma^{s_{i_m}, t_{i_m}}$, where $p\in(n, n+1)$ and $C$ depends on $ \alpha, R$, $n$ and $\|h\|_{L^\infty}$. Since $\gamma_{n}(\nu+\hat{t}_{i_m})\rightarrow+\infty$ as $i$ goes to infinity, by (\ref{f4}), (\ref{eq3.10})-\eqref{eq3.10-111}, and $y_{i_m}\rightarrow y_0$, we can get the limiting equation in \eqref{eq3.13}. If $\hat{\varsigma}\in (-\infty, 0)$, we have $\hat{\psi}(\hat{\varsigma}, y)=0$. According to Proposition \ref{prop8}, we have that equation \eqref{eq3.13} has a unique solution $\hat{\psi}(\nu, y)\equiv0$, which contradicts with $\abs{\hat{\psi}(0, 0)}>0$. Hence, it only happens that $\hat{\varsigma}=-\infty$. Thus, we get the desired result. \textbf{Step} $\mathbf{2}$. We will prove the following orthogonality condition for $\widehat{\psi}$ \begin{equation}\label{eq3.16} \int_{\R}\widehat{\psi}(\nu, y)\omega'(y+y_0){\mathrm d}y=0, \qquad \text{for all}\ \nu\in(-\infty, 0]. \end{equation} In fact, according to (\ref{eq3.6}) and (\ref{eq3.14}), we get that \begin{align*} 0&= \frac{\left[y_{i_m}+\rho(\nu+\hat{t}_{i_m})\right]^{1-n}} {\Phi\big(\hat{t}_{i_m}, y_{i_m}+\rho(\hat{t}_{i_m})\big)} \int_0^\infty \psi^{i_m}(\nu+\hat{t}_{i_m}, r)\omega'\big(r-\rho(\nu+\hat{t}_{i_m})\big)r^{n-1}{\mathrm d}r \\[2mm] &=\int_{-y_{i_m}-\rho(\nu+\hat{t}_{i_m})}^\infty\widehat{\psi}^{i_m}(\nu, y)\omega'(y+y_{i_m}) \left[\frac{y+y_{i_m}+\rho(\nu+\hat{t}_{i_m})} {y_{i_m}+\rho(\nu+\hat{t}_{i_m})}\right]^{n-1}{\mathrm d}y, \end{align*} where $\rho(t)=\gamma_n(t)+h(t)$ with $\|h(t)\|_{L^\infty}<1$ and the function $\omega$ is given by \eqref{eqq}. And using (\ref{eq3.15}), (\ref{eq3.17}) and Lemma \ref{lem1}, we also get that \begin{align*} &\abs{\widehat{\psi}_{i_m}(\nu, y)\omega'(y+y_{i_m})}\leq C(R, n, \alpha, \|h\|_{L^\infty})(1+\abs{y})^p \exp\big\{-\alpha\abs{y+y_{i_m}}\big\}. \end{align*} Since $y_{i_m}\rightarrow y_0$ and $\hat{t}_{i_m}\rightarrow-\infty$, as $m$ goes to infinity, using Lebesgue Dominated Convergence Theorem, we can obtain that (\ref{eq3.16}) holds. \textbf{Step} $\mathbf{3}$. In this step, we will prove the following decay of $\widehat{\psi}(\nu, y)$: there exists a constant $C$ of the form $$ C=C\big(\alpha, n, R, \|h\|_{L^\infty}\big)>0 $$ such that \begin{equation}\label{eq3.18} \abs{\widehat{\psi}(\nu, y)}\leq \frac{C}{\left(1+\abs{y}\right)^2}, \qquad \forall\, (\nu, y)\in (-\infty, 0]\times\R. \end{equation} In fact, for any $(\nu, y)\in B_{t_{i_m}, j}$, by the definition of $\rho(t)$ in \eqref{drho1}, in view of the proof of (\ref{eq3.15}), we have \begin{equation}\label{f3} \abs{\frac{g^{i_m}\big(\nu+\hat{t}_{i_m}, y+y_{i_m}+\rho(\nu+\hat{t}_{i_m})\big)} {\Phi\big(\hat{t}_{i_m}, y_{i_m}+\rho(\hat{t}_{i_m})\big)}} \leq C \|g^{i_m}\|_{C_{\Phi}((s_{i_m},\, t_{i_m})\times(0,+\infty))}(1+\abs{y})^p, \ \end{equation} for all $(\nu, y)\in \Gamma^{s_{i_m}, t_{i_m}}$ given by \eqref{eq3.17}, where $C$ depends on $\alpha, R, n$ and $\|h\|_{L^\infty}$. By formula \eqref{f1}, the solution of equation \eqref{eq3.10} has the form \begin{equation}\label{f2} \widehat{ \psi}_{i_m}(\nu, y)=\int_{0}^{\nu-s_{i_m}+\hat{t}_{i_m}}\int_{\R}e^{-\alpha^4\tau} \mathbf{Q}(\tau, x)\hat{f}_{i_m}\big(\hat{\psi}^{i_m}(\nu-\tau, y-x), \nu-\tau, y-x\big){\mathrm d}x{\mathrm d}\tau, \end{equation} for all $(\nu, y)\in(s_{i_m}-\hat{t}_{i_m}, 0)\times\R$, where $\alpha=\sqrt{W''(1)}$ and the function $\hat{f}_{i_m}$ is given by \begin{align*} \hat{f}_{i_m}=&2\Big[W''\big(\hat{z}(\nu, y)\big)-W''(1)\Big]\widehat{\psi}^{i_m}_{yy} +2W'''\big(\hat{z}(\nu, y)\big)\partial_y\hat{z}(\nu, y)\widehat{\psi}^{i_m}_y -\Big(\big[W''\big(\hat{z}(\nu, y)\big)\big]^2-\big[W''(1)\big]^2\Big)\widehat{\psi}^{i_m} \\[3mm] &+\Big\{\partial_{yy}W''\big(\hat{z}(\nu, y)\big)+W'''\big(\hat{z}(\nu, y)\big)\big[\partial_{yy}\hat{z}(\nu, y)-W'\big(\hat{z}(\nu, y)\big)\big]\Big\}\widehat{\psi}^{i_m} \\[3mm] &-\frac{2(n-1)}{y+y_{i_m}+\rho(\nu+\hat{t}_{i_m})}\widehat{\psi}^{i_m}_{yyy} -\frac{(n-3)(n-1)}{\big(y+y_{i_m}+\rho(\nu+\hat{t}_{i_m})\big)^2}\widehat{\psi}^{i_m}_{yy} \\[3mm] &+\Bigg[\frac{2(n-1)W''\big(\hat{z}(\nu, y)\big)}{y+y_{i_m}+\rho(\nu+\hat{t}_{i_m})} -\frac{(3-n)(n-1)}{\big(y+y_{i_m}+\rho(\nu+\hat{t}_{i_m})\big)^3}\Bigg]\widehat{\psi}^{i_m}_{y} \\[3mm] &+2\frac{n-1}{y+y_{i_m}+\rho(\nu+\hat{t}_{i_m})}W'''\big(\hat{z}(\nu, y)\big)\hat{z}_y\widehat{\psi}^{i_m} +\frac{g^{i_m}\big(\nu+\hat{t}_{i_m}, y+y_{i_m}+\rho(\nu+\hat{t}_{i_m})\big)} {\Phi\big(\hat{t}_{i_m}, y_{i_m}+\rho(\hat{t}_{i_m})\big)} \\[3mm] &-\partial_\nu \rho(\nu+\hat{t}_{i_m})\widehat{\psi}^{i_m}_{y} \end{align*} and $\hat{z}(\nu, y)=z\big(\nu+\hat{t}_{i_m}, y+y_{i_m}+\rho(\nu+\hat{t}_{i_m})\big)$. According to the lemma \ref{lem1}, \eqref{eq3.3}, \eqref{f3}, \eqref{eq3.15} and \eqref{f4}, we have that \begin{align*} \abs{\hat{f}_{i_m}}\leq &C(1+\abs{y})^p\Bigg[\exp\big\{-\alpha\abs{y+y_{i_m}}\big\} +\|g^{i_m}\|_{C_{\Phi}((s_{i_m}, t_{i_m})\times(0,+\infty))} \\[2mm] &\qquad\qquad+\frac{1}{\left[-(\nu+\hat{t}_{i_m})\right]^{3/4}}+\sum_{l=1}^3\left(y+y_{i_m}+\rho(\nu+\hat{t}_{i_m})\right)^{-l}\Bigg] \overline{\chi}_{\left\{y>\frac{\delta_0}{2}-\rho(\nu+\hat{t}_{i_m})-y_{i_m}\right\}}, \end{align*} for all $(\nu, y)\in \big(s_{i_m}-\hat{t}_{i_m}, 0\big)\times(-\rho(\nu+\hat{t}_{i_m})-y_{i_m},+\infty)$, where $C$ depends on $\alpha, R, n$ and $\|h\|_{L^\infty}$, and $\overline{\chi}_{A}$ is the characteristic function of the set $A$. Thus by \eqref{f2}, similar arguments as \eqref{13} and $\abs{y_{i_m}}\leq R+2$, let $m$ goes to infinity, and then we can get the desired result since $\|g^{i_m}\|_{C_{\Phi}((s_{i_m}, \bar{t}_{i_m})\times(0,+\infty))}\rightarrow0$ as index $m$ goes to infinity in \eqref{eq8} and the following fact that \begin{align*} &\int_{0}^{\nu+\check{t}_{i_m}}\int_{\R}e^{-\alpha^4\tau}\abs{\mathbf{Q}(\tau, y-x)}\big(1+\abs{x}\big)^p \exp\big\{-\alpha\abs{x+y_{i_m}}\big\}{\mathrm d}x{\mathrm d}\tau \\[2mm] &\leq C\int_{0}^{\nu+\check{t}_{i_m}}\int_{\R}\exp\left\{-\alpha^4\tau-\frac{1}{4}\abs{x}\right\} \exp\left\{-\frac{\alpha}{2}\abs{y-\tau^{1/4}x}\right\}{\mathrm d}x{\mathrm d}\tau \\[2mm] &\leq C\int_{0}^{\nu+\check{t}_{i_m}}\int_{\R}\exp\left\{-\alpha^4\tau-\frac{1}{4}\abs{x}\right\} \frac{1}{\left(1+\abs{y-\tau^{1/4}x}\right)^2}{\mathrm d}x{\mathrm d}\tau \\[2mm] &\leq C\int_{0}^{\nu+\check{t}_{i_m}}\int_{\R}\exp\left\{-\alpha^4\tau-\frac{1}{4}\abs{x}\right\} \left(\frac{1+\abs{\tau^{1/4}x}}{1+\abs{y}}\right)^2{\mathrm d}x{\mathrm d}\tau \\[2mm] & \leq \frac{C}{\left(1+\abs{y}\right)^2}, \end{align*} where $\check{t}_{i_m}:=\hat{t}_{i_m}-s_{i_m}$ and $C>0$ only depends on $\alpha$. \textbf{Step} $\mathbf{4}$. To proceed further, we need the following result. However, we can not find the proof of this result in references. Here we give a proof in details. \begin{lem}\label{lem2} Considering the Hilbert space $$H=\left\{\phi(y)\in H^2(\R):\int_{\R}\phi(y)\omega'(y){\mathrm d}y=0\right\}, $$ then the following inequality is valid \begin{equation}\label{eqi} \int_\R\abs{\phi''(y)-W''(\omega)\phi(y)}^2{\mathrm d}y\geq c\int_\R\abs{\phi(y)}^2{\mathrm d}y, \qquad \forall\, \phi\in H, \end{equation} where $c>0$. \end{lem} \begin{proof} For any $\phi\in H$, since $\omega'>0$, we can set $\phi(y)=\zeta_1(y)\omega'(y)$, then \begin{align*} \int_\R\abs{\phi''(y)-W''(\omega)\phi(y)}^2{\mathrm d}y=&\int_\R\abs{\omega'''(y)\zeta_1(y)+2\zeta'_1(y)\omega''(y)+\zeta''_1(y)\omega'(y) -W''(\omega)\zeta_1(y)\omega'(y)}^2{\mathrm d}y \\[2mm] =&\int_\R\abs{2\omega''(y)\zeta_1'(y)+\omega'(y)\zeta_1''(y)}^2{\mathrm d}y\geq 0, \end{align*} where we have used $\omega'''=W''(\omega)\omega'$ and integration by parts in the last equality. Hence we have \begin{equation*} \int_\R\abs{\phi''(y)-W''(\omega)\phi(y)}^2{\mathrm d}y=0\quad \text{if and only if}\quad \ 2\omega''(y)\zeta_1'(y)+\omega'(y)\zeta_1''(y)=0. \end{equation*} All solutions of the last equation have the following form $$ \zeta_1(x)= c_1\int_0^x\frac{1}{(\omega'(y))^2}{\mathrm d}y+c_2, \qquad c_1, c_2\in\R. $$ Since $$ \phi(x)=\zeta_1(x)\omega'(x)\in H^2(\R)\qquad \mbox{and}\qquad \int_{\R}\phi(y)\omega'(y){\mathrm d}y=0, $$ we have that $c_1=0$ and $c_2=0$. Now we assume that there exists a sequence $\{\phi_m\}\in H$ such that \begin{equation}\label{esi} \int_{\R}\abs{\phi_m}^2{\mathrm d}y=1\ \ \text{and}\ \ \int_{\R}\abs{\phi''_m-W''(\omega)\phi_m}^2{\mathrm d}y\leq\frac{1}{m}. \end{equation} Thus, $\phi_m\rightharpoonup\phi$ in $H(\R)$ and $\phi_m\rightarrow\phi$ in $L^2(K)$ for any compact subset $K\subset\R$, which implies that $$0=\int_{\R}\phi_m(y)\omega'(y){\mathrm d}y\rightarrow\int_{\R}\phi\omega'{\mathrm d}y=0, $$ for some $\phi\in H(\R)$. By Fatou's Lemma, we have \begin{equation*} \int_{\R}\abs{\phi''-W''(\omega)\phi}^2{\mathrm d}y=0. \end{equation*} Hence $\phi=0$. On the other hand, we first have \begin{align*} \big|W''(1)\big|^2\int_{\R}\abs{\phi_m}^2{\mathrm d}y&\leq\int_{\R} \big|\phi''_m-W''(1)\phi_m\big|^2{\mathrm d}y \\[2mm] &\leq 2\Bigg\{\int_{\R} \big|\phi''_m-W''(\omega)\phi_m\big|^2{\mathrm d}y +\int_{\R} \Big|\big[W''(1)-W''(\omega)\big]\phi_m\Big|^2{\mathrm d}y\Bigg\}. \end{align*} Thus, by \eqref{esi}, we get that $$ \big|W''(1)\big|^2\leq \int_{\R}\Big|\big[W''(1)-W''(\omega)\big]\phi\Big|^2{\mathrm d}y. $$ According to the assumptions on $W$ in \eqref{eq1.4}, $W''(1)\neq 0$. The above last inequality implies that $\phi\neq0$, which is a contradiction. \end{proof} \textbf{Step} $\mathbf{5}$. We will prove that \eqref{eq3.9} holds. Multiplying (\ref{eq3.13}) by $\hat{\psi}(y)$ and integrating with respect to $y$, and using \eqref{eqi}, we have \begin{align*} 0=&\frac{1}{2}\int_\R \big(\widehat{\psi}^2\big)_\nu{\mathrm d}y+\int_\R\abs{\widehat{\psi}_{yy}-W''\big(\omega(y+y_0)\big)\widehat{\psi}(\nu, y)}^2{\mathrm d}y \\[2mm] \geq& \frac{1}{2}\int_\R \big(\widehat{\psi}^2\big)_\nu{\mathrm d}y+c\int_\R\abs{\widehat{\psi}(\nu, y)}^2{\mathrm d}y, \end{align*} where the constant $c>0$. Then, according to the Gronwall inequality, we get that \begin{equation*} a(\nu)\geq a(0)e^{-2c\nu}, \ \ \ \ \forall\, \nu<0,\ \end{equation*} where the function $$ a(\nu):=\int_\R\abs{\widehat{\psi}(\nu, y)}^2{\mathrm d}y. $$ This is a contradiction since (\ref{eq3.18}). The proof of \eqref{eq3.9} is completed. \bigskip \subsection{The projected linear problem involving the operator $L$} We consider the following projection problem: \begin{equation}\label{eq3.19} \left\{ \begin{array}{l} \psi_t=L[\psi]+g(t, r)-c(t)\partial_r\widehat{\omega}(t, r)\qquad \text{in}\ (s, -T]\times(0, \infty), \\[2mm] \psi(s, r)=0, \qquad \forall\, r\in (0, +\infty), \\[2mm] \partial_r^j\psi(t, 0)=0, \qquad \forall\, t\in (s, -T], \ \ j=1, 3, \end{array} \right. \end{equation} where the linear operator $L[\psi]=F'\big(z(t, r)\big)[\psi] $ is defined by \eqref{eq2.13}, the function $g\in C_{\Phi}((s, -T)\times (0, +\infty))$ and $c(t)$ satisfies the following relation \begin{align} &c(t)\int_{0}^\infty\partial_r\widehat{\omega}(t, r)\omega'\big(r-\rho(t)\big)r^{n-1}{\mathrm d}r \nonumber\\[3mm] &= \int_{0}^\infty \Big[\omega'''\big(r-\rho(t)\big)+\frac{n-1}{r}\omega''\big(r-\rho(t)\big) -W''\big(z(t, r)\big)\omega'\big(r-\rho(t)\big)\Big] \nonumber\\[3mm] &\qquad\qquad\times\left[-\psi_{rr}-\frac{n-1}{r}\psi_r+W''\big(z(t, r)\big)\psi\right] r^{n-1}{\mathrm d}r \nonumber\\[3mm] &\quad+\int_{0}^\infty\left[\partial_{rr}z(t, r) +\frac{n-1}{r}\partial_{r}z(t, r)-W'\big(z(t, r)\big)\right]\psi \omega'\big(r-\rho(t)\big)r^{n-1}{\mathrm d}r \nonumber\\[3mm] &\quad+\int_{0}^\infty \psi(t, r)\partial_t\big[\omega'\big(r-\rho(t)\big)\big]r^{n-1}{\mathrm d}r \,+\, \int_{0}^\infty g(t, r)\omega'\big(r-\rho(t)\big) r^{n-1}{\mathrm d}r, \label{eq3.20} \end{align} for all $t\in (s, -T]$, where the functions $\widehat{\omega}(t, r)$ and $z(t, r)$ are defined by \eqref{eq2.3} and \eqref{dz} respectively. If $\psi$ is a solution of (\ref{eq3.19}) and $c(t)$ satisfies (\ref{eq3.20}), integration by parts will imply that $\psi$ satisfies the following the orthogonality condition \begin{equation}\label{eq3.21} \int_0^\infty \psi(t, r)\omega'\big(r-\rho_i(t)\big) r^{n-1}{\mathrm d}r=0, \qquad \forall\, t\in (s, -T). \end{equation} For (\ref{eq3.20}), we have the following result: \begin{lem}\label{lem6} Let $T>0$ big enough and $\psi, \psi_r, \psi_{rr}$ and $g\in C_{\Phi}((s, -T)\times (0, +\infty))$. Then there exists $c(t)$ such that $(\ref{eq3.20})$ holds. Furthermore the following estimates are valid \begin{align} \abs{c(t)}\leq \frac{C_0}{\big[\log\abs{t}\big]^{p-1}\abs{t}^{1/2}}\left\{\frac{1}{\abs{t}^{1/4}}\left[\sum_{l=0}^2 \big\|\partial_r^l\psi\big\|_{C_{\Phi}((s, -T)\times(0, \infty))}\right] \ +\ \|g\|_{C_{\Phi}((s, -T)\times(0, \infty))}\right\}, \label{estimateofc1} \end{align} and \begin{align} \abs{\frac{c(t)\partial_r\widehat{\omega}(t, r)}{\Phi(t, r)}} \,\leq\, &\frac{C_0}{\abs{t}^{1/4}} \sum_{l=0}^2\big\|\partial_r^l\psi\big\|_{C_{\Phi}((s, -T)\times(0, \infty))} \ +\ C_0\|g\|_{C_{\Phi}((s, -T)\times(0, \infty))}, \label{estimateofc2} \end{align} for any $t\in[s, -T]$, where the function $\widehat{\omega}(t, r)$ is given by \eqref{eq2.3}, $p\in(n, n+1]$ and $C_0$ is a positive constant which does not depend on $s, t, T$, $\psi$ and $g$. \end{lem} \begin{proof} We first consider the left hand side of system $(\ref{eq3.20})$. By Lemma \ref{lem1} and the definition of $\widehat{\omega}(t, r)$ in \eqref{eq2.3}, we have that \begin{equation*} \begin{aligned} \int_{0}^\infty\partial_r\widehat{\omega}(t, r)\omega'\big(r-\rho(t)\big)r^{n-1}{\mathrm d}r&= \int_{\delta_0-\rho(t)}^\infty\omega'(r)\omega'(r)\big(r+\rho(t)\big)^{n-1}{\mathrm d}r \ +\ O\left([\gamma_n(t)]^{n-2}\right) \\[3mm] &=\left(\rho(t)\right)^{n-1}\int_\R\big[\omega'(y)\big]^2dy \ +\ O\left([\gamma_n(t)]^{n-2}\right), \end{aligned} \end{equation*} where the functions $\rho(t)$ and $\gamma_n(t)$ are given by \eqref{drho1} and \eqref{dgamma-n} respectively. For the first and second terms in the right hand side of \eqref{eq3.20}, we can obtain that \begin{align*} I(t):=&\int_{0}^\infty\Big[-\omega'''\big(r-\rho_i(t)\big)-\frac{n-1}{r}\omega''\big(r-\rho_i(t)\big) +W''\big(z(t, r)\big)\omega'\big(r-\rho(t)\big)\Big] \\[3mm] &\qquad\qquad \times \left(\psi_{rr}+\frac{n-1}{r}\psi_r-W''\big(z(t, r)\big)\psi\right) r^{n-1} {\mathrm d}r \\[3mm] &+\int_{0}^\infty\left[\partial_{rr}z(t, r) +\frac{n-1}{r}\partial_{r}z(t, r)-W'\big(z(t, r)\big)\right]\psi \omega'\big(r-\rho(t)\big)r^{n-1}{\mathrm d}r \\[3mm] \leq &C\sum_{l=0}^2\big\|\partial_r^l\psi\big\|_{C_{\Phi}((s, -T)\times(0, \infty))} \int_{0}^\infty\Bigg[\abs{\omega'''\big(r-\rho(t)\big) -W''\big(z(t, r)\big)\omega'\big(r-\rho(t)\big)} \\[3mm] &\qquad\qquad\qquad\qquad+\abs{\omega'\big(r-\rho(t)\big)}\abs{\partial_{rr}z(t, r)-W'\big(z(t, r)\big)}\Bigg]\Phi(t, r)r^{n-1}{\mathrm d}r \\[3mm] & \,+\, C\sum_{l=0}^2\big\|\partial_r^l\psi\big\|_{C_{\Phi}((s, -T)\times(0, \infty))}\int_{0}^\infty\Phi(t, r) \sum_{l=1}^3\abs{\omega^{(l)}\big(r-\rho(t)\big)}r^{n-2}{\mathrm d}r, \end{align*} due to the facts that the functions $\psi, \psi_r, \psi_{rr}$ belong to $C_{\Phi}((s, -T)\times (0, +\infty))$. By similar arguments in the proof of Lemma \ref{lem10} and the definition of $\Phi(t, r)$ in \eqref{dpsi}, we have that \begin{align*} I(t)&\leq C\sum_{l=0}^2\big\|\partial_r^l\psi\big\|_{C_{\Phi}((s, -T)\times(0, \infty))} \Bigg\{ \int_{0}^\infty\Phi(t, r) \sum_{l=1}^3\abs{\omega^{(l)}\big(r-\rho(t)\big)}r^{n-2}{\mathrm d}r \\[2mm] &\qquad\qquad +\int_{0}^\infty\Phi(t, r)\abs{\omega'\big(r-\rho(t)\big)}r^{n-3}{\mathrm d}r\Bigg\} \\[2mm] &\leq C\frac{\big[\rho(t)\big]^{n-1}}{\abs{t}^{3/4}\big[\log\abs{t}\big]^{p-1}} \left\{\sum_{l=0}^2\big\|\partial_r^l\psi\big\|_{C_{\Phi}((s, -T)\times(0, \infty))}\right\}. \end{align*} Similarly, by the definitions in \eqref{drho1} and \eqref{dpsi} again, we have that \begin{align*} \int_{0}^\infty \psi(t, r)\partial_t\big[\omega'\big(r-\rho(t)\big)\big]r^{n-1}{\mathrm d}r &= -\rho'(t)\int_{0}^\infty \psi(t, r)\omega''\big(r-\rho(t)\big)r^{n-1}{\mathrm d}r \\[2mm] &=O\left\{\frac{\big[\rho(t)\big]^{n-1}}{\abs{t}^{5/4}}\right\}\|\psi\|_{C_{\Phi}((s, -T)\times(0, \infty))}. \end{align*} And, there holds that \begin{align*} \abs{\int_{0}^\infty g(t, r)\omega'\big(r-\rho(t)\big)r^{n-1}{\mathrm d}r} \leq & C\int_{0}^\infty \Phi(t, r)\abs{\omega'\big(r-\rho(t)\big)}r^{n-1}{\mathrm d}r \\[2mm] \leq & C\frac{\big[\rho(t)\big]^{n-1}}{\abs{t}^{1/2}\big[\log\abs{t}\big]^{p-1}}. \end{align*} According the above estimates and \eqref{eq3.20}, we can get the validity of \eqref{estimateofc1}. Using the following fact \begin{equation*} \abs{\frac{\partial_r\widehat{\omega}(t, r)}{\Phi(t, r)}}\leq C\abs{t}^{1/2}\big[\log\abs{t}\big]^{p-1}, \qquad \text{for all} \ r>0, \end{equation*} where $\widehat{\omega}(t, r)$ is defined by \eqref{eq2.3} and $C$ is a constant which does not depend on $t$, we can obtain \eqref{estimateofc2}. \end{proof} According to Lemma \ref{lem6}, we have \begin{lem}\label{lem7} Let $g\in C_{\Phi}((s, -T)\times (0,\infty))$, then there exists a uniform constant $T_0>0$ and a unique solution $\psi^s$ of problem $(\ref{eq3.19})$. Moreover, $\psi^s$ satisfies the orthogonality conditions in $(\ref{eq3.21})$ with $t\in(s, -T_0)$, and \begin{equation}\label{eq3.27} \sum_{l=0}^3\big\|\partial_r^l\psi^s\big\|_{C_{\Phi}((s, t)\times(0, \infty))}\leq C\|g\|_{C_{\Phi}((s, t)\times(0, \infty))}, \end{equation} where $t\in(s, -T_0)$ and $C$ is a uniform constant which does not depend on $t, s$. \end{lem} \begin{proof} We will use a fixed-point argument to prove this lemma. Let $$ X^s:=\Big\{\psi \,:\, \sum_{l=0}^3\big\|\partial_r^l\psi\big\|_{C_{\Phi}((s, s+1)\times(0, \infty))}<+\infty\Big\}, $$ and $\mathcal{T}^s(g)$ be the solution of (\ref{eq3.5}) with $-T=s+1$. We consider the operator $\mathcal{A}^s: X^s\rightarrow X^s$ defined by $$ \mathcal{A}^s(\psi):=\mathcal{T}^s(g-C(\psi)), \qquad C(\psi)=c(t)\partial_r\widehat{\omega}(t, r), $$ where $\widehat{\omega}(t, r)$ is given by \eqref{eq2.3} and the function $c(t)$ satisfies (\ref{eq3.20}). Hence $c(t)$ depends on $\psi$, and $\psi$ satisfies the orthogonality conditions in \eqref{eq3.21}. By Lemma \ref{lem5}, we have that \begin{equation}\label{eq3.26} \sum_{l=0}^3\big\|\partial_r^l\mathcal{A}^s(\psi)\big\|_{C_{\Phi}((s, s+1)\times(0, \infty))}\leq C_1\|g-C(\psi)\|_{C_{\Phi}((s, s+1)\times(0, \infty))}, \end{equation} for a uniform constant $C_1>0$. Set $$c:=(C_0+C_1)\|g\|_{C_{\Phi}((s, s+1)\times(0, \infty))}$$ and $$ X^s_c:=\big\{\psi:\|\psi\|_{C_{\Phi}((s, s+1)\times(0, \infty))}<2c\big\}, $$ where $C_0$ and $C_1$ are given by Lemma \ref{lem6} and (\ref{eq3.26}) respectively. It is easy to derive the following two facts: \smallskip \noindent{\textbf{(1).}} By (\ref{eq3.26}) and Lemma \ref{lem6}, we have \begin{equation*} \begin{aligned} \sum_{l=0}^3\big\|\partial^l_r\mathcal{A}^s(\psi)\big\|_{C_{\Phi}((s, s+1)\times(0, \infty))} \leq & C_0\Big(\|C(\psi)\|_{C_{\Phi}((s, s+1)\times(0, \infty))}+\|h\|_{C_{\Phi}((s, s+1)\times(0, \infty))}\Big) \\ \leq&\frac{CC_0}{\sqrt[4]{\abs{s+1}}}\Big(\sum_{l=0}^2\big\|\partial_r^l\psi\big\|_{C_{\Phi}((s, s+1)\times(0, \infty))} \Big)+c, \end{aligned} \end{equation*} where $C$ is given by Lemma \ref{lem6}. \smallskip \noindent{\textbf{(2).}} For any $\psi_1, \psi_2\in X^s_c$, by (\ref{eq3.26}) and Lemma \ref{lem6} again, we have \begin{align*} \sum_{l=0}^3\big\|\partial^l_r\left[\mathcal{A}^s(\psi_1)-\mathcal{A}^s(\psi_2)\right]\big\|_{C_{\Phi}((s, s+1)\times(0, \infty))} \leq & C_0\,\|C(\psi_1)-C(\psi_2)\|_{C_{\Phi}((s, s+1)\times(0, \infty))} \\[2mm] \leq & C_0\,\|C(\psi_1-\psi_2)\|_{C_{\Phi}((s, s+1)\times(0, \infty))} \\[2mm] \leq&\frac{CC_0}{\sqrt[4]{\abs{s+1}}}\,\|\psi_1-\psi_2\|_{C_{\Phi}((s, s+1)\times(0, \infty))}, \end{align*} where $C_0$ is given by Lemma \ref{lem6}. Hence, taking $s$ large enough, we have that the operator $\mathcal{A}^s$ is a contraction map from $X^s_c$ to itself. According to the Banach Fixed Theorem, we know that there exists a unique $\psi^s\in X^s_c$ such that $\mathcal{A}^s(\psi^s)=\psi^s$, which is a solution to problem (\ref{eq3.19}) with $-T=s+1$. Next we will extend the solution $\psi^s(t, r)$ in $(s, s+1)\times(0,+\infty)$ to $(s, -T_0)\times(0,+\infty)$, which will satisfy the orthogonality condition in (\ref{eq3.20}) and the a priori estimate in Lemma \ref{lem5}. We choose $T_0$ large enough such that $\frac{CC_0}{\sqrt[4]{T_0}}<1$, where $C$ is given by Lemma \ref{lem6} and $C_0$ is given by (\ref{eq3.26}). Thus the above fixed-point argument can be repeated when $s+1\leq -T_0$. Hence passing finite steps of fixed-point arguments, the solution $\psi^s(t, r)$ can be extended to $(s, -T_0]$. Moreover the solution $\psi^s$ satisfies (\ref{eq3.27}) and the orthogonality condition. \end{proof} \subsection{The solvability of a linear projected problem} In this section, we devote to building the solvability of the following linear parabolic projected problem: \begin{equation}\label{eq3.1} \left\{ \begin{array}{l} \psi_t=L[\psi]+g(t, r)-c(t)\partial_r\widehat{\omega}(t, r) \qquad\text{in}\ (-\infty, -T]\times(0, \infty), \\[2mm] \int_{\R}r^{n-1}\psi(t, r)\omega'\big(r-\rho(t)\big){\mathrm d}r=0, \hspace{0.75cm} \text{for all}\ t\in (-\infty, -T], \end{array} \right. \end{equation} for a bounded function $g$, and $T>0$ fixed sufficiently large. In the above, the linear operator $L$ is given by \eqref{eq2.13}, the functions $\rho(t)$ and $\widehat{\omega}(t, r)$ are defined by (\ref{drho1}) and \eqref{eq2.3} respectively. The function $c(t)$ solves the following relation: \begin{align} &c(t)\int_{0}^\infty\partial_r\widehat{\omega}(t, r)\omega'\big(r-\rho(t)\big)r^{n-1}{\mathrm d}r \nonumber\\[3mm] &= \int_{0}^\infty \Big[\omega'''\big(r-\rho(t)\big)+\frac{n-1}{r}\omega''\big(r-\rho(t)\big) -W''\big(z(t, r)\big)\omega'\big(r-\rho(t)\big)\Big] \nonumber\\[3mm] &\qquad\qquad \times\left[-\psi_{rr}-\frac{n-1}{r}\psi_r+W''\big(z(t, r)\big)\psi\right]r^{n-1}{\mathrm d}r \nonumber\\[3mm] &\quad+\int_{0}^\infty\Bigg[\partial_{rr}z(t, r)+\frac{n-1}{r}\partial_{r}z(t, r)-W'\big(z(t, r)\big)\Bigg]\psi \omega'\big(r-\rho(t)\big)r^{n-1}{\mathrm d}r \nonumber\\[3mm] &\quad+\int_{0}^\infty \psi(t, r)\partial_t[\omega'\big(r-\rho(t)\big)]r^{n-1}{\mathrm d}r \,+\, \int_{0}^\infty g(t, r)\omega'\big(r-\rho(t)\big)r^{n-1}{\mathrm d}r, \label{eq3.2} \end{align} for all $t<-T$, where $z(t, r)$ is given by \eqref{dz}. Indeed, this can be solved uniquely since if $T$ is taken sufficiently large, the coefficient $\int_{0}^\infty\partial_r\widehat{\omega}(t, r)\omega'\big(r-\rho(t)\big)r^{n-1}{\mathrm d}r$ is strictly positive. The main result in this section is as follows: \begin{prop}\label{prop2} There exist positive constants $T_1$ and $C$ such that for any $t\leq -T_1$ and $g\in C_{\Phi}((-\infty, t)\times(0, \infty))$, problem \eqref{eq3.1}-\eqref{eq3.2} has a solution $\psi=\mathcal{A}(g)$, which defines a linear operator of $g$ and satisfies the following estimate \begin{equation}\label{eq3.4} \sum_{l=0}^3\big\|\partial^l_r\psi\big\|_{C_{\Phi}((-\infty, t)\times(0, \infty))} \leq C\|g\|_{C_{\Phi}((-\infty, t)\times(0, \infty))}, \qquad \text{for all}\ t\leq-T_1, \end{equation} where we denote the $l$-th order derivative of $\psi(t, r)$ with respect of $r$ by $\partial^l_r\psi$. \end{prop} $\mathbf{Proof \ of\ Proposition \ \ref{prop2} }$: We choose a sequence $s_j\rightarrow -\infty$. Let $\psi^{s_j}$ be the solution to problem (\ref{eq3.19}) with $s=s_j$, according to Lemma \ref{lem7}. By (\ref{eq3.27}), we can find that the sequence $\{\psi^{s_j}\}$ converges to $\psi$ (up to subsequence) locally uniformly in $(-\infty, -T_1)\times(0, \infty)$. Using (\ref{eq3.7}) and Proposition \ref{prop7}, we have that $\psi$ is a solution of (\ref{eq3.1}) and satisfies (\ref{eq3.4}). The proof is completed. \section{Solving the nonlinear projected problem} \label{sec:nl} In this section, we mainly solve the nonlinear problem (\ref{eq2.10})-(\ref{eq2.11}) by using the fixed-point arguments. According to the result in Proposition \ref{prop2}, $\phi$ is a solution of (\ref{eq2.10})-(\ref{eq2.11}) if only if $\phi\in C_{\Phi}((-\infty, -T_1)\times(0, \infty)) $ is a fixed point of the following operator \begin{equation}\label{eq4.2} \mathbf{ T}(\phi):=\mathcal{A}\big(E(t, r)+N(\phi)-c(t)\partial_r\widehat{\omega}(t, r)\big), \end{equation} where $T_1$ and $\mathcal{A}$ are given by Proposition \ref{prop2}. Let $T\geq T_1$, We define a set \begin{equation}\label{LambdaT} \Lambda_T:=\Big\{h\in C^1(-\infty, -T]:\ \|h\|_{\Lambda_T}<1\Big\}, \end{equation} with the norm \begin{equation}\label{LambdaTNorm} \|h\|_{\Lambda_T}:=\sup_{t\leq-T}|h(t)|+\sup_{t\leq -T}\Big(\frac{|t|}{\log|t|}|h'(t)|\Big), \end{equation} and also a close domain \begin{equation}\label{eq4.9} X_{T}:=\left\{\phi:\phi\in C_{\Phi}((-\infty, -T)\times(0, \infty))\ \ \text{and}\ \ \sum_{l=0}^2\big\|\partial^l_r\phi\big\|_{C_{\Phi}((-\infty, -T)\times(0, \infty))} \leq \frac{2 \widehat{C}}{\log T}\right\}, \end{equation} where the space $ C_{\Phi}((-\infty, -T)\times(0, \infty))$ is defined by \eqref{eq3.3} and $\widehat{C}$ is a fixed constant. The main result is given by the following proposition \begin{prop}\label{prop3} There exists $T_2\geq T_1$ such that for any given function $h(t)$ with each $h\in \Lambda_{T_2}$, there is a solution $\phi(t, r, h)$ to $\phi=\mathbf{ T}(\phi)$ with respect to $\rho(t)=\gamma_n(t)+h(t)$. The solution $\phi(t, r, h)$ satisfies problem (\ref{eq2.10})-(\ref{eq2.11}). Furthermore, the following estimate holds \begin{equation}\label{eq4.3} \sum_{l=0}^2\big\|\partial^l_r\phi(t, r, h_1)-\partial^l_r\phi(t, r, h_2)\big\|_{C_{\Phi}((-\infty, -T_1)\times(0, \infty))} \leq C\frac{1}{\log T_2}\|h_1-h_2\|_{\Lambda_{T_2}}, \end{equation} where $T_1$ and $\mathbf{T}$ are given by Proposition \ref{prop2} and \eqref{eq4.2} respectively, $C$ is a positive constant which does not depend on $h_1, h_2$ and $T_2$. \end{prop} To prove the above proposition, we first prepare two lemmas. We note that the error term $E(t, r)$ and the nonlinear term $N(\phi)$ in \eqref{Error}-(\ref{nonlinearterm}) are all dependent of $h$, due to the setting $\rho(t)=\gamma_n(t)+h(t)$ in \eqref{drho1}. So we denote $E(t, r)$ and $N(\phi)$ by $E(t, r, h)$ and $N(\phi, h)$. \begin{lem}\label{lem8} Let $h_1, h_2\in\Lambda_{T}$ and $\phi_1, \phi_2\in X_{T}$, then there exists a constant $C$ depending on $\widehat{C}$ such that \begin{equation}\label{eq4.4} \begin{aligned} &\big\|N(\phi_1, h_1)-N(\phi_2, h_2)\big\|_{C_{\Phi}((-\infty, -T)\times(0, \infty))} \\[2mm] &\leq \frac{C}{\log T} \Bigg\{\|h_1-h_2\|_{\Lambda_{T}} +\sum_{l=0}^2\big\|\partial_r^l\big(\phi_1-\phi_2\big)\big\|_{C_{\Phi}((-\infty, -T)\times(0, \infty))}\Bigg\} \end{aligned} \end{equation} and \begin{equation}\label{eq4.5} \begin{aligned} \|E(t, r, h_1)-E(t, r, h_2)\|_{C_{\Phi}((-\infty, -T)\times(0, \infty))}\leq &\frac{C}{\log T}\|h_1-h_2\|_{\Lambda_{T}}. \end{aligned} \end{equation} \end{lem} \begin{proof} By the definition of $N(\phi, h)$ in (\ref{nonlinearterm}) and the smoothness of $W$, we have that \begin{align}\label{eq4.6} &\nonumber\abs{N(\phi_1, h)-N(\phi_2, h)} \\[2mm]\nonumber &=\abs{\int_0^1\Big\{F'\big(z(t, r)+y\phi_1+(1-y)\phi_2\big)[\phi_1-\phi_2]-F'\big(z(t, r)\big)[\phi_1-\phi_2]\Big\}{\mathrm d}y} \\[2mm]\nonumber &=\abs{\int_0^1\Big\{F''\big(z(t, r)+\theta\big[y\phi_1+(1-y)\phi_2\big]\big) \big[y\phi_1+(1-y)\phi_2,\, \phi_1-\phi_2\big]\Big\}{\mathrm d}y} \\[2mm] &\leq C\left(\sum_{l=0}^2\big\|\partial_r^l\big(\phi_1-\phi_2\big)\big\|_{C_{\Phi}((-\infty, T_1)\times(0, \infty))}\right) \left(\sum_{l=0}^2\sum_{i=1}^2\big\|\partial_r^l\big(\phi_i\big)\big\|_{C_{\Phi}((-\infty, T_1)\times(0, \infty))} \right)\Phi(t, r), \end{align} and \begin{align} \abs{N(\phi, h_1)-N(\phi, h_2)} & =\abs{\int_0^{1}\int_{0}^{1}F''\big(yz_1(t, r)+(1-y)z_2(t, r)+s\big) \big[z_1(t, r)-z_2(t, r),\, \phi\big]{\mathrm d}yds} \nonumber\\[2mm] &\leq C\|h_1-h_2\|_{\Lambda_{T_1}}\|\phi\|_{C_{\Phi}((-\infty, T_1)\times(0, \infty))}\Phi(t, r), \label{eq4.7} \end{align} where $z_i(t, r)$ is given by \eqref{dz} with $\rho(t)=\gamma_n(t)+h_i(t)$, for $i=1, 2$. In the above, $C$ is a positive constant independent of $h, h_1, h_2$ and $\phi, \phi_1, \phi_2$, and the linear operator $F''(u)[v_1, v_2]$ is defined by \eqref{dFs}. Hence combining (\ref{eq4.6}) and (\ref{eq4.7}), we can obtain that (\ref{eq4.4}) holds. Next we prove that \eqref{eq4.5} also holds. By \eqref{de1} and \eqref{de2}, similar arguments in Lemma \ref{lem10} will imply that \begin{equation*}\begin{aligned} \abs{E_1(t, r, h_1)-E_1(t, r, h_2)} &\leq C\Bigg\{\abs{\rho'_1(t)\omega'\big(r-\rho_1(t)\big) \,-\, \rho'_2(t)\omega'\big(r-\rho_2(t)\big)} \\[2mm] &\quad\qquad\qquad+\frac{(n-1)^2(n-3)}{r^3}\abs{\omega'\big(r-\rho_1(t)\big) \,-\, \omega'\big(r-\rho_2(t)\big)} \Bigg\}\overline{\chi}_{\left\{r\geq\delta_0\right\}} \\[2mm] &\quad +\frac{C}{[\gamma_n(t)]^2}\Big\{\abs{h_1(t)-h_2(t)}+\abs{h'_1(t)-h'_2(t)}\Big\}\overline{\chi}_{\left\{ \frac{ \delta_0}{2}<r<\delta_0\right\}} \\[2mm] &\leq C\|h_1-h_2\|_{\Lambda}\frac{\Phi(t, r)}{\log T}, \end{aligned}\end{equation*} and \begin{align*} &\abs{E_2(t, r, h_1)-E_2(t, r, h_2)} \\[2mm] & \leq C\Bigg\{\abs{\rho'_1(t)\widetilde{\omega}'\big(r-\rho_1(t)\big)-\rho'_2(t)\widetilde{\omega}'\big(r-\rho_2(t)\big)} \,+\, \frac{1}{r^3}\sum_{l=0}^3\abs{\partial_r^l\omega\big(r-\rho_1(t)\big) -\partial_r^l\omega\big(r-\rho_2(t)\big)} \\[2mm] &\qquad\qquad \,+\, \frac{1}{r^3}\sum_{l=0}^3\abs{\partial_r^l\widetilde{\omega}\big(r-\rho_1(t)\big) -\partial_r^l\widetilde{\omega}\big(r-\rho_2(t)\big)} \Bigg\}\overline{\chi}_{\left\{r\geq\delta_0\right\}} \\[2mm] &\quad +\frac{C}{[\gamma_n(t)]^2}\Big\{\abs{h_1(t)-h_2(t)}+\abs{h'_1(t)-h'_2(t)}\Big\}\overline{\chi}_{\left\{ \frac{\delta_0}{2}<r<\delta_0\right\}} \\[2mm] &\leq C\|h_1-h_2\|_{\Lambda}\frac{\Phi(t, r)}{\log T}, \end{align*} where $\Phi(t, r)$ is defined by \eqref{dpsi} and $C>0$ is a uniform constant independent of $h_1, h_2$ and $t$. Using the fact that $E(t, r)=E_1(t, r)+E_2(t, r)$ and combining the above estimates, we can get (\ref{eq4.5}). \end{proof} \begin{lem} Let $h_1, h_2\in \Lambda_{T}$, $\phi_1, \phi_2\in X_{T}$, $c(\phi, h, t)$ satisfy equation $(\ref{eq3.20})$ with respect to $\phi$ and $\rho=\gamma_n+h$. Then \begin{equation}\label{eq4.8}\begin{aligned} \abs{c(\phi_1, h_1, t)-c(\phi_2, h_2, t)} &\leq \frac{C}{T^{3/4}\big[\log T\big]^{p-1}}\|\phi_1-\phi_2\|_{C_{\Phi}((-\infty, -T)\times(0, \infty))} \\[2mm] &\quad+\frac{C}{T^{1/2}\big[\log T\big]^{p-1}}\|h_1-h_2\|_{\Lambda_{T}}, \end{aligned}\end{equation} where $C$ depends on $\widehat{C}$ in $(\ref{eq4.9})$, and $p\in(n, n+1]$. \end{lem} \begin{proof} Proving this lemma just need to do some similar calculations in Lemmas \ref{lem6} and \ref{lem8}, we omit it here. \end{proof} \textbf{Proof of Proposition \ref{prop3}} We consider the operator $\mathbf{T}$ defined by (\ref{eq4.2}) from the domain $X_{T}$ in (\ref{eq4.9}) to itself. We will prove $\mathbf{T}$ is a contraction mapping. Thus by fixed-point theorem, the operator $\mathbf{T}$ has a unique fixed point $\phi$, i.e. $\mathbf{T}(\phi)=\phi$. For any $\phi_1, \phi_2\in X_{T}$, according to Lemma \ref{lem10}, Lemma \ref{lem8} and Proposition \ref{prop2}, we find that \begin{equation*} \sum_{l=0}^2\big\|\partial^l_r\mathbf{T}(0)\big\|_{C_{\Phi}((-\infty, T_4)\times(0, \infty))}\leq \frac{\widehat{C}}{\log T} \end{equation*} and \begin{equation*} \sum_{l=0}^2\big\| \partial^l_r \left[\mathbf{T}(\phi_1)-\mathbf{T}(\phi_2)\right]\big\|_{C_{\Phi}((-\infty, T)\times(0, \infty))} \leq \frac{\widehat{C}}{\log T} \| \phi_1-\phi_2\|_{C_{\Phi}((-\infty, T)\times(0, \infty))}. \end{equation*} Hence, $\mathbf{T}$ is a contraction mapping in $X_{T}$ for any $T$ large enough. Hence, according to Banach fixed-point theorem, there exists a unique $\phi\in X_{T}$ such that $\mathbf{T}(\phi)=\phi$. Next we will prove the estimate (\ref{eq4.3}). Choosing $h_1, h_2\in \Lambda_{T}$, according to the above proof, we know that there exist $\phi_i=\phi(t, r, h_i)$, $i=1, 2$, are two solutions to problem \eqref{eq2.10}-\eqref{eq2.11} with $\rho=\gamma_n+h_i$ respectively. We note that $\phi_1-\phi_2$ does not satisfy the orthogonality condition (\ref{eq2.11}). Let us consider a function $\bar{\phi}=\phi_1-\bar{\phi}_2$, where \begin{equation*} \bar{\phi}_2= \phi_2-\tilde{c}(t)\partial_r\widehat{\omega}_1(t, r), \end{equation*} where $\widehat{\omega}_l(t, r):=\omega\big(r-\rho_1(t)\big)\chi(\gamma_n(t)r)$ with $\rho_1(t)=\gamma_n(t)+h_1(t)$, the cut-off function $\chi$ is defined by \eqref{dcut-off}, and $\tilde{c}(t)$ is defined by the following relation \begin{equation*} \tilde{c}(t)\int_{0}^\infty \partial_r\widehat{\omega}_1(t, r)\omega'\big(r-\rho_1(t)\big)r^{n-1}{\mathrm d}r =\int_{0}^\infty \phi_2(t, r)\omega'\big(r-\rho_1(t)\big)r^{n-1}{\mathrm d}r, \end{equation*} for all $t\leq-T$. According to the proof of Lemma \ref{lem6}, the coefficient of $ \tilde{c}(t)$ in the left hand side of above equality is strictly positive, hence the function $\tilde{c}(t)$ is well-defined. Thus $\bar{\phi}$ satisfies the following problem \begin{equation*} \left\{ \begin{array}{l} \bar{\phi}_t=F'\big(z_1(t, r)\big)[\bar{\phi}] \,+\, \big[E(t, r, h_1)-E(t, r, h_2)\big] \,+\, \big[N(\phi_1, h_1)-N(\phi_2, h_2)\big] \\[3mm] \quad \quad+c_2(t)\big[\partial_r\widehat{\omega}_2(t, r)-\partial_r\widehat{\omega}_1(t, r)\big] \,+\, R(h_1, h_2) \, -\, \big[c_1(t)-c_2(t)\big]\partial_r\widehat{\omega}_1(t, r) \\[3mm] \quad \quad \ \text{in}\ (-\infty, -T)\times(0, \infty), \\[3mm] \int_{0}^\infty \bar{\phi}(t, r)\omega'\big(r-\rho_1(t)\big)r^{n-1}{\mathrm d}r=0, \qquad \text{for all}\ t<-T, \end{array} \right. \end{equation*} where $\widehat{\omega}_i(t, r):=\omega(r-\rho_i(t))\chi(\gamma_n(t)r)$ with $i=1, 2$, and the term $R(h_1, h_2)$ is defined by \begin{align*} R(h_1, h_2)=&\,\tilde{c}'(t)\partial_r\widehat{\omega}_1(t, r) +\tilde{c}(t)\partial_t\big[\partial_r\widehat{\omega}_1(t, r)\big] \\[2mm] &+\tilde{c}(t)F'(z_1(t, r))\big[\widehat{\omega}_1(t, r)\big] +F'\big(z_2(t, r)\big)[\phi_2] -F'\big(z_1(t, r)\big)[\phi_2], \end{align*} where $z_i(t, r)$ is defined by \eqref{dz} with $\rho(t)=\gamma_n(t)+h_i(t)$ for $i=1, 2$. Recall that the linear operator $L[\phi]=F'\big(z(t, r)\big)[\phi]$ is given by \begin{align*} L[\phi]:=&-\phi_{rrrr}-\frac{2(n-1)}{r}\phi_{rrr}+\Bigg[2W''\big(z(t, r)\big)-\frac{(n-3)(n-1)}{r^2}\Bigg]\phi_{rr} -\big(W''\big(z(t, r)\big)\big)^2\phi \\[2mm] &+\Bigg[\frac{2(n-1)W''\big(z(t, r)\big)}{r}-\frac{(3-n)(n-1)}{r^3}\Bigg]\phi_{r} +2W'''\big(z(t, r)\big)\phi_rz_r+W^{(4)}\big(z(t, r)\big)\abs{z_r}^2\phi \\[2mm] &-W'''\big(z(t, r)\big)W'\big(z(t, r)\big)\phi +2W'''\big(z(t, r)\big)z_{rr}\phi+2\frac{n-1}{r}W'''\big(z(t, r)\big)z_r\phi, \end{align*} where the approximate solution $z(t,r)$ is defined by \eqref{dz}. Hence by Proposition \ref{prop2}, the proof of Lemma \ref{lem2}, Lemmas {\ref{lem7}} and {\ref{lem8}}, we have \begin{align} \sum_{l=0}^2 \| \partial^l_r\bar{\phi}\|_{C_{\Phi}((-\infty, -T)\times(0, \infty))}\leq &C\frac{1}{\log T}\Big\{\|h_1-h_2\|_{\Lambda_T}+\|\phi_1-\phi_2\|_{C_{\Phi}((-\infty, -T)\times(0, \infty))}\Big\} \nonumber\\[2mm] &+C\sup_{t\leq -T}\frac{\abs{t}^{1/2}}{\left(\log\abs{t}\right)^{1-p}}\Big(\abs{\tilde{c}(t)}+\abs{\tilde{c}'(t)}\Big). \label{eq4.10} \end{align} By the orthogonality condition (\ref{eq2.11}), we have \begin{equation}\label{eq4.11}\begin{aligned} \abs{\int_{0}^\infty \phi_2(t, r)\partial_r\widehat{\omega}_1(t, r)r^{n-1}{\mathrm d}r} &=\abs{\int_{0}^\infty \phi_2(t, r)\Big[\partial_r\widehat{\omega}_2(t, r)-\partial_r\widehat{\omega}_1(t, r)\Big]r^{n-1}{\mathrm d}r} \\[2mm] &\leq \frac{C}{\big[\log\abs{t}\big]^{p-1}\abs{t}^{\frac{1}{2}}\log T}\|h_1-h_2\|_{\Lambda_{T}}\abs{\gamma_n(t)}^{n-1}, \end{aligned}\end{equation} where we have used the following fact \begin{equation*} \abs{\frac{\partial_r\widehat{\omega}(t, r)}{\Phi(t, r)}}\leq C\abs{t}^{1/2}\big[\log\abs{t}\big]^{p-1}, \qquad \text{for all} \ r>0, \end{equation*} where $\widehat{\omega}(t, r)$ is defined by \eqref{eq2.3}. We consider \begin{equation}\label{eq4.12}\begin{aligned} &\abs{\frac{{\mathrm d}}{{\mathrm d}t}\int_{0}^\infty \phi_2(t, r)\partial_r\widehat{\omega}_1(t, r)r^{n-1}{\mathrm d}r} =\abs{\frac{{\mathrm d}}{{\mathrm d}t}\int_{0}^\infty \phi_2(t, r) \big[\partial_r\widehat{\omega}_2(t, r)-\partial_r\widehat{\omega}_1(t, r)\big]r^{n-1}{\mathrm d}r }. \end{aligned}\end{equation} By the definition of $F'\big(z(t, r)\big)[\phi]$ and integration by parts, we have \begin{align} &\int_{0}^\infty (\phi_2)_t\big[\partial_r\widehat{\omega}_2(t, r)-\partial_r\widehat{\omega}_1(t, r)\big]r^{n-1}{\mathrm d}r \nonumber \\[2mm] &=\int_{0}^\infty \Big[(\phi_2)_t-F'\big(z_2(t, r)\big)[\phi]\Big]\Big(\partial_r\widehat{\omega}_2(t, r)-\partial_r\widehat{\omega}_1(t, r)\Big)r^{n-1}{\mathrm d}r \nonumber\\[2mm] &\quad+\int_0^\infty\left(-\phi_{rr}-\frac{n-1}{r}\phi_r+W''\big(z_2(t, r)\big)\right)\partial_{rr}\big[\partial_r\widehat{\omega}_2(t, r) -\partial_r\widehat{\omega}_1(t, r)\big]r^{n-1}{\mathrm d}r \nonumber \\[2mm] &\quad+(n-1)\int_0^\infty\left(-\phi_{rr}-\frac{n-1}{r}\phi_r+W''\big(z_2(t, r)\big)\right)\partial_r\big[\partial_r\widehat{\omega}_2(t, r)-\partial_r\widehat{\omega}_1(t, r)\big]r^{n-2}{\mathrm d}r \nonumber \\[2mm] &\quad+\int_0^\infty W''\big(z_2(t, r)\big)\Big[\phi_{rr}+\frac{n-1}{r}\phi_r-W''\big(z_2(t, r)\big)\phi\Big]\Big(\partial_r\widehat{\omega}_2(t, r)-\partial_r\widehat{\omega}_1(t, r)\Big)r^{n-1}{\mathrm d}r \nonumber \\[2mm] &\quad+\int_{0}^\infty\left(\partial_{rr}z_2+\frac{n-1}{r}\partial_rz_2-W'\big(z_2(t, r)\big)\right) W'''\big(z_2(t, r)\big)\phi\Big(\partial_r\widehat{\omega}_2(t, r)-\partial_r\widehat{\omega}_1(t, r)\Big)r^{n-1}{\mathrm d}r. \label{eq4.13} \end{align} By the previous fixed-point arguments and the above equality \eqref{eq4.13}, we have \begin{equation}\label{eq4.14}\begin{aligned} \abs{\int_{0}^\infty r^{n-1}(\phi_2)_t\Big[\partial_r\widehat{\omega}_2(t, r) -\partial_r\widehat{\omega}_1(t, r)\Big]{\mathrm d}r} \leq \frac{C}{\log T}\frac{\|h_1-h_2\|_{\Lambda_{T}}}{\abs{t}^{\frac{1}{2}}\big[\log\abs{t}\big]^{p-1}} \abs{\gamma_n(t)}^{n-1}. \end{aligned}\end{equation} By the above estimates (\ref{eq4.11}), (\ref{eq4.12}), (\ref{eq4.14}) and the definition of $\tilde{c}(t)$, we have \begin{equation*} \abs{\tilde{c}(t)}+\abs{\tilde{c}'(t)}\leq \frac{C}{\log T}\frac{\|h_1-h_2\|_{\Lambda_{T}}}{\abs{t}^{\frac{1}{2}}\big[\log\abs{t}\big]^{p-1}}, \qquad \forall\, t<-T. \end{equation*} Hence \begin{equation*} \sum_{l=0}^2 \big\|\partial^l_r \bar{\phi}\big\|_{C_{\Phi}((-\infty, -T)\times(0, \infty))} \leq \frac{C}{\log T}\Big[ \big\|\phi_1-\phi_2\big\|_{C_{\Phi}((-\infty, -T)\times(0, \infty))}+\|h_1-h_2\|_{\Lambda_{T}}\Big], \end{equation*} where $C$ is uniform positive constant independent of $T$. Eventually, we have that \begin{align*} \sum_{l=0}^2\big\| \partial_r^l[\phi_1-\phi_2]\big\|_{C_{\Phi}((-\infty, -T)\times(0, \infty))} &\leq \sum_{l=0}^2 \big\| \partial_r^l\bar{\phi}\big\|_{C_{\Phi}((-\infty, -T)\times(0, \infty))} +C\sup_{t\leq -T}\frac{\abs{t}^{1/2}}{\big(\log\abs{t}\big)^{p-1}}\abs{\tilde{c}(t)} \\[2mm] &\leq C \frac{1}{\log T}\Big[ \| \phi_1-\phi_2\|_{C_{\Phi}((-\infty, -T)\times(0, \infty))}+\|h_1-h_2\|_{\Lambda_{T}}\Big], \end{align*} where we choose $T$ large enough. Thus we can obtain the estimate (\ref{eq4.3}). \section{The reduction procedure: choosing the parameter $h$} \label{sec:tc} As we stated in Section \ref{section2.2}, the left job is to choose suitable $h(t)$, i.e. $\rho(t)=\gamma_n(t)+h(t)$, such that the function $c(t)$ vanishes in equation (\ref{eq2.10}). \subsection{Deriving the reduced equation involving $h(t)$} According to (\ref{eq2.14}), the relation $c(t)=0$ is equivalent to \begin{align}\label{eq7.12} 0=&\int_{0}^\infty \Big[\omega'''\big(r-\rho(t)\big)+\frac{n-1}{r}\omega''\big(r-\rho(t)\big) -W''\big(z(t, r)\big)\omega'\big(r-\rho(t)\big)\Big] \nonumber\\[2mm] &\qquad\qquad \times\left(-\phi_{rr}-\frac{n-1}{r}\phi_r+W''\big(z(t, r)\big)\phi\right) r^{n-1}{\mathrm d}r \nonumber\\[2mm] & +\int_{0}^\infty\left[\partial_{rr}z(t, r) +\frac{n-1}{r}\partial_{r}z(t, r)-W'\big(z(t, r)\big)\right]\phi \omega'\big(r-\rho(t)\big)r^{n-1}{\mathrm d}r \nonumber\\[2mm] & +\int_{0}^\infty \phi(t, r)\partial_t\big[\omega'\big(r-\rho(t)\big)\big]r^{n-1}{\mathrm d}r \,+\, \int_{0}^\infty \big(E(t, r)+N(\phi)\big)\omega'\big(r-\rho(t)\big)r^{n-1}{\mathrm d}r, \end{align} where the error term $E(t, r)$ and the nonlinear term $N(\phi)$ are defined by \eqref{Error}-(\ref{nonlinearterm}), $z(t, r)$ is given by \eqref{dz} and $\phi, \phi_r, \phi_{rr}\in C_{\Phi}((-\infty, -T)\times(0,+\infty))$ defined by \eqref{eq3.3}. The tedious computations of all terms in \eqref{eq7.12} will be given in the following parts. \subsubsection{} We first estimate the projection of the error term $E(t, r)$. According to \eqref{de1} and \eqref{de2}, we have that \begin{align} \int_{0}^{+\infty}E(t, r)\omega'\big(r-\rho(t)\big)r^{n-1}{\mathrm d}r =& \int_{0}^{\delta_0}E(t, r)\omega'\big(r-\rho(t)\big)r^{n-1}{\mathrm d}r \nonumber\\[2mm] &+\sum_{l=1}^5\int_{\delta_0}^{+\infty}\widetilde{E}_l(t, r)\omega'\big(r-\rho(t)\big)r^{n-1}{\mathrm d}r, \label{integerate E} \end{align} where $\gamma_n(t)$ is given by \eqref{dgamma-n} with fixed small positive number $\delta_0$, and the functions $ \widetilde{E}_1(t, r), \cdots, \widetilde{E}_5(t, r)$ are defined as follows \begin{equation}\label{dee}\begin{aligned} \widetilde{E}_1(t, r)&:=\omega'\big(r-\rho(t)\big)\left(\rho'(t)+\frac{(n-3)(n-1)^2}{2r^3}\right), \\[2mm] \widetilde{E}_2(t, r)&:=-\partial_{rr}\Big(\partial_{rr}\widehat{\omega}(t, r)-W'\big(\widehat{\omega}(t, r)\big)\Big) +W''\big(\widehat{\omega}(t, r)\big)\Big(\partial_{rr}\widehat{\omega}(t, r) -W'\big(\widehat{\omega}(t, r)\big)\Big), \\[2mm] \widetilde{E}_3(t, r)&:=\frac{2(n-1)}{r}\Big[W''\big(\widehat{\omega}(t, r)\big)-W''\big(\omega\big(r-\rho(t)\big)\big)\Big]\omega'\big(r-\rho(t)\big), \\[2mm] \widetilde{E}_4(t, r)&:=F'\big(\widehat{\omega}(t, r)\big)\big[\widetilde{z}(t, r)\big]-\frac{(n-1)(n-3)}{r^2}\left[\frac{(n-3)}{2r}\omega'\big(r-\rho(t)\big)+\omega''\big(r-\rho(t)\big)\right], \\[2mm] \widetilde{E}_5(t, r)&:=F''\big(\widehat{\omega}(t, r)+\theta\widetilde{z}(t, r)\big) \big[\widetilde{z}(t, r), \widetilde{z}(t, r)\big]-\frac{\partial\widetilde{z}(t, r)}{\partial t}, \end{aligned}\end{equation} where the functions $\widehat{\omega}(t, r)$ and $\widetilde{z}(t, r)$ are defined by \eqref{eq2.3} and \eqref{dzt} respectively. For the first term in the right hand side of \eqref{integerate E}, by Lemma \ref{lem10}, we have that \begin{equation}\label{intE_1}\begin{aligned} \abs{\int_{0}^{\delta_0}E(t, r)\partial_r\big[\omega\big(r-\rho(t)\big)\big]r^{n-1}{\mathrm d}r} &\leq C \int_{0}^{\delta_0}\Phi(t, r)\omega'\big(r-\rho(t)\big)r^{n-1}{\mathrm d}r \\[2mm] &\leq C\frac{\, \big[\rho(t)\big]^{n-1}\log\abs{t}}{\abs{t}^{1/2}\, }\exp\Bigg\{-\frac{\gamma_n(t)}{2}\Bigg\}, \end{aligned}\end{equation} where $\gamma_n(t)$ is defined by \eqref{dgamma-n} and $C>0$ only depends on $n$ and $\alpha=\sqrt{W''(1)}$. Next we will estimate other terms in the right hand side of \eqref{integerate E}. Using the definition of $\widehat{\omega}(t, r)$ in \eqref{eq2.3} and equation in \eqref{eqq}, we have that \begin{equation*} \widetilde{E}_2(t, r)=\widetilde{E}_3(t, r)=0, \qquad \text{for all}\ r>\delta_0. \end{equation*} For $\widetilde{E}_1(t, r)$ in \eqref{dee}, by Lemma \ref{lem1}, we have that \begin{align} &\int_{\delta_0}^\infty\widetilde{E}_1(t, r)\omega'\big(r-\rho(t)\big)r^{n-1}{\mathrm d}r \nonumber\\[2mm] &= \int_{\delta_0}^\infty\omega'\big(r-\rho(t)\big)\left(\rho'(t)+\frac{(n-3)(n-1)^2}{2r^3}\right)\omega'\big(r-\rho(t)\big)r^{n-1}{\mathrm d}r \nonumber\\[2mm] &=\left(\rho'(t)+\frac{(n-3)(n-1)^2}{2\rho^3(t)} \right)\big[\rho(t)\big]^{n-1}\int_{\R}\big[\omega'(x)\big]^2{\mathrm d}x \,+\, O\left(\frac{\big[\rho(t)\big]^{n-1}\log\abs{t}}{\abs{t}^{5/4}}\right), \label{ese1} \end{align} where we have used the fact that $$ \int_\R[\omega'(y)]^2y{\mathrm d}y=0, $$ in the last equality. Next we estimate the terms $\widetilde{E}_4(t, r)$ given in \eqref{dee}, by the definition of $\widetilde{z}(t, r)$ in \eqref{dzt} and the same arguments in proof of Lemma \ref{lem10}, \eqref{eq2.13}, for all $r>\delta_0$, we have that \begin{align*} \widetilde{E}_4(t, r)&=\frac{(n-1)(n-3)}{r^3} \Bigg\{4\partial_{rrr}\widetilde{\omega}\big(r-\rho(t)\big)-4W''\big(\omega\big(r-\rho(t)\big)\big)\partial_r\widetilde{\omega}\big(r-\rho(t)\big) -\frac{(n-3)}{2}\omega'\big(r-\rho(t)\big) \\[3mm] &\qquad\qquad\qquad-2(n-3)\Big[\partial_{rrr}\widetilde{\omega}\big(r-\rho(t)\big)-W''\big(\omega\big(r-\rho(t)\big)\big)\partial_{r}\widetilde{\omega}\big(r-\rho(t)\big) \omega'\big(r-\rho(t)\big) \\[3mm] &\qquad\qquad\qquad\qquad -W'''\big(\omega\big(r-\rho(t)\big)\big)\widetilde{ \omega}\big(r-\rho(t)\big)\Big]\Bigg\} \\[3mm] &\quad+\frac{1}{r^4}\Bigg\{\Big[12(n-4)-(n-1)(n-3)\Big]\partial_{rr}\widetilde{\omega}\big(r-\rho(t)\big) -4(n-4)W''\big(\omega\big(r-\rho(t)\big)\big)\widetilde{\omega}\big(r-\rho(t)\big)\Bigg\} \\[3mm] &\quad+\rho'(t)\frac{(n-1)(n-3)}{r^2}\widetilde{\omega}'\big(r-\rho(t)\big) \,+\, O\left(\frac{1}{r^5}\sum_{j=1}^ke^{-\frac{\alpha}{2}}\abs{r-\rho(t)}\right), \end{align*} where we have used that $\widetilde{\omega}(x)$ in \eqref{w1} and its derivatives have exponential decay. Thus, by same arguments in above estimate of the projection of $\widetilde{E}_2(t, r)$ and the equalities in \eqref{eq2.1}, \eqref{w1} and \eqref{eqq}, integrating by parts, we have that \begin{align*} &\int_{\delta_0}^\infty\widetilde{E}_4(t, r)\omega'\big(r-\rho(t)\big)r^{n-1}{\mathrm d}r \\[2mm] &=4(n-3)(n-1)\int^\infty_{\frac{\gamma_n}{2}-\rho(t)}\omega'(y) \Big[\widetilde{\omega}'''(y)-W''\big(\omega(y)\big)\widetilde{\omega}'(y)\Big]\big[y+\rho(t)\big]^{n-4}{\mathrm d}y \\[2mm] &\quad+\frac{(n-3)^2(n-1)}{2}\int^\infty_{\delta_0-\rho(t)}\omega'(y) \Big[2y\omega''(y)+\omega'(y)\Big]\big[y+\rho(t)\big]^{n-4}{\mathrm d}y \,+\, \int^\infty_{\delta_0-\rho}\frac{\omega'(y)\big[y+\rho(t)\big]^{n-6}}{\exp\Big\{\alpha\frac{\abs{y}}{2}\Big\}}{\mathrm d}y \\[2mm] &\quad \,+\, \int^\infty_{\delta_0-\rho(t)}\Bigg[\Big(12(n-4)-4(n-1)(n-3)\Big)\widetilde{\omega}''(y) -4(n-4)W''\big(\omega(y)\big)\widetilde{\omega}(y)\Bigg] \omega'(y)\big[y+\rho(t)\big]^{n-4}{\mathrm d}y \\[2mm] &\quad+\frac{\rho'(t)}{\big[\rho(t)\big]^2}\int^\infty_{\delta_0-\rho(t)}\widetilde{\omega}'(y)\omega'(y)dy \\[2mm] &=O\left(\frac{\big[\rho(t)\big]^{n-1}}{\abs{t}^{5/4}}\right), \end{align*} where we have used \eqref{eq2.4}, \eqref{eq2.1}, \eqref{dwtd} and the fact that \begin{equation*} \int_\R \omega'(y)\Big[2y\omega''(y)+\omega'(y)\Big]{\mathrm d}y=0. \end{equation*} For the term $\widetilde{E}_5(t, r)$ in \eqref{dee}, by the definitions in \eqref{dzt} and \eqref{dFs}, the properties of $W$ in \eqref{eq1.4} and the oddness of $\widetilde{\omega}$ in \eqref{w1} respectively, and the equalities in \eqref{eq2.1}, we get that \begin{align*} &\int_{\delta_0}^\infty\widetilde{E}_5(t, r)\omega'\big(r-\rho(t)\big)r^{n-1}{\mathrm d}r \\[2mm] &=\int_{\delta_0}^\infty\Bigg\{ \partial_{rr}\left[W'''\big(\omega\big(r-\rho(t)\big)\big)\big(\omega\big(r-\rho(t)\big)\big)^2\right]-W'''\big(\omega\big(r-\rho(t)\big)\big)W''\big(\omega\big(r-\rho(t)\big)\big)\big(\omega\big(r-\rho(t)\big)\big)^2 \\[2mm] & \quad +2\Big[\partial_{rr}\widetilde{\omega}\big(r-\rho(t)\big)-W''\big(\omega\big(r-\rho(t)\big)\big)\widetilde{\omega}\big(r-\rho(t)\big)\Big] W'''\big(\omega(r-\rho(t)\big)\widetilde{\omega}\big(r-\rho(t)\big)\Bigg\}\omega'\big(r-\rho(t)\big)r^{n-5}{\mathrm d}r \\[2mm] & \quad +O\left(\int^\infty_{\delta_0-\rho(t)}\frac{\omega'(y)\big[y+\rho(t)\big]^{n-6}}{\exp\Big\{\alpha\frac{\abs{y}}{2}\Big\}}{\mathrm d}y \right) \\[2mm] &=O\left( \frac{\big[\rho(t)\big]^{n-1}}{\abs{t}^{5/4}}\right). \end{align*} \subsubsection{ } We next estimate other terms in the right hand side of \eqref{eq7.12}. By $\phi\in C_{\Phi}((-\infty, -T)\times(0, +\infty))$ and the definitions of $\Phi(t, r)$ and $\rho(t)$ in \eqref{dpsi} and \eqref{drho1} respectively, we have that \begin{align*} &\int_{0}^{\infty}\abs{N(\phi)}\omega'\big(r-\rho(t)\big)r^{n-1}{\mathrm d}r \\[3mm] &\leq C\int_{0}^\infty\big(\Phi(t, r)\big)^2\omega'\big(r-\rho(t)\big)r^{n-1}{\mathrm d}r \\[3mm] &\leq C\frac{\big[\log\abs{t}\big]^2} {\abs{t}}\int^{+\infty}_{\frac{\gamma_n(t)}{4}-\rho(t)} \left (1+\abs{x+\rho(t)-\frac{\alpha}{3}\log\abs{t}}\right )^{-2p}\big[x+\rho(t)\big]^{n-1}\omega'(x){\mathrm d}x \\[3mm] &\quad +C\frac{\big[\log\abs{t}\big]^2}{\abs{t}}\big[\rho(t)\big]^{n-1}\exp\Bigg\{-\frac{\gamma_n(t)}{2}\Bigg\}. \end{align*} Thus, by Lemma \ref{lem1} and direct computations, we have that \begin{equation}\label{en} \int_{0}^{\infty}\abs{N(\phi)}\omega'\big(r-\rho(t)\big)r^{n-1}{\mathrm d}r\leq C \frac{\big[\rho(t)\big]^{n-1}}{\abs{t}\big[\log\abs{t}\big]^{2(p-1)}}, \end{equation} where $C$ only depends on $\alpha, n$ and $\|h\|_{L^\infty}$. We estimate the first and second term in \eqref{eq7.12}. According to Taylor expansion and \eqref{eqq}, the same argument as the proof of Lemma \ref{lem2}, we have that \begin{align*} \Bigg{|}&\int_{0}^\infty\Big[-\omega'''\big(r-\rho(t)\big) -\frac{n-1}{r}\omega''\big(r-\rho(t)\big) +W''\big(z(t, r)\big)\omega'\big(r-\rho(t)\big)\Big] \\[3mm] &\qquad\qquad \times\left[\phi_{rr}+\frac{n-1}{r}\phi_{r}-W''\big(z(t, r)\big)\phi\right] r^{n-1}{\mathrm d}r \\[3mm] &\ +\int_{0}^\infty\left[\partial_{rr}z(t, r)+\frac{n-1}{r}\partial_{r}z(t, r)-W'\big(z(t, r)\big)\right]W'''\big(z(t, r)\big)\phi\omega'\big(r-\rho(t)\big)r^{n-1}{\mathrm d}r\Bigg{|} \\[3mm] &\leq C\int_{-\rho(t)}^\infty\abs{\left[\omega^{(4)}(r)-W''\big(\omega(r)\big)\omega''(r)-W'''\big(\omega(r)\big) [\omega'(r)]^2\right]\phi\big(t, r+\rho(t)\big)}\big[r+\rho(t)\big]^{n-1}{\mathrm d}r \\[3mm] &\leq C\frac{\big[\rho(t)\big]^{n-1}}{\abs{t}\big[\log\abs{t}\big]^{p-2}}. \end{align*} By Lemma \ref{lem1} and the definition of $\rho$ in (\ref{drho1}), similar arguments as in \eqref{en}, we have \begin{align*} \abs{\rho'(t)\int_{0}^\infty \phi(t, r)\omega''\big(r-\rho(t)\big)r^{n-1}{\mathrm d}r} &\leq \frac{C}{\abs{t}^{3/4}} \abs{ \int_{0}^\infty \Phi(t, r)\omega''\big(r-\rho(t)\big)r^{n-1}{\mathrm d}r} \\[2mm] & \leq C \frac{\big[\rho(t)\big]^{n-1}}{\abs{t}^{5/4}}. \end{align*} Combining the above estimates, we can obtain that (\ref{eq7.12}) is equivalent to the following ODE: \begin{equation}\label{eq7.13}\begin{aligned} \rho'(t)+\frac{(n-3)(n-1)^2}{2\rho^3(t)}=Q\big(\rho(t), \rho'(t)\big), \end{aligned}\end{equation} for all $t<-T$, where we recall that $\rho(t)=\gamma_n(t)+h(t)$, and $\gamma_n(t)$ is given by \eqref{dgamma-n}. The function $h(t)$ belongs to the set $\Lambda_T$ with $T>T_2$, where $T_2$ is given by Proposition \ref{prop3} and we also recall the following two definitions in \eqref{LambdaT}-\eqref{LambdaTNorm} $$ \Lambda_T=\left\{h(t): h\in C^1(-\infty, -T] \quad \text{and}\quad \|h\|_{\Lambda_T}< 1 \right\}, $$ where \begin{equation}\label{eq7.16} \|h\|_{\Lambda_T}=\sup_{t\leq-T}\abs{h(t)}+\sup_{t\leq-T}\left[\frac{\abs{t}}{\log\abs{t}}\abs{h'(t)}\right]. \end{equation} According to the above arguments, Proposition \ref{prop3} and Lemma \ref{lem10}, we have that \begin{prop}\label{prop1} Let $p\in(n, n+1]$ and $P(h(t), h'(t)):=Q\big(\rho(t), \rho'(t)\big)$ in \eqref{eq7.13}. Then for all $h, h_1, h_2\in\Lambda_T$, there hold that \begin{equation*} \abs{P(h, h')}\leq \frac{C(n, \alpha)}{\abs{t}\big[\log\abs{t}\big]^{2(p-1)}} \end{equation*} and \begin{equation*} \abs{P\big({h}_1, (h_1)'\big)-P\big(h_2, (h_2)'\big)}\leq \frac{C(n, \alpha)}{\abs{t}\big[\log\abs{t}\big]^{2(p-1)}}\|h_1-h_2\|_{\Lambda_T}, \end{equation*} where we recall that $\alpha=\sqrt{W''(1)}>0$ and $C(n, \alpha)$ is an uniform constant depending on $n$ and $\alpha$. \end{prop} \subsection{Solving the reduced equation involving $h$} In the rest of this section we will study the ODE in \eqref{eq7.13}. We look for solutions of \eqref{eq7.13} of the form $\rho(t)=\gamma_n(t)+h(t)$, then $h(t)$ satisfies that \begin{equation*} h'(t)+\gamma'_n(t)+\frac{(n-3)(n-1)^2}{2\big[\gamma_n(t)+h(t)\big]^3}=P\big(h(t), h'(t)\big)\qquad \text{in}\ (-\infty, -\widehat{T}_0], \end{equation*} where $\widehat{T}_0>T_2$, where $T_2$ is given by Proposition \ref{prop3}. By the definition of $\gamma_n(t)$ in \eqref{dgamma-n}, the above equation is equivalent to \begin{equation}\label{eq7.14} h'(t)+\frac{3h(t)}{4t}=\widetilde{P}\big(h(t), h'(t)\big)\qquad \text{in}\ (-\infty, -\widehat{T}_0], \end{equation} where \begin{equation}\label{dpt} \widetilde{P}\big(h(t), h'(t)\big):=P\big(h(t), h'(t)\big)+\frac{(n-3)(n-1)^2}{2}\left[\frac{1}{\big(\gamma_n(t)\big)^3}-\frac{1}{\big(\gamma_n(t)+h(t)\big)^3} -\frac{3h(t)}{\big(\gamma_n(t)\big)^4}\right]. \end{equation} We will solve equation \eqref{eq7.14} by applying the fixed-point theorem in a suitable space with $h(-\widehat{T}_0)=0$. It is easily to check that if $h(t)$ is a solution of \eqref{eq7.14} with initial data $0$, then it has the form \begin{equation}\label{eq7.17} h(t)=-\frac{1}{{(-t)}^{\frac{3}{4}}}\int^{-T_0}_t{(-s)}^{\frac{3}{4}} \widetilde{P}\big(h(s), h'(s)\big)\mathrm{d}s, \end{equation} with $t\leq-\widehat{T}_0$. Let us define two operators as following \begin{equation*} \mathcal{ P}(h(t)):=-\frac{1}{{(-t)}^{\frac{3}{4}}}\int^{-\widehat{T}_0}_t{(-s)}^{\frac{3}{4}} \widetilde{P}\big(h(s), h'(s)\big)\mathrm{d}s \qquad \text{and}\qquad \mathcal{ P}'(h(t)):=\partial_t \mathcal{ P}(h(t)). \end{equation*} Then using Proposition \ref{prop1} and \eqref{dpt}, we have that \begin{equation}\label{eq7.15} \abs{\mathcal{ P}(0)}\leq\frac{ \widetilde{C}(n, \alpha)}{\log \widehat{T}_0}\quad \text{and}\quad \frac{\abs{t}}{\log\abs{t}}\abs{\mathcal{ P}'(0)}\leq\frac{ \widetilde{C}(n, \alpha)}{\log \widehat{T}_0}, \end{equation} with $\widehat{T}_0>e^{2}$, where $\widetilde{C}(n, \alpha)$ is a positive constant depending on $n$ and $\alpha$. We consider the domain \begin{equation*} Y:=\left\{h(t)\in C^1(-\infty, -\widehat{T}_0]\ :\ \|h(t)\|_{\Lambda_{\widehat{T}_0}}\leq \frac{2\widetilde{C}(n, \alpha)}{\log \widehat{T}_0}\right\}, \end{equation*} where the norm $\|\cdot\|_{\Lambda_T}$ is given by \eqref{eq7.16} and $\widetilde{C}(n, \alpha)$ is a positive constant in \eqref{eq7.15}. According to Proposition \ref{prop1} and \eqref{dpt}, we get that for all $h_1, h_2\in Y$, \begin{equation*} \abs{\mathcal{ P}(h_1)-\mathcal{ P}(h_2)}\leq\frac{C(n, \alpha)}{\log \widehat{T}_0} \|h_1(t)-h_2(t)\|_{\Lambda_{\widehat{T}_0}} \end{equation*} and \begin{equation*} \abs{\mathcal{ P}'(h_1)-\mathcal{ P}'(h_2)}\leq\frac{C(n, \alpha)}{\log \widehat{T}_0} \|h_1(t)-h_2(t)\|_{\Lambda_{\widehat{T}_0}}, \end{equation*} where $C(n, \alpha)>0$ only depends on $n$ and $\alpha$. Thus by Banach fix point theorem, there exists $h(t)\in Y$ such that $\mathcal{P}(h(t))=h(t)$, if we choose $\widehat{T}_0$ big enough. Thus we proved that the ODE in \eqref{eq7.13} is solvable. Furthermore, by the formula in \eqref{eq7.17}, we get that \begin{equation*} \abs{h(t)}\leq\frac{C}{\log\abs{t}}, \qquad \text{as} \ t\rightarrow-\infty. \end{equation*}

44133ca26d7362d5c1e47220e3dc2c1dbc83a293

\section{Introduction} Over the course of the last several years, a significant amount of scholarly attention has been drawn to the issue of feature selection. At a high level, feature selection can be considered as a branch of reducing data dimensionality of which the two primary methods are \textit{feature learning} and \textit{feature selection}. The problem of feature learning involves the creation of new features from the original data. In contrast, the feature selection problem does not change the original representation of the data variables, so the physical meaning of each variable is preserved. To be more specific, the feature selection problem can be subdivided into two scenarios: supervised and unsupervised. Since we do not have target variables, selecting unsupervised features is more challenging. Typically, the unsupervised feature selection relies on matrix decomposition \citep{cheng2005compression, liberty2007randomized, martinsson2011randomized, lu2022bayesian}, filter \citep{dash2002feature}, and embeddings \citep{dy2004feature, hou2011feature}. On the other hand, matrix decomposition algorithms such as QR decomposition, and singular value decomposition have been used extensively over the years to reveal hidden structures of data matrices in scientific and engineering areas such as collaborative filtering \citep{marlin2003modeling, lim2007variational, mnih2007probabilistic, lu2022matrix, lu2022bayesian}, recommendation systems \citep{lu2022matrix}, clustering and classification \citep{li2009non, wang2013non}. Low-rank matrix approximations are therefore essential in data science. Due to the Eckart-Young-Misky theorem, low-rank approximation problems can be easily solved with singular value decomposition \citep{golub1987generalization}. However, it is frequently desirable for many applications to operate with a basis consisting of a subset of the original columns of the observed matrix \citep{martinsson2011randomized, kakushadze2016101}. The interpolative decomposition (ID) is one of these low-rank approximations; it reuses columns from the observed matrix, preserving matrix sparsity and nonnegativity while removing redundant information. In this context, the ID of underlying matrices captures our interest. The ID of an $M\times N$ data matrix $\bm{A}$ can be described by $\bm{A}=\bm{C}\bm{W}+\bm{P}$, where the matrix $\bm{A}$ is approximately factorized into a matrix $\bm{C}\in \mathbb{R}^{M\times K}$ reusing $K$ \textit{basis columns} of $\bm{A}$ (thus $\bm{C}$ is also known as a skeleton of $\bm{A}$) and a matrix $\bm{W}\in \mathbb{R}^{K\times N}$ with entries no greater than 1 in magnitude; the error is captured by an $M\times N$ matrix $\bm{P}$. Training such models amounts to finding the optimal rank-$K$ approximation to the observed $M\times N$ data matrix $\bm{A}$ under some loss functions. Let $\bm{r}\in \{0,1\}^N$ be the \textit{state vector} with each entry indicating the type of the corresponding column, i.e., \textit{basis column} or \textit{interpolated (remaining) column}: if $r_n=1$, then the $n$-th column of $\bm{A}$ is a basis column; on the contrary, the $n$-th column is interpolated using the basis columns within a tolerance of error. Suppose further the set $I$ contains the indices of the interpolated columns with $r_n=0$ and the set $J$ contains the indices of the basis columns with $r_n=1$ where $$ J\cap I =\emptyset ; \,\,\,\,\,\,\,\, J \cup I =\{1,2,\ldots, N\}. $$ Then $\bm{C}$ can be described by the Matlab-style notation as $\bm{C}=\bm{A}[:,J]$ where the colon operator implies all indices. The approximation $\bm{A}\approx \bm{C}\bm{W}$ can be equivalently stated that $\bm{A}\approx\bm{C}\bm{W}=\bm{X}\bm{Y}$ where $\bm{X}\in \mathbb{R}^{M\times N}$ and $\bm{Y}\in \mathbb{R}^{N\times N}$ with $$ \begin{aligned} \bm{X}[:,J]&=\bm{C}\in \mathbb{R}^{M\times K}; \,\,\,\,\,\,\,\, &\bm{X}[:,I] &= \mathbf{0}\in \mathbb{R}^{M\times (N-K)};\\ \bm{Y}[J,:]&=\bm{W} \in \mathbb{R}^{K\times N}; \,\,\,\,\,\,\,\, &\bm{Y}[I,:] &= \text{random matrix} \in \mathbb{R}^{(N-K)\times N}. \end{aligned} $$ We also notice that there exists a $K\times K$ identity matrix $\bm{I}$ inside $\bm{W}$ and $\bm{Y}$: \begin{equation}\label{equation:submatrix_bid_identity} \bm{I} = \bm{W}[:,J] = \bm{Y}[J,J]. \end{equation} Having the equivalence of $\bm{C}\bm{W}=\bm{X}\bm{Y}$, the problem of $\bm{A}\approx\bm{C}\bm{W}$ can be stated as finding the approximation $\bm{A}\approx\bm{X}\bm{Y}$ alternatively with the state vector $\bm{r}$ recovering the submatrix $\bm{C}$ (Figure~\ref{fig:id-column}). Mean squared error (MSE) is applied to evaluate the \textit{reconstruction error}: \begin{equation}\label{equation:idbid-per-example-loss} \mathop{\min}_{\bm{W},\bm{Z}} \,\, \frac{1}{MN}\sum_{n=1}^N \sum_{m=1}^{M} \left(a_{mn} - \bm{x}_m^\top\bm{y}_n\right)^2, \end{equation} where $a_{mn}$ is the $(m,n)$-th element of matrix $\bm{A}$, and $\bm{x}_m$, $\bm{y}_n$ are the $m$-th \textbf{row} and $n$-th \textbf{column} of $\bm{X}$, $\bm{Y}$ respectively for simplicity. The magnitude constraint in $\bm{Y}$ or $\bm{W}$ is approached by considering the Bayesian ID model as a latent factor model where we employ Bayesian inference to find the latent components via the specified graphical model. Therefore, no explicit magnitude constraints are considered. In this paper, we introduce a novel Bayesian ID (BID) approach with each column of the observed matrix having its score measuring the importance in the model; the larger the score, the higher the priority to select; hence the name \textit{intervened interpolative decomposition (IID)}. The rest of the paper is organized as follows. We will introduce the vanilla Bayesian ID method in Section~\ref{section:related_iid_word}. Section~\ref{section:iid_main} then presents the proposed IID method. Section~\ref{section:iid_quantaprob} provides one of the applications for the IID method in finding quantitative strategies, followed by the experiments in Section~\ref{section:iid_experiments}. \section{Related Work}\label{section:related_iid_word} \begin{figure*}[h] \centering \vspace{-0.35cm} \subfigtopskip=2pt \subfigbottomskip=9pt \subfigcapskip=-5pt \includegraphics[width=0.95\textwidth]{imgs/id-column.pdf} \caption{Overview of the ID of the matrix $\bm{A}\in\mathbb{R}^{M\times N}$ where the yellow vectors denote the basis columns of matrix $\bm{A}$, white columns denote zero vectors, purple entries denote one, blue and black entries denote elements that are not necessarily zero. The Bayesian ID models get the approximation $\bm{A}\approx\bm{X}\bm{Y}$ and the post processing procedure obtains the approximation $\bm{A}\approx\bm{X}\bm{Y}\approx\bm{C}\bm{W}$.} \label{fig:id-column} \end{figure*} \subsection{Bayesian GBT Model for Interpolative Decomposition} \begin{SCfigure \centering \vspace{-0.55cm} \subfigtopskip=2pt \subfigbottomskip=6pt \subfigcapskip=-2pt \includegraphics[width=0.36\textwidth]{./imgs/bmf_bid_GBT_IID.pdf} \caption{Graphical representation of the GBT model where green circles denote prior variables, orange circles represent observed and latent variables, and plates represent repeated variables. Comma ``," in the cycles represents ``and", and ``/" in the cycles represents ``or". Parameters $a,b$ are fixed with $a=-1,b=1$ in our case; while a weaker construction can set them to $a=-2,b=2$.} \label{fig:bmf_bids_foriid} \end{SCfigure} In this section, we review the Bayesian approach for computing the interpolative decomposition. We consider the data matrix $\bm{A}$ to be generated via the probabilistic generative process (Figure~\ref{fig:bmf_bids_foriid}). The element $a_{mn}$ of matrix $\bm{A}$ is modeled via a Gaussian likelihood function, \begin{equation}\label{equation:iid_data_entry_likelihood} p(a_{mn} | \bm{x}_m^\top\bm{y}_n, \sigma^2) = \mathcal{N}(a_{mn}|\bm{x}_m^\top\bm{y}_n, \sigma^2), \end{equation} where $\bm{x}_m^\top\bm{y}_n$ and $\sigma^2$ are mean and variance respectively. Then, we place an inverse-Gamma prior over the data variance (a conjugate prior), \begin{equation}\label{equation:prior_iid_gamma_on_variance} p(\sigma^2 | \alpha_\sigma, \beta_\sigma) = \mathrm{IG}(\sigma^2 | \alpha_\sigma, \beta_\sigma), \end{equation} where $\mathrm{IG}(x|\alpha_\sigma, \beta_\sigma)= \frac{(\beta_\sigma)^\alpha}{\Gamma(\alpha_\sigma)} x^{-\alpha_\sigma-1}\exp\{-\frac{\beta_\sigma}{x}\}u(x)$ is an inverse-Gamma density with $\Gamma(\cdot)$ being the gamma function and $u(x)$ being the unit step function that has a value of $1$ when $x\geq0$ and 0 otherwise. We treat the latent variables $y_{kl}$'s (with $k,l\in \{1,2,\ldots,N\}$, see Figure~\ref{fig:bmf_bids_foriid}) as random variables. And in order to express beliefs about the values of these latent variables, we need prior densities over them, for example, a constraint with magnitude smaller than 1, even when there are many additional constraints (e.g., semi-nonnegativity in \citet{ding2008convex}, nonnegativity in \citet{lu2022flexible, lu2022robust}, or discreteness in \citet{gopalan2014bayesian, gopalan2015scalable}). Here we assume further that the latent variable $y_{kl}$'s are independently drawn from a general-truncated-normal prior: \begin{equation}\label{equation:rn_prior_bidd} p(y_{kl} | \cdot ) = \mathcal{GTN}(y_{kl} | \mu_{kl}, (\tau_{kl})^{-1}, a=-1, b=1), \end{equation} where $\mathcal{GTN}(x|\mu, \frac{1}{\tau}, a, b)=$ $\frac{\sqrt{\frac{\tau}{2\pi}} \exp \{-\frac{\tau}{2}(x-\mu)^2 \} }{\Phi((b-\mu)\cdot \sqrt{\tau})-\Phi((a-\mu)\cdot \sqrt{\tau})}$$u(x|a,b)$ is a general-truncated-normal (GTN) with zero density below $x=a$ or above $x=b$ and renormalized to integrate to one, $u(x|a,b)$ is a step function that has a value of 1 when $a\leq x\leq b$ and 0 otherwise, and $\Phi(\cdot)$ function is the cumulative distribution function of standard normal density $\mathcal{N}(0,1)$. The parameters $\mu$ and $\tau$ in GTN are known as the ``parent mean" and ``parent precision" of the original normal distribution $\mathcal{N}(\mu, \frac{1}{\tau})$. This GTN prior is thus utilized to enforce the constraint on the components $\bm{Y}$ (or $\bm{W}$) with no entry of $\bm{Y}$ having an absolute value greater than 1, and is conjugate to the Gaussian likelihood. We call the Bayesian ID method discussed above \textit{GBT} where \textit{G} stands for Gaussian density, \textit{B} stands for Beta-Bernoulli density intrinsically, and \textit{T} is short for general-truncated-normal density. \paragraph{Hierarchical prior and automatic relevance determination (ARD)} There is also a hierarchical model on Bayesian inference for ID where we place a joint hyperprior over the hyperparameters $\{\mu_{kl}, \tau_{kl}\}$ of GTN density in Eq.~\eqref{equation:rn_prior_bidd}, i.e., the GTN-scaled-normal-Gamma (GTNSNG) density that can decouple the parameters $\mu_{kl}, \tau_{kl}$, and as a result, their posterior conditional densities are normal and Gamma respectively \citep{lu2022bayesian}. And also the ARD method can determine the number of columns inside the factored component $\bm{X}$ automatically by a special prior on the state vector $\bm{r}$ \citep{lu2022comparative}. The development of the IID method on the non-hierarchical, hierarchical, and ARD models are the same, and we shall only discuss the non-hierarchical and non-ARD versions for simplicity. \paragraph{Post processing} The last step shown in Figure~\ref{fig:id-column} presents a step of post processing, where we enforce the identity submatrix in $\bm{W}$ (Eq.~\eqref{equation:submatrix_bid_identity}). This normally can reduce the MSE to a minor extent \citep{lu2022bayesian}. \subsection{Gibbs Sampler for GBT Model} In this section, we only shortly describe the posterior conditional density for Gibbs sampling to find the Bayesian inference. While a step-by-step derivation is provided in \citet{lu2022bayesian, lu2022comparative} for both hierarchical, non-hierarchical, ARD, and non-ARD versions. Denote all elements of $\bm{Y}$ except $y_{kl}$ as $\bm{Y}_{-kl}$, the conditional density of $y_{kl}$ is also a GTN density and it can be obtained by \begin{equation}\label{equation:posterior_gbt_ykl} \begin{aligned} &\,\,\,\,\,\,\,\, p(y_{kl} | \bm{A}, \bm{X}, \bm{Y}_{-kl}, \mu_{kl}, \tau_{kl}, \sigma^2) \propto p(\bm{A}|\bm{X},\bm{Y}, \sigma^2) \cdot p(y_{kl}|\mu_{kl}, \tau_{kl} )\\ & \propto \mathcal{GTN}(y_{kl}| \widetilde{\mu},( \widetilde{\tau})^{-1}, a=-1,b=1), \end{aligned} \end{equation} where $\widetilde{\tau} =\frac{\sum_{i}^{M} x_{ik} ^2}{\sigma^2} +\tau_{kl}$ is the posterior ``parent precision" of the GTN distribution, and $ \widetilde{\mu} = \big(\frac{1}{\sigma^2} \sum_{i}^{M} x_{ik} \big(a_{il}-\sum_{j\neq k}^{N}x_{ij} y_{jl}\big) +\textcolor{black}{\tau_{kl}\mu_{kl}} \big) \big/ \widetilde{\tau} $ is the posterior ``parent mean" of the GTN distribution. Given the state vector $\bm{r}=[r_1,r_2, \ldots, r_N]^\top\in \mathbb{R}^N$ such that the index set $J = J(\bm{r}) = \{n|r_n = 1\}_{n=1}^N$ and $I = I(\bm{r}) = \{n|r_n = 0\}_{n=1}^N$. To draw a state vector $\bm{r}$, we can select one index $j\in J$ and another index $i\in I$ (where the old values are $r_j=1$ and $r_i=0$) such that \begin{equation}\label{equation:postrerior_gbt_rvector} \begin{aligned} o_j &= \frac{p(r_j=0, r_i=1|\bm{A},\sigma^2, \bm{Y}, \bm{r}_{-ji})} {p(r_j=1, r_i=0|\bm{A},\sigma^2, \bm{Y}, \bm{r}_{-ji})}\\ &= \frac{p(r_j=0, r_i=1)}{p(r_j=1, r_i=0)} \times \frac{p(\bm{A}|\sigma^2, \bm{Y}, \bm{r}_{-ji}, r_j=0, r_i=1)}{p(\bm{A}|\sigma^2, \bm{Y}, \bm{r}_{-ji}, r_j=1, r_i=0)}, \end{aligned} \end{equation} where $\bm{r}_{-ji}$ denotes all elements of $\bm{r}$ except the $j$-th and $i$-th entries. In GBT, every column has same priority so we have $p(r_j=0, r_i=1)=p(r_j=1, r_i=0)$. Then the conditional probability of $p(r_j=0, r_i=1|\bm{A},\sigma^2, \bm{Y}, \bm{r}_{-ji})$ can be obtained by \begin{equation}\label{equation:postrerior_gbt_rvector222} p(r_j=0, r_i=1|\bm{A},\sigma^2, \bm{Y}, \bm{r}_{-ji}) = \frac{o_j}{1+o_j}. \end{equation} Finally, by conjugacy, the conditional posterior density of $\sigma^2$ is an inverse-Gamma distribution: \begin{equation}\label{equation:posterior_gnt_sigma2} \begin{aligned} &\,\,\,\,\,\,\,\, p(\sigma^2 | \bm{X}, \bm{Y}, \bm{A}) = \mathrm{IG}(\sigma^2 | \widetilde{\alpha_\sigma}, \widetilde{\beta_\sigma}), \end{aligned} \end{equation} where $\widetilde{\alpha_\sigma} = \frac{MN}{2}+\alpha_\sigma$, $\widetilde{\beta_\sigma}=\frac{1}{2} \sum_{i,j=1}^{M,N}(a_{ij}-\bm{x}_i^\top\bm{y}_j)^2+\beta_\sigma$ are the posterior shape and scale parameters for the inverse-Gamma density. The procedure for GBT is then formulated in Algorithm~\ref{alg:gbt_iid_gibbs_sampler}. \begin{algorithm}[!htb] \caption{Gibbs sampler for GBT (and IID) model. The procedure presented here can be inefficient but is explanatory. While a vectorized manner can be implemented to find a more efficient algorithm. By default, weak priors are $\alpha_\sigma=0.1, \beta_\sigma=1$, $\{\mu_{kl}\}=0, \{\tau_{kl}\}=1$, and $a=-1, b=1$ are fixed. Set the latent dimension $K$.} \label{alg:gbt_iid_gibbs_sampler} \begin{algorithmic}[1] \FOR{$t=1$ to $T$} \STATE Sample state vector $\bm{r}$ from Eq.~\eqref{equation:postrerior_gbt_rvector222} (based on Eq.~\eqref{equation:postrerior_gbt_rvector} for GBT, or on Eq.~\eqref{equation:posterior_IID} for IID); \STATE Update matrix $\bm{X}$ by $\bm{A}[:,J]$ where index vector $J$ is the index of $\bm{r}$ with value 1 and set $\bm{X}[:,I]=\mathbf{0}$ where index vector $I$ is the index of $\bm{r}$ with value 0; \STATE Sample variance $\sigma^2$ from $p(\sigma^2 | \bm{X},\bm{Y}, \bm{A})$ in Eq.~\eqref{equation:posterior_gnt_sigma2}; \FOR{$k=1$ to $N$} \FOR{$l=1$ to $N$} \STATE Sample factored component $y_{kl}$ from $p(y_{kl} | \bm{A}, \bm{X}, \bm{Y}_{-kl}, \mu_{kl}, \tau_{kl}, \sigma^2)$ in Eq.~\eqref{equation:posterior_gbt_ykl}; \ENDFOR \ENDFOR \STATE Output $||\bm{A}-\bm{X}\bm{Y}||_2$ loss in Eq.~\eqref{equation:idbid-per-example-loss}, and stop iteration if it converges. \ENDFOR \STATE Output mean loss in Eq.~\eqref{equation:idbid-per-example-loss} for evaluation after burn-in iterations. \end{algorithmic} \end{algorithm} \section{Intervened Interpolative Decomposition (IID)}\label{section:iid_main} Going further from the GBT model, we propose the intervened interpolative decomposition (IID) algorithm. The proposed IID algorithm has exactly the same generative process as shown in Eq.~\eqref{equation:iid_data_entry_likelihood}, inverse-Gamma prior on the variance parameter $\sigma^2$ in Eq.~\eqref{equation:prior_iid_gamma_on_variance}, and GTN prior over the latent variables $y_{kl}$'s in Eq.~\eqref{equation:rn_prior_bidd}. However, we consider further that some columns of the observed matrix $\bm{A}$ has a larger importance that \textit{should} be selected with a higher priority over the other columns. Suppose the importance of each column of the observed matrix $\bm{A}$ is captured by a \textit{raw importance vector} $\widehat{\bm{p}}\in \mathbb{R}^N$ where $\widehat{p}_i \in [-\infty, \infty]$ for all $i$ in $\{1,2,\ldots, N\}$. The raw importance vector can then be transformed into the range 0 to 1 $$ \bm{p} = \text{Sigmoid}(\widehat{\bm{p}}), $$ where \textit{Sigmoid($\cdot$)} is the $f(x) = \frac{1}{1+\exp\{-x\}}$ that can return value in the range 0 to 1. The Sigmoid function acts as a squashing function because its domain is the set of all real numbers, and its range is (0, 1). Then we take the $\bm{p}$ vector as the final \textit{importance vector} to indicate the importance of each column in the matrix $\bm{A}$. Going further from Eq.~\eqref{equation:postrerior_gbt_rvector}, the intermediate variable $o_j$ is calculated instead by \begin{equation}\label{equation:posterior_IID} \begin{aligned} o_j &= \frac{p(r_j=0, r_i=1)}{p(r_j=1, r_i=0)} \times \frac{p(\bm{A}|\sigma^2, \bm{Y}, \bm{r}_{-ji}, r_j=0, r_i=1)}{p(\bm{A}|\sigma^2, \bm{Y}, \bm{r}_{-ji}, r_j=1, r_i=0)}\\ &= \textcolor{blue}{ \frac{1-p_j }{p_j} \frac{p_i }{1-p_i} } \times \frac{p(\bm{A}|\sigma^2, \bm{Y}, \bm{r}_{-ji}, r_j=0, r_i=1)}{p(\bm{A}|\sigma^2, \bm{Y}, \bm{r}_{-ji}, r_j=1, r_i=0)}. \end{aligned} \end{equation} And again, the conditional probability of $p(r_j=0, r_i=1|\bm{A},\sigma^2, \bm{Y}, \bm{r}_{-ji})$ can be obtained by \begin{equation}\label{equation:postrerior_gbt_rvector222_IID} p(r_j=0, r_i=1|\bm{A},\sigma^2, \bm{Y}, \bm{r}_{-ji}) = \frac{o_j}{1+o_j}. \end{equation} Since we intervene in the procedure of the Gibbs sampling in Eq.~\eqref{equation:posterior_IID}, hence the name \textit{intervened interpolative decomposition (IID)}. \section{Quantitative Problem Statement}\label{section:iid_quantaprob} After developing the intervened interpolative decomposition algorithm, one may get confused about why it is so important and curious about the applications it can be applied in practice. It is well known that large quantitative hedge funds and asset managers have been recruiting a large number of data miners and financial engineers in order to build effective alphas, and the number of alpha components might climb into the millions or perhaps billions \citep{tulchinsky2019finding}. As a result, creating a meta-alpha from all of the alphas or a large fraction of the alpha pool might be troublesome for the following reasons: a). If we use the same alphas as others, some illiquid alphas with low volume will be traded heavily. This will make the strategy meaningless due to capacity constraints; b). Using too many alphas may result in overfitting, resulting in poor out-of-sample (OS) performance; c). Many alphas might be mutually dependent, and certain machine learning algorithms, such as neural networks, might uncover their limits caused by multi-linear difficulties while attempting to determine the meta-strategy from the entire set of alphas; d). Finding trading signals from the full alpha pool can be time-consuming because of limited computing resources; e). To minimize market risks, we constantly aim to discover a distinct subset of alphas to test alternative methods with low correlation. For the five reasons stated above, there is an urgent need to design algorithms that choose a small subset of alphas from a large pool of them in order to prevent overfitting, make the final approach more scalable, and obtain the findings in a reasonable amount of time. It is trivial to select an appropriate subset by the \textit{RankIC} metric (see definition below), i.e., we select the alphas having the highest RankIC values. However, the problems still remain that the selected subset will not represent the whole pool of alphas, and the selected alphas may be mutually dependent. Our objective is to identify as many representative alpha factors as possible with optimal performance. The selected subset of alphas is representative in the sense that the small subset of alphas can be used to reconstruct other alphas with a small replication error. The traditional ID algorithm, either using a \textit{Randomized algorithm} \citep{liberty2007randomized} or a Bayesian approach we have discussed above, can only help to find the representative ones. However, the end choices may seem to select alphas with low performance. Using the proposed IID method, on the other hand, can help find the representative (that can reconstruct other alphas with small error) and the desirable (high RankIC scores) alphas at the same time. \subsection{Formulaic Alphas} WorldQuant, a quantitative investment management firm, previously disclosed 101 formulaic short-term alpha determinants in 2016 \citep{kakushadze2016101}. Since then, the 191 alpha factors from Guotai Junan Securities \citep{guotaijunan2017} have also been welcomed by many investors and institutions. These formulaic alpha components are derived from several stock data elements, including, among others, volumes, prices, volatilities, and volume-weighted average prices (vwap). As the name implies, a formulaic alpha is a type of alpha that can be expressed as a formula or a mathematical expression. For example, a \textit{mean-reversion} alpha can be expressed in terms of a mathematical expression as follows: $$ \text{Alpha = }- \left( \text{close(today) $-$close(5$\_$days$\_$ago ) } \right)/ \text{close(5$\_$days$\_$ago)}. $$ In this sense, we take the opposite action as indicated by the closing price: we go short if the price has risen during the previous five days, and we go long otherwise. At a high level, the alpha value indicates the trend of the price in the days to come; the higher the alpha value for each stock, the more likely it is that the stock's price will rise in the next few days. \subsection{Evaluation Metrics} Let $r_{t}$ denote the yield rate of stock $s$ on $t$-th day. Suppose further $p_t$ is asset closing price at time $t$ where $t\in \{1,2,\ldots, T\}$, the return of the asset at time $t$ can be obtained by the following equation: \begin{equation} r_t = \frac{p_t - p_{t-1}}{p_t}. \end{equation} We use the Rank information coefficient (\textit{RankIC}) to evaluate the effectiveness of an alpha: \begin{equation} \text{RankIC}(\bm{a}, \bm{r}^h)=\text{Spearman}(\bm{a}, \bm{r}^h), \end{equation} where $\text{Spearman}(\cdot)$ indicates the Spearman correlation, $\bm{a}$ is the sequence of an alpha, $\bm{r}^h$ is the sequence of the return value with holding period $h$ such that the $i$-th element of $\bm{r}^h$ represent the daily return of $h$ days later. The RankIC then can be used as an indicator of the importance of each alpha factor and plugged into Eq.~\eqref{equation:posterior_IID} directly. \section{Experiments}\label{section:iid_experiments} For each stock $s$ (i.e., $s\in \{1,2,\ldots, S\}$ where $S$ is the total number of stocks), we have a matrix $\bm{A}_s$ with shape $\bm{A}_s\in \mathbb{R}^{N\times D}$ where $N$ is the number of alphas and $D$ is the number of dates so that each row of $\bm{A}_s$ is regarded as an alpha series. We want to select a subset of the alphas (here we assume $M$ out of the $N$ alphas are selected). The RankIC between each alpha series and the delayed return series with horizon $h=1$ is then taken as the \textit{important value} directly, a higher RankIC indicates a higher priority. \begin{table}[h] \centering \small \setlength{\tabcolsep}{7.4pt} \begin{tabular}{llllll} \hline Ticker & Type & Sector & Company & Average Amount \\ \hline SH601988 & Share & Bank & Bank of China Limited & 427,647,786 \\ SH601601 & Share & Public Utility & China Pacific Insurance (Group) & 819,382,926 \\ SH600028 & Share & Public Utility & China Petroleum \& Chemical Corporation & 748,927,952\\ SH600016 & Share & Bank & China Minsheng Banking Corporation &285,852,414 \\ SH601186 & Share & Public Utility & China Railway Construction Corporation & 594,970,588\\ SH601328 & Share & Bank & Bank of Communications Corporation & 484,445,915 \\ SH601628 & Share & Public Utility & China Life Insurance Company Limited&368,179,861 \\ SH601939 & Share & Bank & China Construction Bank Corporation &527,876,669 \\ \hline SH510300 & ETF & CSI 300 & Huatai-PineBridge CSI 300 ETF &1,960,687,059 \\ SH510050 & ETF & CSI 50 & ChinaAMC China CSI 50 ETF &2,020,385,879 \\ \hline \end{tabular} \vspace{-0.25cm} \caption{Summary of the underlying portfolios in the China market, ten assets in total. The average amount (in the currency of RMB) is calculated in the period of the test set.} \label{table:iid_cn_data_summary} \end{table} \paragraph{Dataset} To assess the proposed algorithm and highlight the primary benefits of the IID technique, we perform experiments with several analytical tasks and use data for ten assets from the China market and diverse industrial areas, including Bank, Public Utility, and ETF. We obtain publicly available data from tushare \footnote{\url{https://tushare.pro/}.}. The data covers a three-year period, i.e., 2018-07-18 to 2021-07-05 (720 trading days), where the data between 2018-07-18 and 2020-07-09 is considered the training set (480 calendar days); while data between 2020-07-10 and 2021-07-05 is taken as the test set (240 trading days). The underlying portfolios are summarized in Table~\ref{table:iid_cn_data_summary} and Figure~\ref{fig:bid_iid_datasets_ashare} shows the series of different assets where we initialize each portfolio with a unitary value for clarity. The assets are chosen by selecting the ones with high amount values (random ten assets among the fifty assets with highest average amounts in China market during the selected period) so that there are fewer trading restrictions. We obtain 78 alphas from the 101 formulaic alphas \citep{kakushadze2016101}, 94 alphas from the 191 formulaic alphas \citep{guotaijunan2017}, and 19 proprietary alphas. The alphas are chosen to have a value that is neither too large nor too small. In this sense, the alpha matrix $\bm{A}_s$ is of shape $214\times 480$ for each asset. In all scenarios, the same parameter initialization is adopted when conducting different tasks. Experimental evidence demonstrates that post-processing can marginally improve performance. For clarification, we only provide the findings of the GBT and IID models after post processing. The IID model can select the important features (alphas) with a higher priority while keeping the reconstructive error as small as possible, resulting in performance that is as good as or better than the vanilla GBT method in low-rank ID approximation across a wide range of experiments on different datasets. We use mean squared error (MSE, Eq.~\eqref{equation:idbid-per-example-loss}), which measures the similarity between the observed and reconstructive matrices, to evaluate the overall decomposition performance; the smaller the value, the better the performance. \begin{figure*}[h] \centering \vspace{-0.2cm} \subfigtopskip=2pt \subfigbottomskip=0pt \subfigcapskip=-2pt \subfigure[Convergence of the models on the SH510050, SH510300, SH601939, SH601628, and SH601328 datasets, as measured by MSE. The algorithm almost converges in less than 100 iterations.]{\includegraphics[width=1\textwidth]{imgs/iid_alpha_convergence.pdf} \label{fig:iid_alpha_convergence}} \subfigure[Averaged autocorrelation coefficients of samples of $y_{kl}$ computed using Gibbs sampling on the SH510050, SH510300, SH601939, SH601628, and SH601328 datasets.]{\includegraphics[width=1\textwidth]{imgs/iid_alpha_autocorrelation.pdf} \label{fig:iid_alpha_autocorrelation}} \vspace{-0.3cm} \caption{Convergence results (upper), and sampling mixing analysis (lower) on the SH510050, SH510300, SH601939, SH601628, and SH601328 datasets for a latent dimension of $K=10$. } \label{fig:allresults_bids_ard} \end{figure*} \paragraph{Hyperparameters} In those experiments, we use $a=-1, b=1,\alpha_\sigma=0.1, \beta_\sigma=1$, ($\{\mu_{kl}\}=0, \{\tau_{kl}\}=1$) for both GBT and IID models. The adopted parameters are uninformative and weak prior choices and the models are insensitive to them. The observed or unobserved variables are initialized from random draws as long as those hyperparameters are fixed since this initialization method provides a better initial guess of the correct patterns in the matrices. In all cases, we execute 1,000 iterations of Gibbs sampling with a burn-in of 100 iterations and a thinning of 5 iterations, since the convergence analysis indicates the algorithm can converge in fewer than 100 iterations. \begin{table}[] \centering \vspace{-0.35cm} \scriptsize \setlength{\tabcolsep}{3pt} \begin{tabular}{lllllllllll} \hline & SH601988 & SH601601 & SH600028 & SH600016 & SH601186 & SH601328 & SH601628 & SH601939 & SH510300 & SH510050\\ \hline GBT Min & 5.235 & 5.814 & 5.235 & \textbf{6.381}& 5.819 & 5.700 & 5.734 & 5.785 & 5.462 & 6.297 \\ IID Min & \textbf{4.567} &\textbf{5.700}& \textbf{4.843} & 6.490 & \textbf{5.104} & \textbf{5.658} & \textbf{5.445} & \textbf{5.435} & \textbf{4.876} & \textbf{5.767} \\ GBT Mean & 6.476 & \textbf{7.367} & 6.764 & 8.053 & 7.066 & 7.250 & 7.206 & 7.242 & 6.769 & 7.776 \\ IID Mean & \textbf{6.239} & 7.449 & \textbf{6.664} & \textbf{7.831} & \textbf{6.558} & \textbf{7.081} & \textbf{7.002} & \textbf{7.031} & \textbf{6.450} & \textbf{7.492} \\ \hline \end{tabular} \vspace{-0.3cm} \caption{Minimal and mean MSE measures after burn-in across different iterations for GBT and IID models on the 10 alpha matrices from 10 assets. In all cases, $K=10$ is set as the latent dimension. In most cases, the results of IID converge to a smaller value than the GBT model.} \label{table:comparis_gbt_iid_mse} \vspace{-0.1cm} \end{table} \begin{figure*}[h] \centering \vspace{-0.3cm} \subfigtopskip=2pt \subfigbottomskip=9pt \subfigcapskip=-5pt \subfigure[Ten different portfolios where we initialize each portfolio with a unitary value for clarity.]{\includegraphics[width=0.49\textwidth]{imgs/bid_iid_datasets_ashare.pdf} \label{fig:bid_iid_datasets_ashare}} \subfigure[Portfolio values with the same strategy by using different alphas via comparative selection models.]{\includegraphics[width=0.49\textwidth]{imgs/bid_iid_portfolio_ashare.pdf} \label{fig:bid_iid_portfolio_ashare}} \vspace{-0.7cm} \caption{Portfolio values of the ten assets (left), and the portfolio values (right) of different methods where we split by in-sample and out-of-sample periods, and initialize with a unitary value for each period. The proposed IID performs better in the out-of-sample period (see also Table~\ref{table:iid_selected_mean_ic}).} \label{fig:bid_iid_portfolio_ashare_full} \end{figure*} \subsection{Convergence and Comparative Analysis} We first show the rate of convergence over iterations on different assets. Due to space constraints, we omit convergence results for the first five assets and only present those for portfolios SH510050, SH510300, SH601939, SH601628, and SH1303. Results for the other assets are qualitatively similar. We run GBT and IID models with $K=10$ for the five datasets where $214$ is the full rank of the matrices, and the error is measured by MSE. Figure~\ref{fig:iid_alpha_convergence} shows the rate of convergence over iterations. Figure~\ref{fig:iid_alpha_autocorrelation} shows autocorrelation coefficients of samples computed using Gibbs sampling. We observe that the mixings of the IID are close to those of GBT. When the lags are greater than ten, the coefficients are less than 0.1, indicating that the Gibbs sampler mixes well. In all experiments, the algorithm converges in less than 100 iterations. We also observe that the IID model does not converge to a larger error than the vanilla GBT model, though we put more emphasis on selecting the columns with high RankIC. Table~\ref{table:comparis_gbt_iid_mse} presents the minimal MSE and mean MSE after burn-in across different iterations for GBT and IID models on the ten alpha matrices from ten assets. In most cases, the IID can even converge to a smaller MSE value. \begin{algorithm}[!htb] \caption{Alpha selection for portfolio allocation. Gibbs sampler for GBT and GBTN ID models. Select holding period $h$, number of alphas to select $M$. } \label{alg:iid_alpha_selection} \begin{algorithmic}[1] \STATE Split the alpha matrix for in-sample (IS) and out-of-sample (OS) evaluations: $$ \bm{A}_{\text{in}} = \bm{A}_s[:,0:D_{\text{in}}] \in \mathbb{R}^{N\times D_{\text{in}}}, \,\,\,\,\,\,\,\, \bm{A}_{\text{out}} = \bm{A}_s[:,D_{\text{in}}+1:D]\in \mathbb{R}^{N\times (D-D_{\text{in}})}; $$ \STATE Using ID to decide the alphas to be selected on matrix $\bm{A}_{\text{in}}^\top$, with the selected indices $\bm{m}$: $$ \widehat{\bm{A}}_{\text{in}} = \bm{A}_s[\bm{m},0:D_{\text{in}}] \in \mathbb{R}^{\textcolor{blue}{M}\times D_{\text{in}}}, \,\,\,\,\,\,\,\, \widehat{\bm{A}}_{\text{out}} = \bm{A}_s[\bm{m},D_{\text{in}}+1:D]\in \mathbb{R}^{\textcolor{blue}{M}\times (D-D_{\text{in}})}; $$ \FOR{$m=1$ to $M$} \STATE Using the $m$-th IS alpha vector $\bm{a}_m=\widehat{\bm{A}}_{\text{in}}[m,:]\in \mathbb{R}^{D_{\text{in}}}$ to decide the weight $\bm{w}$ and interception $b$ via ordinary least squares (OLS) so that the MSE between the prediction $\bm{a}_m^\top\bm{w}_m +b_m$ and the shifted return vector $\bm{r}^h$ is minimized, i.e., minimizing $\text{MSE}(\bm{a}_m^\top\bm{w}_m +b_m, \bm{r}^h)$. The weight and interception are then used in OS evaluation. \ENDFOR \FOR{$d=1$ to $D-D_{\text{in}}$} \STATE On each day in the OS period, we use the mean evaluation of each prediction from the $M$ alphas to decide to go long or not, i.e., to go long if $\sum_{m=1}^M \bm{a}_m^\top\bm{w}_m +b_m >0$; and do nothing otherwise since we restrict the analysis to long-only portfolios. Though we employ a long-only portfolio, we can favor a \textit{market-neutral strategy}: we open long positions only when we anticipate that at least half of the stocks will rise on the following $h$ day, and we weight each stock equally. \ENDFOR \end{algorithmic} \end{algorithm} \begin{table}[] \centering \vspace{-0.4cm} \setlength{\tabcolsep}{10pt} \small \begin{tabular}{lllll} \hline Methods & Highest RankIC & Randomized ID & BID with GBT & BID with IID \\ \hline Mean RankIC & \textbf{0.1035} & 0.0651 & 0.0553 & \textbf{0.0752} \\ Mean Correlation & 0.2276$\downarrow$ & 0.5741$\downarrow$ & \textbf{0.1132} & \textbf{0.1497} \\ \hline Sharpe Ratio (OS) & 1.0276 & 1.0544 & 0.5045 & \textbf{1.5721} \\ Sharpe Ratio (IS) & \textbf{2.6511} & 1.3019 & 1.4965 & 2.3231 \\ \hline Annual Return (OS) & 0.1043 & 0.0932 & 0.0484 & \textbf{0.1633} \\ Annual Return (IS) & \textbf{0.4390} & 0.2281 & 0.2425 & 0.3805 \\ \hline Max Drawdown (OS) & 0.0632 & \textbf{0.0373} & \textbf{0.0484} & \textbf{0.0552} \\ Max Drawdown (IS) & \textbf{0.0892} & 0.1548 & 0.1232 & \textbf{0.0975} \\ \hline \end{tabular} \vspace{-0.3cm} \caption{Mean RankIC and correlation of the selected alphas across various assets for different methods. A higher mean RankIC and a lower mean correlation are better. The proposed IID method can find the trade-off between the mean RankIC and the mean correlation. In all cases, \textit{IS} means in-sample measurements, and \textit{OS} means out-of-sample measurements. The symbol ``$\downarrow$" means the performance is extremely poor. } \label{table:iid_selected_mean_ic} \end{table} \subsection{Quantitative Strategy} After executing the GBT and IID algorithms for computing the interpolative decomposition of each asset's alpha matrix, the state vector $\bm{r}$ for each asset is saved and the ten alphas with the largest mean selection during the 1,000 iterations are chosen (with a burn-in of 100 iterations, and thinning of 5 iterations). Then we follow the quantitative strategy in Algorithm~\ref{alg:iid_alpha_selection} (in which case $h=1$, $N=214$ alphas, $M=10$ alphas, $D=720$ trading days, and $D_{\text{in}}=480$ trading days). The procedure shown in Algorithm~\ref{alg:iid_alpha_selection} is a very simple quantitative strategy. However, the algorithm can show precisely how the proposed IID method can work in practice. The strategy using the alphas selected by the proposed IID method is only slightly worse than the one selecting the \textit{highest RankIC} alphas for the in-sample (IS) performance in terms of Sharpe ratio, annual return, and maximum drawdown; however, the IID performs better in the out-of-sample (OS) scenario and this is what we actually want (see Table~\ref{table:iid_selected_mean_ic} and Figure~\ref{fig:bid_iid_portfolio_ashare}). To evaluate the strategy, we also adopt the Randomized algorithm to compute the ID for comparison \citep{liberty2007randomized}, termed \textit{Randomized ID}. The Randomized ID performs even worse than BID with GBT (see Table~\ref{table:iid_selected_mean_ic}). Though the IID does not select alphas with the highest RankIC values, this does not mean that the alpha selection procedure is meaningless for the following reasons: 1). \textit{Pool size}: We only use a small alpha pool that only contains 214 alpha factors. When the number of alphas is approaching millions or even billions, the alpha selection procedure is expected to work better. 2). \textit{Correlation}: The mean correlation of selected alphas across the ten assets of the proposed IID method is smaller than the highest RankIC method. In this sense, the alphas of the latter method have high correlations and a low diversity. If the correlated alphas have low liquidity or perform poorly during a given period, the strategy's risk might increase. 3). \textit{Machine learning models}: In our test, we only use OLS to find the weight of each alpha. For more complex models, e.g., neural networks, the correlated alphas can cause multi-linear problems so that the performance and interpretability are hampered. 4). \textit{Diversification}: Even if selecting the alphas with the highest RankIC can work well in practice, we also want to diversify the strategies so that we are not exposed to specific risks. The proposed IID method can help find different strategies. \section{Conclusion} The purpose of this paper is to propose a novel Bayesian identification algorithm that can select the most significant features while still representing the entire feature pool. The proposed IID method is computationally efficient and requires minimal additional processing. Overall, we demonstrate that the convergence results of the presented IID model are comparable to those of the existing GBT model. Similar to vanilla GBT, the IID model can ensure numerical stability by restricting the magnitude of the factored matrix to no more than one.

47ebba210ed7369bb5e594e6b0946f5386ff7e86

\section{INTRODUCTION} Jellium model assumes that the movement of valence electrons in simple metal clusters is carried out in a uniform potential\cite{ekardt_work_1984, ekardt_theory_1986}. The quantum states of valence electrons in the cluster follows the configuration of $1\mathrm{S}^2-1\mathrm{P}^6-1\mathrm{D}^{10}-2\mathrm{S}^2-1\mathrm{F}^{14}-2\mathrm{P}^6-\cdots $, with S, P, D, F the angular momentum. This way, these clusters, also known as superatoms, mimic the chemical behaviour of single atom \cite{shao_density_2015, yan_face-sharing_2019, yan_superatomic_2020, sung_packing_1994, zhang_magnetic_2013, zhang_probing_2010}. Jellium model predicts that superatoms with full electronic shells are endowed with extra stability, giving the magic numbers of 2, 8, 18, 20, 34, $\cdots$. For example, Li$_3$O$^+$ cluster consists of 8 electrons, responsible for a gap as high as 3.72 eV between the highest occupied molecular orbital (HOMO) and the lowest unoccupied molecular orbital (LUMO) \cite{pauna_evolution_2017}. Tetrahedral Au$_{20}$ cluster is of special stability, of which the electronic shell can be viewed as $1\mathrm{S}^2-1\mathrm{P}^6-2\mathrm{S}^2-1\mathrm{D}^{10}$ \cite{cheng_2014_superatom}. Experimental measured HOMO-LUMO gap is 1.77 eV, greater than that in C$_{60}$ \cite{li_au20_2003}. 40-electron Al$_{13}^-$ cluster is known for mimicing hologen atoms, the stability of which originates from closed 2P shell\cite{ bergeron_formation_2004}. \\ Despite the great success that shperical jellium model has achieved on the explanation of stability possessed by closed shell superatoms, the limitation of magic number hinders the diversity of physical and chemical properties of clusters\cite{reveles_2009_designer}. Here we propose that when orbitals are splitted into several subgroups, extra stability can also be achieved if electrons happen to fully or halfly occupy one subgroup. In this work, we report the splitting of originally degenerate superatom orbitals through our investigation of orbitals in Li$_{19}$ clusters, the number of electrons $n=19$ is exactly between magic numbers 18 and 20. When a transition metal cation is located in the center of an octahedral cage of oxygen anions, due to the electric repulsion produced by ligands, its d orbitals splits into two sets, which is the core idea of crystal field theory \cite{kugel_1982_jahn}. Orbitals in superatoms, e.g., 1P and 1D, will also split into several groups, only the intrinsic dirving force is not ligand's electric field but the deformation of cluster shape. \\ Lithium, the simplest alkali metal with electronic shell of $1\mathrm{s}^2-2\mathrm{s}^1$, is an ideal prototype for simple metal \cite{cheng_2013_communication, yan_seventeen-coordinate_2019, yan_face-sharing_2019, yan_superatomic_2020}. Li$_{19}$ cluster owns 19 electrons, and its electrons can fully occupy orbitals from 1S to 1D, providing convenience for analysis of electronic orbitals. Based on a combined searching strategy and DFT calculation, the optimal configuration of Li$_{19}$ cluster is obtained along with eight low-lying isomers. The superatom orbitals are studied and it is found that their electronic configurations can be divided into two types, corresponding to two different cluster shapes, prolate ellipsoid and oblate ellipsoid, respectively. Clusters in the shape of oblate ellipsoid have the electronic configuration of $1\mathrm{S}^2-1\mathrm{P}^6-1\mathrm{D}_{xy/x^2-y^2}^4-2\mathrm{S}^2-1\mathrm{D}_{xz/yz}^4-1\mathrm{D}_{z^2}^1$, while clusters in the shape of prolate ellipsoid have the configuration of $1\mathrm{S}^2-1\mathrm{P}^6-1\mathrm{D}_{z^2}^2-1\mathrm{D}_{xz/yz}^4-1\mathrm{D}_{xy/x^2-y^2}^4-2\mathrm{S}^1$. We conclude that for originally degenerate orbitals, the more the orbital extends in the direction of the cluster extension, the lower the energy becomes. It is found that this relationship can be applied to explain the stability of Li$_{14}$ and predict the shape as well as magnetic moment of alkali clusters. Our results based on superatomic orbital splitting effect analysis is a powerful tool and theoretically sound for explaining and predicting the structure and stability of simple alkali metal clusters. \section{COMPUTATIONAL DETAILS} A large number of initial isomers are generated by the following methods in order to carry out an unbiased search for the global-minimum structures of Li$_{19}$: \\ \begin{enumerate} \item Crystal structure AnaLYsis by Particle Swarm Optimization (CALYPSO) software package developed by Yanchao Wang et.al \cite{lv_2012_particle}. \item Blending smaller clusters. Place several smaller clusters (19 lithium atoms in total) at random relative position, which simulates the formation of large clusters through collision of small clusters. \item Constructing geometrically symmetric system. We randomly fabricated polygons with high mathematical symmetry and fit lithium atoms in geometric vertex properly. \end{enumerate} All Electron Relativistic methods were subsequently performed for further geometric optimization with Dmol3 package in $Materials Studio$. The Perdew and Wang’s 1991 exchange and correlation functional (PW91) functional \cite{wang_1991_correlation} was continually selected since the quality has been verified by our previous work in V@Li$_n$ superatom clusters \cite{zhang_magnetic_2013}. The highest precision basis set of double numerical plus polarization with addition of diffuse functions (DNP+) was chosen. Cut-off energy and the self-consistent field (SCF) convergence tolerance were both set as Fine. The frequencies of the configurations were calculated and there are no imaginary frequencies for the obtained configuration with lowest energy. For odd or even number of valence electrons, the potential spin multiplicity may be 2, 4, 6, 8... or 1, 3, 5, 7..., respectively. Each possible spin multiplicity of the clusters was tested and value with the lowest energy is picked. \\ \section{RESULTS AND DISCUSSION} \subsection{Identification of Geometric Configurations} Although a large variety initial structures were obtained from the above three different methods, the lowest energy structrue obtained by further DFT optimization of these initial configurations are indeed the same. We listed 9 isomers with distinct geometry in FIG. \ref{Li19_stru} sorted by energy from low to high. \\ \begin{figure}[htbp] \centering \includegraphics[scale=0.07]{stru2.png} \caption{Nine low energy structures of Li$_{19}$ marked by letter A to I. Their relative energies are calculated.} \label{Li19_stru} \end{figure} The ground state structure for Li$_{19}$ clusters is an oblate cage with three lithium atoms inside as marked A in FIG. \ref{Li19_stru}, which is consistent with the result predicted by Sung using simulated annealing method \cite{sung_packing_1994}. According to the jellium model, alkali metal clusters with closed-shell electronic configurations usually adopt high symmetry geometries \cite{luo_special_2014}. Since Li$_{19}$ cluster has total 19 valence electrons, which is between two magic number of 18 and 20, it is in line with expectations that structure of the Li19 has low C$_{2v}$ symmetry and does not tend to form a spherical shape. Another intriguing structure is E with the symmetry of D$_{5h}$. Its shape is a prolate cage-like structure with a morphology similar to that of Chinese lanterns. Four atoms sit on its central axis, encircled by three staggered 5-membered rings. C and H structures evolve from E and I structures respectively with one atom moves from the top or buttom of the original structures to their side. These two isomers C and H are considered as the transitiaoal structures because they are because it is somewhere in between prolate ellipsoid and oblate ellipsoid. Binding energies of oblate C and H are lower than their parent structures E and I, respectively, indicating that Li$_{19}$ is more inclined to form oblate ellipsoid clusters. This structural characteristic is also verifed by the most stable isomer A with the oblate structure. \\ \subsection{Superatom Orbitals in Li$_{19}$} \begin{figure*} \centering \includegraphics[scale=0.1]{orbital2.png} \caption{The molecular orbitals (MOs) and MO energy level diagrams of the pristine Li19 cluster for structure C, E and A respectively. The color bar on the scale axis represents the energy range covered by corresponding sets of MOs. The shapes of MOs are drawn outside the scale axis, and the type and number of each MOs are marked nearby. To represent the energy level order for non-degenerate orbitals, the orbitals drawn before are lower in energy while the orbitals in the same radial position have similar energies. The highest occupied molecular orbital (HOMO) and lowest unoccupied molecular orbital (LUMO) energy levels are marked by dotted lines.} \label{orbital2} \end{figure*} \begin{figure*} \centering \includegraphics[scale=0.6]{Li19-level.png} \caption{Energy level of superatom orbitals of Li$_{19}$ clusters where solid lines with different colors represent different energy levels. Orbitals with strange shapes that cannot be classified as known orbitals are classified as Non-Classical.} \label{Li19level} \end{figure*} Orbital analysis reveals that Li$_{19}$ clusters possess superatom orbitals and mimic the s, p, d orbitals of simple atom. As each lithium atom provides one delocalized electron, the number of effective valence electrons is 19. Despite having a open shell, it is found that molecular orbitals of Li$_{19}$ clusters can be discussed under jellium model. FIG. \ref{orbital2} displays orbitals of three typical structures: A, C and E as representatives of oblate cluster, transition shape cluster and prolate cluster respectively. All these three sets of orbitals show a lot of similarities to the atomic orbitals.The spherical 1S orbital is followed by the dumbbell-shaped 1P orbital, then comes the petal-like 1D orbital together with 2S orbital. The identification of orbitals relies on the orientation of the coordinate axes, which is determined by the following method: we take advantage of co-planar 1D$_{xy}$/1D$_{x^2-y^2}$ to fix x-y plane, thus the direction penperdicular to the plane is z axis. \\ As one might note the orbital alignment sequences of three isomers in FIG. \ref{orbital2} differ from one another. For the most stable configuration of pure Li$_{19}$ marked as A, 1D orbitals split into three groups, $1\mathrm{D}_{xy}/1\mathrm{D}_{x^2-y^2}$ with the lowest energy, $1\mathrm{D}_{xz}/1\mathrm{D}_{yz}$ with the second lowest energy, and $1\mathrm{D}_{z^2}$ with the highest energy. 1P orbitals also splitted into two groups and the energy of $1\mathrm{P}_{x/y}^{}$ is lower than the energy of $1\mathrm{P}_z$. The more the orbital stretches in the $z$ direction, the higher the energy become. For prolate configuration E, the sequence of splitted orbials is dramatically opposite. 1D orbitals with lowest energy are $1\mathrm{D}_{z^2}$, followed by $1\mathrm{D}_{xz}/1\mathrm{D}_{yz}$, and highest energy level orbitals are $1\mathrm{D}_{xy}/1\mathrm{D}_{x^2-y^2}$. The situation is same in 1P orbitals. The more the orbital stretches in the z direction, the lower the energy. Orbitals in transition structure C basically follows the pattern of of the prolate ellipsoid, but also begins to show some signs of transition. For example, the energy of $1\mathrm{D}_{z^2}$ is higher than that of $1\mathrm{D}_{xz}/1\mathrm{D}_{yz}$ in isomer C. Is this a random case or is there an underlying pattern? To explore the underlying pattern of superatom orbitals, the detailed molecular orbitals and orbital energies of the clusters are calculated and identified. FIG. \ref{Li19level} gives a general overview of energy levels of nine low energy structures of Li$_{19}$ given in FIG. \ref{Li19_stru}. It shows that the energy levels are nondegenerate and split in certain order. For geometry A, B, D and F, the orbital sequence is $1\mathrm{S}^2-1\mathrm{P}^6-1\mathrm{D}_{xy/x^2-y^2}^4-2\mathrm{S}^2-1\mathrm{D}_{xz/yz}^4-1\mathrm{D}_{z^2}^1$. For geometry E, G and I, the sequence is $1\mathrm{S}^2-1\mathrm{P}^6-1\mathrm{D}_{z^2}^2-1\mathrm{D}_{xz/yz}^4-1\mathrm{D}_{xy/x^2-y^2}^4-2\mathrm{S}^1$. For C, as analysed previously, the orbitals mainly draws out the properties of prolate clusters along with the transition. For H, because of the irregular geometric structure, the order of its energy levels is also not obvious. Besides, the exsistance of nonclassical orbitals caused by irregular geometry also hindered the summarization of the pattern. \\ \begin{figure*} \centering \includegraphics[scale=0.15]{alpha.png} \caption{Fitting results of the flattening of seven lithium clusters, atoms inside the cluster have a weight of 0 in the fit. $\alpha =\left( a-b \right) /a$ describes the flattening of the ellipsoid and $R_s$ is the squared norm of the residual. } \label{alpha} \end{figure*} We linked the pattern to the flattening of ellipsoid of cluster shapes. Function $\left( x^2+y^2 \right) /a^2+z^2/b^2=1$ is applied to fit the shape of cluster as an rotational ellipsoid and the flattening is defined as $\alpha =\left( a-b \right) /a$ as Clemenger pointed out in 1985 \cite{clemenger_ellipsoidal_1985}. For prolate ellipsoid $\alpha <0$ while for oblate ellipsoid $\alpha > 0$. As in FIG. \ref{alpha}, geometry A, B, D, F are oblate ellipsoids with $\alpha$ ranging from 0.33 to 0.47, geometry E, G, I are prolate ellipsoids with $\alpha$ ranging from -0.29 to -0.19. By such classification, we can draw the conclusion that oblate clusters have the orbital sequence of $1\mathrm{S}^2-1\mathrm{P}^6-1\mathrm{D}_{xy/x^2-y^2}^4-2\mathrm{S}^2-1\mathrm{D}_{xz/yz}^4-1\mathrm{D}_{z^2}^1$, while prolate clusters have the sequence of $1\mathrm{S}^2-1\mathrm{P}^6-1\mathrm{D}_{z^2}^2-1\mathrm{D}_{xz/yz}^4-1\mathrm{D}_{xy/x^2-y^2}^4-2\mathrm{S}^1$. For originally degenerate orbitals, the more the orbital extends in the direction of the cluster extension, the lower the energy becomes (e.g. energy of 1D$_{z^2}$ is lower than 1D$_{xz/yz}$ in prolate cluster because 1D$_{z^2}$ has a greater lob in $z$ axis). \\ \subsection{Prediction of Stability} Futhermore, orbital splitting effect can be generalized to explain the stabilities of alkali metal systems with ellipsoidal structures, such as the Li$_{14}$ cluster. In the literature \cite{cheng_2013_communication}, the stability of the Li$_{14}$ cluster is explained as the combination of two Li$_{10}$ clusters with super-covalent bonds, mimicing fluorine molecule. However, superatomic orbitals of Li$_{14}$ from 43(LUMO)-48 do not correspond to any orbital of fluorine molecule according to our molecular orbital analysis. Within the view of orbital splitting effect, the structure and stability can be successfully interpreted. As illustrated in FIG. \ref{Li14}, the prolate structure of Li$_{14}$ will lead to the split of 1D energy levels. As a result, $1\mathrm{D}_{z^2}$ lies at the lowest energy levels, following the next higher energy of $1\mathrm{D}_{xz/yz}$ MOs. The $1\mathrm{D}_{z^2}$ and $1\mathrm{D}_{xz}/1\mathrm{D}_{yz}$ are all doubly occupied resulting in a large HOMO-LUMO gap of 1.60 eV. This interpretation based on splitted orbitals clearly shows the nonspherical clusters can achieve good stability without meeting the requirement of magic number.\\ \begin{figure}[H] \centering \includegraphics[scale=0.08]{JT-Li14.png} \caption{Calculated energy levels and electronic orbitals in Li$_{14}$ clusters.} \label{Li14} \end{figure} \subsection{Prediction of Geometric Structure} It should be pointed out that the orbital splitting effect of Li$_{19}$ cluster proposed here can be extended to describe and predict the geometric shapes of the pure alkali metal species with other size. Based on previous discussion, the electron configuration of oblate, prolate and spherical jellium clusters can be theoretically predicted, and we are able to fill electrons in properly to connect the shape of cluster with the number of valence electrons ($n$). The rule of filling electrons are proposed in analogy with single element atoms: \begin{enumerate} \item In all possible electron configurations, the term with fully occupied (degenerate) HOMOs has the lowest energy. This rule is based on the experimental fact that high peaks in mass spectrum correspond to closed superatom shells \cite{knight_1984_electronic, knight_electronic_1985}. \item When a closed shell cannot be achieved, the configuration with halfly occupied (degenerate) HOMOs has the lowest energy. \item When previous goals can't be satisfied, considering the fact that repulsion of electrons may raise the energy, we suppose HOMOs with less electrons occupied has lower energy. \end{enumerate} \begin{figure*}[htbp] \centering \includegraphics[scale=0.1]{occupy.png} \caption{Preferred electron occupation for (a) prolate, (b) oblate and (c) spherical clusters.} \label{occupy} \end{figure*} As shown in FIG. \ref{occupy}, when $n=3(4)$, electrons can halfly (fullly) fill $1\mathrm{P}_z$ orbital, so the shape of cluster should be prolate. When $n=5$, three electrons halfly occupy three degenerate 1P orbitals, so the cluster will be shperical. When $n=6$, electrons can either halfly occupy $1\mathrm{P}_{x/y}$ orbitals with fully occupied $1\mathrm{P}_z$ in prolate configuration or fully occupy $1\mathrm{P}_{x/y}$ orbitals with empty $1\mathrm{P}_z$ in oblate configuration. Since fully occupied HOMOs are more stable, the cluster will be prolate. When $n=11$, the cluster will be prolate because in this case there will be only one electron in the HOMOs instead of three when the cluster is oblate. To sum up, for prolate clusters, numbers match electron configuration are $n = 3, 4, 9, 10, 11, 14, 15, \cdots$ For a cluster with oblate shape, the occupancy agrees with $n = 6, 7, 12, 16, 17, \cdots$ For spherical clusters, preferred numbers are $n=5, 8, 13, 18 \cdots$ The results are summarized in TABLE.\ref{shape-pref}. \\ \begin{table}[htbp] \centering \caption{Preferred shapes for clusters with different number of valence electrons($n$). '$\surd$' means 'preferred'.} \begin{ruledtabular} \begin{tabular}{cccccccc} n & prolate & oblate & spherical & n & prolate & oblate & spherical \\ \colrule 3 & $\surd$ & & & 11 & $\surd$ & & \\ 4 & $\surd$ & & & 12 & & $\surd$ & \\ 5 & & & $\surd$ & 13 & & & $\surd$ \\ 6 & & $\surd$ & & 14 & $\surd$ & & \\ 7 & & $\surd$ & & 15 & $\surd$ & & \\ 8 & & & $\surd$ & 16 & & $\surd$ & \\ 9 & $\surd$ & & & 17 & & $\surd$ & \\ 10 & $\surd$ & & & 18 & & & $\surd$ \\ \end{tabular}% \end{ruledtabular} \label{shape-pref}% \end{table}% Once the configuration is determined, the magnetic moment of each cluster equal to the number of unpaired electrons (FIG. \ref{mub}). Compared with DFT calculated results, the prediction reproduced odd-even oscillation of magnet moments. In particular, the predicted magnetic moment of Li$_{13}$ clusters is 5$\mu_B$, in good agreement with the DFT theory calculations. It seems that we have overestimated the magnet moment in Li$_{5}$ while underestimated that in Li$_{16}$, the reasons will be analysed later. \\ \begin{figure}[H] \centering \includegraphics[scale=0.5]{mub.png} \caption{Predicted magnetic moments for Li$_n$ clusters compared with the results calculated by DFT.} \label{mub} \end{figure} \begin{figure*}[htbp] \centering \includegraphics[scale=0.1]{pred.png} \caption{Lowest energy configurations of Li$_n(n=3-18)$ clusters \cite{zhang_magnetic_2013, fournier_theoretical_2003}. Fitted $\alpha$ is applied to specify the mophorlogy (prolate, oblate or shperical) of the cluster. } \label{pred} \end{figure*} When considering the lowest energy structures of $\mathrm{Li}_n$ clusters (FIG. \ref{pred}), in line with expectations, most of the shapes satisfy our prediction. When $n = 3$, instead of forming an equilateral triangle, Li$_3$ forms an isosceles triangle which can be viewed as a prolate structure stretched from a equilateral triangle in order to make the cluster more stable. Li$_4$ ($n=4$) has a tendency to be prolate so it will not form a regular tetrahedral (a square) which is close to spherical (oblate). For Li$_{10}$, Li$_{11}$, Li$_{14}$ and Li$_{15}$, the optimized equilibrium structures are prolate ellipsoid with their orbital configurations matching FIG. \ref{occupy} perfectly. \\ The most stable structures of Li$_{6}$ and Li$_{7}$ are all oblate ellipsoids, namely, Li$_{6}$ and Li$_{7}$ are squared pyramidal and pentagonal bipyramid respectively, which are consistent with theoretical prediction.\\ Li$_{8}$, Li$_{13}$ and Li$_{18}$ clusters have quite spherical shape and with high $\mathrm{D}_{5\mathrm{h}}$, $\mathrm{I}_{\mathrm{h}}$, $\mathrm{C}_{\mathrm{s}}$ symmetry respectively. Li$_{13}$ has 5 parallel 1D orbitals resulting in a high magnetic moment of 5$\mu_B$, as forecasted in FIG. \ref{mub}. We can see that the model of orbital splitting can be successfully applied to interpret the structure and stability of other alkali metal clusters and the model is likely to be extended to $n>18$ clusters after taking into accunt of 1F, 2S, 2P... orbitals. It is also expected to be useful for the study of other types of doped metal clusters.\\ Although the geometric structures of the pure lithium clusters is overwhelmingly consistent with our predictions, there are also a few deviations. These exceptions may arise due to the multiple factors that determine the cluster structure. First of all, Jahn-Teller distortion is known to reduce the symmetry of small clusters and cause a low spin state \cite{khanna_1987_jahn}. Li$_{5}$ was set to be a high spin sphere with $3\mu_B$ but transformed to be a prolate spheroid with $1\mu_B$. Li$_{9}$ was predicted to be prolate but turned out spherical with no explicit $z$ direction. Likewise, Li$_{12}$ transformed from a prolate sphroid to a tri-axial ellipsoid with $a<b<c$ which leads to nondegeneration of the $1\mathrm{D}_{xz}$ and $1\mathrm{D}_{yz}$. The final electron configuration for Li$_{12}$ is $1\mathrm{S}^2-1\mathrm{P}_{z}^{2}-1\mathrm{P}_{y}^{2}-1\mathrm{P}_{x}^{2}-1\mathrm{D}_{z^2}^{2}-1\mathrm{D}_{yz}^{2}\left( \mathrm{HOMO} \right) -1\mathrm{D}_{xz}\left( \mathrm{LOMO} \right) $. Moreover, Hund's rule is reported to favour high spin state in cluster \cite{medel_hunds_2011, geguzin_1981_spin}, which may be responsible for prolate Li$_{16}$ with $2\mu_B$. In addition, we ignored the intercalation of 2S orbitals. It has been calculated that 2S inserted between $1\mathrm{D}_{xy/x^2-y^2}^{4}$ and $1\mathrm{D}_{xz/yz}^{4}$ in Li$_{19}$ while it inserted between 1P and 1D in Li$_{6}$. Li$_{17}$ is a prolate ellipsoid which has 2S orbitals intercalated in $1\mathrm{D}_{xy/x^2-y^2}^{}$, causing a orbital configuration of $1\mathrm{S}^2-1\mathrm{P}_{z}^{2}-1\mathrm{P}_{x/y}^{4}-1\mathrm{D}_{z^2}^{2}-1\mathrm{D}_{xz/yz}^{4}-1\mathrm{D}_{xy/x^2-y^2}^{3}\left( \mathrm{HOMO} \right) -2\mathrm{S}\left( \mathrm{LOMO} \right)-1\mathrm{D}_{xy/x^2-y^2}^{} $. There only exsist a small gap (0.148 eV) between HOMO and LUMO. It is predictable that this effect will be negligible for $n>20$ clusters.\\ \section{CONCLUSION} In conclusion, our work offered a method to predict the magnetic moment and shape of alkali clusters through the analysis of superatom orbitals. The molecular orbitals of nine low-lying Li$_{19}$ clusters have been calculated, identified and compared. It has been found that clusters in the shape of oblate ellipsoid have the electronic configuration of $1\mathrm{S}^2-1\mathrm{P}^6-1\mathrm{D}_{xy/x^2-y^2}^4-2\mathrm{S}^2-1\mathrm{D}_{xz/yz}^4-1\mathrm{D}_{z^2}^1$, while clusters in the shape of prolate ellipsoid have the sequence of $1\mathrm{S}^2-1\mathrm{P}^6-1\mathrm{D}_{z^2}^2-1\mathrm{D}_{xz/yz}^4-1\mathrm{D}_{xy/x^2-y^2}^4-2\mathrm{S}^1$. For originally degenerate orbitals, the more the orbital extends in the direction of the cluster extension, the lower the energy becomes. The applications of the orbital splitting are promoted in two following ways. The stability of the Li$_{14}$ cluster are explained in the framework of orbital splitting effect without resorting to the SVB model. Meanwhile, it is proposed that the orbital splitting effect can be used to predict the shape and magnetic moment of clusters. 11 out of 16 predicted shapes of Li$_n(n=3-18)$ are consistent with the results obtained by the principle of minimum energy. \begin{acknowledgements} The authors thank ShangHai Students' Innovation and Entrepreneurship Program (No.S20091) for supporting this research. \end{acknowledgements} \bibliographystyle{unsrt}

f8cb122b523167ea439df8a71e8d404b7d506d97

\section{Introduction} Numerical flow simulations of super- and transcritical conditions require appropriate thermodynamic models. A state-of-the-art thermodynamic model is based on a cubic equation of state (EoS) and departure function formalism for the evaluation of enthalpy and energy. This can be found in early~\citep{zong2004numerical,kim2011numerical,park2012and,petit2013large} as well as recent computational fluid dynamics (CFD) investigations~\citep{fathi2022large,traxinger2018experimental,Matheis:2017dx,muller2016large,sharan2021investigation,lagarza2019large,poblador2022volume}. For the cubic EoS mostly the Peng-Robinson~\citep{PR} (PR) or the Soave-Redlich-Kwong EoS~\citep{SRK,soave1972equilibrium} (SRK) are employed. Both assume a universal critical compressibility and are therefore only suitable for fluids with a matching critical compressibility. Volume translation methods~\citep{abudour2012volume,martin1979cubic,harstad1997efficient,matheis2016volume} represent one possible solution to improve the density prediction for fluids, that are not well described by SRK or PR. \citet{Cismondi} suggested the Redlich-Kwong-Peng-Robinson EoS (RKPR), introducing a third EoS parameter and formulating all three EoS parameters as a function of the critical compressibility. Building upon their work, \citet{Kim} presented a thermodynamic modeling approach based on the RKPR and demonstrated its advantages and suitability for different fluids. Despite its advantages, only a few recent studies~\citep{fathi2022large,jung2020real} have employed the RKPR EoS recently for real gas CFD simulations of n-dodecane injections. Within the suggestion of the RKPR, \citet{Cismondi} and earlier Mollerup (see \citet{michelsen2004thermodynamic}), also proposed a general formulation of the cubic EoS by which all of the well-known cubic EoS can be described with a particular set of values. Such a formulation allows for a modularized implementation of all these cubic EoS, thus, less code duplication and a better readability. An alternative to a cubic EoS is the PC-SAFT EoS (perturbed-chain statistical associating fluid theory)~\citep{gross2001perturbed}, which has been successfully employed by \citet{rodriguez2019simulation,rodriguez2018simulation} for two-dimensional CFD simulations. Another approach is the usage of tabulated reference data, yielding higher accuracy~\citep{Koukouvinis} and also a potentially faster evaluation of the thermodynamic data~\citep{Doehring,jafari2022exploring}. However, cubic EoS are still mostly used due to their simplicity and overall good accuracy. In addition to the EoS and the relations for thermodynamic properties, CFD simulations also require relations for the transport properties viscosity and thermal conductivity. \citet{Chung} proposed correlations for these transport properties, which are often employed for such real gas simulations~\citep{matheis2016volume,Matheis:2017dx, Matheis, fathi2022large,traxinger2018experimental, muller2016large}. Alternative methods are for example, the Lucas method~\citep{Poling} for the viscosity and the Stiel--Thodos method~\citep{Poling} for the thermal conductivity. These methods have been recently employed by \citet{sharan2021investigation} for nitrogen and showed a good agreement with NIST reference data. Alternatively, the residual entropy scaling technique can be used for the calculation of the thermal conductivity~\citep{hopp2017thermal} and the viscosity~\citep{looetgering2015group} as done by \citet{Koukouvinis}. In general, it is an important step in CFD simulations to check the accuracy and suitability of a thermodynamic model in advance, as included in several studies~\citep{Matheis:2017dx,Koukouvinis,traxinger2018experimental,doehring2022nicfd}. For different fluids, pressure and temperature ranges, such an evaluation can be complicated and time-consuming. Apart from that, such a verification requires an already successful implementation of the thermodynamic model. This is usually not the case during the development or further development of a CFD solver. The open source library CoolProp~\citep{bell2014pure} provides implementations for the SRK and PR EoS. Bell et al.~\citep{Bell2016Thermo} wrote a comprehensive thermodynamic library to evaluate chemical properties specifically targeted for chemical engineering. Our new tool \texttt{realtpl}, on the other hand, has been specifically designed for applications in the context of CFD simulations and evaluates the entire thermodynamic model required for these simulations. In this paper, we aim to further promote the idea of the generalized formulation of cubic EoS by Mollerup~\citep{Cismondi,michelsen2004thermodynamic}, i.e., one formulation for all three cubic EoS (PR, SRK, and RKPR). To this end, we describe in detail how this formulation is solved and present the resulting relations for the thermodynamic properties. We also outline the overall thermodynamic model based on this generalized formulation. For the thermodynamic model, we employ the Chung correlations for the evaluation of the transport properties. We apply the thermodynamic model to selected fluids and study its suitability. Therewith, we also demonstrate the good applicability of the RKPR for all fluids with different critical compressibility factors. In order to test the proposed thermodynamic model and to apply it to different configurations, we provide an open source Python tool called \texttt{realtpl} for real gas thermodynamic python library. This tool can be used to evaluate and compare the results for a wide range of different fluids. Additionally, we also provide the implementation of the generalized form in OpenFOAM. The paper is structured as follows: \Cref{s:ThermoMod} presents the thermodynamic model based on the generalized cubic EoS. Then, in \cref{s:Accuracy}, the applicability and suitability of the thermodynamic model is assessed for selected fluids comparing the model using SRK, PR, and RKPR. \Cref{s:SupMat} contains information about the additionally provided Python tool \texttt{realtpl} and the validation of the proposed OpenFOAM implementation based on the generalized formulation. Finally, the paper is summarized in \cref{s:Con}. \begin{table*}[!tb] \caption{Parameters of the EoS adopted from \citet{Kim}.} \label{tab:KimTabelle} \centering \begin{tabular}{p{0.5cm}p{3.7cm}p{3.7cm}p{6.5cm}} \hline & \textbf{\small SRK \citep{SRK}} & \textbf{\small PR \citep{PR}} & \textbf{\small RKPR \citep{Cismondi}} \\ \hline $\delta_{1}$ & $1$ & $1+\sqrt{2}$ & $d_{2}+d_{1}(d_{3}-c_z Z_{c})^{d_{4}}+d_{5}(d_{3}-c_z Z_{c})^{d_{6}}\newline \mathrm{with}\; c_z = 1.168, \newline d_1 = 0.428363,\; d_2 = 18.496215, \newline d_3 = 0.338426,\; d_4 = 0.660000, \newline d_5 = 789.723105,\; d_6 = 2.512392$\\ \hline $\delta_{2}$ & $0$ & $1-\sqrt{2}$ & $(1-\delta_{1})/(1+\delta_{1})$\\ \hline $a$ & $0.42747\left(\frac{{R}^{2}T_{c}^{2}}{p_{c}}\right)$ & $0.45724\left(\frac{{R}^{2}T_{c}^{2}}{p_{c}}\right)$ & $\frac{3y^{2}+3yd+d^{2}+d-1}{(3y+d-1)^{2}}\left(\frac{{R}^{2}T_{c}^{2}}{p_{c}}\right)$ \\ \hline $b$ & $0.08664\left(\frac{{R}T_{c}}{p_{c}}\right)$ & $0.0778\left(\frac{{R}T_{c}}{p_{c}}\right)$ & $\frac{1}{3y+d-1}\left(\frac{{R}T_{c}}{p_{c}}\right)$ \\ \multicolumn{1}{l}{} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{} & \multicolumn{1}{l}{with \enspace $d=\frac{1+\delta_{1}^{2}}{1+\delta_{1}}$} \\ \multicolumn{1}{l}{} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{$y = 1+[2(1+\delta_{1})]^{\frac{1}{3}}+\left(\frac{4}{1+\delta_{1}}\right)^{\frac{1}{3}}$} \\ \hline $\alpha$ & $\left(1+\kappa(1-\sqrt{T/T_{c}})\right)^{2}$ & $\left(1+\kappa(1-\sqrt{T/T_{c}})\right)^{2}$ & $(3/(2+T/T_{c}))^{\kappa}$ \\ \hline $\kappa$ & $0.48508 \newline + 1.55171\,\omega \newline- 0.15613\,\omega^2$ & $0.37464 \newline + 1.54226\,\omega \newline - 0.26992\,\omega^2$ & $(66.125\, c_z Z_{c} - 23.359)\omega^{2} \newline +(-40.594\, c_z Z_{c} + 16.855)\omega \newline +(5.27345\, c_z Z_{c} - 0.25826) $\\ \hline \end{tabular}% \end{table*}% \section{Thermodynamic model based on a generalized cubic equation of state} \label{s:ThermoMod} We present a thermodynamic model based on the generalized cubic EoS. First, we present the EoS and describe in detail how it is solved. Then, the correlations to evaluate the thermodynamic properties are presented and, finally, the Chung correlations for the transport properties are briefly described. \subsection{Generalized cubic equation of state} \label{ss:eos} We here consider the generalized formulation of a cubic EoS suggested by \citet{Cismondi} and already earlier by Mollerup~\citep{michelsen2004thermodynamic} \begin{equation} p({v},T)=\frac{{R}T}{{v}-b}-\frac{a \alpha}{({v}+\delta_1b)({v}+\delta_2b)}\,. \label{eq:genEq} \end{equation} The pressure $p$ is a function of the molar volume ${v}$ and the temperature $T$. ${R}$ denotes the universal gas constant with ${R} = 8314.472\;\si{J/(kmol\,K)}$. $a$ and $b$ represent the two traditional EoS parameters, considering attractive forces with $a$ and repulsive forces by the the effective molecular volume $b$. Both are determined by a proportionality factor and the critical properties $p_c$ and $T_c$ of the fluid, see \cref{tab:KimTabelle}. Further, $a$ is multiplied by a correction factor $\alpha$ that is a function of reduced temperature $T/T_c$ and the acentric factor $\omega$. It is worth noting that for $a=0$ and $b=0$, the cubic EoS collapses to the ideal gas law. As a consequence, mathematically, and also physically, the molar volume ${v}$ has to be larger than the co-volume $b$ (${v} > b$). The common cubic EoS can be described with special pairs of the values $\delta_1$ and $\delta_2$ , where $\delta_2$ is a supplementary parameter defined as $(1-\delta_1)/(1+\delta_1)$. Multiplying the denominator out results in the well-known and often used formulation of \begin{equation} p({v},T)=\frac{{R}T}{{v}-b}-\frac{a \alpha}{{v}^2 + u b {v} + w b^2 }, \label{eq:genEq_old} \end{equation} where $u=\delta_1+\delta_2$ and $w=\delta_1\delta_2$. However, the first formulation (\cref{eq:genEq}) yields simpler expressions of the derivations required for evaluating the thermodynamic properties (see \cref{ss:caloric}) than \cref{eq:genEq_old}. For the widely used EoS SRK and PR, the proportionality factor in $a$ and $b$ is constant and $\delta_1$, or respectively $u$ and $w$, are constants with $\delta_1 = 1$ ($u= 1$, $w = 0$) for SRK and $\delta_1 = 1 + \sqrt{2}$ ($u=2$, $w=-1$) for PR. Hence, for SRK and PR a universal critical compressibility has been assumed, which is about 0.285 for SRK and 0.263 for PR~\citep{Kim}. Therefore, these two EoS are only well suited for a certain set of fluids with a corresponding similar critical compressibility. To overcome this limitation, \citet{Cismondi} suggested to evaluate the EoS parameters as a function of the critical compressibility resulting in the RKPR EoS. For the detailed evaluation of the EoS parameters see \cref{tab:KimTabelle}. For the RKPR, a different correlation than for SRK and PR is used to evaluate $\alpha$. Consequently, also the derivatives by temperature $\partial \alpha/\partial T$ and $\partial^2 \alpha/\partial T^2$, required for the evaluation of the thermodynamic properties, differ for the EoS. Concluding, all three cubic EoS SRK, PR, and RKPR can be described by \cref{eq:genEq}, where only the EoS parameters $a$, $b$, and $\delta_1$ as well as the evaluation of $\alpha$ changes. In order to solve the cubic EoS (\cref{eq:genEq}), the equation is reformulated using the dimensionless compressibility factor \begin{equation} Z = \frac{p {v}}{{R}T}, \label{eq:Z} \end{equation} as well as dimensionless expressions for $a$ and $b$ with $A=pa\alpha(RT)^{-2}$ and $B=pb(RT)^{-1}$. Therewith, one obtains from \cref{eq:genEq} \begin{equation} Z =\frac{1}{1-B/Z}-\frac{A}{B}\;\frac{B/Z}{(1 + \delta_1B/Z)(1+ \delta_2B/Z)}\,. \label{eq:Zeq2} \end{equation} Recasting results in the cubic form of all considered EoS in terms of $Z$, which reads \begin{equation} Z^3 + a_2 Z^2+ a_1 Z + a_0 =0 \label{eq:cubicFormZ} \end{equation} with the coefficients \begin{equation} a_2 = B(\delta_1+\delta_2-1)-1, \label{eq:a2} \end{equation} \begin{equation} a_1 = A +\delta_1\delta_2 B^2-(\delta_1+\delta_2) B (B+1), \label{eq:a1} \end{equation} \begin{equation} a_0 = -B(\delta_1\delta_2 B^2+\delta_1\delta_2 B+A). \label{eq:a0} \end{equation} The obtained cubic equation can be solved for real roots, which is well described in the literature~\citep{elliott2012introductory,education_psu,Matheis}. A cubic equation has either one real and two imaginary roots or three real roots. At supercritical conditions usually only one real root is present. For this reason, we recommend first checking for the existence of one real root when evaluations are focused on supercritical conditions. Three real roots are generally associated with the two-phase region present at sub-critical conditions. As mentioned above, the volume has to be larger than the co-volume $b$ (${v} > b$), or expressed in terms of the compressibility factor $Z>B$. If the physical constraint $Z>B$ is full-filled, the smallest root represents the liquid state and the largest root the vapor or gas state. From a thermodynamic perspective, the center root is not stable and therefore physically meaningless. The correct root of the two physically stable roots can be identified by comparing the Gibbs energy, see \cref{ss:caloric}. An alternative approach is to take always the largest root, which usually corresponds to the vapor/gaseous state (exception see below). At very high and low pressures, the smallest root can be smaller than the co-volume ($Z\leq B$). In this case, there is only one physical meaningful root, which is the largest one. At high pressures, this root then corresponds to a liquid state, while at very low pressures it corresponds to the gaseous/vapor state. It is important to note that this can also occur in a clearly supercritical regime, such as for n-dodecane at the ECN-Spray A condition with $p=8\,\si{MPa}$ in the temperature range $T = 1097 - 1500 \, \si{K}$ using PR or RKPR. Consequently, a missing check for $Z> B$ results in high deviations for both EoS, while including the check yields moderate deviations of about 2.5\% for the RKPR, see \cref{fig:val_of}. The presented EoS can be extended to model a homogeneous mixture for an arbitrary number of components. To this end, the EoS parameter $a\alpha$ and $b$ have to be evaluated as a function of the mixture. Details can be found in the literature~\citep{Kim, Matheis, Matheis:2017dx, fathi2022large}. \subsection{Thermodynamic properties} \label{ss:caloric} Besides the correlation of density, pressure and temperature, also expressions for thermodynamic properties, such as the internal energy ${e}$, entropy $s$, enthalpy $h$ and specific heats ${c}_p$ and ${c}_v$, are needed for CFD simulations. The evaluation of these quantities includes several thermodynamic derivatives, which can be solved using the departure function formalism. For more detailed information we refer to \citet{Poling, elliott2012introductory}, and for details on the formulations for the RKPR to \citet{Kim,fathi2022large}. The presented formulations are obtained from \citet{Matheis} and recast to be valid for the more generic formulation. For the internal energy this can be written as \begin{equation} {e}({v},T)={e}_0 (T)+ \int_\infty^{{v}} \left(T \frac{\partial p}{\partial T}\bigg \vert_{{v}}-p\right) \mathrm{d} {v}, \label{eq:innEnergy} \end{equation} where the subscript $0$ refers to the ideal reference state. The solution of the integral reads \begin{equation} {e}-{e}_0= \left(a\alpha - T \frac{\partial a\alpha}{\partial T}\right) \,K, \label{eq:innEnergy2} \end{equation} where the term $K$ contains the following expression \begin{center} \begin{equation} K = \frac{1}{b(\delta_1 - \delta_2)} ln \left( \frac{{v} + \delta_1 b}{{v} + \delta_2 b}\right). \label{eq:K} \end{equation} \end{center}% Consequently, the enthalpy ${h}$ is calculated with \begin{equation} {h} - {h}_0 = {e} - {e}_0 + p{v} - {R}T\,. \label{eq:enthalpy} \end{equation}% resulting in the expression \begin{equation} \begin{split} {h} - {h}_0 = \left(a\alpha - T \frac{\partial a\alpha}{\partial T}\right)\,K + p{v}- {R}T. \end{split} \end{equation} The entropy ${s}$ is obtained with \begin{equation} {s}({v},T)={s}_0(T)+ \int_\infty^{{v}}\left(\frac{\partial p}{\partial T}\bigg \vert_{{v}}-\frac{{R}}{{v}}\right)\mathrm{d}{v} +{R}\ln(Z), \label{eq:entropy} \end{equation} resulting in \begin{equation} {s} - {s}_0 = - K T \frac{\partial a\alpha}{\partial T} + {R}\ln\left(1-\frac{b}{{v}}\right)\,. \end{equation} Finally, the Gibbs energy $g$ is calculated using \cref{eq:enthalpy} and \cref{eq:entropy} \begin{equation} {g} - {g}_0 = {h} - {h}_0 - T({s} - {s}_0) \end{equation} \begin{equation} {g} - {g}_0 = a\alpha K + p{v} - {R}T\left(1 + \ln\left(1-\frac{b}{{v}}\right)\right). \end{equation} As mentioned above, the Gibbs energy can be used to determine the most stable root out of three real roots. If the smallest root is larger than $B$ ($min(Z) > B$), then the smallest root represents the liquid state ($Z_l = min (Z)$) and the largest one the vapor state ($Z_v = max(Z)$). The relative difference between the Gibbs energy of the two solutions can be evaluated with: \begin{equation} \begin{split} d{g} = \frac{{g}_v - {g}_l}{{R}T} = \frac{A}{B (\delta_1 - \delta_2)} ln\left( \frac{(Z_l + \delta_1 B)(Z_v + \delta_2 B)}{(Z_l + \delta_2 B)(Z_v + \delta_1 B)} \right) \\ - (Z_l - Z_v) + ln\left( \frac{Z_l - B }{Z_v -B}\right). \end{split} \end{equation} If $d{g}<0$ (${g}_v < {g}_l$), the vapor state is stable. Contrary, if $d{g}>0$ (${g}_v > {g}_l$), the liquid state is stable. The heat capacity at constant volume ${c}_v$ is calculated with \begin{equation} ({c}_{v} - {c}_{v0}) = -T\frac{\partial^{2}a\alpha}{\partial T^2}K\,, \label{eq:cv} \end{equation} where ${c}_{v0}$ is evaluated using ${c}_{v0} = {c}_{p0} - R$. ${c}_{p0}$, the heat capacity at constant pressure at ideal reference state, is determined with the 7-coefficient or the 9-coefficient NASA polynomials. For the corresponding data for the 7-coefficient polynomials, we refer to \citet{Goos_Burcat}, and for the 9-coefficient ones to \citet{mcbride2002nasa}. Special attention should be paid to the fact that the polynomials are adapted for certain temperature ranges and that for a smooth calculation over several temperature ranges an appropriate implementation has to be done. Then, the heat capacity at constant pressure ${c}_p$ can be evaluated using \begin{equation} c_{p} = c_{v} - T \left(\frac{\partial p}{\partial T} \bigg \vert _{{v}} \right)^{2} \bigg / \frac{\partial p}{\partial {v}} \bigg \vert _{_{T}}, \label{eq:cp} \end{equation} with \begin{equation} \frac{\partial p}{\partial T} \bigg \vert _{{v}} = \frac{{R}}{{v} -b} + \frac{\partial a \alpha}{\partial T}\frac{1}{{D}} \label{eq:dpdT} \end{equation} and \begin{equation} \frac{\partial p}{\partial {v}} \bigg \vert _{p} = - \frac{{R}T}{({v} -b)^2} + \frac{a\alpha ( 2 {v} + (\delta_1 + \delta_2) b)}{{D}^2}, \label{eq:dpdT} \end{equation} with the denominator $D$ \begin{equation} D= ({v} + \delta_1 b)({v} + \delta_2 b) = {v}^2 + (\delta_1 + \delta_2) b {v} + \delta_1 \delta_2 b^2. \label{eq:dpdT} \end{equation} The speed of sound $c$ is calculated using \begin{equation} c = \sqrt{\frac{{c}_{p}}{{c}_{v}}\frac{\partial p}{\partial {v}} \bigg \vert _{_{T}} \frac{{v}^{2}}{M}}. \label{eq:speed_of_sound} \end{equation} \begin{figure*}[!tb] \centering \subfigure[]{\includegraphics[trim={50 0 250 0}, clip, width=0.49\linewidth]{Methane_.png}} \subfigure[]{\includegraphics[trim={50 0 250 0}, clip,width=0.49\linewidth]{nHexane_.png}}\\ \subfigure[]{\includegraphics[trim={50 0 250 0}, clip,width=0.49\linewidth]{nDodecane_.png}} \subfigure[]{\includegraphics[trim={50 0 250 0}, clip,width=0.49\linewidth]{Cyclopentane_.png}} \caption{ Comparison of the modeled density $\rho$, heat capacity $c_p$, viscosity $\mu$ and heat conductivity $\lambda$ using the thermodynamic model employing different EoS with reference values from CoolProp~\citep{bell2014pure}. (a) Methane ($Z_c = 0.2863$), (b) n-hexane ($Z_c = 0.2664$), (c) n-dodecane ($Z_c = 0.2497$), and (d) cyclopentane ($Z_c = 0.2813$). All data has been evaluated at a relative pressure of $p/p_c = 1.5$. } \label{fig:compare} \end{figure*} \subsection{Transport properties viscosity and heat conductivity with the Chung correlations} \label{ss:transport} For CFD simulations also suitable relations for the transport properties viscosity $\mu$ and heat conductivity $\lambda$ are necessary, where the correlations by \citet{Chung} are often employed~\citep{matheis2016volume,Matheis:2017dx, Matheis, fathi2022large,traxinger2018experimental, muller2016large}. Both quantities are composed of two terms referring to different pressure levels: \begin{equation} \mu = \mu_{k} + \mu_{p} \; \mathrm{and} \label{eq:mu} \end{equation} \begin{equation} \lambda = \lambda_{k} + \lambda_{p}. \label{eq:lambda} \end{equation} The first summand $\mu_{k}$, and $\lambda_{k}$ respectively, dominates at low pressures and is based on the Chapman-Enskog theory for diluted gases. The second term, $\mu_{p} $, and $\lambda_{p}$, dominates at higher pressures and is based on empirical correlations. The input for the model is composed of the fluid properties, the temperature, the density and the heat capacity ${c}_v$, where the latter only affects the evaluation of $\lambda$. For a detailed description see \citet{Chung,Poling}. \section{Assessment of the accuracy of the thermodynamic model} \label{s:Accuracy} \Cref{fig:compare} shows the density $\rho$, the heat capacity $c_p$, the viscosity $\mu$ and heat conductivity $\lambda$ for the n-alkanes methane ($Z_c = 0.2863$), n-hexane ($Z_c = 0.2664$), n-dodecane ($Z_c = 0.2497$) and the cycloalkane cyclopentane ($Z_c = 0.2813$). All data has been evaluated at a relative pressure of $p/p_c = 1.5$. Overall, the thermodynamic modeling is able to reproduce the non-linear behavior for all depicted quantities. As expected, the fluids with a critical compressibility close to 0.285 are well described by SRK, while for n-hexane with $Z_c = 0.2664$ PR (optimized for $Z_c = 0.263$) gives good results. For n-dodecane RKPR yields the best results. In all cases, the density is modeled very well with the RKPR EoS. It is either comparable to SRK and PR or much better, if the critical compressibility differs from the values for which the two were designed. The specific heat capacity $c_p$ is evaluated with \cref{eq:cp}. The peak at the pseudo-boiling is well captured by all EoS, but the maximum value is underestimated. The overall behavior of the transport properties ($\mu$, $\lambda$) is well described by the Chung correlations. In comparison with the density evolution for the different cubic EoS, one can see that the error in the modeling of the density $\rho$ corresponds to the error in the modeling of these two quantities, see PR in \cref{fig:compare}~(a) or \cref{fig:compare}~(d). This is due to the fact that the density $\rho$ is an input parameter for the Chung model, which directly affects the calculation of the high pressure empirical terms $\mu_p$ and $\lambda_p$. Hence, the error of the Chung correlations increases with an increasing modeling error of the density. Further, for $\lambda$, the error of ${c}_v$ affects the calculation of $\lambda_k$. As a consequence, the Chung correlations yield better results the more accurate $\rho$ and ${c}_v$ are modeled. \section{Supplementary material} \label{s:SupMat} We provide an open source Python tool called \texttt{realtpl} (real gas thermodynamic python library) for the presented thermodynamic model. Additionally, we also provide the implementation of the generalized form in OpenFOAM. \subsection{Python framework \texttt{realtpl}} Checking the accuracy of a thermodynamic model in advance is a central step before conducting CFD simulations. For different fluids as well as different pressure and temperature ranges, such an evaluation can be complicated and especially time consuming. To this end, we have written an open source Python tool to easily compare the results obtained with the here described thermodynamic model based on cubic EoS. The Python tool is called \texttt{realtpl} standing for our real gas thermodynamic python library. It is directly coupled to the open source library CoolProp~\citep{bell2014pure} obtaining experimental reference data, as well as fluid properties, such as for example molar mass and critical properties. In addition, a database was created for the NASA coefficients, which is also directly coupled to \texttt{realtpl}. Using \texttt{realtpl}, thermodynamic modeling based on the cubic EoS PR, SRK, RKPR can be compared and also contrasted with the reference data from CoolProp. The current implementation is designed to evaluate results over a temperature range (with defined number of temperature steps) for a given pressure level. The data is displayed graphically and can also be exported to a \textit{csv} file for further processing. Moreover, also evaluations over temperature and pressure ranges can be done, which allows for table generation~\citep{Doehring,jafari2022exploring,Koukouvinis}. To this end, the ranges and also the step width can be specified as configuration parameter. Apart from that, this open source Python tool can serve as an inspiration for implementing the present model into an internal flow solver. \Cref{tab:Performance} lists the process time for the evaluation of a representative configuration to provide estimates for the evaluation times using \texttt{realtpl}. The data has been evaluated for $10$, $10^2$, $10^3$ and $10^4$ temperature steps using a standard laptop (Intel i5, 7th generation). \textit{Start-up} refers to the reading of the config files and corresponding fluid properties. \textit{Ref. data} stands for extracting the reference data from CoolProp. In the current version, the five quantities density, heat capacity, speed of sound, viscosity and heat conductivity are extracted. This extraction from CoolProp can not be vectorized and, thus, it has to be looped over the temperature steps. Therefore, the evaluation time required scales roughly linearly with the number of temperature evaluations. This is the main time consumer when about $\num{3e3}$ temperature evaluations are exceeded. For the thermodynamic model, we have here listed the average of all three cubic EoS named \textit{Thermo. model}. To improve performance, the implementation of the thermodynamic models has been recast as vectors, avoiding time-consuming loops. For this reason, there is no linear scaling of the evaluation process. Also at $10^4$ temperature evaluations the time per thermodynamic model is still about 0.05 s. Here it has to be noted that five quantities are evaluated. Among different cubic EoS, we have seen that it varies depending on how often the check for B and the Gibbs evaluation has to be done. The next contributions are then output and postprocessing related. \textit{Figures} refers to the visualization of all five quantities including the write out of the figures. For less than $\num{3e3}$, this is the most time consuming part. The last part \textit{Data-output} is the write out of all evaluated data to a \textit{csv} file and does not consume significant time. For the entire evaluation at e.g. $10^3$ temperature steps, approx. 6 s are required with writing out of the figures and approx. 2.2 s without. \texttt{realtpl} is available as a PIP Python package and also on github github.com/ttrummler/realtpl. \begin{table} \center \caption{Overall process time for different numbers of temperature evaluations (t-ev), for details see text. } \label{tab:Performance} \begin{tabular}{lrrrr} \hline \multicolumn{1}{l}{} & \multicolumn{1}{c}{10 t-ev [s]} & \multicolumn{1}{c}{$10^2$ t-ev [s]} & \multicolumn{1}{c}{$10^3$ t-ev [s]}& \multicolumn{1}{c}{$10^4$ t-ev [s]} \\ \hline Start-up & 0.020 & 0.020 & 0.020 & 0.020 \\ Ref. data & 0.037& 0.1787& 1.489 &15.894\\ Thermo. model & 0.002 &0.002 &0.006 &0.046\\ Figures & 3.860 & 3.860 &3.860 &3.860\\ Data-output & 0.018& 0.034 & 0.146 &0.365\\ \hline Total & 3.941 & 4.099 & 5.533 & 20.276\\ \end{tabular} \end{table} \subsection{Generalized Formulation in OpenFOAM$^{\tiny{\textregistered}}$} OpenFOAM is a widely used open source software for simulations, where currently the most recent versions are the foundation version OpenFOAM-10~\citep{OF10} and the ESI version OpenFOAM2206~\citep{OF2206}. In both, the PR EoS is available as \textit{PengRobinsonGas}. In some in-house extensions of OpenFOAM~\citep{traxinger2020pressure,traxinger2019single,Traxinger}, the SRK has been additionally implemented. Despite the identical structure of SRK and PR, these two EoS are often hard coded and thus, lead to code duplicates. Following the generalized formulation proposed above, we propose a more general implementation of cubic EoS to avoid code duplication and to improve readability. We provide this extension for OpenFOAM under github.com/ttrummler/realFOAM. In order to keep the traditional OpenFOAM code structure and to not change the input files, we have created three separate folders for the different EoS. \Cref{fig:val_of} shows a validation of our OpenFOAM implementations comparing the density distribution with that obtained using the Python tool \texttt{realtpl}. As test configuration, we consider n-dodecane at a pressure of $p=8\,\si{MPa}$ and a temperature range of $T=500$ -- $1500 \,\si{K}$, matching the conditions of one operating point of the ECN Spray A~\citep{Matheis:2017dx, Koukouvinis}. For RKPR a very small deviation is visible, which is due to rounding errors. \begin{figure}[!htb] \centering \subfigure{\includegraphics[trim={0 0 0 0}, clip, width=0.99\linewidth]{Density.png}} \subfigure{\includegraphics[trim={0 0 0 0}, clip, width=0.99\linewidth]{deviation_rho.png}} \caption{ Comparison of the modeled density for n-dodecane at a pressure of $p=8\,\si{MPa}$. Reference values are taken from CoolProp~\citep{bell2014pure}. } \label{fig:val_of} \end{figure} \section{Conclusions} \label{s:Con} We have presented a thermodynamic model for real gas CFD simulations based on the generalized formulation of cubic EoS. Using this generalized formulation all of the well-known cubic EoS can be described with a particular value for $\delta_1$ ($\delta_2=(1-\delta_1)/(1+\delta_1)$). We have provided a detailed presentation of the resulting generalized cubic equation in $Z$ and practical hints for solving it. To evaluate the thermodynamic properties, we presented formulations of the derivatives. The transport properties are modeled with the Chung correlations. The thermodynamic model allows for a modularized implementation of several EoS. For the cubic EoS we have considered the well-known formulations SRK and PR. These two are specifically designed for an assumed critical compressibility factor and therefore their suitability is limited. Additionally, we also considered the RKPR, where the EoS parameters are functions of the critical compressibility factor. In this study, we have assessed the applicability of the three EoS for selected fluids and showed that the RKPR could be a good universally applicable choice for the EoS. In addition to this, we have demonstrated that overall the presented thermodynamic model can capture and reproduce the non-linear behavior of relevant thermodynamic quantities with an acceptable error. As supplementary material for the paper we provide an open source Python tool that can be used to evaluate and compare the results for a wide range of different fluids. Moreover, we also provide the implementation of the generalized EoS form in OpenFOAM. \section*{Acknowledgment} This project received funding by dtec.bw - Digitalization and Technology Research Center of the Bundeswehr - under the project MaST: Macro/Micro-simulation of Phase Decomposition in the Transcritical Regime. \section*{Data Availability} The data that support the findings of this study are available from the corresponding author upon reasonable request. Further, the data supporting the findings of this study can be generated with the openly available open source code provided within this paper. \section*{References}

0c7258538cffef8dd5cacb5f5c3ae32df9e8d452

\section{Introduction} \label{sec:Introduction} Highly directive radiation from a finite aperture is desired in many electromagnetics applications such as radar, communications, sensing, and imaging \cite{casares2006high,dash2018circularly,rana2017microstrip}. In this context, Fabry P\'{e}rot cavity (FPC) antennas have captured significant research interest because they offer high gain with a small aperture while retaining a simple feed structure for ease of fabrication and conformal deployment \cite{gardelli2006array,orr2013design,guerin2006met}. The FPC consists of a dielectric cavity encased by a metal ground plane on one side and a partially reflecting surface (PRS) on the other. Excitation from a primary radiator within the cavity undergoes multiple reflections between the PRS and ground plane before emanating from the antenna, resulting in an enhanced gain. The primary FPC radiators have been conventionally linearly polarised (LP) microstrip patches or slot antennas. The design of circularly polarized (CP) FPC antennas is relatively very challenging. In applications such as communications, biomedical imaging, and sensing, CP mode is desired to prevent the possible polarisation mismatch between the transmit and receive signals \cite{gu2019noncontact}. This ensures maximum power from the transmit to the receive antenna irrespective of their orientations. A common way to achieve CP FPC is to employ a CP primary radiator such as a feed point at a particular position on an asymmetric and rectangular patch \cite{qing2010compact,sharma1983analysis}. However, this technique suffers from low fabrication error tolerances and narrow bandwidths. Simple patches with dual-orthogonal feed structures have also been proposed \cite{kim2019compact} but a complicated feeding mechanism negates the purpose of realizing a simple single-feed high gain FPC antenna \cite{carver1981microstrip}. Alternatively, a linearly polarized primary radiator is used and the polarization is subsequently changed to circular by the PRS \cite{zhu2013linear,liu2015compact}. Here, a single-feed LP microstrip patch serves as the primary radiator. But the superstrate PRS is specifically designed to convert linear to circular polarization. In \cite{orr2013design}, the PRS was designed with a periodic array of unit cells where each unit cell consists of a rectangular loop with a diagonal. The resulting axial ratio of the structure was fairly low, but the antenna system was narrowband and of low gain. Our preliminary work in \cite{jain2020circularly} showed that further enhancements in the bandwidth and gain could be realized by incorporating peripheral roughness in the form of bricks along the metal edges of the unit cell. However, there are considerable degrees of freedom (DoFs) in the number and distribution of the bricks along the periphery. This design is not optimal because the evaluation of each candidate design, through either electromagnetic (EM) simulations or measurements, is slow and laborious. In this paper, we propose accelerating the design optimization process by replacing the time-consuming EM simulations with rapid neural network-based surrogate models. In general, the antenna design often involves optimization of several complicated and irregular geometry parameters in order to meet multiple objectives pertaining to the resonant frequency, gain, polarization, bandwidth, and size constraints. The procedure involves two time-consuming steps before the optimal design is realized: The first step is the synthesis of multiple candidate designs for evaluation. Traditionally, antenna parameters were optimized through trial and error. Later works have employed evolutionary algorithms, such as genetic algorithms and particle swarm optimization, wherein geometry parameters were optimized through an iterative synthesis of antenna designs constrained by a fitness function on the antenna characteristics \cite{jin2005parallel,silva2013adaptive}. In the second step, the antenna characteristics for each design are simulated using computationally-expensive full-wave EM solvers involving finite difference time domain techniques, finite element methods, or method of moments. Several recent studies have shown to reduce the computational workload during the evaluation process through machine learning (ML) techniques, including artificial neural networks \cite{sedaghat2022compressed}, support vector regression \cite{prado2019wideband}, and deep learning (DL) networks \cite{hodge2019rf,hodge2019joint,liaqat2021hybrid,xiao2018deep}. These methods map the non-linear relationship between the geometric parameters and antenna characteristics using data from EM solvers. Once trained, the design process is significantly accelerated with the surrogate model replacing the EM solver for rapidly generating the antenna behaviors for any given set of geometry parameters. In this context, deep generative adversarial networks (GANs) have emerged as a preferred DL technique to solve a wide variety of EM problems \cite{alnujaim2021synthesis,liu2018generative,hodge2019multi,liu2021generative, vijayamohanan2020antenna,ye2020inhomogeneous,zhang2022image}. A GAN works as a zero-sum game between two deep networks: generator and discriminator. Its objective is to \emph{implicitly} learn the probabilistic distribution of a set of training samples and, subsequently, create samples of the distribution during the prediction stage \cite{goodfellow2014generative}. The generator produces samples of a distribution from the training data while the discriminator assesses the samples and decides if they are real or fake (produced by the generator). The primary advantage of using the GAN framework with two competing neural networks is that the GAN is semi-supervised and requires smaller, less diverse training data set \cite{creswell2018generative}. In this work, we propose to accelerate the CP FPC parameter optimization through a GAN-aided design procedure. The generator serves as a surrogate model for producing samples of antenna characteristics using training antenna patterns obtained from an EM solver. The inputs to the generator are the unit cell geometric parameters. Once the surrogate model is trained, we use it to simultaneously evaluate the antenna characteristics of several hundred candidate designs, eventually facilitating the choice of the optimal design with bandwidth, axial ratio, and gain as 269 MHz, 0.4 dB, and 7.5 dBi respectively. We demonstrate the feasibility of our proposed method through validation via full-wave simulations, fabrication of optimized antenna, and hardware measurements. \section{FPC Structure} \label{sec:sysmod} In FPC, the primary source of excitation is introduced within the cavity. The height of the cavity is carefully chosen such that the multiple reflections within the cavity are in phase with each other when they emanate from the antenna, thereby enhancing the gain of the primary radiator. The polarization of the resulting radiation is determined by either the polarization of the primary radiating source or by the unit cell in the PRS. Consider a basic FPC structure (Fig.~\ref{fig:Antenna}a), where a single feed patch antenna - the primary radiator resonating at $2.4$ GHz - is mounted on a Rogers 4350B substrate and impedance matched to 50 $\Omega$ through a three-stage quarter-wave transformer (Fig.~\ref{fig:Antenna}b(i)). The other side of the substrate is a partial ground metal plane of copper (Fig.~\ref{fig:Antenna}b(ii)). The patch radiates into a polystyrene-based dielectric cavity which is enclosed, on the other end, by a Rogers 4350B superstrate. The inner side of the superstrate is printed with a periodic array of $4 \times 4$ unit cells in copper to form a PRS (Fig.~\ref{fig:Antenna}b(iii)). Each unit cell of the PRS is a rectangular loop with a diagonal. The dimensions of this basic antenna structure (Fig.~\ref{fig:Antenna}) remain fixed across all candidate designs. Then, peripheral roughness is introduced to metallic edges along each dimension of the unit cell through 36 $0.5 \times 0.5$ mm$^2$ metal bricks (Fig.~\ref{fig:Antenna}b(iv)). The positions of the bricks along the peripheries of the rectangular loop become the DoFs for reducing the axial ratio while enhancing the gain and bandwidth. The position of each brick is indicated in two-dimensional (2-D) Cartesian coordinates, with the origin assumed to be at the left lower corner of the unit cell. The antenna system with each unique unit cell design is simulated in CST Microwave Studio to obtain the electrical characteristics from 2 to 3 GHz. Since this is a 3-D antenna structure, there are approximately 6.4 million mesh cells for each design with a simulation duration of each around 75 minutes, which is excessive. \section{GAN Architecture for CP FPC Design} \label{sec:GAN Architecture for Antenna Design} We propose to reduce the resource-expensive EM simulations by replacing them with a GAN at the prediction stage. This network is a semi-supervised learning architecture comprising a generator ($G_{\theta}$) and a discriminator ($D_{\phi}$) (Fig.~\ref{fig:TrainingGAN}a). During the training phase, the 2-D position coordinates of 36 bricks in the unit cell in the PRS are reshaped to a single column vector, $\xb$, of size $[72 \times 1]$. \begin{figure}[t] \centering \includegraphics[width=3.5in,height=2.5in]{Antenna_structure.pdf} \vspace{-5mm} \caption{(a) Three-dimensional (3-D) view of CP FPC antenna. (b) Top-view of (i) patch with a three-stage quarter-wave transformer, (ii) partial reflecting ground plane, (iii) partially reflecting superstrate with $4\times 4$ array of uniform unit cells with the diagonal, and (iv) unit cell with peripheral roughness. (c) As in (b), but for our fabricated FPC antenna with optimized design.} \label{fig:Antenna} \end{figure} \begin{figure}[t] \centering \includegraphics[scale=0.52]{Block_diagram.pdf} \vspace{-5mm} \caption{GAN architecture for (a) training and (b) prediction stages.} \label{fig:TrainingGAN} \end{figure} This is concatenated with a latent noise vector, $\mathbf{z}$, of $[100 \times 1]$ size, to prevent overfitting, and provided as input to $G_{\theta}$. The output of the generator ($G_{\theta}(\mathbf{z}+\xb)=\y$) are antenna performance metrics (as a function of frequency): axial ratio, return loss, and gain, each of size $[101 \times 1]$ concatenated to form a single column vector. These are input to $D_{\phi}$ along with the real antenna characteristics, $\yb$, obtained from CST Microwave Studio for the same set of input antenna designs $\xb$. Both GAN networks compete adversarially to optimize the weights $\theta$ and $\phi$ of, respectively, $G_{\theta}$ and $D_{\theta}$ based on the value function \par\noindent\small \begin{multline} \label{eqn:objectivefn} V(G_{\theta},D_{\phi})= log(D_{\phi}(\yb))+\log(1-D_{\phi}(G_{\theta}(\mathbf{z}+ \xb))). \end{multline}\normalsize The optimal $G$ and $D$ are obtained by solving the minimax problem, i.e., $\{D^{\ast}_{\phi}, G^{\ast}_{\theta}\} = \underset{G_{\theta}}{\text{arg min}}\;\underset{D_{\phi}}{\text{arg max}}\; V(D_{\phi},G_{\theta})$. The training process involved iterative simultaneous stochastic gradient descent based on Adam optimization on batches of 16 samples of $\xb, \yb$ and $\hat{\yb}$. In each iteration, ${\theta}$ were updated while ${\phi}$ and biases were kept constant and vice versa. The weights were normalized during each update to prevent overfitting and weighted by a regularizer $\lambda=0.01$. \begin{figure}[t] \centering \includegraphics[scale=0.15]{Histogram_merged.png} \caption{Histograms depicting variations of antenna characteristics for (a-c) 300 samples that were used to generate GAN and CST results; both training (blue) and test (red) sets are shown (d-f) 500 candidate antenna designs from which an optimized design was selected. } \label{fig:Histogram} \end{figure} The total number of iterations was set to 10000. The learning rate of the stochastic gradient descent operation for both networks was $5\times10^{-4}$ and batch normalization was applied with a momentum of 0.8. The network $G$ had 128, 256, and 512 nodes in the first, second, and third layers, respectively, with \emph{Leaky ReLU} activation functions. The network $D$ had similar 512 and 256 nodes in two respective hidden layers. The output layer had one node with $sigmoid$ function. \section{Numerical Experiments} \label{sec:numexp} We implemented the DL network with Keras 2.7 and Python on an Intel Core i7-10510U processor running at 1.80 GHz and NVIDIA GeForce MX250. \subsection{Validation of surrogate model via simulations} Consider the architecture in Fig.~\ref{fig:TrainingGAN}b where the input antenna designs (distinct from those used for training) correspond to the brick positions in the unit cell form $\mathbf{\hat{x}}$ and the output are the corresponding antenna characteristics $\y = G(\mathbf{z}+ \mathbf{\hat{x}})$. The (fake) output $\y$ of $G$ was then compared with $\yb$ generated from the EM solver to monitor the training process. We used $90\%$ and $10\%$ of a total of 300 antenna designs-pattern pairs for training and validation, respectively, with 10-fold cross-validation. We compared the performance metrics as a function of frequency from 2 to 3 GHz. \begin{table}[t] \centering \caption{\small NMSE of antenna characteristics } \label{Comparison_table} \begin{tabular}{p{1cm}|p{2cm}|p{2cm}|p{2.0cm}} \hline \noalign{\vskip 1pt} Methods & Axial Ratio (dB) & Return Loss (dB) & Gain (dBi)\\[1pt] \hline \hline \noalign{\vskip 1pt} ANN& 5.4&7.5& 5.67\\ CNN& 1.35&3.65& 5.3\\ GAN&0.05&2.28&2.9\\ \hline \end{tabular} \end{table} Training and test data in Fig.~\ref{fig:Histogram}a-c show that these metrics vary with the spatial distribution of the peripheral roughness features and the data were not overfitted. Finally, comparisons (Table.~\ref{Comparison_table}) with a fully connected artificial neural network (ANN) and a convolutional neural network (CNN) trained with the same data demonstrate that GAN has the lowest normalized mean square error (NMSE) between the real and fake antenna characteristics. \subsection{Fabrication and measurement results} \label{subsec:GAN Rsults} The simple rectangular patch is essentially an LP narrowband (20 MHz bandwidth) and low gain (3.4 dBi) antenna. The PRS structure enhances the gain and bandwidth of the structure to 188.5 MHz and 9.4 dBi gain (Fig.~\ref{fig:ExperimentalResult}) with a resonant frequency shifted from 2.4 GHz. \begin{figure}[t] \centering \includegraphics[scale=0.44]{Comparison_experimental.pdf} \caption{\small Antenna performance metrics for (top panel) simple patch and FPC with PRS with unit cell comprising a rectangular loop with a diagonal (without any roughness), and (bottom panel) surrogate GAN model, CST, and fabricated device measurements.} \label{fig:ExperimentalResult} \end{figure} The unit cell design of a \emph{rectangular loop with a diagonal} without roughness transforms the patch’s LP signal to an elliptically polarised wave with an axial ratio of 7.4 dB. The peripheral roughness features reduce the axial ratio without accounting for DoF in the metal brick distribution. The trained generator is able to analyze the antenna characteristics of 500 such designs in 20 seconds (as against 625 hours with CST). From these 500 antenna characteristics (shown in Fig.~\ref{fig:Histogram}d-f), we chose the design with a wide bandwidth of 269 MHz and axial ratio of 0.4 dB. We validated this with CST and fabricated the corresponding antenna (Fig.~\ref{fig:Antenna}c). The measurements from the actual antenna show very good agreement with GAN and CST results (Fig.~\ref{fig:ExperimentalResult}). \begin{table}[t] \centering \caption{\small Comparison with competing antenna structure } \ \label{table: Comparison of antenna characteristics} \begin{tabular}{p{3.3cm}|p{1.25cm}|p{1.35cm}|p{1.0cm}} \hline \noalign{\vskip 1pt} Antenna & Bandwidth (MHz) & Axial Ratio (dB)& Gain (dBi) \\[1pt] \hline \hline \noalign{\vskip 1pt} Simple patch & 20 & 40 & 3.4\\ FPC with PRS & 88.5 & 7.6 & 9.4\\ GAN-based FPC, rough PRS & 269 & 0.4 & 7.5\\ \hline \end{tabular} \end{table} Quantitative comparisons (Table.\ref{table: Comparison of antenna characteristics}) with FPC and the simple patch further demonstrate the performance enhancements with a GAN-aided design. \section{Summary} \label{sec:Conclusion} Peripheral roughness along the edges of the unit cell of the PRS of a CP-FPC offers several DoFs for improving antenna performance. We proposed a versatile GAN-based FPC design strategy where we train a GAN to serve as a surrogate model using input-output pairs of antenna designs and their corresponding patterns obtained from the solver. The proposed design strategy enables rapid evaluation of a large number of candidate designs. Our GAN-optimized unit cell yielded axial ratio, gain, and bandwidth of 0.4 dB, 7.5 dBi, and 269 MHz, respectively, thereby considerably improving the performance of the original FPC structure. Fabrication and experimental validations supported the GAN results. The design files and GAN codes are available at \url{https://essrg.iiitd.edu.in/?page_id=4355}. \bibliographystyle{ieeetr} \section{Introduction} \label{sec:Introduction} Highly directive radiation from a finite aperture is desired in many electromagnetics applications such as radar, communications, sensing, and imaging \cite{casares2006high,dash2018circularly,rana2017microstrip}. In this context, Fabry P\'{e}rot cavity (FPC) antennas have captured significant research interest because they offer high gain with a small aperture while retaining a simple feed structure for ease of fabrication and conformal deployment \cite{gardelli2006array,orr2013design,guerin2006met}. The FPC consists of a dielectric cavity encased by a metal ground plane on one side and a partially reflecting surface (PRS) on the other. Excitation from a primary radiator within the cavity undergoes multiple reflections between the PRS and ground plane before emanating from the antenna, resulting in an enhanced gain. The primary FPC radiators have been conventionally linearly polarised (LP) microstrip patches or slot antennas. The design of circularly polarized (CP) FPC antennas is relatively very challenging. In applications such as communications, biomedical imaging, and sensing, CP mode is desired to prevent the possible polarisation mismatch between the transmit and receive signals \cite{gu2019noncontact}. This ensures maximum power from the transmit to the receive antenna irrespective of their orientations. A common way to achieve CP FPC is to employ a CP primary radiator such as a feed point at a particular position on an asymmetric and rectangular patch \cite{qing2010compact,sharma1983analysis}. However, this technique suffers from low fabrication error tolerances and narrow bandwidths. Simple patches with dual-orthogonal feed structures have also been proposed \cite{kim2019compact} but a complicated feeding mechanism negates the purpose of realizing a simple single-feed high gain FPC antenna \cite{carver1981microstrip}. Alternatively, a linearly polarized primary radiator is used and the polarization is subsequently changed to circular by the PRS \cite{zhu2013linear,liu2015compact}. Here, a single-feed LP microstrip patch serves as the primary radiator. But the superstrate PRS is specifically designed to convert linear to circular polarization. In \cite{orr2013design}, the PRS was designed with a periodic array of unit cells where each unit cell consists of a rectangular loop with a diagonal. The resulting axial ratio of the structure was fairly low, but the antenna system was narrowband and of low gain. Our preliminary work in \cite{jain2020circularly} showed that further enhancements in the bandwidth and gain could be realized by incorporating peripheral roughness in the form of bricks along the metal edges of the unit cell. However, there are considerable degrees of freedom (DoFs) in the number and distribution of the bricks along the periphery. This design is not optimal because the evaluation of each candidate design, through either electromagnetic (EM) simulations or measurements, is slow and laborious. In this paper, we propose accelerating the design optimization process by replacing the time-consuming EM simulations with rapid neural network-based surrogate models. In general, the antenna design often involves optimization of several complicated and irregular geometry parameters in order to meet multiple objectives pertaining to the resonant frequency, gain, polarization, bandwidth, and size constraints. The procedure involves two time-consuming steps before the optimal design is realized: The first step is the synthesis of multiple candidate designs for evaluation. Traditionally, antenna parameters were optimized through trial and error. Later works have employed evolutionary algorithms, such as genetic algorithms and particle swarm optimization, wherein geometry parameters were optimized through an iterative synthesis of antenna designs constrained by a fitness function on the antenna characteristics \cite{jin2005parallel,silva2013adaptive}. In the second step, the antenna characteristics for each design are simulated using computationally-expensive full-wave EM solvers involving finite difference time domain techniques, finite element methods, or method of moments. Several recent studies have shown to reduce the computational workload during the evaluation process through machine learning (ML) techniques, including artificial neural networks \cite{sedaghat2022compressed}, support vector regression \cite{prado2019wideband}, and deep learning (DL) networks \cite{hodge2019rf,hodge2019joint,liaqat2021hybrid,xiao2018deep}. These methods map the non-linear relationship between the geometric parameters and antenna characteristics using data from EM solvers. Once trained, the design process is significantly accelerated with the surrogate model replacing the EM solver for rapidly generating the antenna behaviors for any given set of geometry parameters. In this context, deep generative adversarial networks (GANs) have emerged as a preferred DL technique to solve a wide variety of EM problems \cite{alnujaim2021synthesis,liu2018generative,hodge2019multi,liu2021generative, vijayamohanan2020antenna,ye2020inhomogeneous,zhang2022image}. A GAN works as a zero-sum game between two deep networks: generator and discriminator. Its objective is to \emph{implicitly} learn the probabilistic distribution of a set of training samples and, subsequently, create samples of the distribution during the prediction stage \cite{goodfellow2014generative}. The generator produces samples of a distribution from the training data while the discriminator assesses the samples and decides if they are real or fake (produced by the generator). The primary advantage of using the GAN framework with two competing neural networks is that the GAN is semi-supervised and requires smaller, less diverse training data set \cite{creswell2018generative}. In this work, we propose to accelerate the CP FPC parameter optimization through a GAN-aided design procedure. The generator serves as a surrogate model for producing samples of antenna characteristics using training antenna patterns obtained from an EM solver. The inputs to the generator are the unit cell geometric parameters. Once the surrogate model is trained, we use it to simultaneously evaluate the antenna characteristics of several hundred candidate designs, eventually facilitating the choice of the optimal design with bandwidth, axial ratio, and gain as 269 MHz, 0.4 dB, and 7.5 dBi respectively. We demonstrate the feasibility of our proposed method through validation via full-wave simulations, fabrication of optimized antenna, and hardware measurements. \section{FPC Structure} \label{sec:sysmod} In FPC, the primary source of excitation is introduced within the cavity. The height of the cavity is carefully chosen such that the multiple reflections within the cavity are in phase with each other when they emanate from the antenna, thereby enhancing the gain of the primary radiator. The polarization of the resulting radiation is determined by either the polarization of the primary radiating source or by the unit cell in the PRS. Consider a basic FPC structure (Fig.~\ref{fig:Antenna}a), where a single feed patch antenna - the primary radiator resonating at $2.4$ GHz - is mounted on a Rogers 4350B substrate and impedance matched to 50 $\Omega$ through a three-stage quarter-wave transformer (Fig.~\ref{fig:Antenna}b(i)). The other side of the substrate is a partial ground metal plane of copper (Fig.~\ref{fig:Antenna}b(ii)). The patch radiates into a polystyrene-based dielectric cavity which is enclosed, on the other end, by a Rogers 4350B superstrate. The inner side of the superstrate is printed with a periodic array of $4 \times 4$ unit cells in copper to form a PRS (Fig.~\ref{fig:Antenna}b(iii)). Each unit cell of the PRS is a rectangular loop with a diagonal. The dimensions of this basic antenna structure (Fig.~\ref{fig:Antenna}) remain fixed across all candidate designs. Then, peripheral roughness is introduced to metallic edges along each dimension of the unit cell through 36 $0.5 \times 0.5$ mm$^2$ metal bricks (Fig.~\ref{fig:Antenna}b(iv)). The positions of the bricks along the peripheries of the rectangular loop become the DoFs for reducing the axial ratio while enhancing the gain and bandwidth. The position of each brick is indicated in two-dimensional (2-D) Cartesian coordinates, with the origin assumed to be at the left lower corner of the unit cell. The antenna system with each unique unit cell design is simulated in CST Microwave Studio to obtain the electrical characteristics from 2 to 3 GHz. Since this is a 3-D antenna structure, there are approximately 6.4 million mesh cells for each design with a simulation duration of each around 75 minutes, which is excessive. \section{GAN Architecture for CP FPC Design} \label{sec:GAN Architecture for Antenna Design} We propose to reduce the resource-expensive EM simulations by replacing them with a GAN at the prediction stage. This network is a semi-supervised learning architecture comprising a generator ($G_{\theta}$) and a discriminator ($D_{\phi}$) (Fig.~\ref{fig:TrainingGAN}a). During the training phase, the 2-D position coordinates of 36 bricks in the unit cell in the PRS are reshaped to a single column vector, $\xb$, of size $[72 \times 1]$. \begin{figure}[t] \centering \includegraphics[width=3.5in,height=2.5in]{Antenna_structure.pdf} \vspace{-5mm} \caption{(a) Three-dimensional (3-D) view of CP FPC antenna. (b) Top-view of (i) patch with a three-stage quarter-wave transformer, (ii) partial reflecting ground plane, (iii) partially reflecting superstrate with $4\times 4$ array of uniform unit cells with the diagonal, and (iv) unit cell with peripheral roughness. (c) As in (b), but for our fabricated FPC antenna with optimized design.} \label{fig:Antenna} \end{figure} \begin{figure}[t] \centering \includegraphics[scale=0.52]{Block_diagram.pdf} \vspace{-5mm} \caption{GAN architecture for (a) training and (b) prediction stages.} \label{fig:TrainingGAN} \end{figure} This is concatenated with a latent noise vector, $\mathbf{z}$, of $[100 \times 1]$ size, to prevent overfitting, and provided as input to $G_{\theta}$. The output of the generator ($G_{\theta}(\mathbf{z}+\xb)=\y$) are antenna performance metrics (as a function of frequency): axial ratio, return loss, and gain, each of size $[101 \times 1]$ concatenated to form a single column vector. These are input to $D_{\phi}$ along with the real antenna characteristics, $\yb$, obtained from CST Microwave Studio for the same set of input antenna designs $\xb$. Both GAN networks compete adversarially to optimize the weights $\theta$ and $\phi$ of, respectively, $G_{\theta}$ and $D_{\theta}$ based on the value function \par\noindent\small \begin{multline} \label{eqn:objectivefn} V(G_{\theta},D_{\phi})= log(D_{\phi}(\yb))+\log(1-D_{\phi}(G_{\theta}(\mathbf{z}+ \xb))). \end{multline}\normalsize The optimal $G$ and $D$ are obtained by solving the minimax problem, i.e., $\{D^{\ast}_{\phi}, G^{\ast}_{\theta}\} = \underset{G_{\theta}}{\text{arg min}}\;\underset{D_{\phi}}{\text{arg max}}\; V(D_{\phi},G_{\theta})$. The training process involved iterative simultaneous stochastic gradient descent based on Adam optimization on batches of 16 samples of $\xb, \yb$ and $\hat{\yb}$. In each iteration, ${\theta}$ were updated while ${\phi}$ and biases were kept constant and vice versa. The weights were normalized during each update to prevent overfitting and weighted by a regularizer $\lambda=0.01$. \begin{figure}[t] \centering \includegraphics[scale=0.15]{Histogram_merged.png} \caption{Histograms depicting variations of antenna characteristics for (a-c) 300 samples that were used to generate GAN and CST results; both training (blue) and test (red) sets are shown (d-f) 500 candidate antenna designs from which an optimized design was selected. } \label{fig:Histogram} \end{figure} The total number of iterations was set to 10000. The learning rate of the stochastic gradient descent operation for both networks was $5\times10^{-4}$ and batch normalization was applied with a momentum of 0.8. The network $G$ had 128, 256, and 512 nodes in the first, second, and third layers, respectively, with \emph{Leaky ReLU} activation functions. The network $D$ had similar 512 and 256 nodes in two respective hidden layers. The output layer had one node with $sigmoid$ function. \section{Numerical Experiments} \label{sec:numexp} We implemented the DL network with Keras 2.7 and Python on an Intel Core i7-10510U processor running at 1.80 GHz and NVIDIA GeForce MX250. \subsection{Validation of surrogate model via simulations} Consider the architecture in Fig.~\ref{fig:TrainingGAN}b where the input antenna designs (distinct from those used for training) correspond to the brick positions in the unit cell form $\mathbf{\hat{x}}$ and the output are the corresponding antenna characteristics $\y = G(\mathbf{z}+ \mathbf{\hat{x}})$. The (fake) output $\y$ of $G$ was then compared with $\yb$ generated from the EM solver to monitor the training process. We used $90\%$ and $10\%$ of a total of 300 antenna designs-pattern pairs for training and validation, respectively, with 10-fold cross-validation. We compared the performance metrics as a function of frequency from 2 to 3 GHz. \begin{table}[t] \centering \caption{\small NMSE of antenna characteristics } \label{Comparison_table} \begin{tabular}{p{1cm}|p{2cm}|p{2cm}|p{2.0cm}} \hline \noalign{\vskip 1pt} Methods & Axial Ratio (dB) & Return Loss (dB) & Gain (dBi)\\[1pt] \hline \hline \noalign{\vskip 1pt} ANN& 5.4&7.5& 5.67\\ CNN& 1.35&3.65& 5.3\\ GAN&0.05&2.28&2.9\\ \hline \end{tabular} \end{table} Training and test data in Fig.~\ref{fig:Histogram}a-c show that these metrics vary with the spatial distribution of the peripheral roughness features and the data were not overfitted. Finally, comparisons (Table.~\ref{Comparison_table}) with a fully connected artificial neural network (ANN) and a convolutional neural network (CNN) trained with the same data demonstrate that GAN has the lowest normalized mean square error (NMSE) between the real and fake antenna characteristics. \subsection{Fabrication and measurement results} \label{subsec:GAN Rsults} The simple rectangular patch is essentially an LP narrowband (20 MHz bandwidth) and low gain (3.4 dBi) antenna. The PRS structure enhances the gain and bandwidth of the structure to 188.5 MHz and 9.4 dBi gain (Fig.~\ref{fig:ExperimentalResult}) with a resonant frequency shifted from 2.4 GHz. \begin{figure}[t] \centering \includegraphics[scale=0.44]{Comparison_experimental.pdf} \caption{\small Antenna performance metrics for (top panel) simple patch and FPC with PRS with unit cell comprising a rectangular loop with a diagonal (without any roughness), and (bottom panel) surrogate GAN model, CST, and fabricated device measurements.} \label{fig:ExperimentalResult} \end{figure} The unit cell design of a \emph{rectangular loop with a diagonal} without roughness transforms the patch’s LP signal to an elliptically polarised wave with an axial ratio of 7.4 dB. The peripheral roughness features reduce the axial ratio without accounting for DoF in the metal brick distribution. The trained generator is able to analyze the antenna characteristics of 500 such designs in 20 seconds (as against 625 hours with CST). From these 500 antenna characteristics (shown in Fig.~\ref{fig:Histogram}d-f), we chose the design with a wide bandwidth of 269 MHz and axial ratio of 0.4 dB. We validated this with CST and fabricated the corresponding antenna (Fig.~\ref{fig:Antenna}c). The measurements from the actual antenna show very good agreement with GAN and CST results (Fig.~\ref{fig:ExperimentalResult}). \begin{table}[t] \centering \caption{\small Comparison with competing antenna structure } \ \label{table: Comparison of antenna characteristics} \begin{tabular}{p{3.3cm}|p{1.25cm}|p{1.35cm}|p{1.0cm}} \hline \noalign{\vskip 1pt} Antenna & Bandwidth (MHz) & Axial Ratio (dB)& Gain (dBi) \\[1pt] \hline \hline \noalign{\vskip 1pt} Simple patch & 20 & 40 & 3.4\\ FPC with PRS & 88.5 & 7.6 & 9.4\\ GAN-based FPC, rough PRS & 269 & 0.4 & 7.5\\ \hline \end{tabular} \end{table} Quantitative comparisons (Table.\ref{table: Comparison of antenna characteristics}) with FPC and the simple patch further demonstrate the performance enhancements with a GAN-aided design. \section{Summary} \label{sec:Conclusion} Peripheral roughness along the edges of the unit cell of the PRS of a CP-FPC offers several DoFs for improving antenna performance. We proposed a versatile GAN-based FPC design strategy where we train a GAN to serve as a surrogate model using input-output pairs of antenna designs and their corresponding patterns obtained from the solver. The proposed design strategy enables rapid evaluation of a large number of candidate designs. Our GAN-optimized unit cell yielded axial ratio, gain, and bandwidth of 0.4 dB, 7.5 dBi, and 269 MHz, respectively, thereby considerably improving the performance of the original FPC structure. Fabrication and experimental validations supported the GAN results. The design files and GAN codes are available at \url{https://essrg.iiitd.edu.in/?page_id=4355}. \bibliographystyle{ieeetr}

1a26ebf271ca55087509ca2a790e239d5b9663f7

\section{Introduction} \label{1} Two features of atomic systems that emerge from quantum theory are the presence of highly inhomogeneous shell structure and the spontaneous breaking of spherical symmetry. These two attributes play an important role in determining atomic size and facilitating chemical bonding. An intuitive understanding of the mechanism behind them would be helpful in considerations of molecules and bulk materials. On one hand, quantum superposition arguments can be made in support of the point of view that all isolated atoms, even open shell atoms, are spherically symmetric \cite{Cohen1965, gil1972}. On the other hand, the concept of an ``isolated'' atom is purely theoretical, and operationally, it is nowadays accepted that open-shell atoms used as building blocks for molecular calculations are not generally treated as have spherically symmetric electron densities \cite{Fertig2000}. In particular, the recent density functional theory (DFT) study of open-shell atoms by Chowdhury and Perdew \cite{Chowdhury_Perdew2021} has examined the implications of spherical symmetry-breaking for molecular bonding and atomization energies. An alternative to the Kohn-Sham (KS) DFT used by Chowdhury and Perdew \cite{Chowdhury_Perdew2021} is polymer self-consistent field theory (SCFT). Based on a quantum-classical isomorphism introduced by Feynman \cite{Feynman1953AtomicTO, Chandler_Wolynes1981}, classical statistical mechanics is used to represent quantum particles as ring ``polymers'', that is, extended non-point like objects, embedded in an extra thermal dimension \cite{thompson2019.article}. SCFT has been shown to be equivalent to KS-DFT, assuming the Pauli-exclusion principle holds \cite{thompson2019.article}, and incorporates all quantum effects \cite{Thompson2022} in terms of classical quantities which enables explanations of quantum phenomena more in tune with classical intuition. The use of a Pauli potential and lack of explicit orbitals means that SCFT is related to orbital-free (OF) DFT, and as is typical in OF-DFT, the exclusion principle can be incorporated using approximations \cite{thompson2020.article}. SCFT predicts molecular bonding \cite{sillaste_thompson.article}, atomic shell structure \cite{LeMaitre_Thompson_2022}, includes temperature dependence \cite{thompson2019.article}, and can be related to dynamical quantum mechanics \cite{Thompson2022}. SCFT also has foundational implications due to its classical ensemble formulation \cite{Thompson2022}. Like DFT, SCFT is based on an energy functional, and since the quantities are real and classical, it can be decomposed into thermodynamic components in order to explain the origins of predicted structures \cite{Matsen_Bates1997}. The purpose of this paper is to show that SCFT spontaneously predicts shell structure and spherical symmetry-breaking in isolated atoms, and to give the thermodynamic explanations for non-spherical ground-state structures. We partition the SCFT free energy functional for the electron density into components: the translational entropy contribution, polymer configurational entropy contribution (equivalent to the quantum kinetic energy), and the internal energy which includes electron-electron, electron-nucleus, and Pauli-exclusion interactions. We find that for the atoms we studied from hydrogen to neon, all thermodynamic contributions cause the free energy to increase for spherical symmetry-breaking except the electron-nucleus potential. In other words, the overall energy is reduced when the electrons can move closer to the nucleus by breaking spherical symmetry and this more than compensates for all other factors which would tend to increase the energy. Neutral atoms in their ground-state were studied in a previous work \cite{LeMaitre_Thompson_2022} using a spherical-averaging approximation, which amounts to representing only the radial part of the electron density, as well as the adoption of partial atomic shell information. These approximations were useful to see how well the model could replicate atomic trends for a large range of elements, most importantly if it could produce the proper shell structure. We will generalize that work by restoring the full angular distribution of the electron density. We organize the paper as follows. In section \ref{2}, the conceptual basis of the theory is outlined and a new derivation of the model equations is given. The fields in the model are introduced, the free energy components are given, and the section ends with a discussion about the numerical methods used. The atomic binding energies for the elements hydrogen to neon and the angular electron density contour plots for the elements boron to neon are presented in section \ref{3}, along with tables of data for the free energy components and various density constraints that the model satisfies with a proof of the equivalence with KS-DFT. The Discussion in section \ref{4} demonstrates that the external potential is responsible for the spherical symmetry breaking. The paper concludes in section \ref{5} and some future research directions are discussed. \section{Theory} \label{2} The pioneering work of Feynman in developing the path integral formulation of quantum theory led to the deduction of the isomorphism between quantum mechanics and the statistical mechanics of ring-like molecules, using a Wick rotation of time $t=-i\hbar\beta$ to a parameter $\beta$ \cite{Feynman1953AtomicTO}. The Wick rotation essentially allows one to transform a dynamics problem into a statics problem in one higher dimension, where temporal variables are typically exchanged for spatial variables \cite{lancaster2014}. Part of the intuition behind why the Wick rotation takes quantum mechanics to the statistical mechanics of ring-like molecules stems from the nature of the trace operator in the partition function, which is a sum of the Boltzmann factors over configurations starting and ending at the same position. This cyclic aspect is crucial to the interpretation, especially after having transferred to the path integral picture, because the path integral allows one to see that each of the trajectories followed through the imaginary time configuration parameter space correspond to the different possible arrangements of quantum particles comprising a system of quantum particles \cite{Feynman1953AtomicTO,Chandler_Wolynes1981}. The correspondence with the mathematics of polyatomic fluids noted by Chandler \cite{Chandler_Wolynes1981} lends itself to label each of the quantum particles in the system as atoms comprising a molecule. The probabilistic uncertainty in the position and momenta of the quantum system are then manifested as the collection of system arrangements associated to a given configuration. The connection to polymer SCFT is then made clear through the insights of Matsen \cite{matsen.incollection}. The periodicity of the imaginary time parameter $\beta$ also happens to be a part of one of the conditions to be a Kubo-Martin-Schwinger (KMS) state, which is a general notion of being in a thermal state \cite{haag1967}. Although the mathematics is essentially identical between quantum mechanical and polymeric systems, the interpretation of the fundamental constituents and the mechanisms that govern their behaviour has changed dramatically. In the quantum mechanical case, the time evolution of the system could trace out many different paths, forming a distribution of them that expresses the uncertainty in which path will be followed. In the statistical mechanical case, the system could explore many different energy configurations, forming a distribution of them where thermal fluctuations represent the uncertainty in which configuration will be chosen. A set of postulates has been proposed to bridge between finite-temperature quantum DFT and ring polymer SCFT \cite{Thompson2022}. In the static case considered here, there are two postulates needed, in addition to those from statistical mechanics, to describe quantum many-body systems of Fermions: 1) ``pairs'' (i.e. 1-2) of quantum particles are classically modelled as stochastic chains in a 4-dimensional thermal space and 2) the chains have excluded-volume with respect to 4-dimensional thermal space. The first postulate can be seen to be equivalent to the Heisenberg uncertainty principle \cite{thompson2019.article,thompson2020.article}, and the introduction of the term ``pairs'' is to account for Fermion spin, while the second postulate is a (conjectured) statement of the Pauli-exclusion principle. Similar to ring polymer SCFT, the base entities in this theory are fundamentally distinguishable from each other, so the notion of particle ``pairs'' with higher-dimensional excluded-volume is the mechanism that emulates particle statistics from quantum statistical mechanics. To recover the standard electron density, field, and electron number that appear in DFT, we simply sum up all of the pairs, which we denote with a Greek index. The two postulates in combination with the usual approximations used in the study of many-body systems (i.e. Born-Oppenheimer approximation, point-particle nuclei) are then enough to derive the expression for the partition function of the system. Working in the canonical ensemble for simplicity, with $N$-body potential $U$, the $N$-body partition function $Q_N$ can be expressed as a path integral in configuration space as \begin{align} Q_N(\beta) = \prod^N_i\int\mathrm{d}\bm{r}_i\int\mathcal{D}[\bm{r}_i] e^{-\int^{\beta}_0 \mathrm{d}\tau \left[\sum_j \frac{m}{2\hbar^2}\left|\frac{\mathrm{d}\bm{r}_j(\tau)}{\mathrm{d}\tau}\right|^2+U[\hat{n}](\bm{r}_1(\tau), \ldots, \bm{r}_N(\tau))\right]} \label{expec_bolt} \end{align} where the parameter $\beta$ is the Lagrange multiplier that ensures the expectation value of the free energy remains constant, which will end up being the reciprocal thermal energy $\beta=1/k_BT$ (where $k_B$ is Boltzmann's constant and $T$ is the temperature) due to the KMS condition \cite{haag1967}; $\bm{r}_i(s)$ is the parametrized curve representing the $i$th quantum particle as depicted in figure \ref{poly_contour}, with the parameter $s$ running from $0$ (a high classical temperature) to $\beta$ (a lower temperature); and the $N$-body potential $U$ is expressed as a functional of an electron density operator $\hat{n}(\bm{r})$, which is defined to be \begin{align} \hat{n}(\bm{r}) = \sum_\mu \hat{n}_\mu(\bm{r}) = \sum_\mu\sum^{N_\mu}_i \delta(\bm{r}-\bm{r}_i)\,. \label{elec_denop} \end{align} $U[\hat{n}]$ can be re-expressed in terms of the fields $\mathcal{N}_\mu(\bm{r})$ using the functional Dirac delta $\delta[\mathcal{N}_\mu-\hat{n}_\mu]$, which can in-turn be expressed in terms of its functional Fourier transform representation with respect to conjugate fields $W_\mu(\bm{r})$, allowing eqn. \ref{expec_bolt} to be expressed as \begin{align} Q_N(\beta) &= \prod_\mu\int \int \mathcal{D}[\mathcal{N}_\mu] \mathcal{D}[W_\mu]e^{-\beta U[\mathcal{N}] + \beta\int \mathrm{d}\bm{r}' W_\mu(\bm{r}')\mathcal{N}_\mu(\bm{r}')}\prod^{N_\mu}_i\int\mathrm{d}\bm{r}_i \nonumber \\ &\qquad \qquad \times\int\mathcal{D}[\bm{r}_i] e^{-\int^\beta_0 \mathrm{d}\tau \left[ \frac{m}{2\hbar^2}\left|\frac{\mathrm{d}\bm{r}_i(\tau)}{\mathrm{d}\tau}\right|^2 + \int \mathrm{d}\bm{r}' W_\mu(\bm{r}')\hat{n}_\mu(\bm{r}')\right]}\,. \label{part1} \end{align} \begin{figure}[h] \centering \includegraphics[height=10cm, width=7cm]{./Figures/Pair_Diagram.png} \caption{A visual schematic of a quantum particle trajectory in $s$ parameter space, where the horizontal axes are the spatial coordinates and the vertical axes are the $s$ coordinates. The quantum particle starts out in a position $\bm{r}_0$ at $s=0$ and follows the $s$-space trajectory to the same starting position but at $s=\beta$, representing a ``ring''.} \label{poly_contour} \end{figure} The $\hat{n}$ operator implicitly carries the $N$ coordinate dependencies with it, so inserting eqn. \ref{elec_denop} into eqn. \ref{part1} allows one to see, after some manipulations, that the argument of the exponential in the configuration path integral is now completely separable into $N$ one-body terms, producing a product of $N$ separable path integrals. The configuration integrals can then be evaluated one at a time with the result being $N$ identical terms. The $N$-body partition function $Q_N(\beta)$ can then finally be expressed as \begin{align} Q_N(\beta) = \prod_\mu\int \int \mathcal{D}[\mathcal{N}_\mu] \mathcal{D}[W_\mu]\,e^{-\beta F[\mathcal{N}_\mu, W_\mu]} \label{part_fin} \end{align} where \begin{align} F[\mathcal{N}_\mu, W_\mu] = -\frac{1}{\beta} N_\mu\ln (Q_\mu[W](\beta))+U[\mathcal{N}_\mu]- \int \mathrm{d}\bm{r}' W_\mu(\bm{r}')\mathcal{N}_\mu(\bm{r}') \label{free_func} \end{align} and $Q_\mu[W](\beta)$ is a single-pair partition function that we have defined according to the expression \begin{align} Q_\mu[W](\beta) = \int \mathrm{d}\bm{r}\, q_\mu(\bm{r}, \bm{r}, \beta) \label{1part} \end{align} where $q_\mu(\bm{r}, \bm{r}, \beta)$ represents the propagation of a single pair from initial position $\bm{r}$ at $s=0$ to final position $\bm{r}$ at $s=\beta$. The single-pair propagator $q_\mu(\bm{r}, \bm{r}', s)$ can be expressed as a path integral: \begin{align} q_\mu(\bm{r}, \bm{r}', s) = \mathcal{A}\int \mathcal{D}[\bm{r}] e^{-\int^s_0 \mathrm{d}\tau \left[\frac{m}{2\hbar^2}\left|\frac{\mathrm{d}\bm{r}(\tau)}{\mathrm{d}\tau}\right|^2 + W_\mu(\bm{r}(\tau))\right]} \label{prop1} \end{align} where $\mathcal{A}$ is a formally infinite normalization constant coming from the kinetic degrees of freedom, whose value we shall not be concerned with since it will not appear in any of the quantities of interest. As a propagator, it can be shown that $q_\mu(\bm{r}, \bm{r}', s)$ equivalently satisfies the modified diffusion equation \begin{align} \frac{\partial q_\mu(\bm{r}, \bm{r}', s)}{\partial s} = -H^{\text{eff}}_\mu q_\mu(\bm{r}, \bm{r}', s) = \frac{\hbar^2}{2m}\nabla^2q_\mu(\bm{r}, \bm{r}', s) - W_\mu(\bm{r})q_\mu(\bm{r}, \bm{r}', s) \label{diffuse} \end{align} with initial condition $q_\mu(\bm{r}, \bm{r}', 0) = \delta(\bm{r}-\bm{r}')$, which makes it possible to evaluate $Q_\mu[W](\beta)$ and thus $F[\mathcal{N}, W]$ exactly. It is worth pointing out that the Hamiltonian $H_{\text{eff}}$ above is the same Hamiltonian appearing in the Kohn-Sham equation from KS-DFT, a fact we use to prove the equivalence of the two theories in appendix \ref{AppendixC}. Although it is possible to analytically evaluate the functional eqn. \ref{free_func}, we are not so fortunate with eqn. \ref{part_fin}, whose path integrals are too unwieldy to perform exact calculations with to get the free energy \cite{Matsen_Schick1994.article}. However, eqn. \ref{free_func} plays the same role that the action does in the real-time quantum mechanical path integral, so a solution can be sought which extremizes $F[\mathcal{N}, W]$ by setting its first variation to zero and then approximating the integrand with the extremum of $F$. The free energy can then be calculated from $F[n, w]$, where $n$ and $w$ are the mean-fields for which the functional $F$ has a saddle point \cite{matsen.incollection,Matsen_Schick1994.article,Kim_Yang_Lee2012}. We can further justify the preservation of exactness in the model from varying eqn. \ref{free_func}, since any neglected higher-order contributions can be packaged into the unknown functional $U$, whose approximations occupy a large portion of current DFT research \cite{Jones2015,witt_del_rio_dieterich_carter2018,Karasiev_Trickey2012}. The total kinetic energy functional $K[n, w]$ can also be calculated from the expectation value of the kinetic term in the many-body Hamiltonian, after some manipulation, to be \begin{align} K[n, w] &= -\sum^N_j\bigg\langle \frac{\hbar^2}{2m}\nabla^2_j\bigg\rangle = -\frac{\hbar^2}{2m}\sum_\mu\sum^N_j\frac{1}{Q_\mu[w](\beta)} \int \mathrm{d}\bm{r}_j \nabla^2_{\bm{r}'_j}q_\mu(\bm{r}_j, \bm{r}'_j, \beta)\bigg|_{\bm{r}'_j=\bm{r}_j} \,. \label{kin_func} \end{align} Proceeding from the variational principle outlined above, using the path integral form of the single-pair propagator, the mean-field density $n_\mu(\bm{r}, \beta)$ and field $w_\mu(\bm{r}, \beta)$ corresponding to each pair are found to be \begin{align} w_\mu(\bm{r}, \beta) = \frac{\delta U[n]}{\delta n_\mu(\bm{r}, \beta)} \ \ \text{and} \ \ n_\mu(\bm{r}, \beta) = \frac{N_\mu}{Q_\mu[w](\beta)}q_\mu(\bm{r}, \bm{r}, \beta)\,.\label{field_dens} \end{align} The potential $U[n]$ is the only remaining quantity yet to be specified, which will give us the expressions for the fields $w_\mu(\bm{r}, \beta)$ experienced by each pair in the system, and finally the expression for the free energy $F[n, w]$. In our model, the electron pairs experience four fields in the vicinity of the atomic nucleus: the Coulomb field between the nucleus and the electron pairs $w_\mu^{e-n}(\bm{r}, \beta)$, the Coulomb field between the electrons $w_\mu^{e-e}(\bm{r}, \beta)$, and the exchange field between electron pairs $w_\mu^{x}(\bm{r}, \beta)$ representing two separate fields. The first of these two fields is the electron self-interaction field $w_\mu^{sic}(\bm{r}, \beta)$, which corrects for the interaction of the electron with its own field that is not accounted for in the electron-electron Coulomb field $w_\mu^{e-e}(\bm{r}, \beta)$; the self-interaction correction introduced in prior work \cite{LeMaitre_Thompson_2022} is employed in this work as well. The second of these two fields, as is commonplace in all OF-DFT approaches, is the Pauli-exclusion field $w_\mu^P(\bm{r}, \beta)$, which accounts for the repulsion felt by electron pairs with the same configuration attempting to occupy the same location at the same (imaginary) time, as stipulated by the Pauli-exclusion principle. As will be discussed later, the Pauli-exclusion field used in this work accounts for some exchange effects but does not account for correlations. The electron-nucleus potential $U_{e-n}[n]$ takes the form \begin{align} U_{e-n}[n] = -\int \int \mathrm{d}\bm{r}\, \mathrm{d}\bm{r}'\, n(\bm{r}, \beta)\frac{\rho_{\text{nuc}}(\bm{r}')}{\left|\bm{r}-\bm{r}' \right|}\,, \end{align} where $\rho_{\text{nuc}}(\bm{r})$ is the nuclear density, which we take to be $\rho_{\text{nuc}}(\bm{r}) = N\delta(\bm{r})$. The electron-nucleus field for each pair $w^{e-n}_\mu(\bm{r}, \beta)$ is then found to be \begin{align} w^{e-n}_\mu(\bm{r}, \beta) = - \int \mathrm{d}\bm{r}'\, \frac{\rho_{\text{nuc}}(\bm{r}')}{\left|\bm{r}-\bm{r}' \right|} = -\frac{N}{|\bm{r}|}\,. \label{wen} \end{align} The potential due to electron-electron Coulomb-type interactions $U_{e-e}[n]$ is similarly given by \begin{align} U_{e-e}[n] = \frac{1}{2}\int \int \mathrm{d}\bm{r}\, \mathrm{d}\bm{r}'\, n(\bm{r}, \beta)\frac{n(\bm{r}', \beta)}{\left|\bm{r}-\bm{r}' \right|}\,, \end{align} and the electron-electron field for each pair $w^{e-e}_\mu(\bm{r}, \beta)$ is then found to be \begin{align} w^{e-e}_\mu(\bm{r}, \beta) = \int \mathrm{d}\bm{r}' \frac{n(\bm{r}', \beta)}{|\bm{r}'-\bm{r}|}\,. \label{wee} \end{align} Both eqn. \ref{wen} and eqn. \ref{wee} indicate that each pair experiences the exact same electron-electron and electron-nucleus field. The total field experienced by each pair will however not be the same, thanks to the exchange effects introduced by the other two fields. Following previous work \cite{thompson2020.article}, the closest classical analogue of the Pauli-exclusion principle is the notion of excluded-volume, which in polymer SCFT, is often implemented as a Dirac delta energy penalty for overlapping polymer segments. If we are to be truly faithful to the exclusion principle however, the energy penalty should be for overlapping polymer segments from \textit{different} polymer contours representing pairs of quantum particles to account for spin. Since the position along the polymer contour is parametrized by a parameter $s$, the energy penalty due to overlapping polymer contours occurs only for contours at the same value of $s$. Recall from the quantum-classical correspondence that the parameter $s$ can be interpreted as an imaginary time, so the Pauli-exclusion repulsion is akin to a particle pair feeling the excluded-volume repulsion when at the same place and (imaginary) time as another pair \cite{thompson2020.article}. This idea is difficult to implement in practice however, so we approximate it by projecting out the degrees of freedom from the $s$ parameter space, which effectively amounts to imposing the excluded-volume energy penalties for every value of $s$. The downside to this approximation is that it ignores the inter-contour correlations and will clearly overestimate the excluded-volume felt between the pairs \cite{thompson2020.article}. The approximate Pauli-exclusion potential is then given by \begin{align} U_P[n] = \frac{1}{2g_0}\sum_{\substack{\mu,\nu \\ \mu \neq \nu}} \int \int \mathrm{d}\bm{r}\mathrm{d}\bm{r}' n_\mu(\bm{r}, \beta)\delta(\bm{r}-\bm{r}')n_\nu(\bm{r}', \beta) = \frac{1}{2g_0}\sum_{\substack{\mu,\nu \\ \mu \neq \nu}} \int \mathrm{d}\bm{r} n_\mu(\bm{r}, \beta)n_\nu(\bm{r}, \beta) \label{Paulipot_approx} \end{align} where $g_{0}$ is a constant with the same units as a density of states \cite{thompson2020.article}. In principle, since the excluded-volume interaction is independent of the system under study, $g_0$ should be a universal constant whose value can be determined by comparing the Pauli potential for a very simplistic system (e.g. a uniform gas with only excluded-volume interactions) to experimental results. However, because the Pauli potential is being approximated in this work, $g_0$ is taken to be arbitrary and we choose its value once for all calculations; the value chosen and how it was chosen will be discussed in the Results section \ref{3}. The approximate form of the Pauli-exclusion field for each pair $w^{P}_\mu(\bm{r}, \beta)$ is then calculated as \begin{align} w^{P}_\mu(\bm{r}, \beta) = \frac{1}{g_0} \sum_{\substack{\gamma \\ \gamma \neq \mu}}n_\gamma(\bm{r}, \beta)\,. \label{wpauli} \end{align} In previous work \cite{LeMaitre_Thompson_2022}, some constraints on the exact form of the Pauli-exclusion field were given which allowed for the accuracy of the approximate expression eqn. \ref{wpauli} to be assessed. The relevant constraints were as follows: \begin{align} &w^P(\bm{r}, \beta) \geq 0\ \ , \ \ \Lim{|\bm{r}|\rightarrow \infty} w^P(\bm{r}, \beta) = 0\ \ , \ \ \int \mathrm{d}\bm{r}\, w^P(\bm{r}, \beta)n(\bm{r}, \beta) < \infty\ \ , \ \ w^P(\bm{r}, \beta) = 0 \ \ \text{for} \ \ N=2\ \ , \nonumber \\[1.5ex] &w^P[\lambda^3n](\lambda\bm{r}, \beta) = \lambda^2w^P[n](\lambda\bm{r}, \beta) \label{pauli_cons} \end{align} where in the last criterion $\lambda$ is a scale factor and the functional dependence of the field on the density has been explicitly reinstated \cite{Levy_Ou-Yang1988.article}. We found that all but the last of the constraints in eqn. \ref{pauli_cons} were satisfied by eqn. \ref{wpauli}, with $w^P(\bm{r}, \beta)$ overestimating the excluded-volume interactions by precisely the amount required to fulfill the last constraint in eqn. \ref{pauli_cons}; a point also discussed in a previous work \cite{thompson2020.article}. The electron self-interaction field $w_\mu^{\text{sic}}(\bm{r}, \beta)$ used in this work was first introduced in reference \cite{LeMaitre_Thompson_2022}, and is essentially a Fermi-Amaldi self-interaction correction applied to each particle pair; see reference \cite{LeMaitre_Thompson_2022} for further discussion. It has the form \begin{align} w_\mu^{\text{sic}}(\bm{r}, \beta) = -\frac{1}{N_\mu}\int \mathrm{d}\bm{r}'\frac{n_\mu(\bm{r}', \beta)}{|\bm{r}-\bm{r}'|}\label{self_int} \end{align} where the corresponding potential $U_{\text{sic}}[n]$ is simply \begin{align} U_{\text{sic}}[n] = -\sum_\mu\frac{1}{2N_\mu}\int \int \mathrm{d}\bm{r}\mathrm{d}\bm{r}'\frac{n_\mu(\bm{r}', \beta)n_\mu(\bm{r}, \beta)}{|\bm{r}-\bm{r}'|}\,. \label{self_int_pot} \end{align} In this form, eqn. \ref{self_int} directly preserves the desirable qualities of the original Fermi-Amaldi electron self-interaction correction for hydrogen and helium \cite{ayers_morrison_parr.article}. Furthermore, because eqn. \ref{self_int} acts on electron pairs and $N_\mu = \{1, 2\}$, then eqn. \ref{self_int} effectively accounts for the self-interaction of every electron in the atomic system. Those familiar with the Hartree-Fock model will immediately recognize that our model with eqn. \ref{self_int} is \textit{identical} to the situation in the Hartree-Fock model, which models exchange effects exactly. However, because we are approximating the Pauli-exclusion field by projecting out the degrees of freedom from the $s$ parameter space, our model in its current implementation will not reproduce the precise binding energies predicted by Hartree-Fock theory. This is because our Pauli-exclusion field overestimates the excluded-volume felt by the polymer contours in the $s$-parameter space, hence electron pairs feel too much repulsion between each other and the electron shells will be too distant from their neighbours, raising the free energy substantially in some cases. The Hartree-Fock model on the other hand, is a wavefunction-based model, so the Pauli-exclusion effect is automatically encoded into the wavefunction due to the spin-statistics theorem \cite{szabo2012modern}. Implementing the exact expression for the Pauli potential would then allow our model to coincide exactly with Hartree-Fock theory. As was mentioned previously, electron pairs are taken to be ring polymers embedded in a 3+1-dimensional thermal space under the influence of a potential $U$. The ring polymers are confined to explore the thermal space according to this potential, which they can do using two different mechanisms: translation and configuration. The translational degrees of freedom refer to the three dimensional motion of the polymer as a whole, while the configurational degrees of freedom refer to how the polymer confirmations change while holding one point of the polymer fixed in space. Each of these mechanisms has an entropy associated to them, which we will denote as $S_t$ and $S_c$, respectively. The behaviour of the polymer can then be explained by looking at the competition between the degrees of freedom encoded in the entropies and those restricted by the potential $U$. Therefore, by rephrasing the free energy per pair $F[n_\mu, w_\mu]$ in terms of these entropies, we can exactly describe the process in the polymer picture by which the electrons surrounding the atomic nucleus would break spherical-symmetry; and with the notion of pairs, we can pinpoint exactly which pairs affect this change. Following previous work \onlinecite{thompson2019.article} and reference \onlinecite{Matsen_Bates1997}, the free energy per pair $F[n_\mu, w_\mu]$ can be re-expressed as \begin{align} F[n_\mu, w_\mu] &= -\frac{1}{\beta} \int \mathrm{d}\bm{r}' n_\mu(\bm{r}', \beta)\ln{\left(\frac{n_\mu(\bm{r}', \beta)}{N_\mu}\right)}+U[n_\mu] \nonumber \\[1.5ex] &\quad+\frac{1}{\beta}\int \mathrm{d}\bm{r}' n_\mu(\bm{r}', \beta)\left[\ln{\left(q_\mu(\bm{r}', \bm{r}', \beta)\right)} + \beta w_\mu(\bm{r}', \beta)\right] \label{free_func_ent} \end{align} where we can identify the last term with the free energy contribution coming from the configurational entropy $S_c[n_\mu, w_\mu]$ \footnote{This configurational entropy expression corrects a typo in reference \onlinecite{thompson2019.article} which is missing a factor of $-\beta$.} and the first term with the translational entropy $S_t[n_\mu]$, which we write as \begin{align} S_c[n_\mu, w_\mu] &= -\int \mathrm{d}\bm{r}' n_\mu(\bm{r}', \beta)\left[\ln{\left(q_\mu(\bm{r}', \bm{r}', \beta)\right)} + \beta w_\mu(\bm{r}', \beta)\right]\\[1.5ex] S_t[n_\mu] &= \int \mathrm{d}\bm{r}' n_\mu(\bm{r}', \beta)\ln{\left(\frac{n_\mu(\bm{r}', \beta)}{N_\mu}\right)}\,. \label{ent_funcs} \end{align} We shall examine the intuitive polymeric interpretation of spherical symmetry-breaking in the Results \ref{3} and Discussion \ref{4} sections. To solve the set of self-consistent equations \ref{diffuse},\ref{1part}, and \ref{field_dens}, the biggest obstacle is the solution to the modified diffusion equation eqn. \ref{diffuse}, which needs to be solved a number of times corresponding to the total number of pairs $N_p$ in the system --- per self-consistent iteration --- for every value of both spatial positions $\bm{r}$ and $\bm{r}'$. This double spatial dependence of $q_\mu(\bm{r}, \bm{r}', s)$ means that traditional real-space methods are impractical for computational efficiency \cite{thompson2019.article,Matsen_Schick1994.article,matsen.incollection}. Instead, what is usually done in the polymer SCFT community, is to use the spectral method: all spatially-dependent functions are decomposed in terms of a set of basis functions, for which the position dependence of each function can be integrated out and the resulting equations become matrix equations for the unknown expansion coefficients. The problem encountered earlier with real-space methods is then made to vanish and is replaced with solving a matrix equation $N_p$ times per self-consistent iteration \cite{Matsen_Schick1994.article,thompson2019.article}. In this work we use the spectral method with non-orthogonal Gaussian basis sets outlined in reference \onlinecite{LeMaitre_Thompson_2022} to solve the modified diffusion equation. The method, along with the numerical procedure, has been discussed in reference \onlinecite{LeMaitre_Thompson_2022}, however, only the case of spherical Gaussian basis functions was treated. To solve the angular problem, we will need the full angular Gaussian basis functions, which are given by the (normalized) expression \begin{align} f_{i} (\bm{r}) = \left(2(2c_{pl})^{l+\frac{3}{2}}\left[\Gamma\left(l+\frac{3}{2}\right)\right]^{-1}\right)^{\frac{1}{2}}Z^m_{l} (\theta,\phi) r^l e^{- c_{pl} r^2 }\,, \label{gauss} \end{align} where the index $i$ represents the tuple of indices $(p, l, m)$ and $Z^l_{m} (\theta,\phi)$ are the real spherical harmonics defined in appendix \ref{AppendixA}. From here, the basis-specific quantities outlined in appendix \ref{AppendixA} of the overlap matrix (eqn. \ref{over_gauss}), Laplace matrix (eqn. \ref{new_lap_gauss}), and Gamma tensor (eqn. \ref{gamma_gauss}) must be recalculated, but the general non-orthogonal spectral equations listed in appendix \ref{AppendixB} remain unchanged. Although the differences from the spherical case only appear in the basis-specific quantities, it should be pointed out that the addition of angularity increases the dimensions of the basis-specific quantities substantially, severely increasing the computational runtime for even a modest sized basis set, and also introduces a complicated quantity into the expression for the Gamma tensor known as the real Gaunt coefficients; all details can be found in appendix \ref{AppendixA} and references \onlinecite{LeMaitre_Thompson_2022,thompson2019.article}. The Gaussian exponents $c_{pl}$ are chosen according to an “even-tempered”-type scheme outlined in reference \onlinecite{LeMaitre_Thompson_2022}. In this work $N_b=425$ Gaussian basis functions were used for the angular results since this number provided excellent resolution and converged far enough to the infinite basis set limit, although more basis functions are used in this case compared with reference \onlinecite{LeMaitre_Thompson_2022} because we assigned numbers for each $l$ value of the real spherical harmonics, which in turn have $2l+1$ types of basis functions (for the number of $m$ values associated to each $l$). Convergence here is judged according to the spectral convergence criterion used in reference \onlinecite{LeMaitre_Thompson_2022}. The number 425 comes from assigning 150 basis functions to $l=0$, 50$\times$3 basis functions to $l=1$, and 25$\times$5 basis functions to $l=2$. Increasing $l$ values are assigned smaller numbers of basis functions because the corresponding basis functions become much more diffuse and start to represent smaller portions of the electron density profile. The angular results are also only presented for the first 10 elements, so 150 basis functions for the $l=0$ portion is more than enough to achieve good resolution; the approximations used in this work also limit our accuracy a lot more than the basis set truncation error does. The set of minimum exponents we chose was $c_{1,1,1}=(10^{-15}, 10^{-10}, 10^{-6})$ and the set of maximum exponents was $c_{150,50,25}=(10^{11}, 10^{5}, 10^{3})$, where each entry corresponds to $l$ values in increasing order, respectively. Anything higher than 425 would only add a small amount of resolution and would require a relatively large increase in computation time. For 425 basis functions, every element was able to satisfy a tolerance of at least $10^{-6}$, with some going as far as $10^{-8}$. We decided to keep the same value for $g_0$ that was used in reference \onlinecite{LeMaitre_Thompson_2022} of $g_0=0.1$ for the arbitrary constant $g_0$ associated to the Pauli-exclusion field eqn. \ref{wpauli}, since this will make comparison with previous results easier. The full angular Gaussian basis set used in this work does not introduce or change any previously encountered numerical considerations addressed in reference \onlinecite{LeMaitre_Thompson_2022}. \section{Results} \label{3} The atomic binding energies corresponding to $g_0=0.1$ for the elements hydrogen to neon, strictly enforcing the maximum occupancy of 2 electrons per pair, are shown in table \ref{tab2} for the spectral expansion of the density with angular basis functions, and table \ref{tab3} for the restriction to spherically-symmetric basis functions; atomic units are used unless otherwise specified. The binding energies predicted by our model are contrasted with those predicted by Hartree-Fock theory, since in our neglect of correlations and use of an exact self-interaction correction, Hartree-Fock theory should be considered as ``exact''. One difference with the results from reference \onlinecite{LeMaitre_Thompson_2022} can be seen by looking at the percentage deviation with Hartree-Fock for the two tables: The model with angular dependence is much closer to Hartree-Fock for the first 6 elements, but rapidly worsens due to an overabundance of Pauli-exclusion repulsion felt between electron pairs. In the angular case considered here, the atomic shell configurations are not prebuilt into the code, so the overabundance of the Pauli-exclusion force causes the pairs after carbon to spread too far apart. However, the observed shell structure does arise solely from the information provided by the electron pair configurations and still somewhat resembles what we expect (figures \ref{ang_boron}-\ref{ang_neon})\footnote{The density contour plots of the first four elements are not shown as they are all spherical and have identical pair to shell structure. Reference \onlinecite{LeMaitre_Thompson_2022} can be consulted for plots of hydrogen to beryllium using the model from this work.}. Moreover, we do see spontaneous spherical symmetry-breaking, which is first predicted to occur at carbon as opposed to boron. The shapes of the pair densities after boron do not match naive expectations; this is addressed later in this section. The magnitude of the symmetry-breaking effect can also be modified through the value of $g_0$ (i.e. a smaller value produces more noticeable deviations). In particular, if $g_0$ is given pair dependence, then a simple arithmetic sequence of increasing values chosen from the neighbourhood around $0.1$ such that the smallest value is assigned to the first pair, is sufficient for boron to break spherical symmetry. Although nature predicts spherical symmetry-breaking to first occur at boron, this does not mean the polymer excluded-volume picture is invalid: symmetry-breaking has a subtle effect on the binding energies, as can be seen by comparing tables \ref{tab3} and \ref{tab2}, so the approximation used for the Pauli-exclusion field may be too coarse to allow such a prediction, in which case the exact field would need to be implemented to sufficiently test this. The fact that spontaneous spherical symmetry-breaking does occur, and only one element off from where it is supposed to be, is a very encouraging result. \begin{longtable}{ p{.09\textwidth} p{.12\textwidth} p{.15\textwidth} p{.14\textwidth} } \hline Element & SCFT & Hartree-Fock & \% Deviation \\ \hline H & 0.4999999 & 0.5000000000 & 0.000002\\ He & 2.861679 & 2.861679996 & 0.000000014\\ Li & 7.46842 & 7.432726931 & 0.47792\\ Be & 14.70219 & 14.57302317 & 0.87856\\ B & 24.66954 & 24.52906073 & 0.56944\\ C & 37.655254 & 37.68861896 & 0.088607\\ N & 53.65814 & 54.40093421 & 1.38431\\ O & 72.8257 & 74.80939847 & 2.7239\\ F & 95.2256 & 99.40934939 & 4.3935\\ Ne & 120.9975 & 128.5470981 & 6.2394\\ \hline \caption{Atomic binding energies for hydrogen to neon using full angular basis sets with pair occupancy restricted to a maximum of 2. The number of decimal places for the binding energies predicted by our model correspond to the numerical accuracy of our calculations, where the numerical uncertainty is expressed in the last digit. The Hartree-Fock binding energies are taken from \cite{koga_thakker.article}.} \label{tab2} \end{longtable} \begin{longtable}{ p{.09\textwidth} p{.12\textwidth} p{.15\textwidth} p{.14\textwidth} } \hline Element & SCFT & Hartree-Fock & \% Deviation \\ \hline H & 0.4999999 & 0.5000000000 & 0.000002\\ He & 2.861679 & 2.861679996 & 0.000000014\\ Li & 7.46842 & 7.432726931 & 0.47792\\ Be & 14.70219 & 14.57302317 & 0.87856\\ B & 24.66954 & 24.52906073 & 0.56944\\ C & 37.567740 & 37.68861896 & 0.321764\\ N & 53.40706 & 54.40093421 & 1.86094\\ O & 72.3335 & 74.80939847 & 3.4229\\ F & 94.3264 & 99.40934939 & 5.3887\\ Ne & 119.5084 & 128.5470981 & 7.5633\\ \hline \caption{Atomic binding energies for hydrogen to neon using spherically-symmetric basis sets with pair occupancy restricted to a maximum of 2. The number of decimal places for the binding energies predicted by our model correspond to the numerical accuracy of our calculations, where the numerical uncertainty is expressed in the last digit. The Hartree-Fock binding energies are taken from \cite{koga_thakker.article}.} \label{tab3} \end{longtable} There is no difference in the binding energies between the spherical and angular results for the first 4 elements since these elements are known to have spherical ground-state distributions and minimal Pauli-exclusion repulsion. The lack of a binding energy difference between the spherical and non-angular cases of boron is attributed to the approximate Pauli potential used in this work, which predicts symmetry-breaking to occur at carbon instead of boron. Carbon is the first element where any difference in the binding energy between the spherical and non-angular cases can be seen: the percent difference of the SCFT model with the prediction from Hartree-Fock theory in the angular case is approximately an order of magnitude smaller than in the spherical case. The agreement for carbon is due in part to the cancellation of certain errors as opposed to a genuine agreement with Hartree-Fock theory, but the other trends in tables \ref{tab3} and \ref{tab2} suggest that the angular case does yield an improvement in the agreement with Hartree-Fock theory. The rest of the elements from tables \ref{tab3} and \ref{tab2} display very minor changes in the binding energy, which agrees nicely with the results of Chowdhury and Perdew \cite{Chowdhury_Perdew2021}, who report that the effect of symmetry-breaking has a small impact on the binding energy. These results suggest that the spherical-averaging approximation used in reference \onlinecite{LeMaitre_Thompson_2022} performs quite well in most scenarios, and that it is physically reasonable to use it for isolated atoms provided the aim is not to investigate delicate features that arise from angularity. \begin{figure} \centering \includegraphics[height=8cm, width=16cm]{./Figures/Angular_Boron_Density_Contour_Plot_for_Constant_Theta.png} \caption{Boron angular electron pair density contour plots for fixed values of $\theta$. The axes correspond to $x$ and $y$ Cartesian coordinates and the colour bar indicates the magnitude of the density.} \label{ang_boron} \end{figure} \begin{figure} \centering \includegraphics[height=8cm, width=16cm]{./Figures/Angular_Carbon_Density_Contour_Plot_for_Constant_Theta.png} \caption{Carbon angular electron pair density contour plots for fixed values of $\theta$. The axes correspond to $x$ and $y$ Cartesian coordinates and the colour bar indicates the magnitude of the density.} \label{ang_carbon} \end{figure} \begin{figure} \centering \includegraphics[height=11cm, width=16cm]{./Figures/Angular_Nitrogen_Density_Contour_Plot_for_Constant_Theta.png} \caption{Nitrogen angular electron pair density contour plots for fixed values of $\theta$. The axes correspond to $x$ and $y$ Cartesian coordinates and the colour bar indicates the magnitude of the density.} \label{ang_nitrogen} \end{figure} \begin{figure} \centering \includegraphics[height=11cm, width=16cm]{./Figures/Angular_Oxygen_Density_Contour_Plot_for_Constant_Theta.png} \caption{Oxygen angular electron pair density contour plots for fixed values of $\theta$. The axes correspond to $x$ and $y$ Cartesian coordinates and the colour bar indicates the magnitude of the density.} \label{ang_oxygen} \end{figure} \begin{figure} \centering \includegraphics[height=13cm, width=16.2cm]{./Figures/Angular_Fluorine_Density_Contour_Plot_for_Constant_Theta.png} \caption{Fluorine angular electron pair density contour plots for fixed values of $\theta$. The axes correspond to $x$ and $y$ Cartesian coordinates and the colour bar indicates the magnitude of the density.} \label{ang_fluorine} \end{figure} \begin{figure} \centering \includegraphics[height=13cm, width=16.2cm]{./Figures/Angular_Neon_Density_Contour_Plot_for_Constant_Theta.png} \caption{Neon angular electron pair density contour plots for fixed values of $\theta$. The axes correspond to $x$ and $y$ Cartesian coordinates and the colour bar indicates the magnitude of the density.} \label{ang_neon} \end{figure} In figures \ref{ang_boron}-\ref{ang_neon}, we see that the first pair density always remains spherically-symmetric, which makes sense as this pair corresponds to the innermost electrons in the atom whose density profile is dominated by the spherical 1s contribution. In fact, because the magnitude of the density for the first pair is so much larger than the rest, the total density profile only marginally deviates from a spherical distribution. The pair densities beyond the first pair do not correspond to the orbital picture we get from other DFT approaches, and the non-spherical pair densities only resemble a single lobe in contrast to the multiple lobes expected from orbital pictures such as Hartree-Fock theory. However, pair densities do not correspond to the squared modulus of individual orbitals from KS-DFT, rather, they correspond to sums of squared moduli of orbitals, as shown in appendix \ref{AppendixC}. The pair density profiles corresponding to non-spherical pairs somewhat resemble the situation in polymer macro-phase separation \cite{matsen.incollection}, where one pair occupies one of the lobes in one region and the other occupies the partner lobe across from it, together forming a hybrid 2s-2p-like structure. This macro-phase behaviour is unsurprising given that the system is being modelled as a classical polymeric system with higher-dimensional excluded-volume, a model which is equivalent to the wavefunction picture through the theorems of DFT \cite{Thompson2022}. The total densities of the atoms seen in figures \ref{ang_carbon}-\ref{ang_neon} have approximately the same profiles as the densities predicted by quantum mechanics in the wavefunction picture, but due to the inexact Pauli-exclusion field used in this work, we do not expect to produce identical profiles. In light of the macro-phase-like behaviour seen in figures \ref{ang_carbon}-\ref{ang_neon}, there are a number of constraints that the true electron density must satisfy in order to be considered physically acceptable, which we can use as benchmarks to assess the density profiles predicted by the model used in this work. The two main constraints on the electron density are that it must be positive for all positions $\bm{r}$ and that its integral over all space must yield the electron number $N$ (or the pair electron number if we are dealing with an individual pair density). Two further constraints are \begin{align} 1 \geq \frac{3\pi}{4K}\left[\frac{\pi}{2}\int \mathrm{d}\bm{r}\, n^3(\bm{r}, \beta)\right]^{\frac{1}{3}}\ \ \text{and} \ \ 1 \geq \frac{1}{2K}\int \mathrm{d}\bm{r}\, \left(\nabla \sqrt{n(\bm{r}, \beta)}\right)^2 \label{den_constr} \end{align} where $K$ represents the kinetic energy of the system \cite{lieb2002}, which is given by eqn. \ref{kin_en} in the model. The first constraint of eqns. \ref{den_constr} is the requirement that the electron density be contained in the function space $L^3$, while the second is the requirement that the kinetic energy associated with the electron density be bounded below by the von Weizsacker kinetic energy \cite{lieb2002}. \begin{longtable}{ p{.09\textwidth} p{.13\textwidth} p{.13\textwidth} } \hline Element & Constraint 1 & Constraint 2 \\ \hline H & 0.85127 & 0.99985 \\ He & 0.87446 & 0.99983 \\ Li & 0.85268 & 0.95681 \\ Be & 0.83296 & 0.92839 \\ B & 0.82156 & 0.91523 \\ C & 0.80450 & 0.90587 \\ N & 0.79150 & 0.89451 \\ O & 0.78058 & 0.88589 \\ F & 0.77055 & 0.87868 \\ Ne & 0.76157 & 0.87234 \\ \hline \caption{The values for the right-hand side of eqns. \ref{den_constr} for hydrogen to neon using full angular basis sets and populating the pairs according to their original definition.} \label{tab5} \end{longtable} The pair density profiles corresponding to figures \ref{ang_boron}-\ref{ang_neon} clearly demonstrate that the pair densities, and thus the total density, are non-negative for all positions $\bm{r}$, since we know the density goes to zero for large $r$. Likewise, the expected electron numbers corresponding to each pair were obtained from the pair density integrals to within numerical accuracy (e.g. basis set truncation), adding up to the desired total electron number in every case. Table \ref{tab5} displays the right-hand side values of eqns. \ref{den_constr} for the elements hydrogen to neon, where we can see that the density predicted by the model always satisfies these two inequalities. In particular, one can notice that the right-hand side values of the second inequality in eqns. \ref{den_constr} for hydrogen and helium are almost exactly 1, which is the statement that the von Weizsacker kinetic energy functional is exact for one and two electron systems. \begin{longtable}{ p{.07\textwidth} p{.05\textwidth} p{.11\textwidth} p{.1\textwidth} p{.1\textwidth} p{.09\textwidth} p{.11\textwidth} p{.1\textwidth} p{.09\textwidth} p{.1\textwidth} } \hline & & $U_{e-n}$ & $U_{e-e}$ & $U_{\text{sic}}$ & $U_{P}$ & $U$ & $-S_{c}/\beta$ & $-S_{t}/\beta$ & $F$ \\ \hline \multirow{2}{*}{Pair 1} & Sph. & -69.75559 & 9.91254 & -3.62593 & 0.31151 & -63.15747 & 33.78870 & 0.00376 & -29.36501 \\ & Ang. & -69.70639 & 10.05499 & -3.62318 & 0.29381 & -62.98077 & 33.74174 & 0.00373 & -29.23530 \\ \hline \multirow{2}{*}{Pair 2} & Sph. & -10.06770 & 3.81522 & -0.63213 & 0.48931 & -6.39530 & 1.29296 & -0.01316 & -5.11551 \\ & Ang. & -8.43783 & 3.47576 & -0.58622 & 0.32278 & -5.22551 & 1.02975 & -0.01423 & -4.20999 \\ \hline \multirow{2}{*}{Pair 3} & Sph. & -5.92133 & 2.56803 & -0.35264 & 0.27463 & -3.43131 & 0.36323 & -0.01914 & -3.08723 \\ & Ang. & -8.43782 & 3.47575 & -0.58622 & 0.32278 & -5.22551 & 1.02975 & -0.01423 & -4.20998 \\ \hline \multirow{2}{*}{Total} & Sph. & -85.74463 & 16.29579 & -4.61070 & 1.07545 & -72.98408 & 35.44489 & -0.02854 & -37.56774 \\ & Ang. & -86.58204 & 17.00649 & -4.79562 & 0.93938 & -73.43179 & 35.80124 & -0.02473 & -37.65527 \\ \hline \caption{Free energy, potential energy, and the entropic contributions to the free energy corresponding to each pair in the element carbon.} \label{tab4} \end{longtable} \begin{longtable}{ p{.07\textwidth} p{.05\textwidth} p{.11\textwidth} p{.1\textwidth} p{.1\textwidth} p{.09\textwidth} p{.11\textwidth} p{.1\textwidth} p{.09\textwidth} p{.11\textwidth} } \hline & & $U_{e-n}$ & $U_{e-e}$ & $U_{\text{sic}}$ & $U_{P}$ & $U$ & $-S_{c}/\beta$ & $-S_{t}/\beta$ & $F$ \\ \hline \multirow{2}{*}{Pair 1} & Sph. & -161.60445 & 16.80288 & -5.63552 & 0.91353 & -149.52355 & 80.61756 & 0.00821 & -68.89778 \\ & Ang. & -161.7626 & 17.26576 & -5.64371 & 0.91018 & -149.23038 & 80.79204 & 0.0081 & -68.43023 \\ \hline \multirow{2}{*}{Pair 2} & Sph. & -24.96449 & 7.58292 & -1.08034 & 1.72185 & -16.74006 & 4.12557 & -0.00759 & -12.62207 \\ & Ang. & -21.39702 & 7.25637 & -1.11264 & 1.0224 & -14.23088 & 3.69921 & -0.00781 & -10.53948 \\ \hline \multirow{2}{*}{Pair 3} & Sph. & -10.73121 & 4.40419 & -0.43666 & 0.71924 & -6.04444 & 0.53809 & -0.01693 & -5.52329 \\ & Ang. & -21.39702 & 7.25638 & -1.11264 & 1.0224 & -14.23088 & 3.69922 & -0.00781 & -10.53948 \\ \hline \multirow{2}{*}{Pair 4} & Sph. & -10.73121 & 4.40419 & -0.43666 & 0.71924 & -6.04444 & 0.53809 & -0.01693 & -5.52329 \\ & Ang. & -7.67115 & 3.47493 & -0.35324 & 0.21603 & -4.33343 & 0.3893 & -0.01902 & -3.96315 \\ \hline \multirow{2}{*}{Pair 5} & Sph. & -3.40092 & 1.53289 & -0.13416 & 0.15004 & -1.85215 & 0.10301 & -0.01092 & -1.76006 \\ & Ang. & -3.38691 & 1.55932 & -0.17473 & 0.08222 & -1.9201 & 0.17628 & -0.00962 & -1.75344 \\ \hline \multirow{2}{*}{Total} & Sph. & -211.43228 & 34.72707 & -7.72335 & 4.22391 & -180.20465 & 85.92232 & -0.04415 & -94.32649 \\ & Ang. & -215.6147 & 36.81275 & -8.39695 & 3.25323 & -183.94568 & 88.75605 & -0.03616 & -95.22579\\ \hline \caption{Free energy, potential energy, and the entropic contributions to the free energy corresponding to each pair in the element fluorine.} \label{tab6} \end{longtable} Tables \ref{tab4} and \ref{tab6} list the pairwise components of the potential energy terms, the energy contributions from the configurational and translational entropies, and the total free energies for the elements carbon and fluorine, respectively. In both elements we can see that the lowering of the free energy due to spherical symmetry-breaking is produced from pairs 2 and 3, which both adopt opposing lobe shapes, in contrast to the purely spherical distributions in the spherically-averaged case, that distributes their free energy contribution uniformly amongst the two. It is this feature in particular that accounts for the free energy difference between the spherical and angular cases. Looking more closely at tables \ref{tab4} and \ref{tab6}, the cause of this feature is the fact that the third pair density can occupy a region closer to the atomic nucleus by violating spherical symmetry, which is evidenced by the much lower $U_{e-n}$ value in both elements for the angular case. The electron-electron plus self-interaction correction potential, the Pauli potential, and both entropic contributions to the free energy for pair 3 are all worse in the angular case, suggesting that the move towards the nucleus more than compensates for the interaction penalties with other pairs. In order for pair 3 to accomplish the move from a spherical distribution to a lobe distribution, pair 2 must also transform to a mirroring lobe distribution so that its original spherical shape does not overlap as much with the new lobe shape of pair 3; this effect is propagated with the other pairs (except the first), converting them into non-spherical distributions as well. \section{Discussion} \label{4} Despite the failure of the Pauli-exclusion field to satisfy the coordinate scaling relation eqn. \ref{pauli_cons}, the scaling argument presented in reference \onlinecite{LeMaitre_Thompson_2022} shows that the picture of higher-dimensional excluded-volume interactions between pairs of threads recovers the Thomas-Fermi quantum kinetic energy term and the Dirac exchange term in the uniform density limit. The analysis from section \ref{3} also demonstrates that the pair densities making up the total density satisfy all constraints necessary to guarantee a physically acceptable electron density. Together with the proof that pair densities do not necessarily correspond to individual squared moduli of orbitals, and the formal equivalence of the polymer-thread picture with quantum DFT through the quantum-classical isomorphism \cite{Feynman1953AtomicTO}, the macro-phase-like behaviour exhibited in figures \ref{ang_carbon}-\ref{ang_neon} show that spontaneous shell structure and spherical-symmetry breaking are robust predictions. The symmetry-breaking arises from the energetic benefit of electrons distributing asymmetrically closer to the nucleus. It is also clear from adding up the individual pair density profiles for any given atom, that the total density profiles only deviate slightly from spherical symmetry, which is consistent with the findings of Chowdhury and Perdew \cite{Chowdhury_Perdew2021} that asymmetries in the electron density have a small effect. Together, these two observations highlight an important distinction: the ability of the pairs to individually break spherical-symmetry allows the atom to lower its binding energy in all cases, but this does not necessarily mean that the total electron density also breaks spherical symmetry. One should consider whether the macro-phase-like behaviour is simply an artefact of the specific approximation for the Pauli potential being used in this work, as it is reasonable to speculate that the exaggerated repulsion produced by the approximate Pauli-exclusion field causes the pairs to clear their local neighbourhood. That is, are we only observing isolated atoms to be spherically-asymmetric because we are using an approximate model? If we implemented the exact Pauli potential, would total electron densities always be found to be spherically symmetric, consistent with the arguments of references \onlinecite{Cohen1965, gil1972}?. This seems unlikely, since this would require an unjustifiable perfect balance between the Pauli potential and other factors. Other frustrated systems induce spontaneous symmetry breaking, for example in true polymeric systems, SCFT is used to predict the micro-phases of block copolymers \cite{Matsen_Schick1994.article, Matsen_Bates1997, matsen.incollection}. Returning to the present model, the shell structure for carbon displayed in figure \ref{ang_carbon} demonstrates non-spherical structure yet maintains nearly identical shell structure to that predicted by Hartree-Fock theory. As was mentioned earlier, there is probably some cancellation of errors happening within carbon due in part to competing Pauli pair repulsions, but it seems unlikely based on the trends from the other atoms combined with the difference in sensitivity between the binding energies and the density profiles, that the magnitude of this effect would be large enough to account for the macro-phase-like structure while also producing the minute differences with Hartree-Fock theory that are observed. \section{Conclusions and Future Work} \label{5} The ring polymer SCFT formulation of quantum mechanics predicts the spontaneous emergence of atomic shell structure and spherical symmetry-breaking in isolated neutral atoms hydrogen to neon Using postulated pair structure of the model and ideas of higher-dimensional excluded-volume in cooperation with an exact self-interaction correction, the model shows excellent agreement with Hartree-Fock theory for the atomic binding energies and density profiles of the first six elements, providing exact matches for the elements hydrogen and helium. However, due to the approximation made on the Pauli-exclusion field, the predicted shell structure starts to deviate significantly past the element neon and the symmetry-breaking is first predicted to occur at carbon instead of boron. Consistent with Chowdhury and Perdew \cite{Chowdhury_Perdew2021}, the symmetry-breaking effect is found to have a very small impact on the binding energies, which suggests that the spherical-averaging approximation is physically reasonable when investigating atomic systems. The pair density contour plots also display behaviour similar to polymer macro-phase separation, where individual electron pairs occupy single lobe structures that together form a dumbbell shape analogous to the 2p orbital shape. It is further shown that the predicted densities satisfy known constraints and still produce the same total electronic density profile that is predicted by quantum mechanics. There are a number of future directions to consider, now that the basic engine from reference \onlinecite{LeMaitre_Thompson_2022} has been constructed. One possible direction could be to extend the work of reference \onlinecite{sillaste_thompson.article} in modelling systems of diatomic molecules to arbitrary formations of molecules or even solid-state lattices, since the Gaussian methodology is easily adapted to any number of complex geometries. The initial groundwork involved in this direction would involve switching to contracted Gaussian basis sets \cite{Helgaker_Taylor1995,huzinaga2012gaussian,Hehre_Pople_Stewart1969}, since they allow for many fewer basis functions to be used while still maintaining roughly the same level of precision and resolution; the price tag associated to the contracted sets comes in the form of additional minimization routines that either minimize the spectral representation of the free energy eqn. \ref{free_en} with respect to even-tempered parameters \cite{Schmidt_Ruedenberg1979,Helgaker_Taylor1995} or fit the Gaussians to a Slater-type function \cite{Hehre_Pople_Stewart1969,Helgaker_Taylor1995}. Both of the these methods typically require derivative information to perform the minimization, which is undesirable because the derivatives may not be well-defined or could possibly lead to numerical instabilities yielding false minima. A method that uses only the Nelder-Mead algorithm was originally developed for this work in anticipation of investigating molecular systems, which uses the spectral coefficients of the density and the Gaussian exponents from the uncontracted result to solve for the contraction coefficients, and then minimizes the sum of squared deviations between the two to solve for the exponents of the contracted set. After implementing the contracted sets and updating the current computational engine, one would need to generalize the centres of the Gaussian basis sets to arbitrary positions and modify the structure of the computation to accommodate multiple atoms. The shifting of the centres of the Gaussians from the origin to arbitrary positions is not so easily done with the spherical harmonic representation used in this work and might be better facilitated using a Cartesian representation of the Gaussians \cite{Helgaker_Taylor1995}, which only entails re-deriving the basis specific matrices and the spectral components of the electron-nucleus field. Fortunately, the expressions for these quantities have already been derived, although the expressions are much more complicated. In the case of solid-state lattices, pseudo-potentials and other modifications would need to be introduced as well \cite{Karasiev_Trickey2012,witt_del_rio_dieterich_carter2018}. Further increases in accuracy could also be achieved by combining the molecular dynamics framework of Car and Parrinello \onlinecite{car1985unified.article} with the model, to better account for the nuclear degrees of freedom. Another possible direction is to implement an exact Pauli-exclusion field, so that the comparison of the present model with Hartree-Fock theory can be completed, and the equivalence of higher-dimensional excluded-volume with the Pauli-exclusion principle can be verified. This direction would be an important test of a foundational aspect of the theory \cite{Thompson2022}, and would provide further evidence in support of the symmetry breaking mechanism presented here, in which electrons lower the free energy by breaking spherical symmetry to form electron distributions that approach closer to the atomic nucleus. Other possible directions include adding correlation terms and relativistic corrections, or clarifying the mechanism for electron spin in the model. Correlation fields are stipulated to be the only thing missing from this model that prevents it from completely agreeing with the predictions of non-relativistic quantum mechanics for neutral atoms not in the presence of any other fundamental fields. Therefore, finding a mechanism for correlations within the polymer picture would also be of foundational importance to the model. An investigation of quantum correlations would also naturally lead to the topic of quantum entanglement, which further connects with information-theoretic approaches to DFT \cite{Nagy2015,ALIPOUR2015} that may be useful in learning more about properties of correlations in many-body systems from the perspective of DFT. Investigating the mechanism that represents electron spin would also complement the study of correlations. Lastly, relativistic corrections including fine structure \cite{Liu2020,Autschbach2012} and finite-size nuclear centres \cite{Visscher_Dyall} could be added to the model in order to study heavier elements, which may yield useful information on how relativistic effects manifest themselves in the polymer picture. \section*{Acknowledgements} This research was financially supported by the Natural Sciences and Engineering Research Council of Canada (NSERC).

52cb31f1cf872e117b0d60f711fb2554853a3960

\section{Introduction}\label{Intro} Online news articles have, over time, started to replace traditional print and radio media as a primary source of information \cite{dallmann2015a}. A varying word choice may have a major effect on the public and individual perception of societal issues, especially since regular news consumers are mostly not fully aware of the degree and scope of bias \cite{spinde2020b}. As shown in existing research~\cite{park2009newscube, baumer2015a}, detecting and highlighting media bias might be relevant for media analysis and to mitigate the effects of biased reports on readers. Also, the detection of media bias can assist journalists and publishers in their work \cite{Spinde2021}. To date, only a few research projects focus on the detection and aggregation of bias \cite{lim2020annotating, Spinde2020INRA}. Even though bias embodies a complex structure, contributions \cite{hube2019neural, chen_analyzing_2020} often neglect annotator background and use crowdsourcing to collect annotations. Therefore, existing data sets exhibit low annotator agreement and inferior quality. Our study holds both theoretical and practical significance. We propose BABE (\textbf{B}ias \textbf{A}nnotations \textbf{B}y \textbf{E}xperts), a data set of media bias annotations, which is built on top of the MBIC data set \cite{spinde2021mbic}. MBIC offers a balanced content selection, annotations on a word and sentence level, and is with 1,700 annotated sentences one of the largest data sets available in the domain. BABE improves MBIC, and other data sets, in two aspects. First, annotations are performed by trained experts and in a larger number. Second, the corpus size is expanded considerably with additional 2,000 sentences. The resulting labels are of higher quality and capture media bias better than labels gathered via crowdsourcing. In sum, BABE consists of 3,700 sentences with gold standard expert annotations on the word and sentence level.\footnote{We also provide another 1,000 yet unlabeled sentences for future work. We have not labeled them to date due to resource restrictions.} To analyze the ideal trade-off between the number of sentences, annotations, and human annotation cost, we divide our gold standard into 1,700 and 2,000 sentences, which are annotated by eight and five experts, respectively.\footnote{With the 1,700 stemming from MBIC \cite{spinde2021mbic}.} Lastly, we train and present a neural BERT-based classifier that outperforms existing approaches such as the one by \citet{Spinde2021}. Even though neural network architectures have been applied to the media bias domain \cite{hube2019neural,chen_analyzing_2020}, their data sets created using crowdsourcing do not exhibit similar quality as our expert data set. In addition, we include five state-of-the-art neural models in our comparison and extend two of them in a distant supervision approach \cite{tang2014,deriu2017}. Leveraging large amounts of distantly labeled data, we formulate a pre-training task helping the model to learn bias-specific embeddings by considering bias information when optimizing its loss function. For the classification presented in this paper, we focus on sentence level bias detection, which is the current standard in related work (Section \ref{sec:relatedwork})\footnote{Our data set is in English language, which is also currently most common in the domain \cite{Spinde2021e}.}. We address future work on word level bias in Section \ref{sec:discussion}. We publish all our code and resources on \url{https://github.com/Media-Bias-Analysis-Group/Neural-Media-Bias-Detection-Using-Distant-Supervision-With-BABE}. \section{Related Work}\label{sec:relatedwork} Media bias can be defined as slanted news coverage or internal news article bias \cite{recasens2013a}. While there are multiple forms of bias, e.g., bias by personal perception or by the omission of information \cite{PUGLISI2015647}, our focus is on bias caused by word choice, in which different words refer to the same concept. For a detailed explanation of the types of media bias, we refer to \citet{Spinde2021}. In the following, we summarize the existing literature on bias data sets and media bias classification. \subsection{Media Bias Data Sets}\label{sec:data sets} \citet{lim2018b} present 1,235 sentences labeled for word and sentence level bias by crowdsource workers. All the sentences in their data set focus on one event. Another data set focusing on just one event is presented by \citet{10.1145/3340531.3412876}. It consists of 2,057 sentences from 90 news articles, annotated with bias labels on article and sentence levels, and contains labels such as overall bias, hidden assumption, and framing. The annotators agree with a Krippendorff's $\alpha$ = -0.05. \citet{lim2020annotating} also provide a second data set with 966 sentences labeled on the sentence level. However, their reported interrater-agreement (IRR) of Fleiss' Kappa on different topics averages at zero. \citet{baumer2015a} classify framing in political news. Using crowdsourcing, they label 74 news articles from eight US news outlets, collected from politics-specific RSS feeds on two separate days. \citet{chen_analyzing_2020} create a data set of 6,964 articles containing political bias, unfairness, and non-objectivity labels at the article level. Altogether, they present 11 different topics such as “presidential election”, “politics”, and “white house”. \citet{fan2019a} present 300 news articles containing annotations for lexical and informational bias made by two experts. They define lexical bias as bias stemming from specific word choice, and informational bias as sentences conveying information tangential or speculative to sway readers’ opinions towards entities \cite{fan2019a}. Their data set, BASIL, allows for analysis at the token level and relative to the target, but only 448 sentences are available for lexical bias. Under the name MBIC, \citet{spinde2021mbic} extract 1,700 sentences from 1,000 news articles. Crowdsource workers then label bias and opinion on a word and sentence level using a survey platform that also surveyed the annotators' backgrounds. MBIC covers 14 different topics and yields a Fleiss' Kappa score of 0.21. Even though the referenced data sets contribute valuable resources to the media bias investigation, they still have significant drawbacks, such as (1) a small number of topics \cite{lim2018b, lim2020annotating}, (2) no annotations on the word level \cite{lim2018b}, (3) low inter-annotator agreement \cite{spinde2021mbic, lim2020annotating, baumer2015a, lim2018b}, and (4) no background check for its participants (except \cite{spinde2021mbic}). Also, some related papers focus on framing rather than on bias \cite{baumer2015a, fan2019a}, and results are only partially transferable. Our work aims to address these weaknesses by gathering sentence level annotations about bias by word choice over a balanced and broad range of topics. The annotations are made by trained expert annotators with a higher capability of identifying bias than crowdsource workers. \subsection{Media Bias Classification Systems} Several studies tackle the automated detection of media bias \cite{hube2018detecting, spinde2020a, chen_analyzing_2020}. Most of them use manually created features to detect bias \cite{hube2018detecting}, and are based on traditional machine learning models \cite{Spinde2021}. \citet{recasens2013a} identify sentence level bias in Wikipedia using supervised classification. They use a bias lexicon and a set of various linguistic features (e.g., assertive verbs, sentiment) with a logistic regression classifier, identifying bias-inducing words in a sentence. They also report that crowdsource workers struggle to identify bias words that their classifier is able to detect. \citet{Spinde2021} create a media bias data set (i.e., MBIC) and develop a feature-based tool to detect bias-inducing words. The authors identify and evaluate a wide range of linguistic, lexical, and syntactic features serving as potential bias indicators. Their final classifier returns an $F_{1}$-score of 0.43 and 0.79 AUC. Spinde et al. point out the explanatory power of various feature-based approaches and the performance of their own model on the MBIC data set. Yet, their results indicate that Deep Learning models are promising alternatives for future work. \citet{hube2018detecting} propose a semi-automated approach to extract domain-related bias based on word embeddings properties. The authors combine bias words and linguistic features (e.g., report verbs, assertive verbs) in a random forest classifier to detect sentence level bias in Wikipedia. They achieve an $F_{1}$-score of 0.69 on a newly created ground truth based on Conservapedia.\footnote{\url{https://conservapedia.com/Main\_Page}, accessed on 2021-04-10.} In their following work, \citet{hube2019neural} propose a neural statement-level bias detection approach based on Wikipedia data. Using recurrent neural networks (RNNs) and different attention mechanisms, the authors achieve an $F_{1}$-score of 0.77, indicating a possible advantage of neural classifiers in the domain. \citet{chen_analyzing_2020} train a RNN to classify article-level bias. They also conduct a reverse feature analysis and find that, at the word level, political bias correlates with categories such as negative emotion, anger, and affect. To summarize, most approaches use manually created features, leading to lower performance and poor representation. The few existing contributions on neural models are based on naive data sets (cf. Section \ref{sec:data sets}). Therefore, we decided to develop a neural classifier trained on BABE. Our system incorporates state-of-the-art models and improves their pre-training step through distant supervision \cite{tang2014,deriu2017}, allowing the model to learn bias-specific embeddings, thus improving its representation. Almost all models focus on sentence level bias, describing it as the lowest meaningful level that can be aggregated to higher levels, like the document level. Therefore, we follow the standard practice and construct a sentence level classifier. \section{Data Set Creation}\label{sec:data} Since media bias by word choice rarely depends on context outside the sentences \cite{fan2019a}, we focused on gathering sentences only. To tackle the weaknesses of existing bias data sets, we created a robust and diverse corpus containing \textbf{B}ias \textbf{A}nnotations \textbf{B}y \textbf{E}xperts (BABE). \subsection{Data Collection} The general data collection and annotation pipeline is outlined in Figure \ref{workflow}. Similar to the filtering strategy proposed by \citet{Spinde2021}, the sentences should contain more biased than neutral sentences. BABE contains 3,700 sentences, 1,700 from MBIC \cite{spinde2021mbic} and additional 2,000. Like \citet{spinde2021mbic}, we extracted our sentences from news articles covering 12 predefined controversial topics.\footnote{The list of topics is provided at the repository mentioned in Section \ref{Intro}.} The articles were published on 14 US news platforms from January 2017 until June 2020. We focused on the US media since their political scenario became increasingly polarizing over the last years \cite{atkins2016skewed}. \begin{figure}[H] \centering \small \def0.905{0.905} \input{Figures/workflow_pdf_V2} \caption{\label{workflow}Data collection and annotation pipeline} \end{figure} We selected appropriate left-wing, center, and right-wing news outlets based on the media bias chart provided by Allsides.\footnote{\url{https://www.allsides.com/media-bias/media-bias-chart}, accessed on 2021-04-13.} The sentence collection was performed on the open-source media analysis platform Media Cloud.\footnote{\url{https://mediacloud.org/}, accessed on 2021-04-13.} The collection process was as follows. We defined keywords describing every topic in one word or a short phrase, specified the news outlets, their time frame, and retrieved all available links for the relevant articles.\footnote{The keywords can be found at the repository mentioned in Section \ref{Intro}.} Then, we extracted sentences by manually inspecting the provided list of articles. The sentence selection was based on our media bias annotation guidelines comprising diverse examples of biased and neutral text instances (see Section \ref{sec:data_annotation}). \subsection{Data Annotation} \label{sec:data_annotation} As laid out in Section \ref{sec:relatedwork}, high-quality annotations are often obtained if the participants are properly instructed and have sufficient training \cite{fan2019a,Spinde2021}. We compare our expert annotations with the crowdsourced labels provided by \citet{spinde2021mbic} to further analyze quality differences between the two groups. Our results show that expert annotators render more qualitative bias labels than MBIC's crowdsourcers. We define as an expert a person with at least six months of experience in the media bias domain and underwent sufficient training to (1) reliably identify biased wording, (2) distinguish between bias and plain polarizing language, and (3) take on a politically neutral viewpoint when annotating.\footnote{Note: We cannot guarantee that a media bias expert is fully neutral, but we assume that an expert is able to leave political viewpoints aside to a substantial extent.} To build up such experience, we developed detailed instruction guidelines that are presented before the annotation task.\footnote Available on the repository mentioned in Section \ref{Intro}.} The instructions are substantially more comprehensive than instructions in a crowdsourcing setting. Considering that the annotation of bias on a fine-grained linguistic level is a complex task, and cognitive and language abilities likely have an impact on text perception \cite{kause2019framing}, we hired only master students from programs completely held in English, who were among the top 20\% with respect to their grade. Based on an iterative feedback loop between all annotators and us, we refined the guidelines multiple times with richer and clearer details. We discussed and evaluated existing annotations weekly as a group during the first three weeks of each annotator's work. We also always asked each annotator to hand in annotations before the discussion sessions, so they could not influence each other. The annotators had to provide basic reasoning about their annotation decisions during our discussions. We maintained the labels only if the annotators were able to elaborate their annotations. Annotations of one annotator were discarded based on this method. Apart from evaluation and instructions, each annotator rated at least 1,700 sentences to improve experience over time.\footnote{The same sentences as in MBIC.} On average, per hour, they were paid 15,00\texteuro ~and labeled 40 sentences, costing approximately 10,000\texteuro . The sum of money required to obtain a sufficient number of reasonable bias labels can be restrictive for media bias research. Therefore, BABE represents a major contribution that alleviates the lack of high-quality annotations in the domain. The annotators were instructed to label carefully and not as fast as possible, even though this resulted in a higher overall cost. The general instructions for the annotation task were identical to the approach by \citet{spinde2021mbic}. First, raters were asked to mark words or phrases inducing bias. Then, we asked them to indicate whether the whole sentence was \textit{biased} or \textit{non-biased}. Lastly, the annotators labeled the sentence as \textit{opinionated}, \textit{factual}, \textit{or mixed}. As our resources were limited and the ideal trade-off between the number of sentences and annotators per sentence is not yet determined, we organized BABE into subgroups (SG), as described below: \begin{itemize} \item \textbf{SG1}. 1,700 sentences annotated by eight expert raters each. \item \textbf{SG2}. 3,700 sentences annotated by five expert raters each. \end{itemize} For SG1, we hired eight raters to annotate the 1,700 sentences (same as MBIC) on word and sentence levels \citep{spinde2021mbic}.\footnote{In the original MBIC data set, each sentence was evaluated by ten crowdsource workers \cite{spinde2021mbic}.} Thereby, we obtained an expert-labeled ground truth comparable to MBIC's crowdsourcing results. For SG2, five of the previous eight annotators also labeled the 2,000 additionally collected sentences. We explored the ideal number of annotators by sampling. 5 annotators is a compromise between the agreement quality for both the bias and opinion labels, assuming that the annotation quality stays the same. To show the difference to 8 annotators, and as an outlook into future extensions of the data set, we also release the codings made by 8 raters\footnote{But recommend to use 5-person ratings when using the full data set.}. We will also add detailed statistics and results about all data and point out our selection process more clearly. As resources and time were limited, we leave the inclusion of further annotators and more sentences to future work. All raters were master students with a background in Data Science, Computer Science, Psychology, or Intercultural Communication. The groups and their annotators are described in detail in the repository mentioned in Section \ref{Intro}. \subsection{Evaluation of Data Sets} The raw labels obtained during the annotation phase were processed as follows. We calculated an aggregated bias/opinion label for every sentence based on a majority vote principle. For instance, if a sentence was labeled as biased by more than four expert annotators in SG1, we assigned the label \textit{biased} to the sentence. Otherwise, the sentence was marked as \textit{non-biased}.\footnote{Note: In SG2, the threshold reduced respectively due to the lower number of expert annotators.} The annotators did not agree on a label (no majority vote) in some sentences. Here, we assigned the label \textit{no agreement}. Our annotation scheme allows respondents to mark biased words. In SG1, a word is marked as biased if at least three annotators label it as such. In SG2, the threshold is subsequently reduced to two expert annotators labeling a word as biased.\footnote{We manually inspected all instances to determine reasonable thresholds.} We compute agreement metrics on the sentence level to acquire knowledge about data quality resulting from all annotation approaches. Our agreement metric choice is Krippendorff's $\alpha$ \cite{krippendorff2011computing}, which is a robust agreement metric for studies including varying numbers of annotators per text instance \citep{antoine-etal-2014-weighted}. We first compared the annotations resulting from MBIC's crowdsourcing approach with our expert-based approach, including eight annotators labeling 1,700 sentences (SG1). Table \ref{results_table} shows the agreement scores for the bias and opinion labels on a sentence level. Considering the bias agreement, SG1 exhibits fair agreement ($\alpha$ = 0.39) and outperforms MBIC's agreement score ($\alpha$ = 0.21).\footnote{The scoring interpretations are based on guidelines published by \citet{landis1977measurement}.} A similar pattern can be observed regarding the opinion labels (i.e., SG1: $\alpha$ = 0.46; MBIC: $\alpha$ = 0.26). Furthermore, MBIC's crowdsourcers labeled more words as biased compared to SG1's experts, i.e., 3,283 vs. 1,530 (absolute) and 2.40 vs. 1.95 (average per biased sentence). Even though media bias detection is generally a difficult task, our inter-annotator agreement is much higher than in existing research in the domain, where $\alpha$ ranges between 0 and 0.20, as shown in Section \ref{sec:relatedwork}. \begin{table}[t] \centering \resizebox{0.4\textwidth}{!}{ \begin{threeparttable} \begin{tabular}{lcc} \hline \multirow{2}{*}{\textbf{Metric}} & \multicolumn{2}{c}{\textbf{Data}}\\ \cline{2-3} & SG1 & MBIC \\ \hline Bias Agreement\textsuperscript{1} & 0.39 & 0.21\\ Opinion Agreement\textsuperscript{1} & 0.46 & 0.26 \\ Total Biased Words & 1530\textsuperscript{3} & 3283\textsuperscript{3} \\ $\varnothing$ Biased Words \textsuperscript{2} & 1.95 & 2.40 \\ \hline \end{tabular} \begin{tablenotes} \small \item[1] Calculated based on Krippendorff's $\alpha$ \item[2] Average of bias words per biased sentence \item[3] Out of 56,826 words in total \end{tablenotes} \end{threeparttable} } \caption{\label{results_table} Annotation results for the expert-annotated (SG1) and crowdsourced (MBIC) approach based on 1,700 sentences.} \end{table} Table \ref{label_dist} shows the label distribution comparison between SG1 and MBIC.\footnote{Absolute numbers for all labels are reported in the code files at the repository mentioned in Section \ref{Intro}.} We can observe that our expert annotators (SG1) are more conservative in their annotation than the crowdsourcers (MBIC). In the expert data, 43.88\% of the sentences are labeled as biased, whereas the crowdsources annotated 59.88\%. The opinion labels' distribution is fairly balanced in both the expert annotator and crowdsourced data. Factual sentences occur slightly more often than opinionated sentences in both data sets. \begin{table} \centering \resizebox{0.36\textwidth}{!}{ \begin{threeparttable} \begin{tabular}{lcc} \hline \multirow{2}{*}{\textbf{Label}} & \multicolumn{2}{c}{\textbf{Data}}\\ \cline{2-3} & SG1 & MBIC \\ \hline Biased & 43.88\% & 59.88\%\\ Non-biased & 47.05\% & 31.35\% \\ No agreement & 9.05\% & 8.76\% \\ \cdashline{1-3} Opinionated & 25.00\% & 30.65\%\\ Factual & 37.59\% & 33.65\% \\ Mixed & 26.64\% & 25.47\% \\ No agreement & 10.76\% & 10.24\% \\ \hline \end{tabular} \end{threeparttable} } \caption{\label{label_dist} Class distribution for SG1's and MBIC's 1700 sentences. } \end{table} Next, we evaluate our expert-based annotation approach, including five expert annotators labeling 3,700 sentences (SG2) in comparison to 1,700 (SG1). We compare metrics between both approaches to ascertain whether the reduced number of annotators in SG2 has a substantial impact on the annotator agreement. The finding could yield implications for future research on our extended dataset (SG2). Table \ref{results_table2} shows agreement metrics for the bias and opinion labels of both expert-annotated approaches, and Table \ref{results_table3} represents label distributions. SG2 exhibits moderate agreement ($\alpha$ = 0.40) in the bias annotation task, and slightly outperforms SG1 ($\alpha$ = 0.39). Regarding the opinion labels, we observe a similar pattern, with SG2 outperforming SG1 more substantially (SG2: $\alpha$ = 0.60; SG2: $\alpha$ = 0.46). The expert annotators of SG1 are more conservative in labeling bias than SG2 (SG1: 43.88\% vs. SG2: 49.26\% labeled as biased).\footnote{Due to the uneven number of annotators in SG2, "no agreement" cases do not exist here.} The opinion labels are distributed marginally skewed in both annotator groups. Factual sentences occur more often than opinionated sentences in both data sets. Further statistics on SG 1 and SG 2 such as bias/opinion distribution per news outlet and topic, the connection between bias and opinion, and the overall topic distribution are provided in the repository mentioned in Section \ref{Intro}. \begin{table}[t] \centering \resizebox{0.4\textwidth}{!}{ \begin{threeparttable} \begin{tabular}{lcc} \hline \multirow{2}{*}{\textbf{Metric}} & \multicolumn{2}{c}{\textbf{Data}}\\ \cline{2-3} & SG1 & SG2 \\ \hline Bias Agreement\textsuperscript{1} & 0.39 & 0.40\\ Opinion Agreement\textsuperscript{1} & 0.46 & 0.60 \\ \hline \end{tabular} \begin{tablenotes} \small \item[1] Calculated based on Krippendorff's $\alpha$ \end{tablenotes} \end{threeparttable} } \caption{\label{results_table2}Data set annotation results for the expert-based approaches (left: eight annotators labeling 1,700 sentences (SG1); right: five annotators labeling 3,700 sentences (SG2)).} \end{table} \begin{table}[t] \centering \resizebox{0.4\textwidth}{!}{ \begin{threeparttable} \begin{tabular}{lcc} \hline \multirow{2}{*}{\textbf{Label}} & \multicolumn{2}{c}{\textbf{Data}}\\ \cline{2-3} & SG1 & SG2 \\ \hline Biased & 43.88\%& 49.26\%\\ Non-biased & 47.05\% & 50.70\% \\ No agreement & 9.05\% & 0.00\% \\ \cdashline{1-3} Opinionated & 25.00\% & 23.35\%\\ Factual & 37.59\% & 43.54\% \\ Mixed & 26.64\% & 27.21\% \\ No agreement & 10.76\% & 5.88\% \\ \hline \end{tabular} \end{threeparttable} } \caption{\label{results_table3}Data set class distribution for the expert-based approaches (left: eight annotators labeling 1,700 sentences (SG1); right: five annotators labeling 3,700 sentences (SG2)).} \end{table} \section{Methodology}\label{sec:methodology} We propose the use of neural classifiers with automated feature learning capabilities to solve the given media bias classification task. A distant supervision framework, similar to \citet{tang2014}, allows us to pre-train the feature extraction algorithms leading to improved language representations, thus, including information about a sample's bias. As obtaining large amounts of pre-training labeled data using humans is prohibitively expensive, we resort to noisy yet abundantly available labels that provide supervisory signals. \subsection{Learning Task} Given a corpus $X$ and a randomly sampled sequence of tokens $x_i \in X$ with $i \in \{1,...,N\}$, the learning task consists of assigning the correct label $y_i$ to $x_i$ where $y_i \in \{0,1\}$ represents the \textit{neutral} and \textit{biased} classes, respectively. The supervised task can be optimized by minimizing the binary cross-entropy loss \begin{equation} \label{eq:loss} \mathcal{L} := - \frac{1}{N} \sum_{i=1}^{N} \sum_{k=\{0,1\}} f_k(x_i) \cdot log(\hat{f_k}(x_i)). \end{equation} where $f_k(\cdot)$ is a binary indicator triggering 0 in the case of neutral labels and 1 in the case of a biased sequence. $\hat{f}_k(\cdot)$ is a scalar representing the language model score for the given sequence. \subsection{Neural Models} We fit $\hat{f}_k(\cdot)$ using a range of state-of-the-art language models. Central to the architectural design of these models is \citet{vaswani2017}'s encoder stack of the Transformer relying solely on the attention mechanism. Specifically, we use the BERT model \citep{devlin2018} and its variants DistilBERT \citep{sanh2019} and RoBERTa \citep{liu2019} that learned bidirectional language representations from the unlabeled text. DistilBERT is a compressed model of the original BERT, and RoBERTa uses a slightly different loss function with more training data than its predecessor. We also evaluate models built on the transformer architecture but differ in the training objective. While DistilBERT and RoBERTa use masked language modeling as a pre-training task, ELECTRA \citep{clark2020} uses a discriminative approach to learn language representations. We also include XLNet \citep{yang2019} in our comparison as an example of an autoregressive model. We systematically evaluate the models' performance on the media bias sentence classification task. We also investigate the impact of an additional pre-training task introduced in the next section on the BERT and RoBERTa models' classification capabilities. \subsection{Distant Supervision} Fine-tuning general language models on the target task has proven beneficial for many tasks in NLP \cite{howard2018}. The language model pre-training followed by fine-tuning allows models to incorporate the idiosyncrasies of the target corpus. For text classification, the authors of ULMFiT \cite{howard2018} demonstrated the superiority of task-specific word embeddings. Before fine-tuning, we introduce an additional pre-training task to improve feature learning capabilities considering media bias content. The typical unsupervised setting used in the general pre-training stage does not include information on language bias in the learning of the embedded space. To remedy this, we incorporate bias information directly in the loss function (equation \ref{eq:loss}) via distant supervision. In this approach, distant or \textit{weak} labels are predicted from noisy sources, alleviating the need for data labeled by humans. Results by \citet{severyn2015} and \citet{deriu2017} demonstrated that pre-training on larger distant datasets followed by fine-tuning on supervised data yields improved performance for sentiment classification. A pre-training corpus is compiled consisting of news headlines of outlets with and without a partisan leaning to learn bias-specific word embeddings. The data source, namely, the news outlets, are leveraged to provide distant supervision to our system. As a result, the large amounts of data necessary to learn continuous word representations are gathered by mechanical means alleviating the burden of collecting expensive annotations. The assumption is that the distribution of biased words is denser in some news sources than in others. Text sampled from news outlets with a partisan leaning according to the Media Bias Chart is treated as biased. Text sampled from news organizations with high journalistic standards is treated as neutral. Thus, the mapping of bias and neutral labels to sequences is automatized. The data collection resembles the collection of the ground-truth data described in Section \ref{sec:data}. The defined keywords reflect contentious issues of the US society, as we assume slanted reporting to be more likely among those topics than in the case of less controversial topics. The obtained corpus consisting of 83,143 neutral news headlines and 45,605 biased instances allows for the encoding of a sequence's bias information in the embedded space. The news headlines corpus serves to learn more effective language representations, it is not suitable for evaluation purposes due to its noisy nature. We ensure that no overlap exists between the distant corpus and BABE to guarantee model to guarantee model integrity with respect to training and testing. \section{Experiments}\label{sec:experiments} \textbf{Training Protocol.} We implement the neural models with HuggingFace's Transformer API \citep{wolf2020}. The model components are instantiated with their pre-trained parameters. Parameters of the classification components are uniformly instantiated and learned. First, we fine-tune and evaluate neural models on BABE. Second, we identify the best performing model of the first run and include the distant supervision pre-training task. \textbf{Implementation.} The hyperparameters remain unchanged for pre-training on the distant corpus and fine-tuning on BABE. Sentences are batched together with 64 sentences per mini-batch because estimating gradients in an online learning situation resulted in less stable estimates. To optimize $\mathcal{L}$, we use the Adam optimization with a learning rate of $5^{-5}$ \citep{kingma2014}. Training on the distantly labeled corpus is performed for one epoch. While training on BABE, convergence can be observed after three to four epochs. A monitoring system is in place that stops training after two epochs without improvement of the loss and restores the parameters of the best epoch. All computations were performed on a single Tesla T4 GPU. All in all, pre-training and training of all models is executed in 5 hours. \textbf{Baseline.} To assess the benefit of modern language models for the domain of media bias, we compare their performance to a traditional feature-based model (Baseline). We use the work by \citet{Spinde2021} as our baseline method, as it offers the most complete set of features for the media bias domain. The authors use syntactic and lexical features related to bias words such as dictionaries of opinion words \citep{hu2004}, hedges \citep{hyland2018a} and assertive and factive verbs \citep{hooper1975a}. \citet{Spinde2021}'s classifier serves as a baseline to evaluate our approach. As feature-based models operate on the word level, we provide comparability by implementing the classification rule that the presence of a predicted biased word leads to the overall sentence being labeled as biased. In contrast, if the baseline model does not label words as biased in a given sequence, the sequence will be classified as neutral. \textbf{Evaluation Metric.} Given the relatively small size of 3,700 sequences in BABE, we report performance metrics averaged on a 5 fold cross-validation procedure to stabilize the results. Because the class distribution in SG1 is slightly unbalanced, we use stratified cross-validation to preserve this imbalance in each fold. Following the standard in the literature, we report a weighted average of $F_1$-scores. \section{Results}\label{sec:results} Table \ref{Tab:class_res} summarizes our performance results. Our baseline using engineered features exhibits low scores of 0.511 and 0.569 for SG1 and SG2, respectively.\footnote{In this Section, we show three decimal places to account for detailed model differences.} BERT improves over the baseline by a large margin of 0.251 points on SG1 and 0.220 points on SG2. DistilBERT exhibits a lower performance for both corpora, whereas RoBERTa is the strongest representative of BERT-based models. Both models based on a different training approach than BERT, namely ELECTRA and XLNet, do not match the performance of BERT and its optimized variants. These results reaffirm established findings of the attention mechanism's advantage over traditional models \cite{e23030283} and indicate the benefits of large pre-trained models' for media bias detection. Models trained and evaluated on SG2 generally perform better due to their bigger corpus size. The increase is around 0.02 points of the macro $F_1$-score for all models except RoBERTa + distant, where it is insignificant. Overall, we believe the improvement indicates that extending the data set in the future will be valuable. Results of the fourth block of table \ref{Tab:class_res} show that the distant supervision pre-training task leads to an improvement over BERT and RoBERTa. Our best performing model BERT + distant on SG2 achieves a macro $F_1$-score of 0.804 and improves over the BERT model by 0.02 points. Media bias can be better captured when word embedding algorithms are pre-trained on the news headlines corpus with distant supervision based on varying news outlets. With the added data, information on a sequence's bias is incorporated in the loss function, which is not the case in "general purpose" language models. \begin{table}[ht] \centering \resizebox{0.49\textwidth}{!}{ \begin{threeparttable} \begin{tabular}{p{3cm}cc} \hline \multirow{2}{*}{\textbf{Model}} & \multicolumn{2}{c}{\textbf{Macro $F_1$}} \\ \cline{2-3} & SG1 & SG2 \\ \hline Baseline & 0.511 (0.008) & 0.569 (0.008) \\ \cdashline{1-3} BERT & 0.762 (0.019) & 0.789 (0.011) \\ DistilBERT & 0.758 (0.029) & 0.777 (0.009)\\ RoBERTa & 0.775 (0.023) & 0.799 (0.011)\\ \cdashline{1-3} ELECTRA & 0.742 (0.020) & 0.760 (0.013)\\ XLNet & 0.760 (0.042) & 0.797(0.015) \\ \cdashline{1-3} BERT + distant & 0.778 (0.017) & \textbf{0.804} (0.014) \\ RoBERTa + distant & \textbf{0.798} (0.022) & 0.799 (0.017)\\ \hline \end{tabular} \begin{tablenotes} \small \item Standard errors across folds in parentheses. \item The first model block shows the best results of feature-based models. The second block of models consists of BERT and optimize variants. The models in the third block use new architectural or training approaches. The fourth block refers to models having learned bias-specific embeddings from the distantly supervised corpora. \item The best results are printed in \textbf{bold}. \end{tablenotes} \end{threeparttable} } \caption{Stratified 5 fold cross-validation results.} \label{Tab:class_res} \end{table} \section{Discussion}\label{sec:discussion} Employing annotators with domain expertise allows us to achieve an inter-annotator agreement of $\alpha$ = 0.40, which is higher than existing data sets \cite{spinde2021mbic}. We believe domain knowledge and training alleviate the difficulty of identifying bias and are imperative to create a strong benchmark due to the complexity of the task. In future work, apart from improving the current data set and classifier, we will also explore why a text passage might be biased, not just its overall classification. Currently, traditional machine learning models are interpretable \cite{Spinde2021} but outperformed by recurrence and attention-based models. Hand-crafted features like static dictionaries cannot adequately address the complexity and context-dependence of bias. We argue that standard metrics (e.g., accuracy and $F_1$) provide a limited perspective into a model's predictive power in case of a complex construct like media bias. Further research needs to tackle these pitfalls to propose systems with better generalization capabilities. A promising starting point might be a more refined evaluation scheme that decomposes the bias detection task into multiple sub-tasks, such as presented in CheckList \citep{ribeiro2020beyond}. This scheme will also allow us to understand how our system performs on different types of bias (e.g., bias by context, by linguistics, by overall reporting). Additionally, we believe that current research on explainable artificial intelligence might increase users' trust in neural-based classifiers. Existing research already presents ways to visualize Transformer-based models and make their results more accessible and interpretable \citep{vig2019}. Lastly, combining neural methods with advances in linguistic bias theory \cite{Spinde2021} to explain a classifier's decision to users will also be part of our future work. For this work, we focused on sentence level bias, which is often used in the media bias domain. Still, in addition to the 3,700 labeled sentences, we also include word level annotations in our data set to encourage solutions focusing on more granular characteristics. We believe that word level bias conveys strong explanatory and structural knowledge and see a detailed word level bias analysis and detection as a promising research direction. \section{Conclusion}\label{sec:conclusion} This work proposes BABE, a new high-quality media bias data set. BABE contains 3,700 labeled sentences, and enables us to compare crowdsourcing and expert annotations directly. Additionally, we propose a sentence level bias classifier based on BERT, which outperforms existing work in the domain. By deriving bias-specific word embeddings using distant supervision, we have improved our classifier even more, achieving a macro $F_1$-score = 0.804. We make all models, data, and code publicly available.\footnote{We publish the link in Section \ref{Intro}.} \section*{Ethics/Broader Impact Statement} Detecting and highlighting media bias instances may have many positive implications and can mitigate the effects of such biases \cite{baumer2015a}. Still, bias is a highly sensitive topic, and some forms of bias especially rely on other factors than the content itself, such as a different perception of any text related to the individual background of a reader. When showing detected bias or news outlet classifications on a political or polarization scale to a reader, every algorithm should be transparent in how the classifications were made. In general, the topic should be handled carefully. We want to point out that it is uncertain if and how actual news consumers would like to obtain such information. Some research groups working on the detection of bias have also started to work on psychological and societal questions related to bias \cite{spinde2020b}. From a social science perspective, it remains to be explored how a classifier can mitigate the negative effects of biased media on society. Generally, when performed in a balanced and transparent way, bias detection might positively affect collective decision-making and opinion formation processes. As such, and to this point, we see no immediate negative ethical or societal impacts of our work beyond what applies to other core building blocks of deep learning. Apart from the system transparency, as mentioned above, one important factor to consider when building, training, and presenting any media bias classifier is a manipulation protection strategy. Participants in any study, especially public ones, should not be able to tweak algorithms and therefore, e.g., flag neutral content as biased to undermine the validity of media bias detection systems. Hence, annotations should always be compared among multiple users, where trustworthiness can at least be largely assured. In open (crowdsourcing) scenarios, collecting user characteristics and consciously implementing specific content (like questions that should give an obvious answer but might be answered differently when users a following any pattern) is important. As a side effect of our project, we experienced that our annotators learned to read the news more critically and reflected more about what they read even after the study ended. We have already started to implement the insights we gained into ways to improve the perception of bias in a game, teaching players to read news with greater care and execute a large study investigating how such a game can affect children, especially in school. Our data set is completely anonymized to preserve the identities of everyone involved. \section*{Acknowledgments} The Hanns-Seidel-Foundation, Germany, supported this work, as did the DAAD (German Academic Exchange Service). We are grateful to our raters, who we will keep anonymous, as we are grateful to Dr. Franz Hahn and Prof. Dr. Jelena Mitrović for helping us to hire a part of the annotators.

3150924fe1ca6704649ad165416200228f508862

\section{Introduction} Let $\scr P(\R^d)$ be the space of all probability measures on $\R^d$ equipped with the weak topology. Consider the following distribution dependent SDE on $\R^d$: \beq\label{E1} \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D X_t= b_t(X_t, \scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{X_t})\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D t+ \sigma} \def\ess{\text{\rm{ess}}_t(X_t, \scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{X_t})\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D B_t,\ \ t\in [0,T], \end{equation} where $T>0$ is a fixed time, $\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{X_t}$ is the distribution of $X_t$, $$b: [0,T]\times\R^d\times\scr P(\R^d)\rightarrow}\def\l{\ell\R^d,\ \ \sigma} \def\ess{\text{\rm{ess}}: [0,T]\times\R^d\times\scr P(\R^d) \rightarrow}\def\l{\ell \R^d\otimes\R^d$$ are measurable, and $B_t$ is a $d$-dimensional Brownian motion on a complete filtration probability space $(\Omega,\F,\{\F_t\}_{t\in[0,T]},\P)$. We investigate the regularity in initial distributions for solutions to \eqref{E1}. More precisely, for $k> 1$ let $$\scr P_k(\R^d):=\big\{\mu\in \scr P(\R^d):\ \|\mu\|_k:= \mu(|\cdot|^k)^{\ff 1 k}<\infty\big\},$$ which is a Polish space under the $L^k$-Wasserstein distance $$\mathbb W_k(\mu,\nu)= \inf_{\pi\in \scr C} \def\aaa{\mathbf{r}} \def\r{r(\mu,\nu)} \bigg(\int_{\R^{d}\times\R^{d}} |x-y|^k \pi(\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D x,\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D y)\bigg)^{\ff 1 {k}},\ \ \mu,\nu\in \scr P_k(\R^d), $$ where $\scr C} \def\aaa{\mathbf{r}} \def\r{r(\mu,\nu)$ is the set of all couplings of $\mu$ and $\nu$. When \eqref{E1} is well-posed for distributions in $\scr P_k(\R^d)$, i.e. for any $\F_0$-measurable initial value $X_0$ with $\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{X_0}\in \scr P_k(\R^d)$ (correspondingly, any initial distribution $\mu\in \scr P_k(\R^d)$), the SDE \eqref{E1} has a unique solution (correspondingly, a unique weak solution) with $\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{X_\cdot}\in C([0,T], \scr P_k(\R^d)),$ we consider the regularity of the maps $$\scr P_k(\R^d)\ni \mu\mapsto P_t^*\mu:=\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{X_t}\ \text{for}\ \scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{X_0}=\mu,\ \ t\in (0,T].$$ Since $P_t^*\mu$ is uniquely determined by \beq\label{PT} P_t f(\mu):=\int_{\R^d} f \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D(P_t^*\mu),\ \ f\in \B_b(\R^d),\end{equation} where $\B_b(\R^d)$ is the space of bounded measurable functions on $\R^d$, we study the regularity of functionals $$\scr P_k(\R^d) \ni \mu \mapsto P_tf(\mu), \ \ \ t\in (0,T], f\in \B_b(\R^d).$$ When the noise is distribution free, i.e. $ \sigma} \def\ess{\text{\rm{ess}}_t(x, \mu)=\sigma} \def\ess{\text{\rm{ess}}_t(x)$ does not depend on the distribution argument $\mu$, the log-Harnack inequality \beq\label{LH} P_t\log f(\mu)\le \log P_tf(\nu)+ \ff{c}t \mathbb W_2(\mu,\nu)^2,\ \ 0<f\in \B_b(\R^d), t\in (0,T], \mu,\nu\in \scr P_2(\R^d),\end{equation} for some constant $c>0$ has been established in \cite{HW18, HW22a, RW, FYW1, FYW3} under different conditions, see also \cite{HRW19,HS} for extensions to the infinite-dimensional case. A crucial application of this inequality is that it is equivalent to the entropy-cost estimate \begin{equation*} \Ent(P_t^*\nu|P_t^*\mu)\le \ff{c}t \mathbb W_2(\mu,\nu)^2,\ \ t\in (0,T], \mu,\nu\in \scr P_2(\R^d),\end{equation*} where $\Ent(\nu|\mu)$ is the relative entropy of $\nu$ with respect to $\mu$. With this estimate, the exponential ergodicity of $P_t^*$ in entropy is proved in \cite{RW} for a class of time-homogeneous distribution dependent SDEs. The study of \eqref{LH} goes back to \cite{W97,W10} where the family of dimension-free Harnack inequalities is introduced, see \cite{Wbook} for various applications of this type inequalities. Another crucial tool characterizing the regularity of $\mu\mapsto P_t^*\mu$ is the following Bismut type formula for the intrinsic derivative $D^I$ in $\mu\in\scr P_k(\R^d)$ (see Definition 2.1 below): \beq\label{BS} \begin} \def\beq{\begin{equation}} \def\F{\scr F{split}&D^I_\phi P_tf(\mu)= \E\bigg[ f(X_t^\mu) \int_0^t\<M_s^{\mu,\phi},\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D B_s\>\bigg],\\ &\ \ \ \qquad t\in (0,T],\ f\in \B_b(\R^d), \phi\in L^{k}(\R^d\rightarrow}\def\l{\ell\R^d;\mu),\end{split} \end{equation} where $\int_0^t\<M_s^{\mu,\phi},\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D B_s\>$ is a martingale depending on $\mu$ and $\phi$. Again, when $\sigma} \def\ess{\text{\rm{ess}}_t(x,\mu)=\sigma} \def\ess{\text{\rm{ess}}_t(x)$ is distribution free, this type formula has been established in \cite{BRW, HSW, RW19, FYW3} under different conditions, but it is open for distribution dependent noise. However, arguments used in the above mentioned references do not apply to distribution dependent noise. The only known log-Harnack inequality for distribution dependent noise is established in \cite{BH} for Ornstein-Ulenbeck type SDEs whose solutions are Gaussian processes and thus easy to manage. This again does not apply to more general case. On the other hand, intrinsic derivative estimates have been presented for a class of SDEs with distribution dependent noise, see \cite{HW21} and references. This convinces us of establishing the log-Harnack inequality and Bimsut formula for SDEs with distribution dependent noise. In this paper, we propose a noise decomposition argument which reduces the study of distribution dependent noise to distribution free noise. For simplicity, we only explain here the idea on establishing the log-Harnack inequality for the following distribution dependent SDE: \beq\label{EMV} \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D X_t= b_t(X_t, \scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{X_t})\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D t+ \sigma} \def\ess{\text{\rm{ess}}_t( \scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{X_t})\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D B_t,\ \ t\in [0,T]. \end{equation} When the equation is well-posed for distributions on $\scr P_2(\R^d)$, let $P_t^*\mu=\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{X_t}$ for the solution with $\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{X_0}=\mu\in\scr P_2(\R^d)$. Assume that $\sigma} \def\ess{\text{\rm{ess}}_t$ is bounded and Lipschitz continuous on $\scr P_2(\R^d)$, such that $$ (\sigma} \def\ess{\text{\rm{ess}}_t\sigma} \def\ess{\text{\rm{ess}}_t^*)(\gg)\ge 2\ll^2 I_d,\ \ \gg\in \scr P_2(\R^d) $$ holds for some constant $\ll>0$, where $I_d$ is the $d\times d$ identity matrix. We take $$\tilde} \def\Ric{\text{\rm{Ric}}\sigma} \def\ess{\text{\rm{ess}}_t(\gg):= \ss{(\sigma} \def\ess{\text{\rm{ess}}_t\sigma} \def\ess{\text{\rm{ess}}_t^*)(\gg)- \ll^2 I_d}.$$ Then $\tilde} \def\Ric{\text{\rm{Ric}}\sigma} \def\ess{\text{\rm{ess}}_t(\gg)\ge \ll I_d$, and \cite[Lemma 3.3]{PW} implies that $\tilde} \def\Ric{\text{\rm{Ric}}\sigma} \def\ess{\text{\rm{ess}}_t(\gg)$ is Lipschitz continuous in $\gg\in\scr P_{2}(\R^d)$ as well. Moreover, for two independent $d$-dimensional Brownian motions $W_t$ and $\tilde} \def\Ric{\text{\rm{Ric}} W_t$, $$\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D B_t:= \sigma} \def\ess{\text{\rm{ess}}_t(\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{X_t})^{-1} \big\{\ll \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D W_t+ \tilde} \def\Ric{\text{\rm{Ric}} \sigma} \def\ess{\text{\rm{ess}}_t(\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{X_t})\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D \tilde} \def\Ric{\text{\rm{Ric}} W_t\big\}$$ is a $d$-dimensional Brownian motion, so that \eqref{EMV} is reduced to \beq\label{E'} \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D X_t= b_t(X_t, \scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{X_t})\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D t+ \ll \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D W_t+ \tilde} \def\Ric{\text{\rm{Ric}} \sigma} \def\ess{\text{\rm{ess}}_t( \scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{X_t})\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D \tilde} \def\Ric{\text{\rm{Ric}} W_t,\ \ t\in [0,T]. \end{equation} Thus, by the well-posedness, \eqref{EMV} and \eqref{E'} provide the same operator $P_t$. Now, consider the conditional probability $\P^{\tilde} \def\Ric{\text{\rm{Ric}} W}$ given $\tilde} \def\Ric{\text{\rm{Ric}} W$, under which $\int_0^t \tilde} \def\Ric{\text{\rm{Ric}} \sigma} \def\ess{\text{\rm{ess}}_s( \scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{X_s})\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D \tilde} \def\Ric{\text{\rm{Ric}} W_s$ is deterministic so that \eqref{E'} becomes an SDE with constant noise $\ll \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D W_t$, and hence its log-Harnack inequality follows from exiting arguments developed for distribution free noise. However, this noise decomposition argument is hard to extend to spatial-distribution dependent noise. So, in the following we only consider \eqref{EMV} or \eqref{E'}, rather than \eqref{E1}. Closely related to the log-Harnack inequality, a very nice entropy estimate has been derived in \cite{BRS} for two SDEs with different noise coefficients. Consider, for instance, the following SDEs on $\R^d$ for $i=1,2$: $$\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D X_t^i= b_i(t,X_t^i)\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D t+ \ss{a_i(t)}\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D B_t,\ \ X_0^i=x\in\R^d, t\ge 0, $$ where $a_i(t)$ is positive definite, and for some constant $K>1$, $$|b_i(t,x)-b_i(t,y)|\le K|x-y|,\ \ \ K^{-1} I_d\le a_i(t)\le K I_d,\ \ x,y\in\R^d,t\ge 0.$$ Then \cite[Theorem 1.1]{BRS} gives the entropy estimate \begin} \def\beq{\begin{equation}} \def\F{\scr F{align*}&\Ent(\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{X_t^2}|\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{X_t^1})\le \ff 1 2 \int_0^t \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D s\int_{\R^d} \big|a_1(s)^{-\ff 1 2 } \Phi(s,y)\big|^2\rr_2(s,y) \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D y,\\ &\Phi(s,y):= (a_1(s)-a_2(s)) \nabla} \def\pp{\partial} \def\E{\mathbb E\log \rr_2(s,y) +b_2(s,y)-b_1(s,y),\ \ s> 0, y\in\R^d,\end{align*} where $\rr_2(s,y):=\ff{ \scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{X_s^2}(\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D y)}{\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D y}$ is the distribution density function of $X_s^2$. Since for elliptic diffusion processes $$\int_{\R^d} \big|\nabla} \def\pp{\partial} \def\E{\mathbb E\log \rr_2(s,y) \big|^2\rr_2(s,y) \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D y $$ behaves like $\ff c s$ for some constant $c>0$ and small $s>0$, to derive finite entropy upper bound from this estimate one my assume \beq\label{TY} \int_0^1 \ff{\|a_1(s)-a_2(s)\|^2}s\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D s <\infty,\end{equation} where $\|\cdot\|$ is the operator norm of matrices. To bound $\Ent(P_t^*\nu|P_t^*\mu)$ for \eqref{EMV}, we take $$a_1(s):= (\sigma} \def\ess{\text{\rm{ess}}_s\sigma} \def\ess{\text{\rm{ess}}_s^*)(P_s^*\mu),\ \ a_2(s):= (\sigma} \def\ess{\text{\rm{ess}}_s\sigma} \def\ess{\text{\rm{ess}}_s^*)(P_s^*\nu).$$ But \eqref{TY} fails when $\|(\sigma} \def\ess{\text{\rm{ess}}_s\sigma} \def\ess{\text{\rm{ess}}_s^*)(P_s^*\mu)-(\sigma} \def\ess{\text{\rm{ess}}_s\sigma} \def\ess{\text{\rm{ess}}_s^*)(P_s^*\nu)\|$ is uniformly positive for small $s.$ \ The remainder of the paper is organized as follows. In Section 2 and Section 3, we establish the log-Harnack inequality and Bismut formula for the non-degenerate case and degenerate cases respectively. In Section 4 we apply the log-Harnack inequality to study the exponential ergodicity in entropy. \section{ Non-degenerate case} In this part, we establish the log-Harnack inequality and Bismut formula for $P_tf$ defined in \eqref{PT}, where $P_t^*\mu:=\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{X_t^\mu}$ for $X_t^\mu$ solving \eqref{E'} with initial distribution $\mu$. \subsection{Log-Harnack inequality} To establish the log-Harnack inequality, we make the following assumption. \begin{enumerate} \item[{\bf (A)}] $\ll >0$ is a constant, and there exists $0\le K\in L^1([0,T])$ such that \begin{align*} &|b_t(x,\mu)-b_t(y,\nu)|^2+ \|\tilde} \def\Ric{\text{\rm{Ric}}\sigma} \def\ess{\text{\rm{ess}}_t (\mu)-\tilde} \def\Ric{\text{\rm{Ric}}\sigma} \def\ess{\text{\rm{ess}}_t (\nu)\|^2 \le K_t (|x-y|^2+\mathbb W_2(\mu, \nu)^2),\\ &|b_t(0,\delta_0)|+\|\tilde} \def\Ric{\text{\rm{Ric}}\sigma_t(\delta_0)\|^2\leq K_t,\ \ \ t\in [0,T],\ x,y\in\R^d,\ \mu,\nu\in \scr P_2(\R^d). \end{align*} \end{enumerate} By \cite[Theorems 2.1 and 3.3]{HRW} or \cite[Theorem 2.1]{FYW1}, assumption {\bf (A)} implies that the SDE \eqref{E'} is well-posed for distributions in $\scr P_2(\R^d),$ and there exists a constant $c>0$ such that \begin{align}\label{WDS}\mathbb W_2(P_t^\ast \nu,P_t^\ast \mu)\leq c\mathbb W_2(\nu, \mu),\ \ \mu,\nu\in\scr P_2(\R^d), t\in[0,T]. \end{align} \begin{thm}\label{Loh} Assume {\bf (A)} and let $P_t$ be defined in $\eqref{PT}$ for the SDE $\eqref{E'}$. Then there exists a constant $c>0$ such that \begin{equation*} P_t\log f(\nu) \le \log P_t f(\mu) +\ff c t \mathbb W_2(\mu,\nu)^2,\ \ 0<f\in \B_b(\R^d), \mu,\nu \in \scr P_2(\R^d), t\in (0,T]. \end{equation*} Equivalently, $$\Ent(P_t^*\nu|P_t^*\mu) \leq \frac{c}{t}\mathbb W_2(\mu,\nu)^2,\ \ \mu,\nu \in \scr P_2(\R^d), t\in (0,T].$$ \end{thm} \begin{proof} As explained in Introduction, we will use coupling by change of measure under the conditional expectation given $\tilde} \def\Ric{\text{\rm{Ric}} W$, which will be enough for the proof of the log-Harnack inequality. But for the study of Bismut formula later on, we will use the conditional probability and the conditional expectation given both $\tilde} \def\Ric{\text{\rm{Ric}} W$ and $\F_0$: $$\P^{\tilde} \def\Ric{\text{\rm{Ric}} W,0}:= \P(\ \cdot\ |\tilde} \def\Ric{\text{\rm{Ric}} W,\F_0),\ \ \E^{\tilde} \def\Ric{\text{\rm{Ric}} W,0}:= \E(\ \cdot\ | \tilde} \def\Ric{\text{\rm{Ric}} W,\F_0).$$ (a) For any $t\in [0,T], \mu\in \scr P_2(\R^d)$ and $ f\in\B_b(\R^d),$ let $$ P_t^{\tilde} \def\Ric{\text{\rm{Ric}} W,0} f(X_0^\mu):= \E^{\tilde} \def\Ric{\text{\rm{Ric}} W,0} [f(X_t^\mu)]= \E\big[f(X_t^\mu)\big|\tilde} \def\Ric{\text{\rm{Ric}} W,\F_0\big],$$ where $X_t^\mu$ solves \eqref{E'} with $\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{X_0^\mu}=\mu$. By \eqref{PT}, \begin{equation}\label{PTW} P_tf(\mu)= \E \big[P_t^{\tilde} \def\Ric{\text{\rm{Ric}} W,0} f(X_0^\mu)\big], \ \ \ t\in [0,T], \mu\in \scr P_2(\R^d), f\in \B_b(\R^d).\end{equation} Next, let \begin{equation}\label{XM} \xi_t^\mu:= \int_0^t \tilde} \def\Ric{\text{\rm{Ric}} \sigma} \def\ess{\text{\rm{ess}}_s(P_s^*\mu)\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D \tilde} \def\Ric{\text{\rm{Ric}} W_s,\ \ t\in [0,T], \mu\in \scr P_2(\R^d).\end{equation} By {\bf (A)}, BDG's inequality and \eqref{WDS}, we find constants $C_1,C_2>0$ such that \beq\label{-2} \E\Big[\sup_{t\in [0,T]} |\xi_{t}^\mu-\xi_{t}^\nu|^2\Big]\le C_1 \mathbb W_2(\mu,\nu)^2\int_0^TK_s\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D s\le C_2 \mathbb W_2(\mu,\nu)^2,\ \ \mu,\nu\in \scr P_2(\R^d). \end{equation} (b) For fixed $\mu,\nu\in \scr P_2(\R^d)$, we take $\F_0$-measurable $X_0^\mu$ and $X_0^\nu$ such that \beq\label{0} \scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{X_0^\mu}=\mu,\ \ \scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{X_0^\nu}=\nu,\ \ \E[|X_0^\mu-X_0^\nu|^2] =\mathbb W_2(\mu,\nu)^2.\end{equation} Since $X_t^\mu$ solves \eqref{E'} with $\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{X_0^\mu}=\mu$, we have $\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{X_t^\mu}=P_t^*\mu$ and the SDE becomes \beq\label{1} \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D X_t^\mu = b_t(X_t^\mu, P_t^*\mu)\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D t + \ll \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D W_t +\tilde} \def\Ric{\text{\rm{Ric}}\sigma} \def\ess{\text{\rm{ess}}_t(P_t^*\mu)\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D\tilde} \def\Ric{\text{\rm{Ric}} W_t,\ \ t\in [0,T].\end{equation} For fixed $t_0\in (0,T]$, consider the following SDE: \beq\label{2} \begin} \def\beq{\begin{equation}} \def\F{\scr F{split}&\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D Y_t = \Big\{b_t(X_t^\mu, P_t^*\mu)+\ff 1 {t_0}\big[\xi_{t_0}^\mu-\xi_{t_0}^\nu+X_0^\mu-X_0^\nu\big]\Big\} \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D t + \ll \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D W_t +\tilde} \def\Ric{\text{\rm{Ric}}\sigma} \def\ess{\text{\rm{ess}}_t(P_t^*\nu)\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D\tilde} \def\Ric{\text{\rm{Ric}} W_t,\\ &\quad \ t\in [0,t_0],\ Y_0=X_0^\nu.\end{split} \end{equation} By \eqref{XM}, \eqref{1} and \eqref{2}, we obtain \beq\label{3} Y_t-X_t^\mu = \ff{t_0-t}{t_0} (X_0^\nu-X_0^\mu) +\ff t{t_0} \big(\xi_{t_0}^\mu-\xi_{t_0}^\nu\big) +\xi_t^\nu-\xi_t^\mu,\ \ t\in [0,t_0].\end{equation} To formulate $P_{t_0} f(\nu)$ using $Y_{t_0}$, we make Girsanov's transform as follows. Let \beq\label{ETA} \eta_t:=b_t(Y_t,P_t^*\nu)-b_t(X_t^\mu,P_t^*\mu) + \ff 1 {t_0} \big[\xi_{t_0}^\nu-\xi_{t_0}^\mu+X_0^\nu-X_0^\mu\big],\ \ t\in [0,t_0].\end{equation} By {\bf (A)} and \eqref{WDS}, we find a constant $c_1>0$ such that \begin{equation*} \begin} \def\beq{\begin{equation}} \def\F{\scr F{split} |\eta_t|^2\le &\,c_1K_t\big( \mathbb W_2(\mu,\nu)^2 + |\xi_t^\nu-\xi_t^\mu|^2\big)\\ & + c_1\Big(\ff {t^2K_t+1}{t_0^2} |\xi_{t_0}^\mu-\xi_{t_0}^\nu|^2 +\ff 1 {t_0^2} |X_0^\mu-X_0^\nu|^2 \Big),\ \ t\in [0,t_0].\end{split}\end{equation*} Since $\int_0^TK_t\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D t<\infty$, we find a constant $c_2>0$ uniform in $t_0\in (0,T]$, such that \beq\label{*N} \ff 1 {2\ll^2} \int_0^{t_0} |\eta_t|^2\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D t \le c_2\mathbb W_2(\mu,\nu)^2+\ff{c_2}{t_0}\Big(|X_0^\mu-X_0^\nu|^2+ \sup_{t\in [0,t_0]} |\xi_t^\mu-\xi_t^\nu|^2\Big).\end{equation} Let $\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D\Q^{\tilde} \def\Ric{\text{\rm{Ric}} W,0}:= R^{\tilde} \def\Ric{\text{\rm{Ric}} W,0}\,\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D\P^{\tilde} \def\Ric{\text{\rm{Ric}} W,0}$, where \beq\label{RW} R^{\tilde} \def\Ric{\text{\rm{Ric}} W,0}:= \text{\rm{e}}} \def\ua{\underline a} \def\OO{\Omega} \def\oo{\omega^{\int_0^{t_0}\<\ff 1 \ll \eta_s, \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D W _s\> -\ff 1 2 \int_0^{t_0} |\ff 1 \ll \eta_s|^2\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D s}.\end{equation} By Girsanov's theorem, under the weighted conditional probability $\Q^{\tilde} \def\Ric{\text{\rm{Ric}} W,0},$ $$ \hat{W}_t := W_t-\int_0^t \ff 1 \ll \eta_s\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D s,\ \ t\in [0,t_0]$$ is a $d$-dimensional Brownian motion. By \eqref{2}, $\hat Y_t:= Y_t-\xi_t^\nu$ solves the SDE $$\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D \hat Y_t = b_t(\hat Y_t+\xi_t^\nu, P_t^*\nu)\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D t + \ll \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D \hat{W}_t,\ \ t\in [0,t_0], \hat Y_0=X_0^\nu.$$ On the other hand, let $X_t^\nu$ solve \eqref{E'} with initial value $X_0^\nu$. Then $$\hat X_t^\nu:= X_t^\nu-\xi_t^\nu,\ \ t\in [0,t_0]$$ solves the same SDE as $\hat Y_t$ for $W$ replacing $\hat W$. Then the weak uniqueness of this equation ensured by {\bf (A)} implies $$\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{\hat Y_{t_0} |\Q^{\tilde} \def\Ric{\text{\rm{Ric}} W,0}} =\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{\hat X_{t_0}^\nu|\P^{\tilde} \def\Ric{\text{\rm{Ric}} W,0}},$$ where $\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{\hat Y_{t_0} |\Q^{\tilde} \def\Ric{\text{\rm{Ric}} W,0}} $ is the law of $\hat Y_{t_0}$ under $\Q^{\tilde} \def\Ric{\text{\rm{Ric}} W,0}$, while $\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{\hat X_{t_0}^\nu|\P^{\tilde} \def\Ric{\text{\rm{Ric}} W,0}}$ is the law of $\hat X_{t_0}^\nu$ under $\P^{\tilde} \def\Ric{\text{\rm{Ric}} W,0}$. Since $\xi_{t_0}^\nu$ is deterministic given $\tilde} \def\Ric{\text{\rm{Ric}} W$, it follows that $$\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{Y_{t_0}|\Q^{\tilde} \def\Ric{\text{\rm{Ric}} W,0}}= \scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{\hat Y_{t_0} +\xi_{t_0}^\nu|\Q^{\tilde} \def\Ric{\text{\rm{Ric}} W,0}} = \scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{\hat X_{t_0}^\nu +\xi_{t_0}^\nu|\P^{\tilde} \def\Ric{\text{\rm{Ric}} W,0}} = \scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{X_{t_0}^\nu |\P^{\tilde} \def\Ric{\text{\rm{Ric}} W,0}}.$$ Combining this with $X_{t_0}^\mu=Y_{t_0}$ due to \eqref{3}, we obtain \beq\label{NB} P_{t_0}^{\tilde} \def\Ric{\text{\rm{Ric}} W,0} f(X_0^\nu):= \E^{\tilde} \def\Ric{\text{\rm{Ric}} W,0}[ f(X_{t_0}^\nu)] = \E_{\Q^{\tilde} \def\Ric{\text{\rm{Ric}} W,0}} [ f(Y_{t_0}) ] = \E^{\tilde} \def\Ric{\text{\rm{Ric}} W,0}[ R^{\tilde} \def\Ric{\text{\rm{Ric}} W,0} f(X_{t_0}^\mu) ],\ \ f\in\B_b(\R^d).\end{equation} By Young's inequality \cite[Lemma 2.4]{ATW}, we derive \begin} \def\beq{\begin{equation}} \def\F{\scr F{align*} &P_{t_0}^{\tilde} \def\Ric{\text{\rm{Ric}} W,0}\log f(X_0^\nu):= \E^{\tilde} \def\Ric{\text{\rm{Ric}} W,0}[ \log f(X_{t_0}^\nu)] = \E_{\Q^{\tilde} \def\Ric{\text{\rm{Ric}} W,0}} [\log f(Y_{t_0}) ]\\ &= \E^{\tilde} \def\Ric{\text{\rm{Ric}} W,0}[ R^{\tilde} \def\Ric{\text{\rm{Ric}} W,0} \log f(X_{t_0}^\mu) ] \le \log \E^{\tilde} \def\Ric{\text{\rm{Ric}} W,0} [ f(X_{t_0}^\mu)] + \E^{\tilde} \def\Ric{\text{\rm{Ric}} W,0} [R^{\tilde} \def\Ric{\text{\rm{Ric}} W,0}\log R^{\tilde} \def\Ric{\text{\rm{Ric}} W,0}] \\ & = \log P_{t_0}^{\tilde} \def\Ric{\text{\rm{Ric}} W,0} f(X_{0}^\mu) + \ff 1 2 \int_0^{t_0}\ff 1 {\ll^2}\E_{\Q^{\tilde} \def\Ric{\text{\rm{Ric}} W,0}} [| \eta_t|^2]\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D t,\ \ 0< f\in \B_b(\R^d).\end{align*} This together with \eqref{*N} gives \beq\label{*W0} P_{t_0}^{\tilde} \def\Ric{\text{\rm{Ric}} W,0}\log f(X_0^\nu)\le \log P_{t_0}^{\tilde} \def\Ric{\text{\rm{Ric}} W,0} f(X_{0}^\mu) + c_2\mathbb W_2(\mu,\nu)^2+\ff{c_2}{t_0}\Big(|X_0^\mu-X_0^\nu|^2+ \sup_{t\in [0,t_0]} |\xi_t^\mu-\xi_t^\nu|^2\Big). \end{equation} Taking expectation for both sides, by \eqref{PTW}, \eqref{-2}, \eqref{0} and Jensen's inequality, we find a constant $c>0$ such that \begin} \def\beq{\begin{equation}} \def\F{\scr F{align*} P_{t_0}\log f(\nu)&= \E\big[P_{t_0}^{\tilde} \def\Ric{\text{\rm{Ric}} W,0}\log f(X_{0}^\nu)\big] \le \E\big[ \log P_{t_0}^{\tilde} \def\Ric{\text{\rm{Ric}} W,0} f(X_{0}^\mu)\big]+ \ff{c}{t_0} \mathbb W_2(\mu,\nu)^2\\ &\le \log P_{t_0} f(\mu) + \ff{c}{t_0} \mathbb W_2(\mu,\nu)^2,\ \ t_0\in (0,T], \mu,\nu\in \scr P_2(\R^d).\end{align*} \end{proof} \subsection{Bismut formula} We aim to establish the Bismut type formula \eqref{BS} for the intrinsic derivative of $P_tf$. To this end, we first recall the definition of intrinsic derivative, see \cite{RW0} for historical remarks on this derivative and links to other derivatives for functions of measures. \begin{defn} Let $k\in (1,\infty)$. \begin} \def\beq{\begin{equation}} \def\F{\scr F{enumerate} \item[$(1)$] A continuous function $f$ on $\scr P_k(\R^d)$ is called intrinsically differentiable, if for any $\mu\in \scr P_k(\R^d)$, $$T_{\mu,k}(\R^d):=L^k(\R^d\rightarrow}\def\l{\ell\R^d;\mu)\ni\phi\mapsto D_\phi^If(\mu):= \lim_{\vv\downarrow 0} \ff{f(\mu\circ(id+\varepsilon\phi)^{-1})-f(\mu)}{\vv}\in\mathbb{R} $$ is a well defined bounded linear operator. In this case, the norm of the intrinsic derivative $D^I f(\mu)$ is given by $$\|D^If(\mu)\|_{L^{k^*}(\mu)} :=\sup_{\|\phi\|_{L^k(\mu)}\le 1} |D^I_\phi f(\mu)|.$$ \item [$(2)$] $f$ is called $L$-differentiable on $\scr P_k(\R^d)$, if it is intrinsically differentiable and \begin{equation*} \lim_{\|\phi\|_{T_{\mu,k}(\R^d)}\downarrow0}\ff{|f(\mu\circ(id+\phi)^{-1})-f(\mu)-D_\phi^If(\mu)| }{\|\phi\|_{T_{\mu,k}(\R^d)}}=0,\ \ \mu\in \scr P_k(\R^d). \end{equation*} We denote $f\in C^1(\scr P_k(\R^d))$, if it is $L$-differentiable such that $D^If(\mu)(x)$ has a jointly continuous version in $(x,\mu)\in \R^d\times \scr P_k(\R^d)$. \item[$(3)$] We denote $g\in C^{1,1}(\R^d\times \scr P_k(\R^d))$, if $g: \R^d\times\scr P_k(\R^d)\rightarrow}\def\l{\ell\R$ is $C^1$ in $x\in \R^d$ and $\mu\in \scr P_k(\R^d)$ respectively, such that $$\nabla} \def\pp{\partial} \def\E{\mathbb E g(x,\mu):= \nabla} \def\pp{\partial} \def\E{\mathbb E \{g(\cdot,\mu)\}(x),\ \ \ D^Ig(x,\mu)(y):= D^I\{g(x,\cdot)\}(\mu)(y)$$ are jointly continuous in $(x,y,\mu)\in \R^d\times\R^d\times\scr P_k(\R^d)$. \end{enumerate} \end{defn} In this part, we consider \eqref{E'} with coefficients $$\tilde{\sigma} \def\ess{\text{\rm{ess}}}: [0,T]\times \scr P_k(\R^d)\rightarrow}\def\l{\ell \R^d\otimes\R^d,\ \ b: [0,T]\times \R^d\times \scr P_k(\R^d)\rightarrow}\def\l{\ell\R^d$$ satisfying the following assumption. \begin} \def\beq{\begin{equation}} \def\F{\scr F{enumerate} \item[{\bf (B)}] $\ll>0$ and $k\in (1,\infty)$ are constants, denote $k^*:=\ff{k}{k-1}$. For any $t\in [0,T]$, $b_t\in C^{1,1}(\R^d\times\scr P_k(\R^d)), $ $\tilde} \def\Ric{\text{\rm{Ric}}\sigma_t\in C^1(\scr P_k(\R^d))$, and there exists $0\le K\in L^1([0,T])$ such that \begin} \def\beq{\begin{equation}} \def\F{\scr F{align*} & |D^I b_t(x,\cdot)(\mu)(y)|+\|D^I \tilde} \def\Ric{\text{\rm{Ric}}\sigma_t (\mu)(y) \| \le \ss{K_t} (1+|y|^{k-1}) ,\\ & |b_t(0,\delta_0)|+ |\nabla} \def\pp{\partial} \def\E{\mathbb E b_t(\cdot,\mu)(x)| \le \ss {K_t},\ \ (t,x,\mu)\in [0,T]\times \R^d\times\scr P_k(\R^d),\ y\in\R^d.\end{align*} \end{enumerate} By \cite[Lemma 3.1]{FYW3}, {\bf (B)} implies {\bf (A)} for $(\scr P_k(\R^d),\mathbb W_k)$ replacing $(\scr P_2(\R^d),\mathbb W_2).$ So, according to \cite[Theorem 3.3]{HRW}, the SDE \eqref{E'} is well-posed for distributions in $\scr P_k(\R^d)$, and there exists a constant $c>0$ such that \begin{align}\label{WBS}\mathbb W_k(P_t^\ast \mu,P_t^\ast \nu)\leq c \mathbb W_k(\mu,\nu),\ \ \mu,\nu\in\scr P_k(\R^d), t\in[0,T]. \end{align} By this estimate and {\bf (A)} for $(\scr P_k(\R^d),\mathbb W_k)$ replacing $(\scr P_2(\R^d),\mathbb W_2),$ the argument leading to \eqref{*W0} yields that there exists a constant $c>0$ such that for any $t\in(0,T], 0<f\in\scr B_b(\R^d)$, \beq\label{*W}P_{t}^{\tilde} \def\Ric{\text{\rm{Ric}} W,0}\log f(X_0^\nu)\le \log P_{t}^{\tilde} \def\Ric{\text{\rm{Ric}} W,0} f(X_{0}^\mu)+c \mathbb W_k(\mu,\nu)^2+\ff{c }{t}\Big(|X_0^\mu-X_0^\nu|^2+ \sup_{s\in [0,t]} |\xi_s^\mu-\xi_s^\nu|^2\Big).\end{equation} To calculate $D_\phi^IP_tf(\mu)$ for $\mu\in \scr P_k(\R^d)$ and $\phi\in T_{\mu,k}(\R^d)$, let $X_0^\mu$ be $\F_0$-measurable such that $\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{X_0^\mu}=\mu.$ Then $$\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{X_0^{\mu}+\vv\phi(X_0^\mu)}=\mu^\vv:= \mu\circ (id+\vv\phi)^{-1},\ \ \ \vv\in [0,1].$$ For any $\vv\in [0,1],$ let $X_t^{\mu^\vv}$ solve \eqref{E'} with $ X_0^{\mu^\vv}=X_0^{\mu}+\vv\phi(X_0^\mu),$ i.e. \begin{equation*}\begin} \def\beq{\begin{equation}} \def\F{\scr F{split}& \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D X_t^{\mu^\vv} = b_t(X_t^{\mu^\vv}, P_t^*\mu^\vv) \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D t +\ll \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D W_t+ \tilde} \def\Ric{\text{\rm{Ric}}\sigma} \def\ess{\text{\rm{ess}}_t(P_t^*\mu^\vv) \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D \tilde} \def\Ric{\text{\rm{Ric}} W_t,\\ &\ X_0^{\mu^\vv}=X_0^\mu+\vv\phi(X_0^\mu), t\in [0,T],\vv\in [0,1].\end{split} \end{equation*} Consider the spatial derivative of $X_t^{\mu}$ along $\phi$: \begin{equation*} \nabla} \def\pp{\partial} \def\E{\mathbb E_\phi X_t^{\mu}:=\lim_{\vv\downarrow 0} \ff{X_t^{\mu^\vv}-X_t^{\mu}}\vv,\ \ \ \ t\in [0,T], \phi\in T_{\mu,k}(\R^d).\end{equation*} For any $0\le s<t\le T,$ define \begin} \def\beq{\begin{equation}} \def\F{\scr F{align*} &N^{\mu,\phi}_{s,t}:=\ff{t-s} t\phi(X_0^\mu)+\int_0^s\Big\<\E\big[\<D^I\tilde{\sigma} \def\ess{\text{\rm{ess}}}_r(P_r^*\mu)(X_r^\mu), \nabla} \def\pp{\partial} \def\E{\mathbb E_\phi X_r^\mu\>\big],\ \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D\tilde} \def\Ric{\text{\rm{Ric}} W_r\Big\>\\ &\qquad \qquad -\ff s t \int_0^t \Big\<\E\big[\<D^I\tilde{\sigma} \def\ess{\text{\rm{ess}}}_r(P_r^*\mu)(X_r^\mu), \nabla} \def\pp{\partial} \def\E{\mathbb E_\phi X_r^\mu\>\big],\ \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D\tilde} \def\Ric{\text{\rm{Ric}} W_r\Big\>,\\ &M^{\mu,\phi}_{s,t}:=\E\big[\big\< \{D^Ib_s(y, \cdot)\}(P_s^\ast \mu)(X_s^\mu),\nabla_{\phi}X_s^\mu\big\>\big]_{y=X_s^{\mu}}+ \ff 1 {t}\phi(X_0^\mu)\\ &\qquad\qquad +\ff 1 {t} \int_0^t \Big\<\E\big[\<D^I\tilde{\sigma} \def\ess{\text{\rm{ess}}}_r(P_r^*\mu)(X_r^\mu), \nabla} \def\pp{\partial} \def\E{\mathbb E_\phi X_r^\mu\>\big],\ \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D\tilde} \def\Ric{\text{\rm{Ric}} W_r\Big\>.\end{align*} The main result in this part is the following. \begin} \def\beq{\begin{equation}} \def\F{\scr F{thm}\label{TA2'} Assume {\bf (B)}. \begin} \def\beq{\begin{equation}} \def\F{\scr F{enumerate} \item[$(1)$] For any $\mu\in \scr P_k(\R^d)$ and $\phi\in T_{\mu,k}(\R^d)$, $(\nabla} \def\pp{\partial} \def\E{\mathbb E_\phi X_\cdot^{\mu})$ exists in $L^k(\Omega\rightarrow}\def\l{\ell C([0,T],\R^d),\P)$ such that for some constant $c>0$, $$\E\Big[\sup_{t\in[0,T]}|\nabla_{\phi}X_t^\mu|^k\Big]\leq c\|\phi\|_{L^k(\mu)}^k,\ \ \mu\in \scr P_k(\R^d), \phi\in T_{\mu,k}(\R^d).$$ \item[$(2)$] For any $f\in \B_b(\R^d), t\in (0,T]$, $\mu\in \scr P_k(\R^d)$ and $\phi\in T_{\mu,k}(\R^d)$, $D^I_\phi P_tf(\mu)$ exists and satisfies \beq\label{BSMI} D_\phi^I P_tf(\mu)=\ff 1 \ll \E\bigg[f(X_t^{\mu})\int_0^t \Big\< \nabla} \def\pp{\partial} \def\E{\mathbb E_{N_{s,t}^{\mu,\phi}}b_s(\cdot, P_s^*\mu)(X_s^\mu) +M_{s,t}^{\mu,\phi},\ \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D W_s\Big\>\bigg]. \end{equation} Consequently, $P_tf$ is intrinsically differentiable and for some constant $c>0$, \beq\label{EST}\begin} \def\beq{\begin{equation}} \def\F{\scr F{split}& \|D^I P_t f(\mu)\|_{L^{k^*}(\mu)} \le \ff c {\ss t}\big( P_t |f|^{k^*}(\mu)\big)^{\ff 1 {k^*}},\\ &\qquad\quad f\in \B_b(\R^d), \mu\in \scr P_k(\R^d), t\in (0,T].\end{split}\end{equation}\end{enumerate} \end{thm} \begin} \def\beq{\begin{equation}} \def\F{\scr F{proof} The first assertion follows from \cite[Lemma 5.2]{BRW}. By the first assertion, {\bf (B)} and the definition of $(N_{s,t}^{\mu,\phi}, M_{s,t}^{\mu,\phi})$, we deduce \eqref{EST} from \eqref{BSMI}. So, it remains to prove \eqref{BSMI}. (a) Since {\bf (B)} implies {\bf (A)} for $\scr P_k(\R^d)$ replacing $\scr P_2(\R^d)$, the argument in the proof of Theorem \ref{Loh} up to \eqref{NB} still applies. For fixed $t_0\in (0,T], \mu\in \scr P_k(\R^d)$ and $\phi\in T_{\mu,k}(\R^d)$, let $X_t^\mu$ solve \eqref{1}. Next, for any $\vv\in (0,1]$, let $ Y_t^\vv$ solve \eqref{2} for $$\nu=\mu^\vv,\ \ Y_0=Y_0^\vv:= X_0^\mu+\vv\phi(X_0^\mu).$$ Then \eqref{3} with $(Y_t,\nu)=(Y_t^\vv,\mu^\vv)$ becomes \beq\label{YX} Y_t^\vv-X_t^\mu= \ff{t_0-t}{t_0}\vv\phi(X_0^\mu) +\ff t{t_0} (\xi_{t_0}^\mu-\xi_{t_0}^{\mu^\vv}) +\xi_t^{\mu^\vv}-\xi_t^\mu,\ \ t\in [0,t_0].\end{equation} Let $$H_t:= \int_0^t \Big\<\E\big[\big\<D^I\tilde{\sigma} \def\ess{\text{\rm{ess}}}_s(P_s^*\mu)(X_s^\mu), \nabla} \def\pp{\partial} \def\E{\mathbb E_\phi X_s^\mu\big\>\big],\ \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D\tilde} \def\Ric{\text{\rm{Ric}} W_s\big\>,\ \ t\in [0,T].$$ By {\bf (B)} and \eqref{WBS}, we obtain \beq\label{a1} \big\|\tilde} \def\Ric{\text{\rm{Ric}}\sigma} \def\ess{\text{\rm{ess}}_s(P_s^*\mu^\vv)-\tilde} \def\Ric{\text{\rm{Ric}}\sigma} \def\ess{\text{\rm{ess}}_s(P_s^*\mu)\big\|^2\le \vv^2 c^2 K_s \|\phi\|_{L^k(\mu)}^2,\ \ \vv\in [0,1], s\in [0,T].\end{equation} So, by {\bf (B)}, the chain rule in \cite[Theorem 2.1(1)]{BRW}, \eqref{XM}, BDG's inequality and the dominated convergence theorem, we obtain \beq\label{a2} \lim_{\vv\downarrow 0} \E\bigg[\sup_{t\in [0,T]} \Big|\ff{\xi_t^{\mu^\vv}-\xi_t^\mu}\vv -H_t\Big|^2\bigg]=0.\end{equation} Let $(\eta_t^\vv, R^\vv)=(\eta_t, R^{\tilde} \def\Ric{\text{\rm{Ric}} W,0})$ be defined in \eqref{ETA} and \eqref{RW} for $(Y_t,\nu)=(Y_t^\vv,\mu^\vv)$. By {\bf (B)} and \eqref{YX}, we find a constant $\kk>0$ such that \begin} \def\beq{\begin{equation}} \def\F{\scr F{align*}& \ff{|\eta_s^\vv|^2}{\vv^2} \le\kk K_s \bigg( \|\phi\|_{L^k(\mu)}^2 +|\phi(X_0^\mu)|^2+\sup_{t\in [0,t_0]} \ff{|\xi_t^{\mu^\vv}-\xi_t^\mu|^2}{\vv^2}\bigg)=:\LL_s,\\ &\lim_{\vv\downarrow 0} \ff{\eta_s^\vv}\vv = \nabla} \def\pp{\partial} \def\E{\mathbb E_{N_{s,t_0}^{\mu,\phi}}b_s(\cdot, P_s^*\mu)(X_s^\mu)+M_{s,t_0}^{\mu,\phi},\ \ s\in [0,t_0].\end{align*} Since $\LL_s$ is deterministic given $\tilde} \def\Ric{\text{\rm{Ric}} W$ and $\F_0$, this together with \eqref{NB} and the dominated convergence theorem yields \beq\label{*W2}\begin} \def\beq{\begin{equation}} \def\F{\scr F{split}& \lim_{\vv\downarrow 0} \ff{P_{t_0}^{\tilde} \def\Ric{\text{\rm{Ric}} W,0} f(X_0^{\mu^\vv})-P_{t_0}^{\tilde} \def\Ric{\text{\rm{Ric}} W,0} f(X_0^\mu)}\vv =\lim_{\vv\downarrow 0} \E^{\tilde} \def\Ric{\text{\rm{Ric}} W,0}\Big[f(X_{t_0}^\mu)\ff{R^\vv-1}\vv\Big]\\ &= \frac{1}{\lambda}\E^{\tilde} \def\Ric{\text{\rm{Ric}} W,0} \bigg[f(X_{t_0}^\mu)\int_0^{t_0} \Big\< \nabla} \def\pp{\partial} \def\E{\mathbb E_{N_{s,t_0}^{\mu,\phi}}b_s(\cdot, P_s^*\mu)(X_s^\mu) + M_{s,t_0}^{\mu,\phi},\ \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D W_s\Big\>\bigg].\end{split}\end{equation} (b) Let $\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{\xi|\P^{\tilde} \def\Ric{\text{\rm{Ric}} W,0}}$ be the conditional distribution of a random variable $\xi$ under $\P^{\tilde} \def\Ric{\text{\rm{Ric}} W,0}$. By Pinsker's inequality and \eqref{*W}, we have \begin} \def\beq{\begin{equation}} \def\F{\scr F{align*}& \sup_{\|f\|_\infty\le 1} \big|P_{t_0}^{\tilde} \def\Ric{\text{\rm{Ric}} W,0} f(X_0^{\mu^\vv})-P_{t_0}^{\tilde} \def\Ric{\text{\rm{Ric}} W,0} f(X_0^\mu)\big|^2 \le 2\, \Ent\big(\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{X_{t_0}^{\mu^\vv}|\P^{\tilde} \def\Ric{\text{\rm{Ric}} W,0}}\big|\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{X_{t_0}^\mu|\P^{\tilde} \def\Ric{\text{\rm{Ric}} W,0}}\big)\\ &\le c \mathbb W_k(\mu^\vv,\mu)^2+ \ff{c}{t_0}\Big(\vv^2|\phi(X_0^\mu)|^2 + \sup_{t\in [0,t_0]} |\xi_t^\mu-\xi_t^{\mu^\vv}|^2\Big).\end{align*} This together with $\mathbb W_k(\mu^\vv,\mu)\le \vv \|\phi\|_{L^k(\mu)}$ implies that for some constant $c(t_0)>0$, \begin} \def\beq{\begin{equation}} \def\F{\scr F{align*} &\ff{|P_{t_0}^{\tilde} \def\Ric{\text{\rm{Ric}} W,0} f(X_0^{\mu^\vv})-P_{t_0}^{\tilde} \def\Ric{\text{\rm{Ric}} W,0} f(X_0^\mu)|}{\vv} \\ &\le \|f\|_\infty c(t_0)\Big(\|\phi\|_{L^k(\mu)}+ |\phi(X_0^\mu)| + \sup_{t\in [0,t_0]}\ff{ |\xi_t^{\mu^\vv}-\xi_t^\mu|}\vv\Big), \ \ \vv\in (0,1].\end{align*} Combining this with \eqref{-2} and \eqref{a1}, we may apply the dominated convergence theorem to \eqref{*W2} to derive \begin} \def\beq{\begin{equation}} \def\F{\scr F{align*} &D^I_\phi P_{t_0} f(\mu):=\lim_{\vv\downarrow 0} \E\bigg[\ff{P_{t_0}^{\tilde} \def\Ric{\text{\rm{Ric}} W,0} f(X_0^{\mu^\vv})-P_{t_0}^{\tilde} \def\Ric{\text{\rm{Ric}} W,0} f(X_0^\mu)}\vv\bigg] = \E\bigg[\lim_{\vv\downarrow 0} \ff{P_{t_0}^{\tilde} \def\Ric{\text{\rm{Ric}} W,0} f(X_0^{\mu^\vv})-P_{t_0}^{\tilde} \def\Ric{\text{\rm{Ric}} W,0} f(X_0^\mu)}\vv\bigg]\\ &= \frac{1}{\lambda}\E \bigg[f(X_{t_0}^\mu)\int_0^{t_0} \Big\< \nabla} \def\pp{\partial} \def\E{\mathbb E_{N_{s,t_0}^{\mu,\phi}}b_s(\cdot, P_s^*\mu)(X_s^\mu) + M_{s,t_0}^{\mu,\phi},\ \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D W_s\Big\>\bigg].\end{align*} \end{proof} \section{Degenerate case } Consider the following distribution dependent stochastic Hamiltonian system for $X_t=(X_t^{(1)}, X_t^{(2)})\in \mathbb{R}^{m+d}$: \beq\label{E0} \begin{cases} \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D X_t^{(1)}=\big\{AX^{(1)}_t+MX_t^{(2)}\big\}\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D t, \\ \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D X_t^{(2)}=b_t(X_t,\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{X_t})\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D t+\sigma_t(\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{X_t})\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D B_t,\ \ t\in [0,T], \end{cases} \end{equation} where $B=(B_t)_{t\in [0,T]}$ is a $d$-dimensional standard Brownian motion, $A$ is an $m\times m$ and $M$ is an $m\times d$ matrix, and $$\sigma:[0,T]\times \scr P (\R^{m+d})\rightarrow}\def\l{\ell \mathbb{R}^{d}\otimes\R^d,\ \ b:[0,T]\times\R^{m+d}\times \scr P(\R^{m+d})\rightarrow}\def\l{\ell\mathbb{R}^d$$ are measurable, where $\scr P(\R^{m+d})$ is the space of probability measures on $\R^{m+d}$ equipped with the weak topology. For any $k\ge 1$, let $$\scr P_k(\R^{m+d}):=\big\{\mu\in \scr P(\R^{m+d}):\ \|\mu\|_k:=\mu(|\cdot|^k)^{\ff 1 k}<\infty\big\},$$ which is a Polish space under the $L^k$-Wasserstein distance $\mathbb W_k$. When \eqref{E0} is well-posed for distributions in $\scr P_k(\R^{m+d})$, let $P_t^*\mu=\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{X_t}$ for the solution with initial distribution $\mu\in \scr P_k(\R^{m+d})$. We aim to establish the log-Harnack inequality and Bismut formula for $$P_tf(\mu):= \int_{\R^{m+d}}f\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D (P_t^*\mu),\ \ f\in \B_b(\R^{m+d}).$$ By the same reason reformulating \eqref{EMV} as \eqref{E'}, instead of $\eqref{E0}$ we consider \beq\label{E00} \begin{cases} \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D X_t^{(1)}=\big\{AX^{(1)}_t+MX_t^{(2)}\big\}\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D t, \\ \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D X_t^{(2)}=b_t(X_t,\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{X_t})\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D t+\ll \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D W_t+ \tilde} \def\Ric{\text{\rm{Ric}}\sigma_t(\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{X_t})\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D \tilde} \def\Ric{\text{\rm{Ric}} W_t,\ \ t\in [0,T], \end{cases} \end{equation} where $W_t,\tilde} \def\Ric{\text{\rm{Ric}} W_t$ are two independent $d$-dimensional Brownian motions, and $$\tilde} \def\Ric{\text{\rm{Ric}} \sigma} \def\ess{\text{\rm{ess}}: [0,T]\times \scr P(\R^{m+d})\rightarrow}\def\l{\ell \R^d\otimes \R^d$$ are measurable. \subsection{Log-Harnack inequality} To establish the log-Harnack inequality, we make the following assumption. \begin{enumerate} \item[\bf{(C)}] $\ll>0$ is a constant, $(\tilde} \def\Ric{\text{\rm{Ric}} \sigma} \def\ess{\text{\rm{ess}},b)$ satisfies conditions in {\bf (A)} for $(x,\mu)\in \R^{m+d}\times \scr P_2(\R^{m+d})$, and the following Kalman's rank condition holds for some integer $1\leq l\leq m$: \begin{align}\label{RRS}\mathrm{Rank}[A^iM, 0\le i\le l-1]=m,\end{align} where $A^0:=I_m$ is the $m\times m$-identity matrix. \end{enumerate} By \cite[Theorem 2.1]{FYW1}, {\bf (C)} implies that \eqref{E00} is well-posed for distributions in $\scr P_2(\R^{m+d}),$ and there exists a constant $c>0$ such that $$\mathbb W_2(P_t^*\mu,P_t^*\nu)\le c\mathbb W_2(\mu,\nu), \ \ \ \mu,\nu\in \scr P_2(\R^{m+d}), t\in [0,T].$$ So, as in \eqref{-2}, we find a constant $C>0$ such that \beq\label{*N1} \E\bigg[\sup_{t\in [0,T]} |\xi_t^\mu-\xi_t^\nu|^2\bigg] \le C\mathbb W_2(\mu,\nu)^2,\ \ \mu,\nu\in \scr P_2(\R^{m+d}).\end{equation} To distinguish the singularity of $P_t$ in the degenerate component $x^{(1)}$ and the non-degenerate one $x^{(2)}$, for any $t>0$ we consider the modified distance $$\rr_t(x,y):= \ss{t^{-2}|x^{(1)}-y^{(1)}|^2+|x^{(2)}-y^{(2)}|^2},\ \ \ x,y\in\R^{m+d},$$ and define the associated $L^2$-Wasserstein distance $$\mathbb W_{2,t}(\mu,\nu):=\inf_{\pi\in \scr C} \def\aaa{\mathbf{r}} \def\r{r(\mu,\nu)}\bigg( \int_{\R^{m+d}\times \R^{m+d}} \rr_t(x,y)^2 \pi(\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D x,\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D y)\bigg)^{\ff 1 2}. $$ It is clear that \beq\label{W0} \ff 1 {T^2\lor 1} \mathbb W_2^2\le \mathbb W_{2,t}^2 \le \ff{1\lor T^2}{t^2} \mathbb W_2^2,\ \ \ \ t\in (0,T].\end{equation} For $t\in (0,T]$, let $$Q_{t} :=\int_0^{t} \ff{s(t-s)}{t^2}\text{\rm{e}}} \def\ua{\underline a} \def\OO{\Omega} \def\oo{\omega^{-sA}MM^*\text{\rm{e}}} \def\ua{\underline a} \def\OO{\Omega} \def\oo{\omega^{-sA^*}\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D s.$$ According to \cite{S}, see also \cite[Proof of Theorem 4.2(1)]{WZ1}, the rank condition \eqref{RRS} implies \beq\label{QT} \|Q_t^{-1}\|\le c_0 t^{1-2l},\ \ \ t\in (0,T]\end{equation} for some constant $c_0>0$. \begin{thm}\label{LHI} Assume {\bf (C)} and let $P_t^*$ be associated with the degenerate SDE $\eqref{E00}.$ Then there exists a constant $ c>0$ such that \beq\label{EC2} \begin} \def\beq{\begin{equation}} \def\F{\scr F{split}&P_t\log f(\nu) - \log P_tf(\mu) \le \ff{c}{t^{4l-3}}\mathbb W_{2,t}(\mu,\nu)^2 \le \ff{c(1\lor T^2)}{t^{4l-1}}\mathbb W_{2}(\mu,\nu)^2,\\ &t\in (0,T],\ \mu,\nu\in \scr P_2(\R^{m+d}),\ \ 0<f\in \B_b(\R^{m+d}).\end{split}\end{equation} Equivalently, for any $ t\in (0,T]$ and $ \mu,\nu\in \scr P_2(\R^{m+d}),$ $$ \Ent(P_t^*\nu|P_t^*\mu)\le \ff{c}{t^{4l-3}}\mathbb W_{2,t}(\mu,\nu)^2 \le \ff{c(1\lor T^2)}{t^{4l-1}}\mathbb W_{2}(\mu,\nu)^2.$$ \end{thm} \begin{proof} For any $t_0\in (0,T]$ and $\mu,\nu\in \scr P_2(\R^{m+d})$, let $X_0,Y_0$ be $\F_0$-measurable such that \beq\label{0'} \scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{X_0}=\mu,\ \ \scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{Y_0}=\nu,\ \ \E[\rr_{t_0}(X_0,Y_0)^2]=\mathbb W_{2,t_0}(\mu,\nu)^2.\end{equation} Let $X_t $ solve \eqref{E00} with initial value $X_0$, we have $P_t^*\mu= \scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{X_t}.$ Let \beq\label{TTX} v= (v^{(1)}, v^{(2)}):=(Y_0^{(1)}-X_0^{(1)}, Y_0^{(2)}- X_0^{(2)})=Y_0-X_0. \end{equation} For fixed $t_0\in (0,T],$ let \beq\label{A1}\begin} \def\beq{\begin{equation}} \def\F{\scr F{split} &\aa_{t_0}(s) := \ff{s}{t_0} \big(\xi_{t_0}^\mu-\xi_{t_0}^\nu -v^{(2)}\big) - \ff{s(t_0-s)}{t_0^2} M^*\text{\rm{e}}} \def\ua{\underline a} \def\OO{\Omega} \def\oo{\omega^{-sA^*}Q_{t_0}^{-1}\big(v^{(1)}+ V_{t_0}^{\mu,\nu}\big),\\ &V_{t_0}^{\mu,\nu}:= \int_0^{t_0}\text{\rm{e}}} \def\ua{\underline a} \def\OO{\Omega} \def\oo{\omega^{-rA}M\Big\{ \ff{t_0-r}{t_0} v^{(2)} +\ff r {t_0}\big(\xi_{t_0}^\mu-\xi_{t_0}^\nu\big) +\xi_r^\nu- \xi_r^\mu\Big\}\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D r.\end{split} \end{equation} By \eqref{QT}, we find a constant $c_1>0$ independent of $t_0\in (0,T]$ such that \beq\label{AP2} \sup_{t\in [0,t_0]} \big\{t_0|\aa_{t_0}'(t)|+|\aa_{t_0}(t)|\big\}\le \ff{c_1}{t_0^{2(l-1)}} \Big(t_0^{-1}|v^{(1)}|+ |v^{(2)}| + \sup_{t\in [0,t_0]} |\xi_t^\mu-\xi_t^\nu|\Big).\end{equation} Let $Y_t$ solve the SDE with initial value $Y_0$: \beq\label{A2} \begin{cases} \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D Y_t^{(1)}=\big\{AY_t^{(1)} +MY_t^{(2)}\big\}\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D t, \\ \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D Y_t^{(2)}=\big\{b_t(X_t,P_t^*\mu)+ \aa_{t_0}'(t) \big\}\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D t+\ll\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D W_t+\tilde} \def\Ric{\text{\rm{Ric}}\sigma} \def\ess{\text{\rm{ess}}_t(P_t^*\nu)\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D\tilde} \def\Ric{\text{\rm{Ric}} W_t,\ \ t\in [0,t_0]. \end{cases}\end{equation} This and \eqref{E00} imply \beq\label{A3}\begin} \def\beq{\begin{equation}} \def\F{\scr F{split} & Y_t^{(2)}- X_t^{(2)} = \aa_{t_0} (t)+v^{(2)} +\xi_t^\nu-\xi_t^\mu,\\ & Y_t^{(1)}- X_t^{(1)} = \text{\rm{e}}} \def\ua{\underline a} \def\OO{\Omega} \def\oo{\omega^{tA} v^{(1)} + \int_0^t \text{\rm{e}}} \def\ua{\underline a} \def\OO{\Omega} \def\oo{\omega^{(t-s)A} M \big\{ \aa_{t_0}(s)+v^{(2)} +\xi_s^\nu-\xi_s^\mu\big\} \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D s,\ \ \ t\in [0,t_0].\end{split}\end{equation} Consequently, $$Y_{t_0}^{(2)}-X_{t_0}^{(2)} = \xi_{t_0}^\mu-\xi_{t_0}^\nu-v^{(2)}+v^{(2)} +\xi_{t_0}^\nu-\xi_{t_0}^\mu=0,$$ \begin} \def\beq{\begin{equation}} \def\F{\scr F{align*} &Y_{t_0}^{(1)}-X_{t_0}^{(1)}= \text{\rm{e}}} \def\ua{\underline a} \def\OO{\Omega} \def\oo{\omega^{t_0A} v^{(1)} +\int_0^{t_0} \text{\rm{e}}} \def\ua{\underline a} \def\OO{\Omega} \def\oo{\omega^{(t_0-s)A}M \Big\{\ff s {t_0} \big(\xi_{t_0}^\mu-\xi_{t_0}^\nu -v^{(2)}\big)+v^{(2)}+\xi_s^\nu-\xi_s^\mu\Big\}\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D s\\ &\quad - \text{\rm{e}}} \def\ua{\underline a} \def\OO{\Omega} \def\oo{\omega^{t_0A} Q_{t_0}Q_{t_0}^{-1} \bigg(v^{(1)} +\int_0^{t_0} \text{\rm{e}}} \def\ua{\underline a} \def\OO{\Omega} \def\oo{\omega^{-rA}M \Big\{\ff{t_0-r}{t_0} v^{(2)} +\ff r {t_0} \big(\xi_{t_0}^\mu-\xi_{t_0}^\nu\big) +\xi_r^\nu-\xi_r^\mu\Big\}\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D r\bigg)\\ &=0, \end{align*} so that \beq\label{AP4} Y_{t_0}=X_{t_0}.\end{equation} On the other hand, by \eqref{A3} and \eqref{AP2} we find a constant $c_2>0$ uniform in $t_0\in (0,T]$ such that \beq\label{AP3}\begin} \def\beq{\begin{equation}} \def\F{\scr F{split}&\sup_{t\in [0,t_0]} |Y_t -X_t |^2 \le \ff{c_2}{t_0^{4(l-1)} } \Big\{ t_0^{-2}|v^{(1)}|^2+ |v^{(2)}|^2+\sup_{t\in [0,t_0]}|\xi_t^\mu-\xi_t^\nu|^2\Big\} \\ &=\ff{c_2}{ t_0^{4(l-1)} }\Big\{\rr_{t_0}(X_0,Y_0)^2+ \sup_{t\in [0,t_0]}|\xi_t^\mu-\xi_t^\nu|^2\Big\}. \end{split} \end{equation} To formulate the equation of $Y_t$ as \eqref{E00}, let \beq\label{ETS} \eta_s := \ff 1 \ll \big\{b_s(Y_s, P_s^*\nu)- b_s(X_s, P_s^*\mu) - \aa_{t_0}'(s)\big\},\ \ s\in [0,t_0].\end{equation} By {\bf (C)}, \eqref{AP2} and \eqref{AP3}, we find a constant $c_3>0$ uniformly in $t_0\in (0,T]$ such that \beq\label{AP5} \begin} \def\beq{\begin{equation}} \def\F{\scr F{split} |\eta_s |^2\le &\,c_3 K_s \Big\{\mathbb W_2(\mu,\nu)^2 + t_0^{4(1-l)} \rr_{t_0}(X_0,Y_0)^2 + t_0^{4(1-l)} \sup_{t\in [0,t_0]} |\xi_t^\mu-\xi_t^\nu|^2\Big\}\\ & + c_3 t_0^{2-4l}\Big(\rr_{t_0}(X_0,Y_0)^2+\sup_{t\in [0,t_0]} |\xi_t^\nu-\xi_t^\mu|^2\Big).\end{split}\end{equation} By Girsanov's theorem, $$\hat{W}_t:=W_t- \int_0^t \eta_s \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D s,\ \ t\in [0,t_0]$$ is a $d$-dimensional Brownian motion under the weighted conditional probability measure $\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D \Q^{\tilde} \def\Ric{\text{\rm{Ric}} W,0}:=R^{\tilde} \def\Ric{\text{\rm{Ric}} W,0} \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D\P^{\tilde} \def\Ric{\text{\rm{Ric}} W,0}$, where $$R^{\tilde} \def\Ric{\text{\rm{Ric}} W,0}:= \text{\rm{e}}} \def\ua{\underline a} \def\OO{\Omega} \def\oo{\omega^{\int_0^{t_0} \<\eta_s,\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D W_s\>-\ff 1 2 \int_0^{t_0} |\eta_s|^2\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D s}.$$ Let $\tilde} \def\Ric{\text{\rm{Ric}} \xi_t^\nu=(0,\xi_t^\nu).$ By \eqref{A2}, $\hat Y_t:= Y_t-\tilde} \def\Ric{\text{\rm{Ric}} \xi_t^\nu$ solves the SDE $$\begin{cases} \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D \hat Y_t^{(1)}=\big\{A\hat Y_t^{(1)} +M\hat Y_t^{(2)}+M\xi_t^\nu\big\}\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D t, \\ \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D \hat Y_t^{(2)}= b_t(\hat Y_t+\tilde} \def\Ric{\text{\rm{Ric}} \xi_t^\nu ,P_t^*\nu) \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D t+\ll\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D \hat{W}_t,\ \ t\in [0,t_0],\ \hat Y_0=Y_0. \end{cases}$$ Letting $X_t^\nu$ solve \eqref{E00} with $X_0^\nu=Y_0$, we see that $\hat X_t^\nu:= X_t^\nu-\tilde} \def\Ric{\text{\rm{Ric}}\xi_t^\nu$ solves the same equation as $\hat{Y}_t$ for $W_t$ replacing $\hat{W}_t$. By the weak uniqueness and \eqref{AP4}, \eqref{NB} holds for $\R^{m+d}$ replacing $\R^d$, i.e. for any $f\in \B_b(\R^{m+d})$, \beq\label{NB'} P_{t_0}^{\tilde} \def\Ric{\text{\rm{Ric}} W,0} f(X_{t_0}^\nu):= \E^{\tilde} \def\Ric{\text{\rm{Ric}} W,0} [f(X_{t_0}^\nu)]= \E^{\tilde} \def\Ric{\text{\rm{Ric}} W,0} [R^{\tilde} \def\Ric{\text{\rm{Ric}} W,0} f(Y_{t_0})] = \E^{\tilde} \def\Ric{\text{\rm{Ric}} W,0} [R^{\tilde} \def\Ric{\text{\rm{Ric}} W,0} f(X_{t_0})].\end{equation} Combining this with Young's inequality and \eqref{AP5}, we find constants $c_4>0$ such that \beq\label{YPP} \begin} \def\beq{\begin{equation}} \def\F{\scr F{split} &P_{t_0}^{\tilde} \def\Ric{\text{\rm{Ric}} W,0} \log f(X_{0}^\nu)- \log P_{t_0}^{\tilde} \def\Ric{\text{\rm{Ric}} W,0} f(X_0^\mu) \le \E^{\tilde} \def\Ric{\text{\rm{Ric}} W,0} [R^{\tilde} \def\Ric{\text{\rm{Ric}} W,0}\log R^{\tilde} \def\Ric{\text{\rm{Ric}} W,0}] = \ff 1 2\E_{\Q^{\tilde} \def\Ric{\text{\rm{Ric}} W,0}} \int_0^{t_0}|\eta_t|^2\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D t\\ &\le c_4 \Big\{\mathbb W_2(\mu,\nu)^2 + t_0^{3-4l} \rr_{t_0}(X_0,Y_0)^2+ t_0^{3-4l}\sup_{t\in [0,t_0]} |\xi_t^\nu-\xi_t^\mu|^2\Big\}. \end{split} \end{equation} By taking expectation, using Jensen's inequality, \eqref{*N1}, \eqref{W0} and \eqref{0'}, we prove \eqref{EC2}. \end{proof} \subsection{Bismut formula} We will use Definition 2.1 for $\R^{m+d}$ replacing $\R^d$. The following assumption is parallel to {\bf (B)} with an additional rank condition. \begin} \def\beq{\begin{equation}} \def\F{\scr F{enumerate} \item[{\bf (D)}] $(\tilde} \def\Ric{\text{\rm{Ric}}\sigma} \def\ess{\text{\rm{ess}},b)$ satisfies {\bf (B)} for $\R^{m+d}$ replacing $\R^d$, and the rank condition \eqref{RRS} holds for some $1\le l\le m.$ \end{enumerate} Let $X_0^\mu$ be $\F_0$-measurable such that $\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{X_0^\mu}=\mu\in\scr P_{k}(\R^{m+d})$, and let $X_t^\mu$ solve \eqref{E00} with initial value $X_0^\mu$. For any $\vv\ge 0$, denote $$\mu^\vv:= \mu\circ(id+\vv\phi)^{-1},\ \ X_0^{\mu^\vv}:= X_0^\mu+\vv\phi(X_0^\mu).$$ Let $X_t^{\mu^\vv}$ solve \eqref{E00} with initial value $X_0^{\mu^\vv}$. So, $$X_t^{\mu}=X_t^{\mu^0},\ \ P_t^*\mu^\vv=\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{X_t^{\mu^\vv}},\ \ t\in [0,T], \vv\ge 0.$$ By \cite[Lemma 5.2]{BRW}, for any $\phi=(\phi^{(1)},\phi^{(2)})\in T_{\mu,k}(\R^{m+d}),$ {\bf (D)} implies that $$\nabla_{\phi}X_\cdot^\mu:=\lim_{\vv\downarrow 0}\ff{X_\cdot^{\mu^\vv}-X_\cdot^\mu}\vv$$ exists in $L^k(\OO\rightarrow}\def\l{\ell C([0,T]; \R^{m+d});\P)$, and there exists a constant $c>0$ such that \begin{equation*}\E\bigg[\sup_{t\in [0,T]}|\nabla} \def\pp{\partial} \def\E{\mathbb E_\phi X_t^\mu|^k\bigg]\le c \|\phi\|_{L^k(\mu)}^k,\ \ \mu\in \scr P_k(\R^{m+d}), \phi\in T_{\mu,k}(\R^{m+d}).\end{equation*} Finally, for any $t\in (0,T]$ and $s\in [0,t]$, let \begin} \def\beq{\begin{equation}} \def\F{\scr F{align*} & \gg_t^{\mu,\phi}:= \int_0^t \Big\<\E\big[\big\<D^I\tilde{\sigma} \def\ess{\text{\rm{ess}}}_r(P_r^*\mu)(X_r^\mu),\ \nabla} \def\pp{\partial} \def\E{\mathbb E_\phi X_r^\mu\big\>\big],\ \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D\tilde} \def\Ric{\text{\rm{Ric}} W_r\Big\>\\ &V_t^{\mu,\phi}:= \int_0^t \text{\rm{e}}} \def\ua{\underline a} \def\OO{\Omega} \def\oo{\omega^{-rA}M\Big\{ \ff{t-r}t \phi^{(2)}(X_0^\mu) -\ff r t \gg_{t}^{\mu,\phi} +\gg_r^{\mu,\phi}\Big\} \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D r,\\ &\aa_t^{\mu,\phi}(s):= -\ff{s}t \big\{\phi^{(2)}(X_0^\mu) +\gg_{t}^{\mu,\phi}\big\} -\ff{s(t-s)}{t^2} M^*\text{\rm{e}}} \def\ua{\underline a} \def\OO{\Omega} \def\oo{\omega^{-sA^*}Q_{t}^{-1} \big\{\phi^{(1)}(X_0^\mu) +V_t^{\mu,\phi}\big\},\end{align*} and define \begin} \def\beq{\begin{equation}} \def\F{\scr F{align*} &N_{s,t}^{(1)}:= \text{\rm{e}}} \def\ua{\underline a} \def\OO{\Omega} \def\oo{\omega^{sA}\phi^{(1)}(X_0^\mu)+\int_0^s\text{\rm{e}}} \def\ua{\underline a} \def\OO{\Omega} \def\oo{\omega^{(s-r)A}M\big\{ \aa_t^{\mu,\phi} (r) +\phi^{(2)}(X_0^\mu) +\gg_r^{\mu,\phi}\big\}\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D r\\ &N_{s,t}^{(2)}:= \aa_t^{\mu,\phi}(s) +\phi^{(2)}(X_0^\mu) +\gg_s^{\mu,\phi},\\ & M_{s,t}^{\mu,\phi}:= \E\big[\<D_\phi^I b_s(z,\cdot)(P_s^*\mu)(X_s^\mu),\nabla} \def\pp{\partial} \def\E{\mathbb E_\phi X_s^\mu\>]\big]_{z=X_s^\mu}-(\aa_t^{\mu,\phi})'(s).\end{align*} Then we have the following result. \begin} \def\beq{\begin{equation}} \def\F{\scr F{thm}\label{TA3} Assume {\bf(D)} and let $N_{s,t}^{\mu,\phi} :=\big(N_{s,t}^{(1)}, N_{s,t}^{(2)}\big)\in \R^{m+d}, 0\le s\le t.$ For any $t\in (0,T]$, $\mu\in \scr P_k(\R^{m+d}),$ $\phi\in T_{\mu,k}(\R^{m+d})$ and $f\in \B_b(\R^{m+d})$, \beq\label{BS3} D_\phi^I P_{t}f(\mu)= \ff 1\ll \E\bigg[f(X_{t}^{\mu}) \int_0^{t} \Big\<\nabla} \def\pp{\partial} \def\E{\mathbb E_{N_{s,t}^{\mu,\phi}}b_s(\cdot, P_s^*\mu)(X_s^\mu) + M_{s,t}^{\mu,\phi},\ \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D W_s\Big\> \bigg]. \end{equation} Consequently, $P_{t}f$ is intrinsically differentiable, and there exists a constant $c>0$ such that \beq\label{ESN} \|D^IP_{t}f(\mu)\|_{L^{k^*}(\mu)}\le \ff c {t^{2l-\frac{1}{2}}} \big(P_{t}|f|^{k^\ast}(\mu)\big)^{\ff 1 {k^\ast}} ,\ \ t\in (0,T], f\in\B_b(\R^{m+d}).\end{equation} \end{thm} \begin{proof} It is easy to see that under {\bf (D)}, \eqref{ESN} follows from \eqref{BS3}. So, it suffices to prove \eqref{BS3}. Let $ X_t^\mu$ solve \eqref{E00} with initial value $ X_0^\mu$, and for any $\vv\in (0,1]$, let $ Y_t^\vv$ solve \eqref{A2} for $Y_0=Y_0^\vv:=X_0^\mu+\vv\phi(X_0^\mu)$ and $ \nu=\mu^\vv.$ Then $$\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{Y_0}=\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{Y_0^\vv}=\mu^\vv.$$ Let $\aa_{t_0}^\vv(s)$ be defined in \eqref{A1} for $\nu=\mu^\vv$. By \eqref{a2} and \eqref{TTX}, we have \beq\label{XXZ} \lim_{\vv\downarrow 0} \ff 1 \vv \aa_{t_0}^\vv (s)= \aa_{t_0}^{\mu,\phi}(s),\ \ \ s\in [0,t_0],\end{equation} while \eqref{A3} and \eqref{ETS} reduces to \begin{equation*}\begin} \def\beq{\begin{equation}} \def\F{\scr F{split} & (Y_t^\vv)^{(2)}- (X_t^\mu)^{(2)} = \aa_{t_0}^\vv (t)+\vv\phi^{(2)}(X_0^\mu) +\xi_t^{\mu^\vv} -\xi_t^\mu,\\ & (Y_t^\vv)^{(1)}- (X_t^\mu)^{(1)} = \vv \text{\rm{e}}} \def\ua{\underline a} \def\OO{\Omega} \def\oo{\omega^{tA} \phi^{(1)}(X_0^\mu) + \int_0^t \text{\rm{e}}} \def\ua{\underline a} \def\OO{\Omega} \def\oo{\omega^{(t-s)A} M \big\{ \aa_{t_0}^\vv(s)+\vv\phi^{(2)}(X_0^\mu) +\xi_s^{\mu^\vv}-\xi_s^\mu\big\} \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D s,\end{split}\end{equation*} and \begin{equation*} \eta_t^\vv =\ff 1 \ll \Big\{b_t(Y_t^\vv, P_t^*\mu^\vv)- b_t(X_t^\mu, P_t^*\mu)- \{\aa_{t_0}^{\vv}\}'(t)\Big\},\ \ t\in[0,t_0].\end{equation*} Then by \eqref{a2} and \eqref{XXZ}, we have \beq\label{A5'} \lim_{\vv\downarrow 0} \ff 1 \vv(Y_t^\vv-X_t^\mu)= N_{t,t_0}^{\mu,\phi}. \end{equation} Let $$R^\vv:= \text{\rm{e}}} \def\ua{\underline a} \def\OO{\Omega} \def\oo{\omega^{\int_0^{t_0} \<\eta_t^\vv,\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D W_t\> -\ff 1 2 \int_0^{t_0} |\eta_t^\vv|^2\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D t}.$$ By \eqref{NB'}, we obtain $$ P_{t_0}^{\tilde} \def\Ric{\text{\rm{Ric}} W,0} f(X_0^{\mu^\vv}):= \E^{\tilde} \def\Ric{\text{\rm{Ric}} W,0}[f(X_{t_0}^{\mu^\vv})]= \E^{\tilde} \def\Ric{\text{\rm{Ric}} W,0}[R^\vv f(X_{t_0}^\mu)],\ \ f\in \B_b(\R^{m+d}).$$ As in \eqref{*W2}, by {\bf (D)}, \eqref{A5'} and \eqref{a2}, we derive \beq\label{KM} \begin} \def\beq{\begin{equation}} \def\F{\scr F{split}& \lim_{\vv\downarrow0} \ff{P_{t_0}^{\tilde} \def\Ric{\text{\rm{Ric}} W,0}f(X_0^{\mu^\vv}) - P_{t_0}^{\tilde} \def\Ric{\text{\rm{Ric}} W,0} f(X_0^\mu)}\vv= \lim_{\vv\downarrow 0} \E^{\tilde} \def\Ric{\text{\rm{Ric}} W,0} \Big[f(X_{t_0}^\mu) \ff{R^\vv-1}\vv\Big]\\ &= \frac{1}{\lambda} \E^{\tilde} \def\Ric{\text{\rm{Ric}} W,0}\bigg[ f(X_{t_0}^\mu) \int_0^{t_0} \Big\<\nabla} \def\pp{\partial} \def\E{\mathbb E_{N_{s,t_0}^{\mu,\phi}}b_s(\cdot, P_s^*\mu)(X_s^\mu) + M_{s,t_0}^{\mu,\phi},\ \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D W_s\Big\>\bigg].\end{split}\end{equation} Finally, similarly to the proof of \eqref{*W}, since {\bf (D)} implies {\bf (C)} for $(\scr P_k(\R^{m+d}),\mathbb W_k)$ replacing $(\scr P_2(\R^{m+d}),\mathbb W_2)$, the argument leading to \eqref{YPP} implies $$P_{t_0}^{\tilde} \def\Ric{\text{\rm{Ric}} W,0} \log f(X_{0}^\nu)- \log P_{t_0}^{\tilde} \def\Ric{\text{\rm{Ric}} W,0} f(X_0^\mu) \le c(t_0) \Big\{\mathbb W_k(\mu,\nu)^2 + \rr_{t_0}(X_0^\mu,X_0^\nu)^2+ \sup_{t\in [0,t_0]} |\xi_t^\nu-\xi_t^\mu|^2\Big\} $$ for some constant $c(t_0)>0$. Therefore, as shown in step (b) of the proof of Theorem \ref{TA2'}, this enables us to apply the dominated convergence theorem with \eqref{KM} to derive \begin} \def\beq{\begin{equation}} \def\F{\scr F{align*} &D_\phi^IP_{t_0}f(\mu) = \lim_{\vv\downarrow 0} \ff{\E[P_{t_0}^{\tilde} \def\Ric{\text{\rm{Ric}} W,0} f(X_0^{\mu^\vv})- P_{t_0}^{\tilde} \def\Ric{\text{\rm{Ric}} W,0} f(X_0^\mu)]}\vv= \E\Big\{\lim_{\vv\downarrow 0} \E^{\tilde} \def\Ric{\text{\rm{Ric}} W,0} \Big[f(X_{t_0}^\mu) \ff{R^\vv-1}\vv\Big]\Big\} \\ &= \ff 1\ll \E\bigg[f(X_{t_0}^{\mu}) \int_0^{t_0} \Big\<\nabla} \def\pp{\partial} \def\E{\mathbb E_{N_{s,t_0}^{\mu,\phi}}b_s(\cdot, P_s^*\mu)(X_s^\mu) + M_{s,t_0}^{\mu,\phi},\ \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D W_s\big\> \bigg]. \end{align*} \end{proof} \section{Exponential ergodicity in entropy} Following the line of \eqref{RW}, we may use the log-Harnack inequality to study the exponential ergodicity in entropy. To this end, we consider the time homogeneous equation on $\R^d$ \beq\label{EES} \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D X_t= b(X_t, \scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{X_t})\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D t+\ll\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D W_t+ \tilde} \def\Ric{\text{\rm{Ric}}\sigma} \def\ess{\text{\rm{ess}}(\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{X_t})\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D \tilde} \def\Ric{\text{\rm{Ric}} W_t, \ \ t\ge 0,\end{equation} and the degenerate model on $\R^{m+d}$ \beq\label{EES2} \begin{cases} \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D X_t^{(1)}=\big\{AX^{(1)}_t+MX_t^{(2)}\big\}\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D t, \\ \text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D X_t^{(2)}=b(X_t,\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{X_t})\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D t+\ll\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D W_t+ \tilde} \def\Ric{\text{\rm{Ric}}\sigma} \def\ess{\text{\rm{ess}}(\scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{X_t})\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D \tilde} \def\Ric{\text{\rm{Ric}} W_t,\ \ t\ge 0, \end{cases} \end{equation}where $\ll>0$ is a constant. \subsection{Non-degenerate case} \begin{enumerate} \item[\bf{(E)}] There exist constants $K, \theta_1,\theta_2>0$ with $\theta:=\theta_2-\theta_1>0$, such that for any $\mu,\nu\in \scr P_2(\R^d)$ and $x,y\in \R^d$, \begin{equation*}\begin{split} & |b(x,\mu)-b(y,\nu)|+|\tilde} \def\Ric{\text{\rm{Ric}} \sigma (\mu)-\tilde} \def\Ric{\text{\rm{Ric}} \sigma (\nu)|\leq K (|x-y|+\mathbb W_2(\mu,\nu)),\\ & 2\<b(x,\mu)-b(y,\nu),x-y\>+ \|\sigma(\mu)-\sigma(\nu)\|^2_{HS} \leq -\theta_2|x-y|^2+\theta_1 \mathbb W_2(\mu,\nu)^2, \end{split}\end{equation*}where $\|\cdot\|_{HS}$ is the Hilbert-Schmidt norm. \end{enumerate} By \cite[Theorem 2.1]{FYW1}, this assumption implies that \eqref{EES} is well-posed for distributions in $\scr P_2$, and $P_t^*$ has a unique invariant probability measure $\bar\mu\in \scr P_2(\R^d)$ such that \beq\label{EXP} \mathbb W_2(P_t^*\mu,\bar\mu)^2\le \text{\rm{e}}} \def\ua{\underline a} \def\OO{\Omega} \def\oo{\omega^{-\theta t} \mathbb W_2(\mu,\bar\mu)^2,\ \ t\ge 0.\end{equation} The following result ensures the exponential convergence in entropy. \begin{thm}\label{EXPEN} Assume {\bf (E)} and let $P_t^*$ be associated with $\eqref{EES}$. Then there exists a constant $c>0$ such that \begin{align*}&\max\big\{\mathbb W_2(P_t^\ast\mu, \bar\mu)^2,\mathrm{Ent}(P_t^\ast\mu|\bar\mu)\big\}\\ &\leq c\text{\rm{e}}} \def\ua{\underline a} \def\OO{\Omega} \def\oo{\omega^{-\theta t}\min\big\{\mathbb W_2(\mu, \bar\mu)^2,\mathrm{Ent}(\mu|\bar\mu)\big\},\ \ \mu\in\scr P_2(\R^d),t\ge 1.\end{align*} \end{thm} \begin{proof} According to the proof of \cite[Theorem 2.3]{RW}, {\bf (E)} implies the Talagrand inequality $$\mathbb W_2(\mu, \bar \mu)^2\leq c_1\mathrm{Ent}(\mu|\bar\mu), \ \ \mu\in\scr P_2(\R^d) $$ for some constant $c_1>0$. According to \cite[Theorem 2.1]{RW}, this together with \eqref{EXP} and Theorem \ref{Loh} implies the desired assertion. \end{proof} \subsection{Degenerate case} To study the exponential ergodicity for the degenerate model \eqref{EES2}, we extend the assumption $(A1)$-$(A3)$ in \cite{W17} to the present distribution dependent case. \begin} \def\beq{\begin{equation}} \def\F{\scr F{enumerate} \item[\bf{(F)}] $\tilde} \def\Ric{\text{\rm{Ric}}\sigma} \def\ess{\text{\rm{ess}}$ and $b$ are Lipschitz continuous on $\scr P_2(\R^{m+d})$ and $\R^{m+d} \times \scr P_2(\R^{m+d})$ respectively. $(A,M)$ satisfies the rank condition \eqref{RRS} for some $1\le l\le m$, and there exist constants $r>0, \theta_2>\theta_1\ge 0$ and $r_0\in (-\|M\|^{-1}, \|M\|^{-1})$ such that \begin} \def\beq{\begin{equation}} \def\F{\scr F{align*} &\ff 1 2 \|\tilde} \def\Ric{\text{\rm{Ric}}\sigma} \def\ess{\text{\rm{ess}}(\mu)-\tilde} \def\Ric{\text{\rm{Ric}}\sigma} \def\ess{\text{\rm{ess}}(\nu)\|_{HS}^2 + \big\<b(x,\mu)-b(y,\nu),\ x^{(2)}-y^{(2)}+rr_0M^*(x^{(1)}-y^{(1)})\big\>\\ & +\big\<r^2(x^{(1)}-y^{(1)})+ rr_0M(x^{(2)}-y^{(2)}),\ A(x^{(1)}-y^{(1)})+ M(x^{(2)}-y^{(2)})\big\>\\ &\le \theta_1\mathbb W_2(\mu,\nu)^2-\theta_2|x-y|^2,\ \ x,y\in \R^{m+d},\ \mu,\nu\in \scr P_2(\R^{m+d}).\end{align*}\end{enumerate} In the distribution free case, some examples are presented in \cite[Section 5]{W17}, which can be extended to the present setting if the Lipschitz constant of $\tilde} \def\Ric{\text{\rm{Ric}}\sigma} \def\ess{\text{\rm{ess}}(\mu)$ and $b(x,\mu)$ in $\mu\in \scr P_2(\R^{m+d})$ is small enough. \begin} \def\beq{\begin{equation}} \def\F{\scr F{thm} Assume {\bf (F)}. Then $P_t^*$ associated with $\eqref{EES2}$ has a unique invariant probability measure $\bar\mu$, and there exist constants $c,\ll>0$ such that $$\max\big\{\Ent(P_t^*\mu|\bar\mu), \mathbb W_2(P_t^*\mu,\bar\mu)^2\big\}\le c\text{\rm{e}}} \def\ua{\underline a} \def\OO{\Omega} \def\oo{\omega^{-\ll t} \mathbb W_2(\mu,\bar\mu)^2,\ \ t\ge 1, \mu\in \scr P_2(\R^{m+d}).$$ \end{thm} \begin} \def\beq{\begin{equation}} \def\F{\scr F{proof} Let $$\rr(x):= \ff{r^2} 2 |x^{(1)}|^2 +\ff 1 2 |x^{(2)}|^2 + rr_0 \<x^{(1)}, Mx^{(2)}\>,\ \ x=(x^{(1)},x^{(2)})\in \R^{m+d}.$$ By $r_0\|M\|<1$ and $r>0$, we find a constant $c_0\in (0,1) $ such that \beq\label{N0} c_0 |x|^2 \le \rr(x)\le c_0^{-1} |x|^2,\ \ x\in\R^{m+d}.\end{equation} Let $X_t$ and $Y_t$ solve \eqref{EES2} with initial values \beq\label{N1} \scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{X_0}= \mu,\ \ \scr L}\def\Tt{\tt} \def\TT{\tt}\def\II{\mathbb I_{Y_0}=\nu,\ \ \mathbb W_2(\mu,\nu)^2=\E[|X_0-Y_0|^2].\end{equation} By {\bf (F)} and It\^o's formula, we obtain \begin{align}\label{rhs}\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D \rr (X_t-Y_t)\le \Big\{\theta_1 \mathbb W_2(P_t^*\mu,P_t^*\nu)^2 -\theta_2|X_t-Y_t|^2\Big\}\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D t +\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D M_t \end{align} for some martingale $M_t$, and $$\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D \rr (X_t)\le \Big\{\theta_1 \E[|X_t|^2] -\theta_2|X_t|^2+C+C|X_t|\Big\}\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D t +\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D \tilde{M}_t$$ for some martingale $\tilde{M}_t$ and constant $C>0.$ In particular, by \eqref{N0}, the latter implies \beq\label{N*} \sup_{t\ge 0} \E[|X_t|^2] <\infty.\end{equation} Since \beq\label{N3} \mathbb W_2(P_t^*\mu,P_t^*\nu)^2\le \E[|X_t-Y_t|^2],\end{equation} \eqref{rhs} and \eqref{N0} imply $$\E[\rr(X_t-Y_t)]- \E[\rr(X_s-Y_s)]\le -c_0(\theta_2-\theta_1) \int_s^t \E[\rr(X_r-Y_r)]\text{\rm{d}}} \def\bb{\beta} \def\aa{\alpha} \def\D{\scr D r,\ \ t\ge s\ge 0.$$ By Gronwall''s inequality, we derive $$ \E[\rr(X_t-Y_t)]\le \text{\rm{e}}} \def\ua{\underline a} \def\OO{\Omega} \def\oo{\omega^{-c_0(\theta_2-\theta_1)t} \E[\rr(X_0-Y_0)],\ \ t\ge 0.$$ This together with \eqref{N0}, \eqref{N1} and \eqref{N3} yields \begin} \def\beq{\begin{equation}} \def\F{\scr F{align*} &\mathbb W_2(P_t^*\mu,P_t^*\nu)^2\le \E[|X_t-Y_t|^2] \le c_0^{-1} \E[\rr(X_t-Y_t)]\le c_0^{-1} \text{\rm{e}}} \def\ua{\underline a} \def\OO{\Omega} \def\oo{\omega^{-c_0(\theta_2-\theta_1)t} \E[\rr(X_0-Y_0)]\\ &\le c_0^{-2} \text{\rm{e}}} \def\ua{\underline a} \def\OO{\Omega} \def\oo{\omega^{-c_0(\theta_2-\theta_1)t} \E[|X_0-Y_0|^2]= c_0^{-2} \text{\rm{e}}} \def\ua{\underline a} \def\OO{\Omega} \def\oo{\omega^{-c_0(\theta_2-\theta_1)t}\mathbb W_2(\mu,\nu)^2,\ \ t\ge 0,\mu,\nu\in \scr P_2(\R^{m+d}).\end{align*} As shown in \cite[Proof of Theorem 3.1(2)]{FYW1}, this together with \eqref{N*} implies that $P_t^*$ has a unique invariant probability measure $\bar\mu\in\scr P_2(\R^d)$, and \beq\label{N4} \mathbb W_2(P_t^*\mu,\bar\mu)^2\le c_0^{-2} \text{\rm{e}}} \def\ua{\underline a} \def\OO{\Omega} \def\oo{\omega^{-c_0(\theta_2-\theta_1)t}\mathbb W_2(\mu,\bar\mu)^2,\ \ t\ge 0,\mu \in \scr P_2(\R^{m+d}). \end{equation} Finally, by the log-Harnack inequality \eqref{EC2}, there exists a constant $c_1>0$ such that $$\Ent(P_1^*\mu|\bar \mu)\le c_1 \mathbb W_2(\mu,\bar \mu)^2.$$ Combining this with \eqref{N4} and using the semigroup property $P_{t}^*= P_{t-1}^*P_1^*$ for $t\ge 1$, we finish the proof. \end{proof} When $b$ is of a gradient type (induced by $\sigma} \def\ess{\text{\rm{ess}}$) as in \cite[(2.21)]{RW} such that the invariant probability measure $\bar\mu$ is explicitly given and satisfies the Talagrand inequality, we may also derive the stronger upper bound as in Theorem \ref{EXPEN}. We skip the details. \begin} \def\beq{\begin{equation}} \def\F{\scr F{thebibliography}{99} \bibitem{ATW} M. Arnaudon, A. Thalmaier, F.-Y. Wang, \emph{Gradient estimates and Harnack inequalities on non-compact Riemannian manifolds, } Stochastic Process. Appl. 119(2009), 3653-3670. \bibitem{BH} Y. Bai, X. Huang, \emph{Log-Harnack inequality and exponential ergodicity for distribution dependent CKLS and Vasicek Model,} arXiv:2108.02623v5. \bibitem{BRW} J. Bao, P. Ren, F.-Y. Wang, \emph{Bismut formula for Lions derivative of distribution-path dependent SDEs,} J. Differential Equations 282(2021), 285-329. \bibitem{BRS} V. I. Bogachev, M. R\"{o}ckner, S. V. Shaposhnikov, \emph{Distances between transition probabilities of diffusions and applications to nonlinear Fokker-Planck-Kolmogorov equations,} J. Funct. Anal. 271 (2016), 1262-1300. \bibitem{HRW} X. Huang, P. Ren, F.-Y. Wang, \emph{Distribution dependent stochastic differential equations, } Front. Math. China 16(2021), 257-301. \bibitem{HRW19} X. Huang, M. R\"{o}ckner, F.-Y. Wang, \emph{Non-linear Fokker--Planck equations for probability measures on path space and path-distribution dependent SDEs,} Discrete Contin. Dyn. Syst. 39(2019), 3017-3035. \bibitem{HS} X. Huang, Y. Song, \emph{Well-posedness and regularity for distribution dependent SPDEs with singular drifts,} Nonlinear Anal. 203(2021), 112167. \bibitem{HSW} X. Huang, Y. Song, F.-Y. Wang,\emph{Bismut formula for intrinsic/Lions derivatives of distribution dependent SDEs with singular coefficients,} Discrete Contin. Dyn. Syst. 42(2022), 4597-4614. \bibitem{HW18} X. Huang, F.-Y. Wang, \emph{Distribution dependent SDEs with singular coefficients,} Stochastic Process. Appl. 129(2019), 4747-4770. \bibitem{HW22a} X. Huang, F.-Y. Wang, \emph{Log-Harnack inequality and Bismut formula for singular McKean-Vlasov SDEs,} arXiv:2207.11536. \bibitem{HW21} X. Huang, F.-Y. Wang, \emph{Derivative estimates on distributions of McKean-Vlasov SDEs,} Electron. J. Probab. 26(2021), 1-12. \bibitem{PW} E. Priola, F.-Y. Wang, \emph{Gradient estimates for diffusion semigroups with singular coefficients,} J. Funct. Anal. 236(2006), 244-264. \bibitem{RW19} P. Ren, F.-Y. Wang, \emph{Bismut formula for Lions derivative of distribution dependent SDEs and applications,} J. Differential Equations 267(2019), 4745-4777. \bibitem{RW} P. Ren, F.-Y. Wang, \emph{Exponential convergence in entropy and Wasserstein for McKean-Vlasov SDEs,} Nonlinear Anal. 206(2021), 112259. \bibitem{RW0} P. Ren, F.-Y. Wang, \emph{Derivative formulas in measure on Riemannian manifolds,} Bull. Lond. Math. Soc. 53(2021), 1786-1800. \bibitem{S} T. Seidman,\emph{ How violent are fast controls?} Math. Control Signals Systems 1(1988), 89-95. \bibitem{W97} F.-Y. Wang, \emph{Logarithmic Sobolev inequalities on noncompact Riemannian manifolds,} Probab. Theory Related Fields 109(1997),417-424. \bibitem{W10} F.-Y. Wang, \emph{Harnack inequalities on manifolds with boundary and applications,} J. Math. Pures Appl. 94(2010), 304-321. \bibitem{Wbook} F.-Y. Wang, \emph{Harnack Inequality for Stochastic Partial Differential Equations,} Springer, New York, 2013. \bibitem{W17} F.-Y. Wang, \emph{Hypercontractivity and applications for stochastic Hamiltonian systems, } J. Funct. Anal. 272(2017), 5360-5383. \bibitem{FYW1} F.-Y. Wang, \emph{Distribution-dependent SDEs for Landau type equations,} Stochastic Process. Appl. 128(2018), 595-621. \bibitem{FYW3} F.-Y. Wang, \emph{Regularity estimates and intrinsic-Lions derivative formula for singular McKean-Vlasov SDEs,} arXiv:2109.02030. \bibitem{WZ1} F.-Y. Wang, X. Zhang, \emph{Derivative formula and applications for degenerate diffusion semigroups,} J. Math. Pures Appl. 99(2013),726-740. \end{thebibliography} \end{document}

0fe221ff09cfb714717cc80989b9bbfeac327854

\section{Introduction} In the standard quantum theory, a set of quantum measurements is called incompatible if these measurements cannot be performed simultaneously on a single copy of a quantum system \cite{review}. This notion of measurement incompatibility is one of the fundamental features of quantum theory that differentiates quantum mechanics from the formulation of classical physics. Quantum measurement incompatibility is at the root of demonstrating various fundamental quantum aspects ranging from Bell-nonlocality \cite{Fine1982,PRL103230402}, Einstein–Podolsky–Rosen steering \cite{PhysRevLett.113.160403,PhysRevLett.113.160402,PhysRevLett.115.230402}, measurement uncertainty relations \cite{mur,saha2020}, quantum contextuality \cite{LIANG20111,Xu2019}, quantum violation of macrorealism \cite{PhysRevA.100.042117}, to temporal and channel steering \cite{Karthik:15,PhysRevA.91.062124,PhysRevA.97.032301}. Bell inequality violation is the most compelling operational witness of incompatible measurements since it relies only on the input-output statistics of bipartite systems \cite{PRL103230402,PhysRevA.87.052125,math4030052}. Further, measurement incompatibility can also be witnessed through Einstein–Podolsky–Rosen steering \cite{PhysRevLett.113.160403,PhysRevLett.113.160402,PhysRevLett.115.230402,PhysRevLett.116.240401}. These protocols, however, rely on entanglement. Only recently, witnessing of quantum measurement incompatibility in the prepare-and-measure scenario based on some state discrimination task \cite{PhysRevLett.122.130402} has been proposed. It is particularly noteworthy that measurement incompatibility is necessary but not sufficient for Bell inequality violations employing fully untrusted devices \cite{Bene_2018,PhysRevA.97.012129}, whereas incompatibility is shown to be necessary as well as sufficient in steering with one-sided trusted devices \cite{PhysRevLett.113.160402,PhysRevLett.115.230402} and in state discrimination task with fully trusted preparations \cite{PhysRevLett.122.130403} (also see \cite{PhysRevA.98.012126,PhysRevA.100.042308}). Notwithstanding, the generic link between measurement incompatibility and nonclassical correlations in the simplest prepare-and-measure scenario is still not fully explored. The present article is motivated towards filling this important gap in the relevant literature. Moreover, the results presented here address whether incompatible quantum measurements are necessary for probing quantum advantage in any one-way communication task. Apart from addressing this fundamental question, this work aims to provide an operational witness of incompatibility for any set of quantum measurements of an arbitrary setting - any set of an arbitrary number of measurements acting on an arbitrary (but finite) given dimension wherein different measurements have different arbitrary number of outcomes. Specifically, we consider the one-way communication scenario consisting of two players, say, Alice (sender) and Bob (receiver). Alice and Bob are given inputs such that each player does not know the input of the other player. Alice, upon receiving her input, sends classical or quantum communication to Bob. Bob, upon receiving his input and the communication sent by Alice, produces the outputs. In such scenario, we show that any quantum advantage in an arbitrary communication task over all possible classical strategies with unlimited shared randomness implies that the quantum measurements performed by Bob to produce the outputs are incompatible. Therefore, any one-way communication task in prepare-and-measure scenario serves as a tool to witness measurement incompatibility in a semi-device independent way. Furthermore, we point out that whenever the figure of merit of any task is a convex function of the input-output statistics, its maximum value in classical communication and quantum communication with compatible measurements are the same. The result \cite{Carmeli2020}, that a pair of quantum measurements is incompatible whenever it provides advantage in random access code task, becomes a corollary of our observation. Subsequently, we focus on a specific quantum communication task in the prepare-and-measure scenario, namely, Random Access Codes (RAC) \cite{racambainis}. Based on the operational figure of merit of this task, we propose a witness of measurement incompatibility of a set of arbitrary number of quantum measurements having arbitrary number of outcomes acting on arbitrary dimensional state. Specifically, we derive upper bound (or, exact value in specific cases) of the average success probability of RAC assisted with the best classical strategy, or equivalently the best quantum strategy involving compatible measurements by the receiver. Therefore, given any set of quantum measurements, if the average success probability of RAC involving the given measurements by the receiver exceeds the above bound, then we can certify that the given measurements are incompatible. Here, it should be noted that RAC, being one of the fundamental quantum communication protocols, has been implemented in a series of experiments \cite{exprac1,exprac2,armin2015,racexpt3,distributedrac,parallelrac}. Hence, the results presented in this study can be used as experimental tool to witness measurement incompatibility based on present day technology. Finally, we identify all sets of three incompatible rank-one projective qubit measurements that can be witnessed by RAC. We next proceed by first explaining the definition of measurement incompatibility, followed by detailed analysis and discussions of illustrative results. \section{Quantum Measurement Incompatibility}\label{sec2} An arbitrary measurement is conceptualized by some Positive Operator-Valued Measure (POVM) defined as $E_y \equiv \{M_{b_y|y}\}_{b_y}$ with $M_{b_y|y} \geqslant 0$ for all $b_y$ and $\sum_{b_y} M_{b_y|y} = \mathbbm{1}$. Here $y$ corresponds to the choice of measurement, and $b_y$ denotes the outcomes of measurement $y$. A set of measurements $\{E_y\}_{y}$ with $y \in [n]$ (here we use the notation $[k]:=\{1,\dots,k\}$) is compatible \cite{review} if there exists a parent POVM $\{G_{\kappa}: G_{\kappa} > 0 \, \forall \kappa, \, \sum_{\kappa} G_\kappa = \mathbbm{1} \}$ and classical post-processing for each $y$ given by $\{P_y(b_y|\kappa)\}$ such that \begin{equation} \label{com} \forall b_y,y, \quad M_{b_y|y} = \sum_\kappa P_y(b_y|\kappa) G_\kappa . \end{equation} Post-processing for each $y$ is defined by $\{P_y(b_y|\kappa)\}$ such that \begin{equation} P_y(b_y|\kappa) \geqslant 0 \, \, \forall y, b_y, \kappa ; \, \, \, \, \sum_{b_y} P_y(b_y|\kappa) = 1 \, \, \forall y, \kappa. \end{equation} \section{Incompatibility is necessary for quantum advantage in communication tasks}\label{sec3} Now, we will show that incompatible measurements are necessary for showing quantum advantage in any communication task. Before proceeding, let us briefly describe a generic communication scenario consisting of two players - Alice and Bob. Alice and Bob are given inputs $x\in [l]$ and $y\in[n]$, respectively. Further, initially neither player has any idea about the other player's input. Alice, upon receiving the input $x$ sends a $d$-dimensional classical or quantum system to Bob. Bob, upon receiving the input $y$ and the message (which is $d$-dimensional classical or quantum system) sent by Alice, outputs $b_y\in [d_y]$. The outcome of this communication task is determined by the set of probabilities distributions $\{p(b_y |x,y)\}$. In classical communication, they can use pre-shared randomness $\lambda$, and therefore, the any typical probability can be expressed as \begin{equation} \label{pc} p(b_y|x,y) = \sum_{m=1}^d \int_\lambda \pi(\lambda) p_a(m|x,\lambda)p_b(b_y|y,m,\lambda)\ d\lambda . \end{equation} Here $\{p_a(m|x,\lambda)\},\{p_b(b_y|y,m,\lambda)\}$ are encoding and decoding functions by Alice and Bob, satisfying non-negativity and \begin{equation} \sum_m p_a(m|x,\lambda) = \sum_{b_y} p_b(b_y|y,m,\lambda) =1 . \end{equation} While in quantum communication \begin{equation} \label{pq} p(b_y|x,y) = \text{Tr} (\rho_xM_{b_y|y} ), \quad \rho_x, M_{b_y|y} \in \mathcal{B}(\mathbbm{C}^d) \end{equation} Here $\mathcal{B}(\mathbbm{C}^d)$ stands for the space of all operators acting on $d$ dimensional complex Hilbert space. Let $l, n, d_y$ be some natural numbers. Given the scenario $x \in [l]$, $y \in [n]$, $b_y \in [d_y]$, we define the set all probabilities obtainable by $d$-dimensional classical communication \begin{equation} \mathcal{C}_d := \{p(b_y|x,y)\} \end{equation} where $p(b_y|x,y)$ is given by \eqref{pc}, and the set of all probabilities in $d$-dimensional quantum communication \begin{equation} \mathcal{Q}_d := \{p(b_y|x,y)\} \end{equation} where $p(b_y|x,y)$ is given by \eqref{pq}. We are interested in another set of probabilities, \begin{equation} \mathcal{Q}^C_d := \{p(b_y|x,y)\} \end{equation} where $p(b_y|x,y)$ is given by \eqref{pq} such that the set of measurements acting on $d$-dimensional quantum states used by Bob $\{M_{b_y|y}\}$ is compatible according to \eqref{com}. \begin{result} Given any scenario, \begin{equation} \mathcal{Q}^C_d \subseteq \mathcal{C}_d , \end{equation} that is, measurement incompatibility is necessary for any advantage over classical communication. However, measurement incompatibility is not sufficient for quantum advantage. \end{result} \begin{proof} Consider the case where Bob performs a single POVM measurement $\{G_\kappa\}$, which is the parent POVM of the measurement set $\{M_{b_y|y}\}$. Thanks to the Frenkel-Weiner theorem \cite{Frenkel2015}, which implies that the set of input-output probabilities $p(\kappa|x)$ with a single quantum measurement on $d$-dimensional quantum states can be always be reproduced by a suitable classical $d$-dimensional communication in the presence of shared randomness. That is, $\forall \rho_x,$ there exists classical strategy $\pi(\lambda),p_a(m|x,\lambda),p_b(\kappa|m,\lambda)$ such that \begin{equation} \label{rxgk} \text{Tr} (\rho_x G_\kappa) = \sum_{m=1}^d \int_\lambda \pi(\lambda) p_a(m|x,\lambda)p_b(\kappa|m,\lambda) \ d\lambda . \end{equation} Here note that the scenario considered by Frenkel-Weiner \cite{Frenkel2015} is a bit different since Bob does not receive any input $y$ therein. That is why Bob's output $\kappa$ depends only on the message $m$ sent by Alice and classical shared randomness $\lambda$. On the other hand, Bob receives the message $m$ as well as an input $y$ in the aforementioned communication scenario. Hence, in the communication task, Bob's output $b_y$ depends on $m$, $y$ and classical shared randomness $\lambda$ (see Eq.(\ref{pc})). Let us now focus on the aforementioned communication scenario and take into account the following decoding function, \begin{equation} \label{newde} p_b(b_y|y,m,\lambda) = \sum_\kappa P_y(b_y|\kappa) p_b(\kappa|m,\lambda) , \end{equation} where $\{P_y(b_y|\kappa)\}$ is the post-processing to obtain $M_{b_y|y}$ from the parent POVM defined in \eqref{com}. One can check that this is indeed a valid decoding function. Next, we show that an arbitrary $p(b_y|x,y) \in \mathcal{Q}^C_d$ in the communication scenario can always be reproduced by a suitable classical strategy involving the decoding function \eqref{newde}. An arbitrary $p(b_y|x,y) \in \mathcal{Q}^C_d$ can always be expressed as $p(b_y|x,y)= \text{Tr} (\rho_x M_{b_y|y})$, where $M_{b_y|y}$ satisfies \eqref{com}. Now, with the help of \eqref{rxgk}, one can show the following, \begin{eqnarray} && \text{Tr} (\rho_x M_{b_y|y}) \nonumber \\ &&= \sum_\kappa P_y(b_y|\kappa) \text{Tr} (\rho_x G_\kappa ) \nonumber \\ &&= \sum_\kappa P_y(b_y|\kappa) \left( \sum_{m} \int_\lambda \pi(\lambda) p_a(m|x,\lambda) p_b(\kappa|m,\lambda) d\lambda \right) \nonumber \\ &&= \sum_{m} \int_\lambda \pi(\lambda) p_a(m|x,\lambda) \left( \sum_\kappa P_y(b_y|\kappa) p_b(\kappa|m,\lambda) \right) d\lambda \nonumber \\ &&= \sum_{m} \int_\lambda \pi(\lambda) p_a(m|x,\lambda) p_b(b_y|y,m,\lambda) \ d\lambda . \end{eqnarray} Therefore, an arbitrary probability distribution $p(b_y|x,y)$ obtainable from compatible set of measurement can be reproduced by a suitable classical strategy, inferring that $\mathcal{Q}^C_d \subseteq \mathcal{C}_d$. Since the figure of merit of any communication task is some arbitrary functions of the probabilities $p(b_y|x,y)$, we can infer that any advantage in such tasks over classical communication can be attained only if the set measurements $\{E_y \equiv \{M_{b_y|y} \}_{b_y} \}_y$ is incompatible. Finally, we note that there exists incompatible qubit measurements such that the probabilities obtained from them for arbitrary quantum states are within $\mathcal{C}_2$ (see Section IV-A of \cite{PRXQuantum2021}). This completes the proof. \end{proof} We are often interested in linear functions of $\{p(b_y|x,y)\}$ due to their practical importance in quantum communication complexity tasks \cite{RevModPhys.82.665,DEWOLF,saha2019}, quantum key distribution \cite{Marcin2011}, quantum randomness generation \cite{randomness,randomness1}, quantum random access codes \cite{brac,Carmeli2020}, oblivious transfer \cite{Spekkens2009,saha2019} and many other applications. To find the optimum value of any linear function of $\{p(b_y|x,y)\}$, it is sufficient to consider classical strategy without shared randomness (see \textit{Lemma} 1 in the Appendix \ref{app1} for detailed explanation). As a consequence, all probability distributions $\{p(b_y|x,y)\}$, that are obtained from classical strategy without shared randomness, can always be reproduced by the following quantum strategy. Upon receiving the input $x$, Alice sends the quantum state $\rho_x$ such that $\rho_x$ is diagonal in some basis. Bob, upon receiving the input $y$ and the state $\rho_x$, performs a fixed measurement $\{G_\kappa\}$, which is independent of $y$ and nothing but the measurement in that basis, followed by some post-processing depending on $y$. Therefore, we have another useful result. \begin{result} The maximum value of any linear function of $\{p(b_y|x,y)\}$ obtained within the two sets $\mathcal{C}_d$ and $\mathcal{Q}^C_d$ is same. \label{result2} \end{result} Above results have profound implications in practice. As a consequence of them we are able to conclude that any arbitrary communication task can serve as a witness of measurement incompatibility. Next, we will propose incompatibility witness for an arbitrary set of measurements for a family of communication tasks, namely, the general version of random access codes \cite{racambainis}. \begin{figure} \centering \includegraphics[scale=0.47]{gqrac.pdf} \caption{An unknown measurement set of arbitrary settings, $\{M_{b_{y}|y}\}_{b_y,y}$, is provided; we only know the dimension ($d$) on which this set of measurements act. Our task is to certify the incompatibility of this set of measurements. Here, the notation $[k]:= \{1,\cdots,k\}$ for any natural number $k$.}\label{fig1} \end{figure} \section{Incompatibility witness for sets of measurements of arbitrary setting} \label{sec4} Take the most general form of a set of measurements. There are $n$ measurements, defined by $\{M_{b_y|y}\}$ where $y\in [n]$ each of which has different outcomes, say, measurement $y$ has $d_y$ outcomes, that is, $b_y \in [d_y]$, and these measurement are acting on $d$-dimensional quantum states (see FIG. \ref{fig1}). In order to witness incompatibility of this set we introduce the most general form of random access codes with Bob having this set of measurements. Alice gets a string of $n$ dits $x=x_1x_2\cdots x_n$ randomly from the set of all possible strings in which $x_y \in [d_y]$ for all $y \in [n]$. While Alice communicates a $d$-dimensional classical or quantum system to encode the information about obtained string, the task for Bob is to guess the $y$-th dit when $y$ is chosen randomly. The figure of merit is the average success probability defined by the following linear function \begin{equation} \label{Pndd} S(n,\Vec{d},d) = \frac{1}{n\prod_y d_y} \sum_{x,y} p(b_y=x_y|x,y) \end{equation} that is fully specified by $n$, $\Vec{d}=(d_1d_2\cdots d_n)$, and $d$. Since this \eqref{Pndd} is a linear function of $p(b_y|x,y)$, by \textit{Result} \ref{result2}, the maximum value over $\mathcal{C}_d$ and $\mathcal{Q}^C_d$ is the same and denoted by $S^C(n,\Vec{d},d)$. Precisely, \begin{align} S^C(n,\Vec{d},d) &= \max_{\{p(b_y|x,y)\} \in \mathcal{C}_d} S(n,\Vec{d},d) \nonumber \\ &= \max_{\{p(b_y|x,y)\} \in \mathcal{Q}^C_d} S(n,\Vec{d},d). \end{align} Hence, $S^C(n,\Vec{d},d)$ can be evaluated by maximizing the average success probability either over all classical strategies, or over all quantum strategies involving compatible measurements only. Whenever a set of measurements in the scenario specified by $n$, $\Vec{d}$, $d$ gives $S(n,\Vec{d},d) > S^C(n,\Vec{d},d)$ in above-introduced general version of the random access codes, we can conclude that the measurements are incompatible. Hence, in order to witness measurement incompatibility, we need to know $S^C(n,\Vec{d},d)$. Now we present an upper bound on $S^C(n,\Vec{d},d)$ for arbitrary $n,\Vec{d},d$. \begin{result}\label{thm:gracb} The following relation holds true for arbitrary $n,\Vec{d},d,$ \begin{equation} \label{SC1} S^C(n,\Vec{d},d) \leqslant \frac{1}{n} \times \min \Bigg\{ 1 + \sum_{\substack{i,j\\i<j}} \frac{d}{ d_id_j } \ , \ n-1+\frac{d}{\prod_y d_y} \Bigg\} . \end{equation} \end{result} This upper bound in Eq. \eqref{SC1} is obtained for $\mathcal{Q}^C_d$, that is, by taking the existence of a parent POVM of the measurements $\{M_{b_y|y}\}_{b_y,y}$ performed by Bob. The proof of this result is presented in the Appendix \ref{app1}. When the outcome of all the measurements are same, which is $d_y = \Tilde{d} $ for all $y$, the above bound simplifies to \begin{equation} S^C(n,\Tilde{d},d) \leqslant \frac{1}{n} \times \min \left\{1 + \frac{n(n-1)d}{2\Tilde{d}^{2}} \ , \ n-1 +\frac{d}{\Tilde{d}^n} \right\} . \end{equation} Hence, in different types of RAC involving an arbitrary set of quantum measurements by Bob, if the average success probability exceeds the aforementioned upper bounds on $S^C$, then we can conclude that the measurements by Bob are incompatible. On the other hand, whenever \begin{equation} \label{ddy} d \leqslant \min_y d_y \end{equation} we find out the exact value of $S^C(n,\Vec{d},d)$. Say, $k_i$ is the number of sets among $[d_1],\cdots,[d_n]$ such that dit $i\in [d_y]$. For example, consider the random access codes with $n=4$ and $d_1=2$, $d_2=3$, $d_3=4$ and $d_4 = 3$. That is, Alice gets a string of four dits $x=x_1 x_2 x_3 x_4$ randomly, where $x_1 \in [2]$, $x_2 \in [3]$, $x_3 \in [4]$ and $x_4 \in [3]$. In this case, $k_1=4$, $k_2=4$, $k_3=3$, $k_4=1$. Also, we denote $d_{\max} = \max_y d_y $. \begin{result}\label{thm:graceb} If \eqref{ddy} holds then \begin{equation} \label{SCex} S^C(n,\Vec{d},d) = \frac{1}{n\prod_y d_y} \sum \left[ \left(\prod_{j=1}^{d_{\max}} C^{\alpha_j}_{n_j} \right)\, \max_{i=1,\cdots,d}\{n_i\} \right] \end{equation} with \begin{align} \alpha_j = k_j - \sum_{i=j+1}^{d_{\max}} n_i, \quad C^{\alpha_j}_{n_j} = \frac{\alpha_j (\alpha_j-1) \cdots (\alpha_j - n_j+1) }{n_j (n_j-1) \cdots 1} \nonumber \end{align} and where the summation is taken over all possible integer solutions of the following equation \begin{equation} \sum_{i=1}^{d_{\max}} n_i = n \end{equation} such that $n_i \leqslant k_i$ for all $i$, \end{result} Note here that (\ref{SCex}) is obtained for $\mathcal{C}_d$ by considering the classical strategies. The detailed proof is given in the Appendix \ref{app2}. For a particular case of {\it Result} \ref{thm:graceb} wherein $d_y = \Tilde{d} = d$ for all $y$, the proof is previously given in \cite{brac}. Hence, when $d \leqslant \min_y d_y$, the necessary criteria for a set of measurements to be compatible is given by, \begin{align} \label{witness-ineq} S(n,\Vec{d},d) \leqslant S^C(n,\Vec{d},d), \end{align} where $S^C(n,\Vec{d},d)$ is given by (\ref{SCex}). For $n=2$, $d_y=\Tilde{d}$ for all $y$, and $d\leqslant \Tilde{d}$, the expression \eqref{SCex} simplifies to (for details, see the Appendix \ref{app3}) \begin{equation} \label{simSCex} S^C(2,\Tilde{d},d) = \frac{1}{2\Tilde{d}^2} \left(d + 2d\Tilde{d} -d^2 \right) . \end{equation} And for $n=3$, $d_y=\Tilde{d}$ for all $y$, and $d\leqslant \Tilde{d}$, the expression \eqref{SCex} simplifies to (for details, see the Appendix \ref{app3}) \begin{equation} \label{3_measurements} S^C(3,\Tilde{d},d) = \frac{d}{3\Tilde{d}^3} \left(d^2 - 1 + 3\Tilde{d} (\Tilde{d} + 1 - d) \right) . \end{equation} The particular case of \textit{Result} \ref{thm:gracb} for $n=2$ can be found in \cite{Carmeli2020}, and moreover, it is shown that any pair of rank-one projective measurements that are incompatible provides advantage in RAC \cite{saha2020}. In order to showcase the generic applicability of {\it Results} \ref{thm:gracb}-\ref{thm:graceb}, we consider an arbitrary set of three rank-one projective qubit measurements, which using the freedom of unitary can be expressed as \begin{eqnarray} \label{3qpm} M_{x_1|1} &=& (1/2) U \left[\mathbbm{1} + (-1)^{x_1} \sigma_z \right] U^\dagger \nonumber \\ M_{x_2|2} &=& (1/2) U \left[\mathbbm{1} + (-1)^{x_2} \left( \alpha \sigma_z + \sqrt{1-\alpha^2} \sigma_x \right) \right] U^\dagger \nonumber \\ M_{x_3|3} &=& (1/2) U \Big[\mathbbm{1} + (-1)^{x_3} \big( \beta \sigma_z + \gamma \sqrt{1-\beta^2} \sigma_x \nonumber \\ && \qquad \qquad \quad \pm \sqrt{1-\beta^2} \sqrt{1-\gamma^2} \sigma_y \big) \Big] U^\dagger \end{eqnarray} where $x,x_2,x_3 \in [2],$ the variables $\alpha,\beta,\gamma \in [-1,1]$, and $U$ can be arbitrary unitary operator acting on $\mathbbm{C}^2$. We obtain the following result. \begin{result}\label{thm:322} Any set of three incompatible rank-one projective qubit measurements, except for the sets defined by (\ref{3qpm}) with \begin{equation} (\alpha,\beta,\gamma ) = \{ (\pm 1/2,\pm 1/2,-1),(\pm 1/2,\mp 1/2,1)\} \nonumber \end{equation} and arbitrary $U$, yields larger value than $S^C(n=3,\Tilde{d} =2,d=2)=3/4$. \end{result} This result is proved with the help of numerical optimizations and the proof is put over to Appendix \ref{app:322}. \section{Conclusion}\label{sec5} By characterizing the set of quantum correlations in prepare-and-measure scenarios produced from any set of compatible measurements, we have shown in this article that incompatible measurements at the receiver's end is necessary for demonstrating quantum advantage in any one-way communication task. Further, based on this result, we have presented a semi-device independent witness of measurement incompatibility invoking generalized random access codes. Interestingly, we have completely characterized the sets of three incompatible projective qubit measurements that can be detected using our proposed witness. It might be noted that some of the results derived in \cite{Carmeli2020,saha2020} appear as natural corollaries of the results obtained here. The significance of the result presented here lies in the fact that the classical bound of the success metric of any one-way communication task becomes an upper bound on the metric of the task under compatible set of measurements. Consequently, violating the classical bound of any one-way communication task can be used as a sufficient criteria to witness measurement incompatibility. Further, the present study establishes that measurement incompatibility is the fundamental quantum resource for non-classicality in any one-way communication task or, more generally, in prepare-and-measure scenarios. Our study opens up the possibilities of several open questions. First of all, deriving more efficient incompatibility witnesses based on different communication tasks is worth for future studies. Secondly, our results may be generalized to propose semi-device witnesses for incompatible quantum channels \cite{doi:10.1063/1.5126496} and quantum instruments \cite{IncomQI,PhysRevA.105.052202}. Though we have proved that $\mathcal{Q}^C_d$ is a subset of $\mathcal{C}_d$ for any $d$, we strongly anticipate that $\mathcal{Q}^C_d$ is in fact a strict subset of $\mathcal{C}_d$. It needs further investigation to prove this. Finally, proposing operational witnesses for all incompatible extremal POVM \cite{Sent_s_2013} is another fundamentally motivated open problem. \begin{acknowledgements} DS and DD acknowledges Science and Engineering Research Board (SERB), Government of India for financial support through National Post Doctoral Fellowship (PDF/2020/001682 and PDF/2020/001358, respectively). During the later phase of this work, the research of DD has been supported by the Royal Society (United Kingdom) through the Newton International Fellowship (NIF$\backslash$R$1\backslash212007$). BB acknowledges INSPIRE fellowship from the Department of Science and Technology, Government of India. AKD and ASM acknowledge support from Project No. DST/ICPS/QuEST/2018/98 from the Department of Science and Technology, Govt. of India. \end{acknowledgements}

8ca717b4475b3837d5512f4445d328341c2cca63

\section*{References}} \usepackage{graphicx} \graphicspath{ {figures/} } \usepackage{amsmath,amssymb} \DeclareMathOperator*{\argmax}{arg\,max} \DeclareMathOperator*{\argmin}{arg\,min} \DeclareMathOperator*{\calL}{\mathcal{L}} \DeclareMathOperator*{\rhovec}{\bm{\rho}} \title{Bayesian Neural Network Versus Ex-Post Calibration For Prediction Uncertainty} \author{% Satya Borgohain \\ Monash University \\ \texttt{[email protected]} \\ \And Klaus Ackermann \\ Monash University \\ \texttt{[email protected]} \\ \AND Ruben Loaiza-Maya \\ Monash University \\ \texttt{[email protected]} \\ } \begin{document} \maketitle \begin{abstract} Probabilistic predictions from neural networks which account for predictive uncertainty during classification is crucial in many real-world and high-impact decision making settings. However, in practice most datasets are trained on non-probabilistic neural networks which by default do not capture this inherent uncertainty. This well-known problem has led to the development of post-hoc calibration procedures, such as Platt scaling (logistic), isotonic and beta calibration, which transforms the scores into well calibrated empirical probabilities. A plausible alternative to the calibration approach is to use Bayesian neural networks, which directly models a predictive distribution. Although they have been applied to images and text datasets, they have seen limited adoption in the tabular and small data regime. In this paper, we demonstrate that Bayesian neural networks yields competitive performance when compared to calibrated neural networks and conduct experiments across a wide array of datasets. \end{abstract} \section{Introduction}\label{sec:intro} Obtaining well calibrated estimates of probabilities is crucial, particularly in domains where decision making could have direct consequences on human life such as medical diagnosis \cite{Huang2020} and credit approval \cite{beque2017approaches} to name a few. In such scenarios, it is often not enough for a model to be accurate but it also needs to capture and quantify the degree of uncertainty with which it makes such decisions. However, many classification tasks still primarily rely on the classifier's error rate (or accuracy) as the key metric for its selection and deployment in real world use-cases which could lead to over/under confident predictions. Coupled with class imbalance it poses challenges that given a set of features, simply predicting class membership does not help us understand. Raw probabilities estimates provide a more complete picture of the underlying decision making process by the classifiers and alternative diagnostic metrics such as AUC-ROC, precision, recall and F1 certainly help in that regard \cite{kuhn2013applied}. Post-hoc calibration methods are one such approach which helps match the predicted probabilities with the expected class distribution of the target. They take the output of any model and map the score to the empirical probabilities \cite{Kull2017}. Fairness and bias is another closely related field in machine learning and an arena that has received much attention recently. Some of the methods predominantly used there involve thresholding in order to de-bias the learned model against certain classes. However, finding optimal thresholds without optimization are only possible if the classifiers are well-calibrated \cite{Kull2017}. Bayesian neural networks (BNN) have recently emerged as an alternative approach to calibration methods \cite{kingma2013auto}. However, adding Bayesian layers to neural networks alone, does not solve the problem of not receiving well calibrated output as shown in \cite{hortua2020parameter}. A BNN creates a probabilistic model by linking the neural net to the conditional distribution of the outcome of interest. Thus, BNNs directly allow the researcher to measure aleatoric uncertainty without the need for any post-hoc processing steps. Because BNNs are probabilistic models with high-dimensional parameters spaces they are estimated using approximate Bayesian methods such as variational inference \citep{kingma2013auto}. The computation of a posterior distribution implies that BNNs also have the ability to capture epistemic uncertainty. Therefore, unlike calibration methods, the coherent probabilistic nature of BNNs implies that both aleatoric and epistemic uncertainty are considered in the production of out-of-sample predictions. Furthermore, methods such as SWA-Gaussian \cite{maddox2019simple} which tries to approximate the posterior over the parameters using information inherent in the trajectory of SGD have provably shown that Bayesian methods do work well with neural networks to provide well-calibrated probabilities. Platt scaling has also seen much adoption in large neural nets \cite{guo2017calibration}. Despite the modelling benefits of BNNs, their uptake in machine learning has been relatively slow. One of the main reasons is that, to our knowledge, no comprehensive studies have been undertaken to demonstrate the effectiveness of BNNs over calibration approaches, particularly for tabular and smaller datasets. The main purpose of this paper is to fill this gap in the literature. Although there have been notable developments using Bayesian approaches in the field, they usually involve non-tabular datasets with a large sample size. Using a subset of the rich amalgam of datasets considered in \cite{Kull2017}, we investigate if BNNs outperform post-hoc calibration methods in terms of probability calibration. The key contributions of the paper can thus be summarized as follows: \begin{itemize} \item We illustrate that BNNs are a plausible alternative to obtaining well-calibrated empirical probabilities for classification. \item We specifically demonstrate the applicability of BNNs in tabular, real-world, small data regime. \end{itemize} \section{Bayesian neural networks} \begin{figure} \centering \includegraphics[width=\textwidth]{figures/toy_example.pdf} \caption{BNN predictive results for simulation example. Panel (a) displays the true classes, with red corresponding to observation with $y_i=1$ and blue to observations with $y_i=0$. Panel (b) presents the BNN point predictions $\hat{y}_i$, with the same color scheme as panel (a). Panel (c) reports the classification probabilities $\text{Pr}(Y_i=1|\bm{x}_i,\bm{\theta})$.} \label{Fig:toyexample} \end{figure} \subsection{The model} \label{sec:model} Denote as $\bm{y} = \left(y_1,\dots,y_N\right)^\top$ a vector of $N$ realizations of the binary variable $y_i\in\{0,1\}$, as $\bm{x}_i=\left(x_{1,i},\dots,x_{p,i}\right)^\top$ a vector of $p$ covariates with explanatory power on $y_i$, and set $\bm{x} = \left(\bm{x}_1^\top,\dots,\bm{x}_N^\top\right)^\top$. A Bayesian neural network (BNN) is a probabilistic model that links a neural network function $f_{NN}\left(\bm{x}_i,\bm{\theta}\right)$ to the conditional distribution function of $y_i$ (see \cite{mullachery2018bayesian} for an overview on BNNs). For the binary outcomes considered in this paper, we link $f_{NN}\left(\bm{x}_i,\bm{\theta}\right)$ to the conditional probability distribution of $y_i$ via the logistic function $g\left(a\right) = \frac{1}{1+e^{-a}}$, so that \begin{equation}\label{EQ1} p\left(y_i|\bm{x}_i,\bm{\theta}\right) = g\left[f_{NN}\left(\bm{x}_i,\bm{\theta}\right)\right]^{I\left(y_i=1\right)}\left\{1-g\left[f_{NN}\left(\bm{x}_i,\bm{\theta}\right)\right]\right\}^{I\left(y_i=0\right)}, \end{equation} where $I\left(.\right)$ denotes an indicator function that is equal to one if its argument is true, and zero otherwise. Because a BNN fully characterizes the conditional distribution $p\left(y_i|\bm{x}_i,\bm{\theta}\right)$, it has the ability to both, produce point classification predictions as $\hat{y}_i = \argmax_{Y\in\{0,1\}} p(Y|\bm{\theta},\bm{x}_i)$, and also capture the level of classification uncertainty over those predictions as $\text{Pr}(\hat{y}_i = y_i|\bm{\theta},\bm{x}_i)$. Using the assumption that the elements in $\bm{y}$ are conditionally independent, we can then express the likelihood function for the probabilistic neural network as \begin{align*} p\left(\bm{y}|\bm{\theta},\bm{x}\right) =& \prod_{i=1}^{N}p\left(y_i|\bm{x}_i,\bm{\theta}\right). \end{align*} The main challenge in using a BNN is inference. The parameter vector $\bm{\theta}$ has generally thousands of elements, which often makes exact Bayesian estimation infeasible. In the following section we discuss how approximate Bayesian estimation of BNNs can be applied instead. \subsection{Variational inference} In Bayesian estimation the density of interest is that of the parameters of the neural network conditional on the data. This density is denoted here by $p\left(\bm{\theta}|\bm{y}\right)\propto p\left(\bm{y}|\bm{\theta}\right)p\left(\bm{\theta}\right)$, where $p\left(\bm{\theta}\right)$ is the prior density and $\bm{x}$ is dropped for ease of notation . Because of the high-dimensionality of $\bm{\theta}$, exact Bayesian estimation methods are computationally costly, and as such not practical for the problem at hand. Instead, we resort to variational inference (VI) methods, where a density $q_\lambda(\bm{\theta})$ - member of some parametric family of densities - is used to approximate $p(\bm{\theta}|\bm{y})$, and where $\bm{\lambda}$ is a vector of parameters known as variational parameters (see for instance \cite{blei2017variational} and \cite{kingma2013auto}). Variational inference can then be described as an optimization problem, where the aim is to minimize the Kullback-Leibler divergence between $q_\lambda(\bm{\theta})$ and $p(\bm{\theta}|\bm{y})$ with respect to $\bm{\lambda}$, defined as \begin{align*} \text{KL}(q_\lambda(\bm{\theta})||p(\bm{\theta}|\bm{y}) ) & = E_{q_\lambda}\left[ \log \frac{q_\lambda(\bm{\theta})}{p(\bm{\theta}|\bm{y})}\right]. \end{align*} This divergence can be re-written as $\text{KL}(q_\lambda(\bm{\theta}) ||p(\bm{\theta}|\bm{y}) ) = \log p(\bm{y})-\calL(\bm{\lambda}) \label{kldexpression} $, with $p(\bm{y})=\int p(\bm{\theta})p(\bm{y}|\bm{\theta}) d\bm{\theta}$, $h\left(\bm{\theta}\right)=p\left(\bm{y}|\bm{\theta}\right)p\left(\bm{\theta}\right)$ and where $\calL(\bm{\lambda})$ is known as the Evidence Lower Bound (ELBO), given as \begin{equation} \calL(\bm{\lambda})=E_{q_\lambda}\left[\log h(\bm{\theta}) - \log q_\lambda(\bm{\theta})\right]\ . \label{eq:lowerbound} \end{equation} Because $\log p(\bm{y})$ does not depend on $\bm{\lambda}$, minimization of the Kullback-Leibler divergence with respect to $\bm{\lambda}$ is equivalent to maximizing the ELBO; however, ELBO optimization is more computationally feasible as it does not require evaluation of the intractable term $\log p(\bm{y})$. Stochastic gradient ascent methods (SGA) can be applied for maximization of the ELBO, by first setting an initial value $\bm{\lambda}^{(0)}$ for $\bm{\lambda}$, and then sequentially iterating over the expression \begin{align*} \bm{\lambda}^{(i+1)} & = \bm{\lambda}^{(i)}+\bm{\rho}_i \circ \widehat{\nabla_\lambda \calL(\bm{\lambda}^{(i)})}, \;\mbox{ for } i=1,2,\ldots\,, \end{align*} where $\rhovec_i=(\rho_{i1},\dots, \rho_{im})^\top$ denotes a vector of learning rates, `$\circ$' denotes the element-wise product of two vectors, and $\widehat{\nabla_\lambda \calL(\bm{\lambda}^{(i)})}$ is an unbiased estimate of the gradient of $\calL(\bm{\lambda})$ evaluated at $\bm{\lambda}=\bm{\lambda}^{(i)}$. Here, the learning rates are set according to the ADAM method proposed by \cite{kingma2014adam}. The selection of an unbiased and low variance estimate of the ELBO gradient is key to the success of VI. We follow \cite{kingma2013auto} and employ the so called ``reparametrization trick'' . This approach requires a generative formula $\bm{\theta}=k(\bm{\varepsilon},\bm{\lambda})$ from the specific approximating density $q_\lambda$, where the vector $\bm{\varepsilon}$ has density $f_\varepsilon$ which does not depend on $\bm{\lambda}$. The reparametrization trick then allows to re-write the ELBO as \begin{align} \calL(\bm{\lambda}) & = E_{f_\varepsilon}\left[\log h(k(\bm{\varepsilon},\bm{\lambda}))-\log q_\lambda(k(\bm{\varepsilon},\bm{\lambda}))\right]\,. \label{lbdrepar} \end{align} By differentiating (\ref{lbdrepar}), the ELBO gradient can be expressed as \begin{align} \nabla_\lambda \calL(\bm{\lambda}) & = E_{f_\varepsilon}\left[\frac{\partial\bm{\theta}}{\partial\bm{\lambda}}^\top \left\{\nabla_{\theta} \log h(\bm{\theta})-\nabla_{\theta}\log q_\lambda(\bm{\theta})\right\}\right]\,. \label{lbdgradexpr} \end{align} Then, an unbiased low variance estimate of the ELBO gradient can be computed by estimating the expectation in (\ref{lbdgradexpr}) using one random draw from $f_\varepsilon$. The last component of VI is the selection of an adequate approximating family. We follow \cite{ong2018gaussian} and employ a Gaussian approximation with a factor covariance structure, so that \begin{equation} q_\lambda(\bm{\theta})=\phi_{m}\left(\bm{\theta};\bm{\mu},BB'+D^2\right)\,, \label{eq:q} \end{equation} where $\phi_{m}\left(\bm{x};\bm{\mu},\Sigma\right)$, denotes the $m-$variate Gaussian density with mean $\bm{\mu}$ and covariance $\Sigma$, $D$ is a diagonal matrix with diagonal elements $\bm{d} = \left(d_1,\dots,d_m\right)^\top$, $B$ is an $m\times K$ matrix, $K<m$ denotes the number of factors, and $\bm{\lambda} = (\bm{\mu}^\top,\bm{d}^\top,\text{vech}(B)^\top)^\top$, where $\text{vech}$ denotes the half-vectorization operator of a rectangular matrix. With this approximation, the generative formula needed for the reparametrization trick is $\bm{\theta}=k(\bm{\varepsilon},\bm{\lambda}) = \bm{\mu}+B\bm{z}+D\bm{\eta}$, where $\bm{\varepsilon} = \left(\bm{z}^\top,\bm{\eta}^\top\right)^\top$. The terms needed to perform SGA are $\frac{\partial\bm{\theta}}{\partial\bm{\lambda}}$, $\nabla_{\theta}\log q_\lambda(\bm{\theta})$ and $\nabla_{\theta} \log h(\bm{\theta})$. The first two terms were provided in \cite{ong2018gaussian} for the family of approximations used here. The third term can be re-written as \begin{align}\label{Eq:logpost_gradient} \nabla_{\theta} \log h(\bm{\theta}) = \nabla_{\theta} \log p\left(\bm{y}|\bm{\theta}\right) + \nabla_{\theta} \log p\left(\bm{\theta}\right). \end{align} We employ uniform priors for $\bm{\theta}$, which implies that $\nabla_{\theta} \log p\left(\bm{\theta}\right)=\bm{0}$. The remaining term, $\nabla_{\theta} \log p\left(\bm{y}|\bm{\theta}\right)$, can be computed as \begin{align*} \nabla_\theta\log p\left(\bm{y}|\bm{\theta}\right) = &\sum_{\left\{i:y_i=1\right\}}\nabla_\theta f_{NN}\left(\bm{x}_i,\bm{\theta}\right)\\ &-\sum_{i}g\left[f_{NN}\left(\bm{x}_i,\bm{\theta}\right)\right]\nabla_\theta f_{NN}\left(\bm{x}_i,\bm{\theta}\right), \end{align*} where $\nabla_\theta f_{NN}\left(\bm{x}_i,\bm{\theta}\right)$ is the gradient of the neural network function, which can be evaluated using any readily available back propagation algorithm. \subsection{Toy example} To provide some intuition about how the BNN in Section~\ref{sec:model} can be used in practice, we applied it to a simple simulated data set. We started by generating $10000$ realizations of the i.i.d covariates $x_{i,1}\sim N(0,1)$ and $x_{i,2}\sim N(0,1)$. Subsequently, we generated the corresponding $10000$ realizations of $y_i$ from the true data generating process (DGP) \begin{equation}\label{Eq:trueDGP} y_i = I(x_{i,1}<t(x_{i,2})+\epsilon_i), \end{equation} where $\epsilon_i\sim N(0,1)$ is an i.i.d Gaussian error, while the function $t()$ is the Yeo-Johnson transformation \citep{yeo2000new} with parameter $-1$, whose role is to induce non-linearity in the way the two covariates determine $y_{i}$. In order to learn the true DGP in \eqref{Eq:trueDGP}, we applied the BNN to a random subset of $8000$ observations, and used the remaining $2000$ observations to produce out-of-sample predictions. For the choice of $f_{NN}\left(\bm{x}_i,\bm{\theta}\right)$ we used a feedforward neural network with two nodes and three layers. Panel (a) in Figure~\ref{Fig:toyexample} presents the true binary categories of the out-of-sample points as a function of the covariates. The red dots indicate the observations for which $y_i = 1$, while the blue dots display the observations for which $y_i=0$. Although there is some overlap in the classification regions, the red points are concentrated in the top-left quadrant while the blue dots are concentrated in the bottom-right. Panel (b) in Figure~\ref{Fig:toyexample} presents the point predictions $\hat{y}_i$ from the neural network. The point predictions indicate that there is a clear separation threshold. However, we know from panel (a) that the closer the observations are to the threshold, the more classification overlap will be observed and thus the higher the classification uncertainty. Via the computation of the probabilities $\text{Pr}(Y_i=1|\bm{x}_i,\bm{\theta})$, the BNN naturally allow us to measure the classification uncertainty. These probabilities are presented in panel (c) in Figure~\ref{Fig:toyexample}, where the color scale of the dots indicates the classification uncertainty, with green dots indicating higher uncertainty ($\text{Pr}(Y_i=1|\bm{x}_i,\bm{\theta}) \approx 0.5$). The plot indicates that observations located in the top-left corner are classified as $\hat{y}_i=1$ with high certainty as $\text{Pr}(Y_i=1|\bm{x}_i,\bm{\theta}) \approx 1$. Similarly, observations located in the bottom-right corner are classified as $\hat{y}_i=0$ with high certainty as $\text{Pr}(Y_i=1|\bm{x}_i,\bm{\theta}) \approx 0$. Finally, as the observations get closer to the threshold in the middle, the classification uncertainty increases, which indicates the BNN is able to identify the high probability of class overlap in that region. Figure~\ref{fig:reliability-diagram} shows the reliability diagram for the same highlighting different calibration methods which we discuss in Section~\ref{sec:calibration}. \begin{figure}\ \centering \includegraphics[scale=0.45]{figures/calibration-curve.pdf} \caption{Reliability diagram for the Toy dataset. Also known as Calibration curves, they help visually determine the degree of calibration in probabilistic forecast between different models. Here the x-axis represents the predicted probability and y-axis represents the observed relative frequency of membership to class \cite{brocker2007increasing}. The closer the curves are to perfectly calibrated line (diagonal) the better. The log-losses are denoted in the legend beside each method. We observe here that BNN performs better than the other methods.}. \label{fig:reliability-diagram} \end{figure} \section{Post-Hoc probability calibration} \label{sec:calibration} We draw upon the seminal work by \cite{Kull2017} to compare our findings with the recent advances in the literature and follow their benchmark methodology. We benchmark our approach to well established calibration procedures in the literature. The goal of any calibration method is given the output score $s$ of a classifier, to correctly represent the empirical probabilities of a given example belonging to a class. These empirical probabilities can be visualised using reliability diagrams shown in figure \ref{fig:reliability-diagram}. For a given score range from $0.9-1$ on the test set, 90 to 100 percent of the examples should be members of the class. \paragraph{Isotonic calibration} is a non parametric method that uses the ranking of the output of any classifier to assign bins for mapping from score to probability based on how well a the classifier has ranked the examples \cite{10.1145/775047.775151,fawcett2007pav}. This approach requires enough samples for any given bin, which for small datasets causes this procedure to overfit\cite{Kull2017}. \paragraph{Logistic calibration} was introduced by \cite{platt2000probabilities} for support vector machines. It takes the form of $ \mu(s,\gamma,\delta)=\frac{1}{1+1/exp(\gamma \cdot s + \delta)}$, where $\gamma$ and $\delta$ are real valued parameters. This procedure is commonly referred to as Platt scaling. \cite{Kull2017} showed that this procedure has the tendency to lead to worse results after calibration, compared to only relying on the raw output of a given classifier. \paragraph{Beta calibration} was introduced by \cite{Kull2017}, where the underlying modeling assumption is the beta distribution, a probability distribution bound to the interval $[0,1]$. In general, the beta distribution is used for modeling the behaviour of proportions or percentages, and therefore suited to model the score distribution for a given class. Negative log-likelihood (NLL), also referred to as cross-entropy or log-loss, gives us a measure of how well the methods compare. It penalizes predictions that appear to be more confident than they are. In line with the literature \cite{Kull2017}, we use log-loss as our key metric. A measure of accuracy would require the setting of a threshold and might therefore mask the true probabilities of predictions. \section{Experiments} In order to minimize inductive biases imparted by any complex architectural choice, we consider a simple feedforward neural network consisting of just $2$ hidden layers with $4$ ReLU units each. We use ADAM \cite{kingma2014adam} for parameter optimization in both the networks with a global learning rate of $10^{-3}$ or $10^{-2}$ and the following hyperparameters for the same: $\beta_{1} = 0.9$, $\beta_{2} = 0.999$ and $\epsilon = 10^{-7}$. Minimal hyperparameter optimization was done as we mostly selected the default values\footnote{The computation was performed under Ubuntu 18.04 using a Intel(R) Xeon(R) W-2145 CPU @ 3.70GHz and a NVIDIA Quadro P6000 as GPU}. We devise the following strategy to train both the networks: First, we randomly split the dataset between training ($80\%$) and test ($20\%$) followed by further splitting the train set into validation and calibration sets respectively. For each set, we perform stratified sampling to preserve the original class distribution and standardize the feature matrix with the mean and standard deviation of the train set (avoiding data leakage). During the validation stage, we monitor the cross entropy loss in order to find the optimal number of epochs ($\tau$), which correspond to the minimum validation loss, for each model. Thereafter, we train both the models to $\tau$ and use the base neural network to further calibrate its output probabilities with the post-hoc methods as per Section~\ref{sec:calibration}. We also observed from our experiments that on an average, the standard feedforward neural network needed relatively fewer epochs to converge when compared to the BNN and set their maximum number of epochs during the validation stage to be approximately half that of BNN. We note that early stopping as a regularizing strategy proved to be difficult as the loss curves (particularly for BNNs) sometimes exhibit the double descent phenomena \cite{nakkiran2019deep} to varying degrees. Furthermore, we employ a similar strategy as \cite{Kull2017} and binarize the target variable, considering the majority class as $1$ and the rest to be $0$, in order to convert from a multi-class to binary classification setting. We also note that calibrated neural networks are exposed to roughly $\sim20\%$ more data points than BNN due to their use of calibration set. This also illustrates that BNNs can still work remarkably well given fewer data points than its counterparts. We implement our code using TensorFlow with Python 3.6 to conduct the experiments. In particular, we make use of the fast tensor computation for the Jacobian of errors accessed via the gradient tape on GPU. \subsection{Datasets} We evaluate the calibration of all the models on $20$ well-known datasets from the UCI Machine Learning Repository \cite{Dua2019}. Here we consider datasets representing domains such as financial, medical, social, just to name a few, as listed under Table~\ref{uci-datasets-table}. \begin{table} \caption{UCI datasets for used in the experiments.} \label{uci-datasets-table} \centering \begin{tabular}{lrrll} \toprule Name & Features & Samples & Attribute Type(s) & Domain \\ \midrule Abalone & 8 & 4177 & Categorical, Integer, Real & Life (Marine) \\ Balance Scale & 4 & 625 & Categorical & Social \\ Credit Approval & 15 & 653 & Categorical, Integer, Real & Financial \\ German Credit & 20 & 1000 & Categorical, Integer & Financial \\ Ionosphere & 34 & 351 & Integer, Real & Physical \\ Image Segmentation & 19 & 2310 & Real & N/A \\ Landsat Satellite & 36 & 6435 & Integer & Physical \\ Letter Recognition & 16 & 35000 & Integer & Computer \\ Mfeat (Karhunen) & 64 & 2000 & Integer, Real & Computer \\ Mfeat (Morphological) & 6 & 2000 & Integer, Real & Computer \\ Mfeat (Zernike) & 47 & 2000 & Integer, Real & Computer \\ Mushroom & 22 & 8124 & Categorical & Life \\ Optical Digits Recognition & 64 & 5620 & Integer & Computer \\ Page Blocks & 10 & 5473 & Integer, Real & Computer \\ Spambase & 57 & 4601 & Integer, Real & Computer \\ Vehicle & 18 & 946 & Integer & N/A \\ Waveform-5000 & 40 & 5000 & Real & Physical \\ WDBC & 30 & 569 & Real & Life (Cancer) \\ WPBC & 33 & 194 & Real & Life (Cancer) \\ Yeast & 8 & 1484 & Real & Life (Proteins) \\ \bottomrule \end{tabular} \end{table} Neural networks typically require training samples orders of magnitude more than those of the UCI datasets. Although theoretically, small sample sizes makes the network prone to overfitting, we empirically did not observe any drastic effect on their performances in part due to the simple architecture with a relatively small number of parameters. We also note that due to the stochastic nature of our learning algorithms they are sensitive to initialization of the parameters. \subsection{Evaluation} We observe that BNN yields competitive performance (and on average outperforms) when compared to the other calibration methods across the datasets as outlined in Table~\ref{tab:results}. As expected, BNN also performs better than the uncalibrated vanilla neural network in most cases. Additionally, we also track other metrics such as Brier score and Expected Calibration Error (ECE) with a bin size of $10$. Brier score provides the mean squared error for a probabilistic forecast and is given by $\frac{1}{N}\sum_{i=1}^{N}(f_i - o_i)$, where $f_i$ are probabilities and $o_i$ are observed classes. ECE is a popular binning based approach to measuring calibration error which is quite sensitive to the choice of bins and ultimately not as strongly reliable \cite{naeini2015obtaining}. Figure~\ref{Fig:predictions} illustrates the performance of each model on two of the datasets. Interestingly, we also notice for a few datasets that calibrating with the post-hoc methods sometimes leads to further miscalibration and higher log-loss. \begin{figure} \centering \includegraphics[width=\textwidth]{figures/test-predictions.pdf} \caption{Predictions on the Test set for Balance Scale and Mushroom Datasets. The colour scale represents probabilities between $0-1$. Here we perform PCA on the original feature space and plot on the two principle axes.} \label{Fig:predictions} \end{figure} \begin{table} \caption{Log-loss for each dataset. The best method is marked in bold for each row.} \label{tab:results} \centering \begin{tabular}{lrrrrr} \toprule {} & \multicolumn{5}{c}{Methods} \\ \cmidrule(r){2-6} & Uncalibrated & Beta & Isotonic & Logistic & Var Bayes \\ Dataset & & & & & \\ \midrule Abalone & \textbf{0.597890} & 0.598148 & 0.691808 & 0.602227 & 0.615603 \\ Balance Scale & 0.098068 & 0.080771 & 0.092153 & 0.089347 & \textbf{0.069333} \\ Credit Approval & 0.491384 & 0.522863 & 0.897678 & \textbf{0.437511} & 0.457618 \\ German Credit & 0.525405 & \textbf{0.517368} & 0.532982 & 0.517461 & 0.565519 \\ Ionosphere & 0.385311 & 0.310725 & 0.334547 & \textbf{0.308752} & 0.330256 \\ Image Segmentation & 0.410117 & 0.410120 & 0.410120 & 0.410120 & \textbf{0.012053} \\ Landsat Satellite & 0.549391 & 0.549391 & 0.549391 & 0.549391 & \textbf{0.065981} \\ Letter Recognition & 0.015198 & 0.015066 & 0.015491 & 0.022324 & \textbf{0.012077} \\ Mfeat (Karhunen) & 0.056259 & \textbf{0.052115} & 0.165397 & 0.121016 & 0.067857 \\ Mfeat (Morphological) & 0.325083 & 0.325096 & 0.325096 & 0.325096 & \textbf{0.000206} \\ Mfeat (Zernike) & 0.092498 & 0.070683 & 0.166497 & \textbf{0.048465} & 0.073144 \\ Mushroom & 0.035914 & 0.037741 & 0.037741 & 0.037741 & \textbf{0.033184} \\ Optical Digits Recognition & 0.086171 & 0.056051 & 0.086846 & 0.063662 & \textbf{0.047854} \\ Page Blocks & \textbf{0.071376} & 0.072884 & 0.091231 & 0.082372 & 0.085377 \\ Spambase & 0.225907 & \textbf{0.211673} & 0.212747 & 0.213790 & 0.235523 \\ Toy & 0.376422 & 0.377564 & 0.401984 & 0.391937 & \textbf{0.375313} \\ Vehicle & 0.107625 & 0.331832 & 0.310363 & 0.390933 & \textbf{0.091739} \\ Waveform-5000 & 0.239015 & \textbf{0.234735} & 0.240671 & 0.265519 & 0.238321 \\ WDBC & 0.080406 & 0.164898 & 0.156011 & 0.176233 & \textbf{0.070994} \\ WPBC & 0.584846 & 0.614227 & 1.777795 & 0.594910 & \textbf{0.541680} \\ Yeast & 0.594895 & 0.583243 & 0.806812 & \textbf{0.561925} & 0.569832 \\ \midrule Rank & 2.7619 & 2.6667 & 4.2857 & 3.1905 & \textbf{2.0952} \\ \bottomrule \end{tabular} \end{table} To establish whether the differences between the methods are statistically significant, we follow \cite{demvsar2006statistical} and perform Friedman test across all datasets with the null hypothesis ($H_0$) as there being no significant differences in performance between the classifiers. Considering a significance level of $0.05$, we reject the null with a p-value of $0.000119$ along with a test statistic of $23.13$ and perform a post-hoc analysis based on Wilcoxon-Holm test to further analyse their pairwise differences. Figure~\ref{Fig:critical-diff} illustrates the critical difference diagram with pairwise significance. Here we observe that BNN (Var Bayes) have the highest rank as compared to the other methods. However, we also note that pairwise differences between BNN, Uncalibrated, Beta and Logistic are not as significant as that between BNN and Isotonic. Overall BNNs provide competitive, if not better, performance as its calibrated counterparts. Interestingly, what \cite{Kull2017} observed for Beta calibration when applied to Adaboost and Naive Bayes classifiers does not perfectly hold for neural nets as evident from our experiments. \begin{figure} \centering \includegraphics[scale=0.5]{CD_diagram-crop.pdf} \caption{Critical difference diagram for log-loss along with ranks for each method.} \label{Fig:critical-diff} \end{figure} \section{Conclusion} In this paper, we propose the use of BNNs as a plausible alternative to obtain well-calibrated probabilities when compared with post-hoc calibration methods such as platt scaling, isotonic and beta calibration for tabular datasets. We performed extensive experiments on 20 datasets using the same neural network architecture as our underlying model. With the the normal neural network implementation we applied the common calibration techniques. Our results show that our proposed method works provably well for tabular datasets with small sample size. We made use of recent advances in fast gradient calculation within TensorFlow framework to calculate the full Jacobian matrix with respect to the parameters, as required for Variational Bayes Inference. Albeit, certain datasets required a larger number of epochs for convergence in the Variational framework. Future research will investigate the impact of different priors such as Horseshoe on the parameters and its impact on convergence and calibration. \section{Ethical and societal implication} BNNs could be applied across many domains where neural networks are starting to be heavily utilized in real world settings and strict $0$ or $1$ outcomes are not desired. Although for most machine learning researchers it is clear that output scores obtained via softmax, for example, do not represent true probabilities, many practitioners interpret the same as otherwise. This becomes especially critical, when decision threshold are set, that have major consequences for an individual. In areas such as, bail or no bail, treatment with some medicine or not, having wrongly calibrated models could lead to detrimental or harmful outcomes \cite{rudin2019stop}. Thresholding is also used to try to make machine learning predictions more fair, by setting different thresholds based on protected attributes such as race or gender \cite{hardt2016equality}. Again in such applications, it is highly important that the thresholds represent actual probabilities rather than a ranking. Similarly, in recent years governments have started to scarcely assign resources to its citizen based on algorithmic decisions to make the most efficient use of their limited resources. Hence it becomes increasingly important that resource allocation prioritizes those who truly need such aid. Alternatively, having well calibrated neural networks could also lead to overconfidence and over reliance on algorithms by decision and policy makers. We would encourage further work of the relationship of calibrated threshold in real world scenarios to fully understand the impact miscalibration has. Nevertheless, we hope our work helps counter these issues and ultimately creates a positive social impact across domains.

a0f8ab7c3dc1886c565ed76afe76422ffac87f8e

\section{Introduction} The Circular Electron Positron Collider (CEPC) \cite{CEPCStudyGroup:2018ghi} is a large-scale collider facility proposed after the discovery of Higgs boson in 2012. It is designed with a circumference of $100\,$km with two interaction points. It can operate at multiple center-of-mass energies, including $240\,$GeV as a Higgs factory, $160\,$GeV for the $W^+W^-$ threshold scan, $91\,GeV$ as a Z factory, and $360\,$GeV for the $t\bar{t}$ threshold scan. Table~\ref{CECPope} \cite{CEPCPhysicsStudyGroup:2022uwl} summarizes its baseline operating scheme and the corresponding boson yields. In the future, it can be upgraded to a proton-proton collider to directly scan new physics signals at a center-of-mass energy of about $100\,$TeV, which is an order of magnitude higher than the LHC. \begin{table}[htbp] \centering \caption{\label{CECPope}The operation scheme of the CEPC, including the center-of-mass energy, the instantaneous luminosity, the total integrated luminosity, and the event yields.} \smallskip \begin{tabular}{ccccc} \hline Operation mode & Z factory & WW & Higgs factory & $t\bar{t}$ \\ \hline $\sqrt{s}$ (GeV) & 91.2 & 160 & 240 & 360 \\ Run time (year) & 2 & 1 & 10 & 5 \\ Instantaneous luminosity & \multirow{2}{*}{191.7} & \multirow{2}{*}{26.6} & \multirow{2}{*}{8.3} & \multirow{2}{*}{0.83} \\ ($10^{34}cm^{-2} s^{-1}$, per IP) & & & & \\ Integrated luminosity & \multirow{2}{*}{100} & \multirow{2}{*}{6} & \multirow{2}{*}{20} & \multirow{2}{*}{1} \\ ($ab^{-1}$, 2 IPs) & & & & \\ Event yields & $3\times 10^{12}$ & $1\times 10^8$ & $4\times 10^6$ & $5 \times 10^{5}$ \\ \hline \end{tabular} \end{table} The main scientific objective of the CEPC is the precise measurement of the Higgs properties. When CEPC operates as a Z-boson factory, trillions of generated $Z\ensuremath{\rightarrow}\xspace q\bar{q}$ events can provide a great opportunity for measuring flavor physics, where particle identification (PID) is essential. The baseline CEPC detector uses a TPC as the main tracker, which can record the dE/dx or even dN/dx \cite{Cuna:2021sho, Chiarello:2019eny} information from tracks. In addition, the CEPC detector is proposed to have TOF capability with cluster-level precision of $50\,$ps. The TOF and dE/dx information depend on the particle species and its momentum, and enable {\ensuremath{\kaon^\pm}}\xspace/{\ensuremath{\pion^\pm}}\xspace/\kern 0.18em \optbar{\kern -0.18em p}{}\xspace identification. In this paper, the dE/dx performance of the CEPC TPC is investigated based on Monte Carlo (MC) simulation, which induces the intrinsic dE/dx resolution. In real experiments, both detector effects and imperfect calibration can deteriorate the dE/dx resolution. Based on the potential degradation in previous TPCs by comparing their experimental achievements with the corresponding intrinsic dE/dx resolutions obtained from MC simulation, we quantify the performance requirements of dE/dx using full simulated Z-pole samples corresponding to the Z-pole operation of the CEPC. This article is organized as follows. Section~\ref{Det} introduces the detector, software, and samples used in this analysis. Section~\ref{sec:PID} describes the measurement of TOF and dE/dx and investigates the {\ensuremath{\kaon^\pm}}\xspace/{\ensuremath{\pion^\pm}}\xspace/\kern 0.18em \optbar{\kern -0.18em p}{}\xspace separation performance. Section~\ref{sec:K} presents the performance of $K^{\pm}$ identification, $D^0\ensuremath{\rightarrow}\xspace \pi^+K^-$ and $\phi\ensuremath{\rightarrow}\xspace K^+K^-$ reconstruction efficiency/purity using the inclusive hadronic Z-pole samples. A brief conclusion is given in Section~\ref{sec:Con}. \section{The reference detector, software, and samples} \label{Det} The layout of the baseline CEPC detector design is shown in Fig.~\ref{det}. It is designed following the particle flow principle \cite{arbor}, which emphasizes on the separation of final state particles and measures each final state particle in the most suited subdetector. From innermost part of the detector layout to outermost part, the baseline concept consists of a silicon pixel vertex detector, a silicon inner tracker, a TPC surrounded by a silicon external tracker, a silicon-tungsten sampling ECAL, a steel-Glass Resistive Plate Chambers sampling Hadronic Calorimeter (HCAL), a 3-Tesla superconducting solenoid, and a flux return yoke embedded with a muon detector. The PID performance in this paper is derived by dE/dx and TOF information collected by TPC and ECAL of baseline CEPC detector. The TOF information recorded by ECAL can reach a time resolution of $50\, $ps \cite{TOF}, which can be achieved with modern silicon sensing technologies. Figure~\ref{TPC} shows the structure of the TPC, which is located between the inner and outer silicon trackers and contains 220 radial layers with $6\, $mm increments. This TPC has a cylindrical drift volume that provides a very homogeneous electric field. The drift volume is filled with gases combining Ar/CF$_4$/C$_4$H$_{10}$ at atmospheric pressure with ratios of 95\%/3\%/2\%. A detailed description of the TPC can be found in Ref.~\cite{CEPCStudyGroup:2018ghi}. \begin{figure}[htbp] \centering \includegraphics[width=.55\textwidth]{Det.png} \caption{\label{det}The structure of the baseline CEPC detector design.} \end{figure} \begin{figure}[htbp] \centering \includegraphics[width=.5\textwidth]{TPC.png} \caption{\label{TPC}The structure of the TPC in the CEPC baseline detector design. This TPC is a cylindrical gas detector with an axial electric field formed between the end-plates (yellow) and a central cathode plane/membrane (light blue). The cylindrical walls of the volume form the electric field cage (dark blue). Gas ionization electrons due to charged particles drift to the end-plates where they are collected by readout modules (yellow).} \end{figure} A baseline reconstruction software chain has been developed to evaluate the physics performance of the CEPC baseline detector. The data flow of CEPC baseline software starts from the event generators of Whizard~\cite{Stienemeier:2021cse, Moretti:2001zz} and Pythia~\cite{Bierlich:2022pfr}. The detector geometry is implemented into the MokkaPlus, a GEANT4~\cite{geant4} based full simulation module. The MokkaPlus calculates the energy deposition in the detector sensitive volumes and creates simulated hits. For each sub-detector, the digitization module converts the simulated hits into digitized hits by convoluting the corresponding sub-detector responses. The reconstruction modules include tracking, Particle Flow, and high-level reconstruction algorithms. The digitized tracker hits are reconstructed into tracks via the tracking algorithm. The Particle Flow algorithm, Arbor, reads the reconstructed tracks and the calorimeter hits to build particle events. High-level reconstruction algorithms reconstruct composite physics objects and identify the flavor of the jets. Using the CEPC baseline detector geometry and software chain, we simulated $6.313\times 10^{6}$ inclusive hadronic Z-pole events. These events include all the different quark flavors according to the SM predicts, whose detail are shown in Table~\ref{zpole}. \begin{table}[htbp] \centering \caption{\label{zpole}The branching ratio and the number of simulated events of the Z-pole samples.} \smallskip \begin{tabular}{cc|c} \hline Process & $\mathcal{B}$ & Sample used \\ \hline $Z\ensuremath{\rightarrow}\xspace u\overline{u}$ & 11.17\% & \multirow{5}{*}{ $6.313\times 10^{6}$}\\ \cline{1-2} $Z\ensuremath{\rightarrow}\xspace d\overline{d}$ & 15.84\% & \\ \cline{1-2} $Z\ensuremath{\rightarrow}\xspace s\overline{s}$ & 15.84\% & \\ \cline{1-2} $Z\ensuremath{\rightarrow}\xspace c\overline{c}$ & 12.03\% & \\ \cline{1-2} $Z\ensuremath{\rightarrow}\xspace b\overline{b}$ & 15.12\% & \\ \hline \end{tabular} \end{table} \section{PID performance using TOF and dE/dx} \label{sec:PID} This section presents the measurements of TOF and dE/dx information by ECAL and TPC, respectively, and the corresponding {\ensuremath{\kaon^\pm}}\xspace/{\ensuremath{\pion^\pm}}\xspace/\kern 0.18em \optbar{\kern -0.18em p}{}\xspace separation performance. The separation power is defined as \begin{equation} S_{AB} = \frac{|O_A - O_B|}{\sqrt{\sigma_{A}^2 + \sigma_{B}^2}}, \end{equation} where $O_{A(B)}$ is the TOF or dE/dx of particle A(B) and $\sigma_{A(B)}$ is the TOF resolution or dE/dx resolution of particle A(B). There are also other ways to define the separation power using different conventions. To be consistent with the definition in the previous paper~\cite{An:2018jtk}, we use the definition above. \subsection{PID performance using TOF} The TOF of a given particle can be calculated based on its momentum and its flight distance along its trajectory. In magnetic field, the trajectory of a charged particle would be a helix that can interact with the barrel region or the end-cap region of ECAL, depending on the four-momentum of the incident particle. The inner radius of ECAL is represented by R and the radius of the helix is derived as \begin{equation} \label{eq:tof} \begin{split} r = \frac{1000\cdot P_t}{0.3\cdot B\cdot q} (mm), \end{split} \end{equation} where $P_t$, $q$, and B are the transverse momentum, the charge of the incident particle, and the magnetic field strength in the TPC, respectively. Once the incident particle interacts with the barrel of the ECAL, the trajectory of the incident particle can be described by $arc = 2\pi r \cdot \frac{\phi}{2\pi}$, and the corresponding arc angle is represented by $\phi = 2 \cdot arcsin(\frac{R}{2r})$. The velocity of the particle perpendicular/parallel to the B-field is represented by $v_t$/$v_l$, and the length between the Interaction Point (IP) and the ECAL end cap is represented by $L$. At the CEPC, $L = 2.35\, $m and $R = 1.8\, $m. If $2\cdot r$ $>$ R \& $L/v_l > arc/v_t$, the helix would interact with the barrel of ECAL, $TOF = arc/v_t$. Otherwise, the helix would interact with the end cap of ECAL, $TOF = L/v_l$. \begin{figure}[htbp] \centering \includegraphics[width=.5\textwidth]{c2DTOF.png} \caption{\label{2DTOF} The $K^{\pm}/\pi^{\pm}$ separation power as a function of momentum and cosine polar angle with TOF information.} \end{figure} The $K^{\pm}/\pi^{\pm}$ separation power as a function of momentum and polar angle $\theta$ is shown in Fig.~\ref{2DTOF}. The separation power is peaking around $cos\theta = 0.8$ corresponding to the maximum drift length. \subsection{PID performance using dE/dx} The measured dE/dx of a track follow a Landau distribution with a large tail caused by high-energy $\delta$-electrons. The average dE/dx for a track, denoted $I$, is estimated using the usual ``truncated mean" method, where the truncation ratio of 90\% is determined to obtain the optimal dE/dx resolution, denoted $\sigma_I$. For {\ensuremath{\kaon^\pm}}\xspace/{\ensuremath{\pion^\pm}}\xspace/\kern 0.18em \optbar{\kern -0.18em p}{}\xspace, the distributions of $I$ as a function of momentum are shown in the left plot of Fig.~\ref{dEdx}. The distributions of the absolute difference of $I$ for {\ensuremath{\kaon^\pm}}\xspace/{\ensuremath{\pion^\pm}}\xspace and {\ensuremath{\kaon^\pm}}\xspace/\kern 0.18em \optbar{\kern -0.18em p}{}\xspace are shown in the right plot of Fig.~\ref{dEdx}. The $K^{\pm}$ and $\pi^{\pm}$ with the same dE/dx at $1\, $GeV/c ($K^{\pm}$ and \kern 0.18em \optbar{\kern -0.18em p}{}\xspace at $2.5\, $GeV/c) have none separation power. \begin{figure}[htbp] \centering \includegraphics[width=.45\textwidth]{cdEdx_KPi.png} \hfill \includegraphics[width=.45\textwidth]{cDeltadEdx_kpi.png} \caption{\label{dEdx}The distribution of $I$ as a function of momentum for {\ensuremath{\kaon^\pm}}\xspace/{\ensuremath{\pion^\pm}}\xspace/\kern 0.18em \optbar{\kern -0.18em p}{}\xspace is shown in the left plot and the absolute difference of $I$ for {\ensuremath{\kaon^\pm}}\xspace/{\ensuremath{\pion^\pm}}\xspace and {\ensuremath{\kaon^\pm}}\xspace/\kern 0.18em \optbar{\kern -0.18em p}{}\xspace is shown in the right plot.} \end{figure} \begin{figure}[htbp] \includegraphics[width=.32\textwidth]{dedx2Npad.png} \hfill \includegraphics[width=.32\textwidth]{dedx2PadH.png} \hfill \includegraphics[width=.32\textwidth]{dedx2rho.png} \\ \includegraphics[width=.32\textwidth]{dedx2bg.png} \hfill \includegraphics[width=.32\textwidth]{dedx2Cos.png} \caption{\label{differencial}The intrinsic dE/dx resolution versus the number of pad rings (a), the pad height along the radial direction (b), the ratio of gas density $\rho$ over the default gas density $\rho_0$ (equivalent to the ratio of corresponding pressures) (c), the relativistic velocity $\beta\gamma$ (d) and $cos\theta$ (e) of the ionizing particle. The default working point is indicated with a solid star symbol. Solid lines represent the fit projections.} \end{figure} The predictive power of the dE/dx method depends on precise knowledge of the dE/dx energy loss parameterization and its resolution. The resolution induced by the design of the TPC and the momentum of the incident particle, including the number of pad rings $n$, the pad height along the radial direction $h$, the density of the working gas $\rho$, the relativistic velocity $\beta\gamma$, and the polar angle $\theta$ of the particle trajectory, is named ``intrinsic dE/dx resolution". While in real experiments, the dE/dx resolution could be deteriorated by the detector effects arising in the processes of electron drift, signal amplification and readout in TPC. We name the dE/dx resolution in real experiments as ``actual dE/dx resolution". The differential of the dE/dx resolution, denoted $\sigma_I/I$, to the parameters of the TPC and incident particle is shown in Fig.~\ref{differencial}. The dE/dx resolution is a power function of $n$/$h$/$\rho$/$\beta\gamma$, while dE/dx resolution is a constant in the barrel region and becomes better with the increase of particle drift distance. However, due to the relatively low physics performance of end cap, the dE/dx resolution would become worse in the end-cap region. It can be seen that dE/dx resolution for {\ensuremath{\kaon^\pm}}\xspace/{\ensuremath{\pion^\pm}}\xspace/\kern 0.18em \optbar{\kern -0.18em p}{}\xspace can be better than 2.5\% when $\beta\gamma > 10$ in the barrel region. At the CEPC, dE/dx resolution is pursued to be less than 3\%, corresponding to 20\% degradation compared to the largest intrinsic dE/dx resolution in the barrel region. The parameterization of dE/dx resolution can be factorized as \begin{equation} \label{reso} \begin{split} \frac{\sigma_I}{I} = \frac{13.5}{n^{0.5}\cdot(h\rho)^{0.3}}\left[2.05 + 0.8(\beta\gamma)^{-0.3}\right] \\ \times \left[ 2.5 - 1.5(cos\Theta)^4 + 3.9(cos\Theta)^{10}\right], \end{split} \end{equation} where $h$ and $\rho$ are in mm and mg/cm$^3$, respectively. The distribution of dE/dx resolution as a function of momentum and polar angle for {\ensuremath{\kaon^\pm}}\xspace/{\ensuremath{\pion^\pm}}\xspace/\kern 0.18em \optbar{\kern -0.18em p}{}\xspace is shown in Fig.~\ref{resokaon}. The separation power as a function of momentum and polar angle is shown in Fig.~\ref{cAn} for {\ensuremath{\kaon^\pm}}\xspace/{\ensuremath{\pion^\pm}}\xspace (left plot) and {\ensuremath{\kaon^\pm}}\xspace/\kern 0.18em \optbar{\kern -0.18em p}{}\xspace (right plot). The regions with separation power higher than 3 are shown in warm colors. \begin{figure}[htbp] \centering \includegraphics[width=.45\textwidth]{cKsigma.png} \\ \includegraphics[width=.45\textwidth]{cPisigma.png} \hfill \includegraphics[width=.45\textwidth]{cPsigma.png} \caption{\label{resokaon}The distributions of dE/dx resolution as a function of momentum and cosine polar angle for $K^{\pm}$ (top), $\pi^{\pm}$ (down-left), and \kern 0.18em \optbar{\kern -0.18em p}{}\xspace (down-right).} \end{figure} \begin{figure}[htbp] \centering \includegraphics[width=.45\textwidth]{piK.png} \hfill \includegraphics[width=.45\textwidth]{pK.png} \caption{\label{cAn}Separation power as a function of momentum and cosine polar angle for {\ensuremath{\kaon^\pm}}\xspace/{\ensuremath{\pion^\pm}}\xspace (left) and {\ensuremath{\kaon^\pm}}\xspace/\kern 0.18em \optbar{\kern -0.18em p}{}\xspace (right).} \end{figure} \begin{figure*}[htbp] \centering \includegraphics[width=.45\textwidth]{piKcom.png} \hfill \includegraphics[width=.45\textwidth]{pKcom.png} \caption{\label{ccom}Combine dE/dx and TOF, separation power as a function of momentum and cosine polar angle for {\ensuremath{\kaon^\pm}}\xspace/{\ensuremath{\pion^\pm}}\xspace (left) and {\ensuremath{\kaon^\pm}}\xspace/\kern 0.18em \optbar{\kern -0.18em p}{}\xspace (right). \ifcomment {\color{red} 建议在图上画出3$\sigma$的边界线，或者在图像颜色选取时让3$\sigma$之上和之下的区域颜色有明显区分。} \fi } \end{figure*} \ifcomment {\color{red} 后文结论中 $\sigma_I/I$ better than 3\%, corresponds to a degradation of dE/dx resolution less then 20\% 的对应关系如何得到？} \fi \subsection{PID performance combining both} The above analyses investigate the separation power of {\ensuremath{\kaon^\pm}}\xspace/{\ensuremath{\pion^\pm}}\xspace/\kern 0.18em \optbar{\kern -0.18em p}{}\xspace with TOF and dE/dx information independently. The TOF information can enhance the separation power of dE/dx information at about $1\, $GeV/c for {\ensuremath{\kaon^\pm}}\xspace/{\ensuremath{\pion^\pm}}\xspace separation and about $2.5\, $GeV/c for {\ensuremath{\kaon^\pm}}\xspace/\kern 0.18em \optbar{\kern -0.18em p}{}\xspace separation. After combining dE/dx and TOF, the separation power as a function of momentum and polar angle for {\ensuremath{\kaon^\pm}}\xspace/{\ensuremath{\pion^\pm}}\xspace and {\ensuremath{\kaon^\pm}}\xspace/\kern 0.18em \optbar{\kern -0.18em p}{}\xspace is shown in Fig.~\ref{ccom}. The regions with separation power higher than 3 are shown in warm colors. \section{PID evaluation} \label{sec:K} Since the separation power is convention dependent, we can also use the maximal efficiency times purity for {\ensuremath{\kaon^\pm}}\xspace identification at a given sample, inclusive hadronic Z-pole samples detailed in Table~\ref{zpole}, to study the PID performance of the baseline CEPC detector. In addition, the identification efficiency and purity of any particles that decays into charged hadrons can also be used to quantify the PID performance. Therefore, we extract the $D^0\ensuremath{\rightarrow}\xspace \pi^+K^-$ and $\phi\ensuremath{\rightarrow}\xspace K^+K^-$ reconstruction efficiency/purity using inclusive hadronic Z-pole samples. \subsection{$K^{\pm}$ identification at the Z-pole} \begin{figure}[htbp] \centering \includegraphics[width=.45\textwidth]{1088.png} \hfill \includegraphics[width=.45\textwidth]{ceffpur.png} \caption{\label{sep28}By combining dE/dx and TOF information with intrinsic dE/dx resolution and $50\, $ps TOF resolution, the distribution of variable $(R - R_K)/\sigma_R$ for sample with {\ensuremath{\kaon^\pm}}\xspace/{\ensuremath{\pion^\pm}}\xspace/\kern 0.18em \optbar{\kern -0.18em p}{}\xspace at a momentum of $12.2\, $GeV/c and $cos\theta = 0.36$ is shown as the left plot, and the performance of $K^{\pm}$ identification with maximal efficiency times purity at different momentum values and cosine polar angles is shown as the right plot. } \end{figure} As demonstrated in Fig.~\ref{ccom}, the separation power has a strong dependence on the polar angle and momentum of particles. Therefore, for a given reference of momentum and polar angle range, we can optimize the cuts to select candidate with maximal efficiency times purity. For example, the left plot in Fig.~\ref{sep28} shows the distribution of the variable $(R - R_K)/\sigma_R$, where $R$ is the experimental measurement either by dE/dx alone or by the combination of dE/dx and TOF, $R_k$ is the expected value for the $K^{\pm}$ hypothesis, and $\sigma_R$ denotes the experimental resolution for {\ensuremath{\kaon^\pm}}\xspace/{\ensuremath{\pion^\pm}}\xspace/\kern 0.18em \optbar{\kern -0.18em p}{}\xspace in a given range of momentum and polar angle. The two vertical lines indicate $K^{\pm}$ selection with maximal efficiency times purity. The $K^{\pm}$ identification efficiency and purity are defined as \begin{equation} \begin{split} \epsilon_K = \frac{N_{K\ensuremath{\rightarrow}\xspace K}}{N_K} \\ p_K = \frac{N_{K\ensuremath{\rightarrow}\xspace K}}{N_{K\ensuremath{\rightarrow}\xspace K} + N_{\pi \ensuremath{\rightarrow}\xspace K} + N_{p\ensuremath{\rightarrow}\xspace K}}, \end{split} \end{equation} where $N_K$ is the total number of $K^{\pm}$ produced, $N_{K\ensuremath{\rightarrow}\xspace K}$ is the number of $K^{\pm}$ correctly identified, and $N_{\pi(p)\ensuremath{\rightarrow}\xspace K}$ is the number of $\pi^{\pm}$(\kern 0.18em \optbar{\kern -0.18em p}{}\xspace) incorrectly identified as $K^{\pm}$. By combining dE/dx and TOF with intrinsic dE/dx resolution and $50\, $ps TOF resolution, the right plot in Fig.~\ref{sep28} shows the maximal efficiency times purity of $K^{\pm}$ identification at different ranges of momentum and polar angle. The overall efficiency and purity of $K^{\pm}$ identification is calculated as \begin{equation} \begin{split} overall\ efficiency = \frac{\sum_{bins} \epsilon_k \cdot N_k}{\sum_{bins} N_k} \\ overall\ purity = \frac{\sum_{bins} p_k \cdot N_k}{\sum_{bins} N_k}, \end{split} \end{equation} where $\sum_{bin}$ means looping over all momentum and polar angle ranges, $\epsilon_k$, $p_k$ and $N_k$ represents the $K^{\pm}$ identification efficiency, purity and the number of $K^{\pm}$, respectively, in a given momentum and polar angle range. With above convention, the overall efficiency and purity in Fig.~\ref{sep28} is 98.43\% and 97.89\%, respectively. In reality, the dE/dx resolution is affected by detector effects and readout electronics. With the definition $\sigma_{actual} = factor \cdot \sigma_{intrinsic}$, where $\sigma_{actual}$ is the actual dE/dx resolution and $\sigma_{intrinsic}$ is the intrinsic dE/dx resolution, the performance of $K^{\pm}$ identification under different factors with/without combining TOF information are shown in Table~\ref{kaonID}. The factors are selected according to the Table 1 in Ref.~\cite{An:2018jtk}, which shows that several detectors of other high energy experiments have the actual dE/dx resolution decreased by 20\% or 50\% related to intrinsic ones. The results in Table~\ref{kaonID} prove that worse dE/dx resolution leads to worse $K^{\pm}$ identification performance, while TOF information can improve $K^{\pm}$ identification performance on top of dE/dx information. \begin{table}[htbp] \centering \caption{\label{kaonID}The $K^{\pm}$ identification performance with different factors, $\sigma_{actual} = factor \cdot \sigma_{intrinsic}$, with/without combination of TOF information at the Z-pole.} \smallskip \begin{tabular}{cccccc} \hline & factor & 1. & 1.2 &1.5 & 2. \\ \hline \multirow{2}{*}{dE/dx} & $\varepsilon_K$ (\%) & 95.97 & 94.09 & 91.19 & 87.09 \\ & $purity_K$ (\%) & 81.56 & 78.17 & 71.85 & 61.28 \\ \hline \multirow{2}{*}{dE/dx \& TOF} & $\varepsilon_K$ (\%) & 98.43 & 97.41 & 95.52 & 92.3 \\ & $purity_K$ (\%) & 97.89 & 96.31 & 93.25 & 87.33 \\ \hline \end{tabular} \end{table} \subsection{ $D^0\ensuremath{\rightarrow}\xspace \pi^+K^-$ reconstruction at the Z-pole} \label{secD0} $D^0$ mesons can be abundantly produced in the CEPC, especially in the Z-pole operating mode, and $D^0$ mesons are important research object for the study of CP violation. It is important to understand the reconstruction performance of $D^0$ in the baseline design of the CEPC detector. Using the inclusive $Z\ensuremath{\rightarrow}\xspace q\overline{q}$ sample and under the condition of intrinsic dE/dx resolution, the reconstruction performance of $D^0\ensuremath{\rightarrow}\xspace \pi^+K^-$ in the baseline CEPC detector is investigated in this subsection. The method for reconstructing the $D^0\ensuremath{\rightarrow}\xspace \pi^+K^-$ is described below. \begin{enumerate} \item Match all pairs of reconstructed oppositely charged tracks. \item To suppress the significant combinatorial backgrounds, the absolute mass difference between two oppositely charged tracks and $D^0$ is set to less than $10\, $MeV/c$^2$. \item The impact parameter (IMP) of each track is larger than $0.02\,$ mm$^2$. \item The $\chi^2$ of the secondary vertex fit \cite{Suehara:2015ura} is less than 5.15. \item The distance of the secondary vertex to IP is greater than $0.305\, $mm. \item The identified PID of a positively charged particle is a $\pi^+$ and a negative $K^-$. \end{enumerate} The efficiency and purity of the candidate $D^0$ selection are defined in Eq.~\ref{eq:effpur}. The parameters of the above procedure are set to maximize the value of the efficiency times the purity of the signal selection. The details of event selection can be found in Table~\ref{D0}. \begin{equation} \label{eq:effpur} \begin{split} \epsilon & = \frac{Number\ of\ correctly\ reconstructed\ candidate\ D^0}{Number\ of\ D^0\ensuremath{\rightarrow}\xspace \pi^+K^- decays} \\ p & = \frac{Number\ of\ correctly\ reconstructed\ candidate\ D^0}{Number\ of\ candidate\ D^0} \end{split} \end{equation} \begin{table}[htbp] \centering \caption{\label{D0} The efficiency and purity of $D^0\ensuremath{\rightarrow}\xspace \pi^+K^-$ reconstructed using Z-pole sample. The second column ($\epsilon$) lists selection efficiency after each selection step, and the third column (p) lists the corresponding purity.} \smallskip \begin{tabular}{ccc} \hline & $\epsilon$ (\%) & p (\%) \\ \hline $|mass - mass_{D0}| < 0.01\, $GeV/c$^2$ & $90.39\pm 0.24$ & $2.16\pm 0.07$ \\ IMP $> 0.02\, $mm$^2$ & $79.12\pm 0.21$ & $5.04\pm 0.11$ \\ vertex fitted $\chi^2 < 5.15$ & $72.62\pm 0.23$ & $15.36\pm 0.18$ \\ dis of vertex to IP $> 0.305\, $mm & $69.24\pm 0.24$ & $28.41\pm 0.23$ \\ PID & $68.19\pm 0.24$ & $89.05\pm 0.16$ \\ \hline \end{tabular} \end{table} After applying the above selection criteria, the distribution of the invariant mass of the track pair is shown in Fig.~\ref{D0mass}. The signal distribution is modeled with a Breit-Wigner theory model convolving a Gaussian detector response function, and the background distribution is modeled with a one order Chebyshev polynomial. The mean value of fitted $D^0$ mass is $1864.259\pm 0.025\, $MeV/c$^2$. \begin{figure}[htbp] \centering \includegraphics[width=.5\textwidth]{width15.png} \caption{\label{D0mass}Fit of the $D^0$ invariant mass of the track pair with inclusive hadronic Z-pole background included.} \end{figure} $D^0\ensuremath{\rightarrow}\xspace \pi^+K^-$ reconstruction performance as a function of dE/dx resolution is shown in Fig.~\ref{reso2}, which proves that better dE/dx resolution leads to better $D^0\ensuremath{\rightarrow}\xspace \pi^+K^-$ reconstruction performance. The blue/red/green line corresponds to 0\%/20\%/50\% degradation of dE/dx resolution. It can be seen that 20\% degradation of dE/dx resolution would not significantly degrades the $D^0\ensuremath{\rightarrow}\xspace \pi^+K^-$ reconstruction performance. \begin{figure}[htbp] \centering \includegraphics[width=.45\textwidth]{creso.png} \caption{\label{reso2}The distribution of $D^0\ensuremath{\rightarrow}\xspace \pi^+K^-$ reconstruction performance as a function of the factor defined in $\sigma_{actual} = factor \cdot \sigma_{intrinsic}$. The red/blue/green line corresponds to the 0\%/20\%/50\% degradation of dE/dx resolution.} \end{figure} \subsection{$\phi\ensuremath{\rightarrow}\xspace K^+K^-$ reconstruction at the Z-pole} The $\phi$ plays a central role in the reconstruction of many high-level objects, especially at $B_s\ensuremath{\rightarrow}\xspace \phi\nu\overline{\nu}$, where $\phi$ is the only visible component. This subsection investigates the performance of $\phi\ensuremath{\rightarrow}\xspace K^+K^-$ reconstruction with the inclusive hadronic Z-pole samples. The reconstruction chain is given below. \begin{enumerate} \item Match all pairs of reconstructed oppositely charged tracks. \item Set the absolute difference between the invariant mass of two tracks and the $\phi$ mass to less than $8\, $MeV/c$^2$. \item Use the kinematic fitting package to reconstruct the vertex of the two matched tracks and set the value of the vertex $\chi^2$ to less than 9.95. \item The identified positively charged particle must be a $K^+$ and a negative $K^-$. \end{enumerate} Notably, $\phi$ could come from QCD directly or decay from other hadrons. Since there are no requirement to identify these two sources, the selection conditions are not applied in a further step. The event selection process is shown in Table~\ref{phi}, where the definition of efficiency and purity is similar to that in the $D^0$ reconstruction. The effects of dE/dx resolution on $\phi$ reconstruction performance are shown in Fig.~\ref{resoPhi}. \begin{table}[htbp] \centering \caption{\label{phi}The efficiency and purity of $\phi\ensuremath{\rightarrow}\xspace K^+K^-$ reconstructed using Z-pole sample. The second column ($\epsilon$) lists selection efficiency after each selection step, and the third column (p) lists the corresponding purity.} \smallskip \begin{tabular}{ccc} \hline &$\epsilon$(\%) & p (\%)\\ \hline $|mass - mass_{D0}| < 8\, $MeV/c$^2 $ & $89.42 \pm 0.07$ & $6.43 \pm 0.06$ \\ vertex fitted $\chi^2 < 9.95$ & $83.15 \pm 0.09$ & $9.14 \pm 0.07 $ \\ PID & $82.26 \pm 0.09$ & $77.70 \pm 0.10$ \\ \hline \end{tabular} \end{table} \begin{figure}[htbp] \centering \includegraphics[width=.45\textwidth]{cQCDPhi.png} \caption{\label{resoPhi}The distribution of $\phi\ensuremath{\rightarrow}\xspace K^+K^-$ reconstruction as a function of the factor defined in $\sigma_{actual} = factor \cdot \sigma_{intrinsic}$. The red/blue/green line corresponds to the 0\%/20\%/50\% degradation of the dE/dx resolution.} \end{figure} \section{Conclusion} \label{sec:Con} The Z-pole mode of operation of the CEPC provides a great opportunity for flavor physics, where PID performance is essential. The PID studied in this article is derived from the TOF and dE/dx information collected by ECAL and TPC of the CEPC baseline detector. Using a GEANT4-based MC simulation, the dE/dx resolution can be better than 2.5\% in the barrel region with the energy of incident charged particle larger than $2\, $GeV/c. In real experiment, the dE/dx resolution would be degraded by the effects of the detector and electronic readout. At CEPC, dE/dx resolution is pursued to be better than 3\%, which corresponds to a degradation of the dE/dx resolution of less than 20\%. With inclusive hadronic Z-pole samples, the benchmarks of $K^{\pm}$ identification performance as well as $D^0\ensuremath{\rightarrow}\xspace \pi^+K^-$ and $\phi\ensuremath{\rightarrow}\xspace K^+K^-$ reconstruction performance is used to quantify the PID performance of CEPC. The $K^{\pm}$ identification efficiency and purity can reach 95.97\% and 81.56\%, respectively, only with dE/dx information. After combining the TOF information, the $K^{\pm}$ identification efficiency and purity can be improved by 2.56\% and 20.02\%, respectively. At the degradation of dE/dx resolution less than 20\%, the $K^{\pm}$ identification efficiency/purity can be better than 97\%/96\%. \ifcomment {\color{red} degradation of $\sigma_I/I$ less than 20\%? 不是dE/dx吗？} \fi The $D^0\ensuremath{\rightarrow}\xspace \pi^+K^-$($\phi\ensuremath{\rightarrow}\xspace K^+K^-$) reconstruction performance is also relies strongly on the PID performance and can be reconstructed with efficiency/purity of 68.19\%/89.05\% (82.26\%/77.70\%). The PID performance in the CEPC's Z-pole operating mode provides solid support for relevant flavor physics measurements. The analyses described in this article are used to evaluate PID performance and provide a reference for TPC optimization. It would be appreciated if an alternative technology, e.g., dN/dx \cite{Cuna:2021sho, Chiarello:2019eny}, could achieve significantly better resolution. \section*{Acknowledgment} We thank Fenfen An, Taifan Zheng, Zhiyang Yuan, Yuexing Wang, and Yuzhi Che for their supports and helps. We thank Gang Li and Chengdong Fu for producing the samples. This project is supported by the International Partnership Program of Chinese Academy of Sciences (Grant No. 113111KYSB20190030), the Innovative Scientific Program of Institute of High Energy Physics. \bibliographystyle{unsrt}

2408249cab2dfb736d13f67395f31acbabece1f2

\section{Introduction} The phase classification using machine-learning (ML) based techniques has been attracting intense attentions since the pioneering work in 2017\cite{carrasquilla2017}. In addition to the classical phase detections\cite{bedolla2021,mehta2019} where each phase is well defined by the corresponding order parameters, detecting topological phase transitions\cite{bedolla2021} is interesting and challenging\cite{beach2018} due to the lack of local order parameters. Recently, the phase detections and classifications have been performed via different ML techniques for classifying various topological invariants\cite{yoshioka2018,carvalho2018,balabanov2020,balabanov2021,greplova2020,greplova2020,ho2021,zhang2021,narayan2021,yu2021, zhang2017,cheng2018,sun2018,zhang2020,carvalho2018,kerr2021,kaming2021,ho2021,rem2019,che2020,chung2021,ming2019quantum,zhang2018,sun2018, holanda2020,che2020,kerr2021,tsai2021,kerr2021,zhang2020,zhang2017a,mano2019,su2019,lian2019,carvalho2018,ho2021,beach2018,laskowska2018,zhang2019, RN2019,tsai2020,scheurer2020,tsai2021,caio2019}, including the Chern number\cite{zhang2017,cheng2018,sun2018,zhang2020,carvalho2018,kerr2021,kaming2021,ho2021,rem2019,che2020,chung2021,ming2019quantum}, winding number\cite{zhang2018,sun2018,holanda2020,che2020,kerr2021,tsai2021,kerr2021}, $\mathbb{Z}_2$ index\cite{zhang2020,zhang2017a,mano2019,su2019,lian2019,carvalho2018,ho2021,beach2018,laskowska2018,zhang2019, RN2019,tsai2020,scheurer2020,tsai2021,caio2019}, to name a few. In addition to the applied ML architectures, the forms of the inputs for training the machine also play a crucial role in determining the resulting performance of the topological phase detections\cite{beach2018}. For the topological systems with the Chern numbers or the winding numbers as the topological invariants, various types of inputs are used to perform the phase classifications. For instance, researchers in Kim's group introduced quantum loop topography (QLT) to construct multi-dimensional images from raw Hamiltonians or wave functions as inputs\cite{zhang2017,zhang2020}. Zhai's group collected the Bloch Hamiltonians into an arrays to feed their machines\cite{zhang2018,sun2018}. The real-space particle densities and local density of states were used as inputs by Cheng \textit{et al}.\cite{cheng2018}. Carvalho \textit{et al}. fed the local projections of the density matrix to the machine \cite{carvalho2018}. From cold-atom experiments, momentum-space density images were generated as inputs by Rem \textit{et al}.\cite{rem2019}. The time-of-flight images\cite{ho2021,kaming2021}, spatial correlation function\cite{ho2021} and density–density correlation function\cite{ho2021} were also used as inputs. The density profiles formed in quantum walks were proposed as appropriate inputs for training in Ming \textit{et al}'s work\cite{ming2019quantum}. Furthermore, several works had tried the spin configurations\cite{kerr2021} and the Bloch Hamiltonians over the Brillouin zone (BZ) as inputs\cite{che2020,kerr2021}. For these forms of inputs mentioned above, various ML techniques with distinct real-valued neural networks have been applied to discriminate different topological phases. As the development of artificial neural networks becomes mature, generalization from real-valued neural networks to complex-valued ones is undertaken that anticipates a raise of representation capability of machines~\cite{trabelsi2017,gaudet2018deep}. Specifically, a quaternion contains one real part and three imaginary parts so that a quaternion-based neural network\cite{GR2020,isokawa2009,parcollet2020,matsui2004} is expected to have remarkable performance on handling the processing of data with more than two degrees of freedom such as the color images (RGB channels) and the descriptions of 3D systems (xyz coordinates). There have been various proposals about quaternion-based neural networks in ML techniques and applications in computer science, such as the quaternion convolutional neural network (qCNN)\cite{zhu2019,hongo2020,gaudet2018deep}, quaternion recurrent neural network\cite{parcollet2018quaternion}, quaternion generative adversarial networks\cite{grassucci2021}, quaternion-valued variational autoencoder\cite{grassucci2020}, quaternion graph neural networks\cite{nguyen2020}, quaternion capsule networks\cite{ozcan2020} and quaternion neural networks for the speech recognitions\cite{parcollet2018}. However, the ML-related applications of the quaternion-based neural networks on solving problems in physics are still limited, especially in the topological phase detections, even though the quaternion-related concepts have been applied in some fields in physics~\cite{girard1984,girard2007,girard2018}. \begin{figure}[th!] \centering \includegraphics[width=0.46\textwidth]{figure/QWZ.pdf} \caption{Spin vectors in the Brillouin zone, $k_x,k_y\in (-\pi,\pi]$, with Chern number (a) \(C = 1\), (b) \(C=2\), (c) \(C = 3\) and (d) \(C=4\).} \label{fig:spin_conf} \end{figure} In this work, we perform the Chern-insulator classifications from both supervised- and unsupervised-learning aspects based on the inputs transformed via the quaternion algebra. For the unsupervised learning, we firstly encode the quaternion-transformed eigenstates of Chern insulators via a convolution function as inputs and study them using the principal component analysis. We found that using only the first two principal elements is not enough to fully classify the Chern insulators, consistent with Ming's work\cite{ming2019quantum}. Further studies show that the performance can be improved by including more principal components. For the supervised learning, we construct a quaternion-based neural network in which the first layer is a quaternion convolutional layer. We then show that this quaternion-based machine has better performance than a conventional CNN machine. Our machine is good not only at testing but also at identifying data that have no common structures as in training data. The good performance can be partially attributed to the similarities between the formula of the Berry curvatures and the quaternion algebra. Therefore, our work demonstrates the power of the quaternion algebra on extracting relevant information from data, paving the way to applications of quaternion-based algorithm ML techniques in topological phase classifications. The outline of the remaining part of this work is as follows. In section II, we introduce the model Hamiltonian, generating the training data for our classification tasks, and the quaternion convolution layer used in this work. PCA analysis of the quaternion-transformed eigenstates is discussed in Sec. III. The data preparations, the network structures and the performance of the quaternion-based supervised learning task are given in Sec. IV. Some further discussions on our neural networks are given in Sec. V. Finally, a conclusion is provided in Sec. VI. \section{Model and quaternion convolutional layer} \subsection{Model} A generic two-band Bloch Hamiltonian with the aid of the identity matrix $\sigma_0$ and Pauli matrices \(\pmb{\sigma}=(\sigma_1, \sigma_2,\sigma_3)\) is written as \begin{equation}\label{eq:2band} \mathcal{H}(\vec{k}) = h_0(\vec{k}) \sigma_0 + \mathbf{h}(\vec{k})\cdot\pmb{\sigma}, \end{equation} where $\vec{k}=(k_x,k_y)$ is the crystal momentum in the 2D BZ($={k_x,k_y\in(-\pi,\pi]}$), and the vector \(\mathbf{h} = (h_1,h_2,h_3)\) acts as an external magnetic field. Sometimes we will omit arguments $\vec{k}$ when writing $\mathbf{h}(\vec{k})$ for brevity in $k$-space. The eigenstate of the upper (lower) band will represent the spin pointing antiparallel (parallel) to \(\mathbf{n} = \mathbf{h}/\abs{\mathbf{h}}\). Since $h_0(\vec{k})$ does not affect the pointing direction, it will be ignored in the remaining part of this section. The function $\mathbf{n}$ embeds the topology of the system, and we firstly discuss the topological structure in $\mathbf{n}$. Here, the topological invariant is the Chern number $C \in \mathbb{Z}$ with the formula: \begin{equation}\label{chern} C = \frac{1}{4\pi}\int_{\text{BZ}} \mathbf{n}\cdot(\partial_{k_x}\mathbf{n}\times\partial_{k_y}\mathbf{n})d\vec{k}, \end{equation} where the integrand is the Berry curvature and the integration is over the first BZ. \begin{figure}[b] \centering \includegraphics[width=0.45\textwidth]{figure/HighChernNumber.pdf} \caption{The Chern number with various \(m\) and \(c\).} \label{fig:HighChern} \end{figure} We construct the normalized spin configurations $\mathbf{n}(\vec{k})$ based on the following models. For topological systems, we choose the Hamiltonian with \(\mathbf{h}=\mathbf{h}^{(c)}\), where \begin{equation}\label{eq:HighChern} \mathbf{h}^{(c)}(\vec{k})= \begin{pmatrix} \ \mathrm{Re}\big[(\sin{k_x}-i\sin{k_y})^c\big]\\ -\mathrm{Im}\big[(\sin{k_x}-i\sin{k_y})^c\big]\\ \cos{k_x} + \cos{k_y} + m \end{pmatrix} \end{equation} with positive integer \(c\) and real parameter \(m\) to control the Chern number. For \(c = 1\), the model is the Qi-Wu-Zhang (QWZ) model \cite{PhysRevB.74.085308}. For a given \(c\), the Chern number $C$ can be either \(0,~c,\text{ or} -c\) depending on the value of \(m\): \begin{equation} C = \begin{cases} \mathrm{sgn}(m)c, & 0<|m|<2, \\ 0, & |m|>2. \end{cases} \end{equation} The Chern numbers with different c's and m's are shown in Fig.~\ref{fig:spin_conf}. Note that $C=0$ stands for a topologically trivial phase and nonzero $C$'s are for nontrivial phases. The topological phase diagram is shown in Fig.~\ref{fig:HighChern}. In this work, the unsupervised learning involves seven topological phases ($C=0,\pm 1,\pm 2, \pm 3$) in Sec. \ref{sec:PCA}, and the supervised learning involves nine topological phases ($C=0,\pm 1,\pm 2, \pm 3, \pm 4$) in Sec. \ref{sec:qcnn}. \subsection{Quaternion convolutional layer} A quaternion number has four components, the first of which stands for the real part and the other three of which stand for the imaginary parts. Given two quaternions $q_1=(r_1,a_1,b_1, c_1)$ and $q_2=(r_2,a_2,b_2,c_2)$, their product $Q=q_1q_2=(R,A,B,C)$ is given by \begin{equation}\label{eq:q_1q_2} \begin{pmatrix} R\\ A\\ B\\ C \end{pmatrix} = \begin{pmatrix} r_1 r_2 - a_1 a_2 - b_1 b_2 - c_1 c_2\\ a_1 r_2 + r_1 a_2 - c_1 b_2 + b_1 c_2\\ b_1 r_2 + c_1 a_2 + r_1 b_2 - a_1 c_2\\ c_1 r_2 - b_1 a_2 + a_1 b_2 + r_1 c_2 \end{pmatrix}, \end{equation} which can be written as the matrix product form \begin{equation}\label{eq:matrix_q} \begin{pmatrix} R\\ A\\ B\\ C \end{pmatrix} = \begin{pmatrix*}[r] r_1 & -a_1 & -b_1 & -c_1 \\ a_1 & r_1 & -c_1 & b_1 \\ b_1 & c_1 & r_1 & -a_1 \\ c_1 & -b_1 & a_1 & r_1 \end{pmatrix*} \begin{pmatrix} r_2\\ a_2\\ b_2\\ c_2 \end{pmatrix}. \end{equation} To implement a quaternion convolutional (q-Conv) layer in numerical programming, we will regard the two quaternions as a $4\times4$ matrix and a $4\times1$ column matrix, respectively: \begin{equation} \label{eq:rep2} q_1 \doteq \begin{pmatrix*}[r] r_1 & -a_1 &-b_1 & -c_1 \\ a_1 & r_1 & -c_1 & b_1 \\ b_1 & c_1 & r_1 & -a_1 \\ c_1 & -b_1 & a_1 & r_1 \end{pmatrix*}\quad \mathrm{and}\quad q_2 \doteq \begin{pmatrix*}[r] r_2\\ a_2\\ b_2\\ c_2 \end{pmatrix*}. \end{equation} More details of quaternion algebra are described in Appendix~\ref{app:qcnn}. A conventional CNN contains a real-valued convolutional layer to execute the convolution of the input and the kernel. Let the input $F$ have the shape: $H_i\times W_i\times C_i$ (Height \(\times\) Width \(\times\) Channel) and the shape of the kernel $K$ be $H_{k} \times W_{k} \times C_i \times C_f$. The convolution will produce an output \(O\), $O=F\ast K$, whose elements are \begin{equation}\label{eq:C-convo} O_{i',j',t'} =\sum_{i}^{H_k} \sum_{j}^{W_k} \sum_{t}^{C_i} F_{i'+i-1,j'+j-1,t}\cdot K_{i,j,t,t'}. \end{equation} Here the stride is assumed to be \(1\) both in the width and the height directions. The indices $i$ and $j$ are spatial indicators, $t$ is the index of channel in the input feature map and $t'$ is the kernel index. The shape of the output will be $(H_{i}-H_{k}) \times (W_{i}-W_{k}) \times C_f$. Assume that the input has four components. To uncover the entanglement among components through CNN, we will utilize the quaternion product. Now, we introduce another dimension–Depth–which is four, as a quaternion number of four components. Both of the input $F$ and the kernel $K$ have Depth of four as two quaternion numbers. The product of $F$ and $K$ will have Depth of four as a quaternion in Eq.~(\ref{eq:q_1q_2}). Referring to Eq.~(\ref{eq:rep2}) where we show a matrix representation to implement quaternion algebra and thinking of $F$ as $q_1$ and $K$ as $q_2$ in Eq.~(\ref{eq:rep2}), we transform the Depth-four input $F$ into a 4x4 matrix, $F^{(l,s)}$, and keep the kernel $K$ still of Depth 4, $K^{(l)}$, where $l, s=1,…,4$. The product of $F$ and $K$, say $O$, will have Depth of four as shown in Eq.~(\ref{eq:Q-convo}). We remark that the products between components of $F$ and $K$ are convolution operations as Eq.~(\ref{eq:C-convo}). \begin{equation}\label{eq:Q-convo} O_{i',j',t'}^{(s)} = \sum_{l}^{4}\sum_{i,j,t} F^{(s,l)}_{i'+i-1,j'+j-1,t}\cdot K^{(l)}_{i,j,t,t'}, \end{equation} \begin{figure}[b!] \centering \includegraphics[width=0.47\textwidth]{figure/q_conv_layer.pdf} \caption{Illustration of a quaternion convolutional layer. On the left, we start with the input $q_1$ having four quaternion components ((yellow, red, green, blue) stands for ($r_1$, $a_1$, $b_1$, $c_1$)). In the middle, $q_1$ is permuted to construct $\{F^{(\cdot, l)}\}_{l=1}^{4}$ on which the convolution with four kernels $\{K^{(l)}\}_{l=1}^{4}$ is performed. A summation is taken for each Depth dimension to obtain the output feature map $O$ on the right.} \label{fig:Q-filter} \end{figure} More specifically, we consider an input data as $q_1$ (four color squares on the left of Fig. \ref{fig:Q-filter}) and four kernels encoded in $q_2$, given in the following \begin{equation} \left\{\begin{matrix} q_1 \doteq & (r_1~a_1~b_1~c_1)^T &\\ q_2 \doteq & (r_2~a_2~b_2~c_2)^T & =: K^{(\cdot)}. \end{matrix}\right. \end{equation} The output feature maps $O\doteq (R~A~B~C)^T$ is then calculated based on Eq.~(\ref{eq:q_1q_2}). As the first step, we permute the order of $q_1$ to obtain \begin{eqnarray} F^{(\cdot,1)}=: \begin{pmatrix} r_1\\ a_1\\ b_1\\ c_1 \end{pmatrix}, F^{(\cdot,2)}=: \begin{pmatrix*}[r] -a_1\\ r_1\\ c_1\\ -b_1 \end{pmatrix*},\\\nonumber F^{(\cdot,3)}=: \begin{pmatrix*}[r] -b_1\\ -c_1\\ r_1\\ a_1 \end{pmatrix*}, F^{(\cdot,4)}=: \begin{pmatrix*}[r] -c_1\\ b_1\\ -a_1\\ r_1 \end{pmatrix*} \end{eqnarray} (see the four sets of sqaures in the middle of Fig. \ref{fig:Q-filter}). We then convolute those four quaternions ($F^{(\cdot,l)}$ with $l =1,2,3$ and 4) with four kernels ($K^{(l)}$ with $l = 1,2,3$ and 4) in the following way: $$ \left\{ \begin{matrix} F^{(\cdot,1)}K^{(1)} &\doteq& \begin{pmatrix} r_1r_2&a_1r_2&b_1r_2&c_1r_2 \end{pmatrix}^T\\ F^{(\cdot,2)}K^{(2)} &\doteq& \begin{pmatrix} -a_1a_2&r_1a_2&c_1a_2&-b_1a_2 \end{pmatrix}^T\\ F^{(\cdot,3)}K^{(3)} &\doteq& \begin{pmatrix} -b_1b_2&-c_1b_2&r_1b_2&a_1b_2 \end{pmatrix}^T\\ F^{(\cdot,4)}K^{(4)} &\doteq& \begin{pmatrix} -c_1c_2&b_1c_2&-a_1c_2&r_1c_2 \end{pmatrix}^T \end{matrix} \right. $$ , as shown in the middle of Fig. \ref{fig:Q-filter}. Finally, we sum over the above four quaternions to get the output feature maps $O$, as shown on the right of Fig. \ref{fig:Q-filter}. $$ O := \begin{pmatrix} R\\ A\\ B\\ C \end{pmatrix} = \begin{pmatrix} r_1 r_2 - a_1 a_2 - b_1 b_2 - c_1 c_2\\ a_1 r_2 + r_1 a_2 - c_1 b_2 + b_1 c_2\\ b_1 r_2 + c_1 a_2 + r_1 b_2 - a_1 c_2\\ c_1 r_2 - b_1 a_2 + a_1 b_2 + r_1 c_2 \end{pmatrix}. $$ \section{principal component analysis}\label{sec:PCA} Principal component analysis (PCA) is a linear manifold learning that is to find the relevant basis set among data\cite{jolliffe2016principal,ma2012manifold}. We prepare eigenstates $\ket{u_\pm}$ of Eq.~(\ref{eq:2band}), where $+\ (-)$ stands for the upper (lower) band. For a topological state, the phase cannot be well-defined over the whole BZ. Therefore, we can divide the whole BZ into two parts, in each part of them the topological wave function has continuously well-defined phase. We then choose two regions according to the sign of $h_3$ in Eq. (\ref{eq:HighChern}): \begin{align} \begin{split} \ket{u_+} & \doteq \frac{1}{\sqrt{2h_+(h_{+}+h_3)}}\mqty(h_{+}+h_3 \\ h_1+i h_2) \\ \ket{u_-} & \doteq \frac{1}{\sqrt{2h_-(h_{-}+h_3)}} \mqty(-h_1+ih_2 \\ h_{-}+h_3) \end{split} \quad ,~ h_3\geq 0, \label{eq:gauge1} \end{align} and \begin{align} \begin{split} \ket{u_+} & \doteq \frac{1}{\sqrt{2h_{+}(h_{+}-h_3)}}\mqty(h_1-ih_2 \\ h_{+}-h_3) \\ \ket{u_-} & \doteq \frac{1}{\sqrt{2h_{-}(h_{-}-h_3)}} \mqty(h_{-}-h_3 \\ -h_1-ih_2) \end{split} \quad ,~ h_3 < 0, \label{eq:gauge2} \end{align} where $h_{\pm} =\pm \sqrt{h_1^2+h_2^2+h_3^2}$. In this choice of gauge, the first (second) component of $\ket{u_{+}}$ $(\ket{u_-})$ is real-valued when $h_3\geq0$, and the second (first) component of $\ket{u_{+}}$ $(\ket{u_-})$ is real-valued when $h_3<0$. By translating $\ket{u_\pm} \doteq (\alpha_\pm, \beta_\pm)^T$ with $\alpha_\pm, \beta_\pm\in\mathbb{C}$, into a quaternion number of four components, we have \begin{equation}\label{eq:transf} q_\pm := \mathrm{Re}(\alpha_\pm) + \mathrm{Im}(\alpha_\pm)\hat{\mathbf{i}} + \mathrm{Re}(\beta_\pm)\hat{\mathbf{j}} + \mathrm{Im}(\beta_\pm)\hat{\mathbf{k}}. \end{equation} (a brief introduction of the quaternion number can be seen in Appendix \ref{app:qcnn}). To see the correlation of states over $\vec{k}$, we define the quantity $F$ to be the quaternion-based convolutions: \begin{equation}\label{eq:Ne} \begin{split} F(\vec{p}) := q^*_{+}&\circledast q_{+}[\vec{p}] - q^*_{-}\circledast q_{-}[\vec{p}] \quad\mathrm{with}\\ q_{\pm}^*&\circledast q_{\pm}[\vec{p}] := \sum_{\vec{k}\in \mathrm{BZ}} q_{\pm}^*(\vec{k})q_{\pm}(\vec{p} - \vec{k}), \end{split} \end{equation} where $q^*$ is the conjugate of $q$. It can be proved that $F$ is real-valued. Therefore, $F(\vec{p})$ of all $\vec{p}$ in the BZ based on a given Hamiltonian can be analysed by using PCA. \begin{figure}[b!] \centering \includegraphics[width=0.45\textwidth]{figure/nematic2.pdf} \caption{The maps of the function \(F\) without noise in the BZ. Three rows are for \(c = 1, 2\ \mathrm{and}\ 3\) in Eq.~\ref{eq:HighChern} from top to bottom; four columns from left to right are for $m = -3,-1, 1$ and 3. The corresponding Chern number $C$ is tagged with each panel.} \label{fig:nematic} \end{figure} We collected various $F$ of all $\vec{k} \in $BZ within seven topological phases as the dataset for PCA. For each topological phases, 30 $F$'s were prepared, so the total amount of data was 210. The data for six non-trivial phases were generated based on Eq. (\ref{eq:HighChern}) with $m=\pm 1$ (the sign of $m$ determines the sign of $C$). For the trivial phase, we prepared five data points from each of six combinations of $\{c,m\}$, where $c\in\{1,2,3\}$ and $m\in\{3,-3\}$, and then there are totally 30 data. To augment the number of data, we add Gaussian noises $\delta\mathbf{h}$ at every $\vec{k}$ of the model [Eq. (\ref{eq:HighChern})] such that $\mathbf{h}\to\mathbf{h}+\delta\mathbf{h}$ without closing the band gap. In Fig.~\ref{fig:nematic}, we present various noiseless $F$ generated from Eq. (\ref{eq:HighChern}) with different $c$ and $m$. \begin{figure}[t!] \centering \includegraphics[width=0.49\textwidth]{figure/2DPCA.pdf} \caption{PCA of seven topological phases with various noise. The symbols with corresponding Chren numbers are marked in the legend.} \label{fig:2DPCA} \end{figure} It is notable that $F$ for $C = 0$ are featureless, $F$ for $C = \pm 1$ have a dipole moment, and $F$ for $C = \pm 2$ have a quadruple moment, and $F$ for $C = \pm 3$ seemingly have a primary dipole and a secondary quadruple moment. The remarkable features imply that the convolution function $F$ is a good choice for topological classifications. We examine data with the standard deviation (SD) equal to 0, 0.1, 0.2 and 0.3 respectively, and show the first two PCs of 210 pieces of data for each SD in Fig.~\ref{fig:2DPCA}. In Fig.~\ref{fig:2DPCA}, it is evident that data are clustered into four groups and their variances increase with SD. PCA is successful to separate different topological phases into different clusters. However, some clusters contain two topological phases of Chern numbers: $\{+1,-3\}$, $\{-1,+3\}$, and $\{+2,-2\}$. The $C$ modulo 4 resemblance has also be observed in a previous study~\cite{ming2019quantum}. \begin{figure}[b!] \centering \includegraphics[width=0.5\textwidth]{figure/PCs.pdf} \caption{Magnitude of projection (logarithmic scale) from non-trivial data onto first six principal components. Inset: The first 16 principal values of PCA. (normalized by maximal $\lambda_1$)} \label{fig:ninePC} \end{figure} We find that including more PCs helps separate different classes in each cluster. Figure \ref{fig:ninePC} shows first six PCs of data in topologically non-trivial phases, where PCx denotes the x-th PC component. One can find that PC1 and PC2 in each pair of $\{+1,-3\}$, $\{-1,+3\}$, and $\{+2,-2\}$ are nearly identical, as also shown in Fig.~\ref{fig:2DPCA}. By including PC3 ~ PC6 into the analysis, all topological classes are completely classified. Via the proposed convolution, topological states can be successfully classified by using PCA, a linear machine for classification. \begin{figure*}[t!] \centering \includegraphics[width=0.945\textwidth]{figure/qCNN-net2.pdf} \caption{Framework of (a) the CNN and (b) the qCNN classifier. The shape of data is $L_x \times L_y \times$ (number of channels). The number appended after the symbol ``@'' stands for the number of filters convolved with the data, and the number just above is the filter size. The number of tuneable network parameters of CNN (qCNN) is 24,252 (19,350).} \label{fig:QCNN} \end{figure*} \section{Supervised learning of CNN and the qCNN}\label{sec:qcnn} \subsection{Datasets} The input data are normalized spin configurations $\mathbf{n}$, laying on a $40\times 40$ square lattice with periodic boundary conditions, and their corresponding topological phases are labels with one-hot encoding. We prepared four datasets: training, validation, testing and prediction dataset. The first three datasets are well known in conventional deep learning procedure~\cite{chollet2021deep}. To understand whether our machine can also classify unseen spin configuration, we prepare a prediction dataset that includes a few types of spin configurations never seen by the machine during the training process. The data pool containing training and validation datasets is constructed as follows. Based on the Eq.~(\ref{eq:HighChern}), we firstly prepared 5760 data points of $\mathbf{n}$ of nine topological phases with Chern number ranging from -4 to 4 so that each phase contain 640 data points. Additionally, we add 360 data points for spin vortex which belongs to the trivial phase. A spin-vortex has an in-plane spin texture that winds around a center, which is generated by setting one of three components in Eq.~(\ref{eq:HighChern}) to be zero. By including spin vortices, the machine learn the difference between 3D winding (non-trivial) and 2D winding (trivial) spin configurations (more details are described in the Appendix \ref{app:dataset}). After the training process, the trained machine is scored by a testing dataset with the same composition of nine phases as that in the training (and validation) dataset. Importantly, without changing the topologies, the Gaussian distributed random transition and random rotation imposed on these three datasets can increase the diversity of dataset and enhance the ability of generalization of the trained machine. The prediction dataset contains six categories of spin configurations. The first category is generated with $m$ uniformly distributed from $+3$ to $-3$. In the second and the third categories, we change the sign of $n_z$ (the second category) and swapping $n_y$ and $n_z$ of $\mathbf{n}$ (the third category). Finally, we consider three categories for trivial states, which are ferromagnetic (FM), conical $(\epsilon\neq 0)$ and helical $(\epsilon=0)$ states, based on the following formula: \begin{equation}\label{eq:conical} \mathbf{n}_{\mathrm{spiral/conical}}(\vec{k})= \begin{pmatrix} \sqrt{1-\epsilon^2}\cos{(k_x+ k_y)}\\ \sqrt{1-\epsilon^2}\sin{(k_x+ k_y)}\\ \epsilon \end{pmatrix}. \end{equation} FM can be viewed as 1D uncompleted winding configuration while conical and helical can be viewed as 2D uncompleted ones. In total, we prepared six categories for the prediction dataset. More details about data preparations will be described in Appendix \ref{app:dataset}. For the conventional CNN, we use $\mathbf{n}$ as the input data. For the qCNN, in order to feed the input data into the qCNN classifier, we transform the 3D spin vector into an unit pure quaternion, \begin{equation}\label{eq:encoding} (n_x,n_y,n_z)\in\mathbb{R}^3 \mapsto (0, n_x,n_y,n_z)\in\mathbb{H}, \end{equation} where the scalar part (the first component) is zero and the vector part is $\mathbf{n}$. \subsection{network structure and performance} We implement a qCNN classifier with a quaternion convolution (q-Conv) layer as the first layer [see red dotted cube in Fig.~\ref{fig:QCNN}(b)], and the operations in a q-Conv layer are based on the quaternion algebra. Then the next three layers are conventional 3D convolutional layers. Details of the quaternion algebra and the keynote of a qCNN are explained in Appendix~\ref{app:qcnn}. For comparison, we also set up a conventional CNN classifier by replacing the q-Conv layer by a conventional 2D convolutional layer, and appending three conventional 2D convolutional layers to the first layer. The architecture of these two classifiers are shown in Fig.~\ref{fig:QCNN}. Note that the data in the qCNN has one more rank than that in CNN, and in numerical programs, this extra dimension---\emph{Depth} is used to store the quaternion [referred to Eq. (\ref{eq:encoding})]. As you can see in the Fig.~\ref{fig:QCNN}, the architecture of the qCNN is one layer more than the CNN's, the total network parameters of the qCNN is however less than the CNN's. This is one advantage of the qCNN over the conventional CNN. In order for classifiers to satisfy some physically reasonable conditions, three special designs are needed. Firstly, we extend the $k$ points out of the BZ by padding the input data according to the periodic boundary conditions~\cite{PhysRevB.99.075113}. \begin{figure}[b!] \centering \includegraphics[width=0.45\textwidth]{figure/overlap.pdf} \caption{Schematic of ``overlap'' convolution (red solid) and ``non-overlap'' convolution (blue solid) from a $3\times3$ filter (black dotted) over data. The blue solid square is a signal movement from the filter, and the size of stride is the same as the length of filter, thus each movement of this filter is ``non-overlap''.} \label{fig:overlap} \end{figure} Secondly, a convolutional layer (q-Conv layer in the qCNN) with arctangent activation function is then adopted, which contains 27 filters of kernel size $2\times 2$. This layer executes an ``overlapping'' feature mapping. (Figure~\ref{fig:overlap} illustrates how the ``overlapping'' and ``non-overlapping'' feature mapping can be manipulated by varying the size of stride.) Thirdly. we shrink the $k$ space of data into ``a point" by three non-overlapping convolutional layers. In the qCNN, three 3D convolutional layers are applied while keeping four Depths independent. In particular, we additionally append a 2D convolutional layer to make combination Depths-wisely. Specifically, this layer with 9 filters of kernel size $4\times1$ transforms data from $4\times9$ to $1\times9$, where each of nine neurons corresponds to one topological class. In comparison, the conventional CNN shrinks the $k$-space to a point in each of nine \emph{Channels}, where the dimension is integrated out after the convolution. More details and descriptions are relative to the SEC. \ref{sec:discu}. \begin{figure}[b!] \centering \includegraphics[width=0.46\textwidth]{figure/accuracy2.pdf} \caption{Learning curves of the qCNN and CNN classifiers. By applied \emph{Dropout}, the validation is greater than the training.} \label{fig:learningCurve} \end{figure} In the following, we perform both the qCNN and conventional CNN trainings. The learning curves of both machines are shown in Fig.~\ref{fig:learningCurve}. The CNN machine (orange and light orange lines) jumps over a big learning barrier at around the $700^{th}$ epoch. After that, the training and the validation accuracy (orange and light orange line respectively) are separated and do not converge up to end of this training process. Even though the same training (and validation) dataset is used in the training process, the learning curves of the qCNN machine (blue and light blue lines) are qualitatively different. The training and the validation accuracy are separated around $90^{th}$ epoch, but the difference between these two accuracies decreases with increasing epochs. After the training procedure finished, the qCNN (CNN) machine gets 99.67\% (94.12\%) testing accuracy. This difference in accuracy results from the spin-vortex dataset, where the qCNN works well but CNN dose not. The trained machines are ready to do prediction, and the result is shown in Fig.~\ref{fig:performpred}. \begin{figure}[b!] \centering \includegraphics[width=0.46\textwidth]{figure/predictset.pdf} \caption{The performance of the qCNN (blue line) and CNN (red dashed line) on the prediction datasets. Numbers tagged are the values of the accuracies. Standard deviations (by error bars) are also provided. The qCNN outperforms the conventional CNN on all prediction datasets, especially on three spin-vortex ones} \label{fig:performpred} \end{figure} In Fig.~\ref{fig:performpred}, since the first category contains $\mathbf{n}$ of uniformly distributed $m$, where a few data points are very close to the phase boundaries $m \approx \{0, \pm 2 \}$, the accurate rate of the the qCNN is slightly low at $96\%$. For the second and third categories, we choose $m=\pm 1$, away from the phase transition points, and the performance is nearly perfect. For the uncompleted winding configurations, the qCNN, different from the conventional CNN, can accurately classifies FM, helical and conical states after learning the spin-vortex states. This is the main advantage of the qCNN over the conventional CNN, which is expected to result from the quaternion algebra. \section{discussions}\label{sec:discu} On PCA the reason why we choose spinor state vector $\ket{u}$, not the spin normalized vector $\mathbf{n}$ as data is that the gauge discontinuities exist in the non-trivial $\ket{u}$, but not in $\mathbf{n}$'s. It turns out that the function $F$ can detect the existence of this discontinuity so that trivial and non-trivial states can be distinguished by PCA. Above statement is further supported by the relation between $\ket{u}$ and $\mathbf{n}$ in pure state, which is known as \begin{equation}\label{eq:sigma} \mathbf{n}(\vec{k}) = \bra{u_+(\vec{k})}\pmb{\sigma}\ket{u_+(\vec{k})}. \end{equation} In Eq. (\ref{eq:sigma}), the $U(1)$ ambiguity of $\ket{u}$ is eliminated effectively, thus, there is no discontinuities to be detected by PCA in $\mathbf{n}$. There is another issue if we use n as the data in PCA. If we encode $\mathbf{n}=(n_x, n_y,n_z)$ into a quaternion in the following form, \[q := (0,~n_x,~n_y,~n_z),\] the corresponding convolution [Eq. (\ref{eq:Ne})] will not depend on the sign of $m$. It can be shown that there exists one point $\vec{k}'\to\vec{p}_0 - \vec{k}+\pi$ in BZ such that \[n_z(\vec{k}, m_0)n_z(\vec{p}_0-\vec{k}, m_0) = n_z(\vec{k}', -m_0)n_z(\vec{p}_0-\vec{k}', -m_0).\] That is, for two spin configurations with opposite sign $m_0$, their $n_z$ products are equal at $\vec{k}$ and $\vec{k}'$ respectively. As shown in Eq.~(\ref{eq:HighChern}), m shows up only in the $n_z$ component. Therefore, after the integration over the whole BZ, the convolution is independent of the sign of $m$ and there is no feature for PCA to discriminate distinct topological states. \begin{figure}[b!] \centering \includegraphics[width=0.47\textwidth]{figure/filter.pdf} \caption{(a) Three nearest neighborhood spin vectors will contribute a solid angle, and (b) four nearest neighbors are enclosed by the filters in the first convolutional layer.} \label{fig:filter} \end{figure} On qCNN, there are several possible factors promoting the performance of our supervised learning machine. The first one is that the size of filter in the first convolutional layer is 2 × 2 with stride = 1, which means the machine can collect spin information among four nearest neighbors [see Fig. \ref{fig:filter}(b)]. We know that the Chern number is the integral of the Berry curvature in the BZ, and the Berry curvature is twice of the solid angle. A solid angle $\Omega$ subtended by three normalized vectors $\vec{a}$, $\vec{b}$, $\vec{c}$, can be calculated using the following equation: \begin{equation}\label{eq:solid} \tan{\frac{\pmb{\Omega}}{2}}=\frac{\abs{\vec{a}\cdot(\vec{b}\times \vec{c})}}{1+\vec{a}\cdot \vec{b}+\vec{b}\cdot \vec{c}+\vec{c}\cdot \vec{a}}. \end{equation} Our choice of the size of the filter in the first hidden layer is the minimal of $2 \times 2$ that mixes only the nearest-neighboring spins. In this way, it is very possible to enforce the machine to notice the solid angle extended in this plaquette. The second factor is the quaternion product. Recall that the conventional CNN might correlate spins $\mathbf{n}'s$ in neighboring $\vec{k}'s$ due to the feature map through the kernel. However, the map does not mix the components of spins. In comparison, the qCNN is more efficient for it directly entangle spins via the quaternion product. It is this entanglement of spin components by the quaternion product that makes the scalar and vector products in calculating the solid angle (see Eq.~ (\ref{eq:solid})) become possible to be realized by the machine. As a solid angle involves at least three spins and the feature map by the kernel is just linear, a nonlinear transformation is crucial to create high-order (three spins) terms in the expansion. Based on this argument, the third factor is the non-linear activation function \cite{lin2017}, which is an arctangent function in this work. Based on Eq.~ (\ref{eq:solid}), the calculation of a solid angle involves the arctan operation. Therefore, we expect that using the arctangent function as the activation function can further help the machine to learn correct representations. This idea is further supported by the results shown in Fig. \ref{fig:acti}, where the arctangent activation function outperforms the ReLU and tanh activation functions over nine different datasets. In summary, several factors are combined to enhance the performance of our machine as follows. The quaternion-based operations in the q-Conv layer mix not only different components of a spin $\mathbf{n}(\vec{k}_0)$ but also neighbouring spins, say $\mathbf{n}(\vec{k}_0)$ and $\mathbf{n}(\vec{k}_1)$. When these linear combinations are fed into the non-linear activation functions in our qCNN, the output can be viewed as an expansion of a non-linear function, which may contain a term having both the scalar- and vector-product of neighboring spins, similar to that in Eq.~ (\ref{eq:solid}). Therefore, in the optimization process, combined with the effect of choosing the size of filter to be $2 \times 2$, the machine may keep increasing the weight of a solid-angle-related term and eventually learn to classify the topological phases based on the calculated solid angles. Also, adding some noises to the training dataset helped our supervised-learning machine to learn the generic feature of our data. We found that when the training data was generated directly from Eq. (\ref{eq:HighChern}) without adding any noise, the machine worked well for training and testing datasets but had poor performance on all the prediction dataset. This could be understood by noting that topology is determined by the sign of $m$, which appears only in the z component in Eq. (\ref{eq:HighChern}). By using the dataset without noises, the machine may learn to classify states only by looking at the $z-$component but not extracting the full information from the full configuration. When we gave random transition and rotation to each training data, the machine had to examine all three components in order to have correction predictions. From our observations, the performance on the prediction dataset was remarkably enhanced when the noise was added, which supports our ideas. \begin{figure}[thb!] \centering \includegraphics[width=0.49\textwidth]{figure/activation.pdf} \caption{Comparison between three activation function applied in the first layer of the qCNN classifier.} \label{fig:acti} \end{figure} \section{Conclusions} In summary, we classify topological phases with distinct Chern numbers via two types of machine-learning algorithms. For the unsupervised part, we propose a quaternion-based convolution to transform the topological states into the input data. With this convolution, distinct topological states are successfully classified by PCA, a linear machine for classification. We then go to the supervised learning part where, in contrast to the conventional CNN, we successfully use the qCNN to classify different topological phases. This work demonstrates the power of quaternion-based algorithm, especially for the topological systems with the Chern number as the topological invariants. \begin{acknowledgments} This study is supported by the Ministry of Science and Technology (MoST) in Taiwan under grant No. 108-2112-M-110-013-MY3. M.R.L. and W.J.L. contributed equally to this work. \end{acknowledgments} \section{Introduction} The phase classification using machine-learning (ML) based techniques has been attracting intense attentions since the pioneering work in 2017\cite{carrasquilla2017}. In addition to the classical phase detections\cite{bedolla2021,mehta2019} where each phase is well defined by the corresponding order parameters, detecting topological phase transitions\cite{bedolla2021} is interesting and challenging\cite{beach2018} due to the lack of local order parameters. Recently, the phase detections and classifications have been performed via different ML techniques for classifying various topological invariants\cite{yoshioka2018,carvalho2018,balabanov2020,balabanov2021,greplova2020,greplova2020,ho2021,zhang2021,narayan2021,yu2021, zhang2017,cheng2018,sun2018,zhang2020,carvalho2018,kerr2021,kaming2021,ho2021,rem2019,che2020,chung2021,ming2019quantum,zhang2018,sun2018, holanda2020,che2020,kerr2021,tsai2021,kerr2021,zhang2020,zhang2017a,mano2019,su2019,lian2019,carvalho2018,ho2021,beach2018,laskowska2018,zhang2019, RN2019,tsai2020,scheurer2020,tsai2021,caio2019}, including the Chern number\cite{zhang2017,cheng2018,sun2018,zhang2020,carvalho2018,kerr2021,kaming2021,ho2021,rem2019,che2020,chung2021,ming2019quantum}, winding number\cite{zhang2018,sun2018,holanda2020,che2020,kerr2021,tsai2021,kerr2021}, $\mathbb{Z}_2$ index\cite{zhang2020,zhang2017a,mano2019,su2019,lian2019,carvalho2018,ho2021,beach2018,laskowska2018,zhang2019, RN2019,tsai2020,scheurer2020,tsai2021,caio2019}, to name a few. In addition to the applied ML architectures, the forms of the inputs for training the machine also play a crucial role in determining the resulting performance of the topological phase detections\cite{beach2018}. For the topological systems with the Chern numbers or the winding numbers as the topological invariants, various types of inputs are used to perform the phase classifications. For instance, researchers in Kim's group introduced quantum loop topography (QLT) to construct multi-dimensional images from raw Hamiltonians or wave functions as inputs\cite{zhang2017,zhang2020}. Zhai's group collected the Bloch Hamiltonians into an arrays to feed their machines\cite{zhang2018,sun2018}. The real-space particle densities and local density of states were used as inputs by Cheng \textit{et al}.\cite{cheng2018}. Carvalho \textit{et al}. fed the local projections of the density matrix to the machine \cite{carvalho2018}. From cold-atom experiments, momentum-space density images were generated as inputs by Rem \textit{et al}.\cite{rem2019}. The time-of-flight images\cite{ho2021,kaming2021}, spatial correlation function\cite{ho2021} and density–density correlation function\cite{ho2021} were also used as inputs. The density profiles formed in quantum walks were proposed as appropriate inputs for training in Ming \textit{et al}'s work\cite{ming2019quantum}. Furthermore, several works had tried the spin configurations\cite{kerr2021} and the Bloch Hamiltonians over the Brillouin zone (BZ) as inputs\cite{che2020,kerr2021}. For these forms of inputs mentioned above, various ML techniques with distinct real-valued neural networks have been applied to discriminate different topological phases. As the development of artificial neural networks becomes mature, generalization from real-valued neural networks to complex-valued ones is undertaken that anticipates a raise of representation capability of machines~\cite{trabelsi2017,gaudet2018deep}. Specifically, a quaternion contains one real part and three imaginary parts so that a quaternion-based neural network\cite{GR2020,isokawa2009,parcollet2020,matsui2004} is expected to have remarkable performance on handling the processing of data with more than two degrees of freedom such as the color images (RGB channels) and the descriptions of 3D systems (xyz coordinates). There have been various proposals about quaternion-based neural networks in ML techniques and applications in computer science, such as the quaternion convolutional neural network (qCNN)\cite{zhu2019,hongo2020,gaudet2018deep}, quaternion recurrent neural network\cite{parcollet2018quaternion}, quaternion generative adversarial networks\cite{grassucci2021}, quaternion-valued variational autoencoder\cite{grassucci2020}, quaternion graph neural networks\cite{nguyen2020}, quaternion capsule networks\cite{ozcan2020} and quaternion neural networks for the speech recognitions\cite{parcollet2018}. However, the ML-related applications of the quaternion-based neural networks on solving problems in physics are still limited, especially in the topological phase detections, even though the quaternion-related concepts have been applied in some fields in physics~\cite{girard1984,girard2007,girard2018}. \begin{figure}[th!] \centering \includegraphics[width=0.46\textwidth]{figure/QWZ.pdf} \caption{Spin vectors in the Brillouin zone, $k_x,k_y\in (-\pi,\pi]$, with Chern number (a) \(C = 1\), (b) \(C=2\), (c) \(C = 3\) and (d) \(C=4\).} \label{fig:spin_conf} \end{figure} In this work, we perform the Chern-insulator classifications from both supervised- and unsupervised-learning aspects based on the inputs transformed via the quaternion algebra. For the unsupervised learning, we firstly encode the quaternion-transformed eigenstates of Chern insulators via a convolution function as inputs and study them using the principal component analysis. We found that using only the first two principal elements is not enough to fully classify the Chern insulators, consistent with Ming's work\cite{ming2019quantum}. Further studies show that the performance can be improved by including more principal components. For the supervised learning, we construct a quaternion-based neural network in which the first layer is a quaternion convolutional layer. We then show that this quaternion-based machine has better performance than a conventional CNN machine. Our machine is good not only at testing but also at identifying data that have no common structures as in training data. The good performance can be partially attributed to the similarities between the formula of the Berry curvatures and the quaternion algebra. Therefore, our work demonstrates the power of the quaternion algebra on extracting relevant information from data, paving the way to applications of quaternion-based algorithm ML techniques in topological phase classifications. The outline of the remaining part of this work is as follows. In section II, we introduce the model Hamiltonian, generating the training data for our classification tasks, and the quaternion convolution layer used in this work. PCA analysis of the quaternion-transformed eigenstates is discussed in Sec. III. The data preparations, the network structures and the performance of the quaternion-based supervised learning task are given in Sec. IV. Some further discussions on our neural networks are given in Sec. V. Finally, a conclusion is provided in Sec. VI. \section{Model and quaternion convolutional layer} \subsection{Model} A generic two-band Bloch Hamiltonian with the aid of the identity matrix $\sigma_0$ and Pauli matrices \(\pmb{\sigma}=(\sigma_1, \sigma_2,\sigma_3)\) is written as \begin{equation}\label{eq:2band} \mathcal{H}(\vec{k}) = h_0(\vec{k}) \sigma_0 + \mathbf{h}(\vec{k})\cdot\pmb{\sigma}, \end{equation} where $\vec{k}=(k_x,k_y)$ is the crystal momentum in the 2D BZ($={k_x,k_y\in(-\pi,\pi]}$), and the vector \(\mathbf{h} = (h_1,h_2,h_3)\) acts as an external magnetic field. Sometimes we will omit arguments $\vec{k}$ when writing $\mathbf{h}(\vec{k})$ for brevity in $k$-space. The eigenstate of the upper (lower) band will represent the spin pointing antiparallel (parallel) to \(\mathbf{n} = \mathbf{h}/\abs{\mathbf{h}}\). Since $h_0(\vec{k})$ does not affect the pointing direction, it will be ignored in the remaining part of this section. The function $\mathbf{n}$ embeds the topology of the system, and we firstly discuss the topological structure in $\mathbf{n}$. Here, the topological invariant is the Chern number $C \in \mathbb{Z}$ with the formula: \begin{equation}\label{chern} C = \frac{1}{4\pi}\int_{\text{BZ}} \mathbf{n}\cdot(\partial_{k_x}\mathbf{n}\times\partial_{k_y}\mathbf{n})d\vec{k}, \end{equation} where the integrand is the Berry curvature and the integration is over the first BZ. \begin{figure}[b] \centering \includegraphics[width=0.45\textwidth]{figure/HighChernNumber.pdf} \caption{The Chern number with various \(m\) and \(c\).} \label{fig:HighChern} \end{figure} We construct the normalized spin configurations $\mathbf{n}(\vec{k})$ based on the following models. For topological systems, we choose the Hamiltonian with \(\mathbf{h}=\mathbf{h}^{(c)}\), where \begin{equation}\label{eq:HighChern} \mathbf{h}^{(c)}(\vec{k})= \begin{pmatrix} \ \mathrm{Re}\big[(\sin{k_x}-i\sin{k_y})^c\big]\\ -\mathrm{Im}\big[(\sin{k_x}-i\sin{k_y})^c\big]\\ \cos{k_x} + \cos{k_y} + m \end{pmatrix} \end{equation} with positive integer \(c\) and real parameter \(m\) to control the Chern number. For \(c = 1\), the model is the Qi-Wu-Zhang (QWZ) model \cite{PhysRevB.74.085308}. For a given \(c\), the Chern number $C$ can be either \(0,~c,\text{ or} -c\) depending on the value of \(m\): \begin{equation} C = \begin{cases} \mathrm{sgn}(m)c, & 0<|m|<2, \\ 0, & |m|>2. \end{cases} \end{equation} The Chern numbers with different c's and m's are shown in Fig.~\ref{fig:spin_conf}. Note that $C=0$ stands for a topologically trivial phase and nonzero $C$'s are for nontrivial phases. The topological phase diagram is shown in Fig.~\ref{fig:HighChern}. In this work, the unsupervised learning involves seven topological phases ($C=0,\pm 1,\pm 2, \pm 3$) in Sec. \ref{sec:PCA}, and the supervised learning involves nine topological phases ($C=0,\pm 1,\pm 2, \pm 3, \pm 4$) in Sec. \ref{sec:qcnn}. \subsection{Quaternion convolutional layer} A quaternion number has four components, the first of which stands for the real part and the other three of which stand for the imaginary parts. Given two quaternions $q_1=(r_1,a_1,b_1, c_1)$ and $q_2=(r_2,a_2,b_2,c_2)$, their product $Q=q_1q_2=(R,A,B,C)$ is given by \begin{equation}\label{eq:q_1q_2} \begin{pmatrix} R\\ A\\ B\\ C \end{pmatrix} = \begin{pmatrix} r_1 r_2 - a_1 a_2 - b_1 b_2 - c_1 c_2\\ a_1 r_2 + r_1 a_2 - c_1 b_2 + b_1 c_2\\ b_1 r_2 + c_1 a_2 + r_1 b_2 - a_1 c_2\\ c_1 r_2 - b_1 a_2 + a_1 b_2 + r_1 c_2 \end{pmatrix}, \end{equation} which can be written as the matrix product form \begin{equation}\label{eq:matrix_q} \begin{pmatrix} R\\ A\\ B\\ C \end{pmatrix} = \begin{pmatrix*}[r] r_1 & -a_1 & -b_1 & -c_1 \\ a_1 & r_1 & -c_1 & b_1 \\ b_1 & c_1 & r_1 & -a_1 \\ c_1 & -b_1 & a_1 & r_1 \end{pmatrix*} \begin{pmatrix} r_2\\ a_2\\ b_2\\ c_2 \end{pmatrix}. \end{equation} To implement a quaternion convolutional (q-Conv) layer in numerical programming, we will regard the two quaternions as a $4\times4$ matrix and a $4\times1$ column matrix, respectively: \begin{equation} \label{eq:rep2} q_1 \doteq \begin{pmatrix*}[r] r_1 & -a_1 &-b_1 & -c_1 \\ a_1 & r_1 & -c_1 & b_1 \\ b_1 & c_1 & r_1 & -a_1 \\ c_1 & -b_1 & a_1 & r_1 \end{pmatrix*}\quad \mathrm{and}\quad q_2 \doteq \begin{pmatrix*}[r] r_2\\ a_2\\ b_2\\ c_2 \end{pmatrix*}. \end{equation} More details of quaternion algebra are described in Appendix~\ref{app:qcnn}. A conventional CNN contains a real-valued convolutional layer to execute the convolution of the input and the kernel. Let the input $F$ have the shape: $H_i\times W_i\times C_i$ (Height \(\times\) Width \(\times\) Channel) and the shape of the kernel $K$ be $H_{k} \times W_{k} \times C_i \times C_f$. The convolution will produce an output \(O\), $O=F\ast K$, whose elements are \begin{equation}\label{eq:C-convo} O_{i',j',t'} =\sum_{i}^{H_k} \sum_{j}^{W_k} \sum_{t}^{C_i} F_{i'+i-1,j'+j-1,t}\cdot K_{i,j,t,t'}. \end{equation} Here the stride is assumed to be \(1\) both in the width and the height directions. The indices $i$ and $j$ are spatial indicators, $t$ is the index of channel in the input feature map and $t'$ is the kernel index. The shape of the output will be $(H_{i}-H_{k}) \times (W_{i}-W_{k}) \times C_f$. Assume that the input has four components. To uncover the entanglement among components through CNN, we will utilize the quaternion product. Now, we introduce another dimension–Depth–which is four, as a quaternion number of four components. Both of the input $F$ and the kernel $K$ have Depth of four as two quaternion numbers. The product of $F$ and $K$ will have Depth of four as a quaternion in Eq.~(\ref{eq:q_1q_2}). Referring to Eq.~(\ref{eq:rep2}) where we show a matrix representation to implement quaternion algebra and thinking of $F$ as $q_1$ and $K$ as $q_2$ in Eq.~(\ref{eq:rep2}), we transform the Depth-four input $F$ into a 4x4 matrix, $F^{(l,s)}$, and keep the kernel $K$ still of Depth 4, $K^{(l)}$, where $l, s=1,…,4$. The product of $F$ and $K$, say $O$, will have Depth of four as shown in Eq.~(\ref{eq:Q-convo}). We remark that the products between components of $F$ and $K$ are convolution operations as Eq.~(\ref{eq:C-convo}). \begin{equation}\label{eq:Q-convo} O_{i',j',t'}^{(s)} = \sum_{l}^{4}\sum_{i,j,t} F^{(s,l)}_{i'+i-1,j'+j-1,t}\cdot K^{(l)}_{i,j,t,t'}, \end{equation} \begin{figure}[b!] \centering \includegraphics[width=0.47\textwidth]{figure/q_conv_layer.pdf} \caption{Illustration of a quaternion convolutional layer. On the left, we start with the input $q_1$ having four quaternion components ((yellow, red, green, blue) stands for ($r_1$, $a_1$, $b_1$, $c_1$)). In the middle, $q_1$ is permuted to construct $\{F^{(\cdot, l)}\}_{l=1}^{4}$ on which the convolution with four kernels $\{K^{(l)}\}_{l=1}^{4}$ is performed. A summation is taken for each Depth dimension to obtain the output feature map $O$ on the right.} \label{fig:Q-filter} \end{figure} More specifically, we consider an input data as $q_1$ (four color squares on the left of Fig. \ref{fig:Q-filter}) and four kernels encoded in $q_2$, given in the following \begin{equation} \left\{\begin{matrix} q_1 \doteq & (r_1~a_1~b_1~c_1)^T &\\ q_2 \doteq & (r_2~a_2~b_2~c_2)^T & =: K^{(\cdot)}. \end{matrix}\right. \end{equation} The output feature maps $O\doteq (R~A~B~C)^T$ is then calculated based on Eq.~(\ref{eq:q_1q_2}). As the first step, we permute the order of $q_1$ to obtain \begin{eqnarray} F^{(\cdot,1)}=: \begin{pmatrix} r_1\\ a_1\\ b_1\\ c_1 \end{pmatrix}, F^{(\cdot,2)}=: \begin{pmatrix*}[r] -a_1\\ r_1\\ c_1\\ -b_1 \end{pmatrix*},\\\nonumber F^{(\cdot,3)}=: \begin{pmatrix*}[r] -b_1\\ -c_1\\ r_1\\ a_1 \end{pmatrix*}, F^{(\cdot,4)}=: \begin{pmatrix*}[r] -c_1\\ b_1\\ -a_1\\ r_1 \end{pmatrix*} \end{eqnarray} (see the four sets of sqaures in the middle of Fig. \ref{fig:Q-filter}). We then convolute those four quaternions ($F^{(\cdot,l)}$ with $l =1,2,3$ and 4) with four kernels ($K^{(l)}$ with $l = 1,2,3$ and 4) in the following way: $$ \left\{ \begin{matrix} F^{(\cdot,1)}K^{(1)} &\doteq& \begin{pmatrix} r_1r_2&a_1r_2&b_1r_2&c_1r_2 \end{pmatrix}^T\\ F^{(\cdot,2)}K^{(2)} &\doteq& \begin{pmatrix} -a_1a_2&r_1a_2&c_1a_2&-b_1a_2 \end{pmatrix}^T\\ F^{(\cdot,3)}K^{(3)} &\doteq& \begin{pmatrix} -b_1b_2&-c_1b_2&r_1b_2&a_1b_2 \end{pmatrix}^T\\ F^{(\cdot,4)}K^{(4)} &\doteq& \begin{pmatrix} -c_1c_2&b_1c_2&-a_1c_2&r_1c_2 \end{pmatrix}^T \end{matrix} \right. $$ , as shown in the middle of Fig. \ref{fig:Q-filter}. Finally, we sum over the above four quaternions to get the output feature maps $O$, as shown on the right of Fig. \ref{fig:Q-filter}. $$ O := \begin{pmatrix} R\\ A\\ B\\ C \end{pmatrix} = \begin{pmatrix} r_1 r_2 - a_1 a_2 - b_1 b_2 - c_1 c_2\\ a_1 r_2 + r_1 a_2 - c_1 b_2 + b_1 c_2\\ b_1 r_2 + c_1 a_2 + r_1 b_2 - a_1 c_2\\ c_1 r_2 - b_1 a_2 + a_1 b_2 + r_1 c_2 \end{pmatrix}. $$ \section{principal component analysis}\label{sec:PCA} Principal component analysis (PCA) is a linear manifold learning that is to find the relevant basis set among data\cite{jolliffe2016principal,ma2012manifold}. We prepare eigenstates $\ket{u_\pm}$ of Eq.~(\ref{eq:2band}), where $+\ (-)$ stands for the upper (lower) band. For a topological state, the phase cannot be well-defined over the whole BZ. Therefore, we can divide the whole BZ into two parts, in each part of them the topological wave function has continuously well-defined phase. We then choose two regions according to the sign of $h_3$ in Eq. (\ref{eq:HighChern}): \begin{align} \begin{split} \ket{u_+} & \doteq \frac{1}{\sqrt{2h_+(h_{+}+h_3)}}\mqty(h_{+}+h_3 \\ h_1+i h_2) \\ \ket{u_-} & \doteq \frac{1}{\sqrt{2h_-(h_{-}+h_3)}} \mqty(-h_1+ih_2 \\ h_{-}+h_3) \end{split} \quad ,~ h_3\geq 0, \label{eq:gauge1} \end{align} and \begin{align} \begin{split} \ket{u_+} & \doteq \frac{1}{\sqrt{2h_{+}(h_{+}-h_3)}}\mqty(h_1-ih_2 \\ h_{+}-h_3) \\ \ket{u_-} & \doteq \frac{1}{\sqrt{2h_{-}(h_{-}-h_3)}} \mqty(h_{-}-h_3 \\ -h_1-ih_2) \end{split} \quad ,~ h_3 < 0, \label{eq:gauge2} \end{align} where $h_{\pm} =\pm \sqrt{h_1^2+h_2^2+h_3^2}$. In this choice of gauge, the first (second) component of $\ket{u_{+}}$ $(\ket{u_-})$ is real-valued when $h_3\geq0$, and the second (first) component of $\ket{u_{+}}$ $(\ket{u_-})$ is real-valued when $h_3<0$. By translating $\ket{u_\pm} \doteq (\alpha_\pm, \beta_\pm)^T$ with $\alpha_\pm, \beta_\pm\in\mathbb{C}$, into a quaternion number of four components, we have \begin{equation}\label{eq:transf} q_\pm := \mathrm{Re}(\alpha_\pm) + \mathrm{Im}(\alpha_\pm)\hat{\mathbf{i}} + \mathrm{Re}(\beta_\pm)\hat{\mathbf{j}} + \mathrm{Im}(\beta_\pm)\hat{\mathbf{k}}. \end{equation} (a brief introduction of the quaternion number can be seen in Appendix \ref{app:qcnn}). To see the correlation of states over $\vec{k}$, we define the quantity $F$ to be the quaternion-based convolutions: \begin{equation}\label{eq:Ne} \begin{split} F(\vec{p}) := q^*_{+}&\circledast q_{+}[\vec{p}] - q^*_{-}\circledast q_{-}[\vec{p}] \quad\mathrm{with}\\ q_{\pm}^*&\circledast q_{\pm}[\vec{p}] := \sum_{\vec{k}\in \mathrm{BZ}} q_{\pm}^*(\vec{k})q_{\pm}(\vec{p} - \vec{k}), \end{split} \end{equation} where $q^*$ is the conjugate of $q$. It can be proved that $F$ is real-valued. Therefore, $F(\vec{p})$ of all $\vec{p}$ in the BZ based on a given Hamiltonian can be analysed by using PCA. \begin{figure}[b!] \centering \includegraphics[width=0.45\textwidth]{figure/nematic2.pdf} \caption{The maps of the function \(F\) without noise in the BZ. Three rows are for \(c = 1, 2\ \mathrm{and}\ 3\) in Eq.~\ref{eq:HighChern} from top to bottom; four columns from left to right are for $m = -3,-1, 1$ and 3. The corresponding Chern number $C$ is tagged with each panel.} \label{fig:nematic} \end{figure} We collected various $F$ of all $\vec{k} \in $BZ within seven topological phases as the dataset for PCA. For each topological phases, 30 $F$'s were prepared, so the total amount of data was 210. The data for six non-trivial phases were generated based on Eq. (\ref{eq:HighChern}) with $m=\pm 1$ (the sign of $m$ determines the sign of $C$). For the trivial phase, we prepared five data points from each of six combinations of $\{c,m\}$, where $c\in\{1,2,3\}$ and $m\in\{3,-3\}$, and then there are totally 30 data. To augment the number of data, we add Gaussian noises $\delta\mathbf{h}$ at every $\vec{k}$ of the model [Eq. (\ref{eq:HighChern})] such that $\mathbf{h}\to\mathbf{h}+\delta\mathbf{h}$ without closing the band gap. In Fig.~\ref{fig:nematic}, we present various noiseless $F$ generated from Eq. (\ref{eq:HighChern}) with different $c$ and $m$. \begin{figure}[t!] \centering \includegraphics[width=0.49\textwidth]{figure/2DPCA.pdf} \caption{PCA of seven topological phases with various noise. The symbols with corresponding Chren numbers are marked in the legend.} \label{fig:2DPCA} \end{figure} It is notable that $F$ for $C = 0$ are featureless, $F$ for $C = \pm 1$ have a dipole moment, and $F$ for $C = \pm 2$ have a quadruple moment, and $F$ for $C = \pm 3$ seemingly have a primary dipole and a secondary quadruple moment. The remarkable features imply that the convolution function $F$ is a good choice for topological classifications. We examine data with the standard deviation (SD) equal to 0, 0.1, 0.2 and 0.3 respectively, and show the first two PCs of 210 pieces of data for each SD in Fig.~\ref{fig:2DPCA}. In Fig.~\ref{fig:2DPCA}, it is evident that data are clustered into four groups and their variances increase with SD. PCA is successful to separate different topological phases into different clusters. However, some clusters contain two topological phases of Chern numbers: $\{+1,-3\}$, $\{-1,+3\}$, and $\{+2,-2\}$. The $C$ modulo 4 resemblance has also be observed in a previous study~\cite{ming2019quantum}. \begin{figure}[b!] \centering \includegraphics[width=0.5\textwidth]{figure/PCs.pdf} \caption{Magnitude of projection (logarithmic scale) from non-trivial data onto first six principal components. Inset: The first 16 principal values of PCA. (normalized by maximal $\lambda_1$)} \label{fig:ninePC} \end{figure} We find that including more PCs helps separate different classes in each cluster. Figure \ref{fig:ninePC} shows first six PCs of data in topologically non-trivial phases, where PCx denotes the x-th PC component. One can find that PC1 and PC2 in each pair of $\{+1,-3\}$, $\{-1,+3\}$, and $\{+2,-2\}$ are nearly identical, as also shown in Fig.~\ref{fig:2DPCA}. By including PC3 ~ PC6 into the analysis, all topological classes are completely classified. Via the proposed convolution, topological states can be successfully classified by using PCA, a linear machine for classification. \begin{figure*}[t!] \centering \includegraphics[width=0.945\textwidth]{figure/qCNN-net2.pdf} \caption{Framework of (a) the CNN and (b) the qCNN classifier. The shape of data is $L_x \times L_y \times$ (number of channels). The number appended after the symbol ``@'' stands for the number of filters convolved with the data, and the number just above is the filter size. The number of tuneable network parameters of CNN (qCNN) is 24,252 (19,350).} \label{fig:QCNN} \end{figure*} \section{Supervised learning of CNN and the qCNN}\label{sec:qcnn} \subsection{Datasets} The input data are normalized spin configurations $\mathbf{n}$, laying on a $40\times 40$ square lattice with periodic boundary conditions, and their corresponding topological phases are labels with one-hot encoding. We prepared four datasets: training, validation, testing and prediction dataset. The first three datasets are well known in conventional deep learning procedure~\cite{chollet2021deep}. To understand whether our machine can also classify unseen spin configuration, we prepare a prediction dataset that includes a few types of spin configurations never seen by the machine during the training process. The data pool containing training and validation datasets is constructed as follows. Based on the Eq.~(\ref{eq:HighChern}), we firstly prepared 5760 data points of $\mathbf{n}$ of nine topological phases with Chern number ranging from -4 to 4 so that each phase contain 640 data points. Additionally, we add 360 data points for spin vortex which belongs to the trivial phase. A spin-vortex has an in-plane spin texture that winds around a center, which is generated by setting one of three components in Eq.~(\ref{eq:HighChern}) to be zero. By including spin vortices, the machine learn the difference between 3D winding (non-trivial) and 2D winding (trivial) spin configurations (more details are described in the Appendix \ref{app:dataset}). After the training process, the trained machine is scored by a testing dataset with the same composition of nine phases as that in the training (and validation) dataset. Importantly, without changing the topologies, the Gaussian distributed random transition and random rotation imposed on these three datasets can increase the diversity of dataset and enhance the ability of generalization of the trained machine. The prediction dataset contains six categories of spin configurations. The first category is generated with $m$ uniformly distributed from $+3$ to $-3$. In the second and the third categories, we change the sign of $n_z$ (the second category) and swapping $n_y$ and $n_z$ of $\mathbf{n}$ (the third category). Finally, we consider three categories for trivial states, which are ferromagnetic (FM), conical $(\epsilon\neq 0)$ and helical $(\epsilon=0)$ states, based on the following formula: \begin{equation}\label{eq:conical} \mathbf{n}_{\mathrm{spiral/conical}}(\vec{k})= \begin{pmatrix} \sqrt{1-\epsilon^2}\cos{(k_x+ k_y)}\\ \sqrt{1-\epsilon^2}\sin{(k_x+ k_y)}\\ \epsilon \end{pmatrix}. \end{equation} FM can be viewed as 1D uncompleted winding configuration while conical and helical can be viewed as 2D uncompleted ones. In total, we prepared six categories for the prediction dataset. More details about data preparations will be described in Appendix \ref{app:dataset}. For the conventional CNN, we use $\mathbf{n}$ as the input data. For the qCNN, in order to feed the input data into the qCNN classifier, we transform the 3D spin vector into an unit pure quaternion, \begin{equation}\label{eq:encoding} (n_x,n_y,n_z)\in\mathbb{R}^3 \mapsto (0, n_x,n_y,n_z)\in\mathbb{H}, \end{equation} where the scalar part (the first component) is zero and the vector part is $\mathbf{n}$. \subsection{network structure and performance} We implement a qCNN classifier with a quaternion convolution (q-Conv) layer as the first layer [see red dotted cube in Fig.~\ref{fig:QCNN}(b)], and the operations in a q-Conv layer are based on the quaternion algebra. Then the next three layers are conventional 3D convolutional layers. Details of the quaternion algebra and the keynote of a qCNN are explained in Appendix~\ref{app:qcnn}. For comparison, we also set up a conventional CNN classifier by replacing the q-Conv layer by a conventional 2D convolutional layer, and appending three conventional 2D convolutional layers to the first layer. The architecture of these two classifiers are shown in Fig.~\ref{fig:QCNN}. Note that the data in the qCNN has one more rank than that in CNN, and in numerical programs, this extra dimension---\emph{Depth} is used to store the quaternion [referred to Eq. (\ref{eq:encoding})]. As you can see in the Fig.~\ref{fig:QCNN}, the architecture of the qCNN is one layer more than the CNN's, the total network parameters of the qCNN is however less than the CNN's. This is one advantage of the qCNN over the conventional CNN. In order for classifiers to satisfy some physically reasonable conditions, three special designs are needed. Firstly, we extend the $k$ points out of the BZ by padding the input data according to the periodic boundary conditions~\cite{PhysRevB.99.075113}. \begin{figure}[b!] \centering \includegraphics[width=0.45\textwidth]{figure/overlap.pdf} \caption{Schematic of ``overlap'' convolution (red solid) and ``non-overlap'' convolution (blue solid) from a $3\times3$ filter (black dotted) over data. The blue solid square is a signal movement from the filter, and the size of stride is the same as the length of filter, thus each movement of this filter is ``non-overlap''.} \label{fig:overlap} \end{figure} Secondly, a convolutional layer (q-Conv layer in the qCNN) with arctangent activation function is then adopted, which contains 27 filters of kernel size $2\times 2$. This layer executes an ``overlapping'' feature mapping. (Figure~\ref{fig:overlap} illustrates how the ``overlapping'' and ``non-overlapping'' feature mapping can be manipulated by varying the size of stride.) Thirdly. we shrink the $k$ space of data into ``a point" by three non-overlapping convolutional layers. In the qCNN, three 3D convolutional layers are applied while keeping four Depths independent. In particular, we additionally append a 2D convolutional layer to make combination Depths-wisely. Specifically, this layer with 9 filters of kernel size $4\times1$ transforms data from $4\times9$ to $1\times9$, where each of nine neurons corresponds to one topological class. In comparison, the conventional CNN shrinks the $k$-space to a point in each of nine \emph{Channels}, where the dimension is integrated out after the convolution. More details and descriptions are relative to the SEC. \ref{sec:discu}. \begin{figure}[b!] \centering \includegraphics[width=0.46\textwidth]{figure/accuracy2.pdf} \caption{Learning curves of the qCNN and CNN classifiers. By applied \emph{Dropout}, the validation is greater than the training.} \label{fig:learningCurve} \end{figure} In the following, we perform both the qCNN and conventional CNN trainings. The learning curves of both machines are shown in Fig.~\ref{fig:learningCurve}. The CNN machine (orange and light orange lines) jumps over a big learning barrier at around the $700^{th}$ epoch. After that, the training and the validation accuracy (orange and light orange line respectively) are separated and do not converge up to end of this training process. Even though the same training (and validation) dataset is used in the training process, the learning curves of the qCNN machine (blue and light blue lines) are qualitatively different. The training and the validation accuracy are separated around $90^{th}$ epoch, but the difference between these two accuracies decreases with increasing epochs. After the training procedure finished, the qCNN (CNN) machine gets 99.67\% (94.12\%) testing accuracy. This difference in accuracy results from the spin-vortex dataset, where the qCNN works well but CNN dose not. The trained machines are ready to do prediction, and the result is shown in Fig.~\ref{fig:performpred}. \begin{figure}[b!] \centering \includegraphics[width=0.46\textwidth]{figure/predictset.pdf} \caption{The performance of the qCNN (blue line) and CNN (red dashed line) on the prediction datasets. Numbers tagged are the values of the accuracies. Standard deviations (by error bars) are also provided. The qCNN outperforms the conventional CNN on all prediction datasets, especially on three spin-vortex ones} \label{fig:performpred} \end{figure} In Fig.~\ref{fig:performpred}, since the first category contains $\mathbf{n}$ of uniformly distributed $m$, where a few data points are very close to the phase boundaries $m \approx \{0, \pm 2 \}$, the accurate rate of the the qCNN is slightly low at $96\%$. For the second and third categories, we choose $m=\pm 1$, away from the phase transition points, and the performance is nearly perfect. For the uncompleted winding configurations, the qCNN, different from the conventional CNN, can accurately classifies FM, helical and conical states after learning the spin-vortex states. This is the main advantage of the qCNN over the conventional CNN, which is expected to result from the quaternion algebra. \section{discussions}\label{sec:discu} On PCA the reason why we choose spinor state vector $\ket{u}$, not the spin normalized vector $\mathbf{n}$ as data is that the gauge discontinuities exist in the non-trivial $\ket{u}$, but not in $\mathbf{n}$'s. It turns out that the function $F$ can detect the existence of this discontinuity so that trivial and non-trivial states can be distinguished by PCA. Above statement is further supported by the relation between $\ket{u}$ and $\mathbf{n}$ in pure state, which is known as \begin{equation}\label{eq:sigma} \mathbf{n}(\vec{k}) = \bra{u_+(\vec{k})}\pmb{\sigma}\ket{u_+(\vec{k})}. \end{equation} In Eq. (\ref{eq:sigma}), the $U(1)$ ambiguity of $\ket{u}$ is eliminated effectively, thus, there is no discontinuities to be detected by PCA in $\mathbf{n}$. There is another issue if we use n as the data in PCA. If we encode $\mathbf{n}=(n_x, n_y,n_z)$ into a quaternion in the following form, \[q := (0,~n_x,~n_y,~n_z),\] the corresponding convolution [Eq. (\ref{eq:Ne})] will not depend on the sign of $m$. It can be shown that there exists one point $\vec{k}'\to\vec{p}_0 - \vec{k}+\pi$ in BZ such that \[n_z(\vec{k}, m_0)n_z(\vec{p}_0-\vec{k}, m_0) = n_z(\vec{k}', -m_0)n_z(\vec{p}_0-\vec{k}', -m_0).\] That is, for two spin configurations with opposite sign $m_0$, their $n_z$ products are equal at $\vec{k}$ and $\vec{k}'$ respectively. As shown in Eq.~(\ref{eq:HighChern}), m shows up only in the $n_z$ component. Therefore, after the integration over the whole BZ, the convolution is independent of the sign of $m$ and there is no feature for PCA to discriminate distinct topological states. \begin{figure}[b!] \centering \includegraphics[width=0.47\textwidth]{figure/filter.pdf} \caption{(a) Three nearest neighborhood spin vectors will contribute a solid angle, and (b) four nearest neighbors are enclosed by the filters in the first convolutional layer.} \label{fig:filter} \end{figure} On qCNN, there are several possible factors promoting the performance of our supervised learning machine. The first one is that the size of filter in the first convolutional layer is 2 × 2 with stride = 1, which means the machine can collect spin information among four nearest neighbors [see Fig. \ref{fig:filter}(b)]. We know that the Chern number is the integral of the Berry curvature in the BZ, and the Berry curvature is twice of the solid angle. A solid angle $\Omega$ subtended by three normalized vectors $\vec{a}$, $\vec{b}$, $\vec{c}$, can be calculated using the following equation: \begin{equation}\label{eq:solid} \tan{\frac{\pmb{\Omega}}{2}}=\frac{\abs{\vec{a}\cdot(\vec{b}\times \vec{c})}}{1+\vec{a}\cdot \vec{b}+\vec{b}\cdot \vec{c}+\vec{c}\cdot \vec{a}}. \end{equation} Our choice of the size of the filter in the first hidden layer is the minimal of $2 \times 2$ that mixes only the nearest-neighboring spins. In this way, it is very possible to enforce the machine to notice the solid angle extended in this plaquette. The second factor is the quaternion product. Recall that the conventional CNN might correlate spins $\mathbf{n}'s$ in neighboring $\vec{k}'s$ due to the feature map through the kernel. However, the map does not mix the components of spins. In comparison, the qCNN is more efficient for it directly entangle spins via the quaternion product. It is this entanglement of spin components by the quaternion product that makes the scalar and vector products in calculating the solid angle (see Eq.~ (\ref{eq:solid})) become possible to be realized by the machine. As a solid angle involves at least three spins and the feature map by the kernel is just linear, a nonlinear transformation is crucial to create high-order (three spins) terms in the expansion. Based on this argument, the third factor is the non-linear activation function \cite{lin2017}, which is an arctangent function in this work. Based on Eq.~ (\ref{eq:solid}), the calculation of a solid angle involves the arctan operation. Therefore, we expect that using the arctangent function as the activation function can further help the machine to learn correct representations. This idea is further supported by the results shown in Fig. \ref{fig:acti}, where the arctangent activation function outperforms the ReLU and tanh activation functions over nine different datasets. In summary, several factors are combined to enhance the performance of our machine as follows. The quaternion-based operations in the q-Conv layer mix not only different components of a spin $\mathbf{n}(\vec{k}_0)$ but also neighbouring spins, say $\mathbf{n}(\vec{k}_0)$ and $\mathbf{n}(\vec{k}_1)$. When these linear combinations are fed into the non-linear activation functions in our qCNN, the output can be viewed as an expansion of a non-linear function, which may contain a term having both the scalar- and vector-product of neighboring spins, similar to that in Eq.~ (\ref{eq:solid}). Therefore, in the optimization process, combined with the effect of choosing the size of filter to be $2 \times 2$, the machine may keep increasing the weight of a solid-angle-related term and eventually learn to classify the topological phases based on the calculated solid angles. Also, adding some noises to the training dataset helped our supervised-learning machine to learn the generic feature of our data. We found that when the training data was generated directly from Eq. (\ref{eq:HighChern}) without adding any noise, the machine worked well for training and testing datasets but had poor performance on all the prediction dataset. This could be understood by noting that topology is determined by the sign of $m$, which appears only in the z component in Eq. (\ref{eq:HighChern}). By using the dataset without noises, the machine may learn to classify states only by looking at the $z-$component but not extracting the full information from the full configuration. When we gave random transition and rotation to each training data, the machine had to examine all three components in order to have correction predictions. From our observations, the performance on the prediction dataset was remarkably enhanced when the noise was added, which supports our ideas. \begin{figure}[thb!] \centering \includegraphics[width=0.49\textwidth]{figure/activation.pdf} \caption{Comparison between three activation function applied in the first layer of the qCNN classifier.} \label{fig:acti} \end{figure} \section{Conclusions} In summary, we classify topological phases with distinct Chern numbers via two types of machine-learning algorithms. For the unsupervised part, we propose a quaternion-based convolution to transform the topological states into the input data. With this convolution, distinct topological states are successfully classified by PCA, a linear machine for classification. We then go to the supervised learning part where, in contrast to the conventional CNN, we successfully use the qCNN to classify different topological phases. This work demonstrates the power of quaternion-based algorithm, especially for the topological systems with the Chern number as the topological invariants. \begin{acknowledgments} This study is supported by the Ministry of Science and Technology (MoST) in Taiwan under grant No. 108-2112-M-110-013-MY3. M.R.L. and W.J.L. contributed equally to this work. \end{acknowledgments}

656a599a18e74671979f5a604d06de8a15fb1ac1

\section{Introduction} We recall notations and terminologies used in \citep{hou}, where it was shown that the following holds. \begin{theorem}[{{\citep[Theorem 2.1]{hou}}}] Let a Polish group $\Gamma$ continuously act on a Polish space $X$. Let $S\subseteq\Gamma$ be countable compact, with its Schreier graph $G(S)$ on $X$. For any Baire measurable function $f:X\to\omega$, there is a comeager set of $x\in X$ for which $f[N_{G(S)}(x)]$ is finite. In particular, $f$ is not domatic at any such vertex $x$. \end{theorem} To summarize, it was found that some $\omega$-regular Schreier graphs avoid Baire measurable $\omega$-domatic colorings. In this note, we prove a measure-theoretic counterpart to the previous result. \begin{theorem}[Theorem \ref{thm:2.4}] Let $(X,\mu)$ be a standard probability space. Let $G$ be a $\mu$-preserving $\omega$-regular Borel graph on $X$. Then $G$ admits a $\mu$-measurable $\omega$-domatic coloring. \end{theorem} We say that a locally countable Borel graph $G$ on a standard probability space $(X,\mu)$ is \emph{$\mu$-preserving} if $G$ can be covered by countably many $\mu$-preserving Borel automorphisms of $X$. \subsection*{Acknowledgements} The author would like to thank Clinton Conley and Felix Weilacher for their helpful discussions on the subject. \section{Proofs} \begin{lemma}\label{lem:2.1} Let $G$ be an $\omega$-regular Borel graph on a standard Borel space $X$. \begin{enumerate} \item[(1)] If $\mu$ is a Borel probability measure on $X$ such that $G$ is $\mu$-preserving, and if $f:X\to\omega$ is a $\mu$-measurable coloring such that every vertex $x\in X$ sees infinitely many colors in its neighborhood, i.e.\ $\left|f[N_G(x)]\right|=\omega$, then $G$ admits a $\mu$-measurable $\omega$-domatic coloring. \item[(2)] If $\tau$ is a Polish topology on $X$ such that $G$ can be covered by countably many $\tau$-homeomor\-phisms, and if $f:X\to\omega$ is a $\tau$-Baire measurable coloring such that every vertex $x\in X$ sees infinitely many colors in its neighborhood, i.e.\ $\left|f[N_G(x)]\right|=\omega$, then $G$ admits a $\tau$-Baire measurable $\omega$-domatic coloring. \end{enumerate} \end{lemma} \begin{Proof} For (1), start by defining $\nu_0$ to be the unique probability measure on $\omega$ such that for all $n\in\omega$, $\nu_0(\{n\})=2^{-n-1}$. Then let $\nu$ be the Borel probability measure on $\omega^\omega$ defined as the $\omega$-fold i.i.d.\ product of $\nu_0$; explicitly, if $s=(n_0,n_1,\ldots,n_{\ell-1})\in \omega^{<\omega}$ codes the basic open $N_s\subseteq\omega^\omega$, then we put $\nu(N_s)=\prod_{i<\ell}\nu_0(\{n_i\}) =2^{-\Sigma(s)-\ell}$. We claim that for any infinite set $A\subseteq\omega$, the set of $r\in\omega^\omega$ such that $r[A]=\omega$ is $\nu$-conull. For each $n\in\omega$ and $a\in A$, there is a positive probability $\nu_0(\{n\})>0$ that a $\nu$-randomly chosen $r\in\omega^\omega$ satisfies $r(a)=n$. Then for each $n\in\omega$, since the infinitely many events $r(a)=n$ over all $a\in A$ are i.i.d.\ with positive probability, we see that the $\nu$-probability that $n\in r[A]$ has to be one. By $\sigma$-additivity of $\nu$, and since the probability that $r[A]\ne\omega$ is bounded by the sum over all $n\in\omega$ of the probabilities that $n\notin r[A]$ where each summand is $0$, we find that the event $r[A]=\omega$ must be $\nu$-conull as claimed. For each $r\in\omega^\omega$, consider the $\mu$-measurable coloring $r\circ f:X\to\omega$. For every vertex $x\in X$, the set $A_x=f[N_G(x)]\subseteq\omega$ is infinite by assumption, so by the last claim there is a $\nu$-conull set of $r\in\omega^\omega$ such that $\omega=r[A_x]=r\circ f[N_G(x)]$, i.e.\ the coloring $r\circ f$ is $\omega$-domatic at $x$. By Fubini's theorem, the space of $(r,x)\in \omega^\omega\times X$ where $r\circ f$ is $\omega$-domatic at $x$ is then $(\nu\times\mu)$-conull, and so there exists some $r\in\omega^\omega$ such that there is a $\mu$-conull set of $x\in X$ on which $r\circ f$ is $\omega$-domatic. Finally since $G$ is $\mu$-preserving, we may modify such a coloring $r\circ f$ over a $\mu$-null set to produce a $\mu$-measurable $\omega$-domatic coloring as desired. Part (2) is similar, where one replaces the role of $\nu$ in part (1) by considering comeager subsets of $\omega^\omega$ with respect to the standard topology, and one replaces Fubini's theorem with Kuratowski--Ulam \citep[see][Theorem 8.41]{kechris}. We may note that our assumption on $\tau$ implies that the $E_G$-saturation ($E_G$ denoting $G$-connectedness) of $\tau$-meager sets are $\tau$-meager. \end{Proof} Before we use Lemma \ref{lem:2.1} to prove Theorem \ref{thm:2.4}, here are some quick applications. \begin{corollary}[{{\citep[Theorems 3.4 \& 3.5]{hou}}}] Let $G$ be an $\omega$-regular Borel graph, undirected with no self-loops, on a standard Borel space $X$. \begin{enumerate} \item[(1)] If $\mu$ is a Borel probability measure on $X$ such that $G$ is $\mu$-preserving, then $G$ admits a Borel $\omega$-edge-coloring such that $\mu$-almost every $x\in X$ is incident to edges of all colors in $\omega$. \item[(2)] If $\tau$ is a Polish topology on $X$ such that $G$ can be covered by countably many $\tau$-homeomor\-phisms, then $G$ admits a Borel $\omega$-edge-coloring such that a $\tau$-comeager set of $x\in X$ are incident to edges of all colors in $\omega$. \end{enumerate} \end{corollary} \begin{Proof} By Feldman--Moore \citep[see][Proposition 4.10]{kechris-solecki-todorcevic}, $G$ admits a Borel proper $\omega$-edge-coloring, where edges are symmetrically colored with every two edges incident to a common vertex receiving distinct colors. In such a coloring, every vertex is then incident to edges of infinitely many colors, and so we may suitably modify the proof of Lemma \ref{lem:2.1} to produce colorings of the desired forms. \end{Proof} \begin{corollary}[{{\citep[Theorem 4.7]{hou}}}] Let $G_{\succeq}$ be the graph on $[\omega]^\omega$ defined in {\citep[Section 4.4]{hou}}. \begin{enumerate} \item[(1)] If $\mu$ is a Borel probability measure on $[\omega]^\omega$ such that $G_{\succeq}$ is $\mu$-preserving, then $G_{\succeq}$ admits a $\mu$-measurable $\omega$-domatic coloring. \item[(2)] If $\tau$ is the canonical Polish topology on $[\omega]^\omega$, then $G_{\succeq}$ admits a $\tau$-Baire measurable \mbox{$\omega$-domatic} coloring. \end{enumerate} \end{corollary} \begin{Proof} Note that the minimum function $\min:[\omega]^\omega\to\omega$ is Borel and paints every $G_{\succeq}$-neighborhood with infinitely many colors in $\omega$. The result follows by applying Lemma \ref{lem:2.1}. \end{Proof} Now we will prove the promised result. \begin{theorem}\label{thm:2.4} Let $(X,\mu)$ be a standard probability space. Let $G$ be a $\mu$-preserving $\omega$-regular Borel graph on $X$. Then $G$ admits a $\mu$-measurable $\omega$-domatic coloring. \end{theorem} \begin{Proof} For every $n\in\omega$, by \citep[Theorem 4.1]{hou} with $k_n=2^n$ and $\varepsilon_n=2^{-n}$, there is a Borel coloring $f_n:X\to 2^n$ and a Borel set $A_n\subseteq X$ such that $f_n$ is $2^n$-domatic at $A_n$, and $\mu(A_n)\ge 1-2^{-n}$. Fix these objects as such. By the Borel--Cantelli lemma, our assumption that $\mu(X\ssm A_n)\le 2^{-n}$ decays exponentially implies that the set of $x\in X$ lying in infinitely many $X\ssm A_n$ has $\mu$-measure zero. So if we let $A\subseteq X$ be the Borel set of $x\in X$ lying in cofinitely many $A_n$, then $A$ must be $\mu$-conull. For every $n\in\omega$, since the Borel sets $f_n^{-1}(\{i\})$ for $i\in 2^n$ finitely partition $X$, there must be some part $f_n^{-1}(\{i\})\subseteq X$ among the $2^n$ parts with the least $\mu$-measure. We put $D_n=f_n^{-1}(\{i\})$ to be this $\mu$-least measured part, and this gives us that $\mu(D_n)\le 2^{-n}$. By the assumption that $f_n$ is domatic at $A_n$, we also see that $D_n$ dominates $A_n$, i.e.\ for every $x\in A_n$, there exists a neighbor $y\in N_G(x)$ with $y\in D_n$. By the Borel--Cantelli lemma again, the set of $x\in X$ lying in infinitely many $D_n$ is $\mu$-null. If we put $Y\subseteq X$ to be the Borel set of $x\in X$ lying in only finitely many $D_n$, then $Y$ is $\mu$-conull. We define the Borel function $g:Y\to[\omega]^{<\omega}$ such that $n\in g(x)$ iff $x\in D_n$. Since $G$ is $\mu$-preserving, and since $A,Y\subseteq X$ are $\mu$-conull, we may fix an $E_G$-invariant $\mu$-conull Borel subset $B\subseteq A\cap Y$. We claim that for every $x\in B$, the function $g$ paints the neighborhood $N_G(x)$ with infinitely many colors in $[\omega]^{<\omega}$. For any $x\in B$, since $x\in A$, there is some $n_0\in\omega$ such that $x\in A_n$ for every $n>n_0$. Since $A_n$ is dominated by $D_n$, we have for every $n>n_0$ that $N_G(x)\cap D_n\ne\varnothing$, so $n\in \bigcup g[N_G(x)]$, and hence $(n_0,\infty)\subseteq\bigcup g[N_G(x)]$. Also since $N_G(x)\subseteq Y$ by $E_G$-invariance of $B$, $g[N_G(x)]\subseteq[\omega]^{<\omega}$ is a collection of finite sets. Since the union $\bigcup g[N_G(x)]\supseteq(n_0,\infty)$ is infinite, $g[N_G(x)]$ must also be infinite as claimed. Finally, we have that $g:B\to[\omega]^{<\omega}$ is a Borel $\aleph_0$-coloring for the induced subgraph $G\restrict{}B$ that paints every $G\restrict{}B$-neighborhood with infinitely many colors. So by Lemma \ref{lem:2.1}, $G\restrict{}B$ admits a $\mu$-measurable $\omega$-domatic coloring. Since $B$ is $E_G$-invariant $\mu$-conull, we may extend this coloring to the rest of $X$ to produce a $\mu$-measurable $\omega$-domatic coloring for $G$. \end{Proof}

e431bd9b0dab9d889b428aaa6bb94a6c82dc295e

\section{Introduction} In practical problems, a large amount of data comes in the form of spherical images, such as from cosmology \cite{cosmology}, astrophysics \cite{astrophysics}, geophysics \cite{planetary,geo}, neuroscience \cite{neuroscience}, and omnidirectional AR/VR field \cite{omni,omni2}, where images are naturally defined on the 2D spherical surface. Due to the storage bottleneck and observation being costly and infeasible, these spherical images (signals) usually contain very limited pixels (observed data), especially if the observation scales involved are large. Therefore, repairing missing or damaged parts is a fundamental yet challenging task in spherical image processing. Apparently, spherical images take a different inherent domain than planar images in 2D in terms of symmetries, coordinate systems, and translates, which demand special processing methods. In this paper, we are concerned with spherical image restoration, which can further serve as a preliminary for subsequent tasks, like object recognition and segmentation. Mathematically, it aims to estimate $x$ from observation $y$ for the following model \begin{equation}\label{degraded} y=T(x)+\varepsilon, \end{equation} where $T$ is a degradation operator, $\varepsilon$ is assumed to be the additive noise. Different degradation operations correspond to different image restoration (IR) tasks \cite{jia2021structure,ke2020reconstruction}. Typically, the IR task would be image denoising when $T$ is an identity operation, image deblurring when $T$ is a two-dimensional convolution operation, image super-resolution when $T$ is a composite operation of convolution and down-sampling, color image demosaicing when $T$ is a color filter array (CFA) masking operation, and image inpainting when $T$ is the orthogonal projection onto the linear space of matrices. In this paper, we proposed a general model for spherical image inpainting with a new denoiser. Regarding the degradation equation (\ref{degraded}), the IR task model can be solved through the following optimization, \begin{equation}\label{optimization} \hat x=\arg\min_x\|y-T(x)\|+\lambda \Phi(x), \end{equation} where the first term is the data fitting with $\Vert\cdot\Vert$ usually chosen to be the Frobenius norm, the second term $\Phi (\cdot)$ is an operator playing the role of regularity, and $\lambda$ is a positive trade-off parameter. With the aid of the half quadratic splitting (HQS) algorithm, by introducing an auxiliary variable, the optimization problem (\ref{optimization}) can be addressed by iteratively solving the following subproblems \begin{align} \mathbf{x}_{k}=&\arg\underset{\mathbf{x}}{\min }\|\mathbf{y}-\mathcal{T}(\mathbf{x})\|^{2}+\alpha\left\|\mathbf{x}-\mathbf{z}_{k-1}\right\|^{2}\label{sub-problem-1}, \\ \mathbf{z}_{k}=&\arg\underset{\mathbf{z}}{\min} ~ \alpha\left\|\mathbf{z}-\mathbf{x}_{k}\right\|^{2}+\lambda\Phi(\mathbf{z}). \label{sub-problem-2} \end{align} Here $\alpha$ will be set accordingly to specific problems. Equation (\ref{sub-problem-1}) is usually interpreted as the data fitting subproblem and (\ref{sub-problem-2}) as the regularization subproblem. Many research efforts have been devoted to this hot topic and achieved extensive improvements \cite{chang2021overlapping,dong2019fixing,huang2022quaternion,wong2022incorporating,wu2022efficient}. In recent years, deep learning-based models have extensively emerged and achieved state-of-the-art restoration performance \cite{cheng2022snow,li2022adjustable,li2022multiple,malgouyres2019multilinear}. The SeaNet proposed in [6] consists of three sub-nets for single image super-resolution with the help of image soft edge. Liu et al. \cite{mwcnn} proposed MWCNN for image restoration, which is a U shape network with DWT and IWT for downsampling and upsampling, respectively, and thus there is no information loss during subsampling. Both approaches achieved competitive performance in IR tasks. To improve interpretability and effectively use the trained neural networks from various tasks, Plug and Play is one of the choices to combine neural networks and prior knowledge of images with an optimization model. Zhang et al. \cite{ ZhangK2021,ircnn2017} developed the deep prior to handling the IR tasks, named plug-and-play (PnP). Specifically, they regarded the regularization term $\Phi (x)$ as a deep denoiser with the deep CNNs. The optimization problem (\ref{optimization}) was solved by the half quadratic splitting (HQS) algorithm and divided into two subproblems, in which the solution of one of the problems is replaced by the deep CNN, which is the so-called deep denoiser. Furthermore, the term (\ref{sub-problem-2}) is usually termed as denoiser prior and conducted by a single CNN denoiser \cite{Ng2020}, which is trained specifically for denoising prior and to replace solving (\ref{sub-problem-2}) to exploit the advantages of CNN. Following this line, the PnP-based model has wide applications \cite{hou2022truncated}. For example, Wu et al. \cite{Zeng2022} proposed a deep CNN-based PnP framework with MWCNN and has competitive performance in Cauchy noise removal. Zhao et al. \cite{Ng2020} suggested a PnP model for image completion with a low rankness assumption. Fang and Zeng \cite{fang2020} applied the soft edge network \cite{liedge} as a denoiser for image deblurring and denoising and gave a mathematical interpretation of the PnP-based model. Overall, the PnP-based framework has a promising performance. Many research efforts have been devoted to this hot topic and achieved extensive improvements. Particularly, in recent years, deep Plug-and-Play (DPnP) methods have been extensively developed and achieved the state-of-the-art restoration performance, see for instance \cite{Venkatakrishnan2013, Teodoro2016,Chan2017,Wei2020}. Such a hybrid learning strategy is plugging an off-the-shelf image denoiser to resolve IR problems via optimizing the following regularization framework: \begin{equation}\label{optimization} \hat x=\arg\min_x\|y-T(x)\|+\lambda \Phi(x), \end{equation} where $\Phi$ is an operator playing the role of regularity. To implement the optimization, generally two subproblems: the data fitting subproblem and the regularization subproblem is iteratively solved with the aid of certain optimization techniques, such as the alternating direction method of multipliers (ADMM) \cite{Admm} and the half quadratic splitting (HQS) \cite{Hqs}. Motivated by the advantages of the aforementioned PnP models, in this paper, we are going to apply them to image inpainting problem for spherical signals. Precisely, for a spherical signal, with its partially observed samplings, a novel PnP model integrating spherical framelet decomposition is proposed to restore the signal. The proposed model is based on low rank assumption under directional spherical Haar tight framelet, which is designed for testing image texture. In addition, we exploit a newly designed deep convolutional neural network to be the plug-and-play prior denoiser. The network inspired by \cite{li2022convolutional} and \cite{jha2020doubleu} employs two-stage encoder-decoder architecture, which is termed as Double-S2HaarNet. Under ground-truth supervision at each stage it provides progressive and improved denoising. The rest of this paper is organized as follows. In section \ref{sec:pri}, the related works about spherical signal sampling and frame decomposition are reviewed. The proposed scheme and numerical algorithm are given in section \ref{sec:alg}. Numerical results including gray image and color image inpainting are listed in section \ref{sec:ex}. Section \ref{sec:remarks} concludes this paper. \section{Spherical signal sampling and frame decomposition}\label{sec:pri} We employ a Haar tight framelet transform that developed in \cite{li2022convolutional}. Let $L_2(\mathbb{S}^2)$ be a Hilbert space with inner product $\langle\cdot,\cdot\rangle$ and norm $\|\cdot\|$ defined by \begin{align*} \langle f, g \rangle := \int_{\Omega} f(x)g(x) dx ,\\ \|f\| = \left ( \int_{\Omega} |f(x)|^2 dx \right)^{\frac{1}{2}}, \end{align*} where $f,g \in L_2(\mathbb{S}^{2})$ and $\mathbb{S}^{2} \in \mathbb{R}^3$ is the unit sphere. We call a countable collection $\{e_k\}_{k\in\Lambda}\subset L_2(\mathbb{S}^{2})$ a \emph{tight frame} with frame bound $c$ if there exists a constant $c>0$ such that \begin{align*} f = \frac{1}{c} \sum_{k\in\Lambda} \langle f, e_k \rangle e_k \quad \forall \ f\in L_2(\mathbb{S}^{2}). \end{align*} The frame decomposition is a transformation $\mathcal{F}$ given by \begin{align*} \mathcal{F}: f\in L_2(\mathbb{S}^{2}) \rightarrow \{ \langle f, e_i \rangle: e_i \in \{e_k\}_{k\in\Lambda} \}, \end{align*} and the reconstruction $\mathcal{F}^*$ \begin{align*} \mathcal{F}^*: \{ \langle f, e_i \rangle: e_i \in \{e_k\}_{k\in\Lambda} \} \rightarrow f\in L_2(\mathbb{S}^{2}). \end{align*} A Haar tight frame on the sphere can be constructed based on a hierarchical partition. \begin{definition} Let $ \mathbb{N}_0$ be a set of nonnegative integers. We call $\{\mathcal{B}_j\}_{j\in \mathbb{N}_0}$ a \emph{hierarchical partition} of $\mathbb{S}^{2}$ if the following three conditions are satisfied: \begin{enumerate} \item[{a)}] Root property: $\mathcal{B}_0 = \{\mathbb{S}^{2} \}$ and each $\mathcal{B}_j$ is a partition of $\mathbb{S}^{2}$ having finitely many measurable sets with positive measures. \item[{b)}] Nested property: for any $j\in \mathbb{N}$ and any (child) set $R_1 \in \mathcal{B}_j$, there exists a (parent) set $R_0 \in \mathcal{B}_{j-1}$ such that $R_1 \subseteq R_0$. In other word, partition $\mathcal{B}_j$ is a refinement of the partition $\mathcal{B}_{j-1}$. \item[{c)}] Density property: the maximal diameters among the sets in $\mathcal{B}_j$ tend to zero as $j$ tends to infinity. \end{enumerate} \end{definition} Denote $\Lambda_j:= [\ell_1]\times\cdots\times[\ell_j]$ to be an index set for the labeling sets in $\mathcal{B}_j$, where $[N] = \{1,\dots,N\}$ for any positive integer $N$ and \[ \mathcal{B}_j=\{R_{\vec v}\subseteq \mathbb{S}^2, \,\,\, \vec v\in \Lambda_j\}. \] By the nested property, $ R_{(\vec v,i)}\subseteq R_{\vec v}$ for $\vec v\in\Lambda_{j-1}$ and $i\in[\ell_j]$. Now, for each $\vec v\in \Lambda_j$, define a Haar-type scaling function to be \begin{equation} \label{eq:phi} \phi_{\vec v}:=\frac{\chi_{R_{\vec v}}}{\sqrt{|R_{\vec v}|}}, \end{equation} and for some integer $n_j\geq 1$, $n_j$ Haar-type framelet functions to be \begin{equation}\label{eq:Psi} \psi_{(\vec v,k)}=\sum_{i\in[\ell_j]}{a^{(\vec v)}_{k,i}}\phi_{(\vec v,i)},\quad k=1,\ldots, n_j, \end{equation} where $a^{(\vec v)}_{k,i}$ is the $(k,i)$-entry of some matrix $\bm A_{\vec v}\in\mathbb{R}^{n_j\times l_j}$. By setting proper matrices $A_{\vec v}$, one can construct a Haar-type tight frame and develop its fast decomposition and reconstruction algorithms. The following corollary determines the framelet and algorithm we shall use. \begin{corollary}\label{cor} There exists a collection $\{\phi_{\vec u}\}_{\vec u\in \Lambda_L} \cup \{\psi_{(\vec v,k)}, k\in[6]\}_{j \geq L, \vec v \in \Lambda_j}\subset L_2(\mathbb{S}^{2})$ determined by a hierarchical partition with each parent containing four children that forms a Haar tight frame with frame bound $1$, and the corresponding operators $\mathcal{F}$ and $\mathcal{F}^*$ depend on the following matrix $P$ \[ P =\frac 1 { 2} \begin{bmatrix} 1 & 1 & 1 & 1 \\ 1 & -1 & 0 & 0 \\ 1 & 0 & -1 & 0 \\ 1 & 0 & 0 & -1 \\ 0 & 1 & -1 & 0 \\ 0 & 1 & 0 & -1 \\ 0 & 0 & 1 & -1 \end{bmatrix}. \] \end{corollary} As it is well known, computers can only deal with discrete signals. To do spherical signal processing, we first need a proper way to discretize an analog signal. In this work, we take the discretization sampling method based on an area-regular partition of 2-sphere \cite{li2022convolutional}. It was constructed through a bijective mapping and its rotations: $ T:[-1, 1]\times[-1, 1] \rightarrow \mathbb{S}^{2} $ defined by $T(x,y) = \frac{(x, y,1)}{\sqrt{x^2+y^2+1}}$. See Figure \ref{fig:T} below for the illustration. \begin{figure}[htp] \centering \includegraphics[width=3in]{Figures/mappingT} \caption{Visualization of mapping $T$ which maps a square to a spherical cap.} \label{fig:T} \end{figure} Then for any given resolution $J\geq 0$, a $2$-sphere can be divided into equal-area partitions, see Figure \ref{Partitions} for illustration. This forms an algorithm for a hierarchical partition on the 2-sphere. \begin{figure}[htp] \centering \includegraphics[width=4in]{Figures/partition} \caption{Partition Process.} \label{Partitions} \end{figure} By taking the centers of the partition patches, the samplings of an analog signal can be distributed equivalently, which takes advantage over the traditional spherical coordinates discretization. Based on the above discussion, any signal $f \in L_2(\mathbb{S}^{2})$ is discretized to $\bm f$, which depends on a certain resolution $J$. The discrete signal $\bm f$ is actually the set $\{f(\bm x_i): \bm x_i \in S_i, S_i \in \mathcal{B}_j, \bigcup_{i}S_i = \mathbb{S}^{2}, S_j\cap S_k = \empty, \forall j \neq k \}$. We assume that the dataset is defined on some resolution level in the following. Applying the discretization and \corref{cor}, the spherical Haar framelet and fast framelet transform algorithm are exactly constructed. With the help of the fast decomposition and reconstruction algorithms, it, on the one hand, allows our model to capture directional texture details. On the other hand, it can reduce the spatial footprint and granularity of convolutions. \begin{figure}[htp] \centering \includegraphics[width=5in]{Figures/doubleHaar} \caption{Double-S2HaarNet.} \label{doubleHaar} \end{figure} \section{The proposed model and algorithm} \label{sec:alg} In this section, based on the aforementioned off-the-shelf spherical signal sampling and Haar-type framelets, to enhance the inpainting performance, we shall first improve the denoiser and then exploit a plug-and-play model involving fast frame decomposition. \subsection{Improved Denoiser}\label{3.1} As mentioned above, the iteration \ref{sub-problem-2} will be conducted by a denoiser. The performance of the denoiser will affect the resulting restoration. In \cite{li2022convolutional}, a CNN spherical denoising model, S2HaarNet, was developed and achieved a competitive performance. In the present paper, we further exploit a new spherical CNN (illustrated in Figure \ref{doubleHaar}), which partly follows the infrastructure of the Double-Unet \cite{jha2020doubleu} and S2HaarNet \cite{li2022convolutional}, and incorporates the skip connections and spherical frame transformations. Thus we shall call it Double-S2HaarNet. The new network consists of two feature encoder-decoder stages for which each one follows S2HaarNet. We then take ground-truth supervision at each stage for progressive and improved denoising performance. We adopt the feature concatenation by combining the feature maps from the encoder path and decoder path, which can capture multi-scale information and enrich feature representation for a better feature prediction. To bridge the two blocks, we concentrate the input and output of the first block and feed it into the second. Our model can also be readily extended to deal with color images by handling three channels independently. \subsection{Proposed PnP model} To develop a PnP model, besides the above pre-trained denoiser, a proper design for data fitting subproblem plays a crucial role as well. In this work, we attempt to propose a new data fitting operator to simultaneously exploit the strengths of the PnP model. In \cite{Ng2020}, data fitting was operated by using the tensor singular value decomposition and tensor nuclear norm, which promoted the low-rankness of the underlying tensor. Motivated by such an idea, we instead utilize the tight frame decomposition in our data fitting and suggest the novel PnP model as follows: \begin{equation}\label{w1} \begin{aligned} &\min_{\bm x} \|\mathcal{F} \bm x\|_0 + \lambda \Phi(\bm x)\\ &\text{s.t.~~} \ P_{\mathbb{S}^{2}}(\bm x) = P_{\mathbb{S}^{2}}(\bm g), \end{aligned} \end{equation} where $\mathcal{F} x$ is the coefficients of tight frame decomposition as mentioned in section \ref{3.1}, $ \Phi(\bm x)$ is an implicit regularizer by plugging our denoising Double-S2HaarNet, and $\lambda$ is a positive parameter. \subsection{Implementation Details} We apply ADMM framework to solve the optimization problem. Notice that in practice, it is more convenient to replace $\|\cdot\|_0$ by $\|\cdot \|_1$. First, we denote the indicator function as \begin{equation} \mathbf{1}_{\mathbb{S}}(\bm x)=\left\{\begin{array}{ll} 0, & \text { if } \bm x\in\{\bm x~\vert~{P}_{\mathbb{S}^{2}}(\bm x)={P}_{\mathbb{S}^{2}}({\bm g})\}. \\ \infty, & \text { otherwise }. \end{array}\right. \end{equation} Then we reformulate model (\ref{w1}) as \begin{equation}\label{w2} \begin{aligned} &\min _{\bm x}\| \bm y\|_{1}+\lambda \Phi(\bm z)+\mathbf{1}_{\mathbb{S}}(\bm x) \\ &\text { s.t. } \quad \bm y=\mathcal{F} \bm x, \bm z=\bm x. \end{aligned} \end{equation} The augmented Lagrangian function of (\ref{w2}) is \begin{equation}\label{lag} \begin{aligned} \y{\mathcal{L}(\bm x, \bm y, \bm z; \Lambda_1, \Lambda_2)=} \|\bm y\|_{1}+\lambda \Phi(\bm z)+\mathbf{1}_{\mathbb{S}}(\bm x)&+\left\langle \bm y-\mathcal{F} \bm x, \Lambda_{1}\right\rangle+\frac{\beta_{1}}{2}\|\bm y-\mathcal{F} \bm x\|^{2 +\left\langle \bm z- \bm x, \Lambda_{2}\right\rangle+\frac{\beta_{2}}{2}\|\bm z-\bm x\|^2, \end{aligned} \end{equation} \y{ where $\beta_1, \beta_2>0$ are two penalty parameters and $\Lambda_1, \Lambda_2$ are the Lagrange multipliers. The ADMM iteration for solving (\ref{lag}) goes as follows, \begin{equation} \left\{\begin{array}{l} y=\arg \min _{\bm y}\|\bm y\|_{1}+\left\langle \bm y-\mathcal{F} \bm x, \Lambda_{1}\right\rangle+\frac{\beta_{1}}{2}\left\|\bm y-\mathcal{F} \bm x\right\|^{2}, \\ \bm z=\arg \min _{\bm z} \lambda \Phi(\bm z)+\left\langle \bm z-\bm x, \Lambda_{2}\right\rangle+\frac{\beta_{2}}{2}\|\bm z-\bm x\|^{2}, \\ \bm x=\arg \min _{\bm x} \mathbf{1}_{\mathbb{S}}(\bm x)+\frac{\beta_{1}}{2}\Vert\bm y-\mathcal{F} \bm x+\frac{\Lambda_{1}}{\beta_{1}}\Vert^{2}+\left\langle \bm z-\bm x, \Lambda_{2}\right\rangle+\frac{\beta_{2}}{2}\|\bm z-\bm x\|^{2},\\ \Lambda_1=\Lambda_1+(\bm y-\mathcal{F} \bm x),\\ \Lambda_2=\Lambda_2+(\bm z-\bm x). \end{array}\right. \end{equation} Next, we elaborate on how to solve these subproblems respectively. } \begin{itemize} \item \y{The $\bm y$-subproblem is written as} \begin{equation} \begin{aligned} \bm y &=\arg \min _{\bm y}\|\bm y\|_{1}+\left\langle \bm y-\mathcal{F} \bm x, \Lambda_{1}\right\rangle+\frac{\beta_{1}}{2}\left\|\bm y-\mathcal{F} \bm x\right\|^{2} \\ &=\arg \min _{\bm y}\|\bm y\|_{1}+\frac{\beta_{1}}{2}\left\|\bm y-\mathcal{F} \bm x+\frac{\Lambda_{1}}{\beta_{1}}\right\|^{2}. \end{aligned} \end{equation} \y{Then the solution of $y$ can be obtained by} \begin{equation} \begin{aligned} \bm y &=\operatorname{shrink}\left(\mathcal{F} \bm x-\frac{\Lambda_{1}}{\beta_{1}}, \frac{1}{\beta_{1}}\right) \\ &=\max \left(\left\|\mathcal{F} \bm x-\frac{\Lambda_{1}}{\beta_{1}}\right\|_{2}-\frac{1}{\beta_{1}}, 0\right) \frac{\mathcal{F} \bm x-\frac{\Lambda_{1}}{\beta_{1}}}{\left\|\mathcal{F} \bm x-\frac{\Lambda_{1}}{\beta_{1}}\right\|}, \end{aligned} \end{equation} \y{where the $\operatorname{shrink}$ operator is a soft shrinkage operator.} \item \y{The $\bm z$-subproblem is written as} \begin{equation}\label{denoiser} \begin{aligned} \bm z &=\arg \min _{\bm z} \lambda \Phi(\bm z)+\left\langle \bm z-\bm x, \Lambda_{2}\right\rangle+\frac{\beta_{2}}{2}\|\bm z-\bm x\|^{2} \\ &=\arg \min _{\bm z} \lambda \Phi(\bm z)+\frac{\beta_{2}}{2}\left\|\bm z-\bm x+\frac{\Lambda_{2}}{\beta_{2}}\right\|^{2} \end{aligned} \end{equation} {According to Bayes rule, Eq. (\ref{denoiser}) corresponds to denoising the image $\bm x-\Lambda_2/\beta_2$ by the CNN denoiser with noise level $\lambda/\beta_2$. To address this, we rewrite Eq. (\ref{denoiser}) as \begin{equation} \bm z=\mbox{Denoiser}(\bm x-\frac{\Lambda_2}{\beta_2},\sqrt{\frac{\lambda}{\beta_2}}). \end{equation} In this paper, we apply the Double-S2HaarNet as the denoiser. } \item \y{The $\bm x$-subproblem is written as} \begin{equation}\label{x1} \begin{aligned} \bm x &=\arg \min _{\bm x} \mathbf{1}_{\mathbb{S}}(\bm x)+\frac{\beta_{1}}{2}\left\|\bm y-\mathcal{F} \bm x+\frac{\Lambda_{1}}{\beta_{1}}\right\|^{2}+\left\langle \bm z-\bm x, \Lambda_{2}\right\rangle+\frac{\beta_{2}}{2}\|\bm z-\bm x\|^{2} \\ &=\arg \min _{\bm x} \mathbf{1}_{\mathbb{S}}(\bm x)+\frac{\beta_{1}}{2}\left\|\bm y-\mathcal{F} \bm x+\frac{\Lambda_{1}}{\beta_{1}}\right\|^{2}+\frac{\beta_{2}}{2}\left\|\bm z-\bm x+\frac{\Lambda_{2}}{\beta_{2}}\right\|^{2}. \end{aligned} \end{equation} \y{By minimizing the $\bm x$-subproblem, we have $\mathbf{1}_{\mathbb{S}}(\bm x)=0$, i.e., $\bm x\in\mathbb{S}$. Then optimality condition of (\ref{x1}) is given by \begin{equation} \beta_1\mathcal{F}^*(\mathcal{F}\bm x-\bm y-\frac{\Lambda_1}{\beta_1})+\beta_2(\bm x-\bm z-\frac{\Lambda_2}{\beta_2})=0. \end{equation} Since $\mathcal{F}^*\mathcal{F}=I$,we obtain the following linear system, \begin{equation} (\beta_{1}+\beta_{2})\bm x =\beta_{1} \mathcal{F}^{*} \bm y+\beta_{2} \bm z+\mathcal{F}^{*} \Lambda_{1}+\Lambda_{2}. \end{equation} Thus, the closed-form solution of $\bm x$-subproblem is given as follows: \begin{equation} \left\{\begin{array}{l} \mathcal{P}_{\mathbb{S}^{2}}\left(\bm x\right)=\mathcal{P}_{\mathbb{S}^{2}}(\bm g), \\ \mathcal{P}_{(\mathbb{S}^{2})^{c}}\left(\bm x\right)=\mathcal{P}_{(\mathbb{S}^{2})^{c}}\left(\frac{\beta_{1} \mathcal{F}^{*} \bm y+\beta_{2} \bm z+\mathcal{F}^{*} \Lambda_{1}+\Lambda_{2}}{\beta_{1}+\beta_{2}}\right), \end{array}\right. \end{equation} where $(\mathbb{S}^{2})^{c}$ denotes the complementary set of $\mathbb{S}^{2}$.} \end{itemize} \begin{figure}[h] \centering \subfigure[Barbara]{\includegraphics[width=2.5cm]{Figures/Barbara10.pdf}} \subfigure[Boat]{\includegraphics[width=2.5cm]{Figures/Boat10.pdf}} \\ \subfigure[Fingerprint]{\includegraphics[width=2.5cm]{Figures/fingerprint10.pdf}} \subfigure[Hill]{\includegraphics[width=2.5cm]{Figures/Hill10.pdf}} \subfigure[Man]{\includegraphics[width=2.5cm]{Figures/Man10.pdf}} \caption{Five grey images for testing.} \end{figure} \begin{table}[h] \centering \footnotesize \setlength\tabcolsep{1.5pt} \linespread{1.2} \setlength{\tabcolsep}{10pt} \renewcommand{\arraystretch}{1.5} \caption{{Average inpainting results with PSNR/SSIM on F-360iSOD. In Fx, x = 6,7,8, represents the resolution level. The best results are highlighted.}}\label{t1} \begin{tabular}{c|c|c|c|c|c} \hline dataset&methods\ & 50\% &80\% & 90\% &95\% \\ \hline \multirow{5}{*}{F6}&degraded&10.08/0.2087&8.04/0.0806&7.52/0.0445&7.29/0.0268\\ \cline{2-6}&S2HaarNet&23.02/0.7711&20.27/0.5777&19.16/0.4962&18.02/0.4175\\ \cline{2-6}&S2HaarNetPnP&23.96/0.8125&20.81/0.6024&19.47/0.5225&18.33/0.4290\\ \cline{2-6}&DoubleS2HaarNet&24.22/0.8321&20.99/0.6470&19.55/0.5367&18.41/0.4568\\ \cline{2-6}&DoubleS2HaarNetPnP&\bf{24.63/0.8470}&\bf{21.49/0.6756}&\bf{19.96/0.5495}&\bf{18.63/0.4596}\\ \hline \multirow{5}{*}{F7}&degraded& 10.08/0.1857 &8.04/0.0776 &7.53/0.0454 &7.29/0.0290 \\ \cline{2-6}&S2HaarNet&23.94/0.7822&21.12/0.5920&20.03/0.5125&18.89/0.4380\\ \cline{2-6}&S2HaarNetPnP&24.92/0.8213&21.72/0.6200&20.33/0.5445&19.01/0.4553\\ \cline{2-6} &DoubleS2HaarNet&25.15/0.8349&21.83/0.6590&20.41/0.5506&19.30/0.4750\\ \cline{2-6} &DoubleS2HaarNetPnP&\bf{26.14/0.8623} &\bf{22.63/0.7009} &\bf{20.93/0.5901}&\bf{19.60/0.5007}\\ \hline \multirow{5}{*}{F8}&degraded& 10.08/0.1660&8.04/0.0755& 7.53/0.0469 &7.29/0.0313\\ \cline{2-6}&S2HaarNet&26.86/0.8402&23.79/0.6965&22.49/0.6337&21.05/0.5633\\ \cline{2-6}&S2HaarNetPnP&27.35/0.8613&24.18/0.7313&22.84/0.6667&21.49/0.5953\\ \cline{2-6} &DoubleS2HaarNet&28.55/0.8918&24.76/0.7607&23.03/0.6751&21.64/0.6006\\ \cline{2-6} &DoubleS2HaarNetPnP&\bf{28.93/0.8961}&\bf{25.02/0.7713}&\bf{23.31/0.6862}&\bf{21.87/0.6167}\\ \hline \end{tabular} \end{table} \section{Experimental results} \label{sec:ex} In this section, we present experimental results to verify the performance of the proposed model Double-S2HaarNetPnP in image inpainting. As aforementioned, we proposed a plug-and-play model for the image inpainting task. The parameter of the optimization function (\ref{lag}) are set as $\lambda=1$, $\beta_1 \in [0.1, 1]$ with step $0.1$, $\beta_2 \in [1, 5]$ with step $1$. For training Double-S2HaarNet, we use the ADAM algorithm and a mini-batch size of $16$. The learning rate decays exponentially from the beginning value $0.001$ with a multiplicative factor $0.9$ in $100$ epochs. Weight decay is chosen to be $0.001$. Since the contrast of grayscale images is relatively low, we present the visual effects of the image with color so that the image information can be displayed more clearly. \subsection{Datasets} The dataset for training CNN denoisers is produced by applying spherical sampling operation (defined in section \ref{sec:pri}) on the dataset Caltech101 \cite{fei2004learning} with 7677 for training and 1000 for validation. For the testing, we choose the dataset F-360iSOD \cite{zhang2020fixation} which contains 107 omnidirectional images. Additionally, we take five classical images as illustrated in Figure 4 for testing as well. \begin{figure}[h] \begin{minipage}{0.3\linewidth} \centering \centerline{\includegraphics[width=1.6in]{images/107_5_86_de} } \centerline{\small{(a) degraded} \centerline{\small{7.12/0.0532}} \end{minipage} \begin{minipage}{0.3\linewidth} \centering \centerline{\includegraphics[width=1.6in]{images/107_5_86_haar}} \centerline{\small{(b) HaarNet} \centerline{\small{28.32/0.8735}} \end{minipage} \begin{minipage}{0.3\linewidth} \centering \centerline{\includegraphics[width=1.6in]{images/107_5_86_haarpnp}} \centerline{\small{(c) HaarNetPnP}}% \centerline{\small{30.15/0.9349}} \end{minipage} \vspace{0.1in} \begin{minipage}{0.3\linewidth} \centering \centerline{\includegraphics[width=1.6in]{images/107_5_86_double}} \centerline{\small{(d) DoubleHaarNet} \centerline{\small{30.87/0.9346}} \end{minipage} \begin{minipage}{0.3\linewidth} \centering \centerline{\includegraphics[width=1.6in]{images/107_5_86_doublepnp}} \centerline{\small{(e) DoubleHaarNetPnP} \centerline{\small{31.87/0.95542}} \end{minipage} \begin{minipage}{0.3\linewidth} \centering \centerline{\includegraphics[width=1.6in]{images/107_5_86_or}} \centerline{\small{(f) Ground-truth}} \centerline{\small{infinity/1}} \end{minipage} \caption{The inpainting results (PSNR (dB)/SSIM) with random missing ratio 50\%. (a) the degraded image; the recovered results of (b) Haar network only; (c) plug and play with Haar network; (d) DoubleHaar network only; (e) plug and play with DoubleHaar network; (f) the original image.}\label{f4} \end{figure} \begin{table}{h} \centering \footnotesize \setlength\tabcolsep{1.5pt} \linespread{1.2} \setlength{\tabcolsep}{10pt} \renewcommand{\arraystretch}{1.5} \caption{\y{Inpainting results with PSNR/SSIM. The best results are highlighted.}}\label{t2} \begin{tabular}{c|c|c|c|c|c} \hline Images&methods\ & 50\% &80\% & 90\% &95\% \\ \hline \multirow{5}{*}{Barbara}&degraded&8.54/0.1039&6.50/0.0466&5.99/0.0281&5.75/0.0169\\ \cline{2-6}&HaarNet&26.91/0.8523&23.77/0.7092&22.57/0.6386&21.30/0.5653\\%25.70/0.8008/32.54&16.51/0.4186/93.68&15.79/0.4126/101.77&16.04/0.4313/98.93\\ \cline{2-6}&HaarNetPnP&27.56/0.9043&24.36/0.7547&23.14/0.6834&21.70/0.5731\\ \cline{2-6}&DoubleHaarNet&28.81/0.9153&24.79/0.7732&23.17/0.6788&21.87/0.6083\\ \cline{2-6}&DoubleHaarNetPnP&\bf{29.14/0.9283}&\bf{25.15/0.7895}&\bf{23.75/0.6836}&\bf{22.21/0.6799}\\ \hline \multirow{5}{*}{Boat}&degraded&8.57/0.0976&6.52/0.0453&6.02/0.0281&5.78/0.0172\\ \cline{2-6}&HaarNet&30.33/0.8569&26.94/0.7618&25.08/0.6995&22.84/0.6158\\%&28.21/0.8041/24.37&16.75/0.4340/91.14&16.17/0.4512/97.42&16.56/0.4805/93.22\\ \cline{2-6}&HaarNetPnP&32.04/0.9202&28.08/0.8079&25.77/0.7141&23.16/0.6372\\ \cline{2-6}&DoubleHaarNet&32.69/0.9208&28.20/0.8099&25.83/0.7371&23.60/0.6554\\ \cline{2-6}&DoubleHaarNetPnP&3\bf{3.12/0.9299}&\bf{28.81/0.8234}&\bf{26.63/0.7567}&\bf{23.84/0.6946}\\ \hline \multirow{5}{*}{Fingerprint}&degraded&7.39/0.1268&5.35/0.0477&4.83/0.0254&4.60/0.0136\\ \cline{2-6}&HaarNet&27.53/0.9165&23.88/0.8183&21.28/0.7101&17.90/0.4914\\%&25.91/0.8988/31.76&15.82/0.6124/101.45&14.75/0.5579/114.80&14.18/0.4651/122.49\\ \cline{2-6}&HaarNetPnP&29.10/0.9465&24.62/0.8294&21.79/0.7342&18.39/0.5394\\ \cline{2-6}&DoubleHaarNet&29.50/0.9485&25.06/0.8596&22.26/0.7630&18.92/0.5768\\ \cline{2-6}&DoubleHaarNetPnP&\bf{30.09/0.9552} &\bf{25.51/0.8684}&\bf{22.78/0.7773}&\bf{19.38/0.6392}\\ \hline\multirow{5}{*}{Hill}&degraded&10.25/0.0876&8.21/0.0444&7.69/0.0286&7.47/0.0182\\ \cline{2-6}&HaarNet&31.28/0.8541&28.06/0.7314&26.52/0.6656&24.68/0.5880\\%&28.90/0.8039/22.50&17.67/0.4156/81.98&17.02/0.4393/88.39&17.73/0.4744/81.48\\ \cline{2-6}&HaarNetPnP&32.81/0.9076&28.74/0.7739&27.00/0.6909&25.17/0.5973\\ \cline{2-6}&DoubleHaarNet&33.29/0.9119&29.15/0.7836&27.15/0.6991&25.48/0.6266\\ \cline{2-6}&DoubleHaarNetPnP&\bf{33.60/0.9187} &\bf{29.54/0.7947}&\bf{27.51/0.6994}&\bf{25.93/0.6586}\\\iffalse \hline \multirow{5}{*}{Lena}&degraded&8.64/0.0674&6.62/0.0329&6.10/0.0211&5.86/0.0136\\ \cline{2-6}&HaarNet&32.84/0.8981&29.83/0.8341&27.86/0.7930&27.29/0.7363\\ \cline{2-6}&HaarNetPnP&34.92/0.9445&31.27/0.8718&28.09/0.8147&27.94/0.7653\\ \cline{2-6}&DoubleHaarNet&35.32/0.9440&31.40/0.8762&28.98/0.8304&26.50/0.7721\\ \cline{2-6}&DoubleHaarNetPnP&\bf{35.87/0.9512}&\bf{32.04/0.8863}&\bf{29.16/0.8407}&\bf{26.74/0.7763}\\ \fi \hline\multirow{5}{*}{Man}&degraded&9.35/0.0830&7.31/0.0408&6.80/0.0266&6.56/0.0166\\ \cline{2-6}&HaarNet&31.08/0.8782&27.85/0.7833&26.20/0.7242&24.30/0.6521\\%28.89/0.8182/22.55&17.25/0.4111/86.10&16.75/0.4470/91.15&17.09/0.4924/87.66\\ \cline{2-6}&HaarNetPnP&32.70/0.9293&28.65/0.8204&26.77/0.7391&24.76/0.6733\\ \cline{2-6}&DoubleHaarNet&33.00/0.9303&28.99/0.8295&26.93/0.7588&25.03/0.6890\\ \cline{2-6}&DoubleHaarNetPnP&\bf{33.37/0.9375}&\bf{29.38/0.8369}&\bf{27.12/0.7670}&\bf{25.27/0.6903}\\ \hline \end{tabular} \end{table} \begin{figure}[htbp] \begin{minipage}{0.3\linewidth} \centering \centerline{\includegraphics[width=1.6in]{images/106_105_8_de}% } \centerline{\small{(a) degraded} \centerline{\small{7.21/0.0612}} \end{minipage} \begin{minipage}{0.3\linewidth} \centering \centerline{\includegraphics[width=1.6in]{images/106_105_8_haar}} \centerline{\small{(b) HaarNet}}% \centerline{\small{21.78/0.7525}} \end{minipage} \begin{minipage}{0.3\linewidth} \centering \centerline{\includegraphics[width=1.6in]{images/106_105_8_haarpnp}} \centerline{\small{(c) HaarNetPnP}}% \centerline{\small{22.33/0.8095}} \end{minipage} \vspace{0.1in} \begin{minipage}{0.3\linewidth} \centering \centerline{\includegraphics[width=1.6in]{images/106_105_8_double}} \centerline{\small{(d) DoubleHaarNet}} \centerline{\small{23.38/0.8227}} \end{minipage} \begin{minipage}{0.3\linewidth} \centering \centerline{\includegraphics[width=1.6in]{images/106_105_8_doublepnp}} \centerline{\small{(e) DoubleHaarNetPnP} \centerline{\small{23.68/0.8402}} \end{minipage} \begin{minipage}{0.3\linewidth} \centering \centerline{\includegraphics[width=1.6in]{images/106_105_8_or}} \centerline{\small{(f) Ground-truth}} \centerline{\small{infinity/1}} \end{minipage} \caption{{The inpainting results (PSNR (dB)/SSIM) with random missing ratio 80\%. (a) the degraded image; the recovered results of (b) Haar network only; (c) plug and play with Haar network; (d) DoubleHaar network only; (e) plug and play with DoubleHaar network; (f) the original image.}}\label{f5} \end{figure} \iffalse \begin{figure}[htbp] \begin{minipage}{0.3\linewidth} \centering \centerline{\includegraphics[width=1.6in]{images/106_1_8_de}% } \centerline{\small{(a) degraded} \centerline{\small{8.28/0.0616}} \end{minipage} \begin{minipage}{0.3\linewidth} \centering \centerline{\includegraphics[width=1.6in]{images/106_1_8_haar}} \centerline{\small{(b) HaarNet} \centerline{\small{21.49/0.6361}} \end{minipage} \begin{minipage}{0.3\linewidth} \centering \centerline{\includegraphics[width=1.6in]{images/106_1_8_haarpnp}} \centerline{\small{(c) HaarNetPnP} \centerline{\small{21.96/0.7087}} \end{minipage} \vspace{0.1in} \begin{minipage}{0.3\linewidth} \centering \centerline{\includegraphics[width=1.6in]{images/106_1_8_double}} \centerline{\small{(d) DoubleHaarNet} \centerline{\small{22.56/0.7211}} \end{minipage} \begin{minipage}{0.3\linewidth} \centering \centerline{\includegraphics[width=1.6in]{images/106_1_8_doublepnp}} \centerline{\small{(e) DoubleHaarNetPnP} \centerline{\small{22.71/0.7720}} \end{minipage} \begin{minipage}{0.3\linewidth} \centering \centerline{\includegraphics[width=1.6in]{images/106_1_8_or}} \centerline{\small{(f) F6-1}} \centerline{\small{infinity/1/0}} \end{minipage} \caption{The inpainting results (PSNR (dB)/SSIM) with random missing ratio 80\%. (a) the degraded image; the recovered results of (b) Haar network only; (c) plug and play with Haar network; (d) DoubleHaar network only; (e) plug and play with DoubleHaar network; (f) the original image. } \end{figure} \fi \begin{figure}[htbp] \begin{minipage}{0.3\linewidth} \centering \centerline{\includegraphics[width=1.6in]{images/108_9_102_de} } \centerline{\small{(a) degraded}} \centerline{\small{7.77/0.0284}} \end{minipage} \begin{minipage}{0.3\linewidth} \centering \centerline{\includegraphics[width=1.6in]{images/108_9_102_haar}} \centerline{\small{(b) HaarNet}} \centerline{\small{19.12/0.5090}} \end{minipage} \begin{minipage}{0.3\linewidth} \centering \centerline{\includegraphics[width=1.6in]{images/108_9_102_haarpnp}} \centerline{\small{(c) HaarNetPnP}} \centerline{\small{19.34/0.5303}} \end{minipage} \vspace{0.1in} \begin{minipage}{0.3\linewidth} \centering \centerline{\includegraphics[width=1.6in]{images/108_9_102_double}} \centerline{\small{(d) DoubleHaarNet}} \centerline{\small{19.40/0.5446}} \end{minipage} \begin{minipage}{0.3\linewidth} \centering \centerline{\includegraphics[width=1.6in]{images/108_9_102_doublepnp}} \centerline{\small{(e) DoubleHaarNetPnP}} \centerline{\small{19.68/0.6557}} \end{minipage} \begin{minipage}{0.3\linewidth} \centering \centerline{\includegraphics[width=1.6in]{images/108_9_102_or}} \centerline{\small{(f) Ground-truth}} \centerline{\small{infinity/1}} \end{minipage} \caption{The inpainting results (PSNR (dB)/SSIM) with random missing ratio 90\%. (a) the degraded image; the recovered results of (b) Haar network only; (c) plug and play with Haar network; (d) DoubleHaar network only; (e) plug and play with DoubleHaar network; (f) the original image.}\label{f6} \end{figure} \begin{figure}[htbp] \begin{minipage}{0.3\linewidth} \centering \centerline{\includegraphics[width=1.6in]{images/man95_de} } \centerline{\small{(a) degraded}} \centerline{\small{6.56/0.0166}} \end{minipage} \begin{minipage}{0.3\linewidth} \centering \centerline{\includegraphics[width=1.6in]{images/man95_haar}} \centerline{\small{(b) HaarNet}} \centerline{\small{24.30/0.6521}} \end{minipage} \begin{minipage}{0.3\linewidth} \centering \centerline{\includegraphics[width=1.6in]{images/man95_haarpnp}} \centerline{\small{(c) HaarNetPnP}} \centerline{\small{24.76/0.6733}} \end{minipage} \vspace{0.1in} \begin{minipage}{0.3\linewidth} \centering \centerline{\includegraphics[width=1.6in]{images/man95_double}} \centerline{\small{(d) DoubleHaarNet}} \centerline{\small{25.03/0.6890}} \end{minipage} \begin{minipage}{0.3\linewidth} \centering \centerline{\includegraphics[width=1.6in]{images/man95_doublepnp}} \centerline{\small{(e) DoubleHaarNetPnP}} \centerline{\small{25.27/0.6903}} \end{minipage} \begin{minipage}{0.3\linewidth} \centering \centerline{\includegraphics[width=1.6in]{images/man95_or}} \centerline{\small{(f) Ground-truth}} \centerline{\small{infinity/1}} \end{minipage} \caption{The inpainting results (PSNR (dB)/SSIM) with random missing ratio 95\%. (a) the degraded image; the recovered results of (b) Haar network only; (c) plug and play with Haar network; (d) DoubleHaar network only; (e) plug and play with DoubleHaar network; (f) the original image.}\label{f7} \end{figure} \begin{figure}[htbp] \hspace{0.5in} \begin{minipage}{0.16\linewidth} \centering \centerline{\includegraphics[width=1.2in]{color/or00_50_rec.png} } \centerline{\small{ Color1}} \end{minipage} \hspace{0.3in} \begin{minipage}{0.16\linewidth} \centering \centerline{\includegraphics[width=1.2in]{color/de00_50_rec.png} } \centerline{\small{(a) degraded} \centerline{\small{8.39/0.0473}} \end{minipage} \hspace{0.3in} \begin{minipage}{0.12\linewidth} \centering \centerline{\includegraphics[width=1.2in]{color/haar00_50_rec.png}} \centerline{\small{(b) HaarNet} \centerline{\small{31.53/0.9458}} \end{minipage} \hspace{0.3in} \begin{minipage}{0.16\linewidth} \centering \centerline{\includegraphics[width=1.2in]{color/hp00_50_rec.png}} \centerline{\small{(c) HaarNetPnP}}% \centerline{\small{32.96/0.9551}} \end{minipage} \vspace{0.1in} \hspace{1.86in} \begin{minipage}{0.16\linewidth} \centering \centerline{\includegraphics[width=1.2in]{color/double00_50_rec.png}} \centerline{\small{(d) DoubleHaarNet} \centerline{\small{36.98/0.9850}} \end{minipage} \hspace{0.2in} \begin{minipage}{0.16\linewidth} \centering \centerline{\includegraphics[width=1.2in]{color/our00_50_rec.png}} \centerline{\small{(e) DoubleHaarNetPnP} \centerline{\small{37.23/0.9856}} \end{minipage} \hspace{0.2in} \begin{minipage}{0.16\linewidth} \centering \centerline{\includegraphics[width=1.2in]{color/orr00_50_rec.png}} \centerline{\small{(f) Ground-truth}} \centerline{\small{infinity/1}} \end{minipage} \caption{The inpainting results (PSNR (dB)/SSIM) with random missing ratio 50\%. (a) the degraded image; the recovered results of (b) Haar network only; (c) plug and play with Haar network; (d) DoubleHaar network only; (e) plug and play with DoubleHaar network; (f) the original image.}\label{cf5} \end{figure} \begin{figure}[htbp] \begin{minipage}{0.3\linewidth} \centering \centerline{\includegraphics[width=1.6in]{color/de00_50.png} } \centerline{\small{(a) degraded} \centerline{\small{8.39/0.0473}} \end{minipage} \begin{minipage}{0.3\linewidth} \centering \centerline{\includegraphics[width=1.6in]{color/haar00_50.png}} \centerline{\small{(b) HaarNet} \centerline{\small{31.53/0.9458}} \end{minipage} \begin{minipage}{0.3\linewidth} \centering \centerline{\includegraphics[width=1.6in]{color/hp00_50.png}} \centerline{\small{(c) HaarNetPnP}}% \centerline{\small{32.96/0.9551}} \end{minipage} \vspace{0.1in} \begin{minipage}{0.3\linewidth} \centering \centerline{\includegraphics[width=1.6in]{color/double00_50.png}} \centerline{\small{(d) DoubleHaarNet} \centerline{\small{36.98/0.9850}} \end{minipage} \begin{minipage}{0.3\linewidth} \centering \centerline{\includegraphics[width=1.6in]{color/or00_50.png}} \centerline{\small{(e) DoubleHaarNetPnP} \centerline{\small{37.23/0.9856}} \end{minipage} \begin{minipage}{0.3\linewidth} \centering \centerline{\includegraphics[width=1.6in]{color/or00_50.png}} \centerline{\small{(f) Color1}} \centerline{\small{infinity/1}} \end{minipage} \caption{The inpainting results (PSNR (dB)/SSIM) with random missing ratio 50\%. (a) the degraded image; the recovered results of (b) Haar network only; (c) plug and play with Haar network; (d) DoubleHaar network only; (e) plug and play with DoubleHaar network; (f) the original image.}\label{f4} \end{figure} \begin{figure}[htbp] \begin{minipage}{0.3\linewidth} \centering \centerline{\includegraphics[width=1.6in]{color/de04_80.png} } \centerline{\small{(a) degraded} \centerline{\small{8.70/0.0642}} \end{minipage} \begin{minipage}{0.3\linewidth} \centering \centerline{\includegraphics[width=1.6in]{color/haar04_80.png}} \centerline{\small{(b) HaarNet} \centerline{\small{24.01/0.7669}} \end{minipage} \begin{minipage}{0.3\linewidth} \centering \centerline{\includegraphics[width=1.6in]{color/hp04_80.png}} \centerline{\small{(c) HaarNetPnP}}% \centerline{\small{25.96/0.7815}} \end{minipage} \vspace{0.1in} \begin{minipage}{0.3\linewidth} \centering \centerline{\includegraphics[width=1.6in]{color/double04_80.png}} \centerline{\small{(d) DoubleHaarNet} \centerline{\small{26.13/0.8161}} \end{minipage} \begin{minipage}{0.3\linewidth} \centering \centerline{\includegraphics[width=1.6in]{color/our04_80.png}} \centerline{\small{(e) DoubleHaarNetPnP} \centerline{\small{26.56/0.8169}} \end{minipage} \begin{minipage}{0.3\linewidth} \centering \centerline{\includegraphics[width=1.6in]{color/or04_80.png}} \centerline{\small{(f) Color4}} \centerline{\small{infinity/1}} \end{minipage} \caption{The inpainting results (PSNR (dB)/SSIM) with random missing ratio 80\%. (a) the degraded image; the recovered results of (b) Haar network only; (c) plug and play with Haar network; (d) DoubleHaar network only; (e) plug and play with DoubleHaar network; (f) the original image.}\label{f4} \end{figure} \begin{figure}[htbp] \hspace{0.5in} \begin{minipage}{0.16\linewidth} \centering \centerline{\includegraphics[width=1.2in]{color/orr_rec.png} } \centerline{\small{ Color4}} \end{minipage} \hspace{0.3in} \begin{minipage}{0.16\linewidth} \centering \centerline{\includegraphics[width=1.2in]{color/de04_08_rec.png} } \centerline{\small{(a) degraded} \centerline{\small{8.70/0.0642}} \end{minipage} \hspace{0.3in} \begin{minipage}{0.12\linewidth} \centering \centerline{\includegraphics[width=1.2in]{color/haar04_80_rec.png}} \centerline{\small{(b) HaarNet} \centerline{\small{24.01/0.7669}} \end{minipage} \hspace{0.3in} \begin{minipage}{0.16\linewidth} \centering \centerline{\includegraphics[width=1.2in]{color/hp04_80_rec.png}} \centerline{\small{(c) HaarNetPnP}}% \centerline{\small{25.96/0.7815}} \end{minipage} \vspace{0.1in} \hspace{1.86in} \begin{minipage}{0.16\linewidth} \centering \centerline{\includegraphics[width=1.2in]{color/double04_80_rec.png}} \centerline{\small{(d) DoubleHaarNet} \centerline{\small{26.13/0.8161}} \end{minipage} \hspace{0.2in} \begin{minipage}{0.16\linewidth} \centering \centerline{\includegraphics[width=1.2in]{color/our04_80_rec.png}} \centerline{\small{(e) DoubleHaarNetPnP} \centerline{\small{26.56/0.8169}} \end{minipage} \hspace{0.2in} \begin{minipage}{0.16\linewidth} \centering \centerline{\includegraphics[width=1.2in]{color/or04_80_rec.png}} \centerline{\small{(f) Ground-truth}} \centerline{\small{infinity/1}} \end{minipage} \caption{The color image inpainting results (PSNR (dB)/SSIM) with random missing ratio 80\%. Zoomed part of (a) the degraded image; the recovered results of (b) Haar network only; (c) plug and play with Haar network; (d) DoubleHaar network only; (e) plug and play with DoubleHaar network; (f) the original image.}\label{cf8} \end{figure} \begin{figure}[htbp] \hspace{0.5in} \begin{minipage}{0.16\linewidth} \centering \centerline{\includegraphics[width=1.2in]{color/orr_87-41_rec.png} } \centerline{\small{ Color12}} \end{minipage} \hspace{0.3in} \begin{minipage}{0.16\linewidth} \centering \centerline{\includegraphics[width=1.2in]{color/de12_90_rec.png} } \centerline{\small{(a) degraded}} \centerline{\small{6.88/0.0318}} \end{minipage} \hspace{0.3in} \begin{minipage}{0.12\linewidth} \centering \centerline{\includegraphics[width=1.2in]{color/haar12_90_rec.png}} \centerline{\small{(b) HaarNet}} \centerline{\small{21.55/0.7226}} \end{minipage} \hspace{0.3in} \begin{minipage}{0.16\linewidth} \centering \centerline{\includegraphics[width=1.2in]{color/hp12_90_rec.png}} \centerline{\small{(c) HaarNetPnP}}% \centerline{\small{21.89/0.7251}} \end{minipage} \hspace{1.86in} \begin{minipage}{0.16\linewidth} \centering \centerline{\includegraphics[width=1.2in]{color/double12_90_rec.png}} \centerline{\small{(d) DoubleHaarNet}} \centerline{\small{21.9763/0.7485}} \end{minipage} \hspace{0.2in} \begin{minipage}{0.16\linewidth} \centering \centerline{\includegraphics[width=1.2in]{color/our12_90_rec.png}} \centerline{\small{(e) DoubleHaarNetPnP}} \centerline{\small{22.27/0.7666}} \end{minipage} \hspace{0.2in} \begin{minipage}{0.16\linewidth} \centering \centerline{\includegraphics[width=1.2in]{color/orr12_90_rec.png}} \centerline{\small{(f) Ground-truth}} \centerline{\small{infinity/1}} \end{minipage} \caption{The color image inpainting results (PSNR (dB)/SSIM) with random missing ratio 90\%. Zoomed part of (a) the degraded image; the recovered results of (b) Haar network only; (c) plug and play with Haar network; (d) DoubleHaar network only; (e) plug and play with DoubleHaar network; (f) the original image.}\label{cf90} \end{figure} \begin{figure}[htbp] \hspace{0.5in} \begin{minipage}{0.16\linewidth} \centering \centerline{\includegraphics[width=1.2in]{color/orr_94_-10_rec.png} } \centerline{\small{ Color6}} \end{minipage} \hspace{0.3in} \begin{minipage}{0.16\linewidth} \centering \centerline{\includegraphics[width=1.2in]{color/de06_95_rec.png} } \centerline{\small{(a) degraded} \centerline{\small{7.27/0.0108}} \end{minipage} \hspace{0.3in} \begin{minipage}{0.12\linewidth} \centering \centerline{\includegraphics[width=1.2in]{color/haar06_95_rec.png}} \centerline{\small{(b) HaarNet} \centerline{\small{27.72/0.9042}} \end{minipage} \hspace{0.3in} \begin{minipage}{0.16\linewidth} \centering \centerline{\includegraphics[width=1.2in]{color/hp06_95_rec.png}} \centerline{\small{(c) HaarNetPnP}}% \centerline{\small{28.08/0.9056}} \end{minipage} \vspace{0.1in} \hspace{1.86in} \begin{minipage}{0.16\linewidth} \centering \centerline{\includegraphics[width=1.2in]{color/double06_95_rec.png}} \centerline{\small{(d) DoubleHaarNet} \centerline{\small{28.84/0.9193}} \end{minipage} \hspace{0.2in} \begin{minipage}{0.16\linewidth} \centering \centerline{\includegraphics[width=1.2in]{color/our06_95_rec.png}} \centerline{\small{(e) DoubleHaarNetPnP} \centerline{\small{29.09/0.9216}} \end{minipage} \hspace{0.2in} \begin{minipage}{0.16\linewidth} \centering \centerline{\includegraphics[width=1.2in]{color/orr06_95_rec.png}} \centerline{\small{(f) Ground-truth}} \centerline{\small{infinity/1}} \end{minipage} \caption{The color image inpainting results (PSNR (dB)/SSIM) with random missing ratio 95\%. Zoomed part of (a) the degraded image; the recovered results of (b) Haar network only; (c) plug and play with Haar network; (d) DoubleHaar network only; (e) plug and play with DoubleHaar network; (f) the original image.}\label{cf95} \end{figure} \subsection{Evaluation metrics} \y{To demonstrate the effectiveness of the proposed scheme, we test the images with random missing values, i.e., the missing ratio with $50\%, 80\%, 90\%, 95\%$. With the built-in function in numpy, we use the command `numpy.random.rand' to generate random values of the same size as the input image. Then let the value which is great than the missing ratio equal to 1 and less than the missing ratio equal to 0. Then we have the missing operator, with the function `numpy.multiply' mapping the missing operator and the groundtruth image to the observed image. The peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) are used to evaluate the performance of the inpainting results. To be specific, with the reference $x$ and the obtained result $x^*$, the PSNR is defined as \begin{equation} \mathrm{PSNR}(x,x^*)=20 \log_{10} \frac{255}{\frac{1}{m n}\left\|x^{\star}-x\right\|}, \end{equation} where $\Vert\cdot\Vert$ denotes a Frobenius norm. The SSIM is defined as \begin{equation} \operatorname{SSIM}\left(x, x^{\star}\right)=\frac{\left(2 \mu_{x} \mu_{x^{\star}}+C_{1}\right)\left(2 \sigma_{x x^{\star}}+C_{2}\right)}{\left(\mu_{x}^{2}+\mu_{x^{\star}}^{2}+C_{1}\right)\left(\sigma_{x}^{2}+\sigma_{x^{\star}}^{2}+C_{2}\right)} \end{equation} where $\mu_{x}$, $\mu_{x^*}$ and $\sigma_{x}$, $\sigma_{x^*}$, $\sigma_{xx^*}$ are the mean and the standard deviation of $x$ and $x^*$, respectively. The positive constants $C_1$ and $C_2$ are used to avoid a null denominator, which are defaulted by the build-in ssim function. } \subsection{Results} We make a detailed comparison of our method. More specifically, the methods with single net (S2HaarNet and DoubleS2haarNet) and the methods with plug-and-play (S2HaarNetPnP and DoubleS2haarNetPnP) are compared. First of all, the three datasets are tested. The results of datasets `F6', `F7', and `F8' are listed in Table \ref{t1}, from which we know that the proposed methods are robust in different degradation. For example, when the missing ratio is from $50\%$ up to $95\%$, our methods always have competitive restoration results. Besides, the results based on the plug-and-play approach are better than the one with only CNN, which also illustrate the effectiveness of the proposed plug-and-play scheme. Moreover, the testing of two different datasets and different combinations of the proposed model also gives a strong validity to our scheme. On the other hand, five grey images are also tested in this paper. We list the numerical results in Table \ref{t2}, from which the plug-and-play-based models also have better restoration results. The visual results are present in Figures~\ref{f4}--\ref{f7} with missing ratio $50\%, 80\%, 90\%, 95\%$, respectively. Figure~\ref{f4} (a) is the observed image with a low missing ratio ($50\%$). As we can see from the results (b)--(e), most of the objects in the image are recovered. However, with detailed observation, we know that the plug-and-play-based methods have more competitive performance. With the missing ratio up to $80\%$, there are some outlines of the original image that can be seen in Figure \ref{f5} (a). It turns out that the results of the HaarNet and DoubleHaarNet have quite satisfactory results. As the plug-and-play is applied in (b) and (d) respectively, the results of inpainting are greatly improved. For the low sample ratio, from Figure \ref{f6} (a) and Figure \ref{f7} (a), the details of the image are almost disappeared. With this low observation, our models also can recover the images with good quality. The above visual results demonstrate again the robustness and effectiveness of the proposed inpainting models. The experiments on color images are also conducted in Figures~\ref{cf5}--\ref{cf95}, which illustrate the good generalization of our model to color images. Note that our results are slightly over-smoothed for lower sample rates, such as 95$\%$. Figure \ref{cf95} shows that, in contrast to other results, our solutions can better restore the structures of images, which is consistent with the results of the majority of plug-and-play-based works. \section{Conclusion and further remarks} \label{sec:remarks} In this work, we presented the doubleHaarNetPnP model for image inpainting. We remark that (a) low-rank framelet coefficient regularizer is introduced to learn, (b) a new denoiser DoubleHaarNet for spherical image inpainting is proposed and in the experiments, it is powerful for inpainting task with the deep Plug-and-play framework, and (c) Directional spherical Haar framelet is employed to capture directional texture information to enhance the learning ability of the model and the neural network. Experiments evaluated on various images illustrate the effectiveness of the proposed method for spherical image inpainting problems. Since our framework is general and flexible, the corresponding model for other spherical imaging tasks for example spherical image segmentation. \section*{Acknowledgement} This work was supported in part by the National Key R\&D Program of China under Grant 2021YFE0203700, Grant NSFC/RGC N\_CUHK 415/19, Grant NSFC Nos. 11871210, 11971215, and 61971292, Grant ITF MHP/038/20, Grant RGC 14300219, 14302920, 14301121, and CUHK Direct Grant for Research. And in part by Hong Kong Research Grant Council GRF 12300218, 12300519, 17201020, 17300021, C1013-21GF, C7004-21GF and Joint NSFC-RGC N-HKU76921. And in part by HKRGC Grants Nos. CUHK14301718, CityU11301120, C1013-21GF, CityU Grant 9380101. \bibliographystyle{plain}

42d34d13fa729554022ab2ae51de62cb946e1328

\section{Introduction} \begin{paragraph}\noindent Let $\mathfrak{a}$ be an ideal of a commutative unital ring $A$. There are two additive endo functors on the category of $A$-modules called the $\mathfrak{a}$-torsion functor which is given by $\Gamma_{ \mathfrak{a}}(-):=\underset{k}\varinjlim ~\text{Hom}_{A}(A/\mathfrak{ a}^{k},-)$ and the $\mathfrak{a}$-adic completion functor given by $\Lambda_{ \mathfrak{a}}(-):= \underset{k}\varprojlim~ (A/\mathfrak{ a}^{k}\otimes_{A}-) $. These functors have been utilised in \cite {Alonso-Lipman, Barthel1-Local duality, Barthel 2, brodmann2013local, dwyer2002complete, greenlees1992derived, G.Kempt-Cousin Complex, kyomuhangi2020locally, Annet-David : Generalized reduced, porta2014homology, Leonid-Remarks, Peter-Schenzez, P. Schenzel, David-Application-I, vyas2018weak, Yekutieli-flatness} for different purposes and in different contexts. The functors have been used widely to study local (co)homology in algebraic geometry \cite{Alonso-Lipman, Hartshorne-Local cohomology}, topology \cite{Barthel1-Local duality, L. Pol-Homotopy theory}, representation theory \cite{G.Kempt-Cousin Complex}, commutative algebra \cite{brodmann2013local,porta2014homology, Peter-Schenzez, David-Application-I, Amnon-WPR-der comp-prism}, and non-commutative algebra \cite{vyas2018weak} among others. $\Gamma_{\mathfrak{a}}$ and $\Lambda_{\mathfrak{a}}$ are not pairwise adjoint in general. However, when the ideal $\mathfrak{a}$ is weakly proregular, the right derived functor $\text{R}\Gamma_{\mathfrak{a}}$ of the functor $\Gamma_{ \mathfrak{a}}$ and the left derived functor $\text{L}\Lambda_{\mathfrak{a}}$ of the functor $\Lambda_{ \mathfrak{a}}$ form an adjoint pair in the derived category of $A$-modules \cite[Theorem 9]{Erratum to: On the homology of comp and tor}. This is called the \textit{Greenless-May duality}. A notion related to the Greenless-May (GM for short) duality is the Matlis-Greenless-May (MGM for short) equivalence. Let $\mathfrak{a}\subseteq A$ be a weakly proregular ideal, and let $M$ be a complex of $A$-modules. The derived completion of the complex $M$ is the complex $\text{L}\Lambda_{\mathfrak{a}}(M)$ and the derived torsion complex of $M$ is the complex $ \text{R}\Gamma_{\mathfrak{a}}(M)$. A complex $M$ is derived \textit{$\mathfrak{a}$-torsion} (resp. derived \textit{$\mathfrak{a}$-adically complete}) if there is an isomorphism $\text{R}\Gamma_{\mathfrak{a}}(M)\cong M $ (resp. $ M\cong \text{L}\Lambda_{\mathfrak{a}}(M)$) in the derived category of $A$-modules \cite{porta2014homology, Peter-Schenzez, Amnon-WPR-der comp-prism}. In \cite[Theorem 7.11] {porta2014homology} and \cite[Theorem 9.6.7]{Peter-Schenzez} it is shown that the pair $(\text{R}\Gamma_{\mathfrak{a}},\text{L}\Lambda_{\mathfrak{a}})$ forms an equivalence between the triangulated full subcategories consisting of the derived $\mathfrak{a}$-adically complete complexes and the derived $\mathfrak{a}$-torsion complexes. This is called the \textit{MGM equivalence} in the derived category. A version of this for complexes over a non-commutative ring was given in \cite{vyas2018weak}. \end{paragraph} \begin{paragraph}\noindent An $A$-module $M$ is called \textit{$\mathfrak{ a}$-reduced} (resp. \textit{$\mathfrak{ a}$-coreduced}) if $\mathfrak{ a}\Gamma_{ \mathfrak{a}}(M)=0$ (resp. $\mathfrak{ a}\Lambda_{ \mathfrak{a}}(M)=0$). Recently, it was shown in \cite{David-Application-I} that the pair $(\Lambda_{ \mathfrak{a}}, \Gamma_{ \mathfrak{a}})$ is both an adjoint pair between the full subcategories of $\mathfrak{ a}$-reduced modules and $\mathfrak{ a}$-coreduced modules and an equivalence between full subcategories of $A$-modules which are both $\mathfrak{ a}$-reduced and $\mathfrak{ a}$-torsion and those which are both $\mathfrak{ a}$-coreduced and $\mathfrak{ a}$-adically complete. Precisely, if $\text{M}(A)_{\mathfrak{ a}\text{-red}} ~(\text{resp.} ~\text{M}(A)_{\mathfrak{ a}\text{-cor}}) $ denotes the full subcategory of $A$-modules $\text{M}(A)$ consisting of $\mathfrak{ a}$-reduced (resp. $\mathfrak{ a}$-coreduced) $A$-modules and $\text{M}(A)_{\mathfrak{ a}\text{-tor}} ~(\text{resp.} ~\text{M}(A)_{\mathfrak{ a}\text{-com}})$ denotes the full subcategory of $\text{M}(A)$ consisting of $\mathfrak{a}$-torsion (resp. $\mathfrak{ a}$-adically complete) $A$-modules, then we have \begin{enumerate} \item[$(1)$] $\Gamma_{\mathfrak{ a}}: \text{M}(A)_{\mathfrak{ a}\text{-red}}\rightarrow \text{M}(A)_{\mathfrak{ a}\text{-cor}} $ is right adjoint to $\Lambda_{ \mathfrak{a}}: \text{M}(A)_{\mathfrak{ a}\text{-cor}}\rightarrow \text{M}(A)_{\mathfrak{ a}\text{-red}} $ \cite[Theorem 3.2]{David-Application-I} and \item[$(2)$] $\Gamma_{\mathfrak{ a}}: \text{M}(A)_{\mathfrak{ a}\text{-red}}\cap \text{M}(A)_{\mathfrak{ a}\text{-tor}}\rightarrow \text{M}(A)_{\mathfrak{ a}\text{-cor}}\cap \text{M}(A)_{\mathfrak{ a}\text{-com}}$ is an equivalence with quasi-inverse $\Lambda_{ \mathfrak{a}}$ \cite[Theorem 4.2]{David-Application-I}. \end{enumerate} In Section 2, we introduce $\mathfrak{ a}$-reduced and $\mathfrak{ a}$-coreduced complexes and show that the work done in \cite{David-Application-I} holds in the more general setting of the category of chain complexes. In particular, we prove that the functor $\Gamma_{ \mathfrak{a}}$ on $\text{C}(A)$ is representable on the full subcategory of $\mathfrak{ a}$-reduced complexes, i.e., for any $\mathfrak{ a}$-reduced complex $M$, $\Gamma_{ \mathfrak{a}}(M)\cong \text{Hom}_{A}(A/\mathfrak{ a}, M)$. Dually, the functor $\Lambda_{ \mathfrak{a}}$ is representable on the full subcategory of $\mathfrak{ a}$-coreduced complexes, i.e., for any $\mathfrak{ a}$-coreduced complex $N$, $\Lambda_{ \mathfrak{a}}(N)\cong A/\mathfrak{ a} \otimes_{A} N$. \end{paragraph} \begin{paragraph}\noindent Section 3 contains the two main results of our paper. Our first main result is Theorem \ref{Thm 1.1} which says that in the setting of the full subcategories of $\mathfrak{a}$-reduced complexes and $\mathfrak{a}$-coreduced complexes the functors $\Gamma_{\mathfrak{ a}}$ and $\Lambda_{\mathfrak{a}}$ form an adjoint pair. For an ideal $\mathfrak{ a}$ of a ring $A$, denote by $\text{C}(A)_{\mathfrak{a}\text{-red}}$ and $\text{C}(A)_{\mathfrak{a}\text{-cor}}$ the full subcategory of $\text{C}(A)$ consisting of $\mathfrak{a}$-reduced and $\mathfrak{a}$-coreduced complexes respectively; where $\text{C}(A)$ is the category of chain complex of $A$-modules. \end{paragraph} \begin{thm}[GM Duality in $\text{C}(A)$]\label{Thm 1.1} Let $A$ be a ring and $\mathfrak{a}$ an ideal in it. \begin{itemize} \item[$(1)$] The functor {\normalfont\begin{equation*} \Gamma_{\mathfrak{a}}: \text{C}(A)_{\mathfrak{a}\text{-red}}\rightarrow \text{C}(A)_{\mathfrak{a}\text{-cor}}; ~ M\mapsto \Gamma_{\mathfrak{a}}(M) \end{equation*}} is idempotent and {\normalfont $\Gamma_{\mathfrak{a}}(M)\cong \text{Hom}_{A}(A/\mathfrak{a},M)$} for any $\mathfrak{a}$-reduced $A$-complex $M$. \item[$(2)$] The functor {\normalfont \begin{equation*} \Lambda_{\mathfrak{a}}: \text{C}(A)_{\mathfrak{a}\text{-cor}}\rightarrow \text{C}(A)_{\mathfrak{a}\text{-red}};~ M\mapsto \Lambda_{\mathfrak{a}}(M) \end{equation*}} is idempotent and {\normalfont $\Lambda_{\mathfrak{a}}(M)\cong A/\mathfrak{a}\otimes_{A} M$} for any $\mathfrak{a}$-coreduced complex $M$. \item[$(3)$] For any $M\in {\normalfont \text{C}(A)_{\mathfrak{a}\text{-red}}} $ and any {\normalfont $ N\in \text{C}(A)_{\mathfrak{a}\text{-cor}}$}, {\normalfont \begin{equation*} \text{Hom}_{A}(\Lambda_{\mathfrak{a}}(M) ,N)\cong \text{Hom}_{A}( M,\Gamma_{\mathfrak{a}}(N)). \end{equation*}} \end{itemize} \end{thm} This is repeated as Theorem \ref{sec3: GM Duality-Thm} in the body of the paper. \begin{paragraph}\noindent Our second main result is Theorem \ref{Thm 1.2} which tells us when the equivalence between {\normalfont$ \text{C}(A)_{\mathfrak{a}\text{-red}}$} and {\normalfont $\text{C}(A)_{\mathfrak{a}\text{-cor}}$} is established. We call this equivalence the \textit{MGM equivalence} in $\text{C}(A)$. Let {\normalfont $\text{C}(A)_{\mathfrak{a}\text{-com}}$ (resp. $\text{C}(A)_{\mathfrak{a}\text{-tor}}$)} denote the full subcategory of $\text{C}(A)$ consisting of $\mathfrak{ a}$-adically complete (resp. $\mathfrak{a}$-torsion) complexes of $A$-modules. \begin{thm}\label{Thm 1.2} Let $\mathfrak{a}$ be an ideal of a ring $A$. Let {\normalfont $\mathfrak{A}:=\text{C}(A)_{\mathfrak{a}\text{-tor}}\cap \text{C}(A)_{\mathfrak{a}\text{-red}}$ }and {\normalfont$\mathfrak{B}:= \text{C}(A)_{\mathfrak{a}\text{-com}}\cap \text{C}(A)_{\mathfrak{a}\text{-cor}}$}.\\ \begin{itemize} \item [$(1)$] For any $ M {\normalfont \in\text{C}(A)_{\mathfrak{a}\text{-red}}}$, $\Gamma_{\mathfrak{a}}(M)\in \mathfrak{B}$. \item [$(2)$] For any $ M \in {\normalfont\text{C}( A)_{\mathfrak{a}\text{-cor}}}$, $ \Lambda_{\mathfrak{a}}(M)\in \mathfrak{A}$. \item [$(3)$] The functor $\Gamma_{\mathfrak{a}}: {\normalfont\text{C}( A)_{\mathfrak{a}\text{-red}}\rightarrow \text{C}( A)_{\mathfrak{a}\text{-cor}}}$ restricted to $\mathfrak{A}$ is an equivalence between $\mathfrak{A}$ and $\mathfrak{B}$ with quasi-inverse $\Lambda_{\mathfrak{a}}$. \end{itemize} \end{thm} This is Theorem \ref{sec3: MGM equivalence} in the body of the paper. \end{paragraph} \begin{paragraph}\noindent An $A$-module $M$ is \textit{reduced} if the implication $a^{2}m=0\Rightarrow am=0$ holds for every $a\in A$ and for all $m\in M$ \cite{ kyomuhangi2020locally, Annet-David : Generalized reduced, ssevviiri2018nilpotent}. Equivalently, $M$ is $a$-reduced if $a\Gamma_{a}(M)=0$ for all $a\in A$ where $\Gamma_{a}(M)$ is the submodule $\{m\in M~ |~a^{j}m=0 ~\text{for~ some}~ j\in \mathbb Z^{+}\}$ of $M$. Let $a\in A$ and $t\in\mathbb Z^{+}$. We say that $M$ is \textit {$a^{t}$-reduced} or \textit{generalized reduced} \cite{Annet-David : Generalized reduced} if $a^{t}\Gamma_{a}(M)=0$. This is equivalent to saying that for any $k\ge t\in\mathbb N$, $a\in A$ and $m\in M$, $a^{k}m=0$ implies that $a^{t}m=0$. This general form of reduced modules exists in the literature when called different names. In \cite[Definition 1.2]{greenlees1992derived} and \cite[Definition 5.5]{Amnon-WPR-der comp-prism} they are called \textit{modules with bounded $a$-torsion}. In \cite[Proposition 3.1.10] {Peter-Schenzez} $a^{t}$-reduced modules are called \textit{modules whose ascending sequence of submodules $(0:_{M}a^{t})$ becomes stationary.} \end{paragraph} We remark that $a^{t}$-reduced modules is related to\textit{ prisms}; a concept which belongs to the new groundbreaking theory of perfectoid rings, see \cite{B.Bhatt-P.Scholze, Amnon-WPR-der comp-prism}. \section{$\mathfrak{ a}$-reduced and $\mathfrak{ a}$-coreduced complexes} \begin{paragraph}\noindent Throughout the paper $A$ is a commutative ring with unity, and $\mathfrak{ a}$ is an ideal in $A$. In this section, we introduce $\mathfrak{a}$-reduced and $\mathfrak{a}$-coreduced complexes over $A$ and study their basic properties. We denote by $\text{M}(A)$ the category of $A$-modules, and by $\text{C}(A)$ its category of chain complexes. \end{paragraph} \begin{paragraph}\noindent We begin by recalling some definitions and facts to be used in the sequel. Let $A$ be a ring and let $\mathfrak{ a}$ be an ideal of $A$. The complexes $ M:=(M^{i},d_{M}^{i})_{i\in\mathbb{Z}}$ and $ N:=(N^{j},d_{N}^{j})_{j\in\mathbb{Z}} \in C(A)$ are isomorphic if all the homomorphisms $f^{i}: M^{i}\rightarrow N^{i}$ are isomorphisms. When $M$ and $N$ are isomorphic we write $M\cong N$. Following \cite{Peter-Schenzez}, the homomorphism complex $\text{Hom}_{A}(M,N)$ and the tensor product complex $M\otimes_{A}N$ for $A$-complexes $ M=(M^{i},d_{M}^{i})_{i\in\mathbb{Z}}$ and $ N=(N^{j},d_{N}^{j})_{j\in\mathbb{Z}} \in \text{C}(A)$ are respectively given by:\\ $ \text{Hom}_{A}(M,N):=(\text{Hom}_{A}(M, N )^ {n}, d^{n})_{n\in\mathbb{Z}}$;~ \text{where} $\text{Hom}_{A}(M, N )^ {n}=\underset{j-i=n}{\prod} {\text{Hom}}_{A} (M^{i} , N^ {j})$ and $d^{n}= \underset{j-i=n}{\prod}{\text{Hom}}_{A}(d_{M}^ {i-1},N^{j}) + (-1)^{i+1}Hom_{A}(M^{i},d_{N}^{j});$ $M\otimes_{A} N:= ((M\otimes_{A} N)^{n},\sigma^{n})_{n\in\mathbb{Z}}$ ~\text{where} $(M\otimes_{A}N)^{n}=\underset{i+j=n}{\oplus}M^{i}\otimes_{A} N^{j}$ and $\sigma^{n}:= d^{i}_{M}\otimes_{A} \text{id}_{N} +(-1)^{i}~ \text{id}_{M}\otimes_{A} d^{i}_{N}$. \end{paragraph} \begin{lem}\label{sec2: Hom-tensor Lem} Let $\mathfrak{a}$ be an ideal of a ring $A$, and $M$ an $A$-complex. For any $n\in\mathbb{Z}$, {\normalfont \begin{itemize} \item [$(1)$] $\text{ Hom}_{A}(A/\mathfrak{ a}, M)^{n}=\text{ Hom}_{A}(A/\mathfrak{ a}, M^{n})$, \item[$(2)$] $(A/\mathfrak{ a}\otimes_{A} M)^{n}= A/\mathfrak{ a}\otimes_{A} M^{n}$. \end{itemize}} \end{lem} \begin{prf} Let $M:=(M^{n})_{n\in\mathbb{Z}}$. Since for any $k\in\mathbb Z^{+}$ the $A$-module $(A/\mathfrak{ a})^{k}$ is an $A$-complex concentrated in degree zero we get \begin{itemize} \item [$(1)$] {\normalfont \begin{equation*} \text{Hom}_{A}(A/\mathfrak{a},M)^{n}= \underset{j-i=n}{\prod}\text{Hom}_{A}((A/\mathfrak{a})^{i},M^{j})=\underset{j=n}{\prod}\text{Hom}_{A}(A/\mathfrak{a},M^{j}) \end{equation*} \begin{equation*} =\text{Hom}_{A}(A/\mathfrak{a},M^{n}). \end{equation*}} \item[$(2)$] \begin{equation*} (A/\mathfrak{ a}\otimes_{A}M)^{n}=\underset{i+j=n}{\bigoplus} ((A/\mathfrak{ a})^{i}\otimes_{A} M^{j})= \underset{j=n}{\bigoplus} (A/\mathfrak{ a}\otimes_{A} M^{j}) \end{equation*} \begin{equation*} =A/\mathfrak{ a}\otimes_{A} M^{n}. \end{equation*} \end{itemize} \end{prf} \begin{defn} {\normalfont Let $\mathfrak{a}$ be an ideal of $A$. An $A$-module $M$ is} $\mathfrak{a}$-torsion {\normalfont if $\Gamma_{\mathfrak{a}}(M)$ is isomorphic to $M$} {\normalfont and $M$ is called} $\mathfrak{ a}$-adically complete {\normalfont if $M$ is isomorphic to $\Lambda_{ \mathfrak{a}}(M)$, the $\mathfrak{ a}$-adic completion of $M$.} \end{defn} We denote by $M$ the $A$-complex $(M^{n}, d_{M}^{n})_{n\in\mathbb{Z}}$. \begin{defn} {\normalfont Let $\mathfrak{a}$ be an ideal of $A$ and $M$ an $A$-complex.} \begin{itemize} \item[$(1)$] {\normalfont $M$ is }$\mathfrak{a}$-adically complete {\normalfont if each $M^{n}$ is an $\mathfrak{a}$-adically complete $A$-module \cite[Definition 3.11]{L. Pol-Homotopy theory}}. \item [$(2)$] {\normalfont $M$ is} $\mathfrak{a}$-torsion {\normalfont if each $M^{n}$ is an $\mathfrak{a}$-torsion $A$-module}. \end{itemize} \end{defn} \begin{paragraph}\noindent $\text{C}(A)_{\mathfrak{a}\text{-tor}}$ and $\text{C}(A)_{\mathfrak{a}\text{-com}}$ denote the full subcategory of $\text{C}(A)$ consisting of $\mathfrak{a}$-torsion and $\mathfrak{a}$-adically complete complexes respectively. \end{paragraph} \begin{defn} {\normalfont An $A$-complex $ M$ is} $\mathfrak{a}$-reduced{ \normalfont if each component of $M$ is an $\mathfrak{a}$-reduced $A$-module}. \end{defn} \begin{paragraph}\noindent A subcomplex of an $\mathfrak{a}$-reduced complex is $\mathfrak{a}$-reduced. In particular, if $M$ is an $\mathfrak{a}$-reduced complex then the subcomplex $\Gamma_{\mathfrak{a}}(M)$ is $\mathfrak{a}$-reduced.\end{paragraph} We denote by $\text{C}(A)_{\mathfrak{a}\text{-red}}$ the full subcategory of $\text{C}(A)$ consisting of $\mathfrak{a}$-reduced complexes. \begin{prop}\label{sec2: Pro-I-red} {\normalfont Let $\mathfrak{ a}$ be an ideal of $A$ and $M$ be a complex over $A$. The following statements are equivalent: \begin{itemize} \item[$(1)$] $ M\in \text{C}(A)_{\mathfrak{a}\text{-red}}$, \item[$(2)$]$\text{Hom}_{A}(A/\mathfrak{a}, M)\cong \text{Hom}_{A}(A/\mathfrak{a}^{2},M)$, \item[$(3)$]$\Gamma_{\mathfrak{a}}(M)\cong$ $\text{Hom}_{A} (A/\mathfrak{a},M)$, \item[$ (4)$] $\mathfrak{a}\Gamma_{\mathfrak{a}}(M)\cong 0$. \end{itemize}} \end{prop} \begin{prf} \begin{itemize} \item[$(1)\Rightarrow (2)$] Since $M=(M^{n})_{n\in\mathbb{Z}}$ is $\mathfrak{a}$-reduced for each $n$ and a complex $A/\mathfrak{a}^{k}, k\in\mathbb{Z^{+}}$ is concentrated in degree zero, by Lemma \ref{sec2: Hom-tensor Lem} $ \text{Hom}_{A}(A/\mathfrak{a}, M)^{n}=\text{Hom}_{A}(A/\mathfrak{a}, M^{n}) \cong \text{Hom}_{A}(A/\mathfrak{a}^{2}, M^{n}) = \text{Hom}_{A}(A/\mathfrak{a}^{2}, M)^{n}$. So, the complexes $ \text{Hom}_{A}(A/\mathfrak{a}, M) $ and $\text{Hom}_{A}(A/\mathfrak{a}^{2}, M)$ are isomorphic. \item[$ (2) \Rightarrow (3)$] From $(2)$ there is an isomorphism $\text{Hom}_{A}(A/\mathfrak{a},M)^{n}\cong \text{Hom}_{A}(A/\mathfrak{a}^{k}, M)^{n}$ of $A$-modules for any $k\in\mathbb{Z^{+}}$ and all $ n\in\mathbb{Z}$. By \cite[Proposition 2.1 ]{David-Application-I} for each $n$ : \begin{equation*} \Gamma_{\mathfrak{a}}(M)^{n}:=\underset{k}\varinjlim~ \text{Hom}_{A}(A/\mathfrak{a}^{k}, M)^{n}\cong \text{Hom}_{A}(A/\mathfrak{a},M)^{n}. \end{equation*} So, the complexes $\Gamma_{\mathfrak{a}}(M)$ and $\text{Hom}_{A}(A/\mathfrak{a},M)$ are isomorphic. \item[$ (3) \Rightarrow (4)$] For all $n\in\mathbb{Z}$, the $A$-modules $\text{Hom}_{A}(A/\mathfrak{a},M)^{n}$ in the $A$-complex $\text{Hom}_{A}(A/\mathfrak{a},M)$ are annihilated by $\mathfrak{a}$. So, $\mathfrak{a}\text{Hom}_{A}(A/\mathfrak{a},M)\cong \mathfrak{a}\Gamma_{\mathfrak{a}}(M)\cong 0$. \item[$(4)\Rightarrow (1)$] Since $\mathfrak{a}\Gamma_{\mathfrak{a}}(M)\cong 0$ the $A$-modules $\mathfrak{a}\Gamma_{\mathfrak{a}}(M)^{n}=0$ for all $n$. It follows by \cite [Proposition 2.1]{David-Application-I} that $M^{n}$ is $\mathfrak{a}$-reduced for each $n$. Therefore, $M$ is an $\mathfrak{a}$-reduced complex. \end{itemize} \end{prf} \begin{defn} {\normalfont A complex $M$ is } $\mathfrak{a}$-coreduced {\normalfont if each of its $A$-module is $\mathfrak{a}$-coreduced.} \end{defn} Denote by $\text{C}(A)_{\mathfrak{a}\text{-cor}}$ the full subcategory of $\text{C}(A)$ consisting of $\mathfrak{a}$-coreduced complexes. We give the dual of Proposition \ref{sec2: Pro-I-red}. \begin{prop}\label{sec2: Pro I-cor} Let $A$ be a ring, $\mathfrak{a}$ an ideal in $A$ and $M$ a complex of $A$-modules. The following statements are equivalent. \begin{itemize} \item [$(1)$] $M\in {\normalfont \text{C}(A)_{\mathfrak{a}\text{-cor}}}$, \item[$(2)$]$A/\mathfrak{a}\otimes_{A} M\cong A/\mathfrak{a}^{2}\otimes_{A} M$, \item[$(3)$] $\Lambda_{\mathfrak{a}}(M)\cong A/\mathfrak{a}\otimes_{A} M$, \item[$(4)$] $\mathfrak{a}\Lambda_{\mathfrak{a}}(M)\cong 0$. \end{itemize} \end{prop} \begin{prf} \begin{itemize} \item [$(1)\Rightarrow (2)$] Let $M=(M^{n})_{n\in\mathbb{Z}}$ be an $\mathfrak{a}$-coreduced complex. For each $n$ the $A$-modules $M^{n}$ are $\mathfrak{a}$-coreduced and since $A/\mathfrak{a}^{k}, k\in\mathbb{Z^{+}}$ is concentrated in degree zero it follows by Lemma \ref{sec2: Hom-tensor Lem} that \begin{equation*} (A/\mathfrak{a}\otimes_{A}M)^{n}= A/\mathfrak{a}\otimes_{A} M^{n}\cong A/\mathfrak{a}^{2}\otimes_{A} M^{n}= (A/\mathfrak{a}^{2}\otimes_{A}M)^{n}. \end{equation*} So, $(A/\mathfrak{a}\otimes_{A}M)^{n}\cong (A/\mathfrak{a}^{2}\otimes_{A}M)^{n}$ for each $n$. Therefore the complexes $A/\mathfrak{a}\otimes_{A}M$ and $A/\mathfrak{a}^{2}\otimes_{A}M$ are isomorphic. \item[$(2)\Rightarrow (3)$] By $(2)$ for each $n\in\mathbb{Z}$ and for any $k\in\mathbb{Z^{+}}$ there is an isomorphism \\ $(A/\mathfrak{a}\otimes_{A}M)^{n}\cong (A/\mathfrak{a}^{k}\otimes_{A} M)^{n}$ . It follows that \begin{equation*} \Lambda_{\mathfrak{a}}(M)^{n}:=(\underset{k}\varprojlim~ A/\mathfrak{a}^{k}\otimes_{A}M)^{n}\cong (A/\mathfrak{a}\otimes_{A}M)^{n} \end{equation*} for all $n$. So, the complexes $\Lambda_{\mathfrak{a}}(M)$ and $A/\mathfrak{a}\otimes_{A} M$ are isomorphic. \item[$(3)\Rightarrow (4)$] For each $n\in\mathbb{Z}$ we have $A$-module isomorphisms $\mathfrak{a}\Lambda_{\mathfrak{a}}(M)^{n}\cong \mathfrak{a}(A/\mathfrak{a}\otimes_{A}M)^{n}\cong \mathfrak{a}(A/\mathfrak{a}\otimes_{A}M^{n})\cong 0$. So, $\mathfrak{a}\Lambda_{\mathfrak{a}}(M)\cong 0$. \item[$(4)\Rightarrow (1)$] Since $\mathfrak{a}\Lambda_{\mathfrak{a}}(M)\cong 0$ the $A$-modules $ \mathfrak{a}\Lambda_{\mathfrak{a}}(M)^{n}=0$ for all $n$. It follows by \cite[Proposition 2.2]{David-Application-I} that $M^{n}$ is $\mathfrak{a}$-coreduced for each $n$. Therefore $M$ is an $\mathfrak{a}$-coreduced complex. \end{itemize} \end{prf} \begin{prop} Let $\mathfrak{a}$ be a finitely generated ideal of $A$ and let $M$ be a complex over $A$. If $M$ is an $\mathfrak{a}$-coreduced complex then so is the $\mathfrak{a}$-adic completion $\Lambda_{\mathfrak{a}}(M)$ of $M$. \end{prop} \begin{prf} Since $\Lambda_\mathfrak{a}$ is an $A$-linear $\mathfrak{a}\Lambda_{\mathfrak{a}}(\Lambda_{\mathfrak{a}}(M))\cong \Lambda_{ \mathfrak{a}}(\mathfrak{a}\Lambda_{\mathfrak{a}}(M))\cong 0.$ Therefore by Proposition \ref{sec2: Pro I-cor} the $A$-complex $\Lambda_{\mathfrak{a}}(M)$ is $\mathfrak{a}$-coreduced. \end{prf} \begin{lem}\label{sec2: hom-red-cor module} Let $\mathfrak{a}$ be an ideal of $A$, $M$ an $\mathfrak{a}$-reduced $A$-module, and $N$ an $\mathfrak{a}$-coreduced $A$-module. Then for any module $X$ over $A$, \begin{itemize} \item[$(1)$]$\text{Hom}_{A}(X,M)\in (A\text{-Mod})_{\mathfrak{a}\text{-red}}$, \item[$(2)$] $\text{Hom}_{A}(N,X)\in (A\text{-Mod})_{\mathfrak{a}\text{-cor}}$. \end{itemize} \end{lem} \begin{prf} \begin{itemize} \item[$(1)$] Let $\varphi\in \text{Hom}_{A}(X,M)$ and $a^{2}\varphi=0$ for all $a\in A$. It follows that $a^{2}\varphi(x)=0$ for all $a$ and for all $x\in X$. Since $\varphi(x)\in M$ and $M$ is $\mathfrak{a}$-reduced, $a\varphi(x)=0$ for all $x$ and for all $a\in\mathfrak{ a}$. So, $a\varphi=0$ for any $a\in\mathfrak{a}$. The conclusion is that the $A$-module $\text{Hom}_{A}(X,M)$ is $\mathfrak{a}$-reduced by \cite[Definition 2.1]{kyomuhangi2020locally}. \item[$(2)$] Let $N$ be an $\mathfrak{a}$-coreduced $A$-module. In general, $\mathfrak{a}^{2}\text{Hom}(N,X)\subseteq \mathfrak{a}\text{Hom}(N,X)$ for any $A$-module $X$. We prove the reverse inclusion. Let $x\in \mathfrak{a}\text{Hom}(N,X)$. Then $x=\sum_{i=1}^{k}a_{i}f_{i}(n)$ where $k\in\mathbb{Z^{+}}, a_{i}\in \mathfrak{a}, n\in N$ and $f_{i}\in \text{Hom}(N,X)$ for each $i=1,2,\dots, k$. Thus $ x=\sum_{i=1}^{k}f_{i}(a_{i}n)$ with $a_{i}n\in \mathfrak{a}N=\mathfrak{a}^{2}N$ since $N$ is $\mathfrak{a}$-coreduced \cite[Proposition 2.2]{David-Application-I}. So, $a_{i}n\in \mathfrak{a}^{2}N$ for all $i$. Therefore, $a_{i}m=\sum_{j=1}^{t}b_{j}n_{j}$ where $b_{j}\in \mathfrak{a}^{2}$ and $n_{j}\in N$. It follows that $x=\sum_{i=1}^{k}f_{i}(\sum_{j=1}^{t}b_{j}n_{j})=\sum_{i=1}^{k} \sum_{j=1}^{t} b_{j}f_{i}(n_{j})= \sum_{j=1}^{t} b_{j}(\sum_{i=1}^{k}f_{i}(n_{j}))\in \mathfrak{a}^{2}\text{Hom}(N,X)$. \end{itemize} \end{prf} \begin{prop}\label{sec2: Prop Hom(-,M) and Hom(M,-)} Let $M\in {\normalfont\text{C}(A)_{\mathfrak{a}\text{-red}}}$ and $ {\normalfont N\in \text{C}(A)_{\mathfrak{a}\text{-cor}}}$. For any complex $X$, \begin{itemize} \item [$(1)$]$\text{ Hom}_{A}(X,M)\in {\normalfont\text{C}(A)_{\mathfrak{a}\text{-red}}} $, \item[ $(2)$] $\text{Hom}_{A}(N,X)\in {\normalfont \text{C}(A)_{\mathfrak{a}\text{-cor}}}$. \end{itemize} \end{prop} \begin{prf} Let $X:=(X^{i},d^{i}_{X})_{i\in\mathbb{Z}}$ and $M:=(M^{j},d^{j}_{M})_{j\in\mathbb{Z}}$ be $A$-complexes with $M$ $\mathfrak{a}$-reduced. The $n^{th}$ component of the complex $\text{Hom}_{A}(X,M)$ is the $A$-module $\underset{n=j-i}{\prod}Hom_{A}(X^{i},M^{j})$. Since by Lemma \ref{sec2: hom-red-cor module} $\text{Hom}_{A}(X^{i},M^{j})$ is $\mathfrak{a}$-reduced for each $j$ and the direct product of $\mathfrak{a}$-reduced modules is again $\mathfrak{a}$-reduced \cite[Proposition 2.4] {kyomuhangi2020locally} it follows that the complex $\text{Hom}_{A}(X,M)$ is $\mathfrak{a}$-reduced. The proof of (2) is similar. \end{prf} \begin{paragraph}\noindent Proposition \ref{sec2: Prop Hom(-,M) and Hom(M,-)} says that the functors $\text{Hom}(M,-)$ (resp. $\text{Hom}(-,M)$) are endo functors on $\text{C}(A)_{\mathfrak{a}\text{-red}}$ (resp. $\text{C}(A)_{\mathfrak{a}\text{-cor}}$). \end{paragraph} \begin{cor} Let $\mathfrak{ a}$ be an ideal of a ring $A$, and let $N,X$ be complexes over $A$. Denote by $(.)^{\vee}$ the general Matlis duality functor. \begin{itemize} \item[$(1)$] If $X$ is $\mathfrak{ a}$-reduced, then $\text{Hom}_{A}(N/\mathfrak{ a}N, X)$ is $\mathfrak{ a}$-reduced. \item[$(2)$] If $N$ is $\mathfrak{ a}$-coreduced, then $(N\otimes_{A}X)^{\vee}$ is $\mathfrak{ a}$-coreduced. \item[$(3)$] If $X$ is $\mathfrak{ a}$-coreduced, then $(N\otimes_{A}X)^{\vee}$ is $\mathfrak{ a}$-reduced. \end{itemize} \end{cor} \begin{prf} Since $\text{Hom}_{A}(N/\mathfrak{ a}N, X)\cong \text{Hom}_{A}(A/\mathfrak{ a}, \text{Hom}_{A}(N, X))$ $(1)$ follows by Proposition \ref{sec2: Prop Hom(-,M) and Hom(M,-)} . Moreover, since the general Matlis dual for complexes $P, X$ is given by $(P\otimes_{A}X)^{\vee}\cong \text{Hom}_{A}(P, X^{\vee})$ the proof of $(2)$ is again immediate by Proposition \ref{sec2: Prop Hom(-,M) and Hom(M,-)} with $P=N$. Let $ n\in\mathbb Z$ and $M$ an $\mathfrak{a}$-coreduced module. $\text{Hom}_{A}(M^{n},-)$ is $\mathfrak{a}$-reduced module for all $n$ by \cite[Proposition 2.4]{David-Application-I}. It follows that $ X^{\vee}$ is $\mathfrak{ a}$-reduced for any $\mathfrak{ a}$-coreduced complex $X$. Also the general Matlis dual $(N\otimes_{A}X)^{\vee}$ is $\mathfrak{a}$-reduced complex. This proves $(3)$. \end{prf} \begin{cor} Let $\mathfrak{a}$ be an ideal of a ring $A$, and let $f: X\rightarrow Y$ be a quasi-isomorphism between $\text{K}$-injective (resp. $\text{K}$-projective) $A$-complexes. The morphism $ \text{Hom}_{A}(f,Z): \text{Hom}_{A}(Y, Z)\rightarrow \text{Hom}_{A}(X, Z)$ (resp. $ \text{Hom}_{A}(Z',f): \text{Hom}_{A}(Z', X)\rightarrow \text{Hom}_{A}(Z', Y)$ is quasi-isomorphism of $\mathfrak{a}$-reduced (resp. $\mathfrak{a}$-coreduced) complexes for any $\mathfrak{a}$-reduced complex $Z$ and $\mathfrak{a}$-coreduced complex $Z'$. \end{cor} \begin{prf} Let $f: X\rightarrow Y$ be a quasi-isomorphism between $K$-injective (resp. $K$-projective) $A$-complexes. Then by \cite[Corollary 4.4.11]{Peter-Schenzez} $f$ induces quasi-isomorphisms $\text{Hom}_{A}(f,Z): \text{Hom}_{A}(Y, Z)\rightarrow \text{Hom}_{A}(X, Z)$ (resp. $ \text{Hom}_{A}(Z',f): \text{Hom}_{A}(Z',X)\rightarrow \text{Hom}_{A}(Z', Y))$ for any $A$-complexes $Z$ and $Z'$. The conclusion is clear by Proposition \ref{sec2: Prop Hom(-,M) and Hom(M,-)}. \end{prf} \begin{exam}\label{sec2 : Example} Let $\mathfrak{ a}$ be an ideal of a ring $A$, and let $M$ be an $A$-complex. \begin{enumerate} \item[$(1)$] If $M$ is a complex of $\mathfrak{ a}$-torsion-free modules, then $M$ is an $\mathfrak{ a}$-reduced complex. \item[$(2)$] Let $M$ be a complex of semi-simple $A$-modules. Then $M$ is both $\mathfrak{a}$-reduced and $\mathfrak{a}$-coreduced. \item[$(3)$] $\text{Hom}_{A}(M/\mathfrak{ a}M, N)$ is both $\mathfrak{ a}$-reduced and $\mathfrak{ a}$-coreduced for any $A$-complex $N$. In particular, $\text{Hom}_{A}(A/\mathfrak{ a}, N)$ is $\mathfrak{ a}$-reduced. \item[$(4)$] The quotient complex $M/\mathfrak{ a}M$ is $\mathfrak{a}$-coreduced. \item[$(5)$] If $\mathfrak{ a}^{2}=\mathfrak{ a}$, then $M$ is both $\mathfrak{ a}$-reduced and $\mathfrak{ a}$-coreduced. \end{enumerate} \end{exam} \section{GM Duality and MGM Equivalence} \begin{paragraph}\noindent The purpose of this section is to prove the two main theorems of our paper. $\mathfrak{C}(A)$ denotes the full subcategory $ \text{C}(A)_{\mathfrak{a}\text{-red}} ~\cap~ \text{C}(A)_{\mathfrak{a}\text{-cor}}$ of $\text{C}(A)$. \end{paragraph} \begin{lem}\label{sec3: Idempotent-composition} Let $\mathfrak{a}$ be an ideal of a ring $A$, and let $M$ be a complex of $A$-modules. \begin{itemize} \item [$(1)$] The functors $$A/\mathfrak{a}~\otimes_{A}-,\text{ Hom} _{A}(A/\mathfrak{a},-): \text{C}(A)\rightarrow \mathfrak{C}(A)$$ are idempotent. \item[$(2)$] For any complex $M$, $ A/\mathfrak{a} ~\otimes_{A}\text{ Hom}_{A}(A/\mathfrak{a},M)\cong Hom_{A}(A/\mathfrak{a},M)$ \text{and} $$\text{Hom}_{A}(A/\mathfrak{a}, A/\mathfrak{a}\otimes_{A} M)\cong A/\mathfrak{a}\otimes_{A} M.$$ \item[$(3)$] For any $A$-complex $M$, $\text{Hom}_{A}(A/\mathfrak{a},M)\in \text{C}(A)_{\mathfrak{a\text{-tor}}}$ and $A/\mathfrak{a}~\otimes_{A} M\in \text{C}(A)_{\mathfrak{a\text{-com}}}$. \end{itemize} \end{lem} \begin{prf} \begin{enumerate} \item[$(1)$] Let $ M=(M^{n})_{n\in\mathbb{Z}}\in \text{C}(A)$. Set $L=A/\mathfrak{a}\otimes_{A}M$. Since $A/\mathfrak{a}$ is concentrated in degree $0$ by Lemma \ref{sec2: Hom-tensor Lem}, \begin{equation*} (A/\mathfrak{a}~ \otimes_{A} L)^{n}=A/\mathfrak{a}~\otimes_{A} L^{n}= A/\mathfrak{a}~\otimes_{A} (A/\mathfrak{a}~\otimes_{A}~ M)^{n}\cong A/\mathfrak{a}~\otimes_{A} (A/\mathfrak{a}~\otimes_{A}~ M^{n}) \end{equation*} \begin{equation*} \cong A/\mathfrak{a}~\otimes_{A} M^{n} \cong (A/\mathfrak{a}~\otimes_{A}~ M)^{n}. \end{equation*} Thus \begin{equation*} (A/\mathfrak{a}\otimes_{A}(A/\mathfrak{a}\otimes_{A}M))^{n}=(A/\mathfrak{a}\otimes_{A}M)^{n} \end{equation*} for all $n$. It follows that $A/\mathfrak{a}\otimes_{A}(A/\mathfrak{a}\otimes_{A}M)\cong A/\mathfrak{a}\otimes_{A}M$. Therefore, the functor $A/\mathfrak{a}\otimes_{A}-$ is idempotent on $\text{C}(A)$. For the second idempotence it suffices to show that for any $n\in\mathbb{Z}$, \begin{equation*} \text{Hom}_{A}(A/\mathfrak{a},\text{Hom}_{A}(A/\mathfrak{a},M)^{n}= \text{Hom}_{A}(A/\mathfrak{a},M)^{n}. \end{equation*} Set $N:= \text{Hom}_{A}(A/\mathfrak{a},M)$. By Lemma \ref{sec2: Hom-tensor Lem}, $ N^{n}:= \text{Hom}_{A}(A/\mathfrak{a},M^{n}).$ So, $\text{Hom}_{A}(A/\mathfrak{a},\text{Hom}_{A}(A/\mathfrak{a},M))^{n}= \text{Hom}_{A}(A/\mathfrak{a},N)^{n}=\text{Hom}_{A}(A/\mathfrak{a},N^{n}) =\text{Hom}_{A}(A/\mathfrak{a},\text{Hom}_{A}(A/\mathfrak{a},M^{n})).$ By \cite[Proposition 2.6]{David-Application-I} \begin{equation*} \text{Hom}_{A}(A/\mathfrak{a},\text{Hom}_{A}(A/\mathfrak{a},M^{n}))\cong \text{Hom}_{A}(A/\mathfrak{a},M^{n})= \text{Hom}_{A}(A/\mathfrak{a},M)^{n}. \end{equation*} Therefore for any $A$-complex $M$ \begin{equation*} \text{Hom}_{A}(A/\mathfrak{a},\text{Hom}_{A}(A/\mathfrak{a},M))\cong \text{Hom}_{A}(A/\mathfrak{a},M). \end{equation*} For the remaining part, let $M$ be an $A$-complex and $n\in\mathbb{Z}$. By \cite[Proposition 2.6]{David-Application-I} the $A$-modules $ \text{Hom}_{A}(A/\mathfrak{a}, M^{n})$ and $A/\mathfrak{a}\otimes_{A} M^{n}$ are both $\mathfrak{ a}$-reduced and $\mathfrak{ a}$-coreduced for each $n$. By Lemma \ref{sec2: Hom-tensor Lem} it follows that $\text{Hom}_{A}(A/\mathfrak{a}, M)^{n}$ and $(A/\mathfrak{a}\otimes_{A} M)^{n}$ are $\mathfrak{ a}$-reduced and $\mathfrak{ a}$-coreduced for all $n$. This shows that the $A$-complexes $\text{Hom}_{A}(A/\mathfrak{a}, M)$ and $A/\mathfrak{a}~\otimes_{A} M $ are both $\mathfrak{ a}$-reduced and $\mathfrak{ a}$-coreduced. So, they are contained in $\mathfrak{C}(A)$. \item[$(2)$] \begin{itemize} \item[i.]$ (A/\mathfrak{a} ~\otimes_{A} \text{Hom}_{A}(A/\mathfrak{a},M))^{n}= A/\mathfrak{a} ~\otimes_{A} (\text{Hom}_{A}(A/\mathfrak{a},M))^{n} = A/\mathfrak{a} ~\otimes_{A} \text{Hom}_{A}(A/\mathfrak{a},M^{n})$. However by \cite[Proposition 2.6]{David-Application-I} $$A/\mathfrak{a} ~\otimes_{A} (\text{Hom}_{A}(A/\mathfrak{a},M^{n})\cong \text{Hom}_{A}(A/\mathfrak{a},M^{n})~ \text{for all} ~n. $$ So, $$(A/\mathfrak{a} ~\otimes_{A} \text{Hom}_{A}(A/\mathfrak{a},M))^{n}\cong \text{Hom}_{A}(A/\mathfrak{a},M^{n})\cong \text{Hom}_{A}(A/\mathfrak{a},M)^{n}.$$ Therefore for any complex $M$, $ A/\mathfrak{a} ~\otimes_{A} \text{Hom}_{A}(A/\mathfrak{a},M)\cong \text{Hom}_{A}(A/\mathfrak{a},M).$ \item[ii.] We show that $$\text{Hom}_{A}(A/\mathfrak{a}, A/\mathfrak{a}\otimes_{A}M)^{n}=(A/\mathfrak{a}\otimes_{A}M)^{n}~ \text{for all}~ n.$$ $\text{Hom}_{A}(A/\mathfrak{a}, A/\mathfrak{a}~\otimes_{A}M)^{n}= \text{Hom}_{A}(A/\mathfrak{a},( A/\mathfrak{a}~\otimes_{A}M)^{{n}}) =$ $$\text{Hom}_{A}(A/\mathfrak{a}, A/\mathfrak{a}~\otimes_{A}M^{{n}})\cong A/\mathfrak{a}~\otimes_{A}M^{{n}}~\text{by \cite[Proposition 2.6]{David-Application-I} for all $n$}.$$ Therefore there is an isomorphism of $A$-complexes $\text{Hom}_{A}(A/\mathfrak{a}, A/\mathfrak{a}\otimes_{A}M)\cong A/\mathfrak{a}\otimes_{A}M.$ \end{itemize} \item[$(3)$] For an $A$-complex $M$ and $n\in\mathbb{Z}$, $\text{Hom}_{A}(A/\mathfrak{a},M)^{n}= \text{Hom}_{A}(A/\mathfrak{a},M^{n})$ and $(A/\mathfrak{a}\otimes_{A}M)^{n}= A/\mathfrak{a}~\otimes_{A}M^{n}$ [Lemma \ref{sec2: Hom-tensor Lem}]. By \cite[Proposition 2.6]{David-Application-I} the $A$-modules $\text{Hom}_{A}(A/\mathfrak{a},M^{n})$ (resp.$~A/\mathfrak{a}~\otimes_{A}M^{n}$) are $\mathfrak{a}$-torsion (resp. $\mathfrak{a}$-complete $A$-modules) for all $n$. Therefore the $A$-complexes $\text{Hom}_{A}(A/\mathfrak{a},M)$ (resp. $A/\mathfrak{a}~\otimes_{A}M$) are $\mathfrak{a}$-torsion (resp. $\mathfrak{a}$-adically complete) complexes. \end{enumerate} \end{prf} \begin{thm}[GM Duality in $\text{C}(A)$]\label{sec3: GM Duality-Thm} Let $A$ be a ring and $\mathfrak{a}$ an ideal in it. \begin{itemize} \item [$(1)$] The functor {\normalfont\begin{equation*} \Gamma_{\mathfrak{a}}: \text{C}(A)_{\mathfrak{a}\text{-red}}\rightarrow \text{C}(A)_{\mathfrak{a}\text{-cor}}; ~ M\mapsto \Gamma_{\mathfrak{a}}(M) \end{equation*}} is idempotent and {\normalfont $\Gamma_{\mathfrak{a}}(M)\cong \text{Hom}_{A}(A/\mathfrak{a},M)$}. \item[$(2)$] The functor {\normalfont \begin{equation*} \Lambda_{\mathfrak{a}}: \text{C}(A)_{\mathfrak{a}\text{-cor}}\rightarrow \text{C}(A)_{\mathfrak{a}\text{-red}};~ M\mapsto \Lambda_{\mathfrak{a}}(M) \end{equation*}} is idempotent and {\normalfont $\Lambda_{\mathfrak{a}}(M)\cong A/\mathfrak{a}\otimes_{A} M$}. \item[$(3)$] For any $ M\in {\normalfont \text{C}(A)_{\mathfrak{a}\text{-red}}} $ and any {\normalfont $ N\in \text{C}(A)_{\mathfrak{a}\text{-cor}}$}, {\normalfont \begin{equation*} \text{Hom}_{A}( \Lambda_{\mathfrak{a}}(M),N)\cong \text{Hom}_{A}( M,\Gamma_{\mathfrak{a}} (N)). \end{equation*}} \end{itemize} \end{thm} \begin{prf} \begin{itemize} \item [$(1)$] For any $\mathfrak{a}$-reduced module $N$, $\Gamma_{\mathfrak{a}}(N)= \underset{k}\varinjlim~ {\text{Hom}_{A}(A/\mathfrak{a}^{k},N)}$ is an $\mathfrak{ a}$-coreduced module \cite[Proposition 2.7]{David-Application-I}. Let $M$ be an $\mathfrak{ a}$-reduced complex. Applying the functor $\Gamma_{\mathfrak{a}}$ levelwise to $M$ one gets $\Gamma_{\mathfrak{a}}(M)\in \text{C}(A)_{\mathfrak{ a\text{-cor}}}$. By Proposition \ref{sec2: Pro-I-red} $\Gamma_{\mathfrak{a}}(M)\cong \text{Hom}_{A}(A/\mathfrak{a},M)$. The functor $ \text{Hom}_{A}(A/\mathfrak{a},-)$ is idempotent by Lemma \ref{sec3: Idempotent-composition}. \item[$(2)$] For any $\mathfrak{a}$-coreduced complex $M$, $\Lambda_{\mathfrak{a}}(M)\cong A/\mathfrak{a}~\otimes_{A}M$ by Proposition \ref{sec2: Pro I-cor}. Since for any $\mathfrak{a}$-coreduced module $N$ $\Lambda_{\mathfrak{a}}(N) = \underset{k}\varprojlim~ {N/\mathfrak{a}^{k}N}$ is an $\mathfrak{ a}$-reduced $A$- module \cite[Proposition 2.7]{David-Application-I}, levelwise application extends $\Lambda_{ \mathfrak{a}}$ to a complex of $\mathfrak{ a}$-reduced modules. Thus if $M$ is an $\mathfrak{ a}$-coreduced complex then $\Lambda_{\mathfrak{a}}(M)\in \text{C}(A)_{\mathfrak{ a\text{-red}}}$. The idempotence $\Lambda_{ \mathfrak{a}}(M)\cong A/\mathfrak{a}\otimes_{A} M$ follows by Lemma \ref{sec3: Idempotent-composition}. \item[$(3)$] The functor $ A/\mathfrak{a}\otimes_{A}-$ is left adjoint to $\text{Hom}_{A}(A/\mathfrak{a},-)$ in the category of $A$-modules. Levelwise application extends the functors to the category $\text{C}(A)$. Also for any $M\in\text{C}(A)_{\mathfrak{a}\text{-red}}$, $\Gamma_{\mathfrak{a}}(M)\cong \text{Hom}_{A}(A/\mathfrak{a},M)$ and for any $ N\in \text{C}(A)_{\mathfrak{a}\text{-cor}}$, $ A/\mathfrak{a}~\otimes_{A}N\cong \Lambda_{\mathfrak{a}}(N) $. By uniqueness of adjoints $\Lambda_{ \mathfrak{a}}$ is left adjoint to $\Gamma_{ \mathfrak{a}}$. \end{itemize} \end{prf} \begin{lem}\label{sec 3: Lem I-red-I-tor...} Let $\mathfrak{a}$ be an ideal of a ring $A$, and let $M$ be a complex over $A$. \begin{itemize} \item[$(1)$] If $M$ is $\mathfrak{a}$-reduced, then $\Gamma_{\mathfrak{a}}(M) $ is an $\mathfrak{a}$-adically complete complex. \item[$(2)$] If $M$ is $\mathfrak{a}$-coreduced, then $\Lambda_{\mathfrak{a}}(M)$ is an $\mathfrak{a}$-torsion $A$-complex. \end{itemize} \end{lem} \begin{prf} \begin{itemize} \item[$(1)$] By Lemma \ref{sec3: Idempotent-composition} $\Gamma_{\mathfrak{a}}(M)\cong \text{Hom}_{A}( A/\mathfrak{a}, M)$ which is both $\mathfrak{a}$-reduced and $\mathfrak{a}$-coreduced. So, $\Lambda_{\mathfrak{a}}(\Gamma_{\mathfrak{a}}(M))\cong A/\mathfrak{a}\otimes_{A} \text{Hom}_{A}( A/\mathfrak{a}, M)\cong \text{Hom}_{A}( A/\mathfrak{a},M)\cong \Gamma_{\mathfrak{a}}(M)$. \item[$(2)$] If $M$ is an $\mathfrak{a}$-coreduced complex, then $\Lambda_{\mathfrak{a}}(M)\cong A/\mathfrak{a}~\otimes_{A} M$ which is both $\mathfrak{a}$-reduced and $\mathfrak{a}$-coreduced. It follows that $ \Gamma_{\mathfrak{a}}(\Lambda_{\mathfrak{a}}(M))\cong \text{Hom}_{A}(A/\mathfrak{a},\Lambda_{\mathfrak{a}}(M))\cong \text{Hom}_{A}(A/\mathfrak{a}, A/\mathfrak{a}~ \otimes_{A}M)\cong A/\mathfrak{a}~\otimes_{A} M\cong \Lambda_{\mathfrak{a}}(M)$. \end{itemize} \end{prf} \begin{thm}[The MGM Equivalence in $\text{C}(A)$]\label{sec3: MGM equivalence} Let $\mathfrak{a}$ be an ideal of a ring $A$. Let $\mathfrak{A}:=\text{C}(A)_{\mathfrak{a}\text{-tor}}\cap \text{C}(A)_{\mathfrak{a}\text{-red}}$ and $\mathfrak{B}:= \text{C}(A)_{\mathfrak{a}\text{-com}}\cap \text{C}(A)_{\mathfrak{a}\text{-cor}}$.\\ \begin{itemize} \item [$(1)$] For an $ M \in\text{C}(A)_{\mathfrak{a}\text{-cor}}$, the $A$-complex $\Lambda_{\mathfrak{a}}(M)\in \mathfrak{A}$. \item[$(2)$] For any $M \in\text{C}(A)_{\mathfrak{a}\text{-red}}$, the complex $\Gamma_{\mathfrak{a}}(M)\in \mathfrak{B}$. \item [$(3)$] The functor \begin{equation*} \Gamma_{\mathfrak{a}}: \text{C}( A)_{\mathfrak{a}\text{-red}}\rightarrow \text{C}(A)_{\mathfrak{a}\text{-cor}} \end{equation*} restricted to $\mathfrak{A}$ is an equivalence between $\mathfrak{A}$ and $\mathfrak{B}$ with quasi-inverse $\Lambda_{\mathfrak{a}}$. \end{itemize} \end{thm} \begin{prf} {\normalfont \begin{itemize} \item[$(1)$] If $M$ is an $\mathfrak{ a}$-coreduced complex, then $\Lambda_{\mathfrak{a}}(M)\in \text{C}(A)_{\mathfrak{ a}\text{-red}}$ by Theorem \ref{sec3: GM Duality-Thm}. Also, $\Lambda_{\mathfrak{a}}(M)\in \text{C}(A)_{\mathfrak{a}\text{-tor}}$ by Lemma \ref{sec 3: Lem I-red-I-tor...}. So, $\Lambda_{\mathfrak{a}}(M)\in\mathfrak{A}$. \item [$(2)$] Let $M$ be an $\mathfrak{ a}$-reduced complex. By Theorem \ref{sec3: GM Duality-Thm}, $\Gamma_{\mathfrak{a}}(M)\in \text{C}(A)_{\mathfrak{ a}\text{-cor}}$ and by Lemma \ref{sec 3: Lem I-red-I-tor...} $\Gamma_{\mathfrak{ a}}(M)\in \text{C}(\text{A-Mod})_{\mathfrak{a}\text{-com}}$. So, $\Gamma_{\mathfrak{a}}(M)\in \mathfrak{B}$. \item[$(3)$] Let $M\in\mathfrak{A}$. By Lemma \ref{sec 3: Lem I-red-I-tor...}, $\Gamma_{\mathfrak{a}}(M)$ is $\mathfrak{a}$-adically complete. So, $ \Lambda_{\mathfrak{a}}(\Gamma_{\mathfrak{ a}}(M))\cong \Gamma_{\mathfrak{ a}}(M)\cong M$ since $M$ is $\mathfrak{ a}$-torsion by hypothesis. On the other hand, if $M\in\mathfrak{B}$ then $\Lambda_{\mathfrak{a}}(M)$ is an $\mathfrak{ a}$-torsion complex [Lemma \ref{sec 3: Lem I-red-I-tor...}]. Thus $ \Gamma_{\mathfrak{ a}}(\Lambda_{\mathfrak{a}}(M))\cong \Lambda_{\mathfrak{a}}(M)\cong M$ since by hypothesis $M$ is an $\mathfrak{ a}$-adically complete complex. \end{itemize}} \end{prf} \subsection{ A more general condition} \begin{paragraph}\noindent Let $\mathfrak{a}$ be an ideal of a ring $A$. In this subsection we demonstrate that the $\mathfrak{ a}$-reduced and $\mathfrak{ a}$-coreduced notion is a more general setting than that of weakly proregular ideals of $A$ for both GM duality and MGM equivalence to hold. \end{paragraph} \begin{paragraph}\noindent Recall that an inverse system of $A$-modules $\{M_{i}\}_{i\in\mathbb{Z}^{+}}$ is called \textit{pro-zero} if for every $i$ there exists $j\ge i$ such that the homomorphism $M_{j}\rightarrow M_{i}$ is zero. Given a sequence $\textbf{a} = (a_{1} ,\dots , a_{n})$ of elements in $A$, and $i\in \mathbb{Z^{+}}$, we let $\textbf{a}^ {i} :=(a^{i}_{1} ,\dots , a^{i}_{n} )$. A collection of the associated Koszul complexes $\{\text{K}(A;\textbf{a}^{ i} )\}_{ i\in\mathbb{Z^{+}}}$ forms an inverse system \cite{Peter-Schenzez,David-Application-I, Amnon-WPR-der comp-prism}. \end{paragraph} \begin{defn} {\normalfont Let $A$ be a ring. A finite sequence $\textbf{a}$ in $A$ is called} weakly proregular {\normalfont (WPR) if for every $q < 0$ the inverse system of $A$-modules $ \{\text{H}^{q}(\text{K}(A;\textbf{a}^{ i}))\}_{ i\in\mathbb{Z^{+}}}$ is pro-zero.} \end{defn} \begin{defn} {\normalfont An ideal $\mathfrak{a}$ in $A$ is called} weakly proregular {\normalfont if it is generated by some weakly proregular sequence \textbf{a}}. \end{defn} \begin{paragraph}\noindent A WPR ideal $\mathfrak{ a}$ is finitely generated. Moreover, if the ring $A$ is Noetherian then every ideal $\mathfrak{ a}$ in $A$ is weakly progregular \cite[Lemma 2.4]{Hartshorne-Local cohomology}. In the derived category setting, Porta M. et'al \cite[Theorem 7.11]{porta2014homology} and \cite[Theorem 9]{Erratum to: On the homology of comp and tor} among others proved the GM duality and MGM equivalence when the ideal $\mathfrak{a}$ is weakly proregular. However, we proved these notions [Theorems \ref{sec3: GM Duality-Thm} and \ref{sec3: MGM equivalence}] in $\text{C}(A)$ by requiring a setting of $\mathfrak{ a}$-reduced and $\mathfrak{ a}$-coreduced complexes instead of weak proregularity of $\mathfrak{a}$. Following \cite{David-Application-I}, it is instructive to compare the two conditions. We remark that the conditions WPR and that of $\mathfrak{ a}$-reduced and $\mathfrak{ a}$-coreduced are distinct in general. There are however conditions under which the two notions coincide. \end{paragraph}\noindent \begin{paragraph}\noindent An idempotent ideal need not be WPR. However, if $\mathfrak{ a}$ is an idempotent ideal then every complex in $\text{C}(A)$ is both $\mathfrak{ a}$-reduced and $\mathfrak{a}$-coreduced. On the other hand, every ideal $\mathfrak{ a}$ of a Noetherian ring is WPR. However, not all complexes over Noetherian ring are $\mathfrak{ a}$-reduced or $\mathfrak{ a}$-coreduced. Moreover, it is not possible to achieve the MGM equivalence and GM duality in the setting of $\text{M}(A)$ and $\text{C}(A)$ with weak proregularity; which is possible with the conditions of $\mathfrak{ a}$-reduced and $\mathfrak{ a}$-coreduced. \end{paragraph} \begin{paragraph}\noindent Lastly, we recall coincidence of WPR with a notion of reduced $A$-module. Let $a\in A$. An ideal $(a)$ of $A$ is WPR if and only if $A$, as an $A$-module, is $a^{t}$-reduced for some $t\in \mathbb{Z^{+}}$, \cite[Proposition 5.6]{Amnon-WPR-der comp-prism}. It follows therefore that an ideal $(a)$ of a ring $A$ is WPR if and only if $A$, as an $A$-complex concentrated in degree $0$, is $a^{t}$-reduced for some $t\in \mathbb{Z^{+}}$, where a complex $M$ is \textit{ $a^{t}$-reduced} if each $A$-module in $M$ is $a^{t}$-reduced. Since we dealt with general complexes it is reasonable to say that $\mathfrak{ a}$-reduced and $\mathfrak{ a}$-coreduced complexes provide a more general setting for which the GM duality and MGM equivalence hold. \end{paragraph}\noindent \newline\newline \textbf{ ACKNOWLEDGMENT} \begin{paragraph}\noindent We acknowledge support from the Eastern Africa Algebra Research Group (EAALG), the EMS-Simons for Africa Program and the International Science Program (ISP). Part of this work was written while the second author was visiting the third author at Makerere University. The second author is grateful for the hospitality. \end{paragraph} \addcontentsline{toc}{chapter}{Bibliography}

c73d1f3eed1452a8f2c0329cc8cf06cf99d86e58

\section{Introduction} The solar photosphere is home to the abundant metal lines that appear absorptive against the optical continuum background. These lines contain rich information of the photosphere and has been intensively studied since the 20th century. Most studies are devoted to the determination of element abundances from the equivalent widths of these absorption lines \citep[e.g.][]{2021asplund}. Micro- and macroturbulent velocities can also be obtained from the equivalent widths or line profiles \citep{1995takeda,2019sheminova,2022takeda}. Transitions between the \ion{Si}{I} 3p4p $^3$D$_{1,2,3}$ and 3p7d $^3$F$^\mathrm{o}_{2,3,4}$ levels give rise to six allowed lines in the visible waveband, and one of them resides at the H$\alpha$ blue wing, with a wavelength of 6560.58 \AA\ \citep{1965radziemski,1968lambert}. This line has been identified in previous solar atlases as blended with another telluric absorption line from the water vapor \citep{1966moore,1973delbouille}. The Kitt Peak solar atlas has unveiled this line for the first time after a careful evaluation of the atmospheric transmission spectra \citep{1984kurucz}. However, in spite of the only record of the equivalent width \citep{1966moore,1968lambert}, information about this line is still scarce due to line blending. The Chinese H$\alpha$ Solar Explorer \citep[CHASE,][]{2019li,2022li} is a space-based telescope that can perform spectroscopic observations of the full solar disk in the H$\alpha$ waveband. The absorptive \ion{Si}{I} 6560.58 \AA\ line clearly stands out in the sample spectra \citep{2022qiu}, without any distortion or disturbance from the earth atmosphere, which makes it feasible to study this line. In this paper, we use the full disk spectra of the \ion{Si}{I} 6560.58\AA\ line and investigate the formation of this line and its diagnostics of the solar photosphere. We briefly introduce the observations and data reduction methods in Sec.~\ref{sect2}. The results are shown in Sec.~\ref{sect3}, followed by a conclusion in Sec.~\ref{sect4}. \section{Observations and Data Reduction} \label{sect2} CHASE\footnote{https://ssdc.nju.edu.cn} can regularly scan the full solar disk in both the H$\alpha$ and \ion{Fe}{I} wavebands. Each scanning takes $\sim$46 s, with a spectral sampling of 24.2 m\AA. After in-orbit focusing calibration at the beginning of August 2022, a spatial resolution of 1.\arcsec2 has been achieved. Normally a $2\times$ binning is used to reduce the data size. We choose three full-Sun observation periods, all with flares in the northern hemisphere. The selected scanning in each observation period is closest to the flare peak time, as listed in Table~\ref{obs}. The selected scanning OBS1 is near the flare peak time, while OBS2 is in the pre-flare phase and OBS3 is in the decay phase. We show the reconstructed images at the \ion{Si}{I} and H$\alpha$ line centers in Fig.~\ref{image}, note that part of the solar disk in OBS1 is outside the field of view. The signal-to-noise ratio (S/N) is calculated in the H$\alpha$ far wings, and expressed in unit of decibel (dB). \begin{figure} \centering \includegraphics[width=0.5\textwidth]{image_siha-eps-converted-to.pdf} \caption{Reconstructed images at the \ion{Si}{I} and H$\alpha$ line centers. } \label{image} \end{figure} \begin{figure} \centering \includegraphics[width=0.5\textwidth]{si_prof-eps-converted-to.pdf} \caption{Sample disk-center spectra with identified lines from the CHASE H$\alpha$ and \ion{Fe}{I} wavebands after radiometric calibration. The black curve is the original spectra, and the red one is after stray light correction. } \label{prof} \end{figure} \begin{table*} \caption{Basic information of the selected full-Sun scannings. } \label{obs} \centering \begin{tabular}{cccccc} \hline Label & Observation Period & Selected Scanning & Flare Class & Flare Location & S/N (dB) \\ \hline OBS1 & 2022.01.29 11:39:58--11:56:56 & 11:53:57 & C2.5 & N17E13 & 21.03 \\ OBS2 & 2022.01.17 14:51:05--15:08:04 & 15:07:04 & C2.7 & N25W37 & 20.97 \\ OBS3 & 2022.01.20 06:08:39--06:25:37 & 06:08:39 & M5.5 & N08W76 & 21.30 \\ \hline \end{tabular} \end{table*} As shown in \cite{2022qiu}, the enhancement of the observed line center is mainly due to the contamination of stray light (Fig.~\ref{prof}). Following \cite{2020hou}, the level of stray light is estimated by least-square fitting the observed profile at disk center with the convolved standard profile of BASS2000\footnote{http://bass2000.obspm.fr/solar\_spect.php} \citep{1973delbouille}. The spectra are then corrected assuming that the influence of stray light is constant over the full field of view, and we show the corrected disk-center sample spectra in Fig.~\ref{prof}. The solar disk center and radius are determined from the reconstructed H$\alpha$ line wing image, using the method proposed by \cite{2015hao}. \section{Results} \label{sect3} \subsection{Line formation} We use the RH code \citep{2001uitenbroek,2015pereira} to calculate the line profiles of the quiet-sun VALC model \citep{1981vernazza} while assuming no turbulent velocity, no line-of-sight velocity and no magnetic field. The H$\alpha$ line is calculated in non-local thermodynamic equilibrium (non-LTE), while the superposed \ion{Si}{I} 6560.58 \AA\ and \ion{Fe}{I} 6569.21 \AA\ lines are treated in LTE, with their data from the Kurucz line database \citep{2018kurucz}. We calculate the line profiles from the solar disk center to the solar limb, with the cosine of heliocentric angle $\mu$ varying from 1.0 to 0.05. The contribution function is defined as $C_I(z)=j_\nu\exp(-\tau_\nu)$, where $j_\nu$ is the emissivity and $\tau_\nu$ is the optical depth, and an integration along the height $z$ gives the value of emergent intensity. In Fig.~\ref{contrib} we show the contribution functions at the line center of the \ion{Si}{I} 6560.58 \AA\ and \ion{Fe}{I} 6569.21 \AA\ lines. Both lines are formed in the optically thick regime, since the contribution function peaks below the $\tau=1$ height. The formation height is defined as the centroid of the contribution function, and is marked with a dashed vertical line. The \ion{Fe}{I} line forms in the mid-photosphere (around 250 km), similar to other \ion{Fe}{I} lines that are sensitive to magnetic fields \citep{2001shchukina,2018hong}. The \ion{Si}{I} line forms much lower than the \ion{Fe}{I} line, almost at the bottom of the photosphere (around 70 km). The low formation height of the \ion{Si}{I} line also justifies the LTE assumption, which saves us from dealing with the complicated Si model atom as in \cite{2008bard}. \begin{figure} \centering \includegraphics[width=0.3\textwidth]{si_height-eps-converted-to.pdf} \caption{Contribution functions at the line center of the \ion{Si}{I} 6560.58 \AA\ and \ion{Fe}{I} 6569.21 \AA\ lines. Black vertical lines marks the formation height, and red vertical lines denotes the height where $\tau=1$. } \label{contrib} \end{figure} \subsection{Center-to-limb variation} \subsubsection{Equivalent width} The equivalent width (EW) of an absorption line is defined as the integration of the normalized line depth profile over wavelength: \begin{equation} W_\lambda=\int R_\lambda\mathrm{d}\lambda=\int \frac{I_c-I_\lambda}{I_c}\mathrm{d}\lambda. \end{equation} For all these three observations, we calculate the values of EW for each pixel on the solar disk. Data pixels at the solar limb where $\mu <0.15$ are discarded since this line becomes relatively weak. We also calculate the EWs from the VALC model for $\mu=1.0$ until $\mu=0.05$. \begin{table*} \caption{Center-to-limb distribution of the EW and the FWHM of the \ion{Si}{I} line from model calculations and observed mean values. Instrumental broadening is included.} \label{width} \centering \begin{tabular}{c|cccc|cccc} \hline \multirow{2}{*}{$\mu$} & \multicolumn{4}{c|}{EW (m\AA)} & \multicolumn{4}{c}{FWHM (m\AA)} \\ & VALC & OBS1 & OBS2 & OBS3 & VALC & OBS1 & OBS2 & OBS3 \\ \hline 1.0 & 13.59 & 14.56 $\pm$ 1.19 & 14.06 $\pm$ 1.94 & 14.16 $\pm$ 1.78 & 134 & 172 $\pm$ 16 & 171 $\pm$ 14 & 176 $\pm$ 13 \\ 0.9 & 14.18 & 14.93 $\pm$ 1.56 & 14.42 $\pm$ 1.84 & 14.71 $\pm$ 1.68 & 134 & 181 $\pm$ 17 & 174 $\pm$ 15 & 180 $\pm$ 13 \\ 0.8 & 14.82 & 15.48 $\pm$ 1.65 & 14.97 $\pm$ 1.89 & 15.30 $\pm$ 1.78 & 134 & 186 $\pm$ 17 & 179 $\pm$ 15 & 181 $\pm$ 14 \\ 0.7 & 15.54 & 16.05 $\pm$ 1.79 & 15.71 $\pm$ 1.83 & 16.04 $\pm$ 1.77 & 134 & 189 $\pm$ 18 & 184 $\pm$ 16 & 186 $\pm$ 14 \\ 0.6 & 16.32 & 16.88 $\pm$ 1.75 & 16.66 $\pm$ 1.88 & 16.86 $\pm$ 1.74 & 134 & 191 $\pm$ 17 & 188 $\pm$ 15 & 188 $\pm$ 14 \\ 0.5 & 17.15 & 17.61 $\pm$ 1.76 & 17.48 $\pm$ 1.89 & 17.59 $\pm$ 1.87 & 134 & 192 $\pm$ 17 & 191 $\pm$ 17 & 190 $\pm$ 14 \\ 0.4 & 18.08 & 18.31 $\pm$ 1.80 & 18.35 $\pm$ 2.04 & 18.47 $\pm$ 1.87 & 134 & 194 $\pm$ 16 & 196 $\pm$ 16 & 193 $\pm$ 15 \\ 0.3 & 18.77 & 18.88 $\pm$ 1.76 & 19.18 $\pm$ 2.07 & 19.34 $\pm$ 1.96 & 133 & 197 $\pm$ 17 & 197 $\pm$ 16 & 198 $\pm$ 15 \\ 0.2 & 19.10 & 19.04 $\pm$ 1.82 & 19.74 $\pm$ 2.15 & 19.74 $\pm$ 2.24 & 133 & 199 $\pm$ 19 & 201 $\pm$ 19 & 201 $\pm$ 17 \\ \hline \end{tabular} \end{table*} \begin{figure*} \centering \includegraphics[width=\textwidth]{wl_dist_all-eps-converted-to.pdf} \caption{Probability density function of the EW of the \ion{Si}{I} line as function of the cosine of the heliocentric angle. Red and blue lines show the averaged value and 1$\sigma$ range, and the green diamonds show the calculated values from the VALC model with no turbulent velocity. } \label{wl} \end{figure*} \begin{figure*} \centering \includegraphics[width=\textwidth]{fwhm_dist_all-eps-converted-to.pdf} \caption{Same as Fig.~\ref{wl}, but for the FWHM of the \ion{Si}{I} line. An instrumental FWHM of 72.6 m\AA\ is included. } \label{fwhm} \end{figure*} \begin{figure*} \centering \includegraphics[width=\textwidth]{si_turb-eps-converted-to.pdf} \caption{Calculated EW and FWHM from the VALC model with different micro- and macroturbulent velocities. An instrumental FWHM of 72.6 m\AA\ is included. } \label{micro} \end{figure*} The probability density functions of the EW as function of $\mu$ in the three observations are shown in Fig.~\ref{wl} as gray shades. The averaged value and 1$\sigma$ range at each $\mu$ are overplotted as red and blue curves, while the calculated values from the VALC model are shown in green diamonds. These values are also listed in Table~\ref{width} for comparison. It is clear that the center-to-limb variation of EW as revealed from the three observations are quite similar, and agrees well with model calculations. Generally speaking, the EW increases from center to limb, and reaches its peak value near $\mu=0.2$, and then begins to decrease. The variation of the EW could be interpreted in the following way. As $\mu$ increases, both the \ion{Si}{I} line and the continuum are formed in higher layers with a lower local temperature, leading to a decrease in both the line and the continuum intensity. A smaller line intensity tends to decrease the EW, while a smaller continuum intensity tends to increase the EW. Since the line and continuum are formed in different heights, the decreasing percentage of their intensity at a certain $\mu$ could vary. The competition of these two factors leads to the final results: when $\mu$ decreases from 1.0 to 0.2, the line intensity decreases more sharply than continuum; while for $\mu<0.2$ the decrease in continuum dominates. The differences of the averaged values from observations and calculated values from models are less than 1 m\AA, which is within the 1$\sigma$ range. The only recorded EW value of this line in literature is from the revised Rowland Table by \citet{1966moore}, which reads 22 m\AA\ at the disk center. We also measure the EW from the Jungfraujoch atlas \citep[BASS2000,][]{1973delbouille} and the Kitt Peak atlas \citep{1984kurucz}, and they give values of 18.52 and 17.27 m\AA\ at the disk center. All these values are far larger than the observed values from CHASE, even outside the 1$\sigma$ range, which is mostly due to the fact that the \ion{Si}{I} line is blended with a telluric absorption line. Although the Kitt Peak atlas has corrected the atmospheric transmission, it seems that their evaluation is still not accurate. \subsubsection{Full width at half maximum} The full width at half maximum (FWHM) of an absorption profile is determined by various line broadening mechanisms, and is close to the FWHM of the line profile in the optically thin regime. For optically thick lines, the FWHMs of the line profile and the absorption profile are not necessarily the same due to the so-called opacity broadening \citep{2015rathore}, while they are still positively related. Here, we measure the FWHM of the line depth profile $R_\lambda$ for each pixel on the solar disk where $\mu\ge 0.15$. Note that the observed profiles are not corrected with a deconvolution of the instrumental profile. Thus, the measured FWHM has included the instrumental FWHM of 72.6 m\AA\ \citep{2022li}. In Fig.~\ref{fwhm} we show the probability density functions of the FWHM as function of $\mu$ in the three observations. Despite of the spikes and sub-structures, there is still an increasing trend for the FWHM from the disk center to the solar limb. However, the calculated FWHM from the VALC model after convolution with the CHASE instrumental profile, as denoted by green diamonds and also listed in Table~\ref{width}, is quite different from the observed values. Without any turbulent velocities, the calculated FWHM decreases towards the solar limb, since for a smaller $\mu$, the line forms higher in the photosphere where the lower temperature narrows the thermal width. For comparison, the measured FWHM is 166 m\AA\ from the Jungfraujoch atlas \citep[BASS2000,][]{1973delbouille} and 158 m\AA\ from the Kitt Peak atlas \citep{1984kurucz}, after convolution with the CHASE instrumental profile. The discrepancy of observed and calculated FWHM values are explained below. \begin{figure} \centering \includegraphics[width=0.3\textwidth]{si_macro2-eps-converted-to.pdf} \caption{Center-to-limb variations of the macroturbulent velocities derived from observations. A fitted curve with $\xi_r=2.85$ km s$^{-1}$ and $\xi_t=3.58$ km s$^{-1}$ is overplotted. } \label{macro} \end{figure} \subsubsection{Micro- and macroturbulent velocities} Traditionally, non-thermal turbulent velocities are introduced to explain the excess width in the absorption profiles and line profiles. A dichotomy of microturbulence and macroturbulence is employed from the geometric scale of the motions. Microturbulence occurs within the mean free path of photons, so the line absorption is changed, resulting in a larger EW and FWHM. However, macroturbulence does not influence the line absorption and only broadens the line profile. We recalculate the EW of the \ion{Si}{I} line assuming different values of microturbulent velocity in the VALC model. The results for the $\mu=0.6$ case are shown in Fig.~\ref{micro}. The value of $\mu=0.6$ is chosen arbitrarily here for illustration, and the increasing trend is similar for other values of $\mu$. One can see that the EW increases by 1.1 m\AA\ for a microturbulent velocity of 4 km s$^{-1}$. Given the large uncertainties in the EW measurement from the observations, it would be unreliable to derive the microturbulent velocities. As stated above, the FWHM of an absorption line is influenced by both micro- and macroturbulent velocities, and Fig.~\ref{micro} provides a schematic view of their contributions as calculated from the VALC model. The line profile has been convolved with the CHASE instrumental profile in order to compare with observations. The macroturbulent velocity is then added by convolving the line depth profile with a Gaussian velocity distribution: \begin{equation} R^\prime=R\ast \frac{1}{\xi_{\mathrm{macro}}\sqrt{\pi}}\mathrm{e}^{-\xi^2/\xi^2_{\mathrm{macro}}}. \end{equation} As shown in Fig.~\ref{micro}, both turbulent velocities could effectively increase the FWHM. However, the increase of the FWHM is not obvious for small macroturbulent velocities, under the spectral resolution of CHASE. In order to separate the contributions to the FWHM from micro- and macroturbulent velocities, we fix the values for microturbulent velocities at different positions on the solar disk. We take the empirical formula of \citet{2022takeda}, which shows an increasing trend of the microturbulent velocities towards the solar limb, with 1 km s$^{-1}$ at the disk center and 1.97 km s$^{-1}$ at $\mu=0.2$. The macroturbulent velocities are then derived from a similar relation as in Fig.~\ref{micro} after inclusion of the microturbulent velocities. The results from the three observations are shown in Fig.~\ref{macro}. There is also an increasing trend towards the solar limb, with 2.80 km s$^{-1}$ at the disk center and 3.52 km s$^{-1}$ at $\mu=0.2$. The center-to-limb variation of the macroturbulent velocity can be interpreted as the intrinsic anisotropy of photospheric motions. If we consider a radial turbulent velocity $\xi_r$ and a tangential turbulent velocity $\xi_t$, then the total macroturbulent velocity would be $\xi_\mathrm{macro}^2=\xi_r^2\mu^2+\xi_t^2(1-\mu^2)$ \citep{2017takeda}. A least-square fit to the averaged values in our observations (Fig.~\ref{macro}) gives $\xi_r=2.85$ km s$^{-1}$ and $\xi_t=3.58$ km s$^{-1}$. These values are slightly larger than previous ones derived from the \ion{Fe}{I} lines \citep{2017takeda,2022sheminova}, since the \ion{Si}{I} line forms deeper in the photosphere where the unresolved turbulent motions are faster, as revealed from observations and simulations \citep{1977gray,1978gray,1995takeda,2012beeck,2017takeda}. \begin{figure} \centering \includegraphics[width=0.5\textwidth]{000si-eps-converted-to.pdf} \caption{Reconstructed maps from the spectra and fitting parameters. (a)--(b) Intensity maps of the \ion{Si}{I} and H$\alpha$ line center. The arrow points to the disk center. The red plus signs denote positions whose line profiles are shown in Fig.~\ref{prof_region} (c)--(d) Line center depth maps of the \ion{Si}{I} and \ion{Fe}{I} line. (e)--(f) Doppler maps of the \ion{Si}{I} and \ion{Fe}{I} line. (g)--(h) Line width maps of the \ion{Si}{I} and \ion{Fe}{I} line. } \label{map} \end{figure} \begin{figure} \centering \includegraphics[width=0.5\textwidth]{000si5-eps-converted-to.pdf} \caption{Line profiles at different locations. (a)--(b) Sample line profiles at selected points in different regions (marked with plus signs in Fig.~\ref{map}). (c)--(d) Line profiles of the flare kernel at selected time. } \label{prof_region} \end{figure} \begin{figure} \centering \includegraphics[width=0.3\textwidth]{000si3-eps-converted-to.pdf} \caption{Histograms of \ion{Si}{I} line-center intensity, EW, and FWHM for different regions. } \label{dist} \end{figure} \subsection{Variations in an active region} The magnetic structure of an active region is apparently different from that of the quiet Sun, leading to a different atmospheric structure and thus a different line profile. We select the active region in the northern hemisphere in OBS1 with a field of view of 300 pixels $\times$ 300 pixels ($\sim$312\arcsec$\times$312\arcsec), where the C2.5 flare is near its peak time. One can clearly identify the sunspot group in the intensity map of \ion{Si}{I}, and the flare kernel in the intensity map of H$\alpha$ (Fig.~\ref{map}(a)--(b)). We fit the \ion{Si}{I} and \ion{Fe}{I} line profiles using a Gaussian function with a slanted background: \begin{equation} I_\lambda=-A_0\exp(-(\lambda-A_1)^2/A_2^2)+A_3+A_4\lambda, \end{equation} where $A_0$ is the line center depth, $A_1$ is the observed line center, and $A_2$ is the line width. The reference line center is chosen as the averaged $A_1$ of the uppermost quiet region in the field of view. The line width $A_2$ is positively related to the FWHM. The reconstructed maps of the fitting parameters are shown in Fig.~\ref{map}, and the line profiles of selected positions are shown in Fig.~\ref{prof_region}(a)--(b). The line center depth of the \ion{Fe}{I} line is generally larger than that of the \ion{Si}{I} line, and the values at the sunspots are smaller than the values at the quiet Sun (Fig.~\ref{map}(c)--(d)). We also show the line profiles of the flare kernel at different time in Fig.~\ref{prof_region}(c)--(d). We do not find any enhancement of these two lines as a response to flare heating, as judged from the variation of line profiles with time. Given the fact that the \ion{Si}{I} line forms in the deep photosphere, either extremely energetic non-thermal particles or local reconnections are expected to heat the photosphere and give rise to its intensity. The Doppler maps, however, show interesting outflows in the sunspot penumbra, with velocities less than 2 km s$^{-1}$ (Fig.~\ref{map}(e)--(f)), known as the Evershed flows. The outflows revealed from the \ion{Si}{I} line is larger than those from the \ion{Fe}{I} line, with a velocity ratio in the range of 1.5--1.9, indicating a decrease in the flow velocities at larger heights, which agrees with previous inversions and simulations \citep{2017siu,2018siu}. The \ion{Si}{I} line width is also larger than the \ion{Fe}{I} line width, which is attributed to both a larger thermal width and a larger turbulent velocity at a lower height. Histograms of the \ion{Si}{I} line-center intensity, EW, and FWHM for different regions are shown in Fig.~\ref{dist}. No obvious difference is found for the EW in different regions. This implies that although the local temperature varies in different regions, the ratio of line intensity to continuum intensity are still in the same range. However, the FWHMs in sunspot areas are generally larger than those in the quiet Sun, which has been observed previously \citep{1966moore}, indicating larger unresolved turbulent velocities in sunspots. \subsection{Diagnosing potentials} Due to the low formation height, the \ion{Si}{I} line is able to unveil the motions and turbulences in the deep photosphere. Combined with other photospheric lines, say, the \ion{Fe}{I} line that is simultaneously observed by CHASE, a full picture of the velocity field in the photosphere could be reconstructed. Thus, it would be more feasible to catch the initial process of flux emergence, as well as the formation of active regions \citep{2021chen} and following activities, such as the Evershed flows inside the magnetic flux tubes \citep{2016murabito,2017siu,2018siu}. In addition, the full-disk velocity map at different atmospheric layers could be used to measure the solar differential rotation, providing new restrictions to the solar dynamo theory \citep{2000beck}. The enhancement of the \ion{Si}{I} line center intensity usually characterizes local heating in the formation height. Whether the deep photosphere could be effectively heated remains to be investigated with further observational evidence. Possible candidates of heating mechanisms include local magnetic reconnections \citep{2001chen,2020song} and extremely energetic particles \citep{2006xu,2018hong,2019kowalski}. Cross-correlations of the \ion{Si}{I} intensity maps could also reveal possible helioseismic waves that could be evidences of local disturbance in the lower photosphere \citep{2011zhao}. \section{Conclusion} \label{sect4} In this paper, we perform statistical analysis of the \ion{Si}{I} 6560.58 \AA\ line observed with CHASE that is free from line blending. The \ion{Si}{I} line is formed at the bottom of the photosphere, which is even lower than the simultaneously observed \ion{Fe}{I} 6569.21 \AA\ line. The measured EW of the \ion{Si}{I} line increases from the disk center until $\mu=0.2$, and then decreases towards the solar limb, reflecting the variation of the decreasing percentages of line intensity and continuum intensity. The theoretical calculation of the center-to-limb variation of the EW from the VALC model generally agrees well with the observations. However, the FWHM shows a monotonically increasing trend from center to limb which is far from model predictions. The discrepancy can be attributed to the unresolved macroscopic turbulent motions in the photosphere. The macroturbulent velocity is derived to be 2.80 km s$^{-1}$ at the disk center, and increases to 3.52 km s$^{-1}$ at $\mu=0.2$. The center-to-limb variation of the macroturbulent velocity indicates the anisotropy of photospheric motions, which requires future observations from other photospheric lines to reconstruct the full physical picture of the photosphere. In our observations, both \ion{Si}{I} and \ion{Fe}{I} lines do not show any response to flare heating. The Doppler maps of these lines show indications of Evershed flows in the sunspot penumbra. The line width and Doppler velocity from the \ion{Si}{I} line are generally larger than those from the \ion{Fe}{I} line, since the lower layers are more turbulent and has a larger temperature. While the FWHMs in sunspot areas are generally larger than in the quiet Sun, the EWs do not show obvious difference. This indicates larger unresolved turbulent velocities in sunspot areas, while the line-to-continuum intensity ratio stays in the same range. As routinely observed by CHASE, these lines provide a promising tool to study the mass flows and turbulence in the different photospheric layers, especially those connected to differential rotation or flux emergence. The deeply formed \ion{Si}{I} line, when combined with other photospheric lines, also has a good potential in the diagnostics of energy transport in the photosphere, such as white-light flares and helioseismic waves. \begin{acknowledgements} We are grateful to the referee for constructive comments. J.H. would like to thank Yikang Wang for fruitful discussions. The CHASE mission is supported by China National Space Administration. This work was supported by National Key R\&D Program of China under grant 2021YFA1600504 and by NSFC under grants 11903020, 11733003, 12127901, and 11873091. \end{acknowledgements} \bibliographystyle{aa}

e2ad9ac309a28b72f9fdeb6c6497550c5bf968a4

\section{Electron effective potential from the geodesic equation} In their paper on Rydberg atom bounds on the cosmological constant \cite{kundu}, Kundu, Pradhan, and Rosenzweig use an effective potential for the electron inferred from an exponentially expanding de Sitter universe. Our aim in this paper is to show that the same potential can be obtained in a more general way from the geodesic equation for motion of the electron, which gives added insight into how bounds obtained from this potential vary between different models for the cosmological constant action. Weinberg in his text``Gravitation and Cosmology'' \cite{wein} gives the geodesic equation for a general spherical metric. For the line element \begin{equation}\label{linelt} ds^2=B(r) dt^2 -A(r) dr^2 -r^2 d\Omega~~~, \end{equation} Eq. (8.4.19) of \cite{wein} gives \begin{equation}\label{geo} \frac{A(r)}{B^2(r)}\left(\frac{dr}{dt}\right)^2 +\frac{J^2}{r^2}-\frac{1}{B(r)}=-E~~~, \end{equation} where $E_{nr}=(1-E)/2$ plays the role of the non-relativistic energy per unit mass. Comparing this with the nonrelativistic energy formula \begin{equation}\label{nrform} \frac{1}{2} \left(\frac{dr}{dt}\right)^2 +\frac{V_{eff}}{m_e} -E_{nr}=0~~~, \end{equation} and doing algebraic rearrangement to eliminate $(dr/dt)^2$ we get a formula for $V_{eff}$, \begin{equation}\label{Veff} \frac{V_{eff}}{m_e}=\frac{B^2(r)}{2A(r)} \frac{J^2}{r^2} + \frac{B(r)}{2A(r)}[B(r)-1]- E_{nr}\left[\frac{B^2(r)}{A(r)}-1\right]~~~. \end{equation} Since $B(r)$ and $A(r)$ are unity up to small corrections of order the electostatic and gravitational potentials and the cosmological constant, the cosmological constant contribution to the effective potential is given by \begin{equation}\label{Veffcosm} V_{eff}\simeq \frac{m_e}{2} [B(r)_{cosm}-1]~~~. \end{equation} As summarized in Adler \cite{adler}, one can write for a central mass $M$ in geometrized units, \begin{align}\label{paramdef1} A(r)=&1+2M/r-C_A \Lambda r^2 +D_A \Lambda M r+...,\cr B(r)=&1-2M/r-C_B \Lambda r^2 +D_B \Lambda M r+...,\cr \end{align} with the parameters given in Table I for a standard dark energy action and for a Weyl scaling invariant \cite{adlerrev} dark energy action. \begin{table} [ht] \caption{Parameters $C_A,\,C_B,\,D_A,\,D_B$ for the spherically symmetric line element arising from the conventional and the Weyl scaling invariant dark energy actions.} \centering \begin{tabular}{c c c c c c} \hline\hline dark energy type&Equations& $~~C_A~~$ & $~~C_B~~$ &$~~D_A~~$ & $~~D_B~~$ \\ \hline ~~~~~conventional ~~~~~~~~~~& (2),(6) & -1/3 & 1/3 & 4/3 & 0 \\ Weyl scaling invariant & (1), (3)--(5) & 1 & 1 & -10 & -14 \\ \hline\hline \end{tabular} \label{tab1} \end{table} So the general formula for the cosmological constant contribution to $V_{eff}$ is \begin{equation}\label{gen} V_{eff}=-\frac{m_e \,C_B}{2} \Lambda r^2 \end{equation} which for the standard dark energy action gives $V_{eff}=-(1/6) m_e \Lambda r^2$, in agreement with Eq. (7) of \cite{kundu}. For a Weyl scaling invariant dark energy action, one instead gets $V_{eff}=-(1/2) m_e \Lambda r^2$ according to Table I. Thus qualitatively, there is no distinction between the potentials arising in the two cases, as well as for other models of dark energy that take the form of a gravitational action, which will each have a characteristic value of $C_B$. One might ask what happens if the center for measuring $\vec r$ is not taken as the center of the atom, but a displacement $\vec R$ from the atomic center. Then $r^2$ in Eq. \eqref{gen} is replaced by $\vec R^{\,2} + 2 \vec R \cdot \vec r +r^2$. The term $\vec R^{\,2} $ is a constant and does not contribute to energy level differences in Rydberg atoms, and the term $2 \vec R \cdot \vec r$ will average to zero over any inversion invariant squared atomic wave function. So the calculation of energy differences is invariant with respect to the choice of $\vec R$. \section{Perturbation theory of Rydberg atoms} Given $V_{eff}$, the change in the energy level of a Rydberg atom can be calculated from first order perturbation theory, \begin{equation}\label{pert} \Delta E_{cosm}= \langle \Psi| V_{eff} |\Psi \rangle = -\frac{m_e \,C_B}{2} \Lambda \langle \Psi| r^2 |\Psi \rangle \simeq -\frac{m_e \,C_B}{2} \Lambda R^2~~~, \end{equation} where $R$ is the radius of the Rydberg atom orbit. Thus, from an upper bound $|\Delta E_{cosm}|<U$ we get a bound on the cosmological constant \begin{equation}\label{bound} \Lambda < \frac{2U}{m_e C_B R^2}~~~. \end{equation} If $U$ is a bound on an energy level difference, $R^2$ in Eq. \eqref{bound} will be an orbit radius-squared difference. In analogy with the formulas of Eqs. (12)-(15) of \cite{kundu} this can be turned into a quantitative bound on $\Lambda$. As Kundu et al. have noted, the energy scale of their bound is much smaller than the energy scale of the standard model, already implying very substantial cancellations if the cosmological constant is interpreted as a vacuum energy. But as shown above using a geodesic equation derivation, a very similar bound is obtained in a dark energy model \cite{adlerrev} in which the cosmological constant does not arise as a vacuum energy. Thus, although the bounds set by Rydberg atoms are interesting, they do not reveal whether the gravitational vacuum gravitates. As long as $C_B$ is of order unity, any dark energy action coupled to the gravitational metric gives a bound similar in magnitude to that given by the conventional cosmological constant action. \section{Acknowledgement} I wish to thank Suman Kumar Kundu for calling my attention to \cite{kundu} and for an informative email correspondence.

af45e9a8ab227f054adfbae3763798bd37e0878b

\section{Introduction} Two-body baryonic decays of the charmonium states $J/\psi$ and $\psi(3686)$, here both denoted by the symbol $\Psi$, provide an excellent laboratory to study flavour-SU(3) symmetry breaking and test various aspects of quantum chromodynamics (QCD) in the perturbative energy regime~\cite{pQCD}. The amplitudes of $\Psi$ decays to different baryon octet pairs are supposed to be the same under the assumption of flavour-SU(3) symmetry. However, branching fractions are not only determined by strong interaction amplitudes, but also by electromagnetic interactions and interference between the two amplitudes~\cite{Rudaz:1975yu}, although these contributions are much smaller than the expected flavour-SU(3) breaking effects. With a phenomenologically plausible model~\cite{Zhu:2015bha,Ferroli:2019nex, Ferroli:2020mra}, the branching fractions of $\Psi$ decay to baryon octet final states can be described well. Perturbative QCD~\cite{Appelquist:1974zd,DeRujula:1974rkb} predicts the partial widths for $\psi(3686)$ decay into an exclusive hadronic state to be proportional to squares of the wave-function, which are well determined from leptonic widths. Furthermore, the ratio between the branching fractions of $J/\psi$ and $\psi(3686)$ decays to the same final states is expected to obey the so-called ``$12\%$ rule''~\cite{Appelquist:1974zd,DeRujula:1974rkb}. Although a large fraction of exclusive decay channels follow this rule approximately, significant violation has been observed in the $\rho \pi$ channel~\cite{Franklin:1983ve}. The ratio of the branching fraction $\mathcal{B}(\psi(3686) \to \rho \pi)$ to $\mathcal{B}(J/\psi \to \rho \pi)$ is much smaller than the perturbative QCD prediction, and this is called the ``$\rho\pi$ puzzle''. Many explanations~\cite{Ref:review2} of the $\rho\pi$ puzzle have been proposed, including the $J/\psi$-glueball admixture scheme~\cite{Li:2007ky}, the intrinsic-charm-component scheme~\cite{Brodsky:1997fj}, the sequential-fragmentation model~\cite{Karl:1984en}, the exponential form-factor model~\cite{Chaichian:1988kn}, the $S$-$D$ wave-mixing scheme~\cite{Liu:2004un,Rosner:2001nm}, the final-state interaction scheme and others~\cite{Li:1996yn}. However, none of these explanations can account for all existing experimental results. Tests of the $12\%$ rule using the baryonic decay modes are helpful in understanding the $\rho \pi$ puzzle. Experimentally, the branching fractions of $\Psi$ decay into octet baryon pairs have been well measured, except for $\Sigma^-\bar{\Sigma}^+$\cite{pdg}. The angular distribution of a baryon pair can be written as $1 + \alpha_{B}\cos^2\theta_{B}$, where $\alpha_B$ is the angular distribution parameter of the baryon, $\theta_{B}$ is a polar angle between the baryon and the positron beam in the centre-of-mass (c.m.) system. The value of $\alpha_{B}$ is expected to be 1 due to the helicity conservation rule~\cite{Brodsky:1981kj}. In addition, in the theoretical calculations of $\alpha_{B}$, the masses of quarks and baryons have been considered~\cite{Claudson:1981fj, Carimalo:1985mw}. Existing theoretical predictions are not consistent with the experimental measurements. The values of $\alpha_{B}$ should be the same among isospin partners, such as $\alpha_{\Sigma^{+}}$ and $\alpha_{\Sigma^{0}}$~\cite{BESIII:2020fqg, BESIII:2017kqw}, $\alpha_{\Xi^{0}}$ and $\alpha_{\Xi^{-}}$~\cite{BESIII:2016nix, BESIII:2016ssr}. There are no significant differences observed experimentally. However, the value of $\alpha_{\Sigma^{-}}$ has not yet been measured. In this paper, the first observation of the decay $\psi(3686)\to\Sigma^-\bar\Sigma^+$ is reported, where $\Sigma^-$ decays to $n\pi^-$ and $\bar\Sigma^+$ decays to $\bar n\pi^+$. The data samples used in this analysis consist of $(448.1\pm2.9)\times10^6$ $\psi(3686)$ events~\cite{npsp} collected with the BESIII detector. \section{The BESIII detector and Monte Carlo simulation} The BESIII detector~\cite{Ablikim:2009aa} records symmetric $e^+e^-$ collisions provided by the BEPCII storage ring~\cite{Yu:IPAC2016-TUYA01} in the c.m. energy range from 2.0 to 4.94~GeV, with a peak luminosity of $1\times10^{33}$~cm$^{-2}$s$^{-1}$ achieved at $\sqrt{s}=3.77$ GeV. BESIII has collected large data samples in this energy region~\cite{Ablikim:2019hff}. The cylindrical core of the BESIII detector covers 93\% of the full solid angle and consists of a helium-based multilayer drift chamber~(MDC), a plastic scintillator time-of-flight system~(TOF), and a CsI(Tl) electromagnetic calorimeter~(EMC), which are all enclosed in a superconducting solenoidal magnet providing a 1.0~T (0.9~T in 2012) magnetic field. The solenoid is supported by an octagonal flux-return yoke with resistive plate counter muon identification modules interleaved with steel. The charged-particle momentum resolution at $1~{\rm GeV}/c$ is $0.5\%$, and the d$E$/d$x$ resolution is $6\%$ for electrons from Bhabha scattering. The EMC measures photon energies with a resolution of $2.5\%$ ($5\%$) at $1$~GeV in the barrel (end-cap) region. The time resolution in the TOF barrel region is 68~ps, while that in the end-cap region is 110~ps. Monte Carlo~(MC) simulated events are used to determine the detection efficiency, optimize selection criteria, and study possible backgrounds. Simulated data samples produced with a {\sc geant4}-based~\cite{geant4} package, which includes the geometric description of the BESIII detector and the detector response, are used to determine detection efficiencies and to estimate backgrounds. The simulation models the beam-energy spread and initial-state radiation (ISR) in the $e^+e^-$ annihilations with the generator {\sc kkmc}~\cite{KKMC}. The inclusive MC sample $\psi(3686)$ includes the production of the $\psi(3686)$ resonance, the ISR production of the $J/\psi$, and the continuum processes incorporated in {\sc kkmc}. The known decay modes are modeled with BesEvtGen~\cite{evtgen} using branching fractions taken from the Particle Data Group (PDG)~\cite{pdg}, and the remaining unknown charmonium decays are modelled with {\sc lundcharm}~\cite{lundcharm}. Final-state radiation from charged final-state particles is incorporated using the {\sc photos} package~\cite{photos}. The differential cross section of the signal process~($\psi(3686)\to\Sigma^-\bar\Sigma^+$, $\Sigma^-\to n \pi^-$, $\bar\Sigma^+\to \bar n \pi^+$) is expressed with respect to five observables $\boldsymbol{\xi}= ( \theta_{\Sigma^{-}}, \theta_{n}, \phi_{n}, \theta_{\overline{n}}, \phi_{\overline{n}})$, which includes four parameters $\alpha_{\Sigma^{-}}$, $\Delta\Phi$, $\alpha_{-}$ and $\alpha_{+}$~\cite{Faldt:2017kgy}. Here, $\theta_{\Sigma^{-}}$ is the polar angle between the $\Sigma^{-}$ and the positron beam in the reaction c.m. frame, $\theta_{n}, \phi_{n}$ and $\theta_{\overline{n}}, \phi_{\overline{n}}$ are the polar and azimuthal angles of the neutron and anti-neutron measured in the rest frames of their corresponding parent particles. The value of $\alpha_{\Sigma^{-}}$ is determined in this analysis, and $\Delta\Phi$ is set to be 0 by assuming no polarization. The decay asymmetry parameters $\alpha_{-}$ and $\alpha_{+}$ in the differential cross sections are fixed to $-0.068$ and $0.068$ using the PDG~\cite{pdg} values, where $\alpha_{-}$ and $\alpha_{+}$ are used to describe the non-leptonic decays of $\Sigma^-\to n\pi^-$ and $\bar\Sigma^+\to \bar n\pi^+$~\cite{Lee:1957qs}. The uncertainties in the values of these parameters are considered when assigning systematic uncertainties. \section{Event selection} The final state of the signal process is $n\pi^-\bar{n}\pi^+$. Event candidates are required to have two well-reconstructed charged tracks with zero net charge, and one anti-neutron. In order to keep the selection efficiency high there is no attempt made to reconstruct the neutron. Charged tracks detected in the MDC are required to be within a polar angle ($\theta$) range of $|\!\cos\theta|<0.93$ and $\theta$ is defined with respect to the $z$ axis, which is along the symmetry axis of the MDC. For each charged track, the distance of closest approach to the interaction point (IP) must be less than 30\,cm along the $z$ axis, and less than 10\,cm in the transverse plane. Particle identification~(PID) for charged tracks combines measurements of the energy deposited in the MDC~(d$E$/d$x$) and the flight time in the TOF to form likelihoods $\mathcal{L}(h)~(h=p,K,\pi)$ for each hadron $h$ hypothesis. Two pions are identified with the requirements that $\mathcal{L}(\pi)>\mathcal{L}(K)$ and $\mathcal{L}(\pi)>\mathcal{L}(p)$. The anti-neutron candidates are identified using showers in the EMC. The deposited energy of each shower must be more than 600~MeV both in the barrel region ($|\!\cos \theta|< 0.80$) and in the end-cap region ($0.86 <|\!\cos \theta|< 0.92$). To exclude showers that originate from charged tracks, the angle subtended by the EMC shower and the position of the closest charged track at the EMC must be greater than 10 degrees as measured from the IP. The second moment $\sum_i E_i r_i^2 /\sum_i E_i$, where $E_i$ is the energy deposition in the $i^{\rm th}$ crystal and $r_i$ is the radial distance of the $i^{\rm th}$ crystal from the cluster centre, is required to be larger than 20, to suppress the photon background misidentified as anti-neutrons. To suppress electronic noise and showers unrelated to the event, the difference between the EMC time and the event start time is required to be within [0, 700]\,ns. If the number of anti-neutron candidates in an event is more than one, the maximum deposited energy in the EMC is selected as the final event candidate. A one-constraint kinematic fit is performed to the decay $\psi(3686)\to n\pi^-\bar{n}\pi^+$ with the constraints provided by four-momentum conservation, the invariant mass $M_{\bar{n} \pi^{+}}$ equal to the known $\bar{\Sigma}^{+}$ mass. Since the anti-neutron could annihilate with the materials in the EMC, the polar and azimuthal angles of anti-neutron are used in kinematic fit, while the deposited energy of anti-neutron are set free. Considering neutron is hardly to be detected, the three-momentum of neutron is free parameters in the kinematic fit.The $\chi^2$ of the kinematic fit is required to be smaller than 50, which is a value optimized by using the figure-of-merit $S/\sqrt{S+B}$, where $S$ is the number of signal MC events and $B$ is the number of the estimated background events. To suppress background from $\psi(3686) \rightarrow \pi^+ \pi^- J/\psi$ decays, with $J/\psi \rightarrow n \bar{n}$, the recoil mass of the $\pi^{+} \pi^{-}$ pair is required to be less than 2.9 GeV/$c^{2}$. An inclusive MC sample of 506 million $\psi(3686)$ events is used to study possible background channels, with a generic event-type analysis tool, TopoAna~\cite{ref:topo}. The potential sources of peaking background are found to be $\psi(3686)\to\gamma\chi_{cJ}(\chi_{cJ}\to\Sigma^-\bar\Sigma^+)~(J=0,1,2)$ and $\psi(3686)\to\gamma\eta_c(\eta_c\to\Sigma^-\bar\Sigma^+)$, and $\psi(3686)\to\pi^0\Sigma^-\bar\Sigma^+$. To estimate the sizes and distributions of these background processes, samples of 100 million events are generated for each channel. In these, the decay processes $\psi(3686)\to\gamma\chi_{cJ}(\chi_{cJ}\to\Sigma^-\bar\Sigma^+)$, $\psi(3686)\to\gamma\eta_c(\eta_c\to\Sigma^-\bar\Sigma^+)$, and $\psi(3686)\to\pi^0\Sigma^-\bar\Sigma^+$ are generated with the P2GCJ (J=0,1,2), JPE, and phase-space models. When accounting for the branching fractions~\cite{pdg} and detection efficiencies of these decays, the numbers of background events passing the selection in the data sample are predicted to be $562\pm86$ for $\psi(3686)\to\gamma\chi_{cJ}(\chi_{cJ}\to\Sigma^-\bar\Sigma^+)$ and $5\pm1$ for $\psi(3686)\to\gamma\eta_c(\eta_c\to\Sigma^-\bar\Sigma^+)$. The contribution of $\psi(3686)\to\pi^0\Sigma^-\bar\Sigma^+$ decays is negligible. An off-resonance data sample taken at the c.m. energy of 3.65 GeV is used to estimate the non-$\psi(3686)$ background. The size of this contribution, $N_{\mathrm{non-}\psi(3686)}$, is determined according to the formula: $N_{\mathrm{non-}\psi(3686)}=N_{\mathrm{cont}}^{\mathrm{obs}}\cdot\frac{L_{\psi(3686)}}{L_{\mathrm{cont}}}\cdot\frac{s_{\mathrm{cont}}}{s_{\mathrm{\psi(3686)}}}\cdot\frac{\varepsilon_{\mathrm{cont}}}{\varepsilon_{\mathrm{\psi(3686)}}}=92\pm53$ events, where $N_{\mathrm{cont}}^{\mathrm{obs}}=7\pm4$ is the number of events surviving the same selection criteria when applied to the 3.65 GeV sample, $s_{\mathrm{cont}}$ and $s_{\mathrm{\psi(3686)}}$ are the squares of the c.m. energies at 3.65 GeV and 3.686 GeV, $\varepsilon_{\mathrm{cont}}=5.96\%$ and $\varepsilon_{\mathrm{\psi(3686)}}=5.26\%$ are the selection efficiencies at 3.65 GeV and 3.686 GeV, and $L_{\mathrm{cont}} = 44\,{\rm pb}^{-1}$ and $L_{\mathrm{\psi(3686)}}=668.55\,{\rm pb}^{-1}$ and are the integrated luminosities at 3.65 GeV and 3.686 GeV, respectively. \section{Measurement of the branching fraction} The $\Sigma^-$ candidate is reconstructed from the $\pi^-$ and the missing neutron. To determine the number of signal events, an unbinned maximum likelihood fit is performed to the distribution of the invariant mass $n\pi^-$ ($M_{n\pi^{-}}$) in the region of [1.15, 1.25] GeV/$c^{2}$. The signal is described by the shape found in the MC simulations, convoluted with a Gaussian function which accommodates any difference in mass resolution between data and MC simulations. The peaking background is described with the shapes of the MC-simulated exclusive background channels, and the corresponding numbers of events are fixed to the estimated values. The non-peaking background is described with a second-order polynomial function. Figure~\ref{fig:fitting_result} shows the fit of the $n\pi^{-}$ mass distribution. The $\chi^2/{\rm ndf}$ of the fit is 35.64/44, where $\rm ndf$ is the number of degrees of freedom. The branching fraction is calculated according to \begin{equation} \begin{aligned} &Br=\frac{N_{\mathrm{sig}} - N_{\mathrm{non-}\psi(3686)}}{\varepsilon\times \prod{Br_i}\times N_{\mathrm{tot}}} , \\ \end{aligned} \end{equation} where $N_{\mathrm{sig}}$ is the number of signal events and $N_{\mathrm{non-}\psi(3686)}$ is the number of non-$\psi(3686)$ events, $\prod{Br_i}$ is the product of the branching fractions of the intermediate states, and $N_{\mathrm{tot}}$ is the total number of $\psi(3686)$ events~\cite{npsp}. The detection efficiency $\varepsilon$ is estimated from the signal MC simulations. Differences in detection efficiency between data and MC simulations is accounted for using control samples of $J/\psi\to p\bar{p}\pi^+\pi^-$, $J/\psi\to p\bar{n}\pi^-$, and $\psi(3686)\to p\bar{n}\pi^-$ decays. Here, in order to study the difference of anti-neutron efficiency from EMC and kinematic fit, the anti-neutron efficiency ratios between data and MC simulations are determined with different anti-neutron momentum and polar-angle regions. Besides, the efficiency difference, the polar and azimuth angles of anti-neutron, and their error matrices have been corrected based on the data-driven method~\cite{Liu:2021rrx}. The $\pi^+$ and $\pi^-$ efficiency ratios are also determined using the same method. The branching fraction is calculated to be $(2.84\pm 0.04) \times10^{-4}$, where the uncertainty is statistical only. The corresponding number of signal events, non-$\psi(3686)$ events, detection efficiency after correcting for data-MC differences, and branching fraction are listed in Tab.~\ref{our br}. \begin{figure}[hbtp] \centering \includegraphics[trim=0 50 0 50,clip,width=0.8\textwidth]{fitting_data_v4.eps} \caption{ The distribution of $M_{n\pi^-}$ for the signal process $\psi(3686)\to\Sigma^-\bar\Sigma^+$, $\Sigma^-\to n \pi^-$, $\bar\Sigma^+\to \bar n \pi^+$. The black dots with error bars are the data, the red solid line is the total fit function, the red dashed line is the signal function, the blue dotted line is the non-peaking background function, and the green dash-dotted line is the peaking background function. } \label{fig:fitting_result} \end{figure} \begin{table}[hbtp] \centering \caption{The number of signal events, the number of non-$\psi(3686)$ background events and the detection efficiency. } \label{our br} \begin{tabular}{ccccc} \hline \hline Channel &$N_{\mathrm{sig}}$ &$N_{\mathrm{non-}\psi(3686)}$ &$\varepsilon$ \\\hline $\psi(3686)\to\Sigma^-\bar\Sigma^+$&$6765 \pm 106$ & $92\pm53$ &$5.26\%$ \\ \hline \hline \end{tabular} \end{table} \section{Measurement of the angular distribution parameter} To determine the value of the angular parameter $\alpha_{\Sigma^-}$, a least-squares fit is performed to the $\cos\theta_{\Sigma^-}$ distribution in the range of [$-$1, 1]. The numbers of signal events are determined in ten equally sized intervals of $\cos\theta_{\Sigma^-}$ with the same method as used in the branching fraction measurement. The detection efficiency in each interval is determined with MC simulations, which is then corrected to account for data-MC differences. The $\cos\theta_{\Sigma^-}$ distribution after efficiency correction is shown in Fig.~\ref{fig:angular distribution}. Superimposed is the result of a fit to the function $1 + \alpha_{\Sigma^-} \cos^2\theta_{\Sigma^-}$. The parameter $\alpha_{\Sigma^-}$ is measured to be $0.96\pm0.09$, where the uncertainty is statistical, and its lower limit is determined to be larger than 0.835 at 90\% confidence level. The $\chi^2/{\rm ndf}$ of the fit is $13.52/8$, where $\rm ndf$ is the number of degrees of freedom. \begin{figure}[hbtp] \centering \includegraphics[trim=0 50 0 50,clip,width=0.8\textwidth]{thetadis.eps} \caption{ The angular distribution for the signal process $\psi(3686)\to\Sigma^-\bar\Sigma^+$, $\Sigma^-\to n \pi^-$, $\bar\Sigma^+\to \bar n \pi^+$. The black dots with error bars indicate the signal yields after efficiency correction, and the red curve represents the fit function. } \label{fig:angular distribution} \end{figure} \section{Systematic uncertainties} To estimate the systematic uncertainties in the measurement of the branching fraction, we consider the differences of the detection efficiency and resolution between data and MC simulations, the uncertainty associated with the generator models, the background estimations and other sources. An overview of all the systematic uncertainties on the branching fraction measurement is given in Tab.~\ref{table6}. \begin{table}[htbp] \centering \caption{Systematic uncertainties on the branching fraction measurement (\%). } \begin{tabular}{lc}\hline\hline \multicolumn{1}{c}{ Source } & \multicolumn{1}{c}{ Uncertainty~(\%)} \\ \hline MC efficiency correction & 1.9 \\ Decay parameter & 1.2 \\ QED peaking-background estimation & 0.8 \\ Non-peaking background estimation & 0.4 \\ Peaking-background estimation & 0.2 \\ Kinematic fitting & 0.2 \\ Total number of $\psi(3686)$ & 0.7 \\ \hline Total & 2.5 \\\hline \hline \end{tabular} \label{table6} \end{table} The tracking and PID efficiency in MC simulations is corrected in bins of transverse momentum and polar angle to agree with that measured in data. The uncertainty on these corrections, derived from the control channels and averaged over bins, is assigned as a systematic uncertainty on the branching fraction. For charged tracks the study is performed with a control sample of $J/\psi\to p\bar p \pi^+\pi^-$ events, and the relative uncertainty is found to be 0.2\%. The relative uncertainty for the reconstruction of the anti-neutron is set with control samples of $J/\psi\to p\bar{n}\pi^-$ and $\psi(3686)\to p\bar{n}\pi^-$ events and found to be 1.9\%. Hence, the total systematic uncertainty associated with the MC efficiency correction is 1.9\%. In the signal generator model~\cite{Faldt:2017kgy}, the values of $\alpha_{-}$ and $\alpha_{+}$ are set to be $-0.068$ and $0.068$ for $\Sigma^{-}$ and $\bar{\Sigma}^{+}$ respectively. Furthermore, we assume that there is no polarization by setting $\Delta\Phi$ to 0. To evaluate the systematic uncertainty associated with these assumptions, we vary $\alpha_{-}$ and $\alpha_{+}$ by one standard deviation (0.008), and change $\Delta\Phi$ to be $-\pi$ or $+\pi$. We compare the efficiencies after these variations with the baseline efficiency, and take the maximum difference, 1.2\%, as the corresponding systematic uncertainty. Possible systematic effects due to the requirement of $M_{\rm rec}(\pi^+\pi^-) < 2.9$ GeV/$c^{2}$ are investigated by varying the selection criteria between 2.80 and 2.91 GeV/$c^{2}$ in steps of 1 MeV/$c^{2}$. The variations observed are compatible with statistical fluctuations and thus no uncertainty is assigned associated with this requirement~\cite{Barlow:2002yb}. The uncertainty associated with non-peaking background is estimated by changing the order of polynomial function used to describe this background. The difference of 0.4\% with respect to the baseline configuration is taken as the systematic uncertainty arising from this source. The uncertainties associated with the number of non-$\psi(3686)$ background events and the peaking background are assigned by varying the sizes of these backgrounds by one standard deviation, giving contributions of $0.8\%$ and $0.2\%$, respectively. To estimate the size of any potential bias arising from the kinematic fit, we obtain the $\chi^2$ distributions with the track correction method for the helix parameters that are corrected to reduce the differences between data and MC simulations~\cite{BESIII:2012mpj}. Besides, the polar and azimuth angles and error matrix of anti-neutron in kinematic fit have also been corrected~\cite{Liu:2021rrx}. Compared with the baseline value, the difference of 0.2\% is taken as the systematic uncertainty. An uncertainty of 0.7\% is assigned to reflect the knowledge of the number of $\psi(3686)$ events in the sample, which is measured from inclusive hadronic decays, as described in Ref.~\cite{npsp}. The main sources of systematic uncertainty on baryonic angular distribution measurement are associated with knowledge of the signal yields, the efficiency correction, and the fitting process. An overview of all the systematic uncertainties is given in Tab.~\ref{table7}. \begin{table}[htbp] \centering \caption{Systematic uncertainties of angular-distribution measurement (\%). } \begin{tabular}{lc}\hline\hline \multicolumn{1}{c}{ Source } & \multicolumn{1}{c}{ Uncertainty(\%)} \\ \hline MC efficiency correction & 0.5 \\ QED peaking background estimation & negligible \\ Non-peaking background estimation & 1.4 \\ Peaking background estimation & 1.6 \\ Kinematic fitting & 0.3 \\ Number of bins & 0.4 \\ Fitting $\cos\theta_{\Sigma^{-}}$ range & 1.9 \\ \hline Total & 2.9 \\\hline \hline \end{tabular} \label{table7} \end{table} In the angular distribution measurement, the number of signal events in each bin is determined by the same method as for the branching fraction measurement. The uncertainties on this yield determination are associated with the MC efficiency correction, background estimation and kinematic-fitting requirement. These uncertainties are estimated with the same method as for the branching fraction. In doing this, we consider the correlations between the measurements in each bin. We then re-perform the fit to the angular distribution and take the difference with respect to baseline value as the systematic uncertainty for each contribution. The uncertainties associated with the $\alpha_{\Sigma^{-}}$ fit itself are estimated by varying the fitting range in $\cos\theta_{\Sigma^{-}}$ from [$-1.0$, 1.0] to [$-0.8$, 0.8], and also changing the number of bins from ten to eight. In both cases the changes in result are assigned as contributions to the uncertainties. The total systematic uncertainty on $\alpha_{\Sigma^{-}}$ is 0.029. \section{Summary} In summary, based on the $(448.1\pm2.9)\times 10^{6}$ $\psi(3686)$ events collected at BESIII detector, the branching fraction and angular parameter, $\alpha_{\Sigma^-} $, of $\psi(3686) \to\Sigma^-\bar\Sigma^+$ decays are measured for the first time. The measurements yield $(2.84 \pm 0.04_{\rm stat.} \pm 0.08_{\rm syst.})\times10^{-4}$ for the branching fraction and $\alpha_{\Sigma^-} = 0.96 \pm 0.09_{\rm stat.} \pm 0.03_{\rm syst.}$. Table~\ref{table:summary} summarizes measurements of the angular parameter and branching fraction for $\psi(3686)\to\Sigma^+\bar\Sigma^-$, $\Sigma^0\bar\Sigma^0$, and $\Sigma^-\bar\Sigma^+$ channels, and predicted values for the branching fractions. The measured branching fraction is around 2.4$\sigma$ above the theoretical prediction value $(2.46 \pm 0.13)\times10^{-4}$~\cite{Ferroli:2020mra}. There are significant differences between the value of $\alpha_{\Sigma^-}$ and those of its isospin partners $\alpha_{\Sigma^{+}}$ and $\alpha_{\Sigma^{0}}$ , which are worthy of further investigation. Finally, it is noted that the analysis method pursued here can also be used to measure the branching fraction of $J/\psi\to\Sigma^-\bar\Sigma^+$, which in combination with the result reported in this paper will provide an opportunity to further test the ``$12\%$ rule'' in charmonium decays. \begin{table}[hbtp] \centering \caption{Summary of the measured angular parameters and branching fractions of $\psi(3686)\to\Sigma^+\bar\Sigma^-$, $\Sigma^0\bar\Sigma^0$, and $\Sigma^-\bar\Sigma^+$, together with theoretical predictions of the branching fractions.} \label{table:summary} \small \begin{tabular}{cccc} \hline \hline Decay mode&Br($\times 10^{-4}$)&Angular parameter $\alpha_{B}$&Br prediction($\times 10^{-4}$)~\cite{Ferroli:2020mra}\\\hline $\psi(3686)\to\Sigma^+\bar\Sigma^-$&$2.43\pm 0.03 \pm 0.05$~\cite{BESIII:2021wkr}&$0.682 \pm 0.030 \pm 0.011$~\cite{BESIII:2020fqg}&$2.29 \pm 0.15$\\\hline $\psi(3686)\to\Sigma^0\bar\Sigma^0$&$2.44 \pm 0.03 \pm 0.11 $~\cite{BESIII:2017kqw}&$0.71 \pm 0.11 \pm 0.04 $~\cite{BESIII:2017kqw}&$2.37 \pm 0.09$ \\\hline $\psi(3686)\to\Sigma^-\bar\Sigma^+$&$2.84 \pm 0.04 \pm 0.08$&$0.96 \pm 0.09 \pm 0.03$&$2.46 \pm 0.13$\\ \hline \hline \end{tabular} \end{table} \acknowledgments The BESIII collaboration thanks the staff of BEPCII and the IHEP computing center for their strong support. This work is supported in part by National Key R\&D Program of China under Contracts Nos. 2020YFA0406300, 2020YFA0406400; National Natural Science Foundation of China (NSFC) under Contracts Nos. 11635010, 11735014, 11835012, 11935015, 11935016, 11935018, 11961141012, 12022510, 12025502, 12035009, 12035013, 12192260, 12192261, 12192262, 12192263, 12192264, 12192265; the Chinese Academy of Sciences (CAS) Large-Scale Scientific Facility Program; Joint Large-Scale Scientific Facility Funds of the NSFC and CAS under Contract No. U1832207; the CAS Center for Excellence in Particle Physics (CCEPP); 100 Talents Program of CAS; The Institute of Nuclear and Particle Physics (INPAC) and Shanghai Key Laboratory for Particle Physics and Cosmology; Sponsored by Shanghai Pujiang Program(20PJ1401700); ERC under Contract No. 758462; European Union's Horizon 2020 research and innovation programme under Marie Sklodowska-Curie grant agreement under Contract No. 894790; German Research Foundation DFG under Contracts Nos. 443159800, Collaborative Research Center CRC 1044, GRK 2149; Istituto Nazionale di Fisica Nucleare, Italy; Ministry of Development of Turkey under Contract No. DPT2006K-120470; National Science and Technology fund; National Science Research and Innovation Fund (NSRF) via the Program Management Unit for Human Resources \& Institutional Development, Research and Innovation under Contract No. B16F640076; STFC (United Kingdom); Suranaree University of Technology (SUT), Thailand Science Research and Innovation (TSRI), and National Science Research and Innovation Fund (NSRF) under Contract No. 160355; The Royal Society, UK under Contracts Nos. DH140054, DH160214; The Swedish Research Council; U. S. Department of Energy under Contract No. DE-FG02-05ER41374.

9473166da190d701237040b68fa768184ed35cd6

\section{Introduction} Recent research efforts on conditional generative modeling, such as Imagen~\citep{saharia2022photorealistic}, DALL$\cdot$E~2~\citep{ramesh2022hierarchical}, and Parti~\citep{yu2022scaling}, have advanced text-to-image generation to an unprecedented level, producing accurate, diverse, and even creative images from text prompts. These models leverage paired image-text data at Web scale (with hundreds of millions of training examples), and powerful backbone generative models, \textit{i.e.}, autoregressive models~\citep{van2017neural,ramesh2021zero,yu2022scaling}, diffusion models~\citep{ho2020denoising,dhariwal2021diffusion}, \textit{etc.}, and generate highly realistic images. Studying these models' generation results, we discovered their outputs are surprisingly sensitive to the frequency of the entities (or objects) in the text prompts. In particular, when generating text prompts about frequent entities (see Appendix~\ref{frequent}), these models often generate realistic images, with faithful grounding to the entities' visual appearance. However, when generating from text prompts with less frequent entities, those models either hallucinate non-existent entities, or output related frequent entities (see~\autoref{fig:comparison}), failing to establish a connection between the generated image and the visual appearance of the mentioned entity. This key limitation can greatly harm the trustworthiness of text-to-image models in real-world applications and even raise ethical concerns. \begin{figure}[!thb] \centering \includegraphics[width=1.0\linewidth]{comparison.001.jpeg} \caption{{Comparison of images generated by Imagen and {Re-Imagen}\xspace on less frequent entities}. We observe that Imagen hallucinates the entities while {Re-Imagen}\xspace maintains better faithfulness.} \label{fig:comparison} \vspace{-2ex} \end{figure} In this paper, we propose a \textbf{Re}trieval-augmented Text-to-\textbf{Ima}ge \textbf{Gen}erator ({Re-Imagen}\xspace), which alleviates such limitations by searching for entity information in a multi-modal knowledge base, rather than attempting to memorize the appearance of rare entities. Specifically, we define our multi-modal knowledge base encodes the visual appearances and descriptions of entities with a collection of reference \texttt{<}image, text\texttt{>} pairs'. To use this resource, {Re-Imagen}\xspace first uses the input text prompt to retrieve the most relevant \texttt{<}image, text\texttt{>} pairs from the external multi-modal knowledge base, then uses the retrieved knowledge as model additional inputs to synthesize the target images. Consequently, the retrieved references provide knowledge regarding the semantic attributes and the concrete visual appearance of mentioned entities to guide {Re-Imagen}\xspace to paint the entities in the target images. The backbone of {Re-Imagen}\xspace is a cascaded diffusion model \citep{ho2022cascaded}, which contains three independent generation stages (implemented as U-Nets \citep{ronneberger2015u}) to gradually produce high-resolution (\textit{i.e.}, 1024{$\times$}1024) images. In particular, we train {Re-Imagen}\xspace on a dataset constructed from the image-text dataset used by Imagen~\citep{saharia2022photorealistic}, where each data instance is associated with the top-k nearest neighbors within the dataset, based on text-only BM25 score. The retrieved top-k \texttt{<}image, text\texttt{>} pairs will be used as a reference for the model attend to. During inference, we design an interleaved guidance schedule that switches between text guidance and retrieval guidance, which ensures both text alignment and entity alignment. We show some examples generated by {Re-Imagen}\xspace, and compare them against Imagen in~\autoref{fig:comparison}. We can qualitatively observe that the our images are more faithful to the appearance of the reference entity. To further quantitatively evaluate {Re-Imagen}\xspace, we present \textit{zero-shot text-to-image generation} results on two challenging datasets: COCO~\citep{lin2014microsoft} and WikiImages~\citep{chang2022webqa}\footnote{The original WikiImages database contains (entity image, entity description) pairs. It was a crawled from Wikimedia Commons for visual question answering, and we repurpose it here for text-to-image generation.}. {Re-Imagen}\xspace uses an external non-overlapping image-text database as the knowledge base for retrieval and then grounds on the retrieval to synthesize the target image. We show that {Re-Imagen}\xspace achieves the state-of-the-art performance for text-to-image generation on COCO and WikiImages, measured in FID score~\cite{heusel2017gans}, among non-fine-tuned models. For non-entity-centric dataset COCO, the perform gain is coming from biasing the model to generate images with similar styles as the retrieved in-domain images. For the entity-centric dataset WikiImages, the performance gain comes from grounding the generation on retrieved images containing similar entities. We further evaluate {Re-Imagen}\xspace on a more challenging benchmark --- EntityDrawBench, to test the model's ability to generate a variety of infrequent entities (dogs, landmarks, foods) in different scenes. We compare {Re-Imagen}\xspace with Imagen~\citep{saharia2022photorealistic}, DALL-E 2~\citep{ramesh2022hierarchical} and StableDiffusion~\citep{rombach2022high} in terms of faithfulness and photorealism with human raters. We demonstrate that that {Re-Imagen}\xspace can reach around 71\% on the faithfulness score, greatly improving from Imagine's score of 27\% and DALL-E 2's score of 48\% . Analysis shows that the improvement mostly comes from low-frequency entities. To summarize, our key contributions are: {(1)} a novel retrieval-augmented text-to-image model {Re-Imagen}\xspace, which achieves SoTA FID scores on two dasets; {(2)} interleaved classifier-free guidance during sampling to ensure both text alignment and entity fidelity; and {(3)} We introduce EntityDrawBench and show that {Re-Imagen}\xspace can significantly improve faithfulness on less-frequent entities. \section{Related Work} \noindent \textbf{Text-to-Image Diffusion Models} There has been a wide-spread success~\citep{ashual2022knn,ramesh2022hierarchical,saharia2022photorealistic,nichol2021glide} in modeling text-to-image generation with diffusion models, which has outperformed GANs~\citep{goodfellow2014generative} and auto-regressive Transformers~\citep{ramesh2021zero} in photorealism and diversity (under similar model size), without training instability and mode collapsing issues. Among them, some recent large text-to-image models such as Imagen~\citep{saharia2022photorealistic}, GLIDE~\citep{nichol2021glide}, and DALL-E2~\citep{ramesh2022hierarchical} have demonstrated excellent generation from complex prompt inputs. These models achieve highly fine-grained control over the generated images with text inputs. However, they do not perform explicit grounding over external visual knowledge and are restricted to memorizing the visual appearance of every possible visual entity in their parameters. This makes it difficult for them to generalize to rare or even unseen entities. In contrast, {Re-Imagen}\xspace is designed to free the diffusion model from memorizing, as models are encouraged to retrieve semantic neighbors from the knowledge base and use retrievals as context to paint the image. {Re-Imagen}\xspace improves the grounding of the diffusion models to real-world knowledge and is therefore capable of faithful image synthesis. \vspace{1ex} \\ \noindent \textbf{Concurrent Work} There are several concurrent works~\citep{li2022memory,blattmann2022retrieval,ashual2022knn}, that also leverage retrieval to improve diffusion models. RDM~\citep{blattmann2022retrieval} is trained similarly to {Re-Imagen}\xspace, using examples and near neighbors, but the neighbors in RDM are selected using image features, and at inference time retrievals are replaced with user-chosen exemplars. RDM was shown to effectively transfer artistic style from exemplars to generated images. In contrast, our proposed {Re-Imagen}\xspace conditions on both text and multi-modal neighbors to generate the image, includes retrieval at inference time, and is demonstrated to improve performance on rare images (as well as more generally). KNN-Diffusion~\citep{ashual2022knn} is more closely related work to us, as it also uses retrieval to the quality of generated images. However, KNN-Diffusion uses discrete image representations, while {Re-Imagen}\xspace uses the raw pixels, and {Re-Imagen}\xspace's retrieved neighbors can be \texttt{<}image, text\texttt{>} pairs, while KNN-Diffusion's are only images. Quantitatively, {Re-Imagen}\xspace{} outperforms KNN-Diffusion on the COCO dataset significantly. \vspace{1ex}\\ \noindent \textbf{Others} Due to the space limit, we provide an additional literature review in the Appendix~\ref{extended_review}. \section{Model} In this section, we discuss our proposed {Re-Imagen}\xspace in detail. We start with background knowledge, in the form of a brief overview of the cascaded diffusion models used by Imagen~\citep{wang2022high}. Next, we describe the concrete technical details on how we incorporate multi-modal retrieval for {Re-Imagen}\xspace. Finally, we discuss training and the interleaved guidance sampling for {Re-Imagen}\xspace. \subsection{Preliminaries} \noindent \textbf{Diffusion Models} Diffusion models~\citep{sohl2015deep} are latent variable models, parameterized by $\theta$, in the form of $p_{\theta}(\bm{x}_0) := \int p_{\theta}(\bm{x}_{0:T})d\bm{x}_{1:T}$, where $\bm{x}_1, \cdots, \bm{x}_T$ are ``noised'' latent versions of the input image $\bm{x}_0 \sim q(\bm{x}_0)$. Note that the dimensionality of both latents and the image are the same throughout the entire process, with $\bm{x}_{0:T} \in \mathbb{R}^d$ and $d$ equals the product of \texttt{<}height, width, \# of channels\texttt{>}. The process that computes the posterior distribution $q(\bm{x}_{1:T}|\bm{x}_0)$ is also called the forward (or diffusion) process, and is implemented as a predefined Markov chain that gradually adds Gaussian noise to the data according to a schedule $\beta_t$: \begin{equation} q(\bm{x}_{1:T}|\bm{x}_0)=\prod_{t=1}^{T}q(\bm{x}_t | \bm{x}_{t-1}) \quad \quad q(\bm{x}_t | \bm{x}_{t-1}) := \mathcal{N}(\bm{x}_t; \sqrt{1 - \beta_t}\bm{x}_{t-1}, \beta_t \bm{I}) \end{equation} Diffusion models are trained to learn the image distribution by reversing the diffusion Markov chain. Theoretically, this reduces to learning to denoise $\bm{x}_t \sim q(\bm{x}_t|\bm{x}_0)$ into $\bm{x}_0$, with a time re-weighted square error loss---see~\cite{ho2020denoising} for the complete proof: \begin{equation} \label{eq:loss} \mathbb{E}_{\bm{x}_0, \bm{\epsilon}, t} [w_t \cdot ||\hat{\bm{x}}_{\theta}(\bm{x}_t, \bm{c}) - \bm{x}_0||_2^2] \end{equation} Here, the noised image is denoted as $\bm{x}_t := \sqrt{\bar{\alpha}_t} \bm{x}_0 + \sqrt{1-\bar{\alpha}_t} \bm{\epsilon}$, $\bm{x}_0$ is the ground-truth image, $\bm{c}$ is the condition, $\bm{\epsilon} \sim \mathcal{N}(\bf{0}, I)$ is the noise term, $\alpha_t: = 1 - \beta_t$ and $\bar{\alpha}_t := \prod_{s=1}^t \alpha_s$. To simplify notation, we will allow the condition $\bm{c}$ include multiple conditioning signals, such as text prompts $\bm{c}_p$, a low-resolution image input $\bm{c}_x$ (which is used in super-resolution), or retrieved neighboring images $\bm{c}_n$ (which are used in {Re-Imagen}\xspace). Imagen~\citep{saharia2022photorealistic} uses a U-Net~\citep{ronneberger2015u} to implement $\bm{\epsilon}_{\theta}(\bm{x}_{t}, \bm{c}, t)$. The U-Net represents the reversed noise generator as follows: \begin{equation} \label{eq:recover_original} \hat{\bm{x}}_{\theta}(\bm{x}_t, \bm{c}) := (\bm{x}_t - \sqrt{1 - \bar{\alpha}_t} \bm{\epsilon}_{\theta}(\bm{x}_{t}, \bm{c}, t)) / \sqrt{\bar{\alpha}_t} \end{equation} During the training, we randomly sample $t \sim \mathcal{U}([0, 1])$ and image $\bm{x}_0$ from the dataset $\mathcal{D}$, and minimize the difference between $\hat{\bm{x}}_{\theta}(\bm{x}_t, \bm{c})$ and $\bm{x}_0$ according to~\autoref{eq:loss}. At the inference time, the diffusion model uses DDPM~\citep{ho2020denoising} to sample recursively as follows: \begin{equation} \bm{x}_{t-1} = \frac{\sqrt{\bar{\alpha}_{t-1}} \beta_t}{1 - \bar{\alpha}_t} \hat{\bm{x}}_{\theta}(\bm{x}_t, \bm{c}) + \frac{\sqrt{\alpha_t}(1- \bar{\alpha}_{t-1})}{1 - \bar{\alpha}_t}\bm{x}_t + \sqrt{\frac{(1 - \bar{\alpha}_{t-1})\beta_t}{1 - \bar{\alpha}_t}}\bm{\epsilon} \end{equation} The model sets $\bm{x}_T$ as a Gaussian noise with $T$ denoting the total number of diffusion steps, and then keep sampling in reverse until step $T=0$, i.e. $\bm{x}_{T} \rightarrow \bm{x}_{T-1} \rightarrow \cdots$, to reach the final image $\hat{\bm{x}}_0$. For better generation efficiency, cascaded diffusion models~\citep{ho2022cascaded,ramesh2022hierarchical,saharia2022photorealistic} use three separate diffusion models to generate high-resolution images gradually, going from low resolution to high resolution. The three models 64$\times$ model, 256$\times$ super-resolution model and 1024$\times$ super-resolution model gradually increase the model resolution to $1024\times1024$. \begin{figure}[!t] \centering \includegraphics[width=1.0\linewidth]{model.001.jpeg} \caption{ An illustration of the text-to-image generation pipeline in the 64$\times$ diffusion model. Specifically, {Re-Imagen}\xspace learns a UNet to iteratively predict $\bm{\epsilon}(\bm{x_t}, \bm{c}_n, \bm{c}_p, t)$ that denoises the image. ($\bm{c}_n$: a set of retrieved image-text pairs from the database; $\bm{c}_p$: input text prompt; $t$: current time-step) } \label{fig:model} \end{figure} \noindent \textbf{Classifier-free Guidance} \label{sec:cfg} \cite{ho2021classifier} first proposed classifier-free guidance to trade off diversity and sample quality. This sampling strategy has been widely used due to its simplicity. In particular, Imagen~\citep{saharia2022photorealistic} adopts an adjusted $\epsilon$-prediction as follows: \begin{equation} \hat{\bm{\epsilon}} = w \cdot \bm{\epsilon}_{\theta}(\bm{x}_t, \bm{c}, t) - (w - 1) \cdot \bm{\epsilon}_{\theta}(\bm{x}_t, t) \end{equation} where $w$ is the guidance weight. The unconditional $\epsilon$-prediction $\bm{\epsilon}_{\theta}(\bm{x}_t, t)$ is calculated by dropping the condition, i.e. the text prompt. \subsection{Generating Image with Multi-Modal Knowledge} Similar to Imagen~\citep{saharia2022photorealistic}, {Re-Imagen}\xspace is a cascaded diffusion model, consisting of 64$\times$, 256$\times$, and 1024$\times$ diffusion models. However, {Re-Imagen}\xspace augments the diffusion model with the new capability of leveraging multimodal `knowledge' from the external database, thus freeing the model from memorizing the appearance of rare entities. For brevity (and concreteness) we present below a high-level overview of the 64$\times$ model: the others are similar. \noindent \textbf{Main Idea} As shown in~\autoref{fig:model}, during the denoising process, {Re-Imagen}\xspace conditions its generation result not only on the text prompt $\bm{c}_p$ (and also with $\bm{c}_x$ for super-resolution), but on the neighbors $\bm{c}_n$ that were retrieved from the external knowledge base. Here, the text prompt $\bm{c}_p \in \mathbb{R}^{n \times d}$ is represented using a T5 embedding~\citep{raffel2020exploring}, with $n$ being the text length and $d$ being the embedding dimension. Meanwhile, the top-k neighbors $\bm{c}_n:=[\texttt{<}\text{image, text}\texttt{>}_1, \cdots, \texttt{<}\text{image, text}\texttt{>}_k]$ are retrieved from external knowledge base $\mathcal{B}$, using the input prompt $p$ as the query and a retrieval similarity function $\gamma(p, \mathcal{B})$. We experimented with two different choices for the similarity function: maximum inner product scores for BM25~\citep{robertson2009probabilistic} and CLIP~\citep{radford2021learning}. \vspace{1ex} \\ \noindent \textbf{Model Architecture} We show the architecture of our model in~\autoref{fig:detail}, where we decompose the UNet into the downsampling encoder (DStack) and the upsampling decoder (UStack). Specifically, the DStack takes an image, a text, and a time step as the input, and generates a feature map, which is denoted as $f_{\theta}(\bm{x}_t, \bm{c}_p, t) \in \mathbb{R}^{F \times F \times d}$, with $F$ denoting the feature map width and $d$ denoting the hidden dimension. We share the same DStack encoder when we encode the retrieved \texttt{<}image, text\texttt{>} pairs (with $t$ set to zero) which produces a set of feature maps $f_{\theta}(\bm{c}_n, 0) \in \mathbb{R}^{K \times F \times F \times d}$. We then use a multi-head attention module~\citep{vaswani2017attention} to extract the most relevant information to produce a new feature map $f'_{\theta}(\bm{x}_t, \bm{c}_p, \bm{c}_n, t) = Attn(f_{\theta}(\bm{x}_t, \bm{c}_p, t), f_{\theta}(\bm{c}_n, 0))$. The upsampling stack decoder then predicts the noise term $\bm{\epsilon}_{\theta}(\bm{x}_{t}, \bm{c}_p, \bm{c}_n, t)$ and uses it to compute $\hat{\bm{x}}_{\theta}$ with~\autoref{eq:recover_original}, which is either used for regression during training or DDPM sampling. \vspace{1ex} \\ \noindent \textbf{Model Training} In order to train {Re-Imagen}\xspace, we construct a new dataset KNN-ImageText based on the 50M ImageText-dataset used in Imagen. There are two motivations for selecting this dataset. (1) the dataset contains many similar photos regarding specific entities, which is extremely helpful for obtaining similar neighbors, and (2) the dataset is highly sanitized with fewer unethical or harmful images. For each instance in the 50M ImageText-dataset, we search over the same dataset with text-to-text BM25 similarity to find the top-2 neighbors as $\bm{c}_n$ (excluding the query instance). We experimented with both CLIP and BM25 similarity scores, and retrieval was implemented with ScaNN~\citep{guo2020accelerating}. We train {Re-Imagen}\xspace on the KNN-ImageText by minimizing the loss function of~\autoref{eq:loss}. During training, we also randomly drop the text and neighbor conditions independently with 10\% chance. Such random dropping will help the model learn the marginalized noise term $\bm{\epsilon}_{\theta}(\bm{x}_{t}, \bm{c}_p, t)$ and $\bm{\epsilon}_{\theta}(\bm{x}_{t}, \bm{c}_n, t)$, which will be used for the classifier-free guidance. \vspace{1ex} \\ \noindent \textbf{Interleaved Classifier-free Guidance} \label{sec:interleaved} Different from existing diffusion models, our model needs to deal with more than one condition, \textit{i.e.}, text prompts $\bm{c}_t$ and retrieved neighbors $\bm{c}_n$, which allows new options for incorporating guidance. In particular, {Re-Imagen}\xspace could use classifier-free guidance by subtracting the unconditioned $\epsilon$-predictions, or either of the two partially conditioned $\epsilon$-predictions. Empirically, we observed that subtracting unconditioned $\epsilon$-predictions (the standard classifier-free guidance of~\autoref{sec:cfg}) often leads to an undesired imbalance, where the outputs are either dominated by the text condition or the neighbor condition. Hence, we designed an interleaved guidance schedule that balances the two conditions. Formally, we define the two adjusted $\epsilon$-predictions as: \begin{align} \label{eq:sample} \begin{split} \hat{\bm{\epsilon}}_{p} &= w_p \cdot \bm{\epsilon}_{\theta}(\bm{x}_t, \bm{c}_p, \bm{c}_n, t) - (w_p - 1) \cdot \bm{\epsilon}_{\theta}(\bm{x}_t, \bm{c}_n, t) \\ \hat{\bm{\epsilon}}_{n} &= w_n \cdot \bm{\epsilon}_{\theta}(\bm{x}_t, \bm{c}_p, \bm{c}_n, t) - (w_n - 1) \cdot \bm{\epsilon}_{\theta}(\bm{x}_t, \bm{c}_p, t) \end{split} \end{align} where $\hat{\bm{\epsilon}}_{p}$ and $\hat{\bm{\epsilon}}_{n}$ are the text-enhanced and neighbor-enhanced $\epsilon$-predictions, respectfully. Here, $w_p$ is the text guidance weight and $w_n$ is the neighbor guidance weight. We then interleave the two guidance predictions by a certain predefined ratio $\eta$. Specifically, at each guidance step, we sample a [0, 1]-uniform random number $R$, and $R < \eta$, we use $\hat{\bm{\epsilon}}_{p}$, and otherwise $\hat{\bm{\epsilon}}_{n}$. We can adjust $\eta$ to balance the faithfulness w.r.t text description or the retrieved image-text pairs. In EntityDrawBench experiment, we found that $\eta=0.55$ can lead to better quality. \begin{figure}[!t] \centering \includegraphics[width=1.0\linewidth]{kimgen.001.jpeg} \caption{The detailed architecture of our model. The retrieved neighbors are first encoded using the DStack encoder and then used to augment the intermediate representation of the denoising image (via cross-attention). The augmented representation is fed to the UStack to predict the noise.} \label{fig:detail} \vspace{-2ex} \end{figure} \section{Experiments} {Re-Imagen}\xspace consists of three submodels: a 2.5B 64$\times$64 text-to-image model, a 750M 256$\times$256 super-resolution model and a 400M 1024$\times$1024 super-resolution model. We finetune these models on the constructed KNN-ImageText dataset. We evaluate the model under two settings: (1) automatic evaluation on COCO and WikiImages dataset, to measure the model's general performance to generate photorealistic images, and (2) human evaluation on the newly introduced EntityDrawBench, to measure the model's capability to generate long-tail entities. \noindent \textbf{Training and Evaluation details} The fine-tuning was run for 200K steps on 64 TPU-v4 chips and completed within two days. We use Adafactor for the 64$\times$ model and Adam for the 256$\times$ super-resolution model with a learning rate of 1e-4. We set the number of neighbors $k$=2 and set $\gamma$=BM25 during training. For the image-text database $\mathcal{B}$, we consider three different variants: {(1)} the COCO/WikiImages training set, which contains non-overlapping small-scale in-domain image-text pairs, {(2)} the ImageText dataset containing 50M \texttt{<}image, caption\texttt{>} pairs, and {(3)} the LAION dataset~\citep{schuhmann2021laion} containing 400M \texttt{<}image, text\texttt{>} crawled pairs. Since indexing ImageText and LAION with CLIP encodings is expensive, we only considered the BM25 retriever for these databases. For the COCO/WikiImages training set, we used both BM25 and CLIP. \subsection{Evaluation on COCO and WikiImages} In these two experiments, we used the standard non-interleaved classifier-free guidance (\autoref{sec:cfg}) with $T$=1000 steps for both the 64$\times$ diffusion model and 256$\times$ super-resolution model. The guidance weight $w$ for the 64$\times$ model is swept over [1.0, 1.25, 1.5, 1.75, 2.0], while the 256$\times$256 super-resolution models' guidance weight $w$ is swept over [1.0, 5.0, 8.0, 10.0]. We select the guidance $w$ with the best FID score, which is reported in~\autoref{tab:coco}. We also demonstrate examples in~\autoref{fig:dataset}. \vspace{1ex} \\ \noindent \textbf{COCO Results} COCO is the most widely-used benchmark for text-to-image generation models. Although COCO does not contain many rare entities, it does contain unusual combinations of common entities, so it is plausible that retrieval augmentation could also help for some challenging text prompts. We adopt FID~\citep{heusel2017gans} score to measure image quality. Following the previous literature, we randomly sample 30K prompts from the validation set as input to the model. The generated images are compared with the reference images from the full validation set (42K). We list the results in two columns: FID-30K denotes the model with access to the COCO train set (either to fine-tune or retrieve from), while Zero-shot FID-30K does not have access to any COCO data. \begin{table}[!t] \small \centering \begin{tabular}{lccc} \toprule Model & \# of Params & FID-30K & Zero-shot FID-30K \\ \midrule GLIDE~\citep{nichol2021glide} & \hphantom{0}\pd5B & - & 12.24 \\ DALL-E 2~\citep{ramesh2022hierarchical} & $\sim$5B & - & 10.39 \\ VQ-Diffusion~\citep{gu2022vector} & 0.4B & - & 19.75 \\ KNN-Diffusion~\citep{ashual2022knn} & 0.8B & - & 16.66 \\ Stable-Diffusion~\citep{rombach2022high} & \hphantom{.}\pz1B & - & 12.63 \\ Imagen~\citep{saharia2022photorealistic} & \hphantom{.}\pz3B & - & \pz7.27 \\ Make-A-Scene~\citep{gafni2022make} & \hphantom{.}\pz4B & 7.55 & 11.84 \\ Parti~\citep{yu2022scaling} & \pd20B & \textbf{3.22} & \pz7.23 \\ \midrule {Re-Imagen}\xspace ($\gamma$=BM25; $\mathcal{B}$=COCO; $k$=2) & 3.6B & \textbf{5.25}$^\dagger$ & - \\ {Re-Imagen}\xspace ($\gamma$=CLIP; $\mathcal{B}$=COCO; $k$=2) & 3.6B & 5.29$^\dagger$ & - \\ {Re-Imagen}\xspace ($\gamma$=BM25; $\mathcal{B}$=ImageText; $k$=2) & 3.6B & - & \pz7.02 \\ {Re-Imagen}\xspace ($\gamma$=BM25; $\mathcal{B}$=LAION; $k$=2) & 3.6B & - & \hphantom{0} 6.88 \\ \bottomrule \end{tabular} \caption{MS-COCO results for zero-shot text-to-image generation. We use a guidance weight of 1.25 for the 64$\times$ diffusion model and 5 for our 256$\times$ super-resolution model. ($\dagger$: {Re-Imagen}\xspace \textit{is not fine-tuned} on the COCO data---it only uses it as the knowledge base for retrieval.) } \label{tab:coco} \vspace{-2ex} \end{table} {Re-Imagen}\xspace (with the COCO database) can achieve a significant gain on FID-30K without fine-tuning: roughly a 2.0 absolute FID improvement over Imagen. The performance is even better than fine-tuned Make-A-Scene~\citep{gafni2022make}, but slightly worse than fine-tuned 20B Parti. In contrast, {Re-Imagen}\xspace retrieving from out-of-domain databases (LAION) achieves less gain, but still obtains a 0.4 FID improvement over Imagen. {Re-Imagen}\xspace outperforms KNN-Diffusion, another retrieval-augmented diffusion model, by a large margin. Since COCO does not contain infrequent entities, `entity knowledge' is not important. In contrast, retrieving from the training set can provide useful `style knowledge' for the model to ground on. {Re-Imagen}\xspace is able to adapt the generated images to the same style of the COCO distribution, it can achieve a much better FID score. As can be seen in the upper part of from~\autoref{fig:dataset}, {Re-Imagen}\xspace with retrieval generates images of the same style as COCO, while without retrieval, the output is still high quality, but the style is less similar to COCO. \begin{figure}[!t] \centering \includegraphics[width=1.0\linewidth]{dataset.001.jpeg} \caption{{The retrieved top-2 neighbors of COCO and WikiImages and model generation.}} \label{fig:dataset} \vspace{-2ex} \end{figure} \noindent\textbf{WikiImages Results} WikiImages is constructed based on the multimodal corpus provided in Web\-QA~\citep{chang2022webqa}, which consists of \texttt{<}image, text\texttt{>} pairs crawled from Wikimedia Commons\footnote{\url{https://commons.wikimedia.org/wiki/Main_Page}}. We filtered the original corpus to remove noisy data (see the Appendix\ref{wikiimages}), which leads to a total of 320K examples. We randomly sample 22K as our validation set to perform zero-shot evaluation, we further sample 20K prompts from the dataset as the input. Similar to the previous experiment, we also adopt the guidance weight schedule as before and evaluate 256$\times$256 images. We report our experimental results in~\autoref{tab:wiki} and mainly compare with Imagen and Stable-Diffusion. \begin{table}[!t] \small \centering \begin{tabular}{lccc} \toprule Model & \# of Params & FID-30K & Zero-shot FID-20K \\ \midrule Stable-Diffusion~\citep{rombach2022high} & \hphantom{.}\pz1B & - & 7.50 \\ Imagen~\citep{saharia2022photorealistic} & \hphantom{.}\pz3B & - & 6.44 \\ \midrule {Re-Imagen}\xspace ($\gamma$=BM25; $\mathcal{B}$=WikiImages; $k$=2) & 3.6B & 5.88 & - \\ {Re-Imagen}\xspace ($\gamma$=CLIP; $\mathcal{B}$=WikiImages; $k$=2) & 3.6B & 5.85 & - \\ {Re-Imagen}\xspace ($\gamma$=BM25; $\mathcal{B}$=ImageText; $k$=2) & 3.6B & - & 6.04 \\ {Re-Imagen}\xspace ($\gamma$=BM25; $\mathcal{B}$=LAION; $k$=1) & 3.6B & - & 5.94 \\ {Re-Imagen}\xspace ($\gamma$=BM25; $\mathcal{B}$=LAION; $k$=2) & 3.6B & - & 5.82 \\ {Re-Imagen}\xspace ($\gamma$=BM25; $\mathcal{B}$=LAION; $k$=3) & 3.6B & - & \textbf{5.80} \\ \bottomrule \end{tabular} \caption{WikiImages results for zero-shot text-to-image generation. We use a guidance weight of 1.5 for the 64$\times$ diffusion model and 5 for our 256$\times$ super-resolution model.} \label{tab:wiki} \vspace{-2ex} \end{table} From~\autoref{tab:wiki}, we found that using out-of-domain LAION-400M as the database actually achieves better performance than using in-domain WikiImages as the database. Unlike COCO, Wiki\-Images contains mostly entity-focused images, thus the importance of finding relevant entities is the database is more important than distilling the styles from the training set---and since the scale of LAION-400M is 100x larger than WikiImages-300K, the chance of retrieving related entities is much higher, which leads to better performance. One example is depicted in the lower part of~\autoref{fig:dataset}, where the LAION retrieval finds `Island of San Giorgio Maggiore', which helps the model generate the classical Renaissance-style church. When generating without retrieval, the model is not able to generate the specific church. This indicates the importance of having relevant entities in the retrievals for WikiImages dataset and also explains why LAION database achieves the best results. We also present more examples from WikiImages in the Appendix~\ref{wikiimages}. \subsection{Entity Focused Evaluation on EntityDrawBench} \noindent \textbf{Dataset Construction} We introduce EntityDrawBench to evaluate the model's capability to generate diverse sets of entities in different visual scenes. Specifically, we pick three types of visual entities (dog breeds, landmarks, and foods) from Wikipedia Commons and Google Landmarks to construct our prompts. In total, we collect 150 entity-centric prompts for evaluation. These prompts are mostly unique and we cannot find corresponding images with Google Image Search. More construction details are in Appendix~\ref{entitydrawbench}. We use the prompt as the input and its corresponding image-text pairs as the `retrieval' for {Re-Imagen}\xspace, to generate four 1024$\times$1024 images. For the other models, we feed the prompts directly also to generate four images. We will pick the best image of these four samples to rate its Photorealism and Faithfulness. For photorealism, we assign 1 if the image is moderately realistic without noticeable artifacts, otherwise, we assign a score of 0. For the faithfulness measure, we assign 1 if the image is faithful to both the entity source and the text description, otherwise, we assign 0. \noindent \textbf{Experimental Results} We use the proposed interleaved classifier-free guidance (\autoref{sec:interleaved}) for the 64$\times$ diffusion model, which runs for 256 diffusion steps under a strong guidance weight of $w$=30 for both text and neighbor conditions. For the 256$\times$ and 1024$\times$ resolution models, we use a constant guidance weight of 5.0 and 3.0, respectively, with 128 and 32 diffusion steps. The inference speed is 30-40 secs for 4 images on 4 TPU-v4 chips. We demonstrate our human evaluation results for faithfulness and photorealism in~\autoref{tab:faithfulness}. \begin{table}[!t] \small \centering \begin{tabular}{l|cccc|c} \toprule \multirow{2}{*}{Model} & \multicolumn{4}{c|}{\textbf{Faithfulness}} & \textbf{Photorealism} \\ & Dogs & Foods & Landmarks & All & All \\ \midrule \addlinespace Imagen & 0.28 $\pm$ 0.02 & 0.26 $\pm$ 0.02 & 0.27 $\pm$ 0.02 & 0.27 & \textbf{0.98} \\ DALL-E 2 & 0.60 $\pm$ 0.02 & 0.47 $\pm$ 0.02 & 0.36 $\pm$ 0.04 & 0.48 & \textbf{0.98} \\ Stable-Diffusion & 0.16 $\pm$ 0.02 & 0.24 $\pm$ 0.04 & 0.12 $\pm$ 0.06 & 0.17 & 0.92 \\ \midrule \addlinespace {Re-Imagen}\xspace & \textbf{0.68} $\pm$ 0.04 & \textbf{0.70} $\pm$ 0.02 & \textbf{0.74} $\pm$ 0.04 & \textbf{0.71} & 0.97 \\ \bottomrule \end{tabular} \caption{Human evaluation results for different models on different types of entities. } \label{tab:faithfulness} \vspace{-2ex} \end{table} We can observe that {Re-Imagen}\xspace can in general achieve much higher faithfulness than the existing models while maintaining similar photorealism scores. When comparing with our backbone Imagen, we see the faithfulness score jumps from 27\% to 71\%, which indicates that our model is paying attention to the retrieved knowledge and assimilating it into the generation process. \begin{figure}[!h] \centering \begin{tabular}{@{}c@{}c@{}} \includegraphics[height=3cm]{figures/frequent_entity_v3} & \includegraphics[height=3cm]{figures/infrequent_entity_v3.png} \\ \end{tabular} \vspace{-2ex} \caption{The human evaluation scores for both frequent and infrequent entities. } \label{fig:human_evaluation_frequency} \end{figure} We further partition the entities into `frequent' and `infrequent' categories based on their frequency (top 50\% as `frequent') in Imagen's training corpus. We plot faithfulness score for `frequent' and `infrequent' separately in~\autoref{fig:human_evaluation_frequency} . We can see that our model is less sensitive to the frequency of the input entities than the other models, with only a 10-20\% drop on infrequent entities. In contrast, both Imagen and DALL-E 2 drop by 40\%-50\% on infrequent entities. This study reflects the effectiveness of text-to-image generation models on long-tail entities. \noindent \textbf{Comparison to Other Models} We demonstrate some examples from different models in~\autoref{fig:example}. As can be seen, the images generated from {Re-Imagen}\xspace strike a good balance between text alignment and entity fidelity. Unlike image editing to perform in-place modification, {Re-Imagen}\xspace can transform the neighbor entities both geometrically and semantically according to the text guidance. As a concrete example, {Re-Imagen}\xspace generates the \textit{Braque Saint-Germain} (2nd row in \autoref{fig:example}) on the grass, in a different viewpoint from to the reference image. \begin{figure}[!t] \centering \includegraphics[width=0.95\linewidth]{example.001.jpeg} \caption{None-cherry picked examples from EntityDrawBench for different models. } \vspace{-1ex} \label{fig:example} \end{figure} \begin{figure}[!t] \centering \includegraphics[width=0.95\linewidth]{more_examples_for_arxiv.001.jpeg} \includegraphics[width=0.95\linewidth]{more_examples_for_arxiv.002.jpeg} \includegraphics[width=0.95\linewidth]{more_examples_for_arxiv.003.jpeg} \caption{Extra None-cherry picked examples from EntityDrawBench for different models. } \vspace{-1ex} \label{fig:more_example_for_arxiv} \end{figure} \noindent \textbf{Text and Entity Faithfulness Trade-offs} In our experiments, we found that there is a trade-off between faithefulness to the text prompt and faithfulness to the retrieved entity images. Based on~\autoref{eq:sample}, by adjusting $\eta$, \textit{i.e.} the proportion of $\hat{\epsilon}_p$ and $\hat{\epsilon}_n$ in the sampling schedule, we can control {Re-Imagen}\xspace so as to generate images that explore this tradeoff: decreasing $\eta$ will increase the entity's image faithfulness but decrease the text alignment. In contrast, increasing the value $\eta$ will increase the text alignment and decrease the similarity to retrieved image. We demonstrate this in~\autoref{fig:transition}. With small $\eta$, the model ignores the text description and simply copies the retrieved image, while with large $\eta$, the model reverts to standard Imagen, without using to the input image much. We found that having $\eta$ around 0.5 is usually a `sweet spot' that balances both conditions. \begin{figure}[!t] \centering \includegraphics[width=0.95\linewidth]{transition.001.jpeg} \caption{Ablation study of interleaved guidance ratio $\eta$ to show the trade-off. } \vspace{-2ex} \label{fig:transition} \end{figure} \section{Conclusions and Limitations} We present {Re-Imagen}\xspace, a retrieval-augmented diffusion model, and demonstrate its effectiveness in generating realistic and faithful images. We exhibit such advantages not only through automatic FID measures on standard benchmarks (\textit{i.e.}, COCO and WikiImage) but also through human evaluation on the newly introduced EntityDrawBench. We further demonstrate that our model is particularly effective in generating an image from text that mentions rare entities. {Re-Imagen}\xspace still suffers from well-known issues in text-to-image generation, which we review below in \autoref{broader_impact}. In addition, {Re-Imagen}\xspace also has some unique limitations due to the retrieval-augmented modeling. First, because {Re-Imagen}\xspace is sensitive the to retrieved image-text pairs it is conditioned on, when the retrieved image is of low-quality, there will be a negative influence on the generated image. Second, {Re-Imagen}\xspace sometimes still fail to ground on the retrieved entities when the entity's visual appearance is out of the generation space. Third, we noticed that the super-resolution model is less effective, and frequently misses low-level texture details of the visual entities. In future work, we plan to further investigate the above limitations and address them. \section*{Ethics Statement} \label{broader_impact} Strong text-to-image generation models, \textit{i.e.}, Imagen~\citep{saharia2022photorealistic} and Parti~\citep{yu2022scaling}, raise ethical challenges along dimensions such as the \textit{social bias}. {Re-Imagen}\xspace is exposed to the same challenges, as we employed Web-scale datasets that are similar to the prior models. The retrieval-augmented modeling techniques of {Re-Imagen}\xspace has substantially improved the controllability and attribution of the generated image. Like many basic research topics, this additional control could be used for beneficial or harmful purposes. One obvious danger is that {Re-Imagen}\xspace (or similar models) could be used for malicious purposes like spreading misinformation, \textit{e.g.,} by producing realistic images of specific people in misleading visual contexts. On the other side, additional control has many potential benefits. One general benefit is that {Re-Imagen}\xspace can reduce hallucination and increase the faithfulness of the generated image to the users' intent. Another benefit is that the ability to work with tail entities makes the model more useful for minorities and other users in smaller communities: for example, {Re-Imagen}\xspace is more effective at generating images of landmarks famous in smaller communities or cultures, and generating images of indigenous foods and cultural artifacts. We argue that this model can help decrease the frequency-caused bias in current neural network based AI systems. Considering such potential threats to the public, we have currently decide not to release the code or a public demo. In future work, we will explore a framework for responsible use that balances the value of external auditing of research with the risks of unrestricted open access, allowing this work to be used in a safe and beneficial way. \section*{Acknoledgement} We thank Jason Baldridge, Boqing Gong, Keran Rong and Slav Petrov for their valuable comments on an early version of the manuscript, which has helped improve this work. We also thank William Chan and Mohammad Norouzi for providing us with their support and the pre-trained models of Imagen, and Michiel de Jong for suggesting the model name.

7d603d027ea60827f533f199999e90a8386417d1

\section{Introduction} Our aim in this paper is to explain the origin of the problems that have been noticed \cite{MeTi2019} when using Gauss-Radau quadrature upper bounds of the $A$-norm of the error in the Conjugate Gradient (CG) algorithm for solving linear systems $Ax=b$ with a symmetric positive definite matrix of order $N$. \smallskip{} The connection between CG and Gauss quadrature has been known since the seminal paper of Hestenes and Stiefel \cite{HeSt1952} in 1952. This link has been exploited by Gene H.~Golub and his collaborators to bound or estimate the $A$-norm of the error in CG during the iterations; see \cite{DaEiGo1972,DaGoNa1979,GoMe1994,GoMe1997,GoSt1994,Me1997,Me1999,MeTi2013,MeTi2014,MeTi2019,StTi2002,StTi2005}. Using a fixed node $\mu$ smaller than the smallest eigenvalue of $A$ and the Gauss-Radau quadrature rule, an upper bound for the $A$-norm of the error can be easily computed. Note that it is useful to have an upper bound of the error norm to stop the CG iterations. In theory, the closer $\mu$ is to the smallest eigenvalue, the closer is the bound to the norm. However, in many examples, even if the bound is close to the norm at the beginning of the CG iterations, the upper bound becomes worse after a while and almost independent of $\mu$, even in exact arithmetic. Therefore, this problem is not linked to rounding errors and has to be explained. \smallskip{} The Gauss quadrature bounds of the error norm were obtained by using the connection of CG to the Lanczos algorithm and modifications of the tridiagonal matrix which is generated by this algorithm and implicitly by CG. This is why we start in Section~2 with the Lanczos algorithm. In Section~3 we discuss the relation with CG and how the Gauss-Radau upper bound is computed. A model problem showing the problems arising with the Gauss-Radau bound in ``exact'' arithmetic is constructed in Section~4. In Sections~5 to 7 we give an analysis that explains that the problems start when the distance of the smallest Ritz value to the smallest eigenvalue becomes smaller than the distance of $\mu$ to the smallest eigenvalue. We also explain why the Gauss-Radau upper bound becomes almost independent of $\mu$. In Section~8 we give an algorithm for improving the accuracy of the upper bound at iteration $k$ using the information from the next CG iterations. In particular, the algorithm determines adaptively the number of forthcoming CG iterations that are needed to get an estimate with a prescribed relative accuracy. Conclusions are given in Section~9. \section{The Lanczos algorithm} Given a starting vector $v$ and a symmetric matrix $A\in\mathbb{R}^{N\times N}$, one can consider a sequence of nested Krylov subspaces \[ \mathcal{K}_{k}(A,v)\equiv\mathrm{span}\{v,Av,\dots,A^{k-1}v\},\qquad k=1,2,\dots \] The dimension of these subspaces can increase up to an index $n$ called the \emph{grade of $v$ with respect to $A$}, at which the maximal dimension is attained, and $\mathcal{K}_{n}(A,v)$ is invariant under multiplication with $A$. \begin{algorithm}[h] \caption{Lanczos algorithm} \label{alg:lanczos} \begin{algorithmic}[0] \STATE \textbf{input} $A$, $v$ \STATE $\beta_{0}=0$, $v_{0}=0$ \STATE $v_{1}=v/\|v\|$ \FOR{$k=1,\dots$} \STATE $w=Av_{k}-\beta_{k-1}v_{k-1}$ \STATE $\alpha_{k}=v_{k}^{T}w$ \STATE $w=w-\alpha_{k}v_{k}$ \STATE $\beta_{k}=\|w\|$ \STATE $v_{k+1}=w/\beta_{k}$ \ENDFOR \end{algorithmic} \end{algorithm} Assuming that $k<n$, the Lanczos algorithm (Algorithm \ref{alg:lanczos}) computes an orthonormal basis $v_{1},\dots,v_{k+1}$ of the Krylov subspace $\mathcal{K}_{k+1}(A,v)$. The basis vectors $v_{j}$ satisfy the matrix relation \[ AV_{k}=V_{k}T_{k}+\beta_{k}v_{k+1}e_{k}^{T} \] where $e_{k}$ is the last column of the identity matrix of order $k$, $V_{k}=[v_{1}\cdots v_{k}]$ and $T_{k}$ is the $k\times k$ symmetric tridiagonal matrix of the recurrence coefficients computed in Algorithm \ref{alg:lanczos}: \[ T_{k}=\left[\begin{array}{cccc} \alpha_{1} & \beta_{1}\\ \beta_{1} & \ddots & \ddots\\ & \ddots & \ddots & \beta_{k-1}\\ & & \beta_{k-1} & \alpha_{k} \end{array}\right]. \] The coefficients $\beta_{j}$ being positive, $T_{k}$ is a so-called Jacobi matrix. If $A$ is positive definite, then $T_{k}$ is positive definite as well. In the following we will assume for simplicity that the eigenvalues of $A$ are simple and sorted such that \[ \lambda_{1}<\lambda_{2}<\cdots<\lambda_{N}. \] \subsection{Approximation of the eigenvalues} The eigenvalues of $T_{k}$ (Ritz values) are usually used as approximations to the eigenvalues of $A$. The quality of the approximation can be measured using $\beta_{k}$ and the last components of the normalized eigenvectors of $T_{k}$. In more detail, consider the spectral decomposition of $T_{k}$ in the form \[ T_{k}=S_{k}\Theta_{k}S_{k}^{T},\quad\Theta_{k}=\mathrm{diag}\left(\theta_{1}^{(k)},\dots,\theta_{k}^{(k)}\right),\quad S_{k}^{T}S_{k}=S_{k}S_{k}^{T}=I_{k}, \] where $I_k$ is the identity matrix of order $k$, and assume that the Ritz values are sorted such that \[ \theta_{1}^{(k)}<\theta_{2}^{(k)}<\cdots<\theta_{k}^{(k)}. \] Denote $s_{i,j}^{(k)}$ the entries and $s_{:,j}^{(k)}$ the columns of $S_{k}$. Then it holds that \begin{equation} \min_{i=1,\dots,N}|\lambda_{i}-\theta_{j}^{(k)}|\leq\left\Vert A\left(V_{k}s_{:,j}^{(k)}\right)-\theta_{j}^{(k)}\left(V_{k}s_{:,j}^{(k)}\right)\right\Vert =\beta_{k}\left|s_{k,j}^{(k)}\right|,\label{eq:simplebound} \end{equation} $j=1,\dots,k$. Since the Ritz values $\theta_{j}^{(k)}$ can be seen as Rayleigh quotients, one can improve the bound \eqref{eq:simplebound} using the gap theorem; see \cite[p.~244]{B:Pa1998} or \cite[p.~206]{B:De1997}. In particular, let $\lambda_{\ell}$ be an eigenvalue of $A$ closest to $\theta_{j}^{(k)}$. Then \[ |\lambda_{\ell}-\theta_{j}^{(k)}|\leq\frac{\left(\beta_{k}s_{k,j}^{(k)}\right)^{2}}{\mathrm{gap}_{j}^{(k)}},\qquad\mathrm{gap}_{j}^{(k)}=\min_{i\neq\ell}\left|\lambda_{i}-\theta_{j}^{(k)}\right|. \] In the following we will be interested in the situation when the smallest Ritz value $\theta_{1}^{(k)}$ closely approximates the smallest eigenvalue of $A$. If $\lambda_{1}$ is the eigenvalue of $A$ closest to $\theta_{1}^{(k)}>\lambda_{1}$, then using the gap theorem and \cite[Corollary 11.7.1 on p.~246]{B:Pa1998}, \begin{equation} \frac{\left(\beta_{k}s_{k,1}^{(k)}\right)^{2}}{\lambda_{n}-\lambda_{1}}\leq\theta_{1}^{(k)}-\lambda_{1}\leq\frac{\left(\beta_{k}s_{k,1}^{(k)}\right)^{2}}{\lambda_{2}-\theta_{1}^{(k)}},\label{eq:gap} \end{equation} giving the bounds \begin{equation} \lambda_{2}-\theta_{1}^{(k)}\leq\frac{\left(\beta_{k}s_{k,1}^{(k)}\right)^{2}}{\theta_{1}^{(k)}-\lambda_{1}}\leq\lambda_{n}-\lambda_{1}.\label{eq:bound1} \end{equation} It is known (see, for instance, \cite{B:Me2006}) that the squares of the last components of the eigenvectors are given by \[ \left(s_{k,j}^{(k)}\right)^{2}=\left|\frac{\chi_{1,k-1}(\theta_{j}^{(k)})}{\chi_{1,k}^{'}(\theta_{j}^{(k)})}\right|, \] where $\chi_{1,\ell}$ is the characteristic polynomial of $T_{\ell}$ and $\chi_{1,\ell}^{'}$ denotes its derivative, i.e., \[ \left(s_{k,j}^{(k)}\right)^{2}=\frac{\theta_{j}^{(k)}-\theta_{1}^{(k-1)}}{\theta_{j}^{(k)}-\theta_{1}^{(k)}}\cdots\frac{\theta_{j}^{(k)}-\theta_{j-1}^{(k-1)}}{\theta_{j}^{(k)}-\theta_{j-1}^{(k)}}\frac{\theta_{j}^{(k-1)}-\theta_{j}^{(k)}}{\theta_{j+1}^{(k)}-\theta_{j}^{(k)}}\cdots\frac{\theta_{k-1}^{(k-1)}-\theta_{j}^{[k)}}{\theta_{k}^{(k)}-\theta_{j}^{(k)}}. \] The right-hand side is positive due to the interlacing property of the Ritz values for symmetric tridiagonal matrices. In particular, \begin{equation} \left(s_{k,1}^{(k)}\right)^{2}=\frac{\theta_{1}^{(k-1)}-\theta_{1}^{(k)}}{\theta_{2}^{(k)}-\theta_{1}^{(k)}}\cdots\frac{\theta_{k-1}^{(k-1)}-\theta_{1}^{(k)}}{\theta_{k}^{(k)}-\theta_{1}^{(k)}}.\label{eq:s1k} \end{equation} When the smallest Ritz value $\theta_{1}^{(k)}$ converges to $\lambda_{1}$, this last component squared converges to zero; see also \eqref{eq:bound1}. \subsection{Modification of the tridiagonal matrix} \label{sec:modified} The nodes of the Gauss quadrature are given by the eigenvalues of $T_k$. To obtain the Gauss-Radau quadrature rule one has to consider the problem of finding the coefficient $\alpha_{k+1}^{(\mu)}$ such that the modified matrix \begin{equation} T_{k+1}^{(\mu)}=\left[\begin{array}{ccccc} \alpha_{1} & \beta_{1}\\ \beta_{1} & \ddots & \ddots\\ & \ddots & \ddots & \beta_{k-1}\\ & & \beta_{k-1} & \alpha_{k} & \beta_{k}\\ & & & \beta_{k} & \alpha_{k+1}^{(\mu)} \end{array}\right]\label{eq:Tmu} \end{equation} has the prescribed $\mu$ as an eigenvalue. The nodes of the Gauss-Radau quadrature rule are the eigenvalues of this matrix and the weights are obtained from its eigenvectors. In \cite[pp.~331-334]{Go1973} it has been shown that at iteration $k+1$ \[ \alpha_{k+1}^{(\mu)}\,=\,\mu+\zeta_{k}^{(\mu)} \] where $\zeta_{k}^{(\mu)}$ is the last component of the vector $y$ that solves the linear system \begin{equation} (T_{k}-\mu I)y=\beta_{k}^{2}e_{k}.\label{eq-Tkmu-1} \end{equation} From \cite[Section 3.4]{MeTi2013}, the modified coefficients $\alpha_{k+1}^{(\mu)}$ can be computed recursively using \begin{equation} \alpha_{j+1}^{(\mu)}=\mu+\frac{\beta_{j}^{2}}{\alpha_{j}-\alpha_{j}^{(\mu)}},\qquad\alpha_{1}^{(\mu)}=\mu,\qquad j=1,\dots,k.\label{eq:alphaupdate} \end{equation} Based on the spectral factorization of $T_{k}$, we can now prove the following lemma. \smallskip{} \begin{lemma} \label{lem:alphamu} Let $\mu<\theta_{1}^{(k)}$. Then it holds that \begin{equation} \alpha_{k+1}^{(\mu)}\,=\,\mu+\sum_{i=1}^{k}\eta_{i,k}^{(\mu)},\qquad\eta_{i,k}^{(\mu)}\equiv\frac{\left(\beta_{k}s_{k,i}^{(k)}\right)^{2}}{\theta_{i}^{(k)}-\mu}.\label{eq:alphaspectral} \end{equation} If $\mu<\lambda<\theta_{1}^{(k)}$, then $\alpha_{k+1}^{(\mu)}<\alpha_{k+1}^{(\lambda)}$. Consequently, if $\mu<\theta_{1}^{(k+1)},$ then $\alpha_{k+1}^{(\mu)}<\alpha_{k+1}$. \end{lemma} \smallskip{} \begin{proof} Since $\mu<\theta_{1}^{(k)}$ the matrix $T_{k}-\mu I$ in \eqref{eq-Tkmu-1} is positive definite and, therefore, nonsingular. Hence, \begin{equation} \zeta_{k}^{(\mu)}=e_{k}^{T}y=\beta_{k}^{2}e_{k}^{T}(T_{k}-\mu I)^{-1}e_{k}=\sum_{i=1}^{k}\frac{\left(\beta_{k}s_{k,i}^{(k)}\right)^{2}}{\theta_{i}^{(k)}-\mu}\label{eq:zeta} \end{equation} so that \eqref{eq:alphaspectral} holds. From \eqref{eq:alphaspectral} it is obvious that if $\mu<\lambda<\theta_{1}^{(k)}$, then $\alpha_{k+1}^{(\mu)}<\alpha_{k+1}^{(\lambda)}$. Finally, taking $\lambda=\theta_{1}^{(k+1)}<\theta_{1}^{(k)}$ (because of the interlacing of the Ritz values) we obtain \begin{equation} \alpha_{k+1}^{(\lambda)}=\alpha_{k+1}\,=\,\theta_{1}^{(k+1)}+\zeta_{k},\qquad\zeta_{k}=\sum_{i=1}^{k}\frac{\left(\beta_{k}s_{k,i}^{(k)}\right)^{2}}{\theta_{i}^{(k)}-\theta_{1}^{(k+1)}}.\label{eq:alphasp} \end{equation} If $\mu<\theta_{1}^{(k+1)}=\lambda$, then $\alpha_{k+1}^{(\mu)}<\alpha_{k+1}^{(\lambda)}=\alpha_{k+1}$. \end{proof} \section{The conjugate gradient method and error norm estimation} \label{sec:estimation} When solving a linear system $Ax=b$ with a symmetric and positive definite matrix $A$, the CG method (Algorithm \ref{alg:cg}) \begin{algorithm}[h] \caption{Conjugate gradient algorithm} \label{alg:cg} \begin{algorithmic}[0] \STATE \textbf{input} $A$, $b$, $x_{0}$ \STATE $r_{0}=b-Ax_{0}$ \STATE $p_{0}=r_{0}$ \FOR{$k=1,\dots$ until convergence} \STATE $\gamma_{k-1}=\frac{r_{k-1}^{T}r_{k-1}}{p_{k-1}^{T}Ap_{k-1}}$ \STATE $x_{k}=x_{k-1}+\gamma_{k-1}p_{k-1}$ \STATE $r_{k}=r_{k-1}-\gamma_{k-1}Ap_{k-1}$\hspace*{5mm} \smash{$\left.\begin{array}{@{}c@{}}\\{}\\{}\\{}\\{}\\{}\end{array} \right\} \begin{tabular}{l}{\tt cgiter(k-1)}\end{tabular}$} \STATE $\delta_{k}=\frac{r_{k}^{T}r_{k}}{r_{k-1}^{T}r_{k-1}}$ \STATE $p_{k}=r_{k}+\delta_{k}p_{k-1}$ \ENDFOR \end{algorithmic} \end{algorithm} is the method of choice. In exact arithmetic, the CG iterates $x_{k}$ minimize the $A$-norm of the error over the manifold $x_{0}+\mathcal{K}_{k}(A,r_{0})$, \[ \|x-x_{k}\|_{A}=\min_{y\in x_{0}+\mathcal{K}_{k}(A,r_{0})}\|x-y\|_{A}, \] and the residual vectors $r_{k}=b-Ax_{k}$ are proportional to the Lanczos vectors $v_{j}$, \[ v_{j+1}=(-1)^{j}\frac{r_{j}}{\|r_{j}\|}\,,\qquad j=0,\dots,k. \] Thanks to this close relationship between the CG and Lanczos algorithms it can be shown (see, for instance \cite{B:Me2006}) that the recurrence coefficients computed in both algorithms are connected via $\alpha_{1}=\gamma_{0}^{-1}$ and \begin{equation} \beta_{j}=\frac{\sqrt{\delta_{j}}}{\gamma_{j-1}},\quad\alpha_{j+1}=\frac{1}{\gamma_{j}}+\frac{\delta_{j}}{\gamma_{j-1}},\quad j=1,\dots,k-1.\label{eq:CGLanczos} \end{equation} Writing \eqref{eq:CGLanczos} in matrix form, we find out that CG computes implicitly the $LDL^{T}$ factorization $T_{k}=L_{k}D_{k}L_{k}^{T}$, where \begin{equation} L_{k}=\left[\begin{array}{cccc} 1\\ \sqrt{\delta_{1}} & \ddots\\ & \ddots & \ddots\\ & & \sqrt{\delta_{k-1}} & 1 \end{array}\right],\quad D_{k}=\left[\begin{array}{cccc} \gamma_{0}^{-1}\\ & \ddots\\ & & \ddots\\ & & & \gamma_{k-1}^{-1} \end{array}\right].\label{eq:matrixLDL} \end{equation} Hence the matrix $T_{k}$ is known implicitly in CG. \subsection{Modification of the factorization of the tridiagonal matrix} Similarly as in Section~\ref{sec:modified} we can ask how to modify the Cholesky factorization of $T_{k+1}$, that is available in CG, such that the resulting matrix $T_{k+1}^{(\mu)}$ given implicitly in factorized form has the prescribed eigenvalue $\mu$. In more detail, we look for a coefficient $\gamma_{k}^{(\mu)}$ such that \[ T_{k+1}^{(\mu)}=L_{k+1}\left[\begin{array}{cc} D_{k}\\ & \left(\gamma_{k}^{(\mu)}\right)^{-1} \end{array}\right]L_{k+1}^{T}. \] This problem was solved in \cite{MeTi2013} leading to an updating formula for computing the modified coefficients \begin{equation} \gamma_{j+1}^{(\mu)}=\frac{\gamma_{j}^{(\mu)}-\gamma_{j}}{\mu(\gamma_{j}^{(\mu)}-\gamma_{j})+\delta_{j+1}},\ j=1,\dots,k-1,\qquad\gamma_{0}^{(\mu)}=\frac{1}{\mu}.\label{eq:gammamu} \end{equation} Moreover, $\gamma_{k}^{(\mu)}$ can be obtained directly from the modified coefficient $\alpha_{k+1}^{(\mu)}$, \begin{equation} \gamma_{k}^{(\mu)}=\frac{1}{\alpha_{k+1}^{(\mu)}-\frac{\delta_{k}}{\gamma_{k-1}}},\label{eq:gammalpha} \end{equation} and vice-versa, see \cite[p.~173 and 181]{MeTi2013}. \subsection{The Gauss-Radau upper bound} To summarize some results of \cite{GoMe1994,GoSt1994,StTi2002}, and \cite{MeTi2013,MeTi2014,MeTi2019} related to the Gauss and Gauss-Radau quadrature bounds for the $A$-norm of the error in~CG, it has been shown that \begin{equation} \gamma_{k}\|r_{k}\|^{2}\le\|x-x_{k}\|_{A}^{2}<\gamma_{k}^{(\mu)}\|r_{k}\|^{2}<\left(\frac{\|r_{k}\|^{2}}{\mu\|p_{k}\|^{2}}\right)\|r_{k}\|^{2}\label{eq:basic} \end{equation} for $k<n-1$, and $\mu$ such that $0<\mu\leq\lambda_{1}$. Note that in the special case $k=n-1$ it holds that $\|x-x_{n-1}\|_{A}^{2}=\gamma_{n-1}\|r_{n-1}\|^{2}$. If the initial residual $r_{0}$ has a nontrivial component in the eigenvector corresponding to $\lambda_{1}$, then $\lambda_{1}$ is an eigenvalue of $T_{n}$. If in addition $\mu$ is chosen such that $\mu=\lambda_{1}$, then $\gamma_{n-1}=\gamma_{n-1}^{(\mu)}$ and the second inequality in \eqref{eq:basic} changes to equality. The last inequality is strict also for $k=n-1$. The rightmost bound in \eqref{eq:basic} was derived in \cite{MeTi2019}. The norm $\Vert p_k\Vert$ is not available in CG, but the ratio \[ \phi_{k}=\frac{\left\Vert r_{k}\right\Vert ^{2}}{\left\Vert p_{k}\right\Vert ^{2}} \] can be computed efficiently using \begin{equation} \phi_{j+1}^{-1}=1+\phi_{j}^{-1}\delta_{j+1},\qquad\phi_{0}=1.\label{eq:phiupdate} \end{equation} Note that the accuracy of the bounds can be improved using the formula \begin{equation} \|x-x_{k}\|_{A}^{2}=\sum_{j=k}^{\ell-1}\gamma_{j}\left\Vert r_{k}\right\Vert ^{2}+\|x-x_{\ell}\|_{A}^{2}\label{eq:delay} \end{equation} for a convenient positive integer $\ell\geq k$, by applying the basic bounds \eqref{eq:basic} to the last term in \eqref{eq:delay}; see \cite{MeTi2019} for details on the construction of more accurate bounds. In practice, however, one runs the CG algorithm, and estimates the error in a backward way, i.e., $\ell-k$ iterations back. The adaptive choice of the delay $\ell-k$ when using the Gauss quadrature lower bound was discussed recently in \cite{MePaTi2021}. In the following we will we concentrate on the analysis of the behaviour of the basic Gauss-Radau upper bound \eqref{eq:basic} corresponding to $\ell=k$ \begin{equation} \gamma_{k}^{(\mu)}\|r_{k}\|^{2}\label{eq:GR} \end{equation} in dependence of the choice of $\mu$. As said above, in theory, the closer $\mu$ is to $\lambda_{1}$, the closer to the $A$-norm of the error is the upper bound. However, this is not what is observed in many examples, even in exact arithmetic. In the next section, we construct a model problem and perform numerical experiments that will motivate our analysis. \section{The model problem and a numerical experiment} \label{sec:model} In the construction of the motivating example we use results presented in \cite{Gr1989,GrSt1992,B:Me2006,OLStTi2007,St1991}. Based on the work by Chris Paige \cite{Pa1980a}, Anne Greenbaum \cite{Gr1989} proved that the results of finite precision CG computations can be interpreted (up to some small inaccuracies) as the results of the exact CG algorithm applied to a larger system with the system matrix having many eigenvalues distributed throughout ``tiny'' intervals around the eigenvalues of the original matrix. The experiments show that ``tiny'' means of the size comparable to $\mathbf{u}\|A\|$, where $\mathbf{u}$ is the roundoff unit. This result was used in \cite{GrSt1992} to predict the behavior of finite precision CG. Inspired by \cite{Gr1989,GrSt1992,OLStTi2007} we will construct a linear system $Ax=b$ with similar properties as the one suggested by Greenbaum~\cite{Gr1989}. However, we want to emphasize and visualize some phenomenons concerning the behaviour of the basic Gauss-Radau upper bound \eqref{eq:GR}. Therefore, we choose the size of the intervals around the original eigenvalues larger than $\mathbf{u}\|A\|$. We start with the test problem $\Lambda y=w$ from \cite{St1991}, where $w=m^{-1/2}(1,\dots,1)^{T}$ and $\Lambda=\mathrm{diag}(\hat{\lambda}_{1},\dots,\hat{\lambda}_{m})$, \begin{equation} \hat{\lambda}_{i}=\hat{\lambda}_{1}+\frac{i-1}{m-1}(\hat{\lambda}_{m}-\hat{\lambda}_{1})\rho^{m-i},\quad i=2,\ldots,m.\label{eq:strakos} \end{equation} The diagonal matrix $\Lambda$ and the vector $w$ determine the stepwise distribution function $\omega(\lambda)$ with points of increase $\hat{\lambda}_{i}$ and the individual jumps (weights) $\omega_{j}=m^{-1}$, \begin{equation} \omega(\lambda)\equiv\;\left\{ \;\begin{array}{rcl} 0 & \textnormal{for} & \lambda<\hat{\lambda}_{1}\,,\\[1mm] \sum_{j=1}^{i}\omega_{j} & \textnormal{for} & \hat{\lambda}_{i}\leq\lambda<\hat{\lambda}_{i+1}\,,\quad1\leq i\leq m-1\,,\\[1mm] 1 & \textnormal{for} & \hat{\lambda}_{m}\leq\lambda\,. \end{array}\right.\,\label{eq:schema_hermit} \end{equation} We construct a blurred distribution function $\widetilde{\omega}(\lambda)$ having clusters of points of increase around the original eigenvalues $\hat{\lambda}_{i}$. We consider each cluster to have the same radius $\delta$, and let the number $c_i$ of points in the $i$th cluster grow linearly from 1 to $p$, \[ c_{i}=\mathrm{round}\left(\frac{p-1}{m-1}i+\frac{m-p}{m-1}\right),\quad i=1,\dots,m. \] The blurred eigenvalues \[ \widetilde{\lambda}_{j}^{(i)},\quad j=1,\dots,c_{i}, \] are uniformly distributed in $[\hat{\lambda}_{i}-\delta,\hat{\lambda}_{i}+\delta]$, with the corresponding weights given by \[ \widetilde{\omega}_{j}^{(i)}=\frac{\omega_{i}}{c_{i}}\quad j=1,\dots,c_{i}, \] i.e., the weights that correspond to the $i$th cluster are equal, and their sum is $\omega_{i}$. Having defined the blurred distribution function $\widetilde{\omega}(\lambda)$ we can construct the corresponding Jacobi matrix $T\in\real^{N\times N}$ using the Gragg and Harrod rkpw algorithm \cite{GrHa1984}. To construct the above mentioned Jacobi matrix $T$ we used Matlab's vpa arithmetic with 128 digits. Finally, we define the \emph{double precision data} $A$ and $b$ that will be used for experimenting as \begin{equation} A=\mathrm{double}(T),\quad b=e_{1},\label{eq:Abdouble} \end{equation} where $e_{1}\in\real^{N}$ is the first column of the identity matrix. We decided to use double precision input data from two reasons. First, we can easily compare results of our computations performed in Matlab's vpa arithmetic with results obtained using double precision arithmetic for the same input data. Second, the nonzero structure of the double precision input data allows us to exploit results of \cite{SiTi2022} showing that the Lanczos algorithm applied to $A$ and $b$ having the nonzero structure as in \eqref{eq:Abdouble}, computes exactly. Hence, if we want to simulate exact CG computations, then we can use \cite[Algorithm~3]{SiTi2022}, and not all computations have to be performed using the (slow) Matlab vpa arithmetic. In our experiment we choose $m=12$, $\hat{\lambda}_{1}=10^{-6}$, $\hat{\lambda}_{m}=1$, $\rho=0.8$, $\delta=10^{-10}$, and $p=4$, resulting in $N=30.$ Let us run the ``exact'' CGQ algorithm of \cite{MeTi2013} on the model problem \eqref{eq:Abdouble} constructed above, where exact arithmetic is simulated using Matlab's variable precision with \texttt{digits=128}. Let $\lambda_{1}$ be the exact smallest eigenvalue of $A$. We use four different values of $\mu$ for the computation of the Gauss-Radau upper bound \eqref{eq:GR}: $\mu_{3}=(1-10^{-3})\lambda_{1}$, $\mu_{8}=(1-10^{-8})\lambda_{1}$, $\mu_{16}$ which denotes the double precision number closest to $\lambda_{1}$ such that $\mu_{16}\leq\lambda_{1}$, and $\mu_{50}=(1-10^{-50})\lambda_{1}$ which is almost like the exact value. Note that $\gamma_{k}^{(\mu)}$ is updated using \eqref{eq:gammamu}. \smallskip{} Figure~\ref{fig-0} shows the $A$-norm of the error $\|x-x_{k-1}\|_{A}$ (solid curve) and the upper bounds for the considered values of $\mu_{i}$, $i=3,8,16,50$ (dotted solid curves). The dots represent the values $\theta_{1}^{(k)}-\lambda_{1}$, i.e., the distances of the smallest Ritz values $\theta_{1}^{(k)}$ to $\lambda_{1}$. The horizontal dotted lines correspond to the values of $\lambda_{1}-\mu_{i}$, $i=3,8,16$. \begin{figure}[!htbp] \centering{}\includegraphics[width=12cm]{TGRfig_1} \caption{$\|x-x_{k-1}\|_{A}$, upper bounds and the distance of $\theta_{1}^{(k)}$ to the smallest eigenvalue $\lambda_{1}$, {\tt digits=128}.} \label{fig-0} \end{figure} \medskip{} The upper bounds in Figure~\ref{fig-0} first overestimate, and then closely approximate $\|x-x_{k-1}\|_{A}$ (starting from iteration 5). However, at some point, the upper bounds start to differ significantly from $\|x-x_{k-1}\|_{A}$ and represent much worse approximations, except for $\mu_{50}$. We observe that for a given $\mu_{i}$, $i=3,8,16$, the upper bounds are delayed when the distance of $\theta_{1}^{(k)}$ to $\lambda_{1}$ becomes smaller than the distance of $\mu_{i}$ to $\lambda_{1}$, i.e., when \begin{equation} \theta_{1}^{(k)}-\lambda_{1}<\lambda_{1}-\mu_{i}.\label{eq:first} \end{equation} If \eqref{eq:first} holds, then the smallest Ritz value $\theta_{1}^{(k)}$ is a better approximation to $\lambda_{1}$ than~$\mu_{i}$. This moment is emphasized using vertical dashed lines that connect the value $\theta_{1}^{(k)}-\lambda_{1}$ with $\|x-x_{k-1}\|_{A}$ in the first iteration $k$ such that \eqref{eq:first} holds. Moreover, below a certain level, the upper bounds become almost independent of $\mu_{i}$, $i=3,8,16$, and coincide. The closer is $\mu$ to $\lambda_1$, the later this phenomenon occurs. Depending on the validity of \eqref{eq:first}, we distinguish between \emph{phase~1} and\emph{ phase~2 }of convergence. If the inequality \eqref{eq:first} does not hold, i.e., if $\mu$ is a better approximation to $\lambda_{1}$ than the smallest Ritz value, then we say we are in phase~1; see Figure~\ref{fig:phase1}. If \eqref{eq:first} holds, then the smallest Ritz value is closer to $\lambda_{1}$ than $\mu$ and we are in phase~2; see Figure~\ref{fig:phase2}. Obviously, the beginning of phase 2 depends on the choice of $\mu$ and on the convergence of the smallest Ritz value to the smallest eigenvalue. Note that for $\mu=\mu_{50}$ we are always in phase~1 before we stop the iterations. In the following we would like to explain the phenomenon observed in Figure~\ref{fig-0}. \begin{figure}[th] \centering\unitlength 1mm \centering\begin{picture}(120,15) \put(10,10){\line(1,0){100}}\linethickness{0.1pt} \put(15,11){\line(0,-1){3}} \put(16,5){\makebox(0,0)[ct]{$\mu$}} \linethickness{1.0pt} \put(34,10){\circle*{1.2}} \put(35,5){\makebox(0,0)[ct]{$\lambda_{1}$}} \put(80,10){\circle*{1.2}} \put(81,5){\makebox(0,0)[ct]{$\lambda_{2}$}} \put(95,11){\line(0,-1){3}} \put(96,5){\makebox(0,0)[ct]{$\theta_{1}^{(k)}$}} \end{picture} \caption{Phase~1: $\mu$ is a better approximation to $\lambda_{1}$ than $\theta_{1}^{(k)}$.} \label{fig:phase1} \end{figure} \begin{figure}[th] \unitlength 1mm \centering \begin{picture}(120,15) \put(10,10){\line(1,0){100}}\linethickness{0.1pt} \put(15,11){\line(0,-1){3}} \put(16,5){\makebox(0,0)[ct]{$\mu$}} \linethickness{1.0pt} \put(34,10){\circle*{1.2}} \put(34,5){\makebox(0,0)[ct]{$\lambda_{1}$}} \linethickness{0.1pt} \put(42,11){\line(0,-1){3}} \put(43,5){\makebox(0,0)[ct]{$\theta_{1}^{(k)}$}} \put(80,10){\circle*{1.2}} \put(81,5){\makebox(0,0)[ct]{$\lambda_{2}$}} \put(105,11){\line(0,-1){3}} \put(106,5){\makebox(0,0)[ct]{$\theta_{2}^{(k)}$}} \end{picture} \caption{Phase~2: $\theta_{1}^{(k)}$ is a better approximation to $\lambda_{1}$ than $\mu$.} \label{fig:phase2} \end{figure} \smallskip{} \section{Analysis} The upper bounds are computed from the modified tridiagonal matrices \eqref{eq:Tmu} discussed in Section~\ref{sec:modified}, that differ only in one coefficient at the position $(k+1,k+1)$. Therefore, the first step of the analysis is to understand how the choice of $\mu$ and the validity of the condition \eqref{eq:first} influences the value of the modified coefficient \begin{eqnarray} \alpha_{k+1}^{(\mu)} & = & \mu+\sum_{i=1}^{k}\eta_{i,k}^{(\mu)},\qquad\eta_{i,k}^{(\mu)}=\frac{\left(\beta_{k}s_{k,i}^{(k)}\right)^{2}}{\theta_{i}^{(k)}-\mu};\label{eq:terms} \end{eqnarray} see \eqref{eq:alphaspectral}. We will compare its value to a modified coefficient for which phase~2 does not occur; see Figure~\ref{fig-0} for $\mu_{50}$. Based on that understanding we will then address further important questions. First, our aim is to explain the behaviour of the basic Gauss-Radau upper bound \eqref{eq:GR} in phase~2, in particular, its closeness to the rightmost upper bound in \eqref{eq:basic} that will be called \emph{simple upper bound} in the following. Second, for practical reasons, without knowing $\lambda_{1}$, we would like to be able to detect phase~2, i.e., the first iteration $k$ for which the inequality \eqref{eq:first} starts to hold. Finally, we address the problem of how to improve the accuracy of the basic Gauss-Radau upper bound \eqref{eq:GR} in phase 2. \subsection{The modified coefficient $\alpha_{k+1}^{(\mu)}$} We first analyze the relation between two modified coefficients $\alpha_{k+1}^{(\mu)}$ and $\alpha_{k+1}^{(\lambda)}$ where $0<\mu<\lambda<\theta_{1}^{(k)}.$ \smallskip{} \begin{lemma} \label{lem:lemma2}Let $0<\mu<\lambda<\theta_{1}^{(k)}$. Then \begin{equation} \frac{\eta_{i,k}^{(\lambda)}-\eta_{i,k}^{(\mu)}}{\eta_{i,k}^{(\mu)}}=\frac{\lambda-\mu}{\theta_{i}^{(k)}-\lambda}\label{eq:sensitivity} \end{equation} and \begin{equation} \alpha_{k+1}^{(\lambda)}-\alpha_{k+1}^{(\mu)}=\left(\frac{\lambda-\mu}{\theta_{1}^{(k)}-\mu}\right)\eta_{1,k}^{(\lambda)}+\left(\lambda-\mu\right)E_{k}^{(\lambda,\mu)}\label{eq:diff1} \end{equation} where \begin{equation} E_{k}^{(\lambda,\mu)}\equiv1+\sum_{i=2}^{k}\frac{\eta_{i,k}^{(\lambda)}}{\theta_{i}^{(k)}-\mu}\label{eq:Ek} \end{equation} satisfies $E_{k}^{(\lambda,\mu)}=E_{k}^{(\mu,\lambda)}.$ \end{lemma} \begin{proof} The relation \eqref{eq:sensitivity} follows immediately from the definition of $\eta_{i,k}^{(\mu)}$ and $\eta_{i,k}^{(\lambda)}$, \[ \frac{\eta_{i,k}^{(\lambda)}-\eta_{i,k}^{(\mu)}}{\eta_{i,k}^{(\mu)}}=\frac{\frac{\left(\beta_{k}s_{k,i}^{(k)}\right)^{2}}{\theta_{i}^{(k)}-\lambda}-\frac{\left(\beta_{k}s_{k,i}^{(k)}\right)^{2}}{\theta_{i}^{(k)}-\mu}}{\frac{\left(\beta_{k}s_{k,i}^{(k)}\right)^{2}}{\theta_{i}^{(k)}-\mu}}=\frac{\lambda-\mu}{\theta_{i}^{(k)}-\lambda}. \] Note that $0<\eta_{i,k}^{(\mu)}<\eta_{i,k}^{(\lambda)}$. The difference of the coefficients $\alpha$'s is \begin{eqnarray*} \alpha_{k+1}^{(\lambda)}-\alpha_{k+1}^{(\mu)} & = & \left(\lambda-\mu\right)+\sum_{i=1}^{k}\left(\eta_{i,k}^{(\lambda)}-\eta_{i,k}^{(\mu)}\right)\\ & = & (\lambda-\mu)+\sum_{i=1}^{k}\left(\frac{\left(\beta_{k}s_{k,i}^{(k)}\right)^{2}}{\theta_{i}^{(k)}-\lambda} -\frac{\left(\beta_{k}s_{k,i}^{(k)}\right)^{2}}{\theta_{i}^{(k)}-\mu}\right)\\ & = & (\lambda-\mu)+(\lambda-\mu)\sum_{i=1}^{k}\frac{\left(\beta_{k}s_{k,i}^{(k)}\right)^{2}}{\left(\theta_{i}^{(k)}-\lambda\right) \left(\theta_{i}^{(k)}-\mu\right)}\\ & = & (\lambda-\mu)\frac{\eta_{1,k}^{(\lambda)}}{\theta_{1}^{(k)}-\mu}+(\lambda-\mu) +(\lambda-\mu)\sum_{i=2}^{k}\frac{\eta_{i,k}^{(\lambda)}}{\theta_{i}^{(k)}-\mu} \end{eqnarray*} which implies \eqref{eq:diff1}. Since, \[ \frac{\eta_{i,k}^{(\lambda)}}{\theta_{i}^{(k)}-\mu}=\frac{\eta_{i,k}^{(\mu)}}{\theta_{i}^{(k)}-\lambda}, \] it holds that $E_{k}^{(\lambda,\mu)}=E_{k}^{(\mu,\lambda)}$. \end{proof} \smallskip{} Let us describe in more detail the situation we are interested in. Suppose that $\lambda_{1}$ is well separated from $\lambda_{2}$, and that $\mu$ is a tight underestimate to $\lambda_{1}$ such that \begin{equation} \lambda_{1}-\mu\ll\lambda_{2}-\lambda_{1}.\label{eq:separate} \end{equation} We would like to compare $\alpha_{k+1}^{(\lambda_{1})}$ for which phase~2 does not occur with $\alpha_{k+1}^{(\mu)}$for which phase~2 occurs; see Figure~\ref{fig-0}. Using \eqref{eq:sensitivity} and \eqref{eq:separate}, we are able to compare the individual $\eta$-terms. In particular, for $i>1$ we obtain \[ \frac{\eta_{i,k}^{(\lambda_{1})}-\eta_{i,k}^{(\mu)}}{\eta_{i,k}^{(\mu)}}<\frac{\lambda_{1}-\mu}{\lambda_{2}-\lambda_{1}}\ll1, \] i.e., \[ \eta_{i,k}^{(\lambda_{1})}\approx\eta_{i,k}^{(\mu)}\quad\mbox{for}\quad i>1. \] Hence, $\alpha_{k+1}^{(\mu)}$ can significantly differ from $\alpha_{k+1}^{(\lambda_{1})}$ only in the first term of the sum in \eqref{eq:terms} for which \begin{equation} \frac{\eta_{1,k}^{(\lambda_{1})}-\eta_{1,k}^{(\mu)}}{\eta_{1,k}^{(\mu)}}=\frac{\lambda_{1}-\mu}{\theta_{1}^{(k)}-\lambda_{1}}.\label{eq:compare1} \end{equation} If $\theta_{1}^{(k)}$ is a better approximation to $\lambda_{1}$ than $\mu$ in the sense of \eqref{eq:first}, see Figure~\ref{fig:phase2}, then \eqref{eq:compare1} shows that $\eta_{1,k}^{(\lambda_{1})}$ can be much larger than $\eta_{1,k}^{(\mu)}$. As a consequence, $\alpha_{k+1}^{(\lambda_{1})}$ can differ significantly from $\alpha_{k+1}^{(\mu)}$ . On the other hand, if $\mu$ is chosen such that \[ \lambda_{1}-\mu\ll\theta_{1}^{(k)}-\lambda_{1}, \] for all $k$ we are interested in, then phase 2 will not occur, and \[ \alpha_{k+1}^{(\mu)}\approx\alpha_{k+1}^{(\lambda_{1})}. \] In the following we discuss phase~1 and phase~2 in more detail. \subsection{Phase~1} \label{subsec:Phase1} In phase~1, \[ \lambda_{1}-\mu<\theta_{1}^{(k)}-\lambda_{1}, \] and, therefore, all components $\eta_{i,k}^{(\mu)}$ (including $\eta_{1,k}^{(\mu)}$) are not sensitive to small changes of $\mu$; see \eqref{eq:sensitivity}. In other words, the coefficients $\alpha_{k+1}^{(\mu)}$ are approximately the same for various choices of $\mu$. Let us denote \[ h=\frac{\lambda_{1}-\mu}{\theta_{1}^{(k)}-\lambda_{1}}<1. \] In fact, we can write $\theta_{1}^{(k)}-\mu=\theta_{1}^{(k)}-\lambda_{1}+\lambda_{1}-\mu$ and use the Taylor expansion of $1/(1+h)$. It yields \begin{eqnarray*} \frac{1}{\theta_{1}^{(k)}-\mu} & = & \frac{1}{\theta_{1}^{(k)}-\lambda_{1}}\left(\frac{1}{h+1}\right)=\frac{1}{\theta_{1}^{(k)}-\lambda_{1}}\left[1-h+h^{2}-h^{3}+\cdots\right]. \end{eqnarray*} At the beginning of the iterations, $\theta_{1}^{(k)}-\lambda_{1}$ is large compared to $\lambda_{1}-\mu$, $h$ is small and the right-hand side of $1/(\theta_{1}^{(k)}-\mu)$ is almost given by $1/(\theta_{1}^{(k)}-\lambda_{1})$, independent of $\mu$. Moreover, the last components $s_{k,i}^{(k)}$ squared are not very small and the first term of the sum of the $\eta_{i,k}^{(\mu)}$ is the largest one. \subsection{Phase~2\label{subsec:Phase2}} First recall that for any $0<\mu<\lambda_{1}$ it holds that \begin{equation} \alpha_{k+1}^{(\mu)}<\alpha_{k+1}^{(\lambda_{1})}\quad\mbox{and}\quad\eta_{1,k}^{(\mu)}<\eta_{1,k}^{(\lambda_{1})}.\label{eq:coefficients} \end{equation} As before, suppose that $\lambda_{1}$ is well separated from $\lambda_{2}$ and that \eqref{eq:separate} holds. Phase~2 begins when $\theta_{1}^{(k)}$ is a better approximation to $\lambda_{1}$ than $\mu$, i.e., when \eqref{eq:first} holds. Since $\theta_{1}^{(k)}$ is a tight approximation to $\lambda_{1}$ in phase~2, \eqref{eq:bound1} and \eqref{eq:first} imply that \begin{equation}\label{eq:boundedaway} \eta_{1,k}^{(\lambda_{1})}\geq\lambda_{2}-\theta_{1}^{(k)} = \lambda_{2}-\lambda_{1} + \lambda_1-\theta_{1}^{(k)} > (\lambda_{2}-\lambda_{1}) - (\lambda_1-\mu). \end{equation} Therefore, using \eqref{eq:separate}, $\eta_{1,k}^{(\lambda_{1})}$ is bounded away from zero. On the other hand, \eqref{eq:bound1} also implies that \[ \eta_{1,k}^{(\mu)}=\frac{\theta_{1}^{(k)}-\lambda_{1}}{\theta_{1}^{(k)}-\mu}\eta_{1,k}^{(\lambda_{1})}\leq\frac{\theta_{1}^{(k)}-\lambda_{1}}{\theta_{1}^{(k)}-\mu}\left(\lambda_{n}-\lambda_{1}\right) \] and as $\theta_{1}^{(k)}$ converges to $\lambda_{1}$, $\eta_{1,k}^{(\mu)}$ goes to zero. Therefore, \[ \alpha_{k+1}^{(\mu)}\approx\mu+\sum_{i=2}^{k}\eta_{i,k}^{(\mu)}, \] and the sum on the right-hand side is almost independent of $\mu$. Note that having two values $0<\mu<\lambda<\lambda_{1}$ such that \begin{equation} \theta_{1}^{(k)}-\lambda_{1}<\lambda_{1}-\lambda\quad\mbox{and}\quad\lambda-\mu\ll\lambda_{2}-\lambda_{1},\label{eq:assumption2} \end{equation} then one can expect that \begin{equation} \alpha_{k+1}^{(\mu)}\approx\alpha_{k+1}^{(\lambda)}\label{eq:predict-1} \end{equation} because $\eta_{1,k}^{(\mu)}$ and $\eta_{1,k}^{(\lambda)}$ converge to zero and $\eta_{i,k}^{(\mu)}\approx\eta_{i,k}^{(\lambda)}$ for $i>1$ due to \[ \frac{\eta_{i,k}^{(\lambda)}-\eta_{i,k}^{(\mu)}}{\eta_{i,k}^{(\mu)}}=\frac{\lambda-\mu}{\theta_{i}^{(k)}-\lambda}<\frac{\lambda-\mu}{\lambda_{2}-\lambda_{1}}\ll1, \] where we have used \eqref{eq:sensitivity} and the assumption \eqref{eq:assumption2}. Therefore, $\alpha_{k+1}^{(\mu)}$ is relatively insensitive to small changes of $\mu$ and the same is true for the upper bound \eqref{eq:GR}. \smallskip Let us demonstrate the theoretical results numerically using our model problem. To compute the following results, we, again, use Matlab's vpa arithmetic with 128 digits. \begin{figure}[!htbp] \centering{}\includegraphics[width=12cm]{eta1} \caption{First term $\eta_{1,k}^{(\mu)}$, maximum term $\eta_{i,k}^{(\mu)}$, and the sum $\zeta_{k}^{(\mu)}$ for $\mu=\mu_{3}$.} \label{fig-7} \end{figure} We first consider $\mu=\mu_{3}=(1-10^{-3})\lambda_{1}$ for which we have $\lambda_{1}-\mu=10^{-9}$. The switch from phase 1 to phase 2 occurs at iteration 13. Figure~\ref{fig-7} displays the first term $\eta_{1,k}^{(\mu)}$ and the maximum term $\eta_{i,k}^{(\mu)}$ as well as the sum $\zeta_{k}^{(\mu)}$ defined by~\eqref{eq:zeta}, see Lemma~\ref{lem:alphamu}, as a function of the iteration number $k$. In phase 1 the first term $\eta_{1,k}^{(\mu)}$ is the largest one. As predicted, after the start of phase 2, the first term is decreasing quite fast. \begin{figure}[!htbp] \centering{}\includegraphics[width=12cm]{eta2} \caption{First term $\eta_{1,k}^{(\mu)}$, maximum term $\eta_{i,k}^{(\mu)}$, and the sum $\zeta_{k}^{(\mu)}$ for $\mu=\mu_{8}$.} \label{fig-4} \end{figure} Let us now use $\mu=\mu_{8}=(1-10^{-8})\lambda_{1}$ for which we have $\lambda_{1}-\mu=10^{-14}$. The switch from phase 1 to phase 2 occurs at iteration 15; see Figure~\ref{fig-4}. The conclusions are the same as for $\mu_{3}$. \begin{figure}[!htbp] \centering{}\includegraphics[width=12cm]{eta3} \caption{First term $\eta_{1,k}^{(\mu)}$, maximum term $\eta_{i,k}^{(\mu)}$, and the sum $\zeta_{k}^{(\mu)}$ for $\mu=\mu_{50}$.} \label{fig-3} \end{figure} The behavior of the first term is completely different for $\mu=(1-10^{-50})\lambda_{1}$ which almost corresponds to using the exact smallest eigenvalue $\lambda_{1}$; see Figure~\ref{fig:phase2}. The maximum term of the sum is then almost always the first one; see Figure~\ref{fig-3}. Remember that, for this value of $\mu$, we are always in phase 1. \begin{figure}[!htbp] \centering{}\includegraphics[width=11.5cm]{zetas} \caption{Comparison of the sums $\zeta_{k}^{(\mu_{3})}$, $\zeta_{k}^{(\mu_{8})}$, and $\zeta_{k}^{(\mu_{50})}$.} \label{fig-3-1} \end{figure} Finally, in Figure~\ref{fig-3-1} we present a comparison of the sums $\zeta_{k}^{(\mu)}$ for $\mu_{3}$, $\mu_{8}$, and $\mu_{50}$. We observe that from the beginning up to iteration $12$, all sums visually coincide. Starting from iteration 13 we enter phase~2 for $\mu=\mu_{3}$ and the sum $\zeta_{k}^{(\mu_{3})}$ starts to differ significantly from the other sums, in particular from the ``reference'' term $\zeta_{k}^{(\mu_{50})}$. Similarly, for $k=15$ we enter phase~2 for $\mu=\mu_{8}$ and $\zeta_{k}^{(\mu_{8})}$ and $\zeta_{k}^{(\mu_{50})}$ start to differ. We can also observe that $\zeta_{k}^{(\mu_{3})}$ and $\zeta_{k}^{(\mu_{8})}$ significantly differ only in iterations 13, 14, and 15, i.e., when we are in phase~2 for $\mu=\mu_{3}$ but in phase~1 for $\mu=\mu_{8}$. In all other iterations, $\zeta_{k}^{(\mu_{3})}$ and $\zeta_{k}^{(\mu_{8})}$ visually coincide. \subsection{The coefficient $\alpha_{k+1}$\label{subsec:alpha}} The coefficient $\alpha_{k+1}$ can also be written as \[ \alpha_{k+1}=\alpha_{k+1}^{(\mu)}\quad\mbox{for}\quad\mu=\theta_{1}^{(k+1)}, \] and the results of Lemma~\ref{lem:alphamu} and Lemma~\ref{lem:lemma2} are still valid, even though, in practice, $\mu$ must be smaller than $\lambda_{1}$. Using \eqref{eq:diff1} we can express the differences between the coefficients, it holds that \begin{eqnarray} \alpha_{k+1}-\alpha_{k+1}^{(\lambda_{1})} & = & \eta_{1,k}^{(\lambda_{1})}\frac{\theta_{1}^{(k+1)}-\lambda_{1}}{\theta_{1}^{(k)}-\theta_{1}^{(k+1)}}+\left(\theta_{1}^{(k+1)}-\lambda_{1}\right)E_{k}^{(\theta_{1}^{(k+1)},\lambda_{1})}.\label{eq:diff02} \end{eqnarray} If the smallest Ritz value $\theta_{1}^{(k+1)}$ is close to $\lambda_{1}$, then the second term of the right-hand side in \eqref{eq:diff02} will be negligible in comparison to the first one, since \[ E_{k}^{(\theta_{1}^{(k+1)},\lambda_{1})}=\mathcal{O}(1), \] see \eqref{eq:Ek}, and since $\eta_{1,k}^{(\lambda_{1})}$ is bounded away from zero; see \eqref{eq:boundedaway}. Therefore, one can expect that \begin{equation} \alpha_{k+1}-\alpha_{k+1}^{(\lambda_{1})}\ \approx\ \eta_{1,k}^{(\lambda_{1})}\frac{\theta_{1}^{(k+1)}-\lambda_{1}}{\theta_{1}^{(k)}-\theta_{1}^{(k+1)}}.\label{eq:closeness} \end{equation} The size of the term on the right-hand side is related to the speed of convergence of the smallest Ritz value $\theta_{1}^{(k)}$ to $\lambda_{1}$. Denoting \[ \frac{\theta_{1}^{(k+1)}-\lambda_{1}}{\theta_{1}^{(k)}-\lambda_{1}}=\rho_{k}<1, \] we obtain \[ \frac{\theta_{1}^{(k+1)}-\lambda_{1}}{\theta_{1}^{(k)}-\theta_{1}^{(k+1)}}=\frac{\frac{\theta_{1}^{(k+1)}-\lambda_{1}}{\theta_{1}^{(k)}-\lambda_{1}}}{1-\frac{\theta_{1}^{(k+1)}-\lambda_{1}}{\theta_{1}^{(k)}-\lambda_{1}}}=\frac{\rho_{k}}{1-\rho_{k}}. \] For example, if the convergence of $\theta_{1}^{(k)}$ to $\lambda_{1}$ is superlinear, i.e., if $\rho_{k}\rightarrow0$, then $\alpha_{k+1}$ tends to be very close to $\alpha_{k+1}^{(\lambda_{1})}$. \begin{figure}[!htbp] \centering{}\includegraphics[width=12cm]{alphas} \caption{$\alpha_{k}^{(\mu_{3})}$, $\alpha_{k}^{(\mu_{8})}$, $\alpha_{k}^{(\lambda_{1})}$, and $\alpha_{k}$.} \label{fig:1} \end{figure} In Figure~\ref{fig:1} we plot the coefficients $\alpha_{k}^{(\mu_{3})}$, $\alpha_{k}^{(\mu_{8})}$, $\alpha_{k}^{(\lambda_{1})}$ and $\alpha_{k}$, so that we can compare the observed behaviour with the predicted one. Phase~2 starts for $\mu_{3}$ at iteration 13, and for $\mu_{8}$ at iteration 15; see also Figure~\ref{fig-0}. For $k\leq13$ we observe that \[ \alpha_{k}^{(\mu_{3})}\approx\alpha_{k}^{(\mu_{8})}\approx\alpha_{k}^{(\lambda_{1})} \] as explained in Section~\ref{subsec:Phase1} and $\alpha_k$ is larger. For $k\geq16$, the first terms $\eta_{1,k-1}^{(\mu_{3})}$ and $\eta_{1,k-1}^{(\mu_{8})}$ are close to zero, and, as explained in Section~\ref{subsec:Phase2}, \[ \alpha_{k}^{(\mu_{3})}\approx\alpha_{k}^{(\mu_{8})}. \] For $k=14$ and $k=15$, $\alpha_{k}^{(\mu_{3})}$ and $\alpha_{k}^{(\mu_{8})}$ can differ significantly because $\alpha_{k}^{(\mu_{3})}$ is already in phase~2 while $\alpha_{k}^{(\mu_{8})}$ is still in phase~1. We can also observe that $\alpha_{k}$ can be very close to $\alpha_{k}^{(\lambda_{1})}$ when the smallest Ritz value $\theta_{1}^{(k)}$ is a tight approximation to $\lambda_{1}$, i.e., in later iterations. We know that the closeness of $\alpha_{k}$ to $\alpha_{k}^{(\lambda_{1})}$ depends on the speed of convergence of the smallest Ritz value to $\lambda_{1}$; see \eqref{eq:closeness} and the corresponding discussion. \section{The Gauss-Radau bound in phase~2} Our aim in this section is to investigate the relation between the basic Gauss-Radau upper bound \eqref{eq:GR} and the simple upper bound; see \eqref{eq:basic}. Recall the notation \[ \phi_{k}=\frac{\left\Vert r_{k}\right\Vert ^{2}}{\left\Vert p_{k}\right\Vert ^{2}}; \] see \eqref{eq:phiupdate}. In particular, we would like to explain why the two bounds almost coincide in phase~2. Note that using \eqref{eq:gammalpha} we obtain \begin{equation} \alpha_{k+1}^{(\mu)}=\left(\gamma_{k}^{(\mu)}\right)^{-1}+\frac{\delta_{k}}{\gamma_{k-1}}\label{eq:alphamu} \end{equation} and from \eqref{eq:alphaspectral} it follows \begin{eqnarray*} \alpha_{k+1}^{(\mu)} & = & \mu+\beta_{k}^{2}e_{k}^{T}\left(T_{k}-\mu I\right)^{-1}e_{k},\qquad\beta_{k}^{2}=\frac{1}{\gamma_{k-1}}\frac{\delta_{k}}{\gamma_{k-1}}. \end{eqnarray*} Therefore, \begin{eqnarray} \left(\gamma_{k}^{(\mu)}\right)^{-1} & = & \mu+\beta_{k}^{2}\left(e_{k}^{T}\left(T_{k}-\mu I\right)^{-1}e_{k}-\gamma_{k-1}\right).\label{eq:gammatilde} \end{eqnarray} In the following lemma we find another expression for $e_{k}^{T}\left(T_{k}-\mu I\right)^{-1}e_{k}$. \smallskip{} \begin{lemma} Let $0<\mu<\theta_{1}^{(k)}$. Then it holds that \begin{equation} e_{k}^{T}\left(T_{k}-\mu I\right)^{-1}e_{k}=\gamma_{k-1}+\mu\frac{\gamma_{k-1}^{2}}{\phi_{k-1}}+\sum_{i=1}^{k}\left(\frac{\mu}{\theta_{i}^{(k)}}\right)^{2}\frac{\left(s_{k,i}^{(k)}\right)^{2}}{\theta_{i}^{(k)}-\mu}.\label{eq:omega-1} \end{equation} \end{lemma} \begin{proof} Since $\left\Vert \mu T_{k}^{-1}\right\Vert <1$, we obtain using a Neumann series \begin{eqnarray*} \left(T_{k}-\mu I\right)^{-1}=\left(I-\mu T_{k}^{-1}\right)^{-1}T_{k}^{-1} & = & \left(\sum_{j=0}^{\infty}\mu^{j}T_{k}^{-j}\right)T_{k}^{-1} \end{eqnarray*} so that \[ e_{k}^{T}\left(T_{k}-\mu I\right)^{-1}e_{k}=e_{k}^{T}T_{k}^{-1}e_{k}+\mu e_{k}^{T}T_{k}^{-2}e_{k}+\sum_{j=2}^{\infty}\mu^{j}e_{k}^{T}T_{k}^{-(j+1)}e_{k}. \] We now express the terms on the right-hand side using the CG coefficients and the quantities from the spectral factorization of $T_{k}$. Using $T_{k}=L_{k}D_{k}L_{k}^{T}$ we obtain $e_{k}^{T}T_{k}^{-1}e_{k}=\gamma_{k-1}$. After some algebraic manipulation, see, e.g., \cite[p.~1369]{Me2020} we get \[ T_{k}^{-1}e_{k}=\gamma_{k-1}\Vert r_{k-1}\Vert\left[\begin{array}{c} \frac{(-1)^{k-1}}{\Vert r_{0}\Vert}\\ \vdots\\ \frac{1}{\Vert r_{k-1}\Vert} \end{array}\right] \] so that \[ e_{k}^{T}T_{k}^{-2}e_{k}=e_{k}^{T}T_{k}^{-1}T_{k}^{-1}e_{k}=\gamma_{k-1}^{2}\sum_{i=0}^{k-1}\frac{\Vert r_{k-1}\Vert^{2}}{\Vert r_{i}\Vert^{2}}=\gamma_{k-1}^{2}\frac{\|p_{k-1}\|^{2}}{\|r_{k-1}\|^{2}}=\frac{\gamma_{k-1}^{2}}{\phi_{k-1}}. \] Finally, \begin{eqnarray*} e_{k}^{T}\left(\sum_{j=2}^{\infty}\mu^{j}T_{k}^{-(j+1)}\right)e_{k} & = & e_{k}^{T}S_{k}\left(\sum_{j=2}^{\infty}\mu^{j}\Theta_{k}^{-(j+1)}\right)S_{k}^{T}e_{k} \end{eqnarray*} where the diagonal entries of the diagonal matrix \[ \sum_{j=2}^{\infty}\mu^{j}\Theta_{k}^{-(j+1)} \] have the form \begin{eqnarray*} \frac{1}{\theta_{i}^{(k)}}\left(\frac{\mu}{\theta_{i}^{(k)}}\right)^{2}\sum_{j=0}^{\infty}\left(\frac{\mu}{\theta_{i}^{(k)}}\right)^{j} & = & \frac{1}{\theta_{i}^{(k)}}\left(\frac{\mu}{\theta_{i}^{(k)}}\right)^{2}\frac{1}{1-\frac{\mu}{\theta_{i}^{(k)}}}=\left(\frac{\mu}{\theta_{i}^{(k)}}\right)^{2}\frac{1}{\theta_{i}^{(k)}-\mu}. \end{eqnarray*} Hence, \[ e_{k}^{T}\left(\sum_{j=2}^{\infty}\mu^{j}T_{k}^{-(j+1)}\right)e_{k}=\sum_{i=1}^{k}\left(\frac{\mu}{\theta_{i}^{(k)}}\right)^{2}\frac{\left(s_{k,i}^{(k)}\right)^{2}}{\theta_{i}^{(k)}-\mu}. \] \end{proof} \smallskip{} Based on the previous lemma we can now express the coefficient $\gamma_{k}^{(\mu)}$. \smallskip \begin{theorem} \label{lem:gammamu}Let $0<\mu<\theta_{1}^{(k)}$. Then it holds that \begin{equation} \left(\gamma_{k}^{(\mu)}\right)^{-1}=\frac{\mu}{\phi_{k}}+\sum_{i=1}^{k}\left(\frac{\mu}{\theta_{i}^{(k)}}\right)^{2}\eta_{i,k}^{(\mu)}.\label{eq:omega} \end{equation} \end{theorem} \begin{proof} We start with \eqref{eq:gammatilde}. Using the previous lemma \begin{eqnarray*} \left(\gamma_{k}^{(\mu)}\right)^{-1} & = & \mu+\mu\beta_{k}^{2}\gamma_{k-1}^{2}\phi_{k-1}^{-1}+\beta_{k}^{2}e_{k}^{T}\left(\sum_{j=2}^{\infty}\mu^{j}T_{k}^{-(j+1)}\right)e_{k}\\ & = & \mu\left(1+\delta_{k}\phi_{k-1}^{-1}\right)+\beta_{k}^{2}\sum_{i=1}^{k}\left(\frac{\mu}{\theta_{i}^{(k)}}\right)^{2}\frac{\left(s_{k,i}^{(k)}\right)^{2}}{\theta_{i}^{(k)}-\mu}\\ & = & \mu\phi_{k}^{-1}+\sum_{i=1}^{k}\left(\frac{\mu}{\theta_{i}^{(k)}}\right)^{2}\frac{\left(\beta_{k}s_{k,i}^{(k)}\right)^{2}}{\theta_{i}^{(k)}-\mu}, \end{eqnarray*} where we have used relation \eqref{eq:phiupdate}. \end{proof} \smallskip{} Obviously, using \eqref{eq:omega}, the basic Gauss-Radau upper bound \eqref{eq:GR} and the simple upper bound in \eqref{eq:basic} are close to each other if and only if \begin{equation} \sum_{i=1}^{k}\left(\frac{\mu}{\theta_{i}^{(k)}}\right)^{2}\eta_{i,k}^{(\mu)}\,\ll\,\frac{\mu}{\phi_{k}},\label{eq:negligible} \end{equation} which can also be written as \begin{equation} \left(\frac{\mu}{\theta_{1}^{(k)}}\right)^{2}\frac{\eta_{1,k}^{(\mu)}}{\mu}+\sum_{i=2}^{k}\left(\frac{\beta_{k}s_{k,i}^{(k)}}{\theta_{i}^{(k)}}\right)^{2}\frac{\mu}{\theta_{i}^{(k)}-\mu}\,\ll\,\phi_{k}^{-1}.\label{eq:negligible2} \end{equation} Assuming as before that $\lambda_{1}$ is well separated from $\lambda_{2}$, and that $\mu$ is a tight underestimate to $\lambda_{1}$ in the sense of \eqref{eq:separate}, the sum of terms on the left-hand side of \eqref{eq:negligible2} can be replaced by its tight upper bound \begin{equation} \left(\frac{\lambda_{1}}{\theta_{1}^{(k)}}\right)^{2}\frac{\eta_{1,k}^{(\mu)}}{\mu}+\sum_{i=2}^{k}\left(\frac{\beta_{k}s_{k,i}^{(k)}}{\theta_{i}^{(k)}}\right)^{2}\frac{\lambda_{1}}{\theta_{i}^{(k)}-\lambda_{1}}\label{eq:upper2} \end{equation} which simplifies the explanation of the dependence of the sum in \eqref{eq:negligible2} on $\mu$. The second term in \eqref{eq:upper2} is independent of $\mu$ and its size depends only on the behaviour of the underlying Lanczos process. Here \begin{equation} \left(\frac{\beta_{k}s_{k,i}^{(k)}}{\theta_{i}^{(k)}}\right)^{2}=\frac{\left\Vert A\left(V_{k}s_{:,i}^{(k)}\right)-\theta_{i}^{(k)}\left(V_{k}s_{:,i}^{(k)}\right)\right\Vert ^{2}}{\left(\theta_{i}^{(k)}\right)^{2}}\label{eq:term1} \end{equation} can be seen as the relative accuracy to which the $i$th Ritz value approximates an eigenvalue, and the size of the term \begin{equation} \frac{\lambda_{1}}{\theta_{i}^{(k)}-\lambda_{1}},\qquad i\geq2,\label{eq:term2} \end{equation} depends on the position of $\theta_{i}^{(k)}$ relatively to the smallest eigenvalue. In particular, one can expect that the term \eqref{eq:term2} can be of size $\mathcal{O}(1)$ if $\theta_{i}^{(k)}$ approximates smallest eigenvalues, and it is small if $\theta_{i}^{(k)}$ approximates largest eigenvalues. Using the previous simplifications and assuming phase~2, the basic Gauss-Radau upper bound \eqref{eq:GR} and the rightmost upper bound in \eqref{eq:basic} are close to each other if and only if \begin{equation} \frac{\eta_{1,k}^{(\mu)}}{\mu}+\sum_{i=2}^{k}\left(\frac{\beta_{k}s_{k,i}^{(k)}}{\theta_{i}^{(k)}}\right)^{2}\frac{\lambda_{1}}{\theta_{i}^{(k)}-\lambda_{1}}\,\ll\,\phi_{k}^{-1}.\label{eq:negligible3} \end{equation} From Section~\ref{subsec:Phase2} we know that $\eta_{1,k}^{(\mu)}$ goes to zero in phase~2. Hence, if \begin{equation} \eta_{1,k}^{(\mu)}<\mu,\label{eq:lessmu} \end{equation} which will happen for $k$ sufficiently large, then the first term in \eqref{eq:negligible3} is smaller than the term on the right-hand side. As already mentioned, the sum of positive terms in \eqref{eq:negligible3} depends only on approximation properties of the underlying Lanczos process, that are not easy to predict in general. Inspired by our model problem described in Section~\ref{sec:model}, we can just give an intuitive explanation why the sum could be small in phase~2. Phase~2 occurs in later CG iterations and it is related to the convergence of the smallest Ritz value to the smallest eigenvalue. If the smallest eigenvalue is well approximated by the smallest Ritz value (to a high relative accuracy), then one can expect that many eigenvalues of $A$ are relatively well approximated by Ritz values. If the eigenvalue $\lambda_{j}$ of $A$ is well separated from the other eigenvalues and if it is well approximated by a Ritz value, then the corresponding term \eqref{eq:term1} measuring the relative accuracy to which $\lambda_{j}$ is approximated, is going to be small. \begin{figure}[!htbp] \centering{}\includegraphics[width=12cm]{sumtermsd} \caption{Sizes of terms in \eqref{eq:upper2} and \eqref{eq:negligible2}.} \label{fig:1-1} \end{figure} In particular, in our model problem, the smallest eigenvalues are well separated from each other, and in phase~2 they are well approximated by Ritz values. Therefore, the corresponding terms \eqref{eq:term1} are small. Hence, the Ritz values that did not converge yet in phase~2, are going to approximate eigenvalues in clusters which do not correspond to smallest eigenvalues, i.e., for which the terms \eqref{eq:term2} are small; see also Figure~\ref{fig-4} and Figure~\ref{fig-7}. In our model problem, the sum of positive terms in \eqref{eq:negligible3} is small in phase~2 because either \eqref{eq:term1} or \eqref{eq:term2} are small. Therefore, one can expect that the validity of \eqref{eq:negligible3} will mainly depend on the size of the first term in \eqref{eq:negligible3}; see Figure~\ref{fig:1-1}. The size of the sum of positive terms in \eqref{eq:negligible3} obviously depends on the clustering and the distribution of the eigenvalues, and we cannot guarantee in general that it will be small in phase~2. For example, it need not be small if the smallest eigenvalues of $A$ are clustered. \section{Detection of phase~2} For our model problem it is not hard to detect phase~2 from the coefficients that are available during the computations. We first observe, see Figure~\ref{fig:1-1}, that the coefficients \begin{equation} \gamma_{k}^{(\mu)}\quad\mbox{and}\quad\frac{\phi_{k}}{\mu},\label{eq:bracketing} \end{equation} and the corresponding bounds \eqref{eq:GR} and \eqref{eq:basic} visually coincide from the beginning up to some iteration $\ell_{1}$. From iteration $\ell_{1}+1$, the Gauss-Radau upper bound \eqref{eq:GR} starts to be a much better approximation to the squared $A$-norm of the error than the simple upper bound \eqref{eq:basic}. When phase~2 occurs, the Gauss-Radau upper bound \eqref{eq:GR} loses its accuracy and, starting from iteration $\ell_{2}$ (approximately when \eqref{eq:lessmu} holds), it will again visually coincide with the simple upper bound \eqref{eq:basic}. We observe that phase~2 occurs at some iteration $k$ where the two coefficients \eqref{eq:bracketing} significantly differ, i.e., for $\ell_{1}<k<\ell_{2}.$ To measure the agreement between the coefficients \eqref{eq:bracketing}, we can use the easily computable relative distance \begin{equation} \frac{\frac{\phi_{k}}{\mu}-\gamma_{k}^{(\mu)}}{\gamma_{k}^{(\mu)}}=\phi_{k}\left[\left(\frac{\mu}{\theta_{1}^{(k)}}\right)^{2} \frac{\eta_{1,k}^{(\mu)}}{\mu}+\sum_{i=2}^{k}\left(\frac{\beta_{k}s_{k,i}^{(k)}}{\theta_{i}^{(k)}}\right)^{2}\frac{\mu}{\theta_{i}^{(k)} -\mu}\right].\label{eq:crit1} \end{equation} We will consider this relative distance to be small, if it is smaller than 0.5. \begin{figure}[!htbp] \centering{}\includegraphics[width=12cm]{fig7_1} \caption{The behaviour of the relative distance in \eqref{eq:crit1} for various values of $\mu$.} \label{fig:71} \end{figure} The behavior of the term in \eqref{eq:crit1} for various values of $\mu$ is shown in Figure~\ref{fig:71}. The index $\ell_{1}=12$ is the same for all considered values of $\mu$. For $\mu_{3}$ we get $\ell_{2}=15$ (red circle), for $\mu_{8}$ we get $\ell_{2}=18$ (magenta circle), for $\mu_{16}$ $\ell_{2}=25$ (blue circle), and finally, for $\mu_{50}$ there is no index $\ell_{2}$. As explained in the previous section, in more complicated cases we cannot guarantee in general a similar behaviour of the relative distance \eqref{eq:crit1} as in our model problem. For example, in practical problems we sometimes observe staircase behaviour of the $A$-norm of the error, when few iterations of stagnation are followed by few iterations of rapid convergence. In such cases, the quantity \eqref{eq:crit1} can oscillate several times and it is much more difficult to use it for detecting phase~2. In general we can only say that when the smallest Ritz value is close enough to the smallest eigenvalue, the basic Gauss-Radau upper bound \eqref{eq:GR} can lose its accuracy even if $\mu$ is close to $\lambda_1$, and it usually coincides with the simple upper bound \eqref{eq:basic}. The formulas \eqref{eq:omega} and \eqref{eq:crit1} can be helpful in understanding this behaviour. \section{Improving the accuracy} For integers $\ell\geq k\geq0$, let us denote \[ \Delta_{k}=\gamma_{k}\left\Vert r_{k}\right\Vert ^{2},\quad\Delta_{k:\ell}=\sum_{j=k}^{\ell}\Delta_{j},\quad\mbox{and}\quad\Delta_{k:k-1}=0. \] As described in \cite{MePaTi2021}, the accuracy of lower and upper bounds can be improved using \begin{equation} \varepsilon_{k}=\Delta_{k:\ell-1}+\varepsilon_{\ell},\label{eq:HSid} \end{equation} where $\varepsilon_{j} \equiv \| x - x_j \|_A^2$; see also \eqref{eq:delay}. An improved bound at iteration $k$ is obtained such that the last term in \eqref{eq:HSid} is replaced by the basic lower or upper bounds on $\varepsilon_{\ell}$. In particular, the improved Gauss-Radau upper bound can be defined as \begin{equation} \Omega{}_{k:\ell}^{(\mu)}\,=\,\Delta_{k:\ell-1}+\gamma_{\ell}^{(\mu)}\left\Vert r_{\ell}\right\Vert ^{2},\label{eq:improved} \end{equation} and the improved Gauss lower bound is given by $\Delta_{k:\ell}$. To improve the accuracy of the Gauss-Radau upper bound, we would like to choose $\ell$ such that the relative error of the improved Gauss-Radau upper bound is small enough, i.e., \begin{equation} \frac{\Omega{}_{k:\ell}^{(\mu)}-\varepsilon_{k}}{\varepsilon_{k}}\leq\tau\label{eq:crit1-1} \end{equation} where $\tau$ is a prescribed tolerance, say, $\tau=0.25$. Since \[ \frac{\text{\ensuremath{\Omega{}_{k:\ell}^{(\mu)}}}-\varepsilon_{k}}{\varepsilon_{k}}<\frac{\Omega{}_{k:\ell}^{(\mu)}-\Delta_{k:\ell}}{\Delta_{k:\ell}}=\frac{\left\Vert r_{\ell}\right\Vert ^{2}\left(\gamma_{\ell}^{(\mu)}-\gamma_{\ell}\right)}{\Delta_{k:\ell}}, \] we can require $\ell\geq k$ to be the smallest integer such that \begin{equation} \frac{\left\Vert r_{\ell}\right\Vert ^{2}\left(\gamma_{\ell}^{(\mu)}-\gamma_{\ell}\right)}{\Delta_{k:\ell}}\leq\tau.\label{eq:crit2} \end{equation} If \eqref{eq:crit2} holds, then also \eqref{eq:crit1-1} holds. The just described adaptive strategy for obtaining $\ell$ giving a sufficiently accurate upper bound is summarized in Algorithm~\ref{alg:pseudo}. \begin{algorithm} \caption{CG with the improved Gauss-Radau upper bound} \label{alg:pseudo} \algsetup{indent=2em} \begin{algorithmic}[1] \STATE \textbf{input} $A$, $b$, $x_{0}$, $\mu$, $\tau$ \STATE $r_{0}=b-Ax_{0}$, $p_{0}=r_{0}$ \STATE $k=0$, $\gamma_{0}^{(\mu)}=\frac{1}{\mu}$ \FOR{{$\ell=0,\dots,$}} \STATE \texttt{cgiter($\ell$)} \WHILE{$\ell\geq k$ and \eqref{eq:crit2}} \STATE accept $\Omega_{k:\ell}^{(\mu)}$ \STATE $k=k+1$ \ENDWHILE \STATE $\gamma_{\ell+1}^{(\mu)}=\frac{\gamma_{\ell}^{(\mu)}-\gamma_{\ell}}{\mu(\gamma_{\ell}^{(\mu)}-\gamma_{\ell})+\delta_{\ell+1}}$ \ENDFOR \end{algorithmic} \end{algorithm} Note that \[ \frac{\Omega{}_{k:\ell}^{(\mu)}-\varepsilon_{k}}{\varepsilon_{k}}+\frac{\varepsilon_{k}-\Delta_{k:\ell}}{\varepsilon_{k}}<\frac{\Omega{}_{k:\ell}^{(\mu)}-\Delta_{k:\ell}}{\Delta_{k:\ell}}, \] i.e., if \eqref{eq:crit2} holds, then $\tau$ represents also an upper bound on the sum of relative errors of the improved lower and upper bounds. In other words, if $\ell$ is such that \eqref{eq:crit2} is satisfied, then both the improved Gauss-Radau upper bound as well as the improved Gauss lower bound are sufficiently accurate. Note that the index $\ell$, determined by Algorithm~\ref{alg:pseudo} that focuses on improving the accuracy of the Gauss-Radau upper bound, usually represents an overestimate of the optimal index $\ell$ for the improved Gauss lower bound. For an adaptive heuristic strategy focused on improving the accuracy of the Gauss lower bound, see \cite{MePaTi2021}. In the previous sections we have seen that the basic Gauss-Radau upper bound is delayed, in particular in phase~2. The delay of the basic Gauss-Radau upper bound can be defined as the smallest nonnegative integer $j$ such that \begin{equation} \gamma_{k+j+1}^{(\mu)}\left\Vert r_{k+j+1}\right\Vert ^{2}<\varepsilon_{k}.\label{eq:udelay} \end{equation} Having sufficiently accurate lower and upper bounds (e.g., if \eqref{eq:crit2} is satisfied), we can approximately determine the delay of the basic Gauss-Radau upper bound as the smallest $j$ satisfying \eqref{eq:udelay} where $\varepsilon_{k}$ in \eqref{eq:udelay} is replaced by its tight lower bound $\Delta_{k:\ell}$ . \begin{figure}[!htbp] \centering{}\includegraphics[width=12cm]{model-strakos12} \caption{Model problem and $\mu_{3}=\lambda_{1}(1-10^{-3})$.} \label{fig:81-1} \end{figure} \begin{figure}[!htbp] \centering{}\includegraphics[width=12cm]{bcsstk01} \caption{Matrix \texttt{bcsstk01} and $\mu_{4}=\lambda_{1}(1-10^{-4})$.} \label{fig:81} \end{figure} The strategy for improving the accuracy of the Gauss-Radau upper bound implemented in Algorithm~\ref{alg:pseudo} is demonstrated in Figure~\ref{fig:81-1} and Figure~\ref{fig:81}. In Figure~\ref{fig:81-1} we consider our model problem described in Section~\ref{sec:model} and use Matlab's vpa arithmetic with 128 digits. In Figure~\ref{fig:81} we consider the matrix \texttt{bcsstk01} of order $48$ from the SuiteSparse Matrix collection and the unit norm right-hand side $b$ that has equal components in the eigenvector basis, choose $\mu_{4}=\lambda_{1}(1-10^{-4})$, and run the experiment in the standard double precision arithmetic using MATLAB R2019b. In the top parts of figures we plot the $A$-norm of the error (dotted), the square root of the basic Gauss-Radau upper bound \eqref{eq:GR} (dashed), and the square root of the improved Gauss-Radau upper bound \eqref{eq:improved} with $\ell$ satisfying \eqref{eq:crit2}. We can see that the improved upper bound is visually the same as the $A$-norm of the error. In the middle parts we plot the relative error \begin{equation} \frac{\Omega{}_{k:\ell}^{(\mu)}-\varepsilon_{k}}{\varepsilon_{k}}\label{eq:relerr} \end{equation} (solid) together with the prescribed tolerance $\tau=0.25$ (dotted) where $\ell$ was chosen using \eqref{eq:crit2}. For comparison we also plot the relative relative error \eqref{eq:relerr} with the ``ideal'' value of $\ell$, i.e., with the smallest integer $\ell\geq k$ such that \eqref{eq:crit1-1} is satisfied. Note that the ideal value of $\ell$ was determined using the quantities $\varepsilon_{k}$ that are not available in practical computations. Finally, in the bottom parts of Figure~\ref{fig:81-1} and Figure~\ref{fig:81} we plot the difference $\ell-k$ at the individual iterations $k$, where $\ell$ was determined using \eqref{eq:crit2} (solid) or using \eqref{eq:crit1-1} (dashed). We observe that in later iterations, the values of $\ell$ satisfying the criterion \eqref{eq:crit2} are almost ideal (optimal). \section{Conclusions} In this paper we discussed and analyzed the behaviour of the Gauss-Radau upper bound on the $A$-norm of the error in CG. In particular, we concentrated on the phenomenon observed during computations showing that, in later CG iterations, the upper bound loses its accuracy, it is almost independent of $\mu$, and visually coincides with the simple upper bound. We explained that this phenomenon is closely related to the convergence of the smallest Ritz value to the smallest eigenvalue of $A$. It occurs when the smallest Ritz value is a better approximation to the smallest eigenvalue than the prescribed underestimate $\mu$. We developed formulas that can be helpful in understanding this behavior, and suggested an adaptive strategy of how to improve the accuracy of the Gauss-Radau upper bound. Note that the loss of accuracy of the Gauss-Radau upper bound is not directly linked to rounding errors in computations of the bound, but it is related to the finite precision behaviour of the underlying Lanczos process. In more detail, the phenomenon can occur when solving linear systems with clustered eigenvalues. However, the results of finite precision CG computations can be seen (up to some small inaccuracies) as the results of the exact CG algorithm applied to a larger system with the system matrix having clustered eigenvalues. Therefore, one can expect that the discussed phenomenon can occur in practical computations not only when $A$ has clustered eigenvalues, but also whenever orthogonality is lost in the CG algorithm.

c3411abfe97a85b3833dbf14b460ddb86c137c46

\section{Introduction} \label{sec:introduction} The linear-quadratic (LQ) optimal control problem originated by \citet{Kalman1960} is of great importantance in the field of optimal control. The SLQ optimal control problem, pioneered by \citet{Wonham1968}, has been widely considered in the previous literatures \citep{YongZhou1999,SunLiYong2016,SunYong2018,ZhangYan2020}. It is well-known that a conventional way to work out the SLQ problem is to solve the corresponding SARE. However, owing to the nonlinear property of the SARE, it is difficult to obtain its analytical solution. Over the past two decades, researchers have turned to investigate numerical solution to the corresponding SARE of their problems. For example, \citet{WuSunZhang2018} proposed two iterative algorithms to solve an SARE arising in SLQ optimal control problems subject to state-dependent noise. \citet{FengAnderson2010} developed a strategy to study a class of state-perturbed SARE in LQ zero-sum games. With the help of the linear matrix inequalities, \citet{RamiZhou2000} obtained the maximal solution of an SARE for a continuous-time indefinite SLQ problems on infinite horizon. In literatures mentioned above, all parameters of their systems need to be used to solve the corresponding SARE. However, the system coefficients may not be completely known in the real world, especially in applications such as finance and engineering. Therefore, it is valuable to solve the SARE with partially model-free systems, i.e., with partial information of the system coefficient matrices. Recently, the techniques of adaptive dynamic programming (ADP) \citep{Werbo1974} and reinforcement learning \citep{SuttonBarto1998} have been widely used to tackle control problems with model-free or partially model-free system dynamics. For example, about deterministic discrete-time problems, \citet{ZhaoZhang2021} applied the method of Q-learning to solve a discrete-time optimal control problem with unknown system dynamics. In \citet{Tamimi2007}, an optimal strategy was obtained for a class of linear model-free zero-sum games by the method of Q-learning. Regarding to deterministic continuous-time problems, \citet{Vrabie2009} introduced a policy iteration algorithm to investigate partially model-free LQ optimal control problems. By virtue of ADP, \citet{JiangJiang2012} studied a kind of deterministic continuous-time LQ problems with completely unknown dynamics. Based on the ADP approach, \citet{WangSuiTong2017} got an optimal output feedback control for model-free continuous-time nonlinear systems with actuator saturation. As for the stochastic case, \citet{ChenWang2021} obtained an optimal control for a kind of model-free stochastic discrete-time systems by the theory of ADP. Without knowing the information of drift term, \citet{Duncan1999} obtained an adaptive linear-quadratic Gaussian control for a class of linear systems where the diffusion term does not rely on the control and state. Recently, \citet{LiXu2020} proposed a partially model-free policy iteration method to solve a kind of continuous-time SLQ problems on infinite horizon, without using system matrix $A$ (see equation (\ref{eq2}) in Section 2 for the system dynamics). Inspired by the above work, especially \citet{JiangJiang2012} and \citet{LiXu2020}, we propose a novel data-driven policy iteration strategy to work out the infinite-horizon continuous-time SLQ optimal control problem. The main feature of this algorithm is that it does not rely on the information of three coefficient matrices $A, B, C$. Specifically, we employ the ADP technique to iteratively solve the corresponding SARE by utilizing the input and state data. Finally, we provide a simulation example to validate the proposed algorithm. The rest of this paper is organized as follows. In Section \ref{sec2}, the formulation of the SLQ problem is introduced and some preliminaries are given. In Section \ref{sec3}, the data-driven algorithm is developed in detail. Section \ref{sec4} provides a numerical example to validate the data-driven algorithm. Finally, some concluding remarks are given in Section \ref{sec5}. \vspace{2mm} \noindent{\bf Notations.} We denote by $\mathbb{R}$ the set of real numbers, by $\mathbb{Z}^+$ the set of non-negative integers, respectively. The collection of all $p\times q$ real matrices is denoted by $\mathbb{R}^{p\times q}$. $\mathbb{R}^{p}$ represents the $p$-dimensional Euclidean space and $|\cdot|$ is the Euclidean norm for matrix or vector of proper size. For simplicity, we denote zero matrix (or vector) by 0. $diag(l)$ denotes a square diagonal matrix with the elements of vector $l$ on the main diagonal. $M^{T}$ is the transpose of a vector or matrix $M$. We use $\textbf{S}^{p}$, $\textbf{S}^{p}_{+}$ and $\textbf{S}^{p}_{++}$ to denote the collection of all symmetric matrices, positive semidefinite matrices and positive definite matrices in $\mathbb{R}^{p\times p}$, respectively. Moreover, if a matrix $E\in \textbf{S}^{p}_{++}$ (resp. $E\in\textbf{S}^{p}_{+}$) is positive definite (resp. positive semidefinite), we usually write $E>0$ (resp. $\geqslant 0$). If matrices $E\in \textbf{S}^{p}$, $F\in \textbf{S}^{p}$, then we write $E\geqslant F$ (resp. $E>F$) if $E-F\geqslant0$ (resp. $E-F> 0$). ($\Omega$,\,$\mathbb{F}$,\,$\{\mathbb{F}_{t}\}_{t\geqslant 0}$,\,$\mathbb{P}$) is a filtered probability space that satisfies usual conditions, on which a one-dimensional standard Brownian motion $W(\cdot)$ is defined. We define space $L^2_{\mathbb{F}}(\mathbb{R}^{n})$ as \begin{equation*} \begin{split} L^2_{\mathbb{F}}(\mathbb{R}^{n}):=\bigg\{\psi (\cdot):[0,+\infty)\times \Omega\ \to \mathbb{R}^{n}\bigg|&\psi(\cdot) \,\,\text{is}\,\, \mathbb{F}_{t}-\text{adapted,\,\,measureable,}\\ &\text{\,\,and}\,\, \mathbb{E}\int_{0}^{\infty} |\psi(s,\omega )|^2ds<\infty \bigg\}\\ \end{split} \end{equation*} and its norm is defined as \begin{equation*} \left \|\psi (\cdot) \right \|:=\big(\mathbb{E}\int_{0}^{\infty} |\psi(s,\omega )|^2ds\big)^{\frac{1}{2}}. \end{equation*} Furthermore, $\otimes$ denotes the Kronecker product. For any matrix $F$, $vec(F)$ is a vectorization map from the matrix $F$ into a column vector of proper size, which stacks the columns of $F$ on top of one another, i.e., \begin{equation*} vec \begin{bmatrix} f_{11}& f_{12}\\ f_{21}& f_{22}\\ f_{31}& f_{32}\\ \end{bmatrix}:=(f_{11},f_{21},f_{31},f_{12},f_{22},f_{32})^T. \end{equation*} \section{Problem formulation and some preliminaries}\label{sec2} In this section, the SLQ optimal control problem and some preliminaries will be presented. Moreover, some assumptions are given to ensure the well-posedness of the SLQ problem. Consider a stochastic linear system \begin{equation} \label{eq2} \begin{cases} dX(s)=[AX(s)+Bv(s)]ds+[CX(s)+Dv(s)]dW(s),s\in[0,\infty),\\ X(0)=x_0, \end{cases} \end{equation} where $A, C\in\mathbb{R}^{n\times n}$, $B, D\in\mathbb{R}^{n\times m}$ are given constant matrices and $x_0\in \mathbb{R}^{n}$. The cost functional adopted in this paper is \begin{equation}\label{eq3} \begin{split} J(v(\cdot))=\mathbb{E}\int_{0 }^{\infty}[X(s)^TQX(s)+2v(s)^TSX(s)+v(s)^TRv(s)]ds, \end{split} \end{equation} where $Q\in\mathbb{R}^{n\times n}$, $S\in\mathbb{R}^{m\times n}$ and $R\in\mathbb{R}^{m\times m}$ are constant matrices. Now we give the definition of $L^2$-stabilizability, which is indespensable for the well-posedness of infinite-horizon SLQ problems. \vspace{2mm} \noindent{\bf Definition 1.} System (\ref{eq2}) is called $L^2$-stabilizable if there exists a matrix $K\in \mathbb{R}^{m\times n} $ such that, for any initial state $x_0$, the solution of \begin{equation} \label{eq4} \begin{cases} dX(s)=(A+BK)X(s)ds+(C+DK)X(s)dW(s),s\in[0,\infty),\\ X(0)=x_0 \end{cases} \end{equation} satisfies $\lim \limits_{s \to \infty}\mathbb{E}[X(s)^TX(s)]=0$. In this case, the feedback control $v(\cdot)=KX(\cdot)$ is called stabilizing and the matrix $K$ is called a stabilizer of system (\ref{eq2}). \vspace{2mm} \noindent{\bf Assumption 1.} System (\ref{eq2}) is $L^2$-stabilizable. \vspace{2mm} We define \begin{equation*} \mathcal{V}_{ad}:=\{v(\cdot)\in L^2_{\mathbb{F}}(\mathbb{R}^{m})|v(\cdot) \,\, \text{is\, stabilizing}\} \end{equation*} as an admissible control set. The continuous-time SLQ problem is given as follows: \vspace{2mm} \noindent{\bf Problem (SLQ).} For given $x_0\in \mathbb{R}^{n}$, we want to find an optimal control $v^*(\cdot) \in \mathcal{V}_{ad}$ such that \begin{equation*} J(v^*(\cdot))=\inf \limits_{v(\cdot) \in \mathcal{V}_{ad}}J(v(\cdot)). \end{equation*} When $\inf _{v(\cdot) \in \mathcal{V}_{ad}}J(v(\cdot))>-\infty $ is satisfied for any $x_0\in \mathbb{R}^{n}$, Problem (SLQ) is called well-posed. Moreover, the control $v^*(\cdot)$ that achieves $\inf _{v(\cdot) \in \mathcal{V}_{ad}}J(v(\cdot))$ is called \emph{optimal control} and the corresponding trajectory $X^*(\cdot)$ is called \emph{optimal trajectory}. Based on the main results of \citet{LiXu2020}, we introduce the following assumption. \vspace{2mm} \noindent{\bf Assumption 2.} $R>0$ and $Q-S^TR^{-1}S > 0$. \vspace{2mm} Therefore, for any $x_0\in \mathbb{R}^n$, Problem (SLQ) is a well-posed problem under Assumptions 1 and 2. \section{ Data-driven algorithm for the SLQ optimal control Problem }\label{sec3} In this section, we will introduce a data-driven algorithm to solve Problem (SLQ), which does not need the knowledge of coefficient matrices $A$, $B$, $C$ in system (\ref{eq2}). Before giving the algorithm, we first present an iterative method to solve Problem (SLQ). For the proof, please see Lemma 2.3 and Theorems 2.1-2.2 in \citet{LiXu2020}. \vspace{2mm} \noindent{\bf Lemma 1.} Suppose $K_0$ is a stabilizer of system (\ref{eq2}) and $P_{i+1} \in \textbf{S}^{n}_{++}$ is the solution of \begin{equation}\label{eq5} \begin{split} P_{i+1}(A+BK_i)+(A+BK_i)^TP_{i+1} &+(C+DK_i)^TP_{i+1}(C+DK_i)\\ &+K_i^TRK_i +S^TK_i+K_i^TS+Q=0, \end{split} \end{equation} where $K_{i+1}$, $i=0,1,2,\cdots$, are updated by \begin{equation}\label{eq6} \begin{split} K_{i+1}=-(R+D^TP_{i+1}D)^{-1}(B^TP_{i+1}+D^TP_{i+1}C+S). \end{split} \end{equation} Then $P_i$ and $K_i$, $i=1,2,3,\cdots$ can be uniquely determined, and the following conclusions hold: (i) every element of $\{K_i\}_{i=0}^{\infty}$ is a stabilizer of system (\ref{eq2}). (ii) $P^* \leq P_{i+1} \leq P_i $, $i=1,2,3,\cdots$. (iii)$\lim \limits_{i \to \infty}P_i=P^*$, $\lim \limits_{i \to \infty}K_i=K^*$, where $K^*=-(R+D^TP^*D)^{-1}(D^TP^*C+B^TP^*+S)$ and $P^*$ is the solution to the SARE \begin{equation}\label{eq111} \begin{split} PA+A^TP&+C^TPC+Q\\ &-(C^TPD+PB+S^T)(R+D^TPD)^{-1}(D^TPC+B^TP+S)=0. \end{split} \end{equation} Moreover, $ v^*(\cdot)=K^*X^*(\cdot)$ is an optimal control of Problem (SLQ). \vspace{2mm} Though Lemma 1 presents an approximation method to solve SARE (\ref{eq111}), solving $P_{i+1} $ and $K_{i+1}$ from equation (\ref{eq5}) and (\ref{eq6}) requires all information of the system coefficient matrices. As noted in the previous section, it is hard to obtain all information of the system parameters in the real world. In the sequel, we will propose a data-driven algorithm to solve them with partial knowledge of system (\ref{eq2}). In order to get our data-driven algorithm, system (\ref{eq2}) is rewritten as \begin{equation}\label{eq11} \begin{cases} \begin{split} dX(s)=&\big[A_iX(s)+B\big(v(s)-K_iX(s)\big)\big]ds\\ &+\big[C_iX(s)+D\big(v(s)-K_iX(s)\big)\big]dW(s),s\in[0,\infty),\\ \end{split}\\ X(0)=x_0, \end{cases} \end{equation} where $A_i=A+BK_i$ and $C_i=C+DK_i$. Then (\ref{eq5}) can be transformed to \begin{equation}\label{eq7} \begin{split} A_i^TP_{i+1}+P_{i+1}A_i+C_i^TP_{i+1}C_i+Q_i=0, \end{split} \end{equation} where $Q_i=S^TK_i+K_i^TRK_i+K_i^TS+Q$. Now we give the next lemma to illustrate some relationship between $P_{i+1}$ and $K_{i+1}$, $i=0,1,2,\cdots$, generated from (\ref{eq5}) and (\ref{eq6}). \vspace{2mm} \noindent{\bf Lemma 2.} For any $K_i$, $i=0,1,2,\cdots$, $P_{i+1}$ and $K_{i+1}$ generated from (\ref{eq5}) and (\ref{eq6}) satisfy the following equation \begin{equation}\label{eq10} \begin{split} &\mathbb{E}\big[X(t+\triangle t)^TP_{i+1}X(t+\triangle t)-X(t)^TP_{i+1}X(t)\big]\\ &+2\mathbb{E}\int_{t}^{t+\triangle t}\big(v(s)-K_iX(s)\big)^TM_{i+1}X(s)ds\\ &-\mathbb{E}\int_{t}^{t+\triangle t}v(s)^TD^TP_{i+1}Dv(s)ds +\mathbb{E}\int_{t}^{t+\triangle t}X(s)^TK_i^TD^TP_{i+1}DK_iX(s)ds\\ =&-\mathbb{E}\int_{t}^{t+\triangle t}X(s)^TQ_iX(s)ds -2\mathbb{E}\int_{t}^{t+\triangle t}\big(v(s)-K_iX(s)\big)^TSX(s)ds, \end{split} \end{equation} where $M_{i+1}=(R+D^TP_{i+1}D)K_{i+1}$, $\triangle t$ is any positive real number, $t\geq 0$ and $X(\cdot)$ is the trajectory of system (\ref{eq11}) with any control $v(\cdot)$. \vspace{2mm} \noindent{\bf Proof.} By Ito's formula and (\ref{eq11}), one gets \begin{equation}\label{eq8} \begin{split} &d\big(X(s)^TP_{i+1}X(s)\big)\\ =\bigg\{&X(s)^T\big[A_i^TP_{i+1}+P_{i+1}A_i+C_i^TP_{i+1}C_i\big]X(s)\\ &+2\big(v(s)-K_iX(s)\big)^T\big(B^TP_{i+1}+D^TP_{i+1}C_i\big)X(s)\\ &+\big(v(s)-K_iX(s)\big)^TD^TP_{i+1}D\big(v(s)-K_iX(s)\big)\bigg\}ds +\bigg\{\cdots\bigg\}dW(s)\\ \end{split} \end{equation} \begin{equation*} \begin{split} =&\bigg\{X(s)^T\big[A_i^TP_{i+1}+P_{i+1}A_i+C_i^TP_{i+1}C_i\big]X(s)\\ &+2\big(v(s)-K_iX(s)\big)^T\big(B^TP_{i+1}+D^TP_{i+1}C+D^TP_{i+1}DK_i+S-S\big)X(s)\\ &+\big(v(s)-K_iX(s)\big)^TD^TP_{i+1}D\big(v(s)-K_iX(s)\big)\bigg\}ds+\bigg\{\cdots\bigg\}dW(s)\\ =&\bigg\{X(s)^T\big[A_i^TP_{i+1}+P_{i+1}A_i+C_i^TP_{i+1}C_i\big]X(s)\\ &+2\big(v(s)-K_iX(s)\big)^T\big(B^TP_{i+1}+D^TP_{i+1}C+S\big)X(s)\\ &+\big[v(s)^TD^TP_{i+1}Dv(s)-X(s)^TK_i^TD^TP_{i+1}DK_iX(s)\\ &-2\big(v(s)-K_iX(s)\big)^TSX(s)\big]\bigg\}ds + \bigg\{\cdots\bigg\}dW(s).\\ \end{split} \end{equation*} Then it follows from \eqref{eq6} and \eqref{eq7} that\\ \begin{equation*} \begin{split} B^TP_{i+1}+D^TP_{i+1}C+S=-(R+D^TP_{i+1}D)K_{i+1}, \end{split} \end{equation*} \begin{equation*} \begin{split} A_i^TP_{i+1}+P_{i+1}A_i+C_i^TP_{i+1}C_i=-Q_i. \end{split} \end{equation*} Inserting them into \eqref{eq8}, we know \begin{equation}\label{e} \begin{split} &d(X(s)^TP_{i+1}X(s))\\ =&-\bigg\{X(s)^TQ_iX(s)\bigg\}ds-\bigg\{2\big(v(s)-K_iX(s)\big)^T\big(R+D^TP_{i+1}D\big)K_{i+1}X(s)\bigg\}ds\\ &+\bigg\{\big[v(s)^TD^TP_{i+1}Dv(s)-X(s)^TK_i^TD^TP_{i+1}DK_iX(s)\\&\quad-2\big(v(s)-K_iX(s)\big)^TSX(s)\big]\bigg\}ds + \bigg\{\cdots\bigg\}dW(s).\\ \end{split} \end{equation} Thus, integrating from $t$ to $t+\triangle t$ and taking expection $\mathbb{E}$ on both sides of (\ref{e}), we get (\ref{eq10}). The proof is completed. $\hfill\blacksquare$ \vspace{2mm} Next, we define some symbols that will be frequently used in the sequel. For any \begin{equation*} X=[x_1,x_2,x_3,\cdots,x_n]\in\mathbb{R}^n, \end{equation*} and \begin{equation*} P=\begin{bmatrix} p_{11}& p_{12}&\cdots&p_{1n}\\ p_{12}& p_{22}&\cdots&p_{2n}\\ \vdots&\vdots& & \vdots\\ p_{1n}& p_{2n}&\cdots&p_{nn}\\ \end{bmatrix}\in\mathbf{S}^n, \end{equation*} we define \begin{equation*} vech(P)\,:=[p_{11}, 2p_{12},\cdots,2p_{1n},p_{22}, 2p_{23},\cdots,2p_{n-1,n},p_{nn}]^T, \end{equation*} \begin{equation*} \overline{X}:=[x_1^2, x_1x_2,\cdots,x_1x_n,x_2^2, x_2x_3,\cdots,x_{n-1}x_n,x_n^2]^T, \end{equation*} where $p_{ij}$, $i,j=1,2,3,\cdots$, is the $(i,j)$th element of $P$ and $x_i$, $i=1,2,3,\cdots$, is the $i$th element of $X$. By Kronecker product theory, if $D$, $E$ and $F$ are matrices of proper sizes, $P$ is any symmetric matrix and $\theta$ is any column vector, we have \begin{equation*} \begin{split} vec(DEF)=(F^T\otimes D)vec(E),\,\,\, E^T\otimes F^T=(E\otimes F)^T, \end{split} \end{equation*} \begin{equation*} \theta^TP\theta=vec(\theta^TP\theta)=(\theta^T\otimes \theta^T)vec(P)=\bar\theta^T vech(P). \end{equation*} Thus, in (\ref{eq10}), noting that $Dv(s)$ and $DK_iX(s)$ are two column vectors and $P_{i+1}\in\mathbf{S}^n$, one gets \begin{equation*} v(s)^TD^TP_{i+1}Dv(s)=\overline{Dv(s)}^T vech(P_{i+1}), \end{equation*} \begin{equation*} X(s)^TK_i^TD^TP_{i+1}DK_iX(s)=\overline{DK_iX(s)}^T vech(P_{i+1}). \end{equation*} Similarly, from (\ref{eq10}) and the above notations, for any $l \in \mathbb{Z^{+}}$, we know \begin{small} \begin{equation}\label{eq22} \begin{split} &\begin{pmatrix} \mathbb{E}\big[\overline{X(t_1)}^T -\overline{X(t_0)}^T\big]-\mathbb{E}\big[\int_{t_0}^{t_1}\overline{Dv(s)}^Tds\big]+\mathbb{E}\big[\int_{t_0}^{t_1}\overline{DK_iX(s)}^Tds\big]\\ \mathbb{E}\big[\overline{X(t_2)}^T-\overline{X(t_1)}^T\big]-\mathbb{E}\big[\int_{t_1}^{t_2}\overline{Dv(s)}^Tds\big]+\mathbb{E}\big[\int_{t_1}^{t_2}\overline{DK_iX(s)}^Tds\big]\\ \vdots\\ \mathbb{E}\big[\overline{X(t_l)}^T-\overline{X(t_{l-1})}^T\big]-\mathbb{E}\big[\int_{t_{l-1}}^{t_l}\overline{Dv(s)}^Tds\big]+\mathbb{E}\big[\int_{t_{l-1}}^{t_l}\overline{DK_iX(s)}^Tds\big]\\ \end{pmatrix}\times vech(P_{i+1})\\ &+2\begin{pmatrix} \begin{bmatrix} \mathbb{E}\big[\int_{t_0}^{t_1}X^T(s)\otimes v^T(s)ds\big]\\ \mathbb{E}\big[\int_{t_1}^{t_2}X^T(s)\otimes v^T(s)ds\big]\\ \vdots\\ \mathbb{E}\big[\int_{t_{l-1}}^{t_l}X^T(s)\otimes v^T(s)ds\big]\\ \end{bmatrix} -\begin{bmatrix} \mathbb{E}\big[\int_{t_0}^{t_1}X^T(s)\otimes X^T(s)ds\big]\\ \mathbb{E}\big[\int_{t_1}^{t_2}X^T(s)\otimes X^T(s)ds\big]\\ \vdots\\ \mathbb{E}\big[\int_{t_{l-1}}^{t_l}X^T(s)\otimes X^T(s)ds\big]\\ \end{bmatrix}\times (I_n \otimes K_i^T) \end{pmatrix}\times vec\big(M_{i+1}\big)\\ =&\begin{pmatrix} \mathbb{E}\big[\int_{t_0}^{t_1}X^T(s)\otimes X^T(s)ds\big]\\ \mathbb{E}\big[\int_{t_1}^{t_2}X^T(s)\otimes X^T(s)ds\big]\\ \vdots\\ \mathbb{E}\big[\int_{t_{l-1}}^{t_l}X^T(s)\otimes X^T(s)ds\big]\\ \end{pmatrix} \times vec\big(-Q_i+2K_i^TS\big) -2\begin{pmatrix} \mathbb{E}\big[\int_{t_0}^{t_1}X^T(s)\otimes v^T(s)ds\big]\\ \mathbb{E}\big[\int_{t_1}^{t_2}X^T(s)\otimes v^T(s)ds\big]\\ \vdots\\ \mathbb{E}\big[\int_{t_{l-1}}^{t_l}X^T(s)\otimes v^T(s)ds\big]\\ \end{pmatrix} \times vec(S),\\ \end{split} \end{equation} \end{small} where $0\leq t_0<t_1<t_2<\cdots<t_l$ and $I_n\in \mathbb{R}^{n\times n}$ is identity matrix of proper sizes. To rewrite (\ref{eq22}) in a more compact form, we define matrices $\eta _{xx} \in \mathbb{R}^{l\times \frac{n(n+1)}{2}} $, $\delta _{xx} \in \mathbb{R}^{l\times n^2} $,$\delta _{xv} \in \mathbb{R}^{l\times mn} $,$\delta _{\overline{dv}}\in \mathbb{R}^{l\times \frac{n(n+1)}{2}} $,$\delta _{\overline{dk_ix}} \in \mathbb{R}^{l\times \frac{n(n+1)}{2}} $ as follows \begin{equation*} \begin{split} \eta _{xx}:=\mathbb{E}\bigg[\overline{X(t_1)} -\overline{X(t_0)},\,\,\overline{X(t_2)} -\overline{X(t_1)},\,\,\cdots,\,\, \overline{X(t_l)}-\overline{X(t_{l-1})}\bigg]^T,\\ \end{split} \end{equation*} \begin{equation*} \begin{split} \delta _{xx}:=\mathbb{E}\bigg[\int_{t_0}^{t_1}X(s)\otimes X(s)ds,\,\,\int_{t_1}^{t_2}X(s)\otimes X(s)ds,\,\,\cdots,\,\, \int_{t_{l-1}}^{t_l}X(s)\otimes X(s)ds\bigg]^T,\\ \end{split} \end{equation*} \begin{equation*} \begin{split} \delta _{xv}:=\mathbb{E}\bigg[\int_{t_0}^{t_1}X(s)\otimes v(s)ds,\,\,\int_{t_1}^{t_2}X(s)\otimes v(s)ds,\,\,\cdots,\,\, \int_{t_{l-1}}^{t_l}X(s)\otimes v(s)ds\bigg]^T,\\ \end{split} \end{equation*} \begin{equation*} \begin{split} \delta _{\overline{dv}}:=\mathbb{E}\bigg[\int_{t_0}^{t_1}\overline{Dv(s)}ds,\,\,\int_{t_1}^{t_2}\overline{Dv(s)}ds,\,\,\cdots,\,\, \int_{t_{l-1}}^{t_l}\overline{Dv(s)}ds\bigg]^T,\\ \end{split} \end{equation*} \begin{equation*} \begin{split} \delta _{\overline{dk_ix}}:=\mathbb{E}\bigg[\int_{t_0}^{t_1}\overline{DK_iX(s)}ds,\,\,\int_{t_1}^{t_2}\overline{DK_iX(s)}ds,\,\,\cdots,\,\, \int_{t_{l-1}}^{t_l}\overline{DK_iX(s)}ds\bigg]^T.\\ \end{split} \end{equation*} With these symbols, (\ref{eq22}) implies \begin{equation}\label{eq12} \mathbb{V}_i\times \begin{pmatrix} vech({P}_{i+1})\\ vec\big(M_{i+1}\big)\\ \end{pmatrix}=\mathbb{I}_i, \end{equation} where $\mathbb{V}_i \in \mathbb{R}^{l\times (\frac{n(n+1)}{2}+mn)}$ and $\mathbb{I}_i \in \mathbb{R}^{l}$ are defined as \begin{equation*} \begin{split} \mathbb{V}_i:=\big[\eta_{xx}-\delta_{\overline{dv}}+\delta_{\overline{dk_ix}},2\delta_{xv}-2\delta_{xx}(I_n\otimes K_i^T) \big], \end{split} \end{equation*} \begin{equation*} \begin{split} \mathbb{I}_i:=&\big[\delta_{xx}vec(-Q_i+2K_i^TS)-2\delta_{xv}vec(S) \big],\,\,\forall i\in\mathbb{Z}^+. \end{split} \end{equation*} Multiplying $\mathbb{V}_i^T$ on both sides of (\ref{eq12}), we have \begin{equation}\label{eq201} \mathbb{V}_i^T\mathbb{V}_i\times\begin{pmatrix} vech({P}_{i+1})\\ vec\big(M_{i+1}\big)\\ \end{pmatrix}=\mathbb{V}_i^T\mathbb{I}_i, \,\,\forall i\in\mathbb{Z}^+. \end{equation} If $\mathbb{V}_i$ has full column rank, (\ref{eq201}) can be solved by\\ \begin{equation}\label{eq20} \begin{pmatrix} vech({P}_{i+1})\\ vec\big(M_{i+1}\big)\\ \end{pmatrix}=(\mathbb{V}_i^T\mathbb{V}_i)^{-1}\mathbb{V}_i^T\mathbb{I}_i, \,\,\forall i\in\mathbb{Z}^+. \end{equation} If $\mathbb{V}_i$, $\forall i\in\mathbb{Z}^+$, has full column rank, it follows from Lemma 2 and the above procedure that $P_{i+1}$ and $K_{i+1}$ generated from (\ref{eq5}) and (\ref{eq6}) satisfy (\ref{eq20}). Note that (\ref{eq20}) does not use the information of coefficient matrices $A$, $B$, $C$, thus if we can solve $P_{i+1}$ and $K_{i+1}$, $\forall i\in\mathbb{Z}^+$, from (\ref{eq20}), we obtain a partially model-free algorithm. Then, we give a rank condition in the next lemma, under which matrices $\mathbb{V}_i$, $\forall i\in\mathbb{Z}^+$, have full column rank. \vspace{2mm} \noindent{\bf Lemma 3.} If there exists an $l_0 \in \mathbb{Z^{+}}$, such that \begin{equation}\label{rank} rank([\delta_{xx},\,\,\delta_{xv}])=mn+\frac{n(n+1)}{2}, \end{equation} for all $l \geq l_0$, then matrices $\mathbb{V}_i$, $\forall i\in\mathbb{Z}^+$, have full column rank. \vspace{2mm} \noindent{\bf Proof.} Given $i\in\mathbb{Z}^+$, this proof is equivalent to proving that \begin{equation}\label{eq14} \mathbb{V}_iN=0 \end{equation} has only the solution $N=0$. Now we prove it by contradiction. Assume $N=[vech(F)^T,vec(G)^T]^T\in \mathbb{R}^{mn+\frac{n(n+1)}{2}}$ is a nonzero column vector, where $vech(F)\in \mathbb{R}^{\frac{n(n+1)}{2}}$ and $vec(G)\in \mathbb{R}^{mn}$. Applying Ito's formula to $X(s)^TFX(s)$, integrating from $t$ to $t+\triangle t$ and taking expection $\mathbb{E}$, one gets \begin{equation}\label{eq13} \begin{split} &\mathbb{E}\big[X(t+\triangle t)^TFX(t+\triangle t)-X(t)^TFX(t)\big]\\ = &\mathbb{E}\int_{t}^{t+\triangle t}X(s)^T\big(A_i^TF+FA_i+C_i^TFC_i\big)X(s)ds\\ &+2\mathbb{E}\int_{t}^{t+\triangle t}\big(v(s)-K_iX(s)\big)^TB^TFX(s)ds\\ &+2\mathbb{E}\int_{t}^{t+\triangle t}\big(v(s)-K_iX(s)\big)^TD^TFC_iX(s)ds\\ &+\mathbb{E}\int_{t}^{t+\triangle t}\big(v(s)-K_iX(s)\big)^TD^TFD\big(v(s)-K_iX(s)\big)ds,\\ \end{split} \end{equation} where $X(\cdot)$ is the trajectory of system (\ref{eq11}) with control $v(\cdot)$. By (\ref{eq10}), (\ref{eq13}) and the definition of $\mathbb{V}_i$, we have \begin{equation}\label{eq15} \mathbb{V}_iN=\delta_{xx}vec(\mathcal{Y})+\delta_{xv}vec(\mathcal{L}), \end{equation} where \begin{equation}\label{eq16} \begin{split} \mathcal{Y}=&A_i^TF+FA_i+C_i^TFC_i-K_i^T(B^TF+D^TFC_i+G-D^TFDK_i)\\ &-(FB+C_i^TFD+G^T-K_i^TD^TFD)K_i,\\ \end{split} \end{equation} \begin{equation}\label{eq17} \begin{split} \mathcal{L}=2B^TF+2D^TFC_i+2G-2D^TFDK_i. \end{split} \end{equation} Noting that $\mathcal{Y}$ is a symmetric matrix, we know \begin{equation*} \delta_{xx}vec(\mathcal{Y})=\delta_{\overline{x}}vech(\mathcal{Y}), \end{equation*} where $\delta_{\overline{x}}\in\mathbb{R}^{l\times\frac{n(n+1)}{2}}$ is defined as\\ \begin{equation*} \begin{split} \delta_{\overline{x}}=\mathbb{E}\bigg[\int_{t_0}^{t_1}\overline{X(s)}ds,\int_{t_1}^{t_2}\overline{X(s)}ds, \cdots, \int_{t_{l-1}}^{t_l}\overline{X(s)}ds\bigg]^T.\\ \end{split} \end{equation*} Then (\ref{eq14}) and (\ref{eq15}) imply \begin{equation}\label{eq18} [\delta_{\overline{x}},\delta_{xv}]\begin{pmatrix} vech(\mathcal{Y})\\ vec(\mathcal{L})\\ \end{pmatrix}=0. \end{equation} Under the rank condition in Lemma 3, it is easy to see that $[\delta_{\overline{x}},\delta_{xv}]$ has full column rank. As a result, the unique solution to (\ref{eq18}) is $vech(\mathcal{Y})=0,vec(\mathcal{L})=0$. By the definitions of $vec(\cdot)$ and $vech(\cdot)$, we have $\mathcal{Y}=0,\mathcal{L}=0$. It follows from (\ref{eq16}),(\ref{eq17}), $\mathcal{Y}=0$ and $\mathcal{L}=0$ that \begin{equation}\label{eq19} A_i^TF+FA_i+C_i^TFC_i=0. \end{equation} Further, since $K_i$, $i\in\mathbb{Z^{+}}$, is a stabilizer, we can easily see from Definition 1 that the trajectory of \begin{equation} \label{system2} \begin{cases} \begin{split} dx(s)= \,\,A_ix(s)ds+C_ix(s)dw(s), \end{split}\\ x(0)=x_0\neq 0, \end{cases} \end{equation} satisfies $\lim_{s\rightarrow+\infty}\mathbb{E}\big[x(s)^Tx(s)\big]=0$. For any $t>0$, applying Ito's formula to $d\big(x(s)^TFx(s)\big)$, we get \begin{equation}\label{eq88} \begin{split} &\mathbb{E}\big[x^T(t)Fx(t)\big]-x_0^TFx_0\\ =\,\,&\mathbb{E}\int_{0}^{t}x^T(s)\big(A_i^TF+FA_i+C_i^TFC_i\big)x(s)ds,\\ \end{split} \end{equation} where $x(\cdot)$ is governed by (\ref{system2}). Letting $t$ go to positive infinity, it is easy to see from (\ref{eq19}) and (\ref{eq88}) that $x_0^TFx_0=0$. Notice that $x_0$ can be any nonzero element in $\mathbb{R}^n$, thus we know $F=0$. Then it follows from (\ref{eq17}) and $\mathcal{L}=0$ that $G=0$, which contradicts with $N\neq0$. The proof is completed.$\hfill\blacksquare$\\ \vspace{2mm} Using notations defined above, the data-driven algorithm is given in Algorithm \ref{A2}. \begin{algorithm}[h] \caption{} \label{A2} \begin{algorithmic}[1] \State Initial $i=0$ and choose a stabilizer $K_0$ for system (\ref{eq2}). Employ $v(\cdot)=K_0X(\cdot)+e(\cdot)$ as the control and compute $\eta_{xx}$, $\delta_{xx}$, $\delta_{xv}$ and $\delta_{\overline{dv}}$. \State \textbf{repeat} \State Compute $\delta_{\overline{dk_ix}}$ and solve $P_{i+1}$ and $M_{i+1}$ from (\ref{eq20}). \State $K_{i+1}=(R+D^TP_{i+1}D)^{-1}M_{i+1}$. \State $i\leftarrow i+1$ \State \textbf{Until} {$|P_i-P_{i+1}|<\varepsilon$, where $\varepsilon>0$ is a constant that can be predefined as a small threshold.} \end{algorithmic} \end{algorithm} \vspace{2mm} \noindent{\bf Remark 1.} In Algorithm \ref{A2}, $e(\cdot)$ is called the exploration noise. The main purpose of adding exploration noise is to meet the persistent excitation condition \citep{JiangJiang2012,Bradtke1994,Bradtke1993}, and thus rank condition (\ref{rank}) in Lemma 3 is satisfied. To tackle some practical ADP and machine learning problems, researchers usually choose exploration noises such as exponentially decreasing noise \citep{Vamvoudakis2011}, the random noise generated from the normal distribution \citep{Bradtke1993}, the sum of sinusoidal signals \citep{JiangJiang2012} and random noise \citep{Tamimi2007}. During the simulation in Section \ref{sec4}, the exploration noise is selected as a noise generated by Gaussian distribution. \vspace{2mm} Finally, we present the convergence analysis of Algorithm \ref{A2}. \vspace{2mm} \noindent{\bf Theorem 1.} When rank condition (\ref{rank}) is guaranteed, $\{K_i\}_{i=0}^\infty$ and $\{P_i\}_{i=1}^\infty$ defined in Algorithm \ref{A2} converge to $K^*$ and $P^*$, respectively. \vspace{2mm} \noindent{\bf Proof.} Given $K_i$, $\forall i\in\mathbb{Z}^+$, it follows from Lemma 2 that $\big(P_{i+1},M_{i+1}\big)$ generated from iteration (\ref{eq5}) and (\ref{eq6}) satisfy (\ref{eq20}). Moreover, it can be seen from Lemma 3 that (\ref{eq20}) has a unique solution if rank condition (\ref{rank}) holds. Therefore, if condition (\ref{rank}) is satisfied, the solution of equation (\ref{eq20}) is equivalent to the solution of iterations (\ref{eq5}) and (\ref{eq6}). Otherwise, (\ref{eq20}) has at least two different solution pairs. Thus, the convergence of Algorithm \ref{A2} is obtained by Lemma 1. This completes the proof.$\hfill\blacksquare$ \section{Numerical example}\label{sec4} In this section, we give a simulation example to illustrate the data-driven partially model-free algorithm. The system parameters of system (\ref{eq2}) are given as follows \begin{equation*} A= \begin{bmatrix} 0 & -0.6\\ 0.6 & -0.3 \end{bmatrix}, B= \begin{bmatrix} 0.05\\ 0.01 \end{bmatrix}, C= \begin{bmatrix} -0.02 & 0.03\\ -0.05 &0.02 \end{bmatrix}, D= \begin{bmatrix} 0.001\\ 0.03 \end{bmatrix}, \end{equation*} and the initial state is $x_0=[0.5,-0.1]^T$. The coefficients in cost functional (\ref{eq3}) are chosen as $Q=diag(1,0.1)$, $S=0$ and $R=1$. Let $K_0=[0,0]$ and $\triangle t=0.01$ s, i.e., the value of $l$ in equation (\ref{eq22}) is $l=\frac{4}{\triangle t}=\frac{4}{0.01}=40$. We employ $v(\cdot)=K_0X(\cdot)+e(\cdot)$ as the input and collect the input and state information over time interval $[0, 4]$. Then, the collected data and the coefficient matrix $D$ are used to implement Algorithm 1, where rank condition (\ref{rank}) is guaranteed due to the existence of exploration noise $e(\cdot)$. Moreover, we set $P_0=0$ to check the stopping criterion at the first iteration step. By applying the data-driven algorithm, we can obtain two approximation matrices $\widetilde{P}^*$ and $\widetilde{K}^*$ as shown below \begin{equation*} \widetilde{P}^*= \begin{bmatrix} 2.2384 & -0.8272 \\ -0.8272 & 1.8240 \\ \end{bmatrix}, \widetilde{K}^*= \begin{bmatrix} -0.1109 & 0.0408\\ \end{bmatrix}. \end{equation*} \begin{figure*}[!h] \centering \includegraphics[width=0.785\textwidth]{PP} \centering \includegraphics[width=0.785\textwidth]{KK} \caption{Convergence of $P_i$ and $K_i$ during the simulation} \end{figure*} Figure 1 plots the convergence of Algorithm 1. Moreover, to check the error between ($\widetilde{P}^*, \widetilde{K}^*$) and the true values ($P^*, K^*$), we denote the left side of (\ref{eq5}) as \begin{equation*} \begin{split} \mathcal{R}(P, K):=P(A+BK)+(A+BK)^TP &+(C+DK)^TP(C+DK)\\ &+K^TRK+S^TK+K^TS+Q. \end{split} \end{equation*} Then we have $| \mathcal{R}(\widetilde{P}^*,\widetilde{K}^*)|=9.7162\times10^{-4}$, implying that the error of ($\widetilde{P}^*, \widetilde{K}^*$) is of size $10^{-4}$. Furthermore, an optimal trajectory governed by $v(\cdot)=\widetilde{K}^*X(\cdot)$ is plotted in Figure 2, which means that $\widetilde{K}^*$ is indeed a stabilizer. The above simulation results imply that the algorithm proposed in this paper maybe an effective method in solving infinite-horizon SLQ problems with partial knowledge of system parameters. \begin{figure*}[!h] \centering \includegraphics[width=0.785\textwidth]{X1} \centering \includegraphics[width=0.785\textwidth]{X2} \caption{An optimal trajectory of the system states generated by $v(\cdot)=\widetilde{K}^*X(\cdot)$} \end{figure*} \section{Conclusions}\label{sec5} This paper develops a data-driven algorithm to tackle a continuous-time SLQ optimal control problem. The data-driven algorithm relaxes the assumption on the information of system matrix parameters by using input and state data collected over some time interval. The convergence analysis is provided under some mild conditions. An interesting topic is to consider the case that the control weighting matrix in the cost functional to be indefinite. This problem is left for further investigation. \section*{Acknowledgements} The authors would like to thank Professor Guangchen Wang and Miss Yu Wang for their insightful comments on improving the quality of this work. \section*{Disclosure statement} The authors declare no potential conflict of interests. \section*{Funding} Heng Zhang acknowledges the financial support from the National Natural Science Foundation of China ( No.~61821004, No.~11831010, No.~61925306), and the Natural Science Foundation of Shandong Province ( No.~ ZR2019ZD42, No.~ZR2020ZD24). Na Li acknowledges the financial support from the National Natural Science Foundation of China (No.~12171279, No.~11801317), the Natural Science Foundation of Shandong Province (No.~ZR2019MA013), and the Colleges and Universities Youth Innovation Technology Program of Shandong Province (No.~2019KJI011).

1432e76ad278b005b83d26510a45f42f4c50a21b

\section{Introduction} Recent face recognition model recognizes the identity of a given face image from the 1M distractors with an accuracy of 99.087\%~\cite{Kemelmacher-Shlizerman2015}. However, most face recognition benchmarks such as MegaFace~\cite{Kemelmacher-Shlizerman2015}, CASIA~\cite{Yi2014}, and~MS-Celeb-1M \cite{Guo2016} contain high resolution (HR) images that differ significantly from real-world environments, typically captured by surveillance cameras. When deep learning approaches are directly applied to low resolution (LR) images after being trained on HR images, significant performance degradation occurred~\cite{Cheng2018LowResolutionFR,10.5555/1736406.1736429,5634490}. \begin{figure} \centering \includegraphics[height=5.0cm]{figure/concept.png} \caption{Proposed attention similarity knowledge distillation (A-SKD) concept for low resolution (LR) face recognition problem. Well-constructed attention maps from the HR network are transferred to the LR network by forming high similarity between them for guiding the LR network to focus on detailed parts captured by the HR network. Face images and attention maps are from the AgeDB-30~\cite{inproceedings}.} \label{fig:concept} \end{figure} To overcome the LR problem associated with face recognition, prior knowledge extracted from HR face images is used to compensate spatial information loss. Depending on the approach of transferring the prior knowledge to LR image domain, LR face recognition methods are categorized into two types: super-resolution and knowledge distillation based approaches. Super-resolution based approaches utilize generative models to improve LR images to HR before input to recognition networks~\cite{Fookes2012,Gunturk2003,Hennings-Yeomans2008,Kong2019,Tran2017,5634490}. Following the development of super-resolution methods, LR images can be successfully reconstructed into HR images and recognized by a network trained on HR images~\cite{Dong2016AcceleratingTS,7364266,Ledig2017PhotoRealisticSI,8889765}. However, super-resolution models incur high computational costs for both training and inference, even larger than the costs required for recognition networks. Furthermore, generating HR from LR images is an ill-posed problem, i.e., many HR images can match with a single LR image~\cite{srcnn}; hence the identity of a LR image can be altered. To combat this, knowledge distillation based methods have been proposed to transfer prior knowledge from HR images to models trained on LR face images~\cite{Ge_Zhang_Liu_Hua_Zhao_Jin_Wen_2020,Massoli2020,8682926}. When the resolution of face images is degraded, face recognition models cannot capture accurate features for identification due to spatial information loss. In particular, features from detailed facial parts are difficult to be captured from a few pixels on LR images, e.g. eyes, nose, and mouth~\cite{S2LD}. Previous studies mainly focused on feature based knowledge distillation (F-KD) methods to encourage the LR network's features to mimic the HR network's features by reducing the Euclidean distance between them~\cite{Ge_Zhang_Liu_Hua_Zhao_Jin_Wen_2020,Massoli2020,8682926}. The original concept of F-KD was proposed as a lightweight student model to mimic features from over-parameterized teacher models~\cite{F-KD}. Because teacher model's features would generally include more information than the student model, F-KD approaches improve the accuracy of the student model. Similarly, informative features from the HR network are distilled to the LR network in the LR face recognition problems. This study proposes the attention similarity knowledge distillation approach to distill well-constructed attention maps from an HR network into an LR network by increasing similarity between them. The approach was motivated by the observation that humans can approximate an object's regions from LR images based on prior knowledge learned from previously viewed HR images. Kumar et al. proposed that guiding the LR face recognition network to generate facial keypoints (e.g., eyes, ears, nose, and lips) improved recognition performance by directing the network's attention to the informative regions~\cite{S2LD}. Thus, we designed the prior knowledge as an attention map and transferred the knowledge by increasing similarity between the HR and LR networks' attention maps. Experiments on LR face recognition, face detection, and general object classification demonstrated that the attention mechanism was the best prior knowledge obtainable from the HR networks and similarity was the best method for transferring knowledge to the LR networks. Ablation studies and attention analyses demonstrated the proposed A-SKD effectiveness. \section{Related Works} \textbf{Knowledge distillation.} Hinton et al. first proposed the knowledge distillation approach to transfer knowledge from a teacher network into a smaller student network~\cite{KD}. Soft logits from a teacher network were distilled into a student network by reducing the Kullback-Leibler (KL) divergence score, which quantifies the difference between the teacher and student logits distributions. Various F-KD methods were subsequently proposed to distill intermediate representations~\cite{Park2019,FitNet,F-KD,AT}. FitNet reduced the Euclidean distance between teacher and student network's features to boost student network training~\cite{FitNet}. Zagoruyko et al. proposed attention transfer (AT) to reduce the distance between teacher and student network's attention maps rather than distilling entire features~\cite{AT}. Since attention maps are calculated by applying channel-wise pooling to feature vectors, activation levels for each feature can be distilled efficiently. Relational knowledge distillation (RKD) recently confirmed significant performance gain by distilling structural relationships for features across teacher and student networks~\cite{Park2019}. \textbf{Feature guided LR face recognition.} Various approaches that distill well-constructed features from the HR face recognition network to the LR network have been proposed to improve LR face recognition performances~\cite{Ge_Zhang_Liu_Hua_Zhao_Jin_Wen_2020,Massoli2020,8682926}. Conventional knowledge distillation methods assume that over-parameterized teacher networks extract richer information and it can be transferred to smaller student networks. Similarly, LR face recognition studies focused on transferring knowledge from networks trained on highly informative inputs to networks trained on less informative inputs. Zhu et al. introduced knowledge distillation approach for LR object classification \cite{8682926}, confirming that simple logit distillation from the HR to LR network significantly improved LR classification performance, even superior to super-resolution based methods. F-KD~\cite{Massoli2020} and hybrid order relational knowledge distillation (HORKD)~\cite{Ge_Zhang_Liu_Hua_Zhao_Jin_Wen_2020}, which is the variant of RKD~\cite{Park2019}, methods were subsequently applied to LR face recognition problems to transfer intermediate representations from the HR network. Another approach is to guide the LR network by training it to generate keypoints (e.g. eyes, ears, nose, and lips)~\cite{S2LD}. An auxiliary layer is added to generate keypoints, and hence guide the network to focus on specific facial characteristics. It is well known that facial parts such as eyes and ears are important for recognition~\cite{landmark,S2LD}, hence LR face recognition networks guided by keypoints achieve better performance. Inspired by this, we designed the attention distillation method that guides the LR network to focus on important regions of the HR network. However, attention distillation methods have not been previously explored for LR face recognition. We investigated the efficient attention distillation methods for LR settings and proposed the cosine similarity as the distance measure between HR and LR network's attention maps. \section{Method} \subsection{Low resolution image generation} We require HR and LR face image pairs to distill the HR network's knowledge to the LR network. Following the protocol for LR image generation in super-resolution studies~\cite{Dong2016AcceleratingTS,7364266,Ledig2017PhotoRealisticSI,8889765}, we applied bicubic interpolation to down-sample HR images with 2$\times$, 4$\times$, and 8$\times$ ratios. Gaussian blur was then added to generate realistic LR images. Finally, the downsized images were resized to the original image size using bicubic interpolation. Figure \ref{fig:face_images} presents sample LR images. \begin{figure}[h!] \centering \includegraphics[scale=0.43]{figure/face_images.png} \caption{The samples of HR and LR images from the training dataset (CASIA \cite{Yi2014}) with the down-sampling ratios of 2$\times$, 4$\times$, and 8$\times$.} \label{fig:face_images} \end{figure} \subsection{Face recognition with attention modules} \textbf{Face recognition network.} ArcFace \cite{Deng2018} is a SOTA face recognition network comprising convolutional neural network (CNN) backbone and angular margin introduced to softmax loss. Conventional softmax loss can be expressed as \begin{equation} \label{eq:softmax} L_{softmax} = -\frac{1}{N} \sum_{i=1}^{N}log\frac{e^{W^T_{y_i}x_i + b_{y_i}}}{\sum_{j=1}^{n}e^{W^T_jx_i + b_j}} , \end{equation} where $x_i \in \mathbb{R}^d$ is the embedded feature of the $i$-th sample belonging to the $y_i$-th class; $N$ and $n$ are the batch size and the number of classes, respectively; $W_j \in \mathbb{R}^d$ denotes the $j$-th column of the last fully connected layer's weight $W \in \mathbb{R}^{d \times n}$ and $b_j \in \mathbb{R}^n$ is the bias term for the $j$-th class. For simplicity, the bias term is fixed to 0 as in \cite{sphereface}. Then the logit of the $j$-th class can be represented as $W^T_j x_i = \lVert{W_j}\rVert \lVert{x_i}\rVert {cos(\theta_j)}$, where $\theta_j$ denotes the angle between the $W_j$ and $x_i$. Following previous approaches~\cite{sphereface,cosface}, ArcFace set $\lVert{W_j}\rVert=1$ and $\lVert{x_i}\rVert=1$ via $l_2$ normalisation to maximize $\theta_j$ among inter-class and minimize $\theta_j$ among intra-class samples. Further, constant linear angular margin ($m$) was introduced to avoid convergence difficulty. The ArcFace~\cite{Deng2018} loss can be expressed as \begin{equation} \label{eq:arcface} L_{arcface} = -\frac{1}{N} \sum_{i=1}^{N}log\frac{e^{s(cos(\theta_{y_i} + m))}}{e^{s(cos(\theta_{y_i} + m)) + \sum_{j=1,j\ne {y_i}}^{n}e^{s(cos(\theta_j))}}} , \end{equation} where $s$ is the re-scale factor and $m$ is the additive angular margin penalty between $x_i$ and $W_{y_i}$. \textbf{Attention.} Attention is a simple and effective method to guide feature focus on important regions for recognition. Let $\mathbf{f}_i = \mathcal{H}_i(\mathbf{x})$ be intermediate feature outputs from the \textit{i}-th layer of the CNN. Attention maps about $\mathbf{f}_i$ can be represented as the $\mathcal{A}_i(\mathbf{f}_i)$, where $\mathcal{A}_i(\cdot)$ is attention module. Many attention mechanisms have been proposed; AT \cite{AT} simply applied channel-wise pooling to features to estimate spatial attention maps. SENet~\cite{senet} and CBAM \cite{Woo2018} utilized parametric transformations, e.g. convolution layers, to represent attention maps. Estimated attention maps were multiplied with the features and passed to a successive layer. Trainable parameters in attention module are updated to improve performance during back-propagation, forming accurate attention maps. Attention mechanisms can be expressed as \begin{equation} \label{eq:refine1} \mathbf{f^{'}_{i}} = \mathcal{A}^c_{i}(\mathbf{f_{i}}) \otimes \mathbf{f_{i}} \end{equation} and \begin{equation} \label{eq:refine2} \mathbf{f^{''}_{i}} = \mathcal{A}^s_{i}(\mathbf{f^{'}_{i}}) \otimes \mathbf{f^{'}_{i}} , \end{equation} where $\mathcal{A}^c_{i}(\cdot)$ and $\mathcal{A}^s_{i}(\cdot)$ are attention modules for channel and spatial attention maps, respectively. \begin{figure} \centering \includegraphics[height=4.5cm]{figure/model_outline.png} \caption{Proposed A-SKD framework. The LR network formulates precise attention maps by referencing well-constructed channel and spatial attention maps obtained from the HR network, focusing on detailed facial parts which are helpful for the face recognition. We only show the attention distillation for the first block.} \label{fig:model} \end{figure} Features are refined twice by multiplying channel and spatial attention maps in order (\ref{eq:refine1}) and (\ref{eq:refine2}). Any parametric attention transformation could be employed for the proposed A-SKD, and we adopted the popular CBAM~\cite{Woo2018} module, \begin{equation} \label{eq:cam} \mathcal{A}^c(\mathbf{f}) = \sigma (FC(AvgPool(\mathbf{f}))+FC(MaxPool(\mathbf{f}))) \end{equation} and \begin{equation} \label{eq:sam} \mathcal{A}^s(\mathbf{f}) = \sigma (f^{7\times7}(AvgPool(\mathbf{f});MaxPool(\mathbf{f}))) , \end{equation} where $\sigma(\cdot)$ is the sigmoid function; and $FC(\cdot)$ and $f^{7\times7}(\cdot)$ are fully connected and convolution layers with $7\times7$ filters, respectively. \subsection{Proposed attention similarity knowledge distillation framework} Unlike the conventional knowledge distillation, the network size of teacher and student network is same for A-SKD. Instead, the teacher network is trained on HR images whereas the student network is trained on LR images. Due to the resolution differences, features from both networks are difficult to be identical. Therefore, we propose to distill well-constructed attention maps from the HR network into the LR network instead of features. \begin{equation} \begin{split} \rho_i &= 1 - \langle\mathcal{A}_{T,i}(\mathbf{f}_{T,i}), \mathcal{A}_{S,i}(\mathbf{f}_{S,i})\rangle \\ &= 1 - \frac{\mathcal{A}_{T,i}(\mathbf{f}_{T,i})}{\lVert \mathcal{A}_{T,i}(\mathbf{f}_{T,i}) \rVert_2} \cdot \frac{\mathcal{A}_{S,i}(\mathbf{f}_{S,i})}{\lVert \mathcal{A}_{S,i}(\mathbf{f}_{S,i})\rVert_2} , \end{split} \end{equation} where $\rho_i$ is the cosine distance between attention maps from the \textit{i}-th layer of the teacher and student networks; $\langle \cdot, \cdot \rangle$ denotes the cosine similarity; $\lVert \cdot \rVert_2$ denotes L2-norm; $\mathcal{A}_i(\mathbf{f}_i)$ denotes the attention maps for the \textit{i}-th layer features; and $T$ and $S$ denote the teacher and student network, respectively. Thus, $\mathcal{A}_{T,i}(\mathbf{f}_{T,i})$ and $\mathcal{A}_{S,i}(\mathbf{f}_{S,i})$ are attention maps estimated from the \textit{i}-th layer of the teacher and student network's features, respectively. Reducing the cosine distance between HR and LR attention maps increases the similarity between them. Distillation loss for A-SKD is calculated as \begin{equation} \label{eq:kd_loss} \mathcal{L}_{distill} = \sum^{N}_{i=1} \frac{(\rho^s_i + \rho^c_i)}{2} \end{equation} which average the cosine distance for channel and spatial attention maps between the HR and LR networks, and sums them across layers ($i=1,2,3,...,N$) of the backbone. $N$ is the number of layers utilized for the distillation. Total loss for the LR face recognition network is the sum of target task's loss and distillation loss (\ref{eq:kd_loss}) weighted by the factor ($\lambda_{distill}$). In this work, we utilized the ArcFace loss (\ref{eq:arcface}) as a target task's loss. \begin{equation} \label{eq:tot_loss} \mathcal{L}_{total} = \mathcal{L}_{arcface} + \lambda_{distill} * \mathcal{L}_{distill}. \end{equation} Further, our method can be utilized in conjunction with the logit distillation by simply adding the logit distillation loss~\cite{KD} to our loss function (\ref{eq:tot_loss}). Since logit is the final output of the network, incorporating the logit distillation loss allows the LR network to make the same decision as the HR network based on the refined attention maps. \section{Experiments} \subsection{Settings} \textbf{Datasets.} We employed the CASIA~\cite{Yi2014} dataset for training, which is a large face recognition benchmark comprising approximately 0.5M face images for 10K identities. Each sample in CASIA was down-sampled to construct the HR-LR paired face dataset. For the evaluation, the manually down-sampled face recognition benchmark (AgeDB-30~\cite{inproceedings}) and the popular LR face recognition benchmark (TinyFace~\cite{Cheng2018LowResolutionFR}) were employed. Since AgeDB-30 have similar resolution to CASIA, networks trained on down-sampled CASIA images were validated on AgeDB-30 down-sampled images with matching ratio. In contrast, the real-world LR benchmark (TinyFace) comprises face images with the resolution of 24$\times$24 in average when they are aligned. Therefore, they were validated using a network trained on CASIA images down-sampled to 24$\times$24 pixels. \textbf{Task and metrics.} Face recognition was performed for two scenarios: face verification and identification. Face verification is where the network determines whether paired images are for the same person, i.e., 1:1 comparison. To evaluate verification performance, accuracy was determined using validation sets constructed from probe and gallery set pairs following the LFW protocol~\cite{LFW}. Face identification is where the network recognize the identity of a probe image by measuring similarity against all gallery images, i.e., 1:N comparison. This study employed the smaller AgeDB-30 dataset for the face verification; and larger TinyFace dataset for the face identification. \textbf{Comparison with other methods.} Typically, the distillation of intermediate representation is performed concurrently with the target task’s loss. Previous distillation methods in the experiments utilized the both face recognition and distillation loss, albeit face recognition loss of varying forms. In addition, some feature distillation approach reported their performances with the logit distillation loss. In order to conduct a fair comparison, we re-implemented the prior distillation methods with the same face recognition loss (ArcFace~\cite{Deng2018}) and without the logit distillation loss. Further, our method requires the parametric attention modules for the distillation. Therefore, we utilized the same backbone network with CBAM attention modules for all methods; we combined the CBAM modules to all convolution layers, with the exception of the stem convolution layer and the convolution layer with a kernel size of 1. \textbf{Implementation details.} We followed the ArcFace protocol for data preprocessing: detecting face regions using the MTCNN~\cite{Zhang2016} face detector, cropping around the face region, and resizing the resultant portion to 112$\times$112 pixel using bilinear interpolation. The backbone network was ResNet-50 with CBAM attention module. For the main experiments, we distilled the attention maps for every convolution layers with the exception of the stem convolution layer and the convolution layer with a kernel size of 1. Weight factors for distillation (\ref{eq:tot_loss}) $\lambda_{distill} = 5$. This weight factors generally achieved the superior results not only for the face recognition benchmarks, but also for the ImageNet~\cite{imagenet}. Learning rate = 0.1 initially, divided by 10 at 6, 11, 15, and 17 epochs. SGD optimizer was utilized for the training with batch size = 128. Training completed after 20 epochs. The baseline refers to the LR network that has not been subjected to any knowledge distillation methods. For the hyperparameter search, we divided 20\% of the training set into the validation set and conducted a random search. After the hyperparameter search, we trained the network using the whole training set and and performed the evaluation on the AgeDB-30 and TinyFace. \subsection{Face recognition benchmark results} \textbf{Evaluation on AgeDB-30.} Table \ref{tab:performances} shows LR face recognition performance on AgeDB-30 with various down-sample ratios depending on distillation methods. Except for HORKD, previous distillation methods~\cite{Massoli2020,AT} exhibited only slight improvement or even reduced performance when the downsampling ratios increase. This indicates that reducing the L2 distance between the HR and LR network's features is ineffective. In contrast, HORKD improved LR recognition performance by distilling the relational knowledge of the HR network's features. When the input’s resolution decrease, the intermediate features are hard to be identical with the features from the HR network. Instead the relation among the features of the HR network can be transferred to the LR network despite the spatial information loss; this was the reason of HORKD’s superior performances even for the 4$\times$ and 8$\times$ settings. \begin{table}[] \centering \caption{Proposed A-SKD approach compared with baseline and previous SOTA methods on AgeDB-30 with 2$\times$, 4$\times$, and 8$\times$ down-sampled ratios. L, F, SA, and CA indicate distillation types of logit, feature, spatial attention, and channel attention, respectively. Ver-ACC denotes the verification accuracy. Base refers to the LR network that has not been subjected to any knowledge distillation methods.} \resizebox{0.82\columnwidth}{!}{% \begin{tabular}{@{}ccccc@{}} \toprule \multirow{2}{*}{\textbf{Resolution}} & \multirow{2}{*}{\textbf{Method}} & \multirow{2}{*}{\textbf{Distill Type}} & \multirow{2}{*}{\textbf{Loss Function}} & \multirow{2}{*}{\textbf{\begin{tabular}[c]{@{}c@{}}Ver-ACC (\%)\\ (AgeDB-30)\end{tabular}}} \\ & & & & \\ \midrule 1$\times$ & Base & - & - & 93.78 \\ \hline \multirow{7}{*}{2$\times$} & Base & - & - & 92.83 \\ & F-KD \cite{Massoli2020} & F & L2 & 93.05 \\ & AT \cite{AT} & SA & L2 & 92.93 \\ & HORKD \cite{Ge_Zhang_Liu_Hua_Zhao_Jin_Wen_2020} & F & L1+Huber & 93.13 \\ & A-SKD \textbf{(Ours)} & SA+CA & Cosine & \textbf{93.35} \\ & A-SKD+KD \textbf{(Ours)} & SA+CA+L & Cosine+KLdiv & \textbf{93.58} \\ \hline \multirow{7}{*}{4$\times$} & Base & - & - & 87.74 \\ & F-KD \cite{Massoli2020} & F & L2 & 87.72 \\ & AT \cite{AT} & SA & L2 & 87.75 \\ & HORKD \cite{Ge_Zhang_Liu_Hua_Zhao_Jin_Wen_2020} & F & L1+Huber & 88.08 \\ & A-SKD \textbf{(Ours)} & SA+CA & Cosine & \textbf{88.58} \\ & A-SKD+KD \textbf{(Ours)} & SA+CA+L & Cosine+KLdiv & \textbf{89.15} \\ \hline \multirow{7}{*}{8$\times$} & Base & - & - & 77.75 \\ & F-KD \cite{Massoli2020} & F & L2 & 77.85 \\ & AT \cite{AT} & SA & L2 & 77.40 \\ & HORKD \cite{Ge_Zhang_Liu_Hua_Zhao_Jin_Wen_2020} & F & L1+Huber & 78.27 \\ & A-SKD \textbf{(Ours)} & SA+CA & Cosine & \textbf{79.00} \\ & A-SKD+KD \textbf{(Ours)} & SA+CA+L & Cosine+KLdiv & \textbf{79.45} \\ \bottomrule \end{tabular}% } \label{tab:performances} \end{table} However, attention maps from the HORKD exhibit similar pattern to LR baseline network rather than the HR network in the Figure~\ref{fig:attn_example}. HR attention maps are highly activated in facial landmarks, such as eyes, lips, and beard, which are helpful features for face recognition~\cite{S2LD}. In contrast, detailed facial parts are less activated for LR attention maps because those parts are represented with a few pixels. Although HORKD boosts LR recognition performance by transferring HR relational knowledge, it still failed to capture detailed facial features crucial for recognition. The proposed A-SKD method directs the LR network's attention toward detailed facial parts that are well represented by the HR network's attention maps. Based on the refined attention maps, A-SKD outperforms the HORKD and other knowledge distillation methods for all cases. AgeDB-30 verification accuracy increased 0.6\%, 1.0\%, and 1.6\% compared with baseline for 2$\times$, 4$\times$, and 8$\times$ down-resolution ratios, respectively. In addition, when A-SKD is combined with logit distillation (KD), the verification accuracy increased significantly for all settings. From the results, we confirmed that the attention knowledge from the HR network can be transferred to the LR network and led to significant improvements that were superior to the previous SOTA method. \textbf{Evaluation on TinyFace.} Unlike the face verification, the identification task requires to select a target person's image from the gallery set consists of a large number of face images. Therefore, the identification performances decrease significantly when the resolution of face images are degraded. Table \ref{tab:tinyface} showed the identification performances on the TinyFace benchmark. When the AT~\cite{AT} was applied, the rank-1 identification accuracy decreased 13.34\% compared to the baseline. However, our approach improved the rank-1 accuracy 13.56\% compared to the baseline, even outperforming the HORKD method. This demonstrated that the parametric attention modules (CBAM) and cosine similarity loss are the key factors for transferring the HR network's knowledge into the LR network via attention maps. The proposed method is generalized well to real-world LR face identification task which is not manually down-sampled. \begin{table}[] \centering \caption{Evaluation results on TinyFace identification benchmark depending on the distillation methods. Acc@K denotes the rank-K accuracy (\%).} \label{tab:tinyface} \resizebox{0.60\textwidth}{!}{% \begin{tabular}{@{}ccccc@{}} \toprule \textbf{} & \textbf{ACC@1} & \textbf{ACC@5} & \textbf{ACC@10} & \textbf{ACC@50} \\ \midrule Base & 42.19 & 50.62 & 53.67 & 60.41 \\ AT~\cite{AT} & 36.56 & 45.68 & 49.03 & 56.44 \\ HORKD~\cite{Ge_Zhang_Liu_Hua_Zhao_Jin_Wen_2020} & 45.49 & 54.80 & 58.26 & 64.30 \\ A-SKD \textbf{(Ours)} & \textbf{47.91} & \textbf{56.55} & \textbf{59.92} & \textbf{66.60} \\ \bottomrule \end{tabular}% } \end{table} \section{Discussion} \textbf{Attention correlation analysis.} Figure \ref{fig:correlation} shows Pearsons correlation between attention maps from the HR and LR networks for the different distillation methods. Spatial and channel attention maps from the four blocks for models other than A-SKD have a low correlation between the HR and LR networks, with a magnitude lower than 0.5. In particular, spatial attention maps obtained from the first block of the LR baseline and HORKD network have negative correlation with the HR network ($r = -0.39$ and $-0.29$, respectively). \begin{figure} \centering \includegraphics[height=7.3cm]{figure/attention.png} \caption{Normalized spatial attention maps from the first block for different distillation methods. Red and blue regions indicate high and low attention, respectively. Face images and attention maps are from the AgeDB-30.} \label{fig:attn_example} \end{figure} Figure \ref{fig:attn_example} shows that spatial attention maps from LR baseline and HORKD networks are highly activated in skin regions, which are less influenced by resolution degradation, in contrast to the HR network. This guides the LR network to the opposite directions from the HR network. However, spatial attention maps from A-SKD exhibit strong positive correlation with those from the HR network, highlighting detailed facial attributes such as beard, hair, and eyes. Through the A-SKD, the LR network learned where to focus by generating precise attention maps similar to those for the HR network. Consequently, Pearsons correlation, i.e., the similarity measure between HR and LR attention maps, was significantly improved for all blocks, with a magnitude higher than 0.6. Thus the proposed A-SKD approach achieved superior efficacy and success compared with previous feature based SOTA methods. \begin{figure} \centering \includegraphics[height=5.1cm]{figure/correlation.png} \caption{Pixel level Pearsons correlation between the HR and LR network's attention maps for different distillation methods. B\{$i$\}-\{S,C\} indicates Pearsons correlation for spatial or channel attention maps obtained from the $i$-th ResNet block between the HR and LR networks; and $r$ is Pearsons correlation coefficient representing linear relationships between input variables. Base refers to the LR network that has not been subjected to any knowledge distillation methods. Pearsons correlation is measured using the AgeDB-30.} \label{fig:correlation} \end{figure} \textbf{Comparison with attention transfer~\cite{AT}.} Primary distinctions between AT \cite{AT} and A-SKD include the cosine similarity loss, parametric attention modules, and distillation of both channel and spatial attention maps. Correlation analysis for A-SKD confirmed that the cosine similarity loss is an effective strategy for transferring attention knowledge. Distilling AT attention maps using the cosine similarity rather than the L2 loss increased AgeDB-30 verification accuracy by 0.32\%p (Table~\ref{tab:ablation}). AT calculates attention maps using channel-wise pooling, a non-parametric layer; whereas A-SKD calculates attention maps using parametric layers comprising fully connected and convolution layers. When the input image resolution degrades, the student network’s feature representation diverges from that of the teacher network. Therefore, it is difficult to match the attention maps of the student network obtained by the non-parametric module with those of the teacher network. Instead, A-SKD employs the parametric module for the attention maps extraction and the cosine similarity loss for the distillation; therefore, the attention maps from the student network can be adaptively trained to be similar to the attention maps from the teacher network despite the differences in the features. Finally, A-SKD distills both spatial and channel attention maps in contrast to AT which only considered spatial attention maps. We confirmed A-SKD with spatial and channel attention additionally improved AgeDB-30 verification accuracy by 0.34\%p compared with spatial-only attention. This comparison results also confirmed that A-SKD, designed for attention distillation on LR settings, is the most effective approach for transferring attention knowledge. \begin{table}[] \centering \caption{Comparing attention transfer (AT) \cite{AT} and proposed A-SKD on AgeDB-30 benchmark down-sampled with 8$\times$ ratio. AT* indicates the cosine similarity loss was utilized for attention transfer rather than the original L2 loss. SA and CA indicate spatial and channel attention maps, respectively.} \resizebox{0.80\textwidth}{!}{% \begin{tabular}{@{}ccccc@{}} \toprule \textbf{Method} & \textbf{Type} & \textbf{Transformation} & \textbf{Loss Function} & \textbf{\begin{tabular}[c]{@{}c@{}}Ver-ACC (\%)\\ (AgeDB-30)\end{tabular}} \\ \midrule AT \cite{AT} & SA & Non-parametric layer & L2 & 77.40 \\ AT$^*$ & SA & Non-parametric layer & Cosine & 77.72 \\ A-SKD & SA & Parametric layer & Cosine & 78.66 \\ A-SKD & SA + CA & Parametric layer & Cosine & \textbf{79.00} \\ \bottomrule \end{tabular} \label{tab:ablation}% } \end{table} \section{Extension to Other Tasks} \subsection{Object classification} We conducted experiments for object classification on LR images using the 4$\times$ down-sampled ImageNet~\cite{imagenet}. For the backbone network, we utilized the ResNet18 with CBAM attention modules. We compared our method to other knowledge distillation methods (AT~\cite{AT} and RKD~\cite{Park2019}) which are widely utilized in the classification domains. We re-implemented those methods using its original hyperparameters. Usually, AT and RKD were utilized along with the logit distillation for the ImageNet; therefore, we performed the AT, RKD, and A-SKD in conjunction with the logit distillation in the Table \ref{tab:imagenet_result}. Training details are provided in the Supplementary Information. Table \ref{tab:imagenet_result} shows that A-SKD outperformed the other methods on the LR ImageNet classification task. Park et al. demonstrated that introducing the accurate attention maps led the significant improvement on classification performances~\cite{Park2018BAMBA,Woo2018}. When the attention maps were distilled from the teacher network, student network could focus on informative regions by forming precise attention maps similar with the teacher's one. Thus, our method can be generalized to general object classification task, not restricted to face related tasks. \begin{table}[] \centering \caption{Proposed A-SKD performance on low resolution ImageNet classification. All distillation methods were performed in conjunction with the logit distillation.} \label{tab:imagenet_result} \resizebox{0.36\textwidth}{!}{% \begin{tabular}{@{}ccc@{}} \toprule \textbf{Resolution} & \textbf{Method} & \textbf{ACC (\%)} \\ \midrule 1$\times$ & Base & 70.13 \\ \hline \multirow{4}{*}{4$\times$} & Base & 65.34 \\ & AT~\cite{AT} & 65.79 \\ & RKD~\cite{Park2019} & 65.95 \\ & A-SKD & \textbf{66.52} \\ \bottomrule \end{tabular}% } \end{table} \subsection{Face detection} Face detection is a sub-task of object detection to recognize human faces in an image and estimate their location(s). We utilized TinaFace~\cite{TinaFace}, a deep learning face detection model, integrated with the CBAM attention module to extend the proposed A-SKD approach to face detection. Experiments were conducted on the WIDER FACE~\cite{7780965} dataset (32,203 images containing 393,703 faces captured from real-world environments) with images categorized on face detection difficulty: easy, medium, and hard. LR images were generated with 16$\times$ and 32$\times$ down-resolution ratios, and further training and distillation details are provided in the Supplementary Information. \begin{table}[] \centering \caption{Proposed A-SKD performance on LR face detection. mAP is mean average precision; easy, medium, and hard are pre-assessed detection difficulty.} \resizebox{0.46\textwidth}{!}{% \begin{tabular}{ccccc} \toprule \multirow{2}{*}{\textbf{Resolution}} & \multirow{2}{*}{\textbf{Model}} & \multicolumn{3}{c}{\textbf{mAP (\%)}} \\ \cline{3-5} & & \textbf{Easy} & \textbf{Medium} & \textbf{Hard} \\ \midrule 1$\times$ & Base & 95.56 & 95.07 & 91.45 \\ \hline \multirow{2}{*}{16$\times$} & Base & 54.38 & 52.73 & 35.29 \\ & A-SKD & \textbf{62.93} & \textbf{60.19} & \textbf{47.28} \\ \hline \multirow{2}{*}{32$\times$} & Base & 31.15 & 26.68 & 14.00 \\ & A-SKD & \textbf{33.50} & \textbf{30.04} & \textbf{16.02} \\ \bottomrule \end{tabular}% } \label{tab:detection-result} \end{table} Table~\ref{tab:detection-result} shows that A-SKD improved the overall detection performance by distilling well-constructed attention maps, providing significant increases of mean average precision (mAP) for the easy (15.72\% for 16$\times$ and 7.54\% for 32$\times$), medium (14.15\% for 16$\times$ and 12.59\% for 32$\times$), and hard (33.98\% for 16$\times$ and 14.43\% for 32$\times$) level detection tasks. Small faces were well detected in the LR images after distillation as illustrated in Figure \ref{fig:detection_result}. Thus the proposed A-SKD approach can be successfully employed for many LR machine vision tasks. \begin{figure} \centering \includegraphics[height=5.0cm]{figure/detection_result.png} \caption{Qualitative results for LR face detection before and after applying A-SKD. Small faces were better detected after A-SKD. The face images are from the evaluation set of WIDERFACE.} \label{fig:detection_result} \end{figure} \section{Conclusion} We verified that attention maps constructed from HR images were simple and effective knowledge that can be transferred to LR recognition networks to compensate for spatial information loss. The proposed A-SKD framework enabled any student network to focus on target regions under LR circumstances and generalized well for various LR machine vision tasks by simply transferring well-constructed HR attention maps. Thus, A-SKD could replace conventional KD methods offering improved simplicity and efficiency and could be widely applicable to LR vision tasks, which have not been strongly studied previously, without being limited to face related tasks. \\ \textbf{Acknowledgments} This work was supported by the ICT R\&D program of MSIT/IITP[2020-0-00857, Development of Cloud Robot Intelligence Augmentation, Sharing and Framework Technology to Integrate and Enhance the Intelligence of Multiple Robots. And also, this work was partially supported by Korea Institute of Energy Technology Evaluation and Planning (KETEP) grant funded by the Korea government (MOTIE)(No. 20202910100030) and supported by Electronics and Telecommunications Research Institute (ETRI) grant funded by the Korean government. [22ZR1100, A Study of Hyper-Connected Thinking Internet Technology by autonomous connecting, controlling and evolving ways]. \clearpage \bibliographystyle{splncs04} \section{Implementation Details of Previous Methods} Various knowledge distillation approaches have been proposed to transfer the teacher network's knowledge to student network. Usually, the loss term of the knowledge distillation methods can be defined as the summation of the target task's loss and the distillation loss. $L_{total} = L_{target} + L_{distill}$. To compare the A-SKD with the previous distillation methods (HORKD~\cite{Ge_Zhang_Liu_Hua_Zhao_Jin_Wen_2020}, F-KD~\cite{Massoli2020}, and AT~\cite{AT}), we re-implemented those methods using the official implementation code. The distillation loss function of the AT~\cite{AT} is defined as the L2 distance between the teacher and student network's self-attention maps. For the ResNet50, the AT distills the self-attention maps on the four blocks. We performed the experiments using the official implementation of AT (\url{https://github.com/szagoruyko/attention-transfer}). The distillation loss function of the F-KD~\cite{Massoli2020} is defined as the L2 distance between the teacher and student network's feature maps. The penultimate layer's features (output of the backbone network before input to the margin) are utilized for the distillation. We performed the experiments using the official implementation of F-KD (\url{https://github.com/fvmassoli/cross-resolution-face-recognition}). The loss function of the previous SOTA method (HORKD~\cite{Ge_Zhang_Liu_Hua_Zhao_Jin_Wen_2020}) is defined as \begin{equation} \label{eq:horkd_loss} \mathcal{L}_{distill} = L_1 + \alpha L_2 + \beta L_3 + \gamma L_C \end{equation} where $L_1, L_2, L_3$, and $L_C$ indicate the individual-level, pair-level, triplet-level, and group-level knowledge distillation loss, respectively; the different order relational distillation losses are tuned by the factor of $\alpha, \beta, \gamma$ in order. In our implementation, $\alpha=0.1$, $\beta=0.2$, and $\gamma=0.1$ after the hyperparameter search using the same protocol with our method. The penultimate layer's features (output of the backbone network before input to the margin) are utilized for the distillation. Because HORKD has no official implementation code, we referenced the official implementation of RKD~\cite{Park2019} (\url{https://github.com/lenscloth/RKD}). \section{Extension to Other Tasks} \subsection{Object Classification.} We utilized the ResNet50-CBAM backbone for the ImageNet training. The model was trained by SGD optimizer for 90 epochs with learning rate = 0.4, batch size = 1024, momentum = 0.9, and weight decay = 0.0001. Mixed precision was applied for the efficient training. We referenced the code of \url{https://github.com/rwightman/pytorch-image-models} for the training and validation on ImageNet benchmark. \subsection{Face Detection} The WIDER FACE~\cite{7780965} was utilized for training and validation to extend A-SKD to face detection. The WIDER FACE contained 32,203 images from 61 event classes, totaling 393,703 faces. The WIDER FACE categorized images into three face detection difficulties: easy, medium, and hard, considering scale, occlusion, and pose. Face detection performance on WIDER FACE was evaluated by overall mean average precision (mAP), and easy, medium, and hard images were evaluated separately. We trained TinaFace~\cite{TinaFace} on WIDER FACE by combining it with the CBAM attention module. LR images with 16$\times$ and 32$\times$ down-sampling ratio were trained to validate A-SKD effectiveness. TinaFace used ResNet-50 backbone network; and the loss function combined face detection and attention distillation losses. For the detection task, we distilled the attention maps of the four ResNet blocks, not every convolution layer. The model was trained by SGD optimizer for 150 epochs with learning rate = 0.001, momentum = 0.9, and weight decay = 0.0005 on two Titan RTX (24GB) GPUs with batch size = 8. We used non-maximum suppression threshold = 0.4 and confidence level = 0.02. We referenced the code of \url{https://github.com/Media-Smart/vedadet} for the implementation of TinaFace. \clearpage \bibliographystyle{splncs04}

cb67f1b6cc8cda3dbbf72dfd9a37e4fff7c5cc18

\section{Introduction} Cone Beam Computed Tomography (CBCT) is an emerging medical imaging technique to examine the internal structure of a subject noninvasively. A CBCT scanner emits cone-shaped X-ray beams and captures 2D projections at equal angular intervals. Compared with the conventional Fan Beam CT (FBCT), CBCT enjoys the benefits of high spatial resolution and fast scanning speed~\cite{scarfe2006clinical}. Recent years have witnessed the blossoming of low dose CT, which delivers a significantly lower radiation dose during the scanning process. There are two ways to reduce the dose: decreasing source intensity or projection views~\cite{gao2014low}. This paper focuses on the latter, \textit{i.e.}, sparse-view CBCT reconstruction. Sparse-view CBCT reconstruction aims to retrieve a volumetric attenuation coefficient field from dozens of projections. It is a challenging task in two respects. First, insufficient views lead to notable artifacts. As a comparison, the traditional CBCT obtains hundreds of images. The inputs of sparse-view CBCT are 10$\times$ fewer. Second, the spatial and computational complexity of CBCT reconstruction is much higher than that of FBCT reconstruction due to the dimensional increase of inputs. CBCT relies on 2D projections to build a 3D model, while FBCT simplifies the process by stacking 2D slides restored from 1D projections (but in the sacrifice of time and dose). Existing CBCT approaches can be divided into three categories: analytical, iterative and learning-based methods. Analytical methods estimate attenuation coefficients by solving the Radon transform and its inverse. A typical example is the FDK algorithm~\cite{feldkamp1984practical}. It produces good results in an ideal scenario but copes poorly with ill-posed problems such as sparse views. The second family, iterative methods, formulates reconstruction as a minimization process. These approaches utilize an optimization framework combined with regularization modules. While iterative methods perform well in ill-posed problems~\cite{andersen1984simultaneous,sidky2008image}, they require substantial computation time and memory. Recently, learning-based methods have become popular with the rise of AI. They use deep neural networks to 1) predict and extrapolate projections~\cite{anirudh2018lose,tang2019projection,wang2021improving,zang2021intratomo}, 2) regress attenuation coefficients with similar data~\cite{kasten2020end,ying2019x2ct}, and 3) make optimization process differentiable~\cite{adler2018learned,chen2017learned,kang2018deep}. Most of these methods~\cite{anirudh2018lose,kasten2020end,tang2019projection,ying2019x2ct} need extensive datasets for network training. Moreover, they rely on neural networks to remember what a CT looks like. Therefore it is difficult to apply a trained model of one application to another. While there are self-supervised methods~\cite{adler2018learned,zang2021intratomo}, they operate under FBCT settings considering network capacity and memory consumption. Their performance and efficiency drop when applied to the CBCT scenario. Apart from the aforementioned work designated for CT reconstruction, efforts have been made to deal with other ill-posed problems, such as 3D reconstruction in the computer vision field. Similar to CT reconstruction, 3D reconstruction uses RGB images to estimate 3D shapes, which are usually represented as discrete point clouds or meshes. Recent studies propose~\cite{mildenhall2020nerf,park2019deepsdf} Implicit Neural Representation (INR) as an alternative to those discrete representations. INR parameterizes a bounded scene as a neural network that maps spatial coordinates to metrics such as occupancy and color. With the help of position encoder~\cite{mueller2022instant,tancik2020fourier}, INR is capable to learn high-frequency details. This paper proposes Neural Attenuation Fields (NAF), a fast self-supervised solution for sparse-view CBCT reconstruction. Here we use `self-supervised' to highlight that NAF requires no external CT scans but the X-ray projections of the interested object. Inspired by 3D reconstruction work~\cite{mildenhall2020nerf,park2019deepsdf}, we parameterize the attenuation coefficient field as an INR and imitates the X-ray attenuation process with a self-supervised network pipeline. Specifically, we train a Multi-Layer Perceptron (MLP), whose input is an encoded spatial coordinate $(x,y,z)$ and whose output is the attenuation coefficient $\mu$ at that location. Instead of using a common frequency-domain encoding, we adopt hash encoding~\cite{mueller2022instant}, a learning-based position encoder, to help the network quickly learn high-frequency details. Projections are synthesized by predicting the attenuation coefficients of sampled points along ray trajectories and attenuating incident beams accordingly. The network is optimized with gradient descent by minimizing the error between real and synthesized projections. We demonstrate that NAF quantitatively and qualitatively outperforms existing solutions on both human organ and phantom datasets. While most INR approaches take hours for training, our method can reconstruct a detailed CT model within 10-40 minutes, which is comparable to iterative methods. In summary, the main contributions of this work are: \begin{itemize} \item We propose a novel and fast self-supervised method for sparse-view CBCT reconstruction. Neither external datasets nor structural prior is needed except projections of a subject. \item The proposed method achieves state-of-the-art accuracy and spends relatively short computation time. The performance and efficiency of our method make it feasible for clinical CT applications. \item The code will be publicly available for investigation purposes. \end{itemize} \begin{figure}[!b] \centering \includegraphics[width=\linewidth]{figures/framework.pdf} \caption{ NAF pipeline. Gray block: The CBCT scanner captures X-ray projections from different views. Blue block: NAF simulates projections. Orange block: NAF is optimized by comparing real and synthesized projections. Green block: NAF generates a CT model by querying corresponding voxels.} \label{fig:framework} \end{figure} \section{Method} \subsection{Pipeline} The pipeline of NAF is shown in Fig.~\ref{fig:framework}. During a CBCT scanning, an X-ray source rotates around the object and emits cone-shaped X-ray beams. A 2D panel detects X-ray projections at equal angular intervals. NAF then uses the scanner geometry to imitate the attenuation process discretely. It learns CT shapes by comparing real and synthesized projections. After the model optimization, the final CT image is generated by querying corresponding voxels. NAF consists of four modules: ray sampling, position encoding, attenuation coefficient prediction, and projection synthesis. First, we uniformly sample points along X-ray paths based on the scanner geometry. A position encoder network then encodes their spatial coordinates to extract valuable features. After that, an MLP network consumes the encoded information and predicts attenuation coefficients. The last step of NAF is to synthesize projections by attenuating incident X-rays according to the predicted attenuation coefficients on their paths. \subsection{Neural attenuation fields} \subsubsection{Ray sampling} Each pixel value of a projection image results from an X-ray passing through a cubical space and getting attenuated by the media inside. We sample $N$ points at the parts where rays intersect the cube. A stratified sampling method~\cite{mildenhall2020nerf} is adopted, where we divide a ray into $N$ evenly spaced bins and uniformly sample one point at each bin. Setting $N$ greater than the desired CT size ensures that at least one sample is assigned to every grid cell that an X-ray traverses. The coordinates of sampled points are then sent to the position encoding module. \subsubsection{Position encoding} A simple MLP can theoretically approximate any function~\cite{hornik1989multilayer}. Recent studies~\cite{rahaman2019spectral,tancik2020fourier}, however, reveal that a neural network prefers to learn low-frequency details due to ``spectral bias''. To this end, a position encoder is introduced to map 3D spatial coordinates to a higher dimensional space. A common choice is the \textit{frequency encoder} proposed by Mildenhall \textit{et al.}~\cite{mildenhall2020nerf}. It decomposes a spatial coordinate $\mathbf{p}\in\mathbb{R}^{3}$ into $L$ sets of sinusoidal components at different frequencies. While frequency encoder eases the difficulty of training networks, it is considered quite cumbersome. In medical imaging practise~\cite{wu2021irem,zang2021intratomo}, the size of encoder output is set to 256 or greater. The following network must be wider and deeper to cope with the inflated inputs. As a result, it takes hours to train millions of network parameters, which is not acceptable for fast CT reconstruction. Frequency-domain encoding is a dense encoder because it utilizes the entire frequency spectrum. However, dense encoding is redundant for CBCT reconstruction for two main reasons. First, a human body usually consists of several homogeneous media, such as muscles and bones. Attenuation coefficients remain approximately uniform inside one medium but vary between different media. High-frequency features are not necessary unless for points near edges. Second, natural objects favor smoothness. Many organs have simple shapes, such as spindle (muscle) or cylinder (bone). Their smooth surfaces can be easily learned with low-dimensional features. To exploit the aforementioned characteristics of the scanned objects, we use the \textit{hash encoder}~\cite{mueller2022instant}, a learning-based sparse encoding solution. The equation of hash encoder $\mathcal{M_{H}}$ is: \begin{equation} \mathcal{M_{H}}(\mathbf{p};\mathbf{\Theta})=[\mathcal{I}(\mathbf{H}_{1}),\cdots,\mathcal{I}(\mathbf{H}_{L})]^T,~\mathbf{H}=\{\mathbf{c}|h(\mathbf{c})=(\bigoplus c_{j}\pi_{j})~{\rm mod}~T\}. \end{equation} Hash encoder describes a bounded space by $L$ multiresolution voxel grids. A trainable feature lookup table $\mathbf{\Theta}$ with size $T$ is assigned to each voxel grid. At each resolution level, we 1) detect neighbouring corners $\mathbf{c}$ (cubes with different colors in Fig.~\ref{fig:framework}(b)) of the queried point $\mathbf{p}$, 2) look up their corresponding features $\mathbf{H}$ in a hash function fashion $h$~\cite{teschner2003optimized}, and 3) generate a feature vector with linear interpolation $\mathcal{I}$. The output of a hash encoder is the concatenation of feature vectors at all resolution levels. More details of hash function and its symbols can be found in~\cite{mueller2022instant}. Compared with frequency encoder, hash encoder produces much smaller outputs ($32$ in our setting) with competitive feature quality for two reasons. On the one hand, the many-to-one property of hash function conforms to the sparsity nature of human organs. On the other hand, a trainable encoder can learn to focus on relevant details and select suitable frequency spectrum~\cite{mueller2022instant}. Thanks to hash encoder, the subsequent network is more compact. \subsubsection{Attenuation coefficient prediction} We represent the bounded field with a simple MLP $\mathbf{\Phi}$, which takes the encoded spatial coordinates as inputs and outputs the attenuation coefficients $\mu$ at that position. As illustrated in Fig.~\ref{fig:framework}(c), the network is composed of 4 fully-connected layers. The first three layers are 32-channel wide and have ReLU activation functions in between, while the last layer has one neuron followed by a sigmoid activation. A skip connection is included to concatenate the network input to the second layer's activation. By contrast, Zang \textit{et al.}~\cite{zang2021intratomo} use a 6-layer 256-channel MLP to learn features from a frequency encoder. Our network is $10\times$ smaller. \subsubsection{Attenuation synthesis} According to Beer's Law, the intensity of an X-ray traversing matter is reduced by the exponential integration of attenuation coefficients on its path. We numerically synthesize the attenuation process with: \begin{equation} I=I_{0}\exp(-\sum_{i=1}^{N}\mu_{i}\delta_{i}), \end{equation} where $I_{0}$ is the initial intensity and $\delta_{i}=\|\mathbf{p}_{i+1}-\mathbf{p}_{i}\|$ is the distance between adjacent points. \subsection{Model optimization and output} NAF is updated by minimizing the L2 loss between real and synthesized projections. The loss function $\mathcal{L}$ is defined as: \begin{equation} \mathcal{L}(\mathbf{\Theta},\mathbf{\Phi}) = \sum_{\mathbf{r}\in\mathbf{B}}\|I_{r}(\mathbf{r})-I_{s}(\mathbf{r})\|^2, \end{equation} where $\mathbf{B}$ is a ray batch, and $I_{r}$ and $I_{s}$ are real and synthesized projections for ray $\mathbf{r}$ respectively. We update both hash encoder $\mathbf{\Theta}$ and attenuation coefficient network $\mathbf{\Phi}$ during the training process. The final output is formulated as a discrete 3D matrix. We build a voxel grid with the desired size and pass the voxel coordinates to the trained MLP to predict the corresponding attenuation coefficients. A CT model thus is restored. \section{Experiments} \subsection{Experimental settings} \subsubsection{Data} We conduct experiments on five datasets containing human organ and phantom data. Details are listed in Table~\ref{tab:dataset}. \noindent{\textit{\textbf{Human organ}}}: We evaluate our method using public datasets of human organ CTs~\cite{armato2011lung,Klacansky2022Open}, including chest, jaw, foot and abdomen. The chest data are from LIDC-IDRI dataset~\cite{armato2011lung}, and the rest are from Open Scientific Visualization Datasets~\cite{Klacansky2022Open}. Since these datasets only provide volumetric CT scans, we generate projections by a tomographic toolbox TIGRE~\cite{biguri2016tigre}. In TIGRE~\cite{biguri2016tigre}, we capture 50 projections with 3\% noise in the range of 180\degree. We train our model with these projections and evaluate its performance with the raw volumetric CT data. \noindent{\textit{\textbf{Phantom}}}: We collect a phantom dataset by scanning a silicon aortic phantom with GE C-arm Medical System. This system captures 582 500$\times$500 fluoroscopy projections with position primary angle from -103\degree to 93\degree and position secondary angle of 0\degree. A 512$\times$512$\times$510 CT image is also generated with inbuilt algorithms as the ground truth. We only use 50 projections for experiments. \begin{table}[!b] \scriptsize \vspace{-2em} \centering \caption{Details of CT datasets used in the experiments.} \label{tab:dataset} \begin{tabular}{m{0.17\columnwidth}<{\centering}|m{0.15\columnwidth}<{\centering}m{0.15\columnwidth}<{\centering}m{0.15\columnwidth}<{\centering}m{0.15\columnwidth}<{\centering}m{0.15\columnwidth}<{\centering}} \toprule Dataset name & CT dimension & Scanning method & Scanning range & Number of projections & Detector resolution \\ \midrule Chest~\cite{armato2011lung} & 128$\times$128$\times$128 & TIGRE~\cite{biguri2016tigre} & $0\degree\sim180\degree$ & 50 & 256$\times$256 \\ Jaw~\cite{Klacansky2022Open} & 256$\times$256$\times$256 & TIGRE~\cite{biguri2016tigre} & $0\degree\sim180\degree$ & 50 & 512$\times$512 \\ Foot~\cite{Klacansky2022Open} & 256$\times$256$\times$256 & TIGRE~\cite{biguri2016tigre} & $0\degree\sim180\degree$ & 50 & 512$\times$512 \\ Abdomen~\cite{Klacansky2022Open} & 512$\times$512$\times$463 & TIGRE~\cite{biguri2016tigre} & $0\degree\sim180\degree$ & 50 & 1024$\times$1024 \\ Aorta & 512$\times$512$\times$510 & GE C-arm & $-103\degree\sim93\degree$ & 50 (582) & 500$\times$500 \\ \bottomrule \end{tabular} \end{table} \subsubsection{Baselines} We compare our approach with four baseline techniques. \textbf{FDK}~\cite{feldkamp1984practical} is firstly chosen as a representative of analytical methods. The second method \textbf{SART}~\cite{andersen1984simultaneous} is a robust iterative reconstruction algorithm. \textbf{ASD-POCS}~\cite{sidky2008image} is another iterative method with a total-variation regularizer. We implement a CBCT variant of IntraTomo~\cite{zang2021intratomo}, named \textbf{IntraTomo3D}, as an example of frequency-encoding deep learning methods. \subsubsection{Implementation details} Our proposed method is implemented in PyTorch~\cite{NEURIPS2019_9015}. We use Adam optimizer with a learning rate that starts at $1\times10^{-3}$ and steps down to $1\times10^{-4}$. The batch size is 2048 rays at each iteration. The sampling quantity of each ray depends on the size of CT data. For example, we sample $192$ points along each ray for the 128$\times$128$\times$128 chest CT. We use the same hyper-parameter setting for hash encoder as~\cite{mueller2022instant}. More details of hyper-parameters can be found in the supplementary material. All experiments are conducted on a single RTX 3090 GPU. We evaluate five methods quantitatively in terms of peak signal-to-noise ratio (PNSR) and structural similarity (SSIM)~\cite{wang2004image}. PSNR (dB) statistically assesses the artifact suppression performance, while SSIM measures the perceptual difference between two signals. Higher PNSR/SSIM values represent the accurate reconstruction and vice versa. \subsection{Results} \subsubsection{Performance} Our method produces quantitatively best results in both human organ and phantom datasets as listed in Table~\ref{tab:results}. Both PSNR and SSIM values are significantly higher than other methods. For example, the PSNR value of our method in the abdomen dataset is 3.07 dB higher than that of the second-best method \textbf{SART}. We also provide visualization results of different methods in Fig.~\ref{fig:demo}. \textbf{FDK} restores low-quality models with notable artifacts, as analytical methods demand large amounts of projections. \begin{table}[h] \vspace{-1.5em} \caption{PSNR/SSIM measurements of five methods on five datasets.} \label{tab:results} \centering \begin{tabular}{m{0.25\columnwidth}<{\centering}|m{0.14\columnwidth}<{\centering}m{0.14\columnwidth}<{\centering}m{0.14\columnwidth}<{\centering}m{0.14\columnwidth}<{\centering}m{0.14\columnwidth}<{\centering}} \toprule & Chest & Jaw & Foot & Abdomen & Aorta \\ \midrule \textbf{FDK}~\cite{feldkamp1984practical} & 22.89/.78 & 28.59/.78 & 23.92/.58 & 22.39/.59 & 12.11/.21 \\ \textbf{SART}~\cite{andersen1984simultaneous} & 32.12/.95 & 32.67/.93 & 30.13/.93 & 31.38/.92 & 27.31/.77 \\ \textbf{ASD-POCS}~\cite{sidky2008image} & 29.78/.92 & 32.78/.93 & 28.67/.89 & 30.34/.91 & 27.30/.76 \\ \textbf{IntraTomo3D}~\cite{zang2021intratomo} & 31.94/.95 & 31.95/.91 & 31.43/.91 & 30.43/.90 & 29.38/.82 \\ \textbf{NAF (Ours)} & \textbf{33.05}/\textbf{.96} & \textbf{34.14}/\textbf{.94} & \textbf{31.63}/\textbf{.94} & \textbf{34.45}/\textbf{.95} & \textbf{30.34}/\textbf{.88} \\ \bottomrule \end{tabular} \vspace{-3em} \end{table} \begin{figure}[h] \centering \includegraphics[width=\linewidth]{figures/demo.pdf} \caption{Qualitative results of five methods. From left to right: examples of X-ray projections, slices of 3D CT models reconstructed by five methods, and the ground truth CT slices.} \label{fig:demo} \end{figure} Iterative method \textbf{SART} suppresses noise in the sacrifice of losing certain details. The reconstruction results of \textbf{ASD-POCS} are heavily smeared because total-variation regularization encourages removing high-frequency details, including unwanted noise and expected tiny structures. \textbf{IntraTomo3D} produces clean results. However, edges between media are slightly blurred, which shows that the frequency encoder fails to teach the network to focus on edges. With the help of hash encoding, results of the proposed \textbf{NAF} have the most details, clearest edges and fewest artifacts. Fig.~\ref{fig:slice} indicates that \textbf{NAF} outperforms other methods in all slices of the reconstructed CT volume. Figure~\ref{fig:numView} shows the performance of iterative methods and learning-based methods under different number of views. It is clear that the performance increases with the rise of input views. Our methods achieves better results than others under most circumstances. \subsubsection{Time} We record the running time of iterative and learning-based methods as shown in Fig.~\ref{fig:time}. All methods use CUDA~\cite{cuda} to accelerate the computation process. Overall, the methods spend less time on datasets with small projections (chest, jaw and foot) and increasingly more time on big datasets (abdomen and aorta). \textbf{IntraTomo3D} requires more than one hour to train the network. Benefiting from the compact network design, \textbf{NAF} spends similar running time to iterative methods and is 3$\times$ faster than the frequency-encoding deep learning method \textbf{IntraTomo3D}. \begin{figure}[t] \centering \begin{minipage}[t]{0.47\linewidth} \centering \includegraphics[width=\linewidth]{figures/slice_abdomen_50.pdf} \caption{Slice-wise performance of iterative and learning-based methods on the abdomen dataset.} \label{fig:slice} \end{minipage} \hspace{4mm} \begin{minipage}[t]{0.47\linewidth} \centering \includegraphics[width=\linewidth]{figures/numView.pdf} \caption{Performance under different number of views on the abdomen dataset.} \label{fig:numView} \end{minipage} \vspace{-1em} \end{figure} \begin{figure}[t] \centering \includegraphics[width=\linewidth]{figures/time.pdf} \caption{Running time that iterative and learn-based methods take to converge to stable results.} \label{fig:time} \end{figure} \section{Conclusion} This paper proposes NAF, a fast self-supervised learning-based solution for sparse-view CBCT reconstruction. Our method trains a fully-connected deep neural network that consumes a 3D spatial coordinate and outputs the attenuation coefficient at that location. NAF synthesizes projections by attenuating incident X-rays based on the predicted attenuation coefficients. The network is updated by minimizing the projection error. We show that frequency encoding is not computationally efficient for tomographic reconstruction tasks. As an alternative, a learning-based encoder entitled hash encoding is adopted to extract valuable features. Experimental results on human organ and phantom datasets indicate that the proposed method achieves significantly better results than other baselines and spends reasonably short computation time. \bibliographystyle{splncs04}

89b231788e6c4c533322c216731caf90d53fe677

\section{Introduction} In general, there are two main paradigms to do text summarization: \textit{abstractive}~\cite{rush2015neural,nallapati2016abstractive,gehrmann2018bottom} and \textit{extractive}~\cite{cheng2016neural,narayan2018ranking, zhong2019searching, zhong2022unsupervised} methods. For extractive summarization, previous studies~\cite{nallapati2017summarunner, liu2019text} formulate it as a \textit{sentence-level} sequence labeling task. However, there is an inherent gap between the \textit{sentence-level} scoring and the \textit{summary-level} evaluation~\cite{DBLP:conf/acl/ZhongLCWQH20}.This means that some high-scoring sentences may share the same meaning, making them not a qualified summary when combined. Similarly, the previous training paradigm for abstractive summarization models can be viewed as a \textit{token-level} scoring process upon the decoder of sequence-to-sequence model. There also exists the issue of exposure bias~\cite{bengio2015scheduled, paulus2017deep} in the teacher-forcing framework leading to the error accumulation during auto-regressive decoding. Therefore, previous frameworks for both extractive and abstractive methods did not perform \textit{summary-level} optimization. To tackle this problem, state-of-the-art summarization systems~\cite{DBLP:conf/acl/ZhongLCWQH20, liu2021simcls} are enhanced with an additional module (called re-ranker) and follow a two-stage paradigm. They first train a summarizer to model the conditional distribution $p(Y|X)$ where $X$ is the document and $Y$ is the output summary. Then the re-ranker is trained to re-score candidates sampled from the pre-trained summarizer in the second stage. However, this paradigm trades efficiency for accuracy, the auxiliary re-ranking greatly harms the inference efficiency especially for the highly efficient extractive systems. Experimentally, the decoding speed of two-stage re-ranking models is only \textasciitilde{{7.0}} samples/s while removing the re-ranker module will greatly boost the decoding speed to \textasciitilde{\textbf{42.0}} samples/s\footnote{We run these two models on the test set of CNN/DailyMail using single GeForce GTX TITAN XP GPU for 3 times and report the average speed.}. This makes two-stage summarization systems may be unacceptable in real-world scenarios that require timely feedback. The limitations of the existing work motivate us to build a one-stage summarization system that can 1) replace previous naive sentence/token-level score with a summary-level score and 2) do not sacrifice the parameter and inference efficiency. In this paper, we propose a \textbf{Co}ntrastive \textbf{L}earning based re-ranking framework for \textbf{o}ne-stage summarization called \textsc{CoLo} for both extractive and abstractive approach. Contrastive learning has been explored in summarization~\cite{sun2021alleviating,an2021retrievalsum} and generation~\cite{lee2020contrastive, an2022cont}. \textsc{CoLo} uses a contrastive re-ranking training objective. We first present a novel sampling method that can be equipped to any one-stage summarization systems so that it can re-score candidates without the second stage. The existing two-stage models use \textbf{offline sampling} to preprocess samples for training of re-ranker where candidate samples are drawn from a fixed model distribution. This is a huge obstacle to turning \textit{summarize-then-rerank} two-stage framework into an efficient end-to-end model. To solve this issue, we propose an \textbf{online sampling} approach. Concretely, instead of sampling from a fixed distribution, we draw positive and negative samples from a dynamic distribution of model outputs during training, which ultimately eliminates the requirement for additional modules in the overall framework. We then introduce a summary-level optimization strategy in addition to the traditional sentence-level (for extractive systems) or token-level loss (for abstractive systems). As a result, as a one-stage model, \textsc{CoLo} achieves comparable performance to two-stage systems, and greatly improves decoding speed to meet the needs of real-world applications. We summarize our contributions as follows: \begin{itemize} \item We are the first to propose a one-stage re-ranking framework \textsc{CoLo} for both extractive and abstractive summarization systems. \item Results on the popular CNN/DailyMail benchmark show that both the extractive and abstractive versions of \textsc{Colo} outperform previous state-of-the-art one-stage systems by a large margin. Compared to the two-stage systems, \textsc{CoLo} achieves comparable performance without additional pre-trained model. More importantly, \textsc{Colo} do not sacrifice inference speed and thus can be more widely used in real-world scenarios. \end{itemize} \section{Background} \subsection{Preliminary about Two-Stage Systems } Two-stage paradigms~\cite{DBLP:conf/acl/ZhongLCWQH20,liu2021simcls} improve summarization quality by re-ranking and selecting a candidate from a given set of candidates. MatchSum~\cite{DBLP:conf/acl/ZhongLCWQH20} forms a contrastive learning based re-ranking framework where they first generate a set of candidates summaries by a extractive summarization model and then feed them to a re-ranker. The re-ranker is trained to optimize a summary-level score and it can evaluate the candidate summaries holistically. SimCLS~\cite{liu2021simcls} is the abstractive version which replaces the extractive summarizer in \citet{DBLP:conf/acl/ZhongLCWQH20} with a abstractive summarizer. The training objective for summarization models is to estimate a conditional probability distribution $p(Y|X)$, where $X$ is the document and $Y$ is the output summary. Given a summarization model $\mathcal{M}$ that has already tuned under the conventional framework with loss function $\mathcal{L}_{sum}$ where $\mathcal{L}_{sum}$ could be binary cross entropy loss (BCELoss) or negative log likelihood loss (NLLLoss). The two-stage systems should first use a sampling algorithm e.g. beam search to sample a candidate set $\mathcal{C} = \{C_1, C_2, \ldots, C_m\}$ of size $m$ from the fixed model distribution $C_i \sim p_{\mathcal{M}}(Y|X)$. Candidates in $\mathcal{C}$ are sort by their ROUGE score in descending order. Then the they further train a separate re-ranker,e.g., BERT , with a contrastive-style ranking loss $\mathcal{L}_{rank}$ to select the the best candidate from $\mathcal{C}$ as the final output. The ranking loss used in the best re-ranking system for summarization is the triplet margin loss~\cite{kingma2014adam}. For a candidate pair $(C_i, C_j)$ where $i<j$, if $C_i$ has higher ROUGE score and it will be treated as the positive sample: \begin{equation} \mathcal{L}_{i,j} = max\{0, \text{cos}(\mathbf{z}_X, \mathbf{z}_{C_i}) - \text{cos}(\mathbf{z}_X, \mathbf{z}_{C_j}) + \rho \}, \label{eq:loss_rank} \end{equation} where $\mathbf{z}_X, \, \mathbf{z}_{C_i},\, \mathbf{z}_{C_j}$ are the vector feature representation of $X, C_i, C_j$ output by the re-ranker, and $\rho$ is the margin value. The final ranking loss is obtained by summing up all pairs: $\mathcal{L}_{rank} = \sum_j\sum_{i<j} \mathcal{L}_{i,j}$. The ranking loss ensures that candidates with higher ROUGE score is closer to the document in the embedding space. \subsection{A Comparison between Two-Stage Systems and \textsc{CoLo}} Figure~\ref{fig:compare} illustrates the difference between the architecture of two-stage systems and \textsc{CoLo}. Although MatchSum and SimCLS significantly outperform all one-stage models, they mainly suffer from three drawbacks which strongly emphasize the necessity of designing an one-stage model: (1) Training/inference inefficiency. Building the training set of the re-ranker and the second training stage consumes large amounts of GPU and CPU time (see details in Section~\ref{sec:effi}). Moreover, the need of re-feeding generation results to another module also requires unaffordable computational resources. (2) Coupling between the summarizer and re-ranker. Each improvement to one of these modules requires simultaneous updating or retraining of another module, which limits the use of such systems in the real world. For example, to try a larger candidate set or a different decoding method, we have to prepare the training set again for the second stage. In addition, how to tune the hyperparameters to be optimal in both modules at the same time is another tricky issue. Compared with two-stage systems, our one-stage system has a simple and clean implementation. (3) Two-stage systems also face difficulties in long document summarization, because the input length of the re-ranker will drastically increase as the length of candidates increasing (see detailed analysis in Appendix~\ref{sec:long_doc}). Correspondingly, \textsc{CoLo} is not easily affected by length variance. \newcommand{\unskip \vrule width 0.5pt}{\unskip \vrule width 0.5pt} \begin{figure*}[ht!] \centering \subfigure[Two-stage models: MatchSum and SimCLS]{ \label{fig:rerank} \includegraphics[width=0.42\textwidth]{figures/rerank.pdf} } \unskip \vrule width 0.5pt \subfigure[CoLo (this work)]{ \label{fig:CoLo} \includegraphics[width=0.26\textwidth]{figures/CoLo.pdf} } \caption{ A comparison between two-stage models and \textsc{CoLo}. The two-stage models including two training stages and a time-consuming preprocess while \textsc{CoLo} is trained in an end-to-end fashion. (GPU and CPU hours cost in each stage are shown in Table~\ref{tab:training effi}). Two-stage models use offline sampling to build positive-negative pairs while \textsc{CoLo} builds positive-negative pairs with online sampling where we directly get theses pairs from a changing model distribution. } \label{fig:compare} \end{figure*} \section{Method} \subsection{A Naive One-Stage Re-ranking Model} The goal of one-stage re-ranking systems is to enable both training and inference to score candidate summaries holistically without requiring a second stage of computation by a separate model. Ideally, an one-stage summarization model should both function as a summarizer and a re-ranker. A straightforward solution is multi-task learning. The naive training pipeline can be formulated as follows: (i) tuning $\mathcal{M}$ with $\mathcal{L}_{sum}$. (ii) Getting positive and negative samples from $p_{\mathcal{M}}(Y|X)$ via offline sampling for each datapoint $X$ in the training set. (iii) Building the ranking loss with these candidates and further tuning $\mathcal{M}$ with $\mathcal{L}_{rank} + \mathcal{L}_{sum}$. However, in practice, such training method is always suboptimal compared to the state-of-the-art two-stage models. We denote the model after multi-task learning as $\mathcal{M'}$. There is a serious \textit{generalization error} in the naive methods: via multi-task learning, $\mathcal{M}'$ is only able to rank candidates drawn from the original model distribution $p_{\mathcal{M}}(Y|X)$ but not candidates from the new distribution $p_{\mathcal{M'}}(Y|X)$. This error makes the naive approach unable to directly output a good summary in sequence-level generated by itself. \subsection{Our approach: \textsc{CoLo}}\label{ext_detail} The first step of CoLo is also to train the summarization model with $\mathcal{L}_{sum}$ like the naive approach. In CoLo, we discard using positive-negative samples that from a fixed model distribution, instead, we sample these candidates from a constantly shifting model distribution during multi-task learning. By doing so, we can mitigate the above mentioned generalization error as much as possible because candidates are dynamically changing with the parameters of the model distribution $p_{\mathcal{M}}(Y|X)$ updated by gradient descent. To implement this process, at each training step, we sample the newest candidates along with their feature presentations from the summarization model and calculate the ranking loss. We will give a detailed description about how we performing the online sampling process on mainstream extractive and abstractive summarization models in the following parts. \paragraph{Online Sampling for Extractive Model} The task of extractive summarization is to assign a label $y_i \in \{0,1\}$ for each sentence ${sent}_{i}$ from the source document $X$ $ = ({sent}_{1}, {sent}_{2},\ldots,{sent}_{n})$ consisting of $n$ sentences. Figure~\ref{fig:ext_model} gives an example of our one-stage extractive summarization model. Extractive candidates can be viewed as a subset of sentences from the document. In this figure, we sample $sent_1, sent_2$ to form the first candidate $C_1 = \{sent_1, sent_2\}$, and $C_2$ is consisting of $\{sent_2, sent_3\}$. After constructing these candidates, the next step is to represent them in the embedding space. In our one-stage model, we employ a heuristic way to obtain the feature presentations of candidates: pooling results of the sentence embedding from the extractive model. Concretely, we denote the sentence embedding for the $i$-th sentence as $\mathbf{h}_i$. The hidden representation of a candidate is created by pooling the sentence representations belong to it. For example $\mathbf{z}_{C_1}$ is the average pooling result of $\mathbf{h}_1$ and $\mathbf{h}_2$. Suppose $C_2$ has higher ROUGE score than $C_1$, then $C_2$ is treated as a positive sample and $C_1$ is treated as a negative sample for this pair. Finally, the whole system is trained by the sum of $\mathcal{L}_{rank}$ and $\mathcal{L}_{sum}$. Sampling informative candidates is essential in re-ranking systems. The first step of the sampling method is to determine ${\mathcal{N}}$ which represents the number of candidate sentences. ${\mathcal{N}}$ is set depending on the number of summary sentences of downstream datasets. Take CNN/DailyMail as an example, we set ${\mathcal{N}}$ to $\{2,3\}$ because most gold summaries consist of 2$\sim$3 sentences. At each training step, we iterate over $\mathcal{N}$ by combination and form $m$ different candidates $\mathcal{C} = \{C_1, C_2, \ldots, C_m\}$. $ m$ is equal to $\sum_iC_{n}^{num_i}$ where $num_i$ is the $i$-th element in ${\mathcal{N}}$ and $n$ is number of sentences of the document. For CNN/DailyMail whose ${\mathcal{N}}$ is set to $\{2,3\}$, we can sample $C_n^2 + C_n^3$ different candidates. However, in practice, we always face the combination explosion problem when the number of sentences $n$ grows larger. The two-stage system~\cite{DBLP:conf/acl/ZhongLCWQH20} pre-trained an extractive model to clip the origin number of sentences to an acceptable size. Notice that our extractive summarizer is also supervised with the BCELoss, so that we can clip the sampling space to $n'$ (a hyperparameter) with the output distribution over the sentences at each training step. Then the total size of the final candidate set decreases to $m' = \sum_iC_{n'}^{num_i}$. For CNN/DailyMail, $n'$ is set to 5, and we can get $C_5^2 + C_5^3 = 20 $ different extractive candidates. Details about the setting of $\mathcal{N}$ and $n'$ can be found in Table~\ref{tab:candidate_size} in Appendix. Notably, the offline sampling needs to feed each candidate into the pre-trained encoder. In real-life setting, when summarizing some long documents, the number of sentences in the input document and output summary will increase significantly. It will bring a polynomial level increase to the computation and GPU overhead of the two-stage model. But our one-stage system with online sampling is robust to the length variance. \paragraph{Inference Stage of Extractive Model} Since we have modeled a summary-level score during training, it is easy to directly generate summaries according to the summary-level semantic score. Concretely, given a candidate set $\mathcal{C}$ built by the combination strategy, we calculate the cosine similarity between each candidate presentation $\mathbf{z}_{C_i}$ and the document representation $\mathbf{z}_X$: \begin{equation} \label{eq:ext_inf} \hat{C} = \max_{C_i \in \mathcal{C}}{cos( \mathbf{z}_{X},\mathbf{z}_{C_i})}. \end{equation} The final output is the candidate with highest cosine similarity score. \begin{figure}[t] \centering \includegraphics[width=0.9\linewidth]{figures/extractive.pdf} \caption{Architecture of our extractive model. Input sequence: The `[doc]' token is used to get vector representation $\mathbf{z}_{X}$ of the document $X$, `[sep]' is used as separator for sentences. We omit the classifier and the BCELoss. $\mathbf{h}_i$ is the sentence embedding the i-$th$ sentence in $X$. $\mathbf{z}_{C_i}$ means the feature representation of the $i$-th candidate.} \label{fig:ext_model} \end{figure} \paragraph{Online Sampling for Abstractive Model} Our method can also be easily adapted in abstractive summarization. Selecting a generated summary maximum a posteriori (MAP) usually result in poor performance~\cite{stahlberg2019nmt}, thus most state-of-the-art generation model usually use the beam search algorithm at inference stage. The online sampling for the abstractive version is much simpler than the extractive version. We use beam search as sampling algorithm and get the feature representations from the encoder/decoder output. We denote the encoder output of source document $X$ as $H_{enc}$ and the decoder hidden states of the target summary as $H_{dec}$. We get the document representation from the encoder output of the 0-th token $\mathbf{z}_{X} = H_{enc}^0$. The feature representation of the $i$-th candidate $C_i$ with length = $|C_i|$ is derived from the last step of the decoder output $\mathbf{z}_{C_i} = H_{dec}^{|C_i|-1}$. Hidden states of other steps can not represent the entire sequence because of the sequence mask in transformer decoder. finally we formulate the ranking loss following Eq.~\ref{eq:loss_rank}. \paragraph{Inference Stage of Abstractive Model} The inference stage of our abstractive version is similar to the extractive version. We save the feature representation of the document and each beam during beam search. The final output is determind by the cosine distance between $\mathbf{z}_{X}$ and $\mathbf{z}_{C_i}$. \section{Experimental Setup} \renewcommand\arraystretch{1.4} \begin{table}[t]\footnotesize\setlength{\tabcolsep}{2.3pt} \centering \begin{tabular}{lccccc} \toprule \qquad &\textbf{CNN/DM} & \textbf{Reddit} & \textbf{XSum} & \textbf{SSN} & \textbf{PubMed} \\ \midrule $n'$ & 5 & 5 & 5 & 8 & 8 \\ $ {\mathcal{N}} $ & 2,3 & 1,2 & 1,2 & 6 & 6,7 \\ $|\mathcal{C}$| & 20 & 15 & 15 & 28 & 36 \\ \bottomrule \end{tabular}% \caption{candidate size $|\mathcal{C}|$ of each datasets (extractive). $|n'|$ is the clipped candidate size, ${\mathcal{N}}$ is a set containing all number of possible sentence.} \label{tab:candidate_size} \end{table} \subsection{Datasets} We conduct experiments on five mainstream datasets to evaluate the effectiveness of our approach. \\ \textbf{CNN/DailyMail} \cite{hermann2015teaching} is a classic benchmark which contains articles from the CNN/Daily Mail newspapers. We use the cased version from datasets\footnote{\url{https://github.com/huggingface/datasets}} \\ \textbf{XSum}~\cite{narayan2018don} is a one-sentence summary dataset from BBC News. Gold summaries are professionally written by the authors of documents.\\ \textbf{Reddit}~\cite{kim2019abstractive} is collected from social media platform and we use the TIFU-long version.\\ \textbf{PubMed}~\cite{cohan2018discourse} is a long document summarization dataset from scientific domain whose $avg$ summary length is about 4 times longer than CNN/DM.\\ \textbf{SSN}~\cite{an2021enhancing} consists of papers mainly from math, physics and computer science with the abstract section as gold reference.\\ \subsection{Implementation Details} For the simplity of experimental settings, both extractive model and abstractive mode are based on BART. We use the encoder of BART (170M) as the backbone and a 3-layer MLP as the classifier to implement the extractor. We add two special token `<cls>' to generate the sentence representation and `<sep>' as sentence separator. `<doc>' token is used to generate the document feature representation. candidate size for each dataset can be found in~\ref{tab:candidate_size} We use adam optimizer~\cite{kingma2014adam} learning rate schedule follows the setting in transformer~\cite{vaswani2017attention}. We train our model for 15000 steps with BCELoss and 32000 steps with BCELoss and RankingLoss where each step has a batch size of 36. The margin parameter $\gamma$ is set to 0.01. The size of generated candidates $|\mathcal{C}|$ is set to 20 for CNN/DM. We report the results. Other settings follow the default setting in \citet{liu2019text}. Our model is trained on single GeForce RTX 3090 GPU for 8 hours. Both our abstractive model and extractive model are trained on 24G GeForce RTX 3090 GPUs and the inference process is on 12G GeForce GTX TITAN XP GPUs. For abstractive model, we choose BART initialized with facebook/bart-large-cnn from transformers\footnote{\url{https://github.com/huggingface/transformers}} as the basic summarizer. We further fintune this model by NLLLoss and RankingLoss for 15000 steps where each step with a batch size of 8. Other setting is the same with our extractive version. To encourage diversity, we use the diverse beam search~\cite{vijayakumar2016diverse} to generate the candidates with beam size set to 16 and diversity penalty set to 1.0. Our model is trained on 8 GeForce RTX 3090 GPUs for about 18 hours. \subsection{Evaluation Metrics} We examine our approach with 4 metrics that measure the distance between generated summaries against the gold reference. \textbf{ROUGE}~\cite{lin2004rouge} where R-1 and R-2 measure informativeness based on n-gram overlapping and R-L represents fluency. \textbf{JS-2 Divergence}~\cite{louis2013automatically} measures Jensen-Shannon divergence between the bigram distributions of two input texts. \textbf{BERTScore}~\cite{zhang2019bertscore} measures soft overlap between BERT embeddings of two texts instead of using lexical matching methods. \textbf{MoverScore}~\cite{zhao2019moverscore} is also based on the neural model but applies a earth mover distance measure to contextualized BERT embeddings. \section{Results} We denote the model without contrastive learning as the baseline system. Since the backbone of our extractive model is BART encoder so that we call the baseline model \textsc{BartExt}. The baseline model for abstractive system is BART. Our extractive model is called \textsc{CoLo}$_{Ext}$ and its abstractive version is denoted as \textsc{CoLo}$_{Abs}$. \subsection{Extractive Results} We compare our models with baseline models which has similar amount of parameters and decoding speed of our models in this section. Our extractive results on CNN/DM are shown in Table~\ref{table:ext_cnndm} We compare our model with previous strong extractive baseline built on pre-trained model~\cite{zhong2019searching, bae2019summary, liu2019text} and strong multi-stage systems~\cite{DBLP:conf/acl/ZhongLCWQH20}. From the third section of Table~\ref{table:ext_cnndm}, we can see that our model \textsc{CoLo}$_{Ext}$ beats the baseline model by 1.49 ROUGE-1 score and achieve the state-of-the-art among all end-to-end systems when input length set to 512 and the results can be further improved while extending the input length to 1024. Even compared with the \textsc{BertSum}-large (340M)~\cite{liu2019text} which is built on large PTM, We still have an improvement of 0.42 with only the half number of parameters of theirs. Though RL-based methods hold the motivation of optimizing towards the evaluation metric, but it does not gain much improvement on performance in practice. To verify whether our model is effective on datasets of various lengths, we also evaluate our model on datasets with short summaries (Reddit and XSum) and long document dataset PubMed and results are shown in Table~\ref{tab:ext_other_datasets}. On reddit and XSum, we achieve the advantage of more than 1.0 point ROUGE-1 than baseline systems and close performance with the upper bound ORACLE. We also gain improvements when tested on the long document summarzation dataset PubMed. Detailed results on long document dataset can be found in Appendix~\ref{sec:long_doc}. \renewcommand\arraystretch{1.2} \begin{table}[t] \center \footnotesize \setlength{\tabcolsep}{1.05mm}{ \begin{tabular}{lccc} \toprule {\textbf{Model}} & \textbf{R-1} & \textbf{R-2} & \textbf{R-L} \\ \midrule LEAD & 40.43 & 17.62 & 36.67 \\ ORACLE & 52.59 & 31.23 & 48.87 \\ \midrule Transformer\cite{vaswani2017attention} & 40.90 &18.02 &37.17 \\ \textsc{Bert-Ext}\cite{bae2019summary} & 42.29 & 19.38 & 38.63 \\ \textsc{Bert-Ext} + RL & 42.76 & 19.87 & 39.11 \\ {BertSum}~\cite{liu2019text} & 42.57 & 19.96 & 39.04 \\ {BertSum}-large & \underline{43.85} & 20.34 & 39.90 \\ \midrule \textsc{BartExt} & 42.78 & 20.24 & 39.24 \\ \textsc{BartExt} ($len=1024$) & 43.65 & \underline{20.88} & \underline{40.19} \\ \midrule Naive one-stage & 43.53 & 20.54 & 39.62 \\ \textsc{CoLo}$_{Ext}$ & 44.10 & 20.97 & 40.19 \\ \textsc{CoLo}$_{Ext}$ + BERTScore & 44.27 & 21.01 & 40.34 \\ \textsc{CoLo}$_{Ext}$ ($len=1024$) & \textbf{44.58} & \textbf{21.25} & \textbf{40.65} \\ \bottomrule \end{tabular}} \caption{Extractive results on CNN/DM test set. $len$ means the input length of the document, results without the marker using 512 tokens as input. +RL means the addition of reinforcement learning. +BERTScore means we use BERTScore to determine positive-negative samples. \textsc{CoLo} clearly outperform all previous one-stage summarization systems. The best results are in bold and the second best ones are underlined. } \label{table:ext_cnndm} \end{table} \renewcommand\arraystretch{1.0} \begin{table*}[t] \center \footnotesize \tabcolsep0.13 in \begin{tabular}{lccccccccc} \toprule \multicolumn{1}{c}{\multirow{2}[1]{*}{\textbf{Model}}} & \multicolumn{3}{c}{\textbf{Reddit}} & \multicolumn{3}{c}{\textbf{XSum}} & \multicolumn{3}{c}{\textbf{PubMed}} \\ & \textbf{R-1} & \textbf{R-2} & \textbf{R-L} & \textbf{R-1} & \textbf{R-2} & \textbf{R-L} & \textbf{R-1} & \textbf{R-2} & \textbf{R-L} \\ \cmidrule(lr){1-1} \cmidrule(lr){2-4} \cmidrule(lr){5-7} \cmidrule(lr){8-10} LEAD & 12.38 & 2.17 & 10.12 & 14.40 & 1.46 & 10.59 & 37.58 & 12.22. & 33.44 \\ ORACLE & 29.10 & 11.08 & 23.10 & 25.62 & 7.62 & 18.72 & 45.12 & 20.33 & 40.19 \\ \cmidrule(lr){1-1} \cmidrule(lr){2-4} \cmidrule(lr){5-7} \cmidrule(lr){8-10} \textsc{BertSum} & 23.86 & 5.85 & 19.11 & 22.86 & 4.48 & 17.16 & 41.05 & 14.88 & 36.57 \\ \textsc{BartExt} & 23.97 & 5.68 & 19.24 & 22.96 & 4.70 & 17.29 &41.40 & 16.18 & 37.89 \\ \rowcolor{mygray} \textsc{CoLo}$_{Ext}$ & \textbf{25.06} & \textbf{5.90} & \textbf{19.52} & \textbf{24.51} & \textbf{5.04} & \textbf{18.21} & \textbf{41.93} & \textbf{16.51} & \textbf{38.28} \\ \bottomrule \end{tabular} \caption{Results on test sets of reddit, XSum and PubMed. Our model achieve significant improvement on the baseline model \textsc{BartExt}. LEAD means we select the first $k$ sentences from the source document as the output summary and ORACLE is the upper bound of extractive methods. } \label{tab:ext_other_datasets} \end{table*} \subsection{Abstractive results} Early work also successfully applies reinforcement learning on abstractive summarization~\cite{paulus2017deep,li2019deep}. But we do not find related works that successfully combine reinforcement learning with strong pre-trained models. Therefore, most of our baselines are strong pertrained model finetuned with NLLLoss. Our results is shown in Table~\ref{table:abs_cnndm}, due to the huge cost of using large pre-trained model with length set to 1024, we also report results with 512 input tokens and it is able to significantly outperform other baselines which has longer input length (1024). \textsc{CoLo}$_{Abs}$ has an improvement of \textbf{2.17} R-1 socre on the very strong BART-large baseline without adding additional parameters or modules. Additionally, our method is able to outperform all one-stage baseline systems by a large margin. We also conduct experiments on long document summarzation datasets (see in Table~\ref{tab:abs_long} in Appendix). \renewcommand\arraystretch{1.3} \begin{table}[t] \center \footnotesize \tabcolsep0.07in \setlength{\tabcolsep}{0.75mm}{ \begin{tabular}{lccc} \toprule {\textbf{Model}} & \textbf{R-1} & \textbf{R-2} & \textbf{R-L} \\ \midrule {BertSumAbs}\cite{liu2019text} & 41.72 &19.39 &38.76 \\ Pegasus\cite{zhang2020pegasus} & 44.17 & 21.47& 41.11 \\ BART\cite{lewis2020bart} & 44.16 & 21.28& 40.90 \\ BART+R3F\cite{aghajanyan2020better} & {44.38} & {21.53} & {41.17} \\ BART ($len=512$) & 43.82 & 20.96 & 40.63 \\ \midrule ConSum~\cite{sun2021alleviating} & 44.53 & 21.54 & 41.57 \\ SeqCo~\cite{xu2021sequence} &\underline{45.02} & \underline{21.80} & \underline{41.75}\\ \midrule Naive one-stage (ROUGE, $len=512$) & 43.90 & 20.88 & 40.69 \\ \textsc{CoLo}$_{Abs}$ (ROUGE, $len=512$)& 45.45 & 21.53 & 42.35 \\ \textsc{CoLo}$_{Abs}$(ROUGE) & \textbf{46.33} & \textbf{22.15} & \textbf{43.08} \\ \bottomrule \end{tabular}} \caption{Abstractive results on CNN/DM test set. $len$ means the maximum input length of the encoder, results without the marker using 1024 tokens as the input. ConSum~\cite{sun2021alleviating} and SeqCo~\cite{xu2021sequence} in the second block are also previous contrastive learning based methods without re-ranking. } \label{table:abs_cnndm} \end{table} \subsection{Comparison with Multi-stage Systems}\label{sec:effi} Apart from the one-stage systems, we also compare our model with these powerful multi-stage systems: CTRLSum, multi-stage re-ranking models. CTRLSum needs other systems to previously produce a control signal. \renewcommand\arraystretch{1.2} \begin{table}[t] \center \footnotesize \setlength{\tabcolsep}{0.5mm}{ \begin{tabular}{lccc} \toprule {\textbf{Model}} & \textbf{R-1} & \textbf{R-2} & \textbf{R-L} \\ \midrule \multicolumn{4}{c}{\textit{extractive systems}} \\ \midrule \textsc{CoLo}$_{Ext}$ & 44.27 & 21.01 & 40.34 \\ BERT+BERT$^\mathcal{R}$~\cite{DBLP:conf/acl/ZhongLCWQH20} & 44.22 & 20.62 & 40.38 \\ BERT+RoBERTa$^\mathcal{R}$ ~\cite{DBLP:conf/acl/ZhongLCWQH20} & 44.41 & 20.86 & 40.55\\ \textsc{CoLo}$_{Ext}$ +RoBERTa$^\mathcal{R}$ & \textbf{44.70} & \textbf{21.03} & \textbf{40.74}\\ \midrule \multicolumn{4}{c}{\textit{abstractive systems}} \\ \midrule \textsc{CoLo}$_{Abs}$ & 46.33 & 22.15 & 43.08 \\ {CTRLSum}\cite{he2020ctrlsum} & 45.65 & \textbf{22.35} & 42.50\\ BART+RoBERTa$^\mathcal{R}$~\cite{liu2021simcls} & \textbf{46.67} & 22.15 & \textbf{43.54} \\ \bottomrule \end{tabular}} \caption{ Comparision with the multi-stage systems. RoBERTa$^\mathcal{R}$ means a RoBERTa re-ranker is is added to the summarization model. } \label{table:multi-stage} \end{table} \paragraph{performance} The addition of another pre-trained model implicitly introduces more parameters and knowledge, thus it is usually unfair to directly compare one-stage systems with the two-stage systems. But we show that \textsc{CoLo} is able to achieve comparable performance with the multi-stage systems. As is shown in the first part of Table~\ref{table:multi-stage}, compared with the multi-stage models that ensembles another pre-trained encoder as a re-ranker, \textsc{CoLo}$_{Ext}$ still performs better than their BERT+BERT$^\mathcal{R}$ version without the need to re-feed the generated candidates to another model meanwhile we obtain a \textasciitilde$\times 5$ speed up over the multi-stage systems. We also try concatenating a re-ranker RoBERTa for our model, results shows that \textsc{CoLo}$_{Ext}$ can be further improved by combing another pre-trained re-ranker reaching new extractive SOTA on the test set of CNN/DM. For abstractive models, our end-to-end model still legs behind multi-stage systems but we do not need training another model and keep similar inference speed with baseline models. \paragraph{Inference Efficiency} Despite the fact that multi-stage models outperform all end-to-end systems, they frequently suffer from inefficiency. In this part we mainly focus on analysing the efficiency of 3 kinds of systems: 1) baseline, which is trained only with BCELoss or NLLLoss, 2) \textsc{CoLo}, our end-to-end constrastive learning framework, 3) Rerank, which means the multi-stage re-ranking systems. it has more 110M parameters than baseline model and \textsc{CoLo}. The efficiency experiments for training and inference are respectively conducted on 24G RTX 3090 GPUs and 12G TITAN XP GPUs. For extractve summarization, figures~\ref{fig:cnn1},\ref{fig:cnnm} give a detailed comparison of the inference speed between the three models. Y-axis represents the number of samples processed per second. To give a fair comparison, we test the inference efficiency in two settings: i) all models are tested with batch size fixed to 1. ii) all models are tested with the maximum batch size allowed by the GPU. While the candidate size varies from 4$\sim$32, both our model have a $\mathbf{3\times \sim 8\times}$ speed-up ratio over the multi-stage re-ranking model. When the candidate size is set to 20, the baseline model is able to process \textasciitilde31.2/41.9 (batch = 1/\textit{MAX}) samples per second, the decoding speed of \textsc{CoLo}$_{Ext}$ is \textasciitilde30.4/38.9 samples/s (batch=1/\textit{MAX}) and the decoding speed of the multi-stage re-ranking model is only \textasciitilde4.9/7.0 samples/s(batch=1/\textit{MAX}). Our model almost does no harm on inference speed while the candidate size $|\mathcal{C}|$ is less than 16. However, when the candidate size grows larger there is more time spent on generating the representations of the candidates. Figure~\ref{fig:abs_time} show the comparison of inference time of the abstractive models. While the bottleneck of abstractive models is the auto-regressive generation process. Our abstractive model generally save \textasciitilde0.5 GPU hours compared to the re-ranking model. \begin{figure}[ht!] \centering \subfigure[CNN/DM (batch=1)]{ \label{fig:cnn1} \includegraphics[width=0.22\textwidth]{figures/cnndm_time.pdf} } \subfigure[CNN/DM (batch =\textit{MAX})]{ \label{fig:cnnm} \includegraphics[width=0.22\textwidth]{figures/cnndm_full_time.pdf} } \caption{Inference speed on CNN/DM (extractive). we use the candidate size $|\mathcal{C}|$ as the X-axis. The Y-axis represents the number of samples processed per second. batch=\textit{{MAX}} means we use the maximum batch size allowed by GPU memory. } \label{fig:decoding_time} \end{figure} \begin{table}[t]\footnotesize\setlength{ \tabcolsep}{1.5pt} \centering \begin{tabular}{lcccc} \toprule \textbf{Systems} &\textbf{Stage1} & \textbf{Preprocess} & \textbf{Stage2} & \textbf{Total hours}\\ \midrule Ext+RoBERTa$^\mathcal{R}$ & 4 & 5 (+20) & 128 & 137 (+20) \\ \textsc{CoLo}$_{Ext}$ & 7 & -- & -- & 7 ($\downarrow\mathbf{130}$) \\ \midrule Abs+RoBERTa$^\mathcal{R}$ & 80 & 132 (+18) & 128 & 340 (+18) \\ \textsc{CoLo}$_{Abs}$ & 224 & -- & -- & 224 ($\downarrow\mathbf{116}$) \\ \bottomrule \end{tabular}% \caption{ GPU hours spent on training for each process on the training set of CNN/DM(reported results are rounded down after the decimal point. Ext+RoBERTa$^\mathcal{R}$/Abs+RoBERTa$^\mathcal{R}$ denotes the multi-stage re-ranking systems with an extracitve/abstrastive summarizer. (+18)/(+20) means 18/20 CPU hours are spent on calculate ROUGE score for each candidate with 32 threads. } \label{tab:training effi} \end{table} \paragraph{Training Efficiency} Table~\ref{tab:training effi} gives an overview of the training time of our system and the multi-stage models on the training set of CNN/DM. The general pipeline for the multi-stage models is: i) training a generator {(Stage1)}, ii) {Preprocess}, ii) training a re-ranker (Stage2). The preprocess includes generating the training/dev/test set for training re-ranker and sorting candidates by ROUGE. For extractive system we save \textbf{130} GPU hours compared to the multi-stage systems whose bottleneck is training the re-ranking model. For abstractive model, apart from the 128 GPU hours spent on training the ranker, using beam search to generate the training set for re-ranker model is also very time consuming, generally we obtain \textbf{116} GPU hours and 18 CPU hours saved. \begin{figure}[t] \centering \includegraphics[width=0.7\linewidth]{figures/abs_time.pdf} \caption{ Test inference time with beam size for abstractive model. We use the maximum batch size allowed by GPU memory. } \label{fig:abs_time} \end{figure} \subsection{Ablation for Different Discriminators} In addition to ROUGE, we also select other metrics as the discriminator (shown in Table~\ref{table:cmp_metrics}). ROUGE and JS-2 is based on lexical matching while BERTScore and MoverScore are based on the contextualized embedding from BERT. Our model generally obtains the best results on the metric used in training. Because these metrics are not actually separated, using one of these metrics as the discriminator can also gain significant improvements on other metrics. Overall, the neural evaluation metric BERTScore and MoverScore bring more improvements compared with metrics that based on the lexical matching. But incorporating neural model based metrics in training will obviously increase the training time. \renewcommand\arraystretch{1.2} \begin{table}[t] \center \footnotesize \tabcolsep0.07in \setlength{\tabcolsep}{1.2mm}{ \begin{tabular}{lcccccc} \toprule {\textbf{Metric Used}} & \textbf{R-1} & \textbf{R-2} & \textbf{R-L}& \textbf{JS-2} & \textbf{BS} & \textbf{MS} \\ \midrule Baseline & 42.78 & 20.24 & 39.23 & 54.24 & 43.52 & 58.27 \\ ROUGE-1,2 & \cellcolor[gray]{0.9} 44.10 & \cellcolor[gray]{0.9} 20.97 & 40.19 & 54.07 & 44.26 & 58.63 \\ ROUGE-L & 44.09 & 20.93 & \cellcolor[gray]{0.9} \bf40.34 & 54.06 & 44.32 & 58.60 \\ JS-2 & 43.85 & \textbf{21.13} & 39.98 & \cellcolor[gray]{0.9} \textbf{53.92} & 44.19 & 58.60 \\ BERTScore & \textbf{44.27} & 21.01 & \textbf{40.34} & 54.08 & \cellcolor[gray]{0.9}\textbf{44.85} & 58.71 \\ MoverScore & 44.21 & 20.81 & 40.25 & 54.33 & 44.47 & \cellcolor[gray]{0.9}\textbf{58.78} \\ \bottomrule \end{tabular}} \caption{Extractive results of using different evaluation metrics as the discriminator on CNN/DM test set. } \label{table:cmp_metrics} \end{table} \subsection{Visualization Experiment} We conduct a visualization experiment on our extractive model to get a close look on the distribution of candidates in semantic space. We randomly sample 100 documents with more than 10 sentences from the test set of CNN/DM. We first select the top 10 sentences based on the predicted score from the classifier. We set the possible number of sentences to $\{2, 3\}$ resulting a candidate size of $C_{10}^2 + C_{10}^3 = 165$ for each sample. We visualize the learned embedding of these candidates and the anchor in a two-dimensional space by applying the t-SNE algorithm. As shown in Figure~\ref{fig:viz}, there is an obvious cluster of the top 50 candidates (colored in purple) and the candidates with higher score are closer to the anchor while the distribution of uninformative candidates (gray,cyan points) is relatively random. \begin{figure}[ht!] \centering \subfigure{ \includegraphics[width=0.225\textwidth]{figures/case1.png} } \subfigure{ \includegraphics[width=0.225\textwidth]{figures/case2.png} } \caption{T-SNE Visualization of two examples from CNN/DM test set. We divide the candidates into 3 groups based on ROUGE score: candidates ranking 1\textasciitilde50, candidates ranking 51\textasciitilde100, candidate ranking 101\textasciitilde150. The red point denotes the anchor and the purple/cyan/gray points respectively denote the top 50/100/150 candidates.} \label{fig:viz} \end{figure} \subsection{Human Evaluation}\label{human_eval} We also conduct a human evaluation on our models to get more accurate results . We randomly select 30 articles from the test set of CNN/DM, and each articles have 5 candidate summaries 4 from automatic systems and 1 is the gold reference. We recruit 2 PhD students majoring in computer science and ask them to rank the candidate summries based on the fluency, informativeness. If two of these systems generate the same summary for the source document, this sample will be filtered out. As we can see from Table~\ref{tab:human_evaluation}, the \textsc{CoLo}$_{Ext}$ with the discriminator as BERTScore achieve the best result among all automatic systems. However, using BERTScore will bring much training time. We also suggest taking JS-2 divergence as the discriminator which also does a good job in human evaluation. \renewcommand\arraystretch{1.2} \begin{table}[!ht] \center \footnotesize \tabcolsep0.05in \setlength{\tabcolsep}{0.9mm}{ \begin{tabular}{lcccccc} \toprule \textbf{Metric Used} & \textbf{1st} & \textbf{2nd} & \textbf{3rd} & \textbf{4th} & \textbf{5th} & \textbf{Avg R.} \\ \midrule Baseline & 0\% & 8.3\%& 8.3\% & 23.3\% & 60\% & 4.33\\ JS2 & 6.7\% & 25\% & 33.3\%& 21.7\% & 13.3\% & 3.10\\ R1+R2 & 5\% & 20\% & 28.3\% & 30.3\% & 16.7\% & 3.35\\ BERTScore & 10\% & 35\% & 20\%& 25\% & 10\% & 2.90\\ Gold label & 78.3\% & 11.7\% & 10\% & 0\% & 0\% & 1.32\\ \bottomrule \end{tabular}} \caption{ Results of human evaluation results. Baseline means the \textsc{BartExt} model, Gold-label means the means the human written summary. Avg R. denotes the average ranking of the system. } \label{tab:human_evaluation} \end{table} \section{Limitations and Future Work} Compared with the most well-known contrastive learning framework simCLR~\cite{chen2020simple} which propose to construct positive and negative pairs from training samples in the same batch, Drawing negative-positive pairs from the summarization model requires more training time. Ideally, providing more positive and negative samples will benefit the performance of \textsc{CoLo} . However, decoding with very large beam size in training mode will cost more GPU memory and training time. Future work can search for an efficient way to construct these positive-negative pairs to perform re-ranking during training. \section{Conclusion} We introduce \textsc{CoLo}, a contrastive learning based summarization framework for one-stage summarization where positive-negative pairs are generated directly from the summaizer with online sampling. \textsc{CoLo} can be both easily applied on extractive and abstractive methods. Results show that we greatly exceed previous stage-of-the art one-stage systems with no additional parameters and obivious decline of the inference efficiency. \section*{Acknowledgement} We would like to thank Yixin Liu and the anonymous reviewers for their valuable advice. This work was supported by the National Key Research and Development Program of China (No.2020AAA0106702) and National Natural Science Foundation of China (No.62022027).

9449bd2f9d9c066bf24e748f9a4a4c8d232fe151

\section{Supplementary Material} \subsection{Symmetries with and without static $\pm 2$ charges} \label{subsect1} In this section, we discuss the various symmetries of the model with and without the $\pm 2$ charges. Let us first discuss the case without the charges, where one has the usual lattice symmetries of translations by one lattice spacing, the various reflection and rotation symmetries which form the point group symmetries. The charge conjugation is a global internal $Z_2$ symmetry, which transforms as: $U \rightarrow U^\dagger, U^\dagger \rightarrow U, E \rightarrow -E$. Additionally, there is the global symmetry associated with the large gauge transformations, generated by the operator $W_i = \frac{1}{L_i} \sum_{x} E_{x,i}$ and classifies the physical states in winding sectors taking values in ${\mathbb{Z}}$ for even $L$. Finally, the gauge symmetry, corresponding to the choice of the Gauss Law allows only the six allowed states with zero charge at a vertex. We note that the QLM with spin $S = \frac{1}{2}$ has a different ground state phase diagram from the Wilsonian LGT. The latter uses quantum rotors as degrees of freedom, and thus the electric flux are quantized in integer units, whereas in our case the electric flux is always $\pm \frac{1}{2}$. The QLM ground state diagram has three distinct phases: two different confined phases (C1 and C2), and the staggered phase. For large negative $\lambda$, where phase C1 is stabilized, the ground state has the maximal number of flippable plaquettes (see \Cref{fig:refCL4} (left)). This phase spontaneously breaks lattice translation symmetry by one lattice spacing, as well as charge conjugation symmetry. As $\lambda$ is reduced, the $J$ term governing the flips dominate, and the system goes into a new phase (C2) where one of the two sublattices (for example, the shaded one) is in a coherent superposition of clockwise and anticlockwise flippable plaquettes, often called a resonating valence bond (RVB) solid. C2 has an unbroken charge conjugation symmetry, but a still broken lattice translation symmetry. In terms of magnetization operators (introduced in the Main text), in the phase C1 both sublattices are ordered, while C2 has ordering of one of the two sublattices. \begin{figure} \centering \includegraphics[scale=0.9]{LetterFigs/refC_4x4.pdf} \includegraphics[scale=0.9]{LetterFigs/phases_U1QLM.pdf} \caption{(Left): Reference configuration for a $L=4$ lattice. Such a configuration has the maximum number of flippable plaquettes, which is signalled by large values of both $M_{A},M_{B}$. Both the flux (on the links) and the height variables (in the dual lattices) are shown. (Right) The phase factors $\eta^{X}_{\tilde{x}}$ for the A and B sublattices which need to be multiplied with the height variables $h^X$ to obtain the magnetization.} \label{fig:refCL4} \end{figure} When the $\pm 2$ charges are included in the thermal ensemble (by appropriately modifying the Gauss Law), several of the previous global symmetries do not remain good quantum numbers any more. Since the distribution of charges in the volume is governed thermally, lattice translation, rotation and reflections are not good symmetries. Charge conjugation is explicitly broken in the presence of charges. The total number of allowed states at a vertex is now 8, out of the 16 possible states in total. The other 8 states correspond to the presence of a $\pm 1$ charge, and are realized in the quantum dimer model, but not in our model. With 8 vertex states allowed by the Gauss Law, this is an effective $Z_2$ gauge constraint on the Hamiltonian in \Cref{eq:ham}. It can also be viewed as a quantum eight-vertex model. A more mathematical argument is provided below to make the effective $Z_2$ nature of the Gauss Law explicit. The winding numbers now take values in $\mathbb{Z}/2$ for even $L$. Since the Gauss law for the pure gauge theory annihilates the physical states $\ket{\psi}$, it follows that the action of an arbitrary gauge transformation on the physical states keeps them unchanged: \begin{equation} \begin{split} V\ket{\psi} & = \prod_x \exp (i\theta_x G_x) \ket{\psi} \\ & = \prod_x \left[ 1 + i \theta_x G_x - \frac{\theta_x^2 G_x^2}{2} + \cdots \right] \ket{\psi} = \ket{\psi}. \end{split} \end{equation} The $U(1)$ nature of the gauge symmetry is because $\theta_x \in (0, 2 \pi]$ is an angle. When we allow $Q_x = 0, \pm 2$ in our theory, then we have two classes of states, $\ket{\Psi} = \left\{ \ket{\psi}, \ket{\phi} \right\} $, which satisfy, $G_x \ket{\psi} = 0$ and $G_x \ket{\phi} = \pm 2 \ket{\phi}$. Under the same gauge transformation as before, we now have \begin{equation} \begin{split} V\ket{\Psi} & = \prod_x \exp (i\theta_x G_x) \ket{\Psi} \\ & = \prod_x \left[ 1 + i \theta_x G_x -\frac{\theta_x^2 G_x^2}{2}+ \cdots \right] \ket{\Psi} \end{split} \end{equation} For the states $\ket{\psi}$ it is clearly satisfied, but for the states $\ket{\phi}$, the states are only unchanged when $\theta_x = 0, \pi$. Thus only a $Z_2$ subgroup of the original $U(1)$ survives when we demand gauge invariance for the states having $\pm 2$ charges in addition to the zero charged ones. (Note that the total charge is still zero, we are only referring to the local charges). \begin{figure} \centering \includegraphics[scale=1.0]{LetterFigs/m1m2m3m4.pdf} \\ \includegraphics[scale=0.8]{LetterFigs/qmcalgfig.pdf} \caption{The layout of the height variables and the cluster rules. (a) $m_1, m_2, m_3, m_4$ lie on a timeslice $t$, $m_5$ is on timeslice $t-1$, and $m_6$ on timeslice $t+1$. (b) and (c) Out-of-plane breakups: when there is a reference configuration, one can connect $m_5$ and $m_6$ with a certain probability only if they are the same ($A$-breakup). The $B$-breakup corresponds to the case when $m_5 \neq m_6$ and they cannot be connected. (d) and (e) In-plane breakups: If $m_5 \neq m_6$, $m_1,m_2,m_3,m_4$ are in a reference configuration and must be connected, while for $m_5 = m_6$ the connection of $m_1,m_2,m_3,m_4$ is done with only a certain probablity. } \label{fig:QLMC} \end{figure} \subsection{Dualization, quantum Monte Carlo algorithm, and order parameters} Construction of efficient quantum Monte Carlo algorithms for (lattice) gauge theories is a particularly difficult challenge due to the constraints that need to be satisfied. Stochastic local updates typically get rejected when they cannot satisfy the constraints, while global updates satisfying detailed balance are non-trivial to construct. Our model, being a lattice gauge theory is no exception. However, recently it has become possible to exploit dualization techniques in order to rewrite the problem in terms of different variables which partially solve the gauge constraints. The $U(1)$ theory in $(2+1)-$d can be dualized into a $(2+1)-$d quantum height model, for which efficient cluster algorithms can be constructed. In particular, rewriting the gauge theory in terms of the height variables completely eliminates the odd charges $Q_x = \pm 1$ from the theory written in terms of the height variables. The resulting theory only allows $Q_x = 0, \pm 2$, which explains the algorithmic ease of incorporating the $\pm 2$ charges. However, as was done in \cite{Banerjee2013}, it is easy to impose a further constraint in the cluster algorithm to project out the $\pm 2$ charges while studying the pure gauge theory. We provide an outline of the procedure below, for further details please see \cite{Banerjee2021}. \begin{figure} \includegraphics[scale=0.8]{LetterFigs/Wx0Wy0_8x8.pdf} \includegraphics[scale=0.8]{LetterFigs/Wx2Wy2_8x8.pdf} \caption{(top) A sample configuration on the $L=8$ lattice with the largest magnetization for both $M_A$ and $M_B$, which corresponds to all flippable plaquettes. Both the flux and the height configurations are shown. For this configuration, $M_A = -16$ and $M_B = 16$. (bottom) A different configuration for the $L=8$ lattice with winding numbers $W_x = W_y= 2$, showing both the height and the flux configurations, for which $M_A = M_B = 0$. The links carrying the winding strings, both in the horizontal and vertical directions, are shown in bold.} \label{fig:refCL8} \end{figure} \Cref{eq:ham} and \Cref{eq:pf} display the Hamiltonian and the partition function of our model, respectively. To proceed further, we construct the transfer matrix $\mathbb{T}=e^{-\epsilon H}$, by discretizing the temporal direction into $L_T$ time slices ($\beta = \epsilon L_T$). Moreover, noting that each link participates in two plaquette interactions, it is natural to divide the lattice into even ($A$) and odd ($B$) sublattices, which is the well-known Trotter decomposition. All the plaquettes in sublattice A (or sublattice B) can be considered to be interacting simultaneously. \Cref{fig:refCL4} shows the division into two sublattices A and B (shaded and unshaded) respectively. The Hamiltonian and partition function can now be written as, \begin{equation} H = H_A + H_B \;\;\; \text{and}\;\;\; Z=\text{Tr}\left[ \mathbb{T}_A^{L_T}\mathbb{T}_B^{L_T} P_{G}\right] \end{equation} where $\mathbb{T}_{A(B)}=\text{exp}[-\epsilon H_{A(B)}]$ and we have neglected $O(\epsilon ^2)$ terms. $P_G$ is the projection operator enforcing the Gauss law: $P_G = \prod_x \{ 6\delta(G_x) + \delta(G_x - 2) + \delta(G_x + 2)\} \frac{1}{8}$ when we need to include the charge $\pm 2$, or $P_G = \prod_x \delta (G_x)$, for the case without the charges. We can read off the weights for the various configurations from the single plaquette transfer matrix operator: \begin{equation} \begin{split} \mathbb{T}_\square & = 1 + (U_\square + U_\square^\dagger) {\rm e}^{-\epsilon\lambda J} \text{sinh}(\epsilon J) \\ & + (U_\square + U_\square^\dagger)^2 [{\rm e}^{-\epsilon \lambda J}\text{cosh}(\epsilon J) - 1 ] \end{split} \end{equation} The non flippable plaquettes (fourteen of them) have unit weights. The two flippable plaquettes give diagonal contribution of $e^{-\epsilon\lambda J} \text{cosh} (\epsilon J)$. A configuration where a plaquette is flipped has a weight $e^{-\epsilon\lambda J}\text{sinh}(\epsilon J)$ (the off-diagonal elements). On dualization, the model is reformulated in terms of quantum height variables, which live in the centres of the plaquettes (the dual lattice sites). Thus, if a site is labelled as $x = (x_1,x_2)$, the dual sites (where the dual heights are located) are at $\tilde{x}=(x_1 + \frac{1}{2}, x_2 + \frac{1}{2})$. Moreover, the height variables on different sublattices are distinct: $h_{\tilde{x}}^{\rm A}=0,1$ for even (A) sublattice and $h_{\tilde{x}}^{\rm B}=\pm \frac{1}{2}$ for odd (B) sublattice. The divergence of the height variables are the electric flux variables, \begin{equation} E_{x,\h{i}}= \left[ h_{\tilde{x}}^X-h_{\tilde{x}+\tilde{i}-\tilde{1}-\tilde{2}}^{X^\prime} \right] ~~{\rm mod2} = \pm\frac{1}{2}; ~~~ {\rm X,X}^\prime \in \{{\rm A},{\rm B}\} \end{equation} An example of the flux to height mapping is shown in \Cref{fig:refCL4} (left). We note that the above relation remains unchanged when both the height variables are flipped, and thus there are two distinct height configurations which map to a single flux configuration. Moreover, it is easy to check by constructing the height variables in the presence of charges $Q_x = \pm 1$ that the height assignment does not work out, while $Q_x = \pm 2$ does not present a problem. In other words, the dualization using the height variables projects out the $Q_x = \pm 1$. The cluster algorithm works by building clusters in a single sublattice at a time. Bonds between the variables in a sublattice (say A) are put depending on the value of the height variables in the other sublattice (therefore B), which we now discuss. In \Cref{fig:QLMC} (a), the general layout of height variables across time slices is shown: $m_1,\:m_2\:,m_3$ and $m_4$ lie in time slice $t$, $m_5$ lies in time slice $t-1$ and $m_6$ lies in time slice $t+1$. We shall consider two different breakups, the out-of-plane breakups (whether to bond $m_5$ and $m_6$) and in-plane breakups (whether to bond $m_1,\:m_2,\:m_3$ and $m_4$). If we can bond heights, we call it a $A$-breakup; otherwise it is a $B$-breakup. \begin{itemize} \item \textbf{Out-of-plane breakups}: \begin{itemize} \item If ($m_1,m_2,m_3,m_4$) are not in reference configuration, $m_5$ must be equal to $m_6$, and we must bond ($m_5,m_6$) to forbid disallowed configurations. \item If ($m_1,m_2,m_3,m_4$) are in reference configuration, either $m_5\neq m_6$ (weight is $e^{-\epsilon\lambda J}\sinh(\epsilon J)$) or $m_5=m_6$ (weight is $e^{-\epsilon\lambda J} \cosh(\epsilon J)$). If $m_5\neq m_6$, have case $B$ (see \Cref{fig:QLMC} (c)). If $m_5=m_6$, both $A$ and $B$ can happen (see \Cref{fig:QLMC}(b)). To satisfy detailed balance, we have, $P_B=B/(A+B)=\tanh(\epsilon J)$ and $P_A=1-P_B=\frac{e^{-\epsilon J}}{\cosh(\epsilon J)}$. \end{itemize} \item \textbf{In-plane breakups}: \begin{itemize} \item If $m_5\neq m_6$, ($m_1,m_2,m_3,m_4$) are in a reference configuration and we must bind them together to avoid disallowed configuration. \item If $m_5=m_6$, either ($m_1,m_2,m_3,m_4$) are in reference configuration (with weight $e^{-\epsilon\lambda J}\cosh(\epsilon J)$) or they are not (weight is $1$). If ($m_1,m_2,m_3,m_4$) form a reference configuration, both $A$ and $B$ can be applied, else, we are in case $B$ (see \Cref{fig:QLMC}). Solving the detailed balance equation we get, $P_B=B/(A+B)=\frac{e^{\epsilon\lambda J}}{\cosh(\epsilon J)}$ and $P_A=1- \frac{e^{\epsilon\lambda J}}{\cosh(\epsilon J)}$. \end{itemize} \end{itemize} The cluster building process proceeds as is usual in a spin model. It is possible to consider the single cluster Wolff algorithm, or the multi-cluster Swendsen-Wang variant. For our case, we have implemented both the algorithm and checked that the answers match. Another important point is the construction of order parameters, $M_{A}$ and $M_B$, which are constructed as $M_{X} = \frac{1}{L_T}\sum_{\tilde{x}} \eta^X_{\tilde{x}} h^X_{\tilde{x}}$, where $X=A,B$ and $\tilde{x}$ denote the dual sites (centres of the plaquettes of the original space-time lattice). The phases $\eta^X$, which need to be multiplied with the height variables $h^X$ to get the magnetization, are displayed in \Cref{fig:refCL4} (right) for $L=4$, but for any other lattice (with a multiple of 4), they can be easily tiled. We note that this order parameter is sensitive to the different confined phases: C1 and C2. While C1 has ordering on both sublattices (and hence have maximum values of both $M_A$ and $M_B$), C2 has order on only one sublattice (so either $M_A$ or $M_B$ is maximum and the other is zero). An example of the C1 phase is shown for the $L=8$ in \Cref{fig:refCL8} (top) with both the height and the flux configurations, for which one has $M_A = -16$ and $M_B = 16$. Interestingly, these order parameters also signal deconfinement by \emph{both} going to zero simultaneously, when the lattice volume has many (extensive) winding strings. An example of a state with winding strings is shown in \Cref{fig:refCL8} (bottom), for which $M_A = M_B = 0$. \subsection{Methods: Exact diagonalization and cluster QMC} For our model in two spatial dimensions ($L_X,L_Y$) and spin-$\frac{1}{2}$ degrees of freedom per link, there are $2^{2 L_X L_Y}$ possible configurations, while working in the electric flux basis. Imposing the Gauss law constraint (which is easy in the flux basis) greatly reduces the number of allowed basis states in the Hilbert space. For example, a (4,4) lattice with only zero-charges has 2970 basis states, while the same lattice allowing for additional $\pm 2$-charges has 131072 basis states. A pragmatic way to proceed is to divide the Hilbert space into various winding number sectors ($W_x,W_y$). Since these sectors do not mix up among themselves, we can independently work in each individual sector, thus effectively reducing the Hilbert space dimension. In the presence of only $Q=0$ the individual winding numbers along $x$ or $y$ direction are all the same irrespective of where they are summed. Now, if we allow $Q = \pm 2$ along with $Q=0$, the individual winding numbers along any direction can, in principle, be different; so we should label these sectors as $(\{W_{x,1},W_{x,2},...,W_{x,L_Y}\}, \{W_{y,1},W_{y,2},...,W_{y,L_X}\})$. Interestingly, all the winding numbers along a direction, are either all even or all odd. Thus, there are four sectors: (even, even), (even, odd), (odd, even) and (odd, odd). For a (4,4) lattice, the winding numbers can take values 0, $\pm1$ and $\pm2$. So there are $3^4 \times 3^4$ sub-sectors within (even, even) sector, $3^4\times 2^4$ sub-sectors within (even, odd) or (odd, even) sector and $2^4\times 2^4$ sub-sectors within (odd, odd) sector. Thus, the Hilbert space in a given winding sector is further divided into many small sub-sectors which do not mix among themselves, facilitating ED significantly. For a (4,4) lattice, in the presence of only zero charges we need to diagonalize the largest sector (0,0) which has 990 basis states, whereas, in the addition of $\pm2$ charges we need to diagonalize the largest sector ($\{0,0,0,0\},\{0,0,0,0\}$) with 3464 basis states. Our QMC cluster algorithm can simulate different winding number sectors, but the height representation constrains the winding sectors that can be simulated at a time. For example, using periodic boundary conditions, only even winding sectors can be simulated, while for odd winding sectors one would need anti-periodic boundary conditions. While this is irrelevant for infinite volume in the spatial directions, it is important in the shorter temporal directions to have periodic boundary conditions on the height variables. Periodic boundary conditions on the flux variables translate to both \emph{periodic} and \emph{anti-periodic} boundary conditions on the height variables in time. However, the latter gives rise to an interface, which would become energetically unfavourable when one approaches the thermodynamic limit. Thus, it is completely natural to use the QMC on the height variables and restrict oneself to the even winding sectors. In \Cref{fig:Z2QLMC} we have plotted specific heat, $C_v=\beta^2(\braket{E^2} - \braket{E}^2)$ versus $\beta$, obtained using QMC and ED for both absence and presence of $Q = \pm 2$. Here $\braket{E} = -\frac{\partial \textrm{ln}Z}{\partial \beta}=-\frac{1}{L_T}\frac{\partial \rm{ln}Z}{\partial \epsilon}$ is average energy and $\braket{E^2} = \frac{1}{Z}\frac{\partial^2 Z}{\partial\beta^2}=\frac{1}{Z L_T^2}\frac{\partial^2 Z}{\partial\epsilon^2}$ is the average of energy squared, where the partition function is $Z=\sum_{\rm{config.}}\prod_\square W_\square(\epsilon J,\epsilon \lambda)$. $W_\square$ is the weight associated with a certain 6-height-variables interaction. $W_\square$ can take values $W_\square^1 = e^{-\epsilon J\lambda}\sinh(\epsilon J)$ for a flipped plaquette, $W_\square^2 = e^{-\epsilon J\lambda}\cosh(\epsilon J)$ for an unchanged flippable plaquette and $W_\square^3=1$ otherwise. After simplification, $\braket{E} = -\frac{1}{L_T}\braket{ \sum_\square\frac{\partial \rm{ln}W_\square }{\partial \epsilon} }$ and $\braket{ E^2}$ reads as $\frac{1}{L_T^2} \left( \braket{\sum_\square \frac{\partial^2 \rm{ln}W_\square}{\partial \epsilon^2}} + \braket{\sum_\square (\frac{\partial \rm{ln}W_\square}{\partial \epsilon})^2} \right)$. To compute $\braket{E}$ and $\braket{E^2}$ using QMC, we need to go through all 6-height-variables interactions and check the corresponding weights associated with them and add the appropriate factors in accordance with the simplified forms of $\braket{E}$ and $\braket{E^2}$. We note the nice agreement of results between QMC and ED in \Cref{fig:Z2QLMC}, which indicates the efficiency of the cluster algorithm we have used. \begin{figure} \centering \includegraphics[width=\hsize]{LetterFigs/cvu1z2.pdf} \caption{Comparison of ED and QMC.} \label{fig:Z2QLMC} \end{figure} \begin{figure} \includegraphics[width=\hsize]{LetterFigs/qsqravgLt24.pdf} \caption{The $\braket{Q^2}$ operator for different values of $\lambda$ for a range of $\beta$ on our finest lattice, $L_T = 24$. The vertical dotted line shows the location of critical temperature for the different $\lambda$ values.} \label{fig:Qsqr} \end{figure} \subsection{Variation of local charge density with temperature} While the $U(1)$ QLM has $\braket{Q^2}=0$ (normalized with the spatial volume) at all $T$ and $\lambda$ by definition, this is not the case for the LGT defined by the partition function in \Cref{eq:pf}. For example, it is easy to see that $\braket{Q^2}\rightarrow 1$ as $T \rightarrow \infty$ from \Cref{eq:pf}. On the other hand, the mass gap of $Q=\pm 2$ charges scale as $O(|\lambda|)$ from \Cref{eq:ham} for negative values of $\lambda$ which implies that $\braket{Q^2} \sim \exp(-a |\lambda|/T)$, where $a$ is an $O(1)$ number, as $T \rightarrow 0$. In Fig.~\ref{fig:Qsqr}, we show the variation of $\braket{Q^2}$ as a function of $\beta$ from the QMC data for the $U(1)$ QLM with $Q=\pm 2$ (\Cref{eq:pf}) for $\lambda=-1.0, -0.9,-0.8$ respectively using $L_T=24$. At low $T$ and also in the neighborhood of $T_c=1/\beta_c$, $\braket{Q^2}$ follows an activated behavior of $\exp(-a |\lambda|/T)$ for all the three values of $\lambda$. Interestingly, even though $\braket{Q^2} \approx 0.03$ in the vicinity of $\beta_c$ and thus the $Q=\pm 2$ charges can be considered to be dilute at the deconfinement transition, this thermal population is sufficient to completely change the critical behavior compared to the case where $\braket{Q^2}$ is strictly zero. We thus conjecture that these charged degrees of freedom ($Q=\pm 2$) generate a marginal operator and hence, weak universality in this LGT. \end{document}

d44c7dfb805656dad0ebf470347d26b641a28bb1

\section{Introduction}\label{sec::intro} Chemical abundance variations play a key role in our understanding of galactic formation and evolution. In particular, radial abundance gradients in stellar populations are believed to be the result of the inside-out galaxy formation scenario~\citep{1989MNRAS.239..885M,1997ApJ...477..765C,2012MNRAS.426..690B,2013ApJ...773...43B,2018MNRAS.474.3629G}. However, theoretical models also suggest that the radial abundance gradients are affected by various dynamical processes where stars can migrate far away from their birth radii~\citep[see, e.g.,][]{2002MNRAS.336..785S,2008ApJ...684L..79R,2009MNRAS.397.1599Q,2010ApJ...722..112M,2011A&A...527A.147M}. For instance, stellar radial migration caused by the spiral arms and/or spiral/bar resonances overlap results in flattening of the radial gradient~\citep{2009MNRAS.398..591S,2012A&A...540A..56P,2013A&A...554A..47G,2013A&A...558A...9M,2014A&A...572A..92M,2018MNRAS.481.1645M}. \begin{figure*} \begin{center} \includegraphics[width=0.49\hsize]{figs/xi2_GasMetallicity_0600.pdf}\includegraphics[width=0.49\hsize]{figs/fig3_GasMetallicity_0600.pdf} \caption{{\it Left:} example of the pattern speed measurement for a single spiral arm in model M1. The coloured lines correspond to the sum of the squared-root differences between the stellar density at $t_0=0.6$~Gyr and the ones from $t_0\pm1$~Myr rotated by a certain angle $\phi$~(see Eq.\ref{eq:eps}). The minimum value corresponds to the angle of rotation which being divided by $1$~Myr results in the pattern speed value~(black symbols) at a given \ensuremath{\rm R_{gal}}\,. {\it Right:} the pattern speed of the spiral arms in different models. Values for the individual spiral arms are shown by the errorbars of different colours. For a better representation, the pattern speed for each arm is shifted vertically by 10~km~s\ensuremath{^{-1}}~kpc\ensuremath{^{-1} }\, and compared to the circular frequency~(solid lines) which is also shifted by the same values.}\label{fig::Omega0}\label{fig::xi} \end{center} \end{figure*} \begin{figure*}[t] \begin{center} \includegraphics[width=1\hsize]{figs/fig1_GasMetallicity_0600.pdf} \caption{Structure of different models after $0.6$~Gyr: M1~(no initial gradient), M2~(both ISM enrichment and gradient) and M3~(only gradient). From left to right: stellar density, gas density, the mean ISM metallicity, residual ISM metallicity~(after subtraction of the radial gradient) and the radial metallicity profile. Different models are shown in different rows. Coloured lines highlight the location of the stellar spiral arms measured as a positive overdensity in the first row~(see Section.~\ref{sec::results} for details). In the right column, the mean radial metallicity trend is shown by the white solid line. Exponential fit of the radial gradient is shown by the red lines~(shifted vertically for better visibility) where $R_{25}=12$~kpc.}\label{fig::fig1} \end{center} \end{figure*} Some models also suggest that spiral arms can cause variations of the mean metallicity of stars in the azimuthal direction~\citep[see, e.g.,][]{2016MNRAS.460L..94G,2018A&A...611L...2K} where the metallicity variations are caused by re-shaping of pre-existing stellar populations with different chemo-kinematical properties in the vicinity of the spiral arms. Alternatively, \cite{2019A&A...628A..38S} developed a new 2D analytic model where stars in the Milky Way~(MW)-type disk inherit systematic~($\approx 0.1$~dex) abundance variations from the ISM, where the latter ones are the result of the enrichment near the arms with the most significant azimuthal variations appearing near the corotation radius. Since stars inherit the abundances from the ISM, analysis of the chemical abundance variations of gas plays a fundamental role in understanding the present-day stellar abundance patterns. Observations suggest that the ISM in disk galaxies is well mixed ~(scatter in azimuth of $\approx 0.05$~dex~\citep{Zinchenko2016}) and environmental variations of abundances depend on the observational techniques, disk coverage, and filling factor of individual \ensuremath{\rm HII}\, regions. For instance, analysis of the M101 galaxy has shown evidence for azimuthal variations in gas-phase metallicity~\citep{1996ApJ...456..504K} but later on \cite{2013ApJ...766...17L} found no difference between arm and inter-arm metallicities. Several more recent integral field unit~(IFU) observations are in favour of small but systematic variations of metallicity that appears to correlate with the location of the spiral arms~\citep[see, e.g.,][]{2012A&A...545A..43C,2016ApJ...830....4C,2017A&A...601A..61V, 2018MNRAS.474.1657S}. In particular, by studying the nearby spiral galaxy NGC~1365, \cite{2017ApJ...846...39H} found systematic~($0.2$~dex) azimuthal variations of the \ensuremath{\rm HII}\, region oxygen abundance near the spiral arms imprinted on a negative radial gradient. \cite{2016ApJ...830L..40S} showed that the \ensuremath{\rm HII}\, region oxygen abundances are higher at the trailing edges and lower at the leading edges of the spiral arms which is likely caused by radially outward~(inward) streaming motion at the trailing~(leading) edges of the spiral arms. By analyzing the Very Large Telescope/Multi Unit Spectroscopic Explorer~(VLT–MUSE) data for eight nearby galaxies, \cite{2019ApJ...887...80K} found a low $0.03-0.05$~dex azimuthal abundance scatter where half of the galaxies reveal azimuthal variations which in many cases, however, can not be clearly associated with the spirals. On the other hand, \cite{2016ApJ...827..103K} found that the inter-arm regions of NGC~628 have oxygen abundances similar to the arms which, however, could be the result of poor coverage of the galactic disk. A recent study of $45$ nearby spiral galaxies by \cite{2020MNRAS.492.4149S} suggests that $45-65\%$ of galaxies have more metal-rich \ensuremath{\rm HII}\, regions in spiral arms with respect to the inter-arm area. However, in some cases~($5-20\%$, depending on the calibrator), the opposite trend is seen, particularly more metal-poor \ensuremath{\rm HII}\, regions in the spiral arms compared to the inter-arm region. According to \cite{2020MNRAS.492.4149S} more metal rich spiral arms than the inter-arm area are observed in more massive galaxies with grand-design spiral arms. Finally, \cite{2021arXiv211010697W} have mapped the two-dimensional variations of metals across the disks of 19 nearby galaxies observed with the VLT–MUSE and found no evidence that spiral arms are enriched compared to the disk. Existing models of the ISM mixing in disk galaxies also have not reached a consensus about the abundance variations in azimuth where both turbulence and gravitational instability act towards homogeneity of the ISM~\citep{2002ApJ...581.1047D, 2003PhRvE..67d6311K, 2018MNRAS.475.2236K, 2012ApJ...758...48Y}, while the large-scale models predict variations driven by spirals~\citep{2016ApJ...830L..40S}\footnote{Note that \cite{2016MNRAS.460L..94G} and \cite{2013A&A...553A.102D} discuss the abundance variations in stellar populations but not in the ISM.}. Therefore, the origin of the azimuthal ISM metallicity variations remains unclear: whether they are real and, if so, are they driven by local self-enrichment~\citep{2017ApJ...846...39H} or by radial flows in the disk~\citep{2016ApJ...830L..40S}? In this work, using a set of high-resolution $N$-body hydrodynamical simulations we explore the origin of the ISM abundance variations across the spiral arms of the MW-type disk galaxies. We quantify the impact of the local ISM enrichment by ongoing star formation~(SF) and the transformation of the radial abundance gradient into the azimuthal one due to gas radial displacement~(migration) caused by the spiral arms. The paper is structured as follows. In Sec.~\ref{sec::models} we describe our models setup, subgrid physics and spiral arms kinematics analysis. In Sec.~\ref{sec::results} we discuss both small- and large-scale metallicity variations in the azimuthal direction and along the spirals linking the observed patterns to the properties of individual arms. In Sec.~\ref{sec::conclusions} we discuss and summarize our findings. \begin{figure} \begin{center} \includegraphics[width=1\hsize]{figs/fourier_GasMetallicity_0600.pdf} \caption{Radial structure of the main~($m=1,...,6$) 2D Fourier harmonics computed from the stellar surface density maps~(top) and from the mean metallicity distribution~(bottom).}\label{fig::fourier_amps} \end{center} \end{figure} \begin{figure} \begin{center} \includegraphics[width=1\hsize]{figs/delta_Z_GasMetallicity_0000.pdf} \caption{Cumulative distribution of the residual metallicity~(after subtraction of the radial metallicity gradient) in different models.}\label{fig::delta_z_cumulative} \end{center} \end{figure} \begin{figure*} \begin{center} \includegraphics[width=1\hsize]{figs/fig22_GasMetallicity_0600.pdf} \caption{The residual ISM metallicity maps where the regions with different minimal gas surface density are shown, from left to right: $\Sigma_{gas}>0.4$~M\ensuremath{_\odot}~pc\ensuremath{^{-2} }\,, $\Sigma_{gas}>1.4$~M\ensuremath{_\odot}~pc\ensuremath{^{-2} }\, and $\Sigma_{gas}>5$~M\ensuremath{_\odot}~pc\ensuremath{^{-2} }\,. The rightmost column shows the distribution of the residual metallicity in the corresponding disk regions. Once the regions with lower gas density are masked the remaining ares follow the stellar spiral arms much close, however, the residual metallicity distributions are weakly affected by these selections. The pairs of numbers in the rightmost column show the mean and the standard deviation of the residual metallicity distributions.}\label{fig::metallicity_den_threshold} \end{center} \end{figure*} \section{Models}\label{sec::models} \subsection{Simulations setup}\label{sec::appendix_model} We performed three $N$-body/hydrodynamical simulations of disk galaxies with a total stellar mass and a rotation curve compatible with those of the MW. All three models are the same in terms of the mass model but differ from each other by the initial radial profile of metallicity of the ISM and on/off enrichment of the ISM by newly formed stellar populations. Our models start from a pre-existing axisymmetric stellar disk with gas where we allow the star formation which is complemented by the chemical evolution of stellar populations. In all three models, $5\times 10^6$ stellar particles are initially redistributed following a Miyamoto-Nagai density profile~\citep{1975PASJ...27..533M} that has a characteristic scale length of $3$~kpc, vertical thicknesses of $0.2$~kpc and mass of $6 \times 10^{10}$~M\ensuremath{_\odot}. Our simulation also includes a live dark matter halo~($5\times 10^6$ particles) whose density distribution follows a Plummer sphere~\citep{1911MNRAS..71..460P}, with a total mass of $4\times 10^{11}$~M\ensuremath{_\odot} ~and a radius of $21$~kpc. Gas component is represented by an exponential disk with a scale length of $5$~kpc and the total mass of $2\times 10^{9}$~M\ensuremath{_\odot}. Gas dynamics is treated on a Cartesian grid with $5$~pc uniform spatial resolution. The initial equilibrium state has been generated using the iterative method from AGAMA software~\citep{2019MNRAS.482.1525V}. In our simulations subgrid physics implementation is the same as in \cite{2021MNRAS.501.5176K}. In particular, we include the formation of new star particles which inherit both kinematics and elemental abundances of their parent gas cells. At each time step, for newly formed stars we calculate the amount of gas returned, the mass of the various species of metals, the number of SNII or SNIa for a given initial mass and metallicity, the cumulative yield of various chemical elements, the total metallicity, and the total gas released. Feedback associated with the evolution of massive stars is implemented as an injection of thermal energy in a nearby gas cell proportional to the number of SNII, SNI and AGB stars. The hydrodynamical part also includes gas-metallicity depended radiative cooling~\citep[see][for details]{2021MNRAS.501.5176K}. Since our simulations aim to explore the formation of the azimuthal metallicity variations in three models we test the impact of the local ISM enrichment with and without pre-existing radial metallicity gradient in the gas. Models M1 and M2 include the ISM enrichment, while in model M3 we turn off the metals release to the ISM by newborn stellar populations. Models M2 and M3 start from the initial negative metallicity gradient~($\rm -0.15dex/r_e$, where $r_e=5.1$~kpc is the disk effective radius or $\rm -0.35dex/R_{25}$, where $R_{25}=12$~kpc) while model M1 has a constant initial metallicity of the gas. Therefore, Model M1 allows us to test how much the local enrichment alone is responsible for the azimuthal variations of the gas metallicity. Model M3 aims to quantify how the spiral arms induced redistribution of metals drives the azimuthal gradients, while Model M2 combines both effects. The simulations were evolved with the $N$-body+Total Variation Diminishing~(TVD) hydrodynamical code~\citep{2014JPhCS.510a2011K}. For the $N$-body system integration and gas self-gravity, we used our parallel version of the TREE-GRAPE code~\citep[][]{2005PASJ...57.1009F} with multithread usage under the SSE and AVX instructions. For the time integration, we used a leapfrog integrator with a fixed step size of $0.1$~Myr. In the simulation we adopted the standard opening angle $\theta = 0.7$. In recent years we already used and extensively tested our hardware-accelerator-based gravity calculation routine in several galaxy dynamics studies where we obtained accurate results with a good performance~\citep{2017MNRAS.468..920K,2018MNRAS.481.3534S,2020A&A...638A.144K}. \subsection{Pattern speed measurements}\label{sec::appendix1} In this section, we provide details about the pattern speed measurements for individual spiral arms. In order to measure the pattern speed of the spiral arms, we analyse the stellar surface density morphology in three snapshots, one is the referenced one at $t_0$ and the other two correspond to $\rm t_0 \pm 1~Myr$. Thanks to the small interval between the snapshots we can directly measure how much different radial segments of the individual spirals rotated in azimuthal direction over $1$~Myr. In practice we split each individual spiral arm into $\delta \ensuremath{\rm R_{gal}}\, = 0.1$~kpc segments and calculate the following parameter: \begin{equation} \displaystyle \rm \epsilon(\phi) = \sum_{x,y} \left(\Sigma_0(x,y) - \Sigma^*(\phi,x,y)\right)^2\,,\label{eq:eps} \end{equation} where $\Sigma_0(x,y)$ is the stellar surface density at $t_0$ and $\Sigma^*(\phi,x,y)$ is the stellar surface density at $\rm t_0-1~Myr$~(or $\rm t_0+1~Myr$) rotated by $\phi$ in the azimuthal direction. Since the amplitude of the spiral arms does not change much over $1$~Myr, as the result of the procedure for each segment of the spiral arms we have a curve $\epsilon(\phi)$ with a global minimum at the angle corresponding to the best similarity between the arm at $t_0$ and $t_0\pm1$~Myr~(see Fig.~\ref{fig::xi}~(left) for a single spiral arm in M1 model). In other words, we are trying to find the angle of rotation of a given spiral arms segment over $1$~Myr~(backward and forward in time). Therefore, the angular offset corresponding to the minima of the curves in Fig.~\ref{fig::xi}~(left) divided by $1$~Myr results in the pattern speed of the spiral arms segments, or the pattern speed of individual arms as the function of \ensuremath{\rm R_{gal}}\,~(see the right panel). Since we have an opportunity to measure the rotation of spirals backwards and forward in time we present the curves $\epsilon(\phi)$ for $\rm t_0 - (t_0+1Myr)$ and $\rm (t_0+1Myr) - t_0$ where the minima for each \ensuremath{\rm R_{gal}}\, are marked by circles and squares, respectively. In Fig.~\ref{fig::Omega0}~(right) we show the pattern speed of the individual arms~(crosses) compared to the rotational frequency~(circular velocity divided by \ensuremath{\rm R_{gal}}\,, solid lines) shown for each pattern speed by the same colour. For a better visibility, both the pattern speed and the corresponding rotational frequency are shifted by $10~\rm km~s\ensuremath{^{-1} }\,$ for different arms. As we can see, the spiral arms in our model, similar to a number of other $N$-body/hydrodynamical simulations~\citep[see, e.g.,][]{2011ApJ...735....1W,2012MNRAS.426..167G,2013MNRAS.432.2878R}, do not rotate with the same pattern speed along the radius. This behaviour makes our results qualitatively similar to the ones presented in \cite{2016ApJ...830L..40S}. \section{Results}\label{sec::results} \subsection{Spiral structure properties and small-scale metallicity behaviour} For each simulation, we focus our analyses on a single snapshot at $t_0=0.6$~Gyr of evolution, when a well-formed spiral structure is present. At later times the models are unstable to bar formation which substantially impacts both stars and the ISM, analyses of which is beyond the scope of the present study. In Fig.~\ref{fig::fig1} we show the face-on distributions of stellar~(first column) and gas density~(second column) together with the mean ISM metallicity~(third column), the residual ISM metallicity~(the fourth column, after the subtraction of the radial metallicity gradient (i.e. the mean metallicity at a given \ensuremath{\rm R_{gal}}\, shown by the white solid lines in the fifth column) and the radial metallicity profile~(fifth column). In the fifth column we also provide the exponential fits of the radial metallicity profiles~(red solid lines, shifted vertically). The slopes of the ISM metallicity we measure, while being mainly inherited from the initial setup, are comparable to the ones observed in the nearby disk galaxies~\citep{2014A&A...563A..49S, Belfiore2017,2019ApJ...887...80K,Zinchenko2019,Zinchenko2021,Zurita2021}. Although our initial setup~(mass model and initial equilibrium state) is the same in all the models, the morphology of the spiral structure is slightly different at the same snapshot in time. This is likely the result of stochasticity of the growth of the perturbations in stellar-gaseous disks~\citep[see, e.g.,][]{2011MNRAS.416.1191R} due to different ISM dynamics caused by metallicity-dependent gas cooling rates. Nevertheless, a Fourier analysis of the snapshots suggests that the modal composition of the spiral structure is very similar~(see the radial structure of Fourier harmonics in Fig.~\ref{fig::fourier_amps}~(top row)) and all three models reveal a multi-arm, tightly wound spiral structure similar to the one usually obtained in $N$-body simulations. It is seen that the spiral arms structure, revealed by the Fourier analysis, is dominated by the $m=2$ and $m=3$ modes, which superposition results in a slightly different morphology among our models~(see Fig.~\ref{fig::fig1}~(left)) at a given time. In order to quantify the ISM metallicity behaviour in the vicinity of the spirals we localize the individual arms which we define as the positive part of 2D stellar density perturbation~(overdensity): \begin{equation} \displaystyle \rm \delta \Sigma = \frac{\Sigma(R_{gal},\varphi) - \left\langle\Sigma(R_{gal},\varphi )\right\rangle_\varphi}{\left\langle\Sigma(R_{gal},\varphi)\right\rangle_\varphi}\,,\label{eq::eq1} \end{equation} where \ensuremath{\rm R_{gal}}\, and $\varphi$ are the radius and azimuth in cylindrical coordinates and the brackets $\left\langle\right\rangle_{\varphi}$ indicate the mean -- i.e. azimuthally averaged value at a given \ensuremath{\rm R_{gal}}\,. In Fig.~\ref{fig::fig1} the stellar spiral arms are highlighted by coloured lines spanning $2$~kpc radial \ensuremath{\rm R_{gal}}\, range. Although, the ISM morphology is quite complicated, because of a number of chains of giant clumps near the arms and isolated clouds in between the arms~\citep[see, e.g.,][]{2006MNRAS.371.1663D,2013MNRAS.436.1836R,2014MNRAS.439..936F,2016MNRAS.455.1782K,2017MNRAS.468..920K}, one can see that inside $\ensuremath{\rm R_{gal}}\,\lesssim10$~kpc the maxima of the large-scale gas density distribution correspond to the leading side of the stellar spiral arms, as it has been expected for slowly rotating spirals~\citep[see, e.g.,][]{2006MNRAS.367..873D,2011AstL...37..563K,2016MNRAS.458.3990P}. In order to test the impact of the local enrichment and the radial gradient transformation on the azimuthal variations of metallicity, one needs to be sure that both the strength and pattern speed of the spiral arms are the same in different models. We showed that the amplitudes of the stellar density perturbations are essentially the same in all the models~(see Fig.~\ref{fig::fourier_amps}, top). Also, the spiral arms for all three models rotate slower than the gas in the inner disk~($\lesssim8-10$~kpc) and corotate with the gas in the outer disk~(see Fig.~\ref{fig::Omega0}, the right panel). Therefore, the material stellar arms are non-steady; they are wound and, likely, stretched by the galactic shear in the outer disk. In the inner disk, spirals can bifurcate and merge with the others~\citep[see, e.g.,][]{2011MNRAS.410.1637S,2011ApJ...730..109F}. This makes the ISM structures and abundance patterns associated with the stellar arms also non-steady. \begin{figure}[t] \begin{center} \includegraphics[width=1\hsize]{figs/fig2_GasMetallicity_0600.pdf} \caption{Probability distribution functions for the residual ISM metallicity in arms and inter-arm regions~(top). Bottom row shows the corresponded PDFs for individual spiral arms where the colour of the lines is the same as in Fig.~\ref{fig::fig1}.}\label{fig::fig2} \end{center} \end{figure} \begin{figure}[t] \begin{center} \includegraphics[width=1\hsize]{figs/fig4_offset_GasMetallicity_0600.pdf} \caption{{\it Top:} background maps are the residual ISM metallicity in model M1~(no radial gradient, left) and M3~(no ISM enrichment, right) where the location of the stellar spiral arms are shown by solid lines while the mean location of the metallicity peaks -- by dashed lines. {\it Bottom:} distribution of the radial offset between the mean location of the stellar spiral arms and the metallicity peaks for different models.}\label{fig::fig4} \end{center} \end{figure} In Fig.~\ref{fig::fig1}~(see the third column) the mean metallicity distributions show a spiral-like morphology which however differs from the ones seen in gas and stars. To quantify the strength of the azimuthal metallicty variations, similar to the stellar overdensity~(see Eq.~\ref{eq::eq1}) we derive the residual metallicity: \begin{equation} \displaystyle \rm \delta Z/Z_\odot = \frac{ (Z/Z_\odot)(R_{gal},\varphi) - \left\langle(Z/Z_\odot)(R_{gal},\varphi)\right\rangle_\varphi }{ \left\langle(Z/Z_\odot)(R_{gal},\varphi)\right\rangle_\varphi}\,,\label{eq::eq2} \end{equation} where $\left\langle(Z/Z_\odot)(R_{gal},\varphi)\right\rangle_{\varphi}$ is the azimuthally-averaged radial metallicity profile~(see the white lines in Fig~\ref{fig::fig1}, fifth column). Maps in Fig.~\ref{fig::fig1}~(fourth column) demonstrate that the residual metallicity varies from negative to positive values across the arms. In particular, we find that at a given \ensuremath{\rm R_{gal}}\, the scatter is about $0.05$~dex which is similar to the numbers obtained in different observations~\citep[e.g.,][]{Zinchenko2016,2019ApJ...887...80K}. To demonstrate that in Fig.~\ref{fig::delta_z_cumulative} we show the cumulative distribution of the absolute values of the residual metallicity for different models. The figure shows that the metallicity deviates from the mean value at a given radius by $0.05$~dex for $40-65$\% of cases. \subsection{Large-scale metallicity behaviour across the spiral arms} In Fig.~\ref{fig::fig1} we notice that in M1 the residual metallicity does not reveal a large-scale spiral pattern, but instead consists of a number of patchy segments, while in models M2 and M3 the regions of systematically lower~(higher) metallicity span across the entire disk. This is highlighted more quantitatively in Fig.~\ref{fig::fourier_amps} where the radial profile of Fourier coefficients are presented. In these models one can see a dominant $m=2,3$ mode~(more prominent in negative $\delta Z$, fourth column in Fig.~\ref{fig::fig1}) despite the presence of six spiral arms. This result suggests that the large-scale metallicity pattern, in the presence of radial gradient, likely follows the dominant spiral mode~(see Figs.~\ref{fig::fig1},~\ref{fig::fourier_amps}). In observations, the information about the ISM metallicity distribution is often based on the data from the \ensuremath{\rm HII}\, regions~\citep[see, e.g.,][]{1996ApJ...456..504K,2012A&A...545A..43C,2017A&A...601A..61V,2019ApJ...887...80K}, where the gas density is high enough to sustain recent star-formation. This usually does not allow for a mapping of the metallicity distribution everywhere in the disk and limits the analysis to some sparsely distributed regions. In our simulations we are not able to resolve individual \ensuremath{\rm HII}\, regions, however, to link the observational results with our models we first limit the analysis of the ISM metallicity to the regions with high gas density. In Fig.~\ref{fig::metallicity_den_threshold} we show the distribution of the residual metallicity where we masked the regions with the gas surface density below given values~($0.4$, $1.4$ and $5$~M\ensuremath{_\odot}~pc\ensuremath{^{-2} }\,). Obviously, once we move to the metallicity distribution in regions with higher gas density the coverage of the disk decreases and the remaining regions trace better the spiral arms. However, there is no prominent systematics in the residual metallicity values in the remaining high-gas density regions. In particular, the second and third columns of Fig.~\ref{fig::metallicity_den_threshold} show that the residual metallicity varies in a wide range where the entire arms or their small patches can have systematically either positive or negative values of the residual metallicity. If, following some theoretical expectations~\citep{2016ApJ...830L..40S,2017ApJ...846...39H,2019A&A...628A..38S}, the high-metallicity ISM is associated with the spiral arms then the mean of the metallicity distribution should move towards positive values once we mask low-gas density regions. To test this in the rightmost column in Fig.~\ref{fig::metallicity_den_threshold} we show the distribution of the residual metallicity in the regions with high gas density. In other words, we plot the distribution of the values from maps shown in the first three columns of Fig.~\ref{fig::metallicity_den_threshold}. The resulting distributions however do not vary significantly, and both the mean and its dispersion are weakly impacted by the spatial selections based on the gas surface density. This suggests that similar to a number of observational studies, there is no apparent match of the large-scale metallicity patterns with the spiral arms in our models. However, since Fig.~\ref{fig::fig1} shows the metallicity distributions still demonstrating certain patterns, in the following, we analyse the complete metallicity distributions without masking any low gas-density regions. To analyze the variations of the metallicity more quantitatively, in Fig.~\ref{fig::fig2} we show the residual metallicity distributions for all spiral arms and inter-arm region~(top) and for individual spiral arms~(bottom). We see that the residual metallicities for both arm and inter-arms regions vary in roughly the same range, while spiral arm regions have slightly higher metallicities. The effect is the most prominent in Model M3~(without enrichment) suggesting that the release of metals by newly formed stars does not correlate much with the spiral arms but tends to smear the metallicity distribution at a given \ensuremath{\rm R_{gal}}\,~\citep[see, e.g.,][]{2012ApJ...758...48Y,2018MNRAS.475.2236K}. The distributions in Fig.~\ref{fig::fig2} suggest that systematically negative residuals in metallicity do occur in some of the spiral arms. This contradicts naive expectations that star formation would locally increase the ISM metallicity. To search for the latter effect, we find the location of nearby positive peaks in metallicity next to the arms and, thus, measure the offset to the spiral arms. In Fig.~\ref{fig::fig4}~(top) we show two examples of the offset calculation for the pair of arms in M1 and M3 models. One can see that there is a positive~(larger \ensuremath{\rm R_{gal}}\,) offset of the metallicity peak~(dashed lines) to the blue arm~(solid) in M1 while in model M3 we see the opposite configuration~(peak of metallicity has systematically smaller \ensuremath{\rm R_{gal}}\, relative to the stellar spiral arm). For the red arms in both models, the offset is, on average, zero. Of course, the offset we calculate may not represent a generic connection between the spiral arms and metallicity distribution, especially at local scales, however, it gives an idea of how, on average, the spatial distribution of metals correlates with the spiral arms depending on initial radial gradient and ISM enrichment. In the bottom panel of Fig.~\ref{fig::fig4}, we show the distribution of the offset for all the arms in three models. Models with initial radial gradient~(M2, M3) show on average a negative radial offset between spiral arm and metallicity peak, suggesting the presence of higher metallicity behind the trailing arms. In M1, the offset distribution is roughly symmetric. This behaviour is likely linked to the SF activity where newborn stars can release metals far away from their birthplaces in the spirals or even reach other arms, especially in the inner disk, where spirals rotate slower than the gas, thus breaking a coherent cycle of enrichment/mixing proposed in some previous works~\citep{2017ApJ...846...39H,2019A&A...628A..38S}. To test further the hypothesis about the impact of the local enrichment, in Fig.~\ref{fig::sfr_z} we show the relation between the SF surface density and the metallicity residuals. The relation is based on the $\rm 500\times500~pc^2$ smoothing of both SFR and $\delta Z$ metallicity XY-maps. Although we do not see a strong correlation for any of our models, in M1, there is a clear trend where more metal-rich regions tend to be spatially associated with more intense SF. Finally, we test how much the ISM dynamics affects the appearance of the large-scale abundance variations in the azimuthal direction. In Fig.~\ref{fig::vrvphi_z} we compare the relation between both radial~($V_R$) and residual azimuthal velocity~($\delta V_\varphi$) components of the gas with the residual metallicity. The residual azimuthal velocity component is the gas velocity component after the subtraction of the mean azimuthally-averaged rotation~(rotation curve, see Eq.~\ref{eq::eq1} where $\Sigma$ should be replaced by $V_\varphi$). We see no correlation in the case of M1~(only enrichment) and a very clear correlation for M3~(no enrichment, but initial radial gradient) while model M2 shows an intermediate behaviour. Similar to the $\rm SFR-\delta Z$ analysis in Fig.~\ref{fig::sfr_z}, we calculated the Pearson correlation coefficients which allow us to quantify the relations between the gas velocity components and the residual ISM metallicity~(see numbers in the panels of Fig.~\ref{fig::vrvphi_z}). This result suggests that, even slowly~(compared to the gas) rotating spiral arms drag a more metal-rich gas from the inner parts of the disk to the trailing side of the arms and, thus, increasing the mean metallicity of the ISM behind the spirals. This picture explains why in models M2 and M3 we find a prominent negative offset of the metallicity peaks relative to the spiral arms~(see Fig.~\ref{fig::fig4}). \begin{figure}[t] \begin{center} \includegraphics[width=1\hsize]{figs/fig7_sfr_GasMetallicity_0600.pdf} \caption{Relations between the residual metallicity and star formation surface density. Coloured lines represent the density contours. The corresponding Pearson correlation coefficients are given in each panel.}\label{fig::sfr_z} \end{center} \end{figure} \begin{figure}[t] \begin{center} \includegraphics[width=1\hsize]{figs/fig5_vrvphi_GasMetallicity_0600.pdf} \caption{Relations between the residual metallicity and gas radial velocity component~(top) and residual azimuthal velocity component~(subtracted circular velocity, bottom). The colours represents the mass fraction of the gas.}\label{fig::vrvphi_z} \end{center} \end{figure} \vspace{-\topsep}\section{Summary}\label{sec::conclusions} Using hydrodynamical simulations of isolated spiral galaxies, we study the impact of the local enrichment and pre-existing radial metallicity gradient transformation on the formation of azimuthal metallicity variations in the vicinity of the spiral arms. Similar to some previous studies, in our models, the pattern speed of individual arms varies with the radius being slow~(compared to the disk rotation) in the inner~($<8$~kpc) and corotating in the outer disk. Our main results are as follows. \begin{itemize} \item[-] Both types of models with local enrichment and pre-existing radial abundance gradient are able to produce the azimuthal scatter of the ISM metallicity~($\approx0.05$~dex) in simulated spiral galaxies~(see Fig.~\ref{fig::fig1}). Analysis of the ISM metallicity as a function of the underlying gas density does not reveal any strong relations between the residual metallicity and the location of the stellar spiral arms~(see Fig.~\ref{fig::metallicity_den_threshold}). Although individual arms could have systematically lower or higher~(than the mean) metallicity, gas in the spiral is slightly more metal-rich compared to the inter-arms region~(see Fig.~\ref{fig::fig2}). \item[-] We find that model with local enrichment only~(M1, no pre-existing metallicity gradient) is not able to reproduce large-scale spiral arms-like variations of the mean metallicity~(see Figs.~\ref{fig::fig1},~\ref{fig::fourier_amps}). In this model, the short-scale patches of high-metallicity regions can be found suggesting that the enrichment of the ISM does not correlate much with the large-scale spiral structure. \item[-] Both models with pre-existing radial metallicity gradient~(M2 and M3) show the formation of the spiral-like metallicity pattern but its morphology differs from the stellar spiral structure~(see Figs.~\ref{fig::fig1}). Although we identify a six-armed spiral structure, the large-scale residual metallicity pattern depicts $m=2,3$ structure corresponding to the most significant 2D Fourier harmonics of the stellar density distribution~(see Fig.~\ref{fig::fourier_amps}). \item[-] We found a substantial radial offset between spiral arms and the metal-rich ISM pattern. The amplitude of the offset reaches up to $1-1.5$~kpc~(see Fig~\ref{fig::fig4}). In models with the radial gradient we find no correlation of the residual metallicity with recent star-formation while in model M1~(without) gradient we can see a weak increase of the metallicity with higher star formation. Model without ISM enrichment shows the most prominent correlation of the residual metallicity with the gas velocity components: larger metallicity corresponds to larger negative~(inflow) radial velocities and negative azimuthal velocity residuals. Therefore, we suggest that dynamical effects play a key role in the formation of the large-scale metallicity variations across spiral arms. \end{itemize} Our models, while being rather simplified, nevertheless allow us to propose an explanation of the observational data demonstrating a controversial picture of the systematic azimuthal variations of metallicity around spiral arms. We suggest that the ISM enrichment near the arms alone is unlikely responsible for the systematic azimuthal metallicity pattern, at least in the case of non-steady spirals, while the key ingredient is a pre-existing radial abundance gradient. If the observed radial gradients are small~\citep{2014A&A...563A..49S}, they are likely not enough to be transformed into the azimuthal variations in most of the galaxies. However, once the azimuthal variations are found, more likely they will depict the shape of the most significant spiral mode~($m=2, 3$), which can explain prominent oxygen variations found in some barred galaxies~\citep{2016ApJ...830L..40S,2017ApJ...846...39H}, however, in some cases flattening of the radial gradient should act against the formation of the systematic azimuthal variations. Extending the model presented in \cite{2016ApJ...830L..40S} we suggest that not only corotating spirals are responsible for the azimuthal variations of the ISM metallicity but also slowly~(compared to the gas) rotating patterns provide similar results. Obviously, our predictions will depend on the strength and nature of the spirals and also star formation activity, which, along with the Fourier analysis of the 2D metallicity distribution, may be tested by IFU surveys in the near future. \begin{acknowledgements} SK deeply appreciates Igor Zinchenko (LMU, Munich; MAO, Kyiv) for his contribution to the early versions of the paper. The authors thank the anonymous referee for a constructive report. SK also thanks Daisuke Kawata for useful discussion about the nature of corotating spirals. Numerical simulations were carried by using the equipment of the shared research facilities of HPC computing resources at Lomonosov Moscow State University~(project RFMEFI62117X0011) supported by the Russian Science Foundation~(project no. 19-72-20089). \end{acknowledgements} \bibliographystyle{aa}

031a023e25aaf5e0bb0e2908752852471863eb47

\section{Introduction} In dimensions 2 and 3, the Euler equations \begin{eqnarray} u\cdot\nabla u+\nabla p=0, \label{1.1}\\ \nabla\cdot u=0 \label{1.2} \end{eqnarray} describe steady flows of ideal incompressible fluid. The equations are also of some mathematical interest in an arbitrary dimension. Here $u=\big(u_1(x),\dots,u_n(x)\big)$ is a vector field on an an open set $U\subset{\mathbb R}^n$ (the fluid velocity) and $p$ is a scalar function on $U$ (the pressure). We consider only smooth real solutions to the Euler equations, i.e., $u_i\in C^\infty(U)\ (i=1,\dots,n)$ and $p\in C^\infty(U)$ are assumed to be real functions (the term ``smooth'' is used as the synonym of ``$C^\infty$''). We say that a solution $(u,p)$ to \eqref{1.1}--\eqref{1.2} is a {\it Gavrilov flow}, if it satisfies \begin{equation} u\cdot\nabla p=0, \label{1.3} \end{equation} i.e., the velocity is orthogonal to the pressure gradient at all points of $U$. Equations \eqref{1.1}--\eqref{1.3} constitute an overdetermined system of first order differential equations: $n+2$ equations in $n+1$ unknown function. Therefore every Gavrilov flow is an exception in a certain sense. Nevertheless, such flows exist and deserve study. Such flows satisfy the following important property: a pair of functions $(\tilde u,\tilde p)$ given~by \begin{equation} \tilde u=\varphi(p)u,\quad \nabla\tilde p=\varphi^2(p)\nabla p, \label{1.4} \end{equation} where $\varphi(p)$ is an arbitrary smooth function, is again a Gavrilov flow. This property underlies the following construction that will be called the {\it Gavrilov localization}. Let $p_0\in\mathbb R$ be a regular value of the function $p$ such that $M_{p_0}=\{x\in U\mid p(x)=p_0\}$ is a compact hypersurface in $U$. Then we can construct a compactly supported smooth solution to the Euler equations on the whole of ${\mathbb R}^n$ by choosing $\varphi(p)$ as a cutoff function with support in a small neighborhood of $p_0$. Indeed, the new velocity vector field $\tilde u$ and the gradient $\nabla\tilde p$ are supported in some compact neighborhood $\tilde U\subset U$ of the surface $M_{p_0}$, as seen from \eqref{1.4}, and we can define $\tilde u$ as zero in ${\mathbb R}^n\setminus U$. Thus, the new pressure $\tilde p$ is constant on every connected component of $U\setminus \tilde U$. Since only the gradient $\nabla p$ participates in \eqref{1.1}--\eqref{1.3}, we can assume without lost of generality that $\tilde p=0$ on the ``exterior component'' of $U\setminus \tilde U$. It is now clear that $\tilde p$ can be extended to a compactly supported function $\tilde p\in C^\infty({\mathbb R}^n)$. For some neighborhood $O({\mathcal C})$ of the circle ${\mathcal C}=\{(x_1,x_2,0)\in{\mathbb R}^3\mid x_1^2+x_2^2=R^2\}$, A.V.~Gavrilov \cite{G} proved the existence of a solution $u\in C^\infty\big(O({\mathcal C})\setminus{\mathcal C};{\mathbb R}^3\big),\ p\in C^\infty\big(O({\mathcal C})\setminus{\mathcal C}\big)$ of the Euler equations satisfying \eqref{1.3}, and such that for some regular value $p_0$ of the function $p$, the surface $M_{p_0}\subset O({\mathcal C})\setminus{\mathcal C}$ is diffeomorphic to the torus. Gavrilov's formulas involve an arbitrary positive constant $R$, without lost of generality we set $R=1$. Using the localization procedure described above, Gavrilov proved the existence of a solution $\tilde u\in C^\infty({\mathbb R}^3;{\mathbb R}^3),\ \tilde p\in C^\infty({\mathbb R}^3)$ of the Euler equations supported in a small neighborhood of $M_{p_0}$. Thus, Gavrilov gave a positive answer to the old question: Is there a smooth compactly supported solution to the Euler equations on ${\mathbb R}^3$ that is not identically equal to zero? Unfortunately, \cite{G} is written in terse language and many details are omitted. Actually the same idea is implemented in a subsequent article \cite{CLV} by Constantin -- La -- Vicol. The latter paper starts with the so called Grad -- Shafranov ansatz that has appeared in plasma physics. Unlike \cite{G}, the article \cite{CLV} involves a thorough analysis of nonlinear ODEs that arise while constructing a solution. We emphasize that in both papers \cite{CLV,G} the existence of a Gavrilov {\it axisymmetric} smooth compactly supported flow on ${\mathbb R}^3$ is proved. Indeed, \cite{G} starts with the Euler equations in cylindrical coordinates for axisymmetric solutions. To authors' knowledge, the Grad -- Shafranov ansatz is adapted to the study of axisymmetric solutions only. We~say that two Gavrilov flows $(u,p)$ and $(\tilde u,\tilde p)$, defined on the same open set $U\subset{\mathbb R}^n$, are {\it equivalent} if \eqref{1.4} holds with a smooth non-vanishing function $\varphi(p)$. For example, $(u,p)$ and $(-u,p)$ are equivalent Gavrilov flows. In the present article, Gavrilov flows are considered up to the equivalence. We mostly study the structure of such a flow in a neighborhood of the hypersurface $M_{p_0}\subset U$ for a regular value $p_0$ of the pressure. \vskip2mm The article is organized as follows. In Section~\ref{sec:geom}, we describe a Gavrilov flow in terms of first and second quadratic forms of isobaric hypersurfaces $M_p$. Having received a geometric description, one can forget about the Euler equations. In Section~\ref{sec:ex} we discuss an example of a Gavrilov flow constructed in any even dimension. The example can be modified using the Gavrilov localization. The following observation is widely used in \cite{NV}. If a solution $(u,p)$ to the Euler equations is defined on ${\mathbb R}^3$ and sufficiently fast decays at infinity, then the quadratic form $(\nu\cdot u)(\xi\cdot u)$ integrates to zero over every affine plane $P\subset{\mathbb R}^3$, where $\nu$ is the normal vector to $P$ and $\xi$ is an arbitrary vector parallel to $P$. In Section~\ref{sec:section}, we show that the double integral over $P$ (for a Gavrilov flow) can be reduced to the integral over the curve $P\cap M_p$. This property is interesting in itself, but applications of this property are still unknown. In~Section~\ref{sec:eqs}, we obtain a system of PDEs for a Gavrilov flow in ${\mathbb R}^3$. It is an overdetermined system: 4 equations in 3 unknown functions. The problem of deriving consistency conditions for the system is the main one in the study of these flows. The problem remains open in the general case. Sections~\ref{sec:axi}--\ref{sec:2ex} are devoted to axisymmetric Gavrilov flows in ${\mathbb R}^3$. Unlike \cite{CLV,G}, our analy\-sis of these flows is based on the well-known geometric fact: the equation for geodesics admits a first integral for surfaces of revolution, the Clairaut integral. In~Section~\ref{sec:axi}, using geometric arguments, we reduce the system from the Section~\ref{sec:eqs} to a much simpler system of partial differential equations for axisymmetric Gavrilov flows: two equations in one unknown function of two variables, including also two functions of one variable $p$. The~consistency conditions for the last system of equations are derived in Section~\ref{sec:cond}. The existence of an axisymmetric Gavrilov flow in the open set $O({\mathcal C})\setminus{\mathcal C}$ is proved in \cite{G} such that the pressure function $p(r,z)$ is smooth in a neighborhood of the point $(r,z)=(1,0)$ that is a nondegenerate minimum point of $p$. We~study such flows in Section~\ref{sec:structure}. The~corresponding system of PDEs can be solved in series. We emphasize that \cite{CLV,G} only prove the existence of axisymmetric Gavrilov flows but do not give numerical examples. In our opinion, numerical and geometric examples are of a great importance since they can lead to new hypotheses. In Section~\ref{sec:2ex}, we give two geometric illustrations of the axisymmetric Gavrilov flow: one with isobaric surfaces diffeomorphic to a torus (considered in Section~\ref{sec:structure}) and another such flow with a pressure function $p(r,z)$ periodic in~$z$. Some open questions on Gavrilov flows are posed in Section~\ref{sec:problems}. \section{Geometry of a Gavrilov flow} \label{sec:geom} Let $(u,p)$ be a Gavrilov flow on an open set $U\subset{\mathbb R}^n$. Integral curves of the vector field $u$ are also called \textit{particle trajectories}. \begin{proposition}\label{P2.1} The pressure $p$ is constant on every integral curve of the vector field~$u$. \end{proposition} \begin{proof} Let $\gamma:(a,b)\rightarrow U$ be an integral curve of $u$, i.e., $\dot\gamma(t)=u(\gamma(t))$. Then $ \frac{d(p(\gamma(t))}{dt}=\nabla p(\gamma(t))\cdot\dot\gamma(t)=\nabla p(\gamma(t))\cdot u(\gamma(t))=0. $ The latter equality holds in virtue of \eqref{1.3}. \end{proof} By Proposition~\ref{P2.1}, the Bernoulli law, that is $\frac{\vert u\vert^2}{2}+p={\rm{const}}$ along a particle trajectory, for a Gavrilov flow splits into two conservation laws: \begin{equation} p=c={\rm{const}},\quad \frac{\vert u\vert^2}{2}=C={\rm{const}}\quad\mbox{along a particle trajectory}. \label{2.1} \end{equation} We say that $x\in U$ is a {\it regular point} if $\nabla p(x)\neq0$. The vector field $u$ does not vanish at regular points as is seen from \eqref{1.1}. The sets \[ M_{p_0}=\{x\in U\mid p(x)=p_0={\rm {const}}\} \] will be called {\it isobaric hypersurfaces}. A particle trajectory starting at a point of an isobaric hypersurface $M_p$ does not leave $M_p$ ``for ever''. In the general case, an arbitrary closed subset of $U$ can be an isobaric hypersurface $M_p$. But $M_p$ is indeed a smooth hypersurface of ${\mathbb R}^n$ in a neighborhood of a regular point $x\in M_p$. We say that $M_p$ is a {\it regular isobaric hypersurface} if it consists of regular points. Recall that a vector field $u$ on a manifold $M$ with a Riemannian metric $g$ is called a {\it geodesic vector field} if $\nabla_{\!u}\,u=0$, where $\nabla$ stands for the covariant derivative with respect to the Levi-Civita connection of $(M,g)$. Integral curves of a geodesic vector field are geodesics. In the case of a Gavrilov flow, regular isobaric hypersurfaces $M_p\subset{\mathbb R}^n$ are considered with the Riemannian metric induced by the Euclidean metric of ${\mathbb R}^n$. \begin{proposition} Given a Gavrilov flow $(u,p)$, the restriction of $u$ to every regular isobaric hypersurface $M_p$ is a non-vanishing geodesic vector field. \end{proposition} \begin{proof} It consists of one line: \begin{equation} \nabla_{\!u}\,u=P(u\cdot\nabla u)=-P(\nabla p)=0. \label{2.2a} \end{equation} Nevertheless, the line needs the following comment. On the left-hand side of \eqref{2.2a}, $\nabla$ stands for the covariant derivative corresponding to the induced Riemannian metric on $M_p$. But two other $\nabla$'s stand for the Euclidean gradient, the same operator as in the Euler equations \eqref{1.1}--\eqref{1.2}; and $P$ is the orthogonal projection (of vectors in ${\mathbb R}^n$) onto the tangent hyperplane of $M_p$. The first equality in \eqref{2.2a} is the main relationship between intrinsic geometry of a hypersurface and geometry of the ambient space; it goes back to Gauss and is valid in a more general setting, see \cite[Chapter~VII, Proposition~3.1]{KN}. The second equality in \eqref{2.2a} holds by \eqref{1.1}, and the last equality holds by \eqref{1.3}. \end{proof} Since the vector field $u$ does not vanish at regular points, no integral curve of $u$ living on a regular hypersurface $M_p$ degenerates to a point. Thus, integral curves of $u$ constitute a geodesic foliation of a regular part of any isobaric hypersurface. \begin{proposition} Let a Gavrilov flow $(u,p)$ be defined on an open set $U\subset{\mathbb R}^n$. Let us restrict the vector field $u$ onto a regular isobaric hypersurface $M_p$, and let ${\rm{div}}\,u$ be the $(n-1)$-dimensional divergence of the restriction which is understood in the sense of intrinsic geometry of the hypersurface $M_p$. Then \begin{equation} {\rm {div}}\,u = u(\log \vert \nabla p\,\vert). \label{2.3} \end{equation} On the right-hand side of \eqref{2.3}, the vector field $u$ is considered as a differentiation of the algebra $C^\infty(M_p)$ of smooth functions on $M_p$. \end{proposition} \begin{proof} We will show that the equation \eqref{2.3} is equivalent to the incompressibility equation \eqref{1.2}. To this end we will rewrite the equation \eqref{1.2} in local curvilinear coordinates adapted to the foliation of $U$ into isobaric hypersurfaces. Fix a regular point $x_0\in U$ and set $p_0=p(x_0)$. For $p\in\mathbb R$ sufficiently close to $p_0$, the isobar $M_p$ is a regular hypersurface near $x_0$. Choose local curvilinear coordinates $(z^1,\dots,z^{n-1})$ on the hypersurface $M_{p_0}$. Let $$ {\mathbb R}^{n-1}\supset\Omega\ {\stackrel r\longrightarrow}\ {\mathbb R}^n,\quad r=r(z^1\dots,z^{n-1}) $$ be the parametrization of $M_{p_0}$ in these coordinates. Assume that $0\in\Omega$ and $r(0)=x_0$. In some neighborhood of $x_0$, we introduce local curvilinear coordinates $(z^1,\dots,z^n)$ in ${\mathbb R}^n$ as follows. Define the vector field \begin{equation} \xi=\frac{\nabla p}{\vert\nabla p\,\vert^2}. \label{2.4} \end{equation} For $(z^1,\dots,z^{n-1})\in{\mathbb R}^{n-1}$ sufficiently close to $0$, let $$ R(z^1,\dots,z^{n-1};z^n),\quad(p_0-\varepsilon< z^n<p_0+\varepsilon) $$ be the integral curve of the vector field $\xi$ starting at the point $r(z^1,\dots,z^{n-1})$ at the initial moment $z^n=p_0$. Thus $R(z)\in{\mathbb R}^n$ is the solution to the Cauchy problem \begin{equation} \frac{\partial R}{\partial z^n}(z)=\xi\big(R(z)\big),\quad R(z^1,\dots,z^{n-1};p_0)=r(z^1,\dots,z^{n-1}). \label{2.5} \end{equation} Obviously, $R$ is a diffeomorphism between some neighborhood of the point $(0,\dots,0,p_0)$ and a neighborhood of $x_0$; therefore $(z^1,\dots,z^n)$ constitute a local coordinate system in ${\mathbb R}^n$ near the point $x_0$. By our construction, $R$ satisfies the identity $p\big(R(z)\big)=z^n$, which means that the coordinate $z^n$ coincides with the pressure $p$. Nevertheless, we use the different notation for the coordinate since $z^n$ is considered as an independent variable while $p$ is a function on $U$. For every $z^n$ sufficiently close to $p_0$, $(z^1,\dots,z^{n-1})$ are local coordinates on the isobaric hypersurface $ M_{z^n}=\{x\in U\mid p(x)=z^n\}. $ Let $$ ds_{z^n}^2=g_{\alpha\beta}dz^\alpha dz^\beta,\quad g_{\alpha\beta}=\frac{\partial R}{\partial z^\alpha}\cdot\frac{\partial R}{\partial z^\alpha} $$ be the first quadratic form of $M_{z^n}$. We use the following convention: Greek indices vary from 1 to $n-1$ and the summation from 1 to $n-1$ is assumed over a repeating Greek index; while Roman indices vary from 1 to $n$ with the corresponding summation rule. We also write the Euclidean metric of ${\mathbb R}^n$ in coordinates $(z^1,\dots,z^n)$~as $$ ds^2=h_{ij}dz^idz^j,\quad h_{ij}=\frac{\partial R}{\partial z^i}\cdot\frac{\partial R}{\partial z^j}. $$ Obviously, $ h_{\alpha\beta}=g_{\alpha\beta}, $ and $$ h_{\alpha n}(z)=\frac{\partial R}{\partial z^\alpha}(z)\cdot\frac{\partial R}{\partial z^n}(z) =\frac{\partial R}{\partial z^\alpha}(z)\cdot\frac{\nabla p}{\vert\nabla p\,\vert^2}\big(R(z)\big)=0. $$ The last equality holds since the vector $\frac{\partial R}{\partial z^\alpha}(z)$ is tangent to the hypersurface $M_{z^n}$ at the point $R(z)$ while the vector $\nabla p(R(z))$ is orthogonal to $M_{z^n}$ at the same point. Similarly, we get $ h_{nn}(z)=\big\vert\frac{\partial R}{\partial z^n}(z)\big\vert^2 =\big\vert\nabla p\big(R(z)\big)\big\vert^{-2}. $ Thus, $$ (h_{ij}) =\Big(\begin{array}{cc}g_{\alpha\beta}&0\\0&\vert\nabla p\,\vert^{-2}\end{array}\Big),\quad (h^{ij})=(h_{ij})^{-1} =\Big(\begin{array}{cc}g^{\alpha\beta}&0\\0&\vert\nabla p\,\vert^2\end{array}\Big). $$ Let $ \Gamma^\alpha_{\beta\gamma}=\frac{1}{2}g^{\alpha\delta}\big(\frac{\partial g_{\beta\delta}}{\partial z^\gamma} +\frac{\partial g_{\gamma\delta}}{\partial z^\beta}-\frac{\partial g_{\beta\gamma}}{\partial z^\delta}\big) $ be the Christoffel symbols of $M_{z^n}$ in coordinates $(z^1,\dots,z^{n-1})$ and $G^i_{jk}$ be the Christoffel symbols of the Euclidean metric in coordinates $(z^1,\dots,z^n)$. As follows from the above relations, the Christoffel symbols satisfy \begin{equation} \begin{aligned} G^\alpha_{\beta\gamma}&=\Gamma^\alpha_{\beta\gamma},\quad G^n_{\beta\gamma}=-\frac{1}{2}\vert\nabla p\,\vert^2\,\frac{\partial g_{\beta\gamma}}{\partial z^n},\quad G^\alpha_{\beta n}=\frac{1}{2}g^{\alpha\delta}\,\frac{\partial g_{\beta\delta}}{\partial z^n},\\ G^\alpha_{nn}&=-\frac{1}{2}g^{\alpha\delta}\,\frac{\partial\vert\nabla p\,\vert^{-2}}{\partial z^\delta},\ \ G^n_{\beta n}=\frac{1}{2}\vert\nabla p\,\vert^2\,\frac{\partial\vert\nabla p\,\vert^{-2}}{\partial z^\beta},\ \ G^n_{nn}=\frac{1}{2}\vert\nabla p\,\vert^2\,\frac{\partial\vert\nabla p\,\vert^{-2}}{\partial z^n}. \end{aligned} \label{2.7} \end{equation} The velocity $u$ can be represented in the chosen coordinates as $u(z)=u^\alpha(z)\frac{\partial R(z)}{\partial z^\alpha}$, since it is tangent to $M_{z^n}$. The velocity $u$ can be thought as a smooth vector field either on the $n$-dimensional open set $U$ or on each isobaric hypersurface $M_{z^n}$ smoothly depending on the parameter $z^n$. We remain the notation $\nabla\cdot u$ for the $n$-dimensional divergence of $u$ while the $(n-1)$-dimensional divergence of $u$ on $M_{z^n}$ will be denoted by ${\rm {div}}\,u$. Thus, \begin{equation} {\rm {div}}\,u=\nabla_{\alpha}u^{\alpha}=\frac{\partial u^\alpha}{\partial z^\alpha}+\Gamma^\alpha_{\alpha\beta}u^\beta. \label{2.9} \end{equation} By the same formula, $ \nabla\cdot u=\frac{\partial u^i}{\partial z^i}+G^i_{ij}u^j, $ since $u^n=0$, this becomes $$ \nabla\cdot u=\frac{\partial u^\alpha}{\partial z^\alpha}+G^i_{i\beta}u^\beta. $$ In particular, the incompressibility equation \eqref{1.2} is written in the chosen coordinates as $ \frac{\partial u^\alpha}{\partial z^\alpha}+G^i_{i\beta}u^\beta=0. $ Using this equation, formula \eqref{2.9} becomes $ {\rm {div}}\,u=-(G^i_{i\beta}-\Gamma^\alpha_{\alpha\beta})u^\beta $. By the formulas for Christoffel symbols, $ G^i_{i\beta}-\Gamma^\alpha_{\alpha\beta}=-\frac{\partial(\log\vert\nabla p\,\vert)}{\partial z^\beta}. $ Inserting this expression into the previous formula with ${\rm {div}}\,u$, we obtain $$ {\rm {div}}\,u=\frac{\partial(\log\vert\nabla p\,\vert)}{\partial z^\beta}\,u^\beta. $$ This is equivalent to \eqref{2.3}. \end{proof} Let ${\rm{II}}$ be the second quadratic form of a regular isobaric hypersurface $M_p$. Recall that ${\rm{II}}$ depends on the choice of the unit normal vector $N$ to a hypersurface (changes its sign if $N$ is replaced with $-N$). We choose $N$ for $M_p$ to be a positive multiple of~$\nabla p$. \begin{proposition} Given a Gavrilov flow $(u,p)$, the restriction of the vector field $u$ onto a regular isobaric hypersurface $M_p$ satisfies \begin{equation} {\rm{II}}(u,u)=-\vert\nabla p\,\vert. \label{2.10} \end{equation} \end{proposition} \begin{proof} We write down \eqref{1.1} in local coordinates $(z^1,\dots,z^n)$ (the same as in the previous proof) as $ (u\cdot\nabla u)^i+(\nabla p)^i=0. $ Setting $i=\alpha$ here gives the same result: integral curves of $u$ are geodesics of $M_p$. Thus, we set $i=n$, and by \eqref{2.4}--\eqref{2.5}, get $(\nabla p)^n=\vert\nabla p\,\vert^2$ and \begin{equation} (u\cdot\nabla u)^n=-\vert\nabla p\,\vert^2. \label{2.11} \end{equation} By a well-known formula for covariant derivatives, $ (u\cdot\nabla u)^n=u^i\big(\frac{\partial u^n}{\partial z^j}+G^n_{ij}u^j\big). $ Since $u^n=0$, this becomes $ (u\cdot\nabla u)^n=G^n_{\alpha\beta}u^\alpha u^\beta. $ Using $G^n_{\alpha\beta}=-\frac{1}{2}\vert\nabla p\,\vert^2\,\frac{\partial g_{\alpha\beta}}{\partial z^n}$ from \eqref{2.7}, we get \begin{equation} (u\cdot\nabla u)^n=-\frac{1}{2}\vert\nabla p\,\vert^2\,\frac{\partial g_{\alpha\beta}}{\partial z^n}u^\alpha u^\beta. \label{2.12} \end{equation} Differentiating the equality $ g_{\alpha\beta}=\frac{\partial R}{\partial z^\alpha}\cdot\frac{\partial R}{\partial z^\beta} $ with respect to $z^n$, we obtain $ \frac{\partial g_{\alpha\beta}}{\partial z^n} =\frac{\partial^2 R}{\partial z^\alpha\partial z^n}\cdot\frac{\partial R}{\partial z^\beta} +\frac{\partial^2 R}{\partial z^\beta\partial z^n}\cdot\frac{\partial R}{\partial z^\alpha} $. This can be written as $$ \frac{\partial g_{\alpha\beta}}{\partial z^n} =\frac{\partial}{\partial z^\alpha}\Big(\frac{\partial R}{\partial z^n}\cdot\frac{\partial R}{\partial z^\beta}\Big) +\frac{\partial}{\partial z^\beta}\Big(\frac{\partial R}{\partial z^n}\cdot\frac{\partial R}{\partial z^\alpha}\Big) -2\frac{\partial^2 R}{\partial z^\alpha\partial z^\beta}\cdot\frac{\partial R}{\partial z^n}. $$ Both expressions in parentheses are equal to zero, and we obtain \begin{equation} \frac{\partial g_{\alpha\beta}}{\partial z^n} =-2\frac{\partial^2 R}{\partial z^\alpha\partial z^\beta}\cdot\frac{\partial R}{\partial z^n}. \label{2.13} \end{equation} Let $ N=\frac{\nabla p}{\vert\nabla p\,\vert}=\vert\nabla p\,\vert\,\frac{\partial R}{\partial z^n} $ be the unit normal vector of the hypersurface $M_{z^n}$. Recall the Gauss formula, e.g., \cite[Chapter~7, Section~3]{KN}, $ \frac{\partial^2R}{\partial z^\alpha\partial z^\beta} =\Gamma^\gamma_{\alpha\beta}\frac{\partial R}{\partial z^\gamma}+b_{\alpha\beta}N $ (with the tangent and normal to $M$ components of $\frac{\partial^2R}{\partial z^\alpha\partial z^\beta}$), where $b_{\alpha\beta}$ are the coefficients of II for $M_{z^n}$ in coordinates $(z^1,\dots,z^{n-1})$. Taking the scalar product of this equality with $N$ and using the orthogonality of $N$ to $\frac{\partial R}{\partial z^\gamma}$, we obtain $\frac{\partial^2R}{\partial z^\alpha\partial z^\beta}\cdot N=b_{\alpha\beta}$. Since $N=\vert\nabla p\,\vert\,\frac{\partial R}{\partial z^n}$, we get $ \frac{\partial^2R}{\partial z^\alpha\partial z^\beta}\cdot \frac{\partial R}{\partial z^n}=\vert\nabla p\,\vert^{-1}b_{\alpha\beta} $. Using this equality, formula \eqref{2.13} becomes $ \frac{\partial g_{\alpha\beta}}{\partial z^n}=-2\vert\nabla p\,\vert^{-1}b_{\alpha\beta} $. Substituting this value into \eqref{2.12}, we get $$ (u\cdot\nabla u)^n=\vert\nabla p\,\vert\,b_{\alpha\beta}u^\alpha u^\beta=\vert\nabla p\,\vert\,{\rm{II}}(u,u). $$ Inserting this expression into \eqref{2.11}, we arrive to \eqref{2.10}. \end{proof} Although we use some special coordinates in the above proof, the obtained equations \eqref{2.3} and \eqref{2.10} are independent of a coordinates choice. All our arguments in this section are invertible, i.e., the following statement is valid. \begin{proposition} Let a smooth vector field $u$ and smooth real function $p$ be defined on an open set $U\subset{\mathbb R}^n$. Choose a point $x_0\in U$ such that $\nabla p(x_0)\ne 0$ and set $p_0=p(x_0)$. Assume that in some neighborhood of $x_0$ the integral curves of $u$ are geodesics of level hypersurfaces $M_p$ and the equations \eqref{2.3} and \eqref{2.10} are satisfied. Then $(u,p)$ is a Gavrilov solution to the Euler equations in some neighborhood of $x_0$. \end{proposition} \section{One example of Gavrilov flow} \label{sec:ex} The following example in even dimensions was suggested by Ya. Bazaykin in a private communication. Most likely, this was known before. In the 2D case it is called a {\it vortex}~\cite{MB}. Let $(x_1,\dots,x_{2n})$ be Cartesian coordinates in ${\mathbb R}^{2n}$. Set \begin{equation} u_{2j-1}(x)=-x_{2j},\ u_{2j}(x)=x_{2j-1}\ (j=1,\dots, n),\quad p(x)=\frac{1}{2}\,\vert x\vert^2. \label{3.1} \end{equation} It is easy to check that the equations \eqref{1.1}--\eqref{1.3} hold in ${\mathbb R}^{2n}$. Observe that $\vert u\vert^2=2p$. All points $x\ne0$ are regular. Isobaric hypersurfaces are spheres $M_p=\{x\in{\mathbb R}^{2n}: \vert x\vert^2=2p\}$. Integral curves of $u$ (particle trajectories) are circles centered at the origin. Every sphere $M_p$ is foliated by particle trajectories. This foliation coincides with the Hopf fiber bundle ${\mathbb S}^{2n-1}\rightarrow{\mathbb C}P^{n-1}$ of an odd-dimensional sphere over the complex projective~space. A Gavrilov flow on ${\mathbb R}^{2n+1}$ can be obtained as a direct product of the flow \eqref{3.1} with a constant velocity flow. Namely, \begin{equation} u_{2j-1}(x)=-x_{2j},\ u_{2j}(x)=x_{2j-1}\ (j=1,\dots, n),\quad u_{2n+1}(x)=a=\mbox{const} \label{3.2} \end{equation} and $p(x)=\frac{1}{2}(x_1^2+\dots+x_n^2)$. Isobaric hypersurfaces are cylinders ${\mathbb S}^{2n-1}\times\mathbb R$, and particle trajectories are either circles (when $a=0$) or helices (when $a\ne0$). In aim to apply the Gavrilov localization to the flow \eqref{3.1}, choose a compactly supported smooth function $\alpha:[0,\infty)\rightarrow\mathbb R$ such that $\alpha(r)=0$ for $r\le\varepsilon$ with some $\varepsilon>0$ and define the function $\beta:[0,\infty)\rightarrow\mathbb R$ by $\beta(r)=-\int_r^\infty s\alpha^2(s)\,ds$. Then $$ \tilde u(x)=\alpha(\vert x\vert)u(x),\quad \tilde p(x)=\beta(\vert x\vert) $$ is a smooth compactly supported Gavrilov flow on ${\mathbb R}^{2n}$ satisfying $\vert\tilde u\vert^2=\psi(\tilde p)$ with a function $\psi$ uniquely determined by $\alpha$. In particular, if $\alpha$ is supported in $(r_0-\delta,r_0+\delta)$ for some $r_0>\delta>0$, then the velocity $\tilde u$ is supported in the spherical layer $\{x\in\mathbb R^{2n}: r_0-\delta<\vert x\vert<r_0+\delta\}$, and the pressure $\tilde p$ is supported in the ball $\{x\in\mathbb R^{2n}: \vert x\vert<r_0+\delta\}$ with $\tilde p=\mbox{const}$ in the smaller ball $\{x\in\mathbb R^{2n}: \vert x\vert\le r_0-\delta\}$. Then we can take a linear combination of several such localized flows with disjoints supports. In particular, a periodic Gavrilov flow can be constructed in this way. \section{Plain sections of a Gavrilov flow} \label{sec:section} Let a smooth solution $(u,p)$ of the Euler equations \eqref{1.1}--\eqref{1.2} be defined on ${\mathbb R}^n$. Assume the solution to decay sufficiently fast at infinity together with first order derivatives (e.g., it can be a smooth compactly supported solution). Then, see \cite{NV}, the equality \begin{equation} \int\nolimits_P\big(\xi\cdot u(x)\big)\big(\nu\cdot u(x)\big)\,dx=0 \label{4.1} \end{equation} holds for every affine hyperplane $P\subset{\mathbb R}^n$ and every vector $\xi\in{\mathbb R}^n$ parallel to $P$, where $\nu$ is the normal vector to the hyperplane $P$ and $dx$ stands for the $(n-1)$-dimensional Lebesgue measure on $P$. Actually there are $(n-1)$ independent equations in \eqref{4.1} since the vector $\xi$ can take $(n-1)$ linearly independent values from the space $\nu^\bot=\{\xi\in{\mathbb R}^n\mid \nu\cdot\xi=0\}$. For a Gavrilov flow, the equation \eqref{4.1}, combined with the Gavrilov localization, yields an interesting statement. The following theorem can be easily generalized to the case of an arbitrary dimension. \begin{theorem} Let $(u,p)$ be a smooth Gavrilov flow defined on ${\mathbb R}^3$ and sufficiently fast decaying at infinity together with first order derivatives. Let $M_{p_0}$ be a regular isobaric surface and let an affine plane $P_0$ transversally intersect $M_{p_0}$. Then, for any $p$ sufficiently close to $p_0$ and for any affine plane $P$ sufficiently close to $P_0$, we get \begin{equation} \int\limits_{M_p\cap P}\frac{1}{\vert\nabla q(x)\vert}\big(\xi\cdot u(x)\big)\big(\nu\cdot u(x)\big)\,ds=0, \label{4.2} \end{equation} where $q\in C^\infty(P)$ is the restriction of the function $p$ to the plane $P$, $\nu$ is the unit normal vector to $P$, and $\xi$ is an arbitrary vector parallel to $P$. The integration in \eqref{4.2} is performed with respect to the arc length $ds$ of the curve $M_p\cap P$. \end{theorem} \begin{proof} Since $M_{p_0}$ and $P_0$ intersect transversally, the same is true for $M_p$ and $P$ for any $p$ close to $p_0$ and for any plane $P$ close to $P_0$. We fix such $p$ and $P$, set $q=p\,\vert_P$ and $\gamma=M_p\cap P$. Observe that $\gamma$ is a regular curve on the plane $P$ and the gradient $\nabla q$ does not vanish in some neighborhood of $\gamma$. Therefore the integral in \eqref{4.2} is well defined. We parameterize the curve $\gamma$ by the arc length, $\gamma=\gamma(s)$. Then we choose local coordinates $(s,\tau)$ in a neighborhood $U\subset P$ of $\gamma$ in the same way as in the proof of Proposition~\ref{2.3}. Namely, the coordinates are chosen so that $x(s,0)=\gamma(s)$ and $q(x(s,\tau))=p+\tau$. For every $s_0$, the coordinate line $\delta(\tau)=x(s_0,\tau)$ starts at $\gamma(s_0)$ orthogonally to $\gamma$ with the initial speed $\vert\dot\delta(0)\vert=\frac{1}{\vert\nabla q(\gamma(s_0))\vert}$. Therefore the area form $dx$ of the plane $P$ is written in the chosen coordinates as $dx=\big(\frac{1}{\vert\nabla q(\gamma(s))\vert}+o(\tau)\big)ds\,d\tau$. Fix a smooth function $\mu:\mathbb R\rightarrow\mathbb R$ such that $\mu(r)=0$ for $\vert r\vert\ge1$, $\mu(r)>0$ for $\vert r\vert<1$, and $\int_{-1}^1\mu(r)\,dr=1$. For $\varepsilon>0$, set $\alpha_\varepsilon(r)=\sqrt{\mu((r-c)/\varepsilon)}$. Using the latter function, we define the localized Gavrilov flow $(\tilde u,\tilde p)$ by $\tilde u=\alpha_\varepsilon(p)\,u,\ \nabla\tilde p=\alpha_\varepsilon^2(p)\nabla p$. Applying \eqref{4.1} to $(\tilde u,\tilde p)$, we obtain $\int_P\alpha_\varepsilon(q(x))\big(\xi\cdot u(x)\big)\big(\nu\cdot u(x)\big)\,dx=0$. In the chosen coordinates, this is written as \begin{equation} \int\limits_\gamma\int\limits_{-\varepsilon}^\varepsilon \mu(\tau/\varepsilon)\Big(\frac{1}{\vert\nabla q(\gamma(s))\vert}+o(\tau)\Big) \big(\xi\cdot u(x(s,\tau))\big)\big(m\cdot u(x(s,\tau))\big)\,d\tau\,ds=0. \label{4.3} \end{equation} The integrand can be represented as \[ \begin{aligned} & \mu(\tau/\varepsilon)\Big(\frac{1}{\vert\nabla q(\gamma(s))\vert}+o(\tau)\Big) \big(\xi\cdot u(x(s,\tau))\big)\big(\nu\cdot u(x(s,\tau))\big)\\ &\qquad =\mu(\tau/\varepsilon)\frac{1}{\vert\nabla q(\gamma(s))\vert} \big(\xi\cdot u(\gamma(s))\big)\big(\nu\cdot u(\gamma(s))\big)+o(1). \end{aligned} \] The variables $s$ and $\tau$ are separated up to $o(\tau)$ on the right-hand side of the latter formula. Using this representation and $\int_{-\varepsilon}^\varepsilon\mu(\tau/\varepsilon)\,d\tau=\varepsilon$, we obtain from \eqref{4.3} $ \varepsilon\int\limits_\gamma\frac{1}{\vert\nabla q(\gamma(s))\vert} \big(\xi\cdot u(\gamma(s))\big)\big(\nu\cdot u(\gamma(s))\big)\,ds+o(\varepsilon)=0. $ In the limit as $\varepsilon\rightarrow0$, this gives~\eqref{4.2}. \end{proof} \section{Differential equations for a Gavrilov flow} \label{sec:eqs} Let a Gavrilov flow $(u,p)$ be defined on an open set of ${\mathbb R}^3$ with Cartesian coordinates $(x,y,z)$. Assume that, for $p\in(-p_0,p_0)$, the regular isobaric surface $M_p$ is the graph \begin{equation} z=f(p;x,y)\quad \big((x,y)\in U\big) \label{5.1} \end{equation} of a smooth function $f$ for some open set $U\subset{\mathbb R}^2$. The first and the second quadratic forms of this $M_p$ are \begin{eqnarray} \nonumber && I =(1+f'_x{}^2)dx^2+2f'_xf'_y\,dxdy+(1+f'_y{}^2)dy^2,\\ && II=\pm\frac{1}{\sqrt{1+f'_x{}^2+f'_y{}^2}}\big(f''_{xx}\,dx^2+2f''_{xy}\,dxdy+f''_{yy}\,dy^2\big), \label{5.3} \end{eqnarray} where the sign depends on the choice of the unit vector normal to $M_p$. Christoffel symbols of this metric are \begin{equation} \begin{aligned} \Gamma^x_{xx}&=\frac{f'_xf''_{xx}}{1+f'_x{}^2+f'_y{}^2},\quad \Gamma^x_{xy}=\frac{f'_xf''_{xy}}{1+f'_x{}^2+f'_y{}^2},\quad \Gamma^x_{yy}=\frac{f'_xf''_{yy}}{1+f'_x{}^2+f'_y{}^2},\\ \Gamma^y_{xx}&=\frac{f'_yf''_{xx}}{1+f'_x{}^2+f'_y{}^2},\quad \Gamma^y_{xy}=\frac{f'_yf''_{xy}}{1+f'_x{}^2+f'_y{}^2},\quad \Gamma^y_{yy}=\frac{f'_yf''_{yy}}{1+f'_x{}^2+f'_y{}^2}. \end{aligned} \label{5.4} \end{equation} Let $(u^x,u^y)$ be geometric coordinates of the vector field $u$, i.e., $u=u^x(p;x,y)\frac{\partial}{\partial x}+u^y(p;x,y)\frac{\partial}{\partial y}$, where $\frac{\partial}{\partial x}$ and $\frac{\partial}{\partial y}$ are considered as coordinate vector fields tangent to the surface $M_p$. Using \eqref{5.4}, we compute covariant derivatives \begin{equation} \begin{aligned} \nabla_{\!x}u^x&=\frac{\partial u^x}{\partial x}+\frac{f'_x (f''_{xx}u^x+f''_{xy}u^y)}{1+f'_x{}^2\!+\!f'_y{}^2},\quad \nabla_{\!y}u^x=\frac{\partial u^x}{\partial y}+\frac{f'_x (f''_{xy}u^x+f''_{yy}u^y)}{1+f'_x{}^2\!+\!f'_y{}^2},\\ \nabla_{\!x}u^y&=\frac{\partial u^y}{\partial y}+\frac{f'_y (f''_{xx}u^x+f''_{xy}u^y)}{1+f'_x{}^2\!+\!f'_y{}^2},\quad \nabla_{\!y}u^y=\frac{\partial u^y}{\partial y}+\frac{f'_y (f''_{xy}u^x+f''_{yy}u^y)}{1+f'_x{}^2\!+\!f'_y{}^2}. \end{aligned} \label{5.7} \end{equation} In particular, \begin{equation} \mbox{div}\,u=\nabla_{\!x}u^x{+}\nabla_{\!y}u^y= \frac{\partial u^x}{\partial x}{+}\frac{\partial u^y}{\partial y} +\frac{(f'_xf''_{xx}{+}f'_yf''_{xy})u^x{+}(f'_xf''_{xy}{+}f'_yf''_{yy})u^y}{1\!+\!f'_x{}^2\!+\!f'_y{}^2}. \label{5.8} \end{equation} Substituting expressions \eqref{5.7} into the formula $ \nabla_{\!u}u=(u^x\nabla_{\!x}u^x+u^y\nabla_{\!y}u^x)\frac{\partial}{\partial x} +(u^x\nabla_{\!x}u^y+u^y\nabla_{\!y}u^y)\frac{\partial}{\partial y}, $ we get $$ \begin{aligned} \nabla_{\!u}u&=\Big(u^x\frac{\partial u^x}{\partial x}+u^y\frac{\partial u^x}{\partial y} {+}\frac{f'_x}{1\!+\!f'_x{}^2\!+\!f'_y{}^2}\big(f''_{xx}(u^x)^2{+}2f''_{xy}u^xu^y{+}f''_{yy}(u^y)^2\big)\Big) \frac{\partial}{\partial x}\\ &+\Big(u^x\frac{\partial u^y}{\partial x}+u^y\,\frac{\partial u^y}{\partial y} {+}\frac{f'_y}{1\!+\!f'_x{}^2\!+\!f'_y{}^2}\big(f''_{xx}(u^x)^2{+}2f''_{xy}u^xu^y{+}f''_{yy}(u^y)^2\big)\Big) \frac{\partial}{\partial y}\,. \end{aligned} $$ Being a geodesic vector field, $u$ satisfies the condition $\nabla_{\!u}u=0$. We arrive to the system \begin{equation} \begin{aligned} u^x\,\frac{\partial u^x}{\partial x}+u^y\,\frac{\partial u^x}{\partial y} +\frac{f'_x}{1\!+\!f'_x{}^2\!+\!f'_y{}^2}\big(f''_{xx}(u^x)^2+2f''_{xy}u^xu^y+f''_{yy}(u^y)^2\big)&=0,\\ u^x\,\frac{\partial u^y}{\partial x}+u^y\,\frac{\partial u^y}{\partial y} +\frac{f'_y}{1\!+\!f'_x{}^2\!+\!f'_y{}^2}\big(f''_{xx}(u^x)^2+2f''_{xy}u^xu^y+f''_{yy}(u^y)^2\big)&=0. \end{aligned} \label{5.9} \end{equation} Let us express $\vert\nabla p\,\vert$ in terms of the function $f$. Let $p(x,y,z)$ be the pressure in Cartesian coordinates. We have the identity $p(x,y,f(p;x,y))=p$. Differentiate the identity to get $$ \begin{aligned} p'_x(x,y,f(p;x,y))+p'_z(x,y,f(p;x,y))f'_x(p;x,y)&=0,\\ p'_y(x,y,f(p;x,y))+p'_z(x,y,f(p;x,y))f'_y(p;x,y)&=0,\\ p'_z(x,y,f(p;x,y))f'_p(p;x,y)&=1. \end{aligned} $$ From this we get $p'_x=-\frac{f'_x}{f'_p},\ p'_y=-\frac{f'_y}{f'_p},\ p'_z=\frac{1}{f'_p}$. The derivative $f'_p$ does not vanish on a regular isobaric surface \eqref{5.1}. Thus, \begin{equation} \vert\nabla p\,\vert^2=p'_x{}^2+p'_y{}^2+p'_z{}^2=\frac{1+f'_x{}^2+f'_y{}^2}{f'_p{}^2}. \label{5.11} \end{equation} Substituting expressions \eqref{5.8} and \eqref{5.11} into \eqref{2.3}, we arrive to the equation $$ \begin{aligned} &\frac{\partial u^x}{\partial x}+\frac{\partial u^y}{\partial y} +\frac{(f'_xf''_{xx}+f'_yf''_{xy})u^x+(f'_xf''_{xy}+f'_yf''_{yy})u^y}{1\!+\!f'_x{}^2\!+\!f'_y{}^2}\\ &=\frac{1}{2}\cdot\frac{f'_p{}^2}{1+f'_x{}^2+f'_y{}^2} \Big[u^x\frac{\partial}{\partial x}\Big(\frac{1+f'_x{}^2+f'_y{}^2}{f'_p{}^2}\Big) +u^y\frac{\partial}{\partial y}\Big(\frac{1+f'_x{}^2+f'_y{}^2}{f'_p{}^2}\Big)\Big]. \end{aligned} $$ After the obvious simplification, it becomes \[ \frac{\partial u^x}{\partial x}+\frac{\partial u^y}{\partial y}+\frac{1}{f'_p}(u^xf''_{px}+u^yf''_{py})=0. \] Substituting expressions \eqref{5.3} and \eqref{5.11} into \eqref{2.10}, we arrive to the equation $f''_{xx}(u^x)^2+2f''_{xy}u^xu^y+f''_{yy}(u^y)^2=\frac{1+f'_x{}^2+f'_y{}^2}{\vert f'_p\vert}$. Using this, equations \eqref{5.9} can be written as $$ u^x\,\frac{\partial u^x}{\partial x}+u^y\,\frac{\partial u^x}{\partial y}+\frac{f'_x}{\vert f'_p\vert}=0,\quad u^x\,\frac{\partial u^y}{\partial x}+u^y\,\frac{\partial u^y}{\partial y} +\frac{f'_y}{\vert f'_p\vert}=0. $$ We have thus proved the following \begin{proposition} \label{P5.1} Let a Gavrilov flow $(u,p)$ be defined on an open set of ${\mathbb R}^3$. Assume that for $p\in(-p_0,p_0)$ the isobaric surface $M_p$ is regular and coincides with the graph of a smooth function $z=f(p;x,y)\ \big((x,y)\in U\subset{\mathbb R}^2\big)$. Write the restriction of the vector field $u$ to the surface $M_p$ in the form $u=u^x(p;x,y)\frac{\partial}{\partial x}+u^y(p;x,y)\frac{\partial}{\partial y}$. Then the derivative $f'_p$ does not vanish and the functions $f(p;x,y),u^x(p;x,y)$ and $u^y(p;x,y)$ satisfy the equations \begin{eqnarray} \frac{\partial (f'_pu^x)}{\partial x}+\frac{\partial (f'_pu^y)}{\partial y}=0, \label{5.12}\\ u^x\,\frac{\partial u^x}{\partial x}+u^y\,\frac{\partial u^x}{\partial y}+\frac{f'_x}{\vert f'_p\vert}=0, \label{5.13} \\ u^x\,\frac{\partial u^y}{\partial x}+u^y\,\frac{\partial u^y}{\partial y}+\frac{f'_y}{\vert f'_p\vert}=0, \label{5.14}\\ f''_{xx}(u^x)^2+2f''_{xy}u^xu^y+f''_{yy}(u^y)^2=\frac{1+f'_x{}^2+f'_y{}^2}{\vert f'_p\vert}. \label{5.15} \end{eqnarray} \end{proposition} It is easy to check that the system \eqref{5.12}--\eqref{5.15} has the following solution: \begin{eqnarray} f(x,y,p)=f(x,p)=-\sqrt{r^2(p)-x^2}\quad\big(-r(p)<x<r(p)\big), \label{5.16}\\ u^x(x,y,p)=u^x(x,p)=\frac{\sqrt{r^2(p)-x^2}}{\sqrt{r(p)\vert r'(p)\vert}},\quad u^y(x,y,p)=u^y(p)=b(p), \label{5.17} \end{eqnarray} where $r(p)$ is a smooth positive function with non-vanishing derivative, and $b(p)$ is an arbitrary smooth function. For every $p$, the graph $M_p$ of the function $(x,y)\mapsto f(x,p)$ is the half of the cylinder ${\tilde M}_p=\{(x,y,z)\mid x^2+z^2=r^2(p)\}$. Observe that ${\tilde M}_p$ is a surface of revolution around the $y$-axis. Thus, \eqref{5.16}--\eqref{5.17} is an axisymmetric Gavrilov flow, and we do not know any solution to the system \eqref{5.12}--\eqref{5.15} different of the (modified) solution \eqref{5.16}--\eqref{5.17}. We will study axisymmetric Gavrilov flows in the next section. \section{Axisymmetric Gavrilov flows} \label{sec:axi} Let $(r,z,\theta)$ be cylindrical coordinates in ${\mathbb R}^3$ related to Cartesian coordinates $(x_1,x_2,x_3)$ by $ x_1=r\cos\theta,\ x_2=r\sin\theta,\ x_3=z. $ We study a Gavrilov flow $(u,p)$ invariant under rotations around the $z$-axis. The flow is defined in an open set $\tilde U\subset\{(r,z,\theta): r>0\}$ invariant under rotations around the $z$-axis. Such a rotationally invariant set $\tilde U$ is uniquely determined by the two-dimensional set $U=\tilde U\cap\{\theta=0\}\subset\{(r,z): r>0\}$. For brevity we say that an axisymmetric Gavrilov flow is defined in $U$. A regular isobaric surface $M_p$ is a surface of revolution determined by its {\it generatrix} $\Gamma_p=M_p\cap\,U$. We parameterize the curve $\Gamma_p$ by the arc length $r=R(p,t)>0,\ z=Z(p,t)$, \begin{equation} R'_t{}^2+Z'_t{}^2=1. \label{6.1} \end{equation} The variables $(t,\theta)$ serve as coordinates on the isobaric surface $M_p$. Since the vector field $u$ is tangent to $M_p$, it is uniquely represented as $u=u^t(p,t)\frac{\partial}{\partial t}+u^\theta(p,t)\frac{\partial}{\partial\theta}$, where $\big(u^t(p,t),u^\theta(p,t)\big)$ are {\it geometric coordinates} of $u$. {\it Physical coordinates} of $u$ are given by \begin{equation} u=u_r(r,z)e_r+u_z(r,z)e_z+u_\theta(r,z)e_\theta, \label{6.2} \end{equation} where $e_r,e_z,e_\theta$ are unit coordinate vectors. Such coordinates are related~by \begin{equation} u_r=R'_tu^t,\quad u_z=Z'_tu^t,\quad u_\theta=Ru^\theta. \label{6.3} \end{equation} We are going to write down differential equations for a Gavrilov flow in terms of the functions $(R,Z,u^t, u^\theta)$. First, the first quadratic form of $M_p$ in coordinates $(t,\theta)$ is \begin{equation} I=dt^2+R^2d\theta^2. \label{6.4} \end{equation} Christoffel symbols of the metric are $\Gamma^t_{\theta\theta}=-RR'_t,\ \Gamma^\theta_{t\theta}=\frac{R'_t}{R},\ \Gamma^t_{tt}=\Gamma^t_{t\theta}=\Gamma^\theta_{tt}=\Gamma^\theta_{\theta\theta}=0$. Using these formulas, we calculate $$ \nabla_{\!t}u^t=\frac{\partial u^t}{\partial t},\quad \nabla_{\!\theta}u^t=-RR'_t\,u^\theta,\quad \nabla_{\!t}u^\theta=\frac{\partial u^\theta}{\partial t}+\frac{R'_t}{R}\,u^\theta,\quad \nabla_{\!\theta}u^\theta=\frac{R'_t}{R}\,u^t. $$ In particular, \begin{equation} \mbox{div}\,u=\nabla_{\!t}u^t+\nabla_{\!\theta}u^\theta =\frac{\partial u^t}{\partial t}+\frac{R'_t}{R}\,u^t. \label{6.6} \end{equation} The restriction of $u$ to $M_p$ is a geodesic vector field, i.e., $\nabla_{\!u}u=0$. This gives the system \begin{equation}\label{6.7} u^t\,\frac{\partial u^t}{\partial t}-RR'_t(u^\theta)^2=0,\quad u^t\,\frac{\partial u^\theta}{\partial t}+\frac{2R'_t}{R}\,u^tu^\theta=0. \end{equation} Let us express $\vert\nabla p\,\vert$ in terms of the functions $R(p,t)$ and $Z(p,t)$. Let $p=p(r,z)$ be the pressure in cylindric coordinates (it is independent of $\theta$). We have the identity $ p\big(R(p,t),Z(p,t)\big)=p. $ Differentiating the identity with respect to $p$ and $t$, we arrive to the linear algebraic system with unknowns $p'_r\big(R(p,t),Z(p,t)\big)$ and $p'_z\big(R(p,t),Z(p,t)\big)$ $$ \begin{aligned} R'_p(p,t)p'_r\big(R(p,t),Z(p,t)\big)+Z'_p(p,t)p'_z\big(R(p,t),Z(p,t)\big)&=1,\\ R'_t(p,t)p'_r\big(R(p,t),Z(p,t)\big)+Z'_t(p,t)p'_z\big(R(p,t),Z(p,t)\big)&=0. \end{aligned} $$ Solving the system, we have $$ p'_r\big(R(p,t),Z(p,t)\big)=J^{-1}(p,t)Z'_t(p,t),\ p'_z\big(R(p,t),Z(p,t)\big)=-J^{-1}(p,t)R'_t(p,t), $$ where \begin{equation} J=\left\vert\begin{array}{cc}R'_p&R'_t\\mathbb Z'_p&Z'_t\end{array}\right\vert \label{6.9} \end{equation} is the Jacobian of the transformation $r=R(p,t),\quad z=Z(p,t)$. The Jacobian does not vanish. This implies with the help of \eqref{6.1} \begin{equation} \vert\nabla p\,\vert=\vert J\vert^{-1}. \label{6.11} \end{equation} Substituting expressions \eqref{6.6} and \eqref{6.11} into \eqref{2.3}, we arrive to the equation \begin{equation} \frac{\partial u^t}{\partial t}+\Big(\frac{R'_t}{R}+\frac{J'_t}{J}\Big)u^t=0. \label{6.12} \end{equation} The second quadratic form of the surface $M_p$ is expressed in coordinates $(t,\theta)$ by $ II=-((R'_tZ''_{tt}-Z'_tR''_{tt})\,dt^2+RZ'_t\,d\theta^2). $ The sign on the right-hand side is chosen taking our agreement into account: the unit normal vector to the surface $M_p$ must coincide with $\frac{\nabla p}{\vert\nabla p\,\vert}$. Recall that the curvature $\kappa=\kappa(p,t)$ of the plane curve $r=R(p,t),z=Z(p,t)$ is expressed, under the condition \eqref{6.1}, by \begin{equation} \kappa=R'_tZ''_{tt}-Z'_tR''_{tt}. \label{6.13} \end{equation} Using the latter equality, the previous formula takes the form \begin{equation} II=-(\kappa\,dt^2+RZ'_t\,d\theta^2). \label{6.14} \end{equation} By \eqref{6.11} and \eqref{6.14}, the equation \eqref{2.10} takes the form \begin{equation} \kappa(u^t)^2+RZ'_t(u^\theta)^2=\vert J\vert^{-1}. \label{6.15} \end{equation} We obtained the system of five equations \eqref{6.1}, (\ref{6.7}a,b), \eqref{6.12}, \eqref{6.15} in four unknown functions $(R,Z,u^t,u^\theta)$. The functions $J$ and $\kappa$ participating in the system are expressed through $(R,Z)$ by \eqref{6.9} and \eqref{6.13} respectively. We proceed to the analysis of the system. All isobaric surfaces $M_p$ under consideration are assumed to be regular and connected. The equation \eqref{6.12} implies the following alternative for every $p_0$: either $u^t(p_0,t)\neq0$ for all $t$ or $u^t(p_0,t)\equiv0$. The second case of the alternative is realized in the example \eqref{3.2} with $a=0$. The converse statement is true at least partially: If $u^t(p_0,t)\equiv0$, then $M_{p_0}$ coincides with the cylinder $\{r=\mbox{const}>0\}$ and particle trajectories living on $M_{p_0}$ are horizontal circles. Indeed, if $u^t(p_0,t)\equiv0$ then, as is seen from \eqref{6.1} and (\ref{6.7}b), $R'_t(p_0,t)\equiv0$ and $Z'_t(p_0,t)\equiv\pm1$. Nevertheless, it is possible that $u^t(p_0,t)\equiv0$ but $u^t(p,t)\neq0$ for $p$ close to $p_0$; the corresponding example can be constructed by a slight modification of~\eqref{3.2}. To avoid degenerate cases of the previous paragraph, we additionally assume $u^t(p,t)\neq0$ for all $(p,t)$. Recall that we study flows up to the equivalence \eqref{1.4}. Therefore the latter assumption can be written without lost of generality in the form \begin{equation} u^t(p,t)>0\quad\mbox{for all}\quad(p,t). \label{6.16} \end{equation} The equation (\ref{6.7}a) simplifies under the assumption \eqref{6.16} to the following: \begin{equation} \frac{\partial u^\theta}{\partial t}+\frac{2R'_t}{R}\,u^\theta=0. \label{6.17} \end{equation} If $u^\theta(p,t_0)=0$ for some $t_0$, then \eqref{6.17} implies that $u^\theta(p,t)=0$ for all $t$. On the other hand, assuming that $u^\theta(p,t)\neq0$ for all $t$ and for a fixed $p$, we can rewrite \eqref{6.17} in the form $\frac{\partial(\log u^\theta)}{\partial t}+\frac{\partial(\log R^2)}{\partial t}=0$. From this we get \begin{equation} u^\theta(p,t)=\frac{b(p)}{R^2(p,t)} \label{6.18} \end{equation} with some function $b(p)$. This equality is true also for such $p_0$ that $u^\theta(p_0,t)=0$ for all $t$, just by setting $b(p_0)=0$. The equality \eqref{6.18} implies smoothness of the function $b$. Substituting the expression \eqref{6.18} into (\ref{6.7}a), we get $\frac{\partial (u^t)^2}{\partial t}=2\,\frac{R'_t}{R^3}b^2(p)$, that can be written in the form $\frac{\partial (u^t)^2}{\partial t}=-\frac{\partial}{\partial t}\big(\frac{b^2(p)}{R^2}\big)$. From this we obtain \begin{equation} u^t(p,t)=\frac{\sqrt{d(p)R^2(p,t)-b^2(p)}}{R(p,t)}, \label{6.19} \end{equation} where $d(p)$ is a smooth function satisfying \begin{equation} d(p)R^2(p,t)-b^2(p)>0. \label{6.20} \end{equation} By \eqref{6.4}, $\vert u\vert^2=(u^t)^2+R^2(u^\theta)^2$ is valid. Substituting values \eqref{6.18}--\eqref{6.19}, we get \begin{equation} \vert u(p,t)\vert^2=d(p). \label{6.21} \end{equation} Thus, we have discovered an important phenomenon: for an axisymmetric Gavrilov flow, all particles living on an isobaric surface $M_p$ move with the same speed. In other words, constants $c$ and $C$ in the Bernoulli law \eqref{2.1} can be expressed through each other. For instance, $\vert u\vert^2=3p$ in example \cite{G}, and $\vert u\vert^2=mp$ with an arbitrary positive constant $m$ in example \cite{CLV}. We guess that $d(p)$ in \eqref{6.21} can be an arbitrary smooth positive function. Most likely, the phenomenon is absent for a general (not axisymmetric) Gavrilov flow, at least we cannot derive a relation of \eqref{6.21} type from \eqref{5.12}--\eqref{5.15}. Studying Gavrilov flows up to equivalence, we can multiply $u(p,t)$ by a non-vanishing smooth function $\varphi(p)$. This opportunity was already used to fix the sign of $u^t$ in \eqref{6.16}. We still have the freedom of multiplying $u(p,t)$ by a positive smooth function $\varphi(p)$ together with the corresponding change of the pressure. Choosing $\varphi(p)=d(p)^{-1/2}$ and denoting the new Gavrilov flow by $(u,p)$ again, we simplify \eqref{6.21} to the following: \begin{equation} \vert u(p,t)\vert^2=1. \label{6.22} \end{equation} The inequality \eqref{6.20} becomes now $\vert b(p)\vert<R(p,t)$, and formula \eqref{6.19} takes the form \begin{equation} u^t(p,t)=\frac{\sqrt{R^2(p,t)-b^2(p)}}{R(p,t)}, \label{6.24} \end{equation} We continue our analysis under assumptions \eqref{6.16} and \eqref{6.22}. In virtue of \eqref{6.16}, the equation \eqref{6.12} can be written in the form $ \frac{\partial(\log u^t)}{\partial t}+\frac{\partial(\log (\vert J\vert R))}{\partial t}=0. $ This implies \begin{equation} u^t(p,t)=\frac{\alpha^2(p)}{\vert J(p,t)\vert R(p,t)} \label{6.25} \end{equation} with some positive smooth function $\alpha(p)$. Comparing \eqref{6.24} and \eqref{6.25}, we arrive to the equation $\sqrt{R^2(p,t)-b^2(p)}\,\vert J(p,t)\vert=\alpha^2(p)$. Finally, we simplify the equation \eqref{6.15}. Substituting expressions \eqref{6.18} and \eqref{6.24} for $u^\theta$ and $u^t$ into \eqref{6.15}, we obtain $\kappa R(R^2-b^2)+b^2Z'_t= R^3\vert J\vert^{-1}$. Expressing $\vert J\vert$ from \eqref{6.25} and substituting the expression into the latter formula, we arrive to the equation $ \kappa R(R^2-b^2)+b^2Z'_t=\frac{R^3\sqrt{R^2-b^2}}{\alpha^2}. $ We have thus proved the following \begin{theorem} \label{Th6.1} Let an axisymmetric Gavrilov flow $(u,p)$ be defined on an open set $U\subset\{(r,z)\mid r>0\}$. Assume that every isobaric surface $M_p$ is regular and connected. For the surface of revolution $M_p$, let $r=R(p,t)>0,\ z=Z(p,t)$ be the arc length parametrization of the generatrix $\Gamma_p$ of $M_p$. Assume also that in the representation $u=u^t(p,t)\frac{\partial}{\partial t}+u^\theta(p,t)\frac{\partial}{\partial \theta}$ the function $u^t(p,t)$ does not vanish. Then {\rm (1)} there exists a smooth positive function $d(p)$ such that $\vert u\vert^2=d(p)$. Replacing $(u,p)$ with an equivalent Gavrilov flow and denoting the new flow by $(u,p)$ again, we can assume without lost of generality that $u^t(p,t)>0$ and \begin{equation} \vert u\vert^2=1. \label{6.28} \end{equation} {\rm (2)} under the assumption \eqref{6.28}, the functions $R(p,t)$ and $Z(p,t)$ satisfy the equations \begin{eqnarray} R'_t{}^2+Z'_t{}^2=1, \label{6.29} \\ \sqrt{R^2-b^2(p)}\,\big\vert R'_pZ'_t-R'_tZ'_p\big\vert=\alpha^2(p), \label{6.30} \\ \kappa R\big(R^2-b^2(p)\big)+b^2(p)Z'_t=\frac{R^3\sqrt{R^2-b^2(p)}}{\alpha^2(p)} \label{6.31} \end{eqnarray} with some smooth functions $\alpha(p)>0$ and $b(p)$, where $\kappa=\kappa(p,t)$ is the curvature of the plane curve $r=R(p,t),z=Z(p,t)$. The functions $u^t(p,t)$ and $u^\theta(p,t)$ are expressed through $\big(R(p,t),b(p)\big)$ by \begin{equation} u^t=\frac{\sqrt{R^2-b^2(p)}}{R},\quad u^\theta=\frac{b(p)}{R^2}. \label{6.32} \end{equation} \end{theorem} \begin{remark}\rm Let us make some remarks on Theorem \ref{Th6.1}. 1. First, we attract reader's attention to the hypothesis: $M_p$ are connected surfaces. Otherwise functions $\alpha(p)$ and $b(p)$ can be different on different connected components. 2. Equations \eqref{6.29}--\eqref{6.32} are invariant under some transformations. First, the parameter $t$ is defined up to a shift, i.e., nothing changes after the replacement $R(p,t)=\tilde R\big(p,t+t_0(p)\big),Z(p,t)=\tilde Z\big(p,t+t_0(p)\big)$. Second, the equations are invariant under the changes $Z(p,t)=\tilde Z(p,t)+z_0$ and $Z(p,t)=-\tilde Z(p,t),\ \kappa=-\tilde\kappa$, which mean a vertical shift of the origin in ${\mathbb R}^3$ and the change of the direction of the $z$-axis, respectively. 3. Compared to Proposition \ref{P5.1}, Theorem \ref{Th6.1} has an important advantage. The unknown functions $(f,u^x,u^y)$ are not separated in the system \eqref{5.12}--\eqref{5.15} and, probably, cannot be separated. On the other hand, equations \eqref{6.29}--\eqref{6.31} involve only the functions $(R,Z)$ that determine isobaric surfaces $M_p$ (the equations involve also $\alpha(p)$ and $b(p)$ that appear as integration constants). If the system \eqref{6.29}--\eqref{6.31} was solved, the velocity vector field $u$ would be determined by explicit formulas \eqref{6.32}. Of course, the simplification is possible thanks to the Clairaut integral for the equation of geodesics on a surface of revolution. Although the Clairaut integral is not mentioned in our proof of Theorem~\ref{Th6.1}, formulas \eqref{6.32} are actually equivalent to the Clairaut integral. \end{remark} Theorem \ref{Th6.1} has two other useful forms. {\bf Case A.} Assume, under hypotheses of Theorem \ref{Th6.1}, that the curve $\Gamma_p$ is the graph of a function $r=f(p,z)$. Then $f'_p\ne0$ and \eqref{6.29}--\eqref{6.31} are equivalent to the system \begin{eqnarray} {f'_p}^2 = \frac{\alpha(p)(1+{f'_z}^2)}{f^2-b^2(p)}, \label{6.33}\\ \alpha(p) f f''_{zz} = b^2(p) {f'_p}^2 - f^3f'_p(1+{f'_z}^2) \label{6.34} \end{eqnarray} with the same functions $\alpha(p)>0$ and $b(p)$. The function $f$ must satisfy $f(z,p)>\vert b(p)\vert$. The velocity vector $u$ is now represented as $u=u^z(z,p)\frac{\partial}{\partial z}+u^\theta(z,p)\frac{\partial}{\partial\theta}$, where the functions $u^z$ and $u^\theta$ are expressed through $(f,\alpha,b)$ by $u^z=\mbox{sgn}(u^z)\,\frac{\sqrt{\alpha(p)}}{f\vert f'_p\vert},\ u^\theta=\frac{b(p)}{f^2}$, where $\mbox{sgn}(u^z)=\pm1$ is the sign of $u^z$ that is assumed do not vanish. {\bf Case B.} Assume, under hypotheses of Theorem \ref{Th6.1}, that the curve $\Gamma_p$ is the graph of a function $z=g(p,r)$. Then equations \eqref{6.29}--\eqref{6.31} are equivalent to the system \begin{eqnarray} \big(r^2-b^2(p)\big)g'_p{}^2-\alpha(p)(1+g'_r{}^2)=0, \label{6.36}\\ r\alpha(p)g''_{rr}+b^2(p)g'_r g'_p{}^2+r^3 g'_p(1+g'_r{}^2)=0 \label{6.37} \end{eqnarray} with the same functions $\alpha(p)>0$ and $b(p)$. The function $g(p,r)$ is considered for $r\in\big(r_1(p),r_2(p)\big)$ with $\vert b(p)\vert<r_1(p)$. The velocity vector $u$ is now represented as $u=u^r(p,r)\frac{\partial}{\partial r}+u^\theta(p,r)\frac{\partial}{\partial \theta}$, where the functions $u^r$ and $u^\theta$ are expressed through $(g,\alpha,b)$ by $u^r=\mbox{sgn}(u^r)\,\frac{\sqrt{\alpha(p)}}{r\vert g'_p\vert},\ u^\theta=\frac{b(p)}{r^2}$, where $\mbox{sgn}(u^r)=\pm1$ is the sign of $u^r$ that does not~vanish. \section{Consistency conditions} \label{sec:cond} First, we will recall some basic facts from theory of first order PDEs following \cite[Part~I, Section~14]{Ka2}. Let us consider the system of two first order PDEs \begin{equation} F(x,y,z,p,q)=0,\quad G(x,y,z,p,q)=0, \label{7.1} \end{equation} where $z=z(x,y)$ is an unknown function and $p=z'_x,q=z'_y$. We assume $F$ and $G$ to be sufficiently smooth functions defined for $(x,y,z)\in U$ and for all $(p,q)\in{\mathbb R}^2$, where $U\subset{\mathbb R}^3$ is an open set. The system \eqref{7.1} is supplied with the initial condition \begin{equation} z(x_0,y_0)=z_0 \label{7.2} \end{equation} for a point $(x_0,y_0,z_0)\in U$. The {\it Jacobi brackets} (sometimes also called {\it Mayer brackets}) of functions $F(x,y,z,p,q)$ and $G(x,y,z,p,q)$ are defined by \begin{equation} [F,G]=(F'_x+pF'_z)G'_p-(G'_x+pG'_z)F'_p+(F'_y+qF'_z)G'_q-(G'_y+qG'_z)F'_q. \label{7.3} \end{equation} The system \eqref{7.1} is said to be an {\it involutory system} if $[F,G]\equiv0$ for $(x,y;z,p,q)\in U\times{\mathbb R}^2$. The system \eqref{7.1} is said to be a {\it complete system on the open set} $U\times{\mathbb R}^2$ if the equation \begin{equation} [F,G]=0 \label{7.4} \end{equation} is an algebraic corollary of the system \eqref{7.1}, i.e., if \eqref{7.4} holds for $(x,y;z,p,q)\in U\times{\mathbb R}^2$ satisfying \eqref{7.1}. The equation \eqref{7.4} is called the {\it consistency condition} (or {\it integrability condition}) for the system \eqref{7.1}. In the case of a complete system, for an arbitrary point $(x_0,y_0,z_0)\in U$, the initial value problem \eqref{7.1}--\eqref{7.2} has a unique solution at least in some neighborhood of the point $(x_0,y_0)$. Several methods are known for the numerical solution to the IVP \eqref{7.1}--\eqref{7.2}, the Mayer method is the most popular one \cite{Ka2}. Let us return to axisymmetric Gavrilov flows. Under certain additional conditions, systems \eqref{6.33}--\eqref{6.34} and \eqref{6.36}--\eqref{6.37} are equivalent. We study the system \eqref{6.33}--\eqref{6.34}. Recall that the system is considered in a neighborhood of a regular point, where the transform $(p,z)\mapsto\big(f(p,z),z\big)$ is one-to-one. Therefore the derivative $f'_p$ does not vanish. Note that only $b^2$ is involved in \eqref{6.33}--\eqref{6.34}, not the function $b$ itself. To simplify our formulas a bit, we introduce the function $\beta(p)=b^2(p)\ge0$ and rewrite the system \eqref{6.33}--\eqref{6.34} as \begin{eqnarray} (f^2-\beta){f'_p}^2 -\alpha(1+{f'_z}^2)=0, \label{7.5}\\ \alpha f f''_{zz} -\beta {f'_p}^2 + f^3f'_p(1+{f'_z}^2)=0. \label{7.6} \end{eqnarray} It is assumed that the function $f$ satisfies the inequality \begin{equation} f(p,z)>\sqrt{\beta(p)}. \label{7.7} \end{equation} Given functions $\alpha(p)>0$ and $\beta(p)\ge0$, \eqref{7.5}--\eqref{7.6} is an overdetermined system of two PDEs in one unknown function $f(p,z)$. The overdeterminess is caused by the circumstance mentioned in Introduction: a Gavrilov flow is defined by the overdetermined system \eqref{1.1}--\eqref{1.3}. We pose the question: \textit{What conditions should be imposed on $\big(\alpha(p),\beta(p)\big)$ for solvability of the system \eqref{7.5}--\eqref{7.6} at least locally, i.e., in a neighborhood of a given point} $(p_0,z_0)$? For a fixed $p$, \eqref{7.6} can be considered as a second order ODE with an unknown function $f_p(z)=f(p,z)$. Note that the variable $z$ does not explicitly participate in \eqref{7.6}. Such an equation can be reduced to a first order ODE, see \cite[Ssection~15.3]{Ka}. This observation is realized by the following statement. \begin{lemma} \label{L7.1} Let $C^1$-functions $\alpha(p)>0$ and $\beta(p)\ge0$ be defined on an interval $(p_1,p_2)$, where $-\infty\le p_1<p_2\le\infty$. Then the following {\rm (1)} and {\rm (2)} are valid. {\rm (1)} Let $f(p,z)$ be a solution to the system \eqref{7.5}--\eqref{7.6} on a rectangle \begin{equation} (p,z)\in(p_1,p_2)\times(z_1,z_2), \label{7.8} \end{equation} and let the inequality \eqref{7.7} be valid on the rectangle. Then there exists a $C^1$-function $\gamma(p)$ on the interval $(p_1,p_2)$ satisfying the equation \begin{equation} f^2(\varepsilon f^2+\gamma)^2(1+f'_z{}^2)-4\alpha(f^2-\beta)=0 \label{7.9} \end{equation} and the inequalities \begin{equation} 0 < \varepsilon f^2+\gamma \le \frac{2\alpha^{1/2}\sqrt{f^2 - \beta}}{f}, \label{7.10} \end{equation} where $\varepsilon=\pm1$ is the sign of $f'_p$ that does not vanish. {\rm (2)} Conversely, let $f(p,z)>0$ and $\gamma(p)$ satisfy \eqref{7.5} and \eqref{7.9}--\eqref{7.10}. Assume additionally that, for every $p\in(p_1,p_2)$, the derivative $f'_z(p,z)$ is not identically equal to zero on any interval $(z'_1,z'_2)\subset(z_1,z_2)$. Then $f$ solves \eqref{7.5}--\eqref{7.6} on the rectangle \eqref{7.8}. \end{lemma} \begin{proof} First, we find from \eqref{7.5} \begin{equation} f'_p =\varepsilon\, \frac{\alpha^{1/2}\sqrt{1+{f'_z}^2}}{\sqrt{f^2-\beta}}, \label{7.11} \end{equation} where $\varepsilon=\pm1$ is the sign of $f'_p$, and then substitute the expression into \eqref{7.6} \begin{equation} f''_{zz} = \frac{\beta(1+{f'_z}^2)}{f(f^2-\beta)} -\varepsilon\, \frac{f^2(1+{f'_z}^2)^{3/2}}{\alpha^{1/2}\sqrt{f^2-\beta}}. \label{7.12} \end{equation} Let us show that equations \eqref{7.11} and \eqref{7.12} imply \begin{equation} \frac{\partial}{\partial z}\Big(\frac{\sqrt{f^2-\beta}}{f\sqrt{1+{f'_z}^2}} -\varepsilon\,\frac{f^2}{2\alpha^{1/2}}\Big)=0. \label{7.13} \end{equation} For this, we implement the differentiation in \eqref{7.13}. The result can be written as \begin{equation} \Big(f''_{zz}-\frac{\beta(1+{f'_z}^2)}{f(f^2-\beta)}+\varepsilon\,\frac{f^2(1+{f'_z}^2)^{3/2}}{\alpha^{1/2}\sqrt{f^2-\beta}}\Big)f'_z=0. \label{7.14} \end{equation} By \eqref{7.12}, the left-hand side of \eqref{7.14} is identically zero. This proves \eqref{7.13}. The equation \eqref{7.13} means the existence of a function $\gamma(p)$ such that $ \frac{\sqrt{f^2-\beta}}{f\sqrt{1+{f'_z}^2}} -\varepsilon\,\frac{f^2}{2\alpha^{1/2}}=\frac{\gamma}{2\alpha^{1/2}}. $ This can be written in the form \begin{equation} 2\alpha^{1/2}\sqrt{f^2-\beta}=f\sqrt{1+{f'_z}^2}(\varepsilon f^2+\gamma). \label{7.15} \end{equation} By \eqref{7.7}, $f>0$ and $f^2-\beta>0$. Therefore \eqref{7.15} implies the inequalities \eqref{7.10}. Squaring the equation \eqref{7.15}, we get \eqref{7.9}. We have proved the first statement of the lemma. The second assertion of the lemma is proved by interchanging the above arguments using the following additional remark. To pass from \eqref{7.14} to \eqref{7.12}, we need to remove the factor $f'_z$ on the left-hand side of \eqref{7.14}. To do this, it suffices to assume that for every $p\in(p_1,p_2)$, the derivative $f'_z(p,z)$ is not identically zero on any interval $(z'_1,z'_2)\subset(z_1,z_2)$. \end{proof} By Lemma \ref{L7.1}, the system \eqref{7.5}--\eqref{7.6} can be replaced with the following: \begin{equation} \begin{aligned} (f^2-\beta){f'_p}^2 -\alpha(1+{f'_z}^2)&=0,\\ f^2(\varepsilon f^2+\gamma)^2(1+f'_z{}^2)-4\alpha(f^2-\beta)&=0. \end{aligned} \label{7.16} \end{equation} Introducing the notations $\pi=f'_p,\quad\zeta=f'_z$, we write the system as \begin{equation} F(p,z,f,\pi,\zeta)=0,\quad G(p,z,f,\pi,\zeta)=0, \label{7.17} \end{equation} where \begin{equation} \begin{aligned} F(p,z,f,\pi,\zeta)&=(f^2-\beta)\pi^2-\alpha(\zeta^2+1),\\ G(p,z,f,\pi,\zeta)&=f^2(\varepsilon f^2+\gamma)^2(\zeta^2+1)-4\alpha(f^2-\beta). \end{aligned} \label{7.18} \end{equation} Assume that $\alpha(p)>0,\, \beta(p)\ge0$ and $\gamma(p)$ are defined and smooth on an interval $(p_1,p_2)$, where $-\infty\le p_1<p_2\le\infty$. Then $F$ and $G$ are defined and smooth in $U\times{\mathbb R}^2$, where \begin{equation} U=\big\{(p,z,f)\in{\mathbb R}^3\vert: p_1<p<p_2,\ f>\!\sqrt{\beta},\ 0<\varepsilon f^2+\gamma<2\sqrt{\alpha}\sqrt{f^2-\beta}/f\big\}. \label{7.19} \end{equation} The functions $F$ and $G$ are actually independent of $z$ and depend on $p$ through the functions $\alpha(p),\beta(p),\gamma(p)$ only. Up to notations, the system \eqref{7.17} is of the form \eqref{7.1}. \begin{theorem} \label{Th7.1} Given $C^1$-functions $\alpha(p)>0,\beta(p)\ge0$ and $\gamma(p)$ on an interval $(p_1,p_2)$, define $F$ and $G$ by \eqref{7.18} and consider the system of PDEs \eqref{7.17}, where $f=f(p,z)$ is an unknown function and $\pi=f'_p,\zeta=f'_z$. Define an open set $U\subset{\mathbb R}^3$ by \eqref{7.19}. Then {\rm 1.} The system \eqref{7.17} is complete on $U\times{\mathbb R}^2$ if the functions $\alpha,\beta,\gamma$ satisfy \begin{equation} \alpha(p)=\alpha_0e^{3 p}\quad\mbox{\rm with some constant}\ \alpha_0>0, \label{7.20} \end{equation} \begin{equation} \beta'+2\varepsilon\gamma'+3\beta+\varepsilon\gamma=0, \label{7.21} \end{equation} \begin{equation} \gamma\beta'-2\beta\gamma'+3\beta\gamma-4\varepsilon\alpha+\varepsilon\gamma^2=0, \label{7.22} \end{equation} where either $\varepsilon=1$ or $\varepsilon=-1$ and $\beta+\varepsilon\gamma\ne0$. But \eqref{7.17} is not an involutory system. {\rm 2.} Conversely, assume that \eqref{7.17} is a complete system on $U\times{\mathbb R}^2$, where the open set $U\subset{\mathbb R}^3$ is defined by \eqref{7.19}. Assume additionally that for every $p_0\in(p_1,p_2)$ there exists $z_0$ such that \eqref{7.17} has a solution $f(p,z)$ in a neighborhood of $(p_0,z_0)$ satisfying $f'_z(p_0,z_0)\ne0$. Then the functions $\alpha,\beta,\gamma$ satisfy \eqref{7.20}--\eqref{7.22} with some $\varepsilon=\pm1$. \end{theorem} \begin{proof} To agree \eqref{7.1} and \eqref{7.17}, we need to change the variables in \eqref{7.1} as follows: $x:=p,\ y:=z,\ z:=f,\ p:=\pi$ and $q:=\zeta$. Then the formula \eqref{7.3} takes the form \begin{equation} [F,G]=(F'_p+\pi F'_f)G'_\pi-(G'_p+\pi G'_f)F'_\pi+(F'_z+\zeta F'_f)G'_\zeta-(G'_z+\zeta G'_f)F'_\zeta. \label{7.23} \end{equation} We find the derivatives by differentiating \eqref{7.18}: \begin{equation} \begin{array}{ll} F'_p=-\beta'\pi^2{-}\alpha'(\zeta^2+1),& G'_p=2\gamma'f^2(\varepsilon f^2{+}\gamma)(\zeta^2{+}1)-4\alpha'(f^2{-}\beta)+4\alpha\beta',\\ [5pt] F'_z=0,& G'_z=0,\\ [5pt] F'_f=2f\pi^2,& G'_f=2f(\varepsilon f^2{+}\gamma)^2(\zeta^2{+}1){+}4\varepsilon f^3(\varepsilon f^2{+}\gamma)(\zeta^2{+}1){-}8\alpha f,\\ [5pt] F'_\pi=2(f^2-\beta)\pi,& G'_\pi=0,\\ [5pt] F'_\zeta=-2\alpha\zeta,& G'_\zeta=2f^2(\varepsilon f^2+\gamma)^2\zeta. \end{array} \label{7.24} \end{equation} Since $G'_\pi=F'_z=G'_z=0$, the formula \eqref{7.23} simplifies to the following: \begin{equation} [F,G]=-G'_pF'_\pi-\pi G'_fF'_\pi+\zeta F'_fG'_\zeta-\zeta G'_fF'_\zeta. \label{7.25} \end{equation} Substituting values \eqref{7.24} into \eqref{7.25}, we obtain \begin{equation} \begin{aligned} &\frac{1}{4}[F,G]=-\pi(f^2-\beta)\big(\gamma'f^2(\varepsilon f^2+\gamma)(\zeta^2+1)-2\alpha'(f^2-\beta)+2\alpha\beta'\big)\\ &-\pi^2f(f^2-\beta)\big((\varepsilon f^2+\gamma)^2(\zeta^2+1)+2\varepsilon f^2(\varepsilon f^2+\gamma)(\zeta^2+1)-4\alpha\big)\\ &+\pi^2\zeta^2f^3(\varepsilon f^2+\gamma)^2 +\zeta^2\alpha f\big((\varepsilon f^2+\gamma)^2(\zeta^2+1)+2\varepsilon f^2(\varepsilon f^2+\gamma)(\zeta^2+1)-4\alpha\big). \end{aligned} \label{7.26} \end{equation} The right-hand side of \eqref{7.26} is a 7th degree polynomial in $f$ and the coefficient at $f^7$ is $-\pi^2(2\zeta^2+3)\neq0$. Thus, \eqref{7.17} is not an involutory system. Now, we prove that \eqref{7.17} is a complete system. For this we derive from \eqref{7.17}--\eqref{7.18} \begin{equation} \zeta^2=\frac{4\alpha(f^2-\beta)}{f^2(\varepsilon f^2+\gamma)^2}-1 \label{7.27} \end{equation} and $\pi=\varepsilon\,\frac{\alpha^{1/2}\sqrt{\zeta^2+1}}{\sqrt{f^2-\beta}}$, where $\varepsilon=\pm1$ is the sign of $\pi$. We find from two last equalities \begin{equation} \pi=\frac{2\varepsilon\alpha}{f(\varepsilon f^2+\gamma)}. \label{7.28} \end{equation} Substituting expressions \eqref{7.27}--\eqref{7.28} into \eqref{7.26}, we obtain \begin{equation} \begin{aligned} \frac{1}{16\alpha^2}[F,G]&=-\varepsilon\frac{2(f^2-\beta)}{f(\varepsilon f^2+\gamma)} \big(\frac{\gamma'(f^2-\beta)}{\varepsilon f^2+\gamma}-\frac{\alpha'}{2\alpha}(f^2-\beta)+\frac{\beta'}{2}\big)\\ &-\frac{4\alpha(f^2-\beta)}{f(\varepsilon f^2+\gamma)^2}\big(\frac{f^2-\beta}{f^2} +2\varepsilon \,\frac{f^2-\beta}{\varepsilon f^2+\gamma}-1\big) +f\big(\frac{4\alpha(f^2-\beta)}{f^2(\varepsilon f^2+\gamma)^2}-1\big)\\ &+f\big(\frac{4\alpha(f^2-\beta)}{f^2(\varepsilon f^2+\gamma)^2}-1\big) \big(\frac{f^2-\beta}{f^2}+2\varepsilon \,\frac{f^2-\beta}{\varepsilon f^2+\gamma}-1\big). \end{aligned} \label{7.29} \end{equation} We are interested in the case when $[F,G]=0$. Equating the right-hand side of \eqref{7.29} to zero and multiplying the resulting equality by $f^3(\varepsilon f^2+\gamma)^3$, we arrive to the equation \begin{equation} \begin{aligned} &-2\varepsilon f^2(f^2-\beta)(\varepsilon f^2+\gamma) \big(\gamma'(f^2-\beta)-\frac{\alpha'}{2\alpha}(f^2-\beta)(\varepsilon f^2+\gamma)+\frac{\beta'}{2}(\varepsilon f^2+\gamma)\big)\\ &-4\alpha(f^2-\beta)\big((f^2-\beta)(\varepsilon f^2+\gamma)+2\varepsilon f^2(f^2-\beta)-f^2(\varepsilon f^2+\gamma)\big)\\ &+4\alpha f^2(f^2-\beta)(\varepsilon f^2+\gamma)-f^4(\varepsilon f^2+\gamma)^3\\ &+\big(4\alpha(f^2{-}\beta)-f^2(\varepsilon f^2{+}\gamma)^2\big) \big((f^2{-}\beta)(\varepsilon f^2{+}\gamma)+2\varepsilon f^2(f^2{-}\beta)-f^2(\varepsilon f^2{+}\gamma)\big)=0. \end{aligned} \label{7.30} \end{equation} The left-hand side of the equation \eqref{7.30} is a polynomial of 10th degree in $f$. It is almost unbelievable, but the degree of the polynomial can be decreased to 4. Namely, the equation \eqref{7.30} is equivalent to the following one: \begin{equation} \begin{aligned} -f^2(f^2-\beta)(\varepsilon f^2+\gamma) \big[&\big(3-\frac{\alpha'}{\alpha}\big)f^4 +\big(2\varepsilon\gamma'-\varepsilon\frac{\alpha'}{\alpha}\gamma+\frac{\alpha'}{\alpha}\beta+\beta'+4\varepsilon\gamma\big)f^2\\ &+\big(-2\varepsilon\beta\gamma'+\varepsilon\frac{\alpha'}{\alpha}\beta\gamma+\varepsilon\gamma\beta'-4\alpha+\gamma^2\big)\big]=0. \end{aligned} \label{7.31} \end{equation} Indeed, a simple (though rather cumbersome) calculation verifies that the polynomials on the left-hand sides of \eqref{7.30} and \eqref{7.31} are equal. By \eqref{7.19}, the factor $f^2(f^2-\beta)(\varepsilon f^2+\gamma)$ does not vanish on $U$. Therefore the equation \eqref{7.31} is equivalent to the following: \begin{equation} \big(3-\frac{\alpha'}{\alpha}\big)f^4 +\big(2\varepsilon\gamma'-\varepsilon\frac{\alpha'}{\alpha}\gamma+\frac{\alpha'}{\alpha}\beta+\beta'+4\varepsilon\gamma\big)f^2 +\big(-2\varepsilon\beta\gamma'+\varepsilon\frac{\alpha'}{\alpha}\beta\gamma+\varepsilon\gamma\beta'-4\alpha+\gamma^2\big)=0. \label{7.32} \end{equation} Equating coefficients of the polynomial on the left-hand side of \eqref{7.32} to zero, we arrive to the system of ODEs which is equivalent to \eqref{7.20}--\eqref{7.22}: \begin{equation} \begin{aligned} &\alpha'-3\alpha=0,\\ &\beta'+2\varepsilon\gamma'+3\beta+\varepsilon\gamma=0,\\ &\varepsilon\gamma\beta'-2\varepsilon\beta\gamma'+3\varepsilon\beta\gamma-4\alpha+\gamma^2=0. \end{aligned} \label{7.33} \end{equation} Of course, \eqref{7.33} implies \eqref{7.32} and the determinant is $\beta+\varepsilon\gamma$. This proves the first assertion of Theorem~\ref{Th7.1}. To prove that \eqref{7.32} implies \eqref{7.33}, we need for each $p_0\in(p_1,p_2)$ to have at least three distinct $z_1,z_2,z_3$ such that the values $f^2(p_0,z_1),\,f^2(p_0,z_2),\,f^2(p_0,z_3)$ are pairwise different and \eqref{7.32} holds at $(p_0,z_1),(p_0,z_2)$ and $(p_0,z_3)$. The existence of such $z_1,z_2,z_3$ is guaranteed by the hypothesis of the second assertion of Theorem~\ref{Th7.1}: For each $p_0\in(p_1,p_2)$, there exists $z_0$ such that the system \eqref{7.17} has a solution $f(p,z)$ in a neighborhood of $(p_0,z_0)$ satisfying $f'_z(p_0,z_0)\ne0$. \end{proof} {\bf Remark.} Roughly speaking, Theorem \ref{Th7.1} means that the relations \eqref{7.20}--\eqref{7.22} constitute the consistency condition for the system \eqref{7.17}. Nevertheless, we emphasize that two statements of Theorem \ref{Th7.1} are not exactly converse to each other. For example, the axisymmetric Gavrilov flow \eqref{3.2} (isobaric surfaces are cylinders and particle trajectories are either circles or spiral lines) corresponds to the solution $f(p,z)=2p^{1/2}$ to the system \eqref{7.16} with $(p_1,p_2)=(0,\infty)$ and \begin{equation} \alpha(p)=4a_0,\ \ \beta(p)=4(1-a_0)p,\ \ \gamma(p)=4(a_0-p)\ \ (a_0=\mbox{const},\,0<a_0\le1). \label{7.34} \end{equation} The functions \eqref{7.34} do not satisfy the consistency conditions \eqref{7.20}--\eqref{7.22}. For this solution, $f'_z\equiv0$ holds, and the second statement of Theorem \ref{Th7.1} does not apply. An analog of Lemma~\ref{L7.1} is valid for the system \eqref{6.36}--\eqref{6.37} with minor changes (the variables $(p,z)$ are replaced by $(p,r)$, the function $f$ is replaced by $g$, etc.). The~cor\-responding system of two first order PDEs looks as follows: \begin{equation} \begin{aligned} (r^2-\beta)g'_p{}^2-\alpha(1+g'_r{}^2)&=0,\\ r^2(\gamma-\tau r^2)^2(1+g'_r{}^2)-4\alpha(r^2-\beta)g'_r{}^2&=0, \end{aligned} \label{7.35} \end{equation} where $\alpha(p)>0,\beta(p)\ge0$ and $\gamma(p)$ are the same functions as in \eqref{7.16} and $\tau=\pm1$ is the sign of $g'_p$ that does not vanish. An analog of Theorem \ref{Th7.1} is valid for the system \eqref{7.35} with the same consistency conditions \eqref{7.20}--\eqref{7.22}. \section{Structure of an axisymmetric Gavrilov flow in a neighborhood of a~minimum point of the pressure} \label{sec:structure} In \cite{G}, the existence of a pair $(\tilde u,\tilde p)$ is proved such that (a)~$(1,0)$ is a non-degenerate minimum point of the function $\tilde p=\psi(r,z)\in C^\infty(U)$ and $\psi(1,0)=0$, where $U\subset\{(r,z)\mid r>0\}$ is a neighborhood of $(1,0)$; (b)~$(\tilde u,\tilde p)$ is an axisymmetric Gavrilov flow in $U\setminus\{(1,0)\}$; (c)~the split Bernoulli law for the flow is of the simplest form \begin{equation} \vert\tilde u\vert^2=3\tilde p. \label{8.1} \end{equation} The flow $(\tilde u,\tilde p)$ does not satisfy the normalization condition \eqref{6.28}. To apply our equations, we must replace $(\tilde u,\tilde p)$ with an equivalent Gavrilov flow $(u,p)$ such that $\vert u\vert^2=1$. By the definition \eqref{1.4} of equivalent Gavrilov flows, $u=\varphi(\tilde p)\tilde u, \ \nabla p=\varphi^2(\tilde p)\nabla\tilde p$ with some non-vanishing function $\varphi(\tilde p)$. As is seen from \eqref{8.1}, $\varphi^2(\tilde p)=1/\vert\tilde u\vert^2=1/3\,\tilde p$.~Thus \begin{equation} p(r,z)=\frac{1}{3}\ln\psi(r,z)+C\quad(C=\mbox{const}), \label{8.2} \end{equation} Hence, the pressure function gets a logarithmic singularity at the point $(1,0)$ after normalization. Comparing \eqref{7.20} and \eqref{8.2}, we conclude that \begin{equation} \alpha=c\,\psi\quad(c=\mbox{const}>0). \label{8.3} \end{equation} Since $p\to-\infty$ as $(r,z)\to(1,0)$, it is natural to choose initial conditions for the system \eqref{7.21}--\eqref{7.22} at $p=-\infty$. The conditions are \begin{equation} \beta(-\infty)=\lim\limits_{p\to-\infty}\beta(p)=1/3,\quad \gamma(-\infty)=\lim\limits_{p\to-\infty}\gamma(p)=-1. \label{8.4} \end{equation} Indeed, for $p$ close to $-\infty$, $\Gamma_p$ is a ``small" closed curve around $(1,0)$. The tangent line to $\Gamma_p$ is vertical at some point $(r_1,z_1)=\big(r_1(p),z_1(p)\big)\in \Gamma_p$. In a neighborhood of $(r_1,z_1)$, the curve $\Gamma_p$ is the graph of a function $r=f(p,z)$ solving the system \eqref{7.16} and satisfying $f'_z(p,z_1)=0$. Setting $z=z_1$ in the second equation of \eqref{7.16}, we get \begin{equation} f^2(p,z_1)\big(\varepsilon f^2(p,z_1)+\gamma(p)\big)^2-4\alpha(p)\big(f^2(p,z_1)-\beta(p)\big)=0. \label{8.5} \end{equation} In view of our assumption $f^2(p,z)-\beta(p)>0$, see \eqref{7.7}, the factor $\big(f^2(p,z_1)-\beta(p)\big)$ remains bounded when $p\to-\infty$. Also $\alpha(p)=\alpha_0e^{3p}\to0$ as $p\to-\infty$. Thus, the second term on the left-hand side of \eqref{8.5} runs to 0 as $p\to-\infty$. The same is true for the first term. Taking into account that $f^2(p,z_1)\to1$ as $p\to-\infty$, we obtain $\gamma(-\infty)=-\varepsilon$. Since $\varepsilon=\pm1$ is the sign of $f'_p$ in \eqref{8.5}, by a similar analysis of the system \eqref{7.35}, we demonstrate that $\varepsilon=1$ in our setting. This proves the second equality in \eqref{8.4}. The~first equality in \eqref{8.4} is proved similarly. The ``Cauchy problem'' \eqref{7.21}--\eqref{7.22}, \eqref{8.4} (with $\varepsilon=1$) is easily solved in series $ \beta=\frac{1}{3}+\sum\nolimits_{k=1}^\infty\beta_k\alpha^k$ and $ \gamma=\frac{1}{3}+\sum\nolimits_{k=1}^\infty\gamma_k\alpha^k. $ Equations \eqref{7.21}--\eqref{7.22} imply some recurrent relations that allow us to compute all coefficients. In particular, \begin{eqnarray} \beta=\frac{1}{3}-\frac{7}{6}\alpha{+}\frac{13}{72}\alpha^2 {-}\frac{133}{1728}\alpha^3{+}\frac{575}{13824}\alpha^4 {-}\frac{2077}{82944}\alpha^5{+}\frac{259}{16128}\alpha^6+\dots, \label{8.6} \\ \gamma=-1+\alpha-\frac{1}{8}\alpha^2 +\frac{7}{144}\alpha^3-\frac{115}{4608}\alpha^4 +\frac{67}{4608}\alpha^5-\frac{7}{768}\alpha^6+\dots. \label{8.7} \end{eqnarray} The system \eqref{7.16} can be equivalently written in terms of the function $\psi(r,z)$. Indeed, as is seen from \eqref{8.2}, a solution $f(p,z)$ to the system \eqref{7.16} is related to $\psi$ by $ \frac{1}{3}\ln\psi(f(p,z),z)+C=p. $ Starting with this equation, we derive from \eqref{7.16} the system \begin{equation} \begin{aligned} 2c\psi'_r-3r(r^2+\gamma)&=0,\\ c^2\psi'_z{}^2-9c(r^2-\beta)\psi+\frac{9}{4}r^2(r^2+\gamma)^2&=0, \end{aligned} \label{8.8} \end{equation} where $c$ is the constant from \eqref{8.3}. By the change $\psi=\frac{1}{c}\tilde\psi$ of the unknown function, \eqref{8.8} is transformed to the same system with $c=1$. Therefore we can assume $c=1$ without lost of generality, i.e., \begin{equation} \begin{aligned} 2\psi'_r-3r(r^2+\gamma)&=0,\\ \psi'_z{}^2-9(r^2-\beta)\psi+\frac{9}{4}r^2(r^2+\gamma)^2&=0. \end{aligned} \label{8.9} \end{equation} The equality \eqref{8.3} becomes $\alpha=\psi$. The function $\psi(r,z)$ is defined and smooth in a neighborhood of the point $(1,0)$ and satisfies $\psi(1,0)=0$. Let us show that $\psi(r,z)$ is an even function of $z$. Indeed, the second equation of the system \eqref{8.9} gives $ \psi'_z{}^2(1,0)+\frac{9}{4}r^2(1+\gamma(-\infty))^2=0. $ Since $\gamma(-\infty)=-1$ by \eqref{8.4}, we obtain \begin{equation} \psi'_z(1,0)=0. \label{8.11} \end{equation} Then we differentiate the first equation of the system \eqref{8.9} with respect to $z$ \begin{equation} 2\psi''_{rz}-3r\gamma'_z=0. \label{8.12} \end{equation} Since $\gamma$ depends on $p$ only, $ \gamma'_z=\gamma' p'_z, $ where $\gamma'=d\gamma/dp$. Together with \eqref{8.2}, this gives $\gamma'_z=\frac{\gamma'\psi'_z}{3\psi}$. Substituting this expression into \eqref{8.12} and setting $z=0$, we arrive to the linear first order ODE for the function $\psi'_z(r,0)$: \begin{equation} 2\frac{d\psi'_z(r,0)}{d r}-3\frac{r\gamma'(p(r,0))}{3\psi(r,0)}\,\psi'_z(r,0)=0. \label{8.13} \end{equation} The coefficient of the equation $\frac{r\gamma'(p(r,0))}{3\psi(r,0)}$ is a bounded smooth function in a neighborhood of $r=1$ as is seen from \eqref{8.7}. Together with the initial condition \eqref{8.11}, the equation \eqref{8.13} implies $\psi'_z(r,0)=0$. The system \eqref{8.9} is invariant under the change $z\to-z$. Therefore the equality $\psi'_z(r,0)=0$ implies that $\psi(r,z)$ is an even function of $z$. The system \eqref{8.9} allows us to compute term-by-term all Taylor coefficients of the function $\psi(r,z)$ at the point $(1,0)$. In particular, \begin{equation} \begin{aligned} \psi(r,z)&=\frac{3}{2}(r-1)^2+\frac{3}{2}z^2+\frac{9}{4}(r-1)^3+\frac{9}{4}(r-1)z^2+\frac{57}{32}(r-1)^4\\ &+\frac{45}{16}(r-1)^2z^2 +\frac{33}{32}z^4+\frac{9}{8}(r-1)^5+\frac{9}{4}(r-1)^3z^2+\frac{9}{4}(r-1)z^4+\dots \end{aligned} \label{8.14} \end{equation} Second and 3d order terms on the right-hand side of \eqref{8.14} are easily derived from \eqref{8.9}, and we used Maple for computing 4th and 5th order terms. Recall that physical coordinates $(u_r,u_z,u_\theta)$ of the velocity vector field $u$ are defined by \eqref{6.2}. The function $u_\theta$ is called {\it the swirl}. As follows from \eqref{6.3} and \eqref{6.32}, \begin{equation} (u_r^2+u_z^2)(r,z)=1-\frac{\beta(p)}{r^2},\quad u_\theta(r,z)=\frac{\sqrt{\beta(p)}}{r}. \label{8.15} \end{equation} Together with \eqref{8.6}, these formulas allow us to compute the swirl and two-dimensional vector field $u_re_r+u_ze_z$ tangent to the curves $\Gamma_p$. \section{Two examples of axisymmetric Gavrilov flows} \label{sec:2ex} Here, we discuss the numerical method for constructing axisymmetric Gavrilov flows on the base of the system \eqref{7.16}. For the initial condition $f(p_0,z_0)=f_0$, we can assume without lost of generality that $p_0=z_0=0$ since the system \eqref{7.16} is invariant under the change $p\to p+\mbox{const}$ and $z\to z+\mbox{const}$. Thus, the initial condition for $f$ is \begin{equation} f(0,0)=f_0. \label{9.1} \end{equation} First, we will find the functions $\alpha(p),\beta(p),\gamma(p)$. The function $\alpha$ is given explicitly by \eqref{7.20} with an arbitrary constant $\alpha_0=\alpha(0)>0$. The functions $\beta(p)\ge0$ and $\gamma(p)$ solve the system \eqref{7.21}--\eqref{7.22} supplied with the initial conditions \begin{equation} \beta(0)=\beta_0,\quad \gamma(0)=\gamma_0. \label{9.2} \end{equation} In particular, $\beta_0\ge0$. For being able to write the system \eqref{7.21}--\eqref{7.22} in the form $\beta'=B(\beta,\gamma),\ \gamma'=\Gamma(\beta,\gamma)$, we have to impose the restrictions \begin{equation} \beta\ge0,\quad \beta+\varepsilon\gamma\neq0. \label{9.3} \end{equation} In particular, $\beta_0$ and $\gamma_0$ must satisfy these inequalities. Let $(p_1,p_2)$ be the {maximal interval} such that the solution to the Cauchy problem \eqref{7.21}--\eqref{7.22}, \eqref{9.2} exists on $(p_1,p_2)$ and satisfies \eqref{9.3}. Here, $p_1=p_1(\alpha_0,\beta_0,\gamma_0)<0<p_2=p_2(\alpha_0,\beta_0,\gamma_0)$. By Theorem \ref{Th7.1}, the system \eqref{7.16} is complete on $U\times{\mathbb R}^2$, where $U\subset{\mathbb R}^3$ is defined by \eqref{7.19}. General theory \cite[Chapter~1, Section~14]{Ka} guarantees the existence and uniqueness of a solution to the IVP \eqref{7.16}, \eqref{9.1} for any $(0,0,f_0)\in U$ at least in some neighborhood of $(p_0,z_0)=(0,0)$. Thus, a solution $f(p,z)$ to the IVP \eqref{7.16}, \eqref{9.1} exists in some neighborhood of $(0,0)$ and is uniquely determined by 5 constants $(\alpha_0,\beta_0,\gamma_0,f_0,\varepsilon)$ chosen so~that \begin{equation} \begin{aligned} &\alpha_0>0,\quad \beta_0\ge0,\quad \beta_0+\varepsilon\gamma_0\neq0,\quad f_0>0,\quad\varepsilon=\pm1,\\ &f_0>\beta_0^{1/2},\quad 0<\varepsilon f_0^2+\gamma_0<2\alpha_0^{1/2}f_0^{-1}\sqrt{f_0^2-\beta_0}. \end{aligned} \label{9.4} \end{equation} The inequalities on the second line of \eqref{9.4} come from \eqref{7.19}. Let us denote this unique solution by $f(p,z;\alpha_0,\beta_0,\gamma_0,f_0,\varepsilon)$. In general, $f(p,z;\alpha_0,\beta_0,\gamma_0,f_0,\varepsilon)$ is a {\it local solution}, i.e., is defined in some neighborhood $U(\alpha_0,\beta_0,\gamma_0,f_0,\varepsilon)\subset{\mathbb R}^2$ of the point $(0,0)$. Nevertheless, for some values of the parameters, it can happen that $(p'_1,p'_2)\times\mathbb R\subset U(\alpha_0,\beta_0,\gamma_0,f_0,\varepsilon)$ for some $-\infty\le p'_1<0<p'_2\le\infty$, in such a case we speak on a {\it global solution} defined on $(p'_1,p'_2)\times\mathbb R$. Global solutions are of particular interest. Unfortunately, so far we have neither necessary nor sufficient conditions on the parameters $(\alpha_0,\beta_0,\gamma_0,f_0,\varepsilon)$ for the existence of a global solution. Everything said above in this section is valid for the numerical method based on the system \eqref{7.35}. Global solutions most often appear due to periodicity using the following \begin{lemma}\label{L9.1} Let a solution $f(p,z)$ of the system \eqref{7.16} be defined on a rectangle $(p_1,p_2)\times(z_1,z_2)$ and satisfy \eqref{7.7}. Assume the existence of $(p_0,z_0)\in(p_1,p_2)\times(z_1,z_2)$ such that $f'_z(p_0,z_0)=0$ and $f''_{zz}(p_0,z_0)\ne0$. Then, at least in some neighborhood of $p_0$, the solution is symmetric with respect to $z_0$, i.e., \begin{equation} f(p,z)=f(p,-z+2z_0)\quad\mbox{\rm for}\ p\in(p_0-\delta,p_0+\delta)\ \mbox{\rm with some}\ \delta>0. \label{9.5} \end{equation} Hence, $f$ can be extended to a solution defined in $(p_0-\delta,p_0+\delta)\times(z_0-\Delta,z_0+\Delta)$ such that $(z_1,z_2)\subset(z_0-\Delta,z_0+\Delta)$. \end{lemma} \begin{proof} By the implicit function theorem, there exists a smooth function $z=\zeta(p)$ defined for $p\in(p_0-\delta,p_0+\delta)$ with some $\delta>0$ such that $\zeta(p_0)=z_0$ and \begin{equation} f'_z(p,\zeta(p))=0. \label{9.6} \end{equation} We are going to prove that $\zeta(p)$ is actually a constant function. To this end we differentiate~\eqref{9.6}: $f''_{pz}(p,\zeta(p))+f''_{zz}(p,\zeta(p))\,\zeta'(p)=0$. By choosing a smaller $\delta$, we can assume that $f''_{zz}(p,\zeta(p))\neq0$. Thus, to prove the equality $\zeta'=0$, we have to demonstrate that $f''_{pz}(p,\zeta(p))=0$. To this end we differentiate the first equation of the system \eqref{7.16} with respect to $z$ (recall that $\alpha$ and $\beta$ are independent of $z$) $ (f^2-\beta)f'_pf''_{pz}+f(f'_p)^2f'_z-\alpha f'_zf''_{zz}=0. $ Setting $z=\zeta(p)$ here and using \eqref{9.6}, we obtain $(f^2-\beta)f'_pf''_{pz}\vert_{z=\zeta(p)}=0$. Since $f'_p\neq0$ and $f^2-\beta>0$, this implies $f''_{pz}(p,\zeta(p))=0$. We have thus proved \begin{equation} f'_z(p,z_0)=0\quad\mbox{for}\ p\in(p_0-\delta,p_0+\delta). \label{9.8} \end{equation} The system \eqref{7.16} is invariant under the change $z\to 2z_0-z$. Thus \eqref{9.8} implies~\eqref{9.5}. \end{proof} Now, under hypotheses of Lemma \ref{L9.1}, assume the existence of a second point $(p'_0,z'_0)\in(p_1,p_2)\times(z_1,z_2)\ (z_0\neq z'_0)$ such that $f'_z(p'_0,z'_0)=0$ and $f''_{zz}(p'_0,z'_0)\neq0$. An analog of \eqref{9.5} holds for $z'_0$ with some $\delta'>0$. Assume additionally that $$ ({\tilde p}_1,{\tilde p}_2)=(p_0-\delta,p_0+\delta)\cap(p'_0-\delta',p'_0+\delta')\neq\emptyset. $$ Using symmetries with respect to $z_0$ and $z'_0$, we extend $f(p,z)$ to a global $z$-periodic solution defined on $({\tilde p}_1,{\tilde p}_2)\times\mathbb R$. The period is equal to $\vert z_0-z'_0\vert$ if $p_0=p'_0$, otherwise the period is equal to $2\vert z_0-z'_0\vert$. \smallskip An example of $z$-periodic Gavrilov flow is presented on Figure~\ref{fig:F1}. The solution $ f(p,z)=f(p,z;\alpha_0,\beta_0,\gamma_0,f_0,\varepsilon) $ to the IVP \eqref{7.16}, \eqref{9.1} was computed for $\alpha_0=1,\beta_0=0.01, \gamma_0=0.5, f_0=0.97,\varepsilon=1$. The solution to the Cauchy problem \eqref{7.21}--\eqref{7.22}, \eqref{9.2} exists for $-0.07<p<7.48$. For $0<p< p_c\approx 0.25$, the flow turns out to be $z$-periodic with the period approximately equal to $2.8$. Graphs $\Gamma_i$ of functions $r=f(p_i,\,z)$ for $p_i=0.02\,i$ and $i=0,\dots,8$ are drawn on the left-hand side of Figure~\ref{fig:F1}. Each curve $\Gamma_i$ is the generatrix of the isobaric surface $M_{p_i}$. The velocity vector field $u$ is drawn on the right picture for $\Gamma_0$ and $\Gamma_8$. We used \eqref{8.15} for computing physical coordinates of $u$. The two-dimensional vector field $u_re_r+u_ze_z$ is tangent to $\Gamma_0$ and $\Gamma_8$, and the vector field $u_\theta e_\theta$ (the swirl) is orthogonal to the plane of the picture. The latter vector field is drawn by vectors orthogonal to $\Gamma_0$ and $\Gamma_8$ in order to avoid 3D pictures. In order to make a nice picture, both fields $u_re_r+u_ze_z$ and $u_\theta e_\theta$ are drawn in the scale $1:4$, i.e., $\vert u\vert=0.25$ on the picture. But we remember that actually~$\vert u\vert=1$. As is seen from the picture, for $p$ close to the critical value $p_c$, the flow is close to the $z$-independent solution \eqref{3.2}: isobaric surfaces are close to circular cylinders and particle trajectories are close to spiral lines intersecting parallels approximately at the angle of $\pi/4$. The hypothesis $f''_{zz}(p_0,z_0)\neq0$ of Lemma~\ref{L9.1} is violated at $p_0=p_c$. Therefore our flow does not need to be periodic with the same period $\approx2.8$ for $p>p_c$. \begin{figure} \noindent\includegraphics[scale=0.46] {GAFA2d.eps} \caption{\small Periodic axisymmetric Gavrilov flow. } \label{fig:F1} \end{figure} \begin{figure} \noindent\includegraphics[scale=0.46] {GAFA1.eps} \caption{\small Gavrilov flow in a neighborhood of a minimum point of the pressure. } \label{fig:F2} \end{figure} The Gavrilov flow discussed in Section~\ref{sec:structure} is illustrated on Figure~\ref{fig:F2}. In our calculations we use the 5th order segment of the Taylor series of the function $\psi$, i.e., we ignore the remainder denotes by dots on the right-hand side of \eqref{8.14}. Six isolines $\Gamma_i=\{(r,z): \psi(r,z)=\psi_i=0.04\,i\}\ (i=1,\dots,6)$ are drawn on the left-hand side of Figure~\ref{fig:F2}. Each of the curves $\Gamma_1,\dots, {\Gamma}_5$ consists of two connected components while $\Gamma_6$ has one component. Besides the minimum point $(1,0)$, the function $\psi(r,z)$ has the saddle point at $(r,z)=(1/3,0)$. The same curves $\Gamma_i$ are isolines of the pressure function, see \eqref{8.2}. We set $C=-\frac{1}{3}\ln\psi_1$, thus the formula \eqref{8.2} becomes $p(r,z)=\frac{1}{3}\ln\frac{\psi(r,z)}{\psi_1}$. Each curve $\Gamma_i$ is the generatrix of the isobaric surface $M_{p_i}$, where $p_i=\frac{1}{3}\ln\frac{\psi_i}{\psi_1}$. In our case, $p_i=\frac{1}{3}\ln i$. The~velocity vector field $u$ is drawn on the right picture for $\Gamma_3$ and $\Gamma_5$. \section{Some open questions} \label{sec:problems} In our opinion, the main open question is: are there Gavrilov flows on ${\mathbb R}^3$ which are not axisymmetric? From an analytical point of view, this is a question on the consistency conditions of the system \eqref{5.12}--\eqref{5.15}. The most expected answer to the question is ``yes''. However, the question is not easy because of the following. For an axisymmetric Gavrilov flow, the constants $c$ and $C$ in the Bernoulli law \eqref{2.1} can be expressed through each other. In other words, all particles living on an isobaric surface $M_p$ move with the same speed in the case of an axisymmetric Gavrilov flow. Most likely, this statement is not true for a general (not axisymmetric) Gavrilov flow, at least we cannot prove it by local reasoning. But maybe simple global arguments will help answer the question. For instance, if there exists a particle trajectory dense in $M_p$, then $\vert u\vert=\mbox{const}$ on $M_p$. Except of Sections~\ref{sec:ex} and \ref{sec:section}, we studied Gavrilov flows locally in a neighborhood of a regular point. The most interesting questions relate to Gavrilov flows with regular compact isobaric hypersurfaces $M_p\subset{\mathbb R}^n$. As mentioned in Introduction, application of the Gavrilov localization to such flow gives a compactly supported Gavrilov flow on the whole of ${\mathbb R}^n$. Since a compact regular isobaric hypersurface $M_p$ is endowed with a non-vanishing tangent vector field $u$, the Euler characteristic of $M_p$ is equal to zero. In~the most important 3D-case this means that $M_p$ is diffeomorphic to the two-dimensional torus~${\mathbb T}^2$. The restriction of $u$ onto $M_p$ is a non-vanishing geodesic vector field. There are Riemannian metrics on ${\mathbb T}^2$ admitting a non-vanishing geodesic vector field, the corresponding example can be found in the class of so called double-twisted products \cite{Go}. But we are interested in metrics on ${\mathbb T}^2$ induced from the Euclidean metric by an embedding $i:{\mathbb T}^2\subset{\mathbb R}^3$. Apart from surfaces of revolution, we do not know any example related to the following~problem. \begin{problem} Classify triples $(i,u,\lambda)$, where $i:{\mathbb T}^2\subset{\mathbb R}^3$ is an embedding of the torus, $u$ is a nonzero geodesic vector field on ${\mathbb T}^2$ endowed with the Riemannian metric induced from the Euclidean metric of ${\mathbb R}^3$ by embedding $i$, and $\lambda>0$ is a smooth function on ${\mathbb T}^2$ satisfying the equations $($obtained from \eqref{2.3} and \eqref{2.10} by setting $\lambda=\vert\nabla p\,\vert\,)$ \begin{equation} \mbox{\rm div}\,u=u(\log\lambda),\quad II(u,u)=-\lambda, \label{10.1} \end{equation} where $II$ is the second quadratic form of $\,{\mathbb T}^2$. \end{problem} Equations \eqref{10.1} are obtained from \eqref{2.3} and \eqref{2.10} by setting $\lambda=\vert\nabla p\,\vert$. In this paper, we did not discuss the behavior of the Gavrilov flow near a critical point of the pressure. Such a discussion could be of great interest. For instance, it makes sense to study the Gavrilov flow $(u,p)$ with the Morse function $p$, i.e., all critical points of $p$ are non-degenerate. For such a flow, $M_p$ is still a regular hypersurface for a regular value of the pressure; but $M_p$ undergoes a Morse surgery when $p$ changes near a critical value of the pressure. Which Morse surgeries are compatible with the Euler equations?

574d7a183a08f244fa4fa9a6ef9cf72bbc9434b2

\section{INTRODUCTION} The ability to detect dynamic and stationary obstacles (e.g., cars, trucks, pedestrians, bicycles, hazards) is critical for autonomous vehicles. This is particularly important in semi-urban and urban settings characterized by complex scenes with large amounts of occlusion and varieties of shapes. Previous perception methods rely heavily on utilizing cameras~\cite{liftsplatshoot}~\cite{cameraCenternet}\cite{camera10d} or LiDARs~\cite{mvLidarNet}~\cite{hendy2020fishing}~\cite{fastFurious}~\cite{2018pixor} to detect obstacles. These methods have a number of drawbacks: they are unreliable in cases of heavy occlusions, the sensors may be prohibitively expensive, they can be unreliable in adverse weather conditions~\cite{lidarFog} or at night. Traditional RADAR based obstacle detection methods work well in detecting moving objects that have good reflection properties, but often struggle when estimating object dimensions and orientations and often completely fail in detecting stationary objects or objects with poor RADAR reflectivity. In this paper, we present a deep neural network (DNN) that detects moving and stationary obstacles, computes their orientation and size, and detects drivable free space from RADAR data alone. We do this in top-down bird's-eye view (BEV) for highway and urban scenarios while using readily available automotive RADARs. Our method relies on RADAR peak detections alone~\cite{classicalRadar2dFFT}~\cite{classicalRadarDopplerFFT} since automotive RADAR firmware provides only this data. In contrast, other approaches~\cite{azimuthRangeTensor}~\cite{probOrientedRadar}, require expensive Fast Fourier Transformation operations on the raw RADAR data cube cross-sections that are not available in most commercial automotive sensors. Our deep learning approach is able to accurately distinguish between stationary obstacles, such as cars, versus stationary background noise. This is important when navigating in a cluttered urban environment. In addition, our approach allows us to regress the dimensions and orientations of these obstacles, which classical methods cannot provide. Our DNN can even detect obstacles with poor reflectivity like pedestrians. Finally, our method provides an occupancy probability map to mark unclassified obstacles and regresses drivable free space. We have tested our NVRadarNet DNN in real-world autonomous driving on our vehicles running NVIDIA DRIVE AGX's embedded GPU. Our DNN runs \emph{faster than real-time} at \textbf{1.5~ms} end-to-end and provides sufficient time for the planner to react safely. Our contributions are as follows: \begin{itemize} \item NVRadarNet: A first of its kind multi-class deep neural network that detects dynamic and stationary obstacles end-to-end without post-processing in a top-down bird's-eye view (BEV) using only peak detections coming from automotive RADARs; \item A novel semi-supervised drivable free space detection method using only RADAR peak detections; \item A DNN architecture that runs \emph{faster than real-time} at \textbf{1.5~ms} end-to-end on an embedded GPU. \end{itemize} \section{PREVIOUS WORK} \textbf{Obstacle Detection.} Fast and efficient obstacle perception is a core component of a self-driving vehicle. Automotive RADAR sensors provide a cost-efficient way of obtaining rich 3D positioning and velocity information and are widely available on most modern cars. Several recent papers examined the use of the dense RADAR data cubes in order to perform obstacle detection~\cite{azimuthRangeTensor}~\cite{probOrientedRadar}. However, these methods require high input/output bandwidth to obtain such rich data. This makes them impractical for real-world autonomous vehicles. Thus, most classical methods in automotive RADAR applications utilize post-processed peak detections from the data cube in order to perform classification and occupancy grid detection~\cite{radarClassicalBerha}~\cite{radarClassicalEnsamble}~\cite{occupancySemanticGridBMW}. Others realized that the RADAR peak detections can be viewed as a sparse 3D point cloud and therefore can be used in sensor fusion along with 3D LiDAR points in approaches similar to LiDAR DNNs~\cite{mvLidarNet}~\cite{4DNet}~\cite{radarNet2020}~\cite{hendy2020fishing}~\cite{liraNet}. There were attempts to enhance camera 3D obstacle detection by fusing it with RADAR as well~\cite{centerNetCameraRadarFusion}. \textbf{Free Space Detection.} RADAR-based drivable free space estimation has been attempted in~\cite{ism} and~\cite{occupancyISM2019}. Our DNN performs multi-class detection of dynamic and static obstacles together with the segmentation of drivable free space by using RADAR peak detections alone. Our DNN architecture is lightweight and runs \emph{faster than real-time} at \textbf{1.5~ms} end-to-end on an embedded GPU (on NVIDIA DRIVE AGX). It has been proven to be robust in real-world driving and was tested on over $10000$~km of highway and urban roads as part of our autonomous stack. To date, we are not aware of any RADAR peak detections only DNN that can perform all of these tasks and can run efficiently on autonomous vehicles. \section{METHOD} \subsection{Input Generation} The input to our network is a top-down BEV orthographic projection of accumulated RADAR detections peaks around our ego-vehicle, which is placed at the center of this top-down bird's-eye view (BEV) with its front facing right. To compute this input, we first accumulate RADAR peak detections across all RADAR sensors on our vehicle (8 radars covering 360 degrees field of view) and then transform them to our ego-vehicle rig coordinate system. We also accumulate these peak detections temporally over $0.5$~seconds in order to increase the density of the signal. Each data point gets a relative timestamp to indicate its age, similar to~\cite{4DNet}. Next, we perform ego-motion compensation for the accumulated detections to the latest known vehicle position. We propagate the older points using the known ego-motion of our vehicle to estimate where they will be at at the time of DNN inference (current time). Next, we project each accumulated detection to a top-down BEV grid using the desired space quantization to create an input tensor for our DNN. We set our input resolution to 800$\times$800 pixels with $\pm$ $100$~m range in each direction, resulting in $25$~cm per pixel resolution. Each valid BEV pixel (with data) gets a set of features in its depth channel computed by averaging the raw signal features of the RADAR detections that land in that pixel. Our final input for time $t$ is a tensor $I_t \in \R^{h \times w \times 5}$ where $h=800, w=800$ are height and width of a top-down view. The 5 RADAR features in the depth channel are the averages of: Doppler, elevation angle, RADAR cross section (\textit{RCS}), azimuth angle and the relative detection timestamp. We normalize these values to a $[0,1]$ range for training stability using maximum and minimum values provided by the hardware specifications. The resulting tensor is used as input to our network. \subsection{Label Propagation}\label{label_propogation} We use LiDAR-based human-annotated bounding box labels as the ground truth for training our RADAR DNN. These labels are created for LiDAR data for the same scene on which we train our RADAR DNN. Given how sparse the RADAR signal is, it is practically impossible for humans to distinguish vehicles using RADAR points alone even in top-down BEV view. Hence, we rely on LiDAR to label training data. We capture both LiDAR and RADAR data at different frequencies and select the data closest in time for processing. We then create a top-down BEV projection of the LiDAR scene for humans to annotate objects with bounding box labels and free space with polylines. For each labeled LiDAR BEV frame, we compute the closest accumulated RADAR BEV image via the pre-processing method described above and then transfer the labels to the RADAR top-down view. We further clean up the the ground truth by removing any vehicle labels that contain fewer than $4$ RADAR detections, which empirically demonstrated to increase the network accuracy. Finally, we remove any detections with an \textit{RCS} below $-40$~dBm as we empirically determined that they introduce more noise than signal. An example can be seen in Fig.~\ref{fig:lidar_to_radar}. \begin{figure} \centering \begin{subfigure}{1\linewidth} \includegraphics[width=1.0\linewidth]{radarnet_images/lidar_to_radar.png} \end{subfigure} \caption{Propagating bounding box labels for cars from LiDAR domain to RADAR domain.} \label{fig:lidar_to_radar} \vspace{-2mm} \end{figure} \subsection{Free Space Label Generation} \begin{figure} \centering \begin{subfigure}{1\linewidth} \includegraphics[width=1.0\linewidth]{radarnet_images/radarnet_freespace_labels.png} \end{subfigure} \caption{Visual representation of the free space target: observed and free in black, observed and occupied in white, unobserved in light gray and partially observed in dark gray.} \label{fig:freespace_labels} \vspace{-6mm} \end{figure} The free space target is generated by using the raw LiDAR point cloud. First, we pre-process the point cloud by identifying and removing the points belonging to the drivable surface itself by using surface slope angle estimation of adjacent LiDAR scan lines. We then overlay manually obtained LiDAR free space labels to further clean up this estimate. Next, a set of rays is traced from the ego-vehicle's origin in all angular directions, enabling us to reason about which regions are: \begin{itemize} \item Observed and free. \item Observed and occupied. \item Unobserved. \item Partially observed. \end{itemize} Finally, we overlay our existing 3D obstacle labels on top of the automatically derived occupancy. We explicitly mark obstacles as observed and occupied. See Fig.~\ref{fig:freespace_labels}. \subsection{Dataset}\label{dataset} Our model is trained on a diverse internal dataset with over 300k training frames and over 70k validation frames sampled from hundreds of hours of driving in several geographic regions. The dataset includes a combination of urban and highway data and contains synchronized LiDAR, RADAR and IMU readings. The labels are human annotated and include vehicles, cyclists, pedestrians and drivable free space. \subsection{Network Architecture} \begin{figure} \centering \begin{subfigure}{1\linewidth} \includegraphics[width=1.0\linewidth]{radarnet_images/freespace_inferred_head.png} \end{subfigure} \caption{Inferred dense occupancy probability map showing probabilities from low in dark to high probability in gradients from red to yellow (highest).} \label{fig:occupancy_map} \vspace{-4mm} \end{figure} \begin{figure*} \centering \begin{subfigure}{0.7\linewidth} \includegraphics[width=1\linewidth]{radarnet_images/NVRadarNet_arch.png} \end{subfigure} \caption{Network architecture. Our network uses CNN for the encoder and decoder with skip connections. The network has three heads: a classification head (produces detection probabilities), a shape regression head (produces bounding box parameters) and a free space segmentation head.} \label{fig:network_diagram} \end{figure*} \begin{table} \caption{Network architecture for NVRadarNet} \begin{center} \begin{minipage}[t]{.95\linewidth} \resizebox{\linewidth}{!}{ \begin{tabular}{l|l|l|l} \textbf{Layer} & \textbf{Layer description} & \textbf{Input} & \textbf{Output dimensions} \\ \hline \hline \multicolumn{3}{c}{\textbf{Inputs:}} \\ input & \emph{Input RADAR data} & -- & $ 5 \times 800 \times 800$ \\ \hline \multicolumn{3}{c}{\textbf{Encoder:}} \\ \hline 1 & conv $(7 \times 7), ReLU $ & input & $ 64 \times 400 \times 400 $ \\ 2a & conv $(3 \times 3), ReLU $ & 1 & $ 64 \times 200 \times 200 $ \\ 2b & conv $(3 \times 3), ReLU $ & 2a & $ 64 \times 200 \times 200 $ \\ 3a & conv $(3 \times 3), ReLU $ & 2b & $ 64 \times 200 \times 200 $ \\ 3b & conv $(3 \times 3), ReLU $ & 3a & $ 64 \times 200 \times 200 $ \\ 4a & conv $(3 \times 3), ReLU $ & 3b & $ 128 \times 100 \times 100 $ \\ 4b & conv $(3 \times 3), ReLU $ & 4a & $ 128 \times 100 \times 100 $ \\ 4c & conv $(3 \times 3), ReLU $ & 4b & $ 128 \times 100 \times 100 $ \\ 4d & conv $(3 \times 3), ReLU $ & 4c & $ 128 \times 100 \times 100 $ \\ 5a & conv $(3 \times 3), ReLU $ & 4d & $ 256 \times 50 \times 50 $ \\ 5b & conv $(3 \times 3), ReLU $ & 5a & $ 256 \times 50 \times 50 $ \\ 5c & conv $(3 \times 3), ReLU $ & 5b & $ 256 \times 50 \times 50 $ \\ 5d & conv $(3 \times 3), ReLU $ & 5c & $ 256 \times 50 \times 50 $ \\ 6a & conv $(3 \times 3), ReLU $ & 5d & $ 512 \times 50 \times 50 $ \\ 6b & conv $(3 \times 3), ReLU $ & 6a & $ 512 \times 50 \times 50 $ \\ 6c & conv $(3 \times 3), ReLU $ & 6b & $ 512 \times 50 \times 50 $ \\ 6d & conv $(3 \times 3), ReLU $ & 6c & $ 512 \times 50 \times 50 $ \\ \hline \multicolumn{3}{c}{\textbf{Decoder:}} \\ \hline freespace\textunderscore output & deconv $(4 \times 4), {ReLU} $ & 6d & $ 2 \times 400 \times 400 $ \\ \hline regression\textunderscore output & deconv $(4 \times 4), {ReLU} $ & 6d & $ 6 \times 200 \times 200 $ \\ \hline class\textunderscore output & deconv $(4 \times 4), {ReLU} $ & 6d & $ 4 \times 200 \times 200 $ \end{tabular} }\end{minipage} \end{center} \label{tab:network_arch} \vspace{-4mm} \end{table} We use a DNN architecture similar to Feature Pyramid Network~\cite{fpn}. Our DNN consists of encoder and decoder components and several heads for predicting different outputs. See Fig.~\ref{fig:network_diagram} for high-level structure and Table~\ref{tab:network_arch} for details. Our encoder starts with a \textit{2D convolutional layer} with $64$ filters, stride 2 and $7\times7$ kernels. It is followed by $4$ blocks of $4$ layers each, where each block increases the number of filters by two, while dividing the resolution in half. Each layer in the block contains a \textit{2D convolution} with \textit{batch normalization} and \textit{ReLU activation}. The decoder consists of one \textit{transposed 2D convolution} with stride $4$ and $4\times4$ kernels per head. We also experimented with using two \textit{transposed 2D convolutions} with a skip connection in the middle. The resulting output tensor is at $1/4$ of the spatial resolution of the input. We use the following heads in our network: \begin{itemize} \item \textbf{Class segmentation head} predicts a multi-channel tensor, one channel per class. Each value contains a confidence indicating that a given pixel belongs to a class corresponding to its channel. \item \textbf{Instance regression head} predicts oriented bounding boxes for an object using an $n_r$ ($n_r = 6$) channels of information for each predicted pixel. The $n_r$ element vectors contains: [$\delta_x$, $\delta_y$, $w_0$, $l_0$, $\sin\theta$, $\cos\theta$], where ($\delta_x$, $\delta_y$) points toward the centroid of the corresponding object, $w_0$ $\times$ $l_0$ are the object dimensions, and $\theta$ is the orientation in the top-down BEV. \item \textbf{Inverse sensor model head} (ISM) computes a map of occupancy probabilities for each grid cell~\cite{ism}. \end{itemize} \subsection{Loss} Our loss consists of a standard cross-entropy loss for the classification head, with a larger weight emphasis on the minority classes, $L_1$ loss for instance bounding box regression, and the inverse sensor model loss for free space detection~\cite{ism}. We combine these losses using Bayesian learned weights by modeling each task weight as a homoscedastic task-dependent uncertainty following the method described in \cite{multiTaskLearning}. This approach allows us to efficiently co-train these three diverse tasks without affecting the overall model accuracy. The total loss is defined as: \begin{align} \textit{TotalLoss} = \sum_{i=0}^{K-1} L_i w_i + \mu_w \end{align} where $K$ is the number of tasks/heads, $L_i$ is a loss for task $i$, $w_i$ = $e^{-\delta_i}$, $\delta_i$ is a learned log variance parameter per task, and $\mu_w$ is the mean of $w_i$ weights. \subsection{End-to-end Obstacle Detection} In order to avoid expensive non-maximum suppression (NMS) or clustering at post-processing (e.g. DBSCAN), we employ an end-to-end approach by classifying a single pixel per obstacle, as inspired by OneNet~\cite{oneNet}. First we compute the $L_1$ loss for the regression head and the pixel-wise classification loss for the classification head. Next, for each target obstacle, we select the foreground pixel with the lowest total loss between $(\textit{ClassWeight} * \textit{ClassLossPerPixel}) + \textit{RegressionLossPerPixel}$. This pixel is then selected for the final loss computation while the rest of the foreground pixels are ignored. The losses from the background pixels are then selectively used by utilizing hard negative mining. Finally, we perform batch normalization by dividing the total cross-entropy loss by the number of positive pixels selected during the above process. The regression losses are computed only for the selected positive pixels. At inference time we simply pick all of the candidate pixels above a certain threshold in the classification head, per class. The obstacle dimensions are picked directly from the regression head for each corresponding threshold candidate. By using this technique our network is able to directly output the final obstacle without expensive post-processing. \subsection{Converting ISM Head Output to a Radial Distance Map} Autonomous vehicle applications often represent drivable free space area by its boundary contour. In this sections we describe how to convert the boundary contour to a \textit{radial distance map} (RDM) if needed. The RDM assigns a set of angular directions $\phi_{f}$, in the top-down BEV view around the ego-vehicle, to the distance $d_{f}$ between a reference point $\vec{p}_{ref}$ on the ego-vehicle and the drivable free space boundary. To compute the RDM, we first re-sample the dense occupancy probability map (DNN output) into a polar coordinate system centered around the reference point $\vec{p}_{ref}$. By employing a nearest-neighbour interpolation schema, the re-sampling process can be expressed in terms of an indexing operation. This assigns the value of each pixel $(\phi_{f},d_{f})$ of the polar representation the value of a single pixel of the predicted dense occupancy probability map. Since this mapping only depends on the dimensions of the occupancy map and the position of reference point $\vec{p}_{ref}$, all required indices can be calculated offline and stored in a lookup table. Fig.~\ref{fig:occupancy_map_polar} shows the occupancy probability map. Fig.~\ref{fig:occupancy_map} shows it re-sampled in polar coordinates. After re-sampling, the distance $d_{f}$ for each angular direction $\phi_{f}$ is determined by finding the first pixel along each angular axis, where the occupancy probability reaches some threshold $p_{occ}$. Fig.~\ref{fig:drivable_freespace_boundary} shows the RDM representation of the drivable free space boundary derived by this procedure from the dense occupancy probability map shown in Fig.~\ref{fig:occupancy_map}. \begin{figure} \centering \begin{subfigure}{1\linewidth} \includegraphics[width=1.0\linewidth]{radarnet_images/occupancy_map_polar.png} \end{subfigure} \caption{Predicted dense occupancy map re-sampled into the polar coordinate system centered around the reference point $\vec{p}_{ref}$. Gradient colors from red (low) to yellow (high) show probabilities.} \label{fig:occupancy_map_polar} \end{figure} \begin{figure} \vspace{-4mm} \centering \begin{subfigure}{1\linewidth} \includegraphics[width=1.0\linewidth]{radarnet_images/rdm.png} \end{subfigure} \caption{Radial distance map representation of the drivable free space boundary extracted from the predicted dense occupancy probability map.} \label{fig:drivable_freespace_boundary} \vspace{-2mm} \end{figure} \section{EXPERIMENTS} \subsection{Internal Dataset Experiments} Datasets, benchmarks and published DNNs dedicated to RADAR based obstacle and freespace detection are limited at this time which presents difficulty when evaluating. See Table~\ref{tab:radarmethods} for a list of available methods and their features. The closest works ~\cite{radarNet2020}~\cite{hendy2020fishing}~\cite{centerNetCameraRadarFusion}~\cite{liraNet} use sensor fusion and do not share RADAR only results publicly. Thus, to the best of our knowledge we are setting a baseline for obstacle detection, classification and freespace regression using RADAR peaks alone. Detection of pedestrians and cyclists is a big challenge due to the sparsity of the RADAR signal. \begin{table} \caption{Related RADAR detection methods. Our method (in bold) uses only RADAR data, supports object and free space detection and we provide public results.} \begin{center} \begin{tabular}{l|l|l|l|l} \textbf{Method}&\textbf{Radar Only}&\textbf{Public}&\textbf{Objects}&\textbf{Freespace}\\ \hline \hline RadarNet\cite{radarNet2020} & \ding{55} & \checkmark & \checkmark & \ding{55} \\ FishingNet\cite{hendy2020fishing} & \ding{55} & \checkmark & \checkmark & \ding{55} \\ LiRaNet\cite{liraNet} & \ding{55} & \checkmark & \checkmark & \ding{55} \\ RADModel\cite{azimuthRangeTensor} & \checkmark & \ding{55} & \checkmark & \ding{55} \\ XSense\cite{probOrientedRadar} & \checkmark & \ding{55} & \checkmark & \ding{55} \\ OccupancyNet\cite{occupancyISM2019} & \checkmark & \checkmark & \ding{55} & \checkmark \\ \textbf{Ours} & \textbf{\checkmark} & \textbf{\checkmark} & \textbf{\checkmark} & \textbf{\checkmark} \\ \end{tabular} \end{center} \label{tab:radarmethods} \vspace{-4mm} \end{table} We evaluated our DNN on our internal NVIDIA's RADAR dataset as well as on the nuScenes public dataset. We also compared our DNN to other published works as much as was practically possible and list all results in this section. For our internal NVIDIA's RADAR dataset, we used held out test data mentioned in section~\ref{dataset} for evaluation. It is important to note that, even after filtering out the ground truth bounding boxes, which contain too few RADAR peak detections (as described in section~\ref{label_propogation}), we still end up with noisy ground truth labels. For example, there are many instances where vehicles are occluded by other vehicles and so human labelers are not able to create good ground truth from LiDAR data alone. In such cases, RADAR still produces valid returns and some obstacles are correctly classified by our DNN, but they are marked as false positives at evaluation due to ground truth shortcomings. This lowers our precision. Also, LiDAR sensor is mounted higher on the vehicle (roof) than RADAR sensors (bumpers) and therefore some obstacles may be visible by LiDAR and have limited RADAR visibility which leads to noisy RADAR data and labels. This results in false negatives and lowers our recall. Our results for the object detection task can be seen in Tables~\ref{tab:internal_results_800},~\ref{tab:internal_mAP}. The metrics for the free space detection task (Table~\ref{tab:internal_freespace}) are calculated for the free space region and the free space RDM independently. The free space region is defined by an occupancy probability $p_{o} < 0.4$. \begin{table} \begin{threeparttable}[b] \caption{Our DNN's obstacle detection accuracy on the internal NVIDIA dataset by class and range.} \begin{center} \begin{tabular}{l|l|l|l} \textbf{Class} & \textbf{Range} & \textbf{F-score, small res.}\tnote{1} & \textbf{F-score, large res.}\tnote{2} \\ \hline \hline vehicles & 0 -- 10 m & 0.728 & 0.770 \\ vehicles & 10 -- 25 m & 0.608 & 0.628 \\ vehicles & 25 -- 40 m & 0.728 & 0.485 \\ vehicles & 40 -- 70 m & 0.327 & 0.319 \\ vehicles & 70 -- 100 m & 0.225 & 0.216 \\ \hline pedestrians & 0 -- 10 m & 0.197 & 0.248 \\ pedestrians & 10 -- 25 m & 0.204 & 0.24 \\ pedestrians & 25 -- 40 m & 0.145 & 0.174 \\ pedestrians & 40 -- 70 m & 0.084 & 0.113 \\ pedestrians & 70 -- 100 m & 0.040 & 0.062 \\ \hline cyclists & 0 -- 10 m & 0.145 & 0.264 \\ cyclists & 10 -- 25 m & 0.125 & 0.257 \\ cyclists & 25 -- 40 m & 0.085 & 0.192 \\ cyclists & 40 -- 70 m & 0.064 & 0.137 \\ cyclists & 70 -- 100 m & 0.065 & 0.114 \\ \end{tabular} \begin{tablenotes} \item [1] Input resolution: $800 \times 800 \times 5$. \item [2] Input resolution: $1024 \times 1024 \times 5$. \end{tablenotes} \end{center} \label{tab:internal_results_800} \end{threeparttable} \vspace{-6mm} \end{table} \begin{table} \begin{threeparttable}[b] \caption{Our DNN's obstacle detection accuracy on the internal NVIDIA dataset by class.} \begin{center} \begin{tabular}{l|l|l} \textbf{Class} & \textbf{mAP, small resolution}\tnote{1} & \textbf{mAP, large resolution}\tnote{2} \\ \hline \hline vehicles & 0.438 & 0.473 \\ pedestrians & 0.039 & 0.057 \\ cyclists & 0.032 & 0.066 \\ \end{tabular} \begin{tablenotes} \item [1] Input resolution: $800 \times 800 \times 5$. \item [2] Input resolution: $1024 \times 1024 \times 5$. \end{tablenotes} \end{center} \label{tab:internal_mAP} \end{threeparttable} \end{table} \begin{table} \caption{Our DNN's free space regression accuracy on the internal NVIDIA dataset.} \begin{center} \begin{tabular}{l|l|l|l|l} \textbf{Resolution} & \textbf{Accuracy} & \textbf{IoU} & \textbf{RDM MAE} & \textbf{RDM IoU} \\ \hline \hline $800 \times 800 \times 5$ & 0.970 & 0.597 & $3.129$~m & 0.630 \\ $1024 \times 1024 \times 5$ & 0.968 & 0.576 & $2.674$~m & 0.712 \\ \end{tabular} \end{center} \label{tab:internal_freespace} \end{table} \subsection{NuScenes Dataset Performance} We further evaluate our approach on the public nuScenes dataset~\cite{nuscenes2019}. This dataset contains sensor data from 1 LiDAR and 5 RADARs. However, the sensors in this dataset are from the older generation making direct comparison difficult. The LiDAR sensor used for the nuScenes data collection contains only 32 beams vs 128 beams in our internal dataset. The extra sparsity in this dataset degrades the quality of our auto-generated free space targets. Similarly, the Continental ARS 408-21 RADARs used in nuScenes dataset produce significantly sparser detections when compared to the newer generation Continental ARS430 RADAR sensors we use in our internal dataset. Nonetheless, we demonstrate respectable results, especially at close ranges. See Tables~\ref{tab:nuscenes_results_800},~\ref{tab:nuscnes_map} and~\ref{tab:nuscenes_freespace} for details. \begin{table} \begin{threeparttable}[b] \caption{Our DNN's obstacle detection accuracy on nuScenes dataset by class and range.} \begin{center} \begin{tabular}{l|l|l|l} \textbf{Class} & \textbf{Range} & \textbf{F-score, small}\tnote{1} & \textbf{F-score, large}\tnote{2} \\ \hline \hline vehicles & 0 -- 10 m & 0.520 & 0.563 \\ vehicles & 10 -- 25 m & 0.500 & 0.538 \\ vehicles & 25 -- 40 m & 0.352 & 0.386 \\ vehicles & 40 -- 70 m & 0.180 & 0.199 \\ vehicles & 70 -- 100 m & 0.080 & 0.086 \\ \hline pedestrians & 0 -- 10 m & 0.056 & 0.059 \\ pedestrians & 10 -- 25 m & 0.046& 0.052 \\ pedestrians & 25 -- 40 m & 0.030 & 0.040 \\ pedestrians & 40 -- 70 m & 0.016 & 0.024 \\ pedestrians & 70 -- 100 m & 0.000 & 0.005 \\ \hline cyclists & 0 -- 10 m & 0.050 & 0.030 \\ cyclists & 10 -- 25 m & 0.068 & 0.066 \\ cyclists & 25 -- 40 m & 0.059 & 0.066 \\ cyclists & 40 -- 70 m & 0.044 & 0.041 \\ cyclists & 70 -- 100 m & 0.000 & 0.000 \\ \end{tabular} \begin{tablenotes} \item [1] Input resolution: $800 \times 800 \times 5$. \item [2] Input resolution: $1024 \times 1024 \times 5$. \end{tablenotes} \end{center} \label{tab:nuscenes_results_800} \end{threeparttable} \end{table} \begin{table} \begin{threeparttable}[b] \caption{Our DNN's obstacle detection accuracy on nuScenes dataset by class.} \begin{center} \begin{tabular}{l|l|l} \textbf{Class} & \textbf{mAP, small resolution}\tnote{1} & \textbf{mAP, large resolution}\tnote{2} \\ \hline \hline vehicles & 0.245 & 0.280 \\ pedestrians & 0.002 & 0.003 \\ cyclists & 0.005 & 0.004 \\ \end{tabular} \begin{tablenotes} \item [1] Input resolution: $800 \times 800 \times 5$. \item [2] Input resolution: $1024 \times 1024 \times 5$. \end{tablenotes} \end{center} \label{tab:nuscnes_map} \end{threeparttable} \end{table} \begin{table} \caption{Our DNN's free space regression accuracy on nuScenes dataset.} \begin{center} \begin{tabular}{l|l|l|l|l} \textbf{Resolution} & \textbf{Accuracy} & \textbf{IoU} & \textbf{RDM MAE} & \textbf{RDM IoU} \\ \hline \hline $800 \times 800 \times 5$ & 0.896 & 0.351 & $10.621$~m & 0.394 \\ $1024 \times 1024 \times 5$ & 0.881 & 0.353 & $9.584$~m & 0.441 \\ \end{tabular} \end{center} \label{tab:nuscenes_freespace} \vspace{-2mm} \end{table} We further compare our NVRadarNet DNN free space detection accuracy against the method published in~\cite{occupancyISM2019}, which also presents results on nuScenes dataset. However, this method operates on a grid covering the area in front of the ego vehicle up to $86$~m with $10$~m to each side and, unlike our approach, does not regresses or classifies obstacles. For this comparison we performed the evaluation on the same image region, while converting the predicted occupancy probability to three classes \textit{Occupied}, \textit{Free} and \textit{Unobserved} as follows. \begin{itemize} \item Occupied: $p_{occ} > 0.65$ \item Free: $p_{occ} < 0.35$ \item Unobserved: $0.35 <= p_{occ} <= 0.65$ \end{itemize} The results are given in table~\ref{tab:occupancy_net_freespace}. Our DNN outperforms the other method for occupied space regression (best results in bold) and performs similarly on other tasks. \begin{table} \caption{Comparison to OccupancyNet~\cite{occupancyISM2019} on nuScenes dataset. Best results in bold.} \begin{center} \begin{tabular}{l|l|l|l|l} \textbf{Method} & \textbf{Occupied} & \textbf{Free} & \textbf{Unobs.} & \textbf{mIoU} \\ \hline \hline OccupancyNet~\cite{occupancyISM2019} & 0.108 & \textbf{0.614} & \textbf{0.593} & \textbf{0.439} \\ Ours @ $800 \times 800$ & \textbf{0.237} & 0.564 & 0.436 & 0.412 \\ Ours @ $1024 \times 1024$ & 0.222 & 0.574 & 0.405 & 0.400 \\ \end{tabular} \end{center} \label{tab:occupancy_net_freespace} \end{table} \subsection{NVRadarNet DNN Inference} Our NVRadarNet DNN can be trained in mixed precision mode using INT8 quantization without any loss of accuracy. We export the network using NVIDIA TensorRT and time it on NVIDIA DRIVE AGX's embedded GPU used in our autonomous vehicles. Our DNN is able to achieve \textbf{1.5~ms} end-to-end inference with all three heads. We process all of the surround RADARs, perform obstacle detection and free space segmentation much faster than real-time on the embedded GPU. It was difficult to find other RADAR DNNs inference timings in the literature for direct comparison. We only found that \cite{azimuthRangeTensor} is an order of magnitude slower. \section{CONCLUSION} In this work, we presented NVRadarNet DNN, a real-time deep neural network for obstacle and drivable free space detection from raw RADAR data provided by common automotive RADARs. We benchmarked our DNN on both internal NVIDIA dataset and the public nuScenes dataset and provided accuracy results. Our DNN runs faster than real-time at \textbf{1.5~ms} end-to-end inference time on NVIDIA DRIVE AGX's embedded GPU. To date, we are not aware of any other RADAR only networks that can simultaneously perform obstacle detection and free space regression while running faster than real-time on automotive embedded computers. \section*{Acknowledgment} We would like to thank Sriya Sarathy, Tilman Wekel and Stan Birchfield for their technical contributions. We further would like to acknowledge David Nister, Sangmin Oh and Minwoo Park for their support. \clearpage \bibliographystyle{IEEEtran}

a5480228e825f6f8692f7997c5fddeca05964846

\section{Introduction} Large datasets containing millions of samples have become the standard for obtaining advanced models in many artificial intelligence directions, including natural language processing~\cite{brown2020language}, speech recognition~\cite{baevski2020wav2vec}, and computer vision~\cite{goyal2021self}. Meanwhile, large datasets also raise some issues. For example, data storage and preprocessing are becoming more and more difficult. Also, expensive servers are needed to train models on these datasets, which is not friendly for low-resource environments~\cite{li2022tri}. An effective way to solve these problems is data selection (coreset construction) which identifies representative training samples of large datasets~\cite{bachem2017practical}. However, since some of the original data cannot be discarded, there is an upper limit on the compression rate of the data selection method. \par Recently, dataset distillation as an alternative method to data selection has attracted widespread attention~\cite{wang2018datasetdistillation}. Dataset distillation is the task of synthesizing a small dataset that preserves most information of the original large dataset. The algorithm of dataset distillation takes a sizeable real dataset as input and synthesizes a small distilled dataset. Unlike the data selection method that uses actual data from the original dataset, dataset distillation generates synthetic data with a different distribution from the original one~\cite{dong2022privacy}. Therefore, the dataset distillation method can distill the whole dataset into several images, or even only one image~\cite{sucholutsky2021soft}. Dataset distillation has many application scenarios, such as privacy protection~\cite{li2020soft, song2022federated}, continual learning~\cite{wiewel2021soft, sangermano2022sample}, neural architecture search~\cite{such2020generative, zhao2021datasetcondensation}, etc. \par Since the dataset distillation task was first introduced in 2018 by Wang et al.~\cite{wang2018datasetdistillation}, it has gained increasing attention in the research community. The original dataset distillation algorithm is based on meta-learning and optimizes the distilled images with gradient-based hyperparameter optimization. Subsequently, many works have significantly improved the distillation performance with label distillation~\cite{bohdal2020flexible}, gradient matching~\cite{zhao2021datasetcondensation}, differentiable augmentation~\cite{zhao2021differentiatble}, and distribution/feature matching~\cite{zhao2021distribution, wang2022cafe}. The recently proposed dataset distillation method by matching network parameters has been the new SOTA on several datasets~\cite{cazenavette2022dataset}. However, a network usually has a large number of parameters. And we found that a few parameters are difficult to match in the distillation process and harm the distillation performance, which could be improved. \par In this paper, we propose a new dataset distillation method using parameter pruning. As one of the model pruning approaches, parameter pruning is often used for model compression and accelerated model training. Here, we introduce parameter pruning into dataset distillation to remove the effect of difficult-to-match parameters. The proposed method can synthesize more robust distilled datasets by pruning difficult-to-match parameters in the distillation process, improving the distillation and cross-architecture generalization performance. Experimental results on two benchmark datasets and a real-world COVID-19 chest X-ray (CXR) dataset show the superiority of the proposed method to other SOTA dataset distillation methods. \par Our main contributions can be summarized as follows: \begin{itemize} \item We propose a new dataset distillation method using parameter pruning, which can synthesize more robust distilled datasets and improve the distillation performance. \item The proposed method can outperform other SOTA dataset distillation methods on two benchmark datasets and have better performance in cross-architecture generalization. \item We verify the effectiveness of the proposed method in the real-world application on a COVID-19 CXR dataset. \end{itemize} \begin{figure*}[t] \centering \includegraphics[width=15cm]{Image/Method.png} \caption{Overview of the proposed method. Our method uses a teacher-student architecture, and the objective is to make the student network parameters $\tilde{\theta}'_{i,J}$ match the teacher network parameters $\theta'_{i+K}$. Our method can avoid the influence of the difficult-to-match parameters on the distilled dataset by pruning the parameters in teacher and student networks.} \label{fig1} \end{figure*} \section{Methodology} An overview of the proposed method is shown in Fig.~\ref{fig1}. Our method is constructed on a teacher-student architecture, and the objective is to make the student network parameters trained on the distilled dataset $\mathcal{D}_\textrm{distill}$ match the teacher network parameters trained on the original large dataset $\mathcal{D}_\textrm{original}$. Our method consists of three stages, teacher-student architecture training, dataset distillation using parameter pruning, and optimized distilled dataset generation, which we will show details in the following subsections. \subsection{Teacher-Student Architecture Training} First, we pre-train $N$ teacher networks on $\mathcal{D}_\textrm{original}$ and save their snapshot parameters at each epoch. We define teacher parameters as time sequences of parameters $\{\theta_{i}\}^{I}_{0}$. Meanwhile, student parameters are defined as $\tilde{\theta}_{i}$ who are trained on the distilled dataset $\mathcal{D}_\textrm{distill}$ at each training step $i$. At each distillation step, we first sample parameters from one of the teacher parameters at a random step $i$ and use it to initialize student parameters as $\tilde{\theta}_{i}=\theta_{i}$. We set an upper bound $I^{+}$ on the random step $i$ to ignore the less informative later parts of the teacher parameters. And the number of updates for student parameters and teacher parameters are set to $J$ and $K$, where $J \ll K$. For each student update $j$, we sample a minibatch $b_{i,j}$ from distilled dataset as follows: \begin{equation} b_{i,j} \thicksim \mathcal{D}_\textrm{distill}, \end{equation} Then we perform $j$ updates on the student parameters $\tilde{\theta}$ using the cross-entropy loss $\ell$ as follows: \begin{equation} \tilde{\theta}_{i,j+1} = \tilde{\theta}_{i,j} - \alpha\nabla\ell(\mathcal{A}(b_{i,j});\tilde{\theta}_{i,j}), \end{equation} where $\alpha$ represents the trainable learning rate. $\mathcal{A}$ represents a differentiable data augmentation module proposed in~\cite{zhao2021differentiatble}, which can improve the distillation performance. \begin{algorithm}[t] \caption{Dataset Distillation using Parameter Pruning} \label{alg1} \begin{algorithmic}[1] \REQUIRE $\{\theta_{i}\}^{I}_{0}$ : teacher parameters trained on $\mathcal{D}$; $\alpha_{0}$: initial value for $\alpha$; $\mathcal{A}$: differentiable augmentation function; $\epsilon$: threshold for pruning; $T$: number of distillation step; $J$: number of updates for student network; $K$: number of updates for teacher network; $I^{+}$: maximum start epoch. \ENSURE optimized distilled dataset $\mathcal{D}^{\ast}_\textrm{distill}$ and learning rate $\alpha^{\ast}$. \\ \STATE Initialize distilled dataset: $\mathcal{D}_\textrm{distill} \thicksim \mathcal{D}$ \STATE Initialize trainable learning rate: $\alpha = \alpha_{0}$ \FOR{each distillation step $t = 0$ to $T - 1$} \STATE Choose random start epoch $i < I^{+}$ \STATE Initialize student network with teacher parameter: $\tilde{\theta}_{i}=\theta_{i}$ \FOR{each distillation step $j = 0$ to $J - 1$} \STATE Sample a minibatch of distilled dataset: $b_{i,j} \thicksim \mathcal{D}_\textrm{distill}$ \STATE Update student network with cross-entropy loss: \STATE $\tilde{\theta}_{i,j+1} = \tilde{\theta}_{i,j} - \alpha\nabla\ell(\mathcal{A}(b_{i,j});\tilde{\theta}_{i,j})$ \ENDFOR \IF{parameter similarity in $\tilde{\theta}_{i,J}$ and $\theta_{i+K}$ is less than $\epsilon$} \STATE Prune network parameters: \STATE $\tilde{\theta}'_{i,J}, \theta'_{i+K}, \theta'_{i} = \textrm{Prune}(\tilde{\theta}_{i,J}, \theta_{i+K}, \theta_{i})$ \ENDIF \STATE Compute loss between pruned parameters: \STATE $\mathcal{L} = || \tilde{\theta}'_{i,J}-\theta'_{i+K} ||^{2}_{2} \,\,\,/\,\,\, || \theta'_{i}-\theta'_{i+K} ||^{2}_{2}$ \STATE Update $\mathcal{D}_\textrm{distill}$ and $\alpha$ with respect to $\mathcal{L}$ \ENDFOR \end{algorithmic} \end{algorithm} \begin{table*}[t] \footnotesize \centering \caption{Test results of different methods on CIFAR-10 and CIFAR-100.} \label{tab1} \begin{tabular}{l|c|cccccccc|cc} \hline & IPC & Random & Forgetting~\cite{toneva2019empirical} & Herding~\cite{chen2010super} & DSA~\cite{zhao2021differentiatble} & DM~\cite{zhao2021distribution} & CAFE~\cite{wang2022cafe} & MTT~\cite{cazenavette2022dataset} & Ours & Full Dataset\\\hline \multirow{3}*{CIFAR-10} & 1 & 14.4$\pm$2.0 & 13.5$\pm$1.2 & 21.5$\pm$1.2 & 28.8$\pm$0.7 & 26.0$\pm$0.8 & 31.6$\pm$0.8 & 46.3$\pm$0.8 & \bfseries{46.4$\pm$0.6} & \multirow{3}*{84.8$\pm$0.1} \\ & 10 & 26.0$\pm$1.2 & 23.3$\pm$1.0 & 31.6$\pm$0.7 & 52.1$\pm$0.5 & 48.9$\pm$0.6 & 50.9$\pm$0.5 & 65.3$\pm$0.7 & \bfseries{65.5$\pm$0.3} & \\ & 50 & 43.4$\pm$1.0 & 23.3$\pm$1.1 & 40.4$\pm$0.6 & 60.6$\pm$0.5 & 63.0$\pm$0.4 & 62.3$\pm$0.4 & 71.6$\pm$0.2 & \bfseries{71.9$\pm$0.2} & \\\hline \multirow{3}*{CIFAR-100} & 1 & 4.2$\pm$0.3 & 4.5$\pm$0.2 & 8.4$\pm$0.3 & 13.9$\pm$0.3 & 11.4$\pm$0.3 & 14.0$\pm$0.3 & 24.3$\pm$0.3 & \bfseries{24.6$\pm$0.1} & \multirow{3}*{56.2$\pm$0.3} \\ & 10 & 14.6$\pm$0.5 & 15.1$\pm$0.3 & 17.3$\pm$0.3 & 32.3$\pm$0.3 & 29.7$\pm$0.3 & 31.5$\pm$0.2 & 40.1$\pm$0.4 & \bfseries{43.1$\pm$0.3} & \\ & 50 & 30.0$\pm$0.4 & 30.5$\pm$0.3 & 33.7$\pm$0.5 & 42.8$\pm$0.4 & 43.6$\pm$0.4 & 42.9$\pm$0.2 & 47.7$\pm$0.2 & \bfseries{48.4$\pm$0.3} & \\\hline \end{tabular} \end{table*} \subsection{Dataset Distillation Using Parameter Pruning} Next, we get the student parameters $\tilde{\theta}_{i,J}$ trained on the distilled dataset $\mathcal{D}_\textrm{distill}$ from $J$ updates after initializing the student network. Meanwhile, we can get the teacher parameters $\theta_{i+K}$ trained on the original dataset $\mathcal{D}_\textrm{original}$ from $K$ updates, which are the known parameters that have been pre-trained. If the similarity of parameters in $\tilde{\theta}_{i,J}$ and $\theta_{i+K}$ is less than a threshold $\epsilon$, these parameters are recognized as difficult-to-match parameters and are pruned as follows: \begin{equation} \tilde{\theta}'_{i,J}, \theta'_{i+K}, \theta'_{i} = \textrm{Prune}(\tilde{\theta}_{i,J}, \theta_{i+K}, \theta_{i}), \end{equation} where $\textrm{Prune}$ represents a function that transforms the parameters to a one-dimension vector and prunes the parameters under the threshold at each last distillation step. By pruning difficult-to-match parameters in teacher and student networks, the proposed method can avoid the influence of these parameters on the distilled dataset, which can improve the distillation and cross-architecture generalization performance. The final loss $\mathcal{L}$ calculates the normalized squared $L_{2}$ error between pruned 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} where we normalize the $L_{2}$ error by the distance $\theta'_{i}-\theta'_{i+K}$ moved by the teacher so that we can still obtain proper supervision from the late training period of the teacher network even if it has converged. In addition, the normalization eliminates cross-layer and neuronal differences in magnitude. \subsection{Optimized Distilled Dataset Generation} Finally, we minimize the loss $\mathcal{L}$ using momentum stochastic gradient descent (SGD) and backpropagate the gradients through all $J$ updates to the student network for updating the pixels of the distilled dataset $\mathcal{D}_\textrm{distill}$ and trainable learning rate $\alpha$. Note that the process of searching the optimized learning rate $\alpha^{\ast}$ can act as an automatic adjustment for the number of student and teacher updates (i.e., hyperparameters $J$ and $K$). The distillation process of the proposed method is summarized in Algorithm~\ref{alg1}. After obtaining the optimized distilled dataset $\mathcal{D}^{\ast}_\textrm{distill}$, we can train different neural networks on it for efficiency and use for downstream tasks, such as continual learning and neural architecture search. \section{Experiments} In this section, we conduct three experiments to verify the effectiveness of the proposed method. The experimental settings are shown in subsection 3.1. Subsections 3.2, 3.3, and 3.4 show the results of benchmark comparison, cross-architecture generalization, and real-world dataset verification, respectively. All of our experiments were conducted using the PyTorch framework with an NVIDIA RTX A6000 GPU. \begin{figure}[t] \centering \includegraphics[width=8cm]{Image/CIFAR-10.png} \caption{Visualization results of the distilled CIFAR-10 dataset.} \label{fig2} \end{figure} \begin{table}[t] \footnotesize \centering \caption{Test results of different width KIP~\cite{nguyen2021kipimprovedresults} and our method on CIFAR-10 and CIFAR-100.} \label{tab2} \begin{tabular}{l|c|c|cc} \hline & IPC & KIP-1024 & KIP-128 & Ours-128 \\\hline \multirow{3}*{CIFAR-10} & 1 & 49.9 & 38.3 & \bfseries{46.4} \\ & 10 & 62.7 & 57.6 & \bfseries{65.5} \\ & 50 & 68.6 & 65.8 & \bfseries{71.9} \\\hline \multirow{3}*{CIFAR-100} & 1 & 15.7 & 18.2 & \bfseries{24.6} \\ & 10 & 28.3 & 32.8 & \bfseries{43.1}\\ & 50 & - & - & \bfseries{48.4} \\\hline \end{tabular} \end{table} \subsection{Experimental Settings} We used two benchmark datasets (i.e., CIFAR-10 and CIFAR-100) in the experiments for comparison with other methods. The resolution of images in CIFAR-10 and CIFAR-100 is 32 $\times$ 32. We also used a COVID-19 CXR dataset~\cite{li2022self} for proving the effectiveness of our method in the real-world application. The COVID-19 CXR dataset has four classes, including COVID-19 (3616 images), Lung Opacity (6012 images), Normal (10192 images), and Viral Pneumonia (1345 images). Since CXR images have high resolutions (224 $\times$ 224), they are resized to 112 $\times$ 112 for rapid distillation. \par For comparative methods, we used three data selection methods, including random selection (Random), example forgetting (Forgetting)~\cite{toneva2019empirical}, and herding method (Herding)~\cite{chen2010super}. Also, we used five SOTA dataset distillation methods, including Differentiable Siamese Augmentation (DSA)~\cite{zhao2021differentiatble}, Distribution Matching (DM)~\cite{zhao2021distribution}, Aligning Features (CAFE)~\cite{wang2022cafe}, Matching Training Trajectories (MTT)~\cite{cazenavette2022dataset} and Kernel Inducing Point (KIP)~\cite{nguyen2021kipimprovedresults}. The network used in this study is a sample 128-width ConvNet~\cite{gidaris2018dynamic}, which is often used in current dataset distillation methods. We conducted three experiments to verify the effectiveness of the proposed method, including benchmark comparison, cross-architecture generalization, and real-world dataset verification. We found that pruning too many parameters would cause the model training to crash. Hence parameter pruning threshold $\epsilon$ was set to 0.1, which performed well in all experiments. All experimental results are average accuracy and standard deviation of five networks trained from scratch on the distilled dataset. \subsection{Benchmark Comparison} In this subsection, we verify the effectiveness of the proposed method by comparing it with other SOTA dataset distillation methods on two benchmark datasets, i.e., CIFAR-10 and CIFAR-100. We employed zero-phase component analysis (ZCA) whitening with default parameters and used a 3-depth ConvNet the same as MTT~\cite{cazenavette2022dataset}. We pre-trained 200 teacher networks (50 epochs per teacher) for the distillation process. The number of distillation steps was set to 5,000. And the number of images per class (IPC) was set to 1, 10, and 50, respectively. For KIP~\cite{nguyen2021kipimprovedresults}, we used their original 1024-width ConvNet (KIP-1024) and 128-width ConvNet (KIP-128) for a fair comparison. Also, we used their custom ZCA implementation for distillation and evaluation. \par From Table~\ref{tab1}, we can see that the proposed method outperformed dataset selection methods and SOTA dataset distillation methods in all settings. Especially for CIFAR-100 with IPC = 10, our method has an accuracy increased by 3.0\% compared to the second best method MTT. As shown in Table~\ref{tab2}, the proposed method drastically outperformed KIP using the same 128-width ConvNet. Even for KIP that uses 1024-width ConvNet, our method has higher accuracy except for CIFAR-10 with 1 image per class. For the results of CIFAR-100 with IPC = 50, KIP did not conduct experiments due to the large computational resources and time required, so we only report our results in this paper. Figure~\ref{fig2} shows visualization results of the distilled CIFAR-10 dataset. As shown in Fig.~\ref{fig2}, when we set the number of distilled images to 1, the generated images were more abstract but also more information-dense because all information of a class has to be compressed into only one image in the distillation process. Meanwhile, when the number of distilled images was set to 10, the generated images were more realistic and contained various forms because discriminative features in a class can be compressed into multiple images in the distillation process. For example, we can see various types of dogs and different colored cars. \begin{table}[t] \footnotesize \centering \caption{Cross-architecture generalization results on CIFAR-10 dataset with IPC = 10.} \label{tab3} \begin{tabular}{lcccc} \hline Architecture & ConvNet & AlexNet & VGG & ResNet \\\hline Ours & \bfseries{65.4$\pm$0.4} & \bfseries{35.8$\pm$1.3} & \bfseries{52.9$\pm$0.9} & \bfseries{51.8$\pm$1.1} \\ MTT~\cite{cazenavette2022dataset} & 64.3$\pm$0.7 & 34.2$\pm$2.6 & 50.3$\pm$0.8 & 46.4$\pm$0.6 \\ KIP~\cite{nguyen2021kipimprovedresults} & 47.6$\pm$0.9 & 24.4$\pm$3.9 & 42.1$\pm$0.4 & 36.8$\pm$1.0 \\\hline \end{tabular} \end{table} \begin{table}[t] \footnotesize \centering \caption{Test results on a COVID-19 CXR dataset when using different numbers of distilled images.} \label{tab4} \begin{tabular}{lcccccc} \hline IPC & 1 & 5 & 10 & 20 \\\hline Ours & \bfseries{54.2$\pm$3.7} & \bfseries{81.6$\pm$0.3} & \bfseries{83.5$\pm$0.2} & \bfseries{84.1$\pm$0.6} \\ MTT~\cite{cazenavette2022dataset} & 52.5$\pm$5.5 & 79.3$\pm$0.4 & 82.2$\pm$0.2 & 82.7$\pm$0.5 \\\hline \end{tabular} \end{table} \begin{figure}[t] \centering \includegraphics[width=7.8cm]{Image/COVID.png} \caption{Visualization results of real and distilled CXR images.} \label{fig3} \end{figure} \subsection{Cross-Architecture Generalization} In this subsection, we verify the effectiveness of our method in cross-architecture generalization. Cross-architecture means using distilled images generated by one architecture and testing on other architectures. The distilled images were generated by ConvNet on CIFAR-10 and the number of distilled images was set to 10. We used the same pre-trained teacher networks used in subsection 3.2 for rapid distillation and experimentation. For KIP, we used 128-width ConvNet and their custom ZCA implementation for distillation and evaluation. And we tested the accuracy of ConvNet and three cornerstone networks for evaluation of cross-architecture generalization, i.e., AlexNet~\cite{krizhevsky2012imagenet}, VGG~\cite{simonyan2015very}, and ResNet~\cite{he2016deep}. \par From Table~\ref{tab3}, we can see that our method outperformed the SOTA methods MTT and KIP with all architectures. Especially for ResNet, our method has increased accuracy by 5.2\% compared to MTT. The results indicate that our method generated more robust distilled images. By pruning difficult-to-match parameters in teacher and student networks, the proposed method can avoid the influence of these parameters on the distilled dataset, which improves cross-architecture generalization performance. \subsection{Real-World Dataset Verification} In this subsection, we verify the effectiveness of the proposed method in real-world application on a COVID-19 CXR dataset. We used a 5-depth ConvNet for distillation since the image resolution increases significantly compared to CIFAR-10 and CIFAR-100. We pre-trained 100 teacher networks (50 epochs per teacher) for the distillation process. The number of distillation steps was set to 5,000. And we tested the COVID-19 accuracy when the IPC was set to 1, 5, 10, and 20, respectively. \par Table~\ref{tab4} shows that the proposed method achieved high test accuracy even when using a few distilled CXR images, such as IPC = 20 (80 distilled CXR images). Furthermore, the proposed method outperformed the SOTA method MTT in all IPC settings, indicating the effectiveness of our method in the real-world application for COVID-19 detection. Figure~\ref{fig3} shows visualization results of real and distilled CXR images. The distilled CXR images are generated from noise and have different distributions from the original images. Compared with the original CXR images, the distilled images are visually different, showing the potential of dataset distillation for anonymization and privacy preservation. \section{Conclusion} This paper has proposed a novel dataset distillation method using parameter pruning. The proposed method can synthesize more robust distilled datasets by pruning difficult-to-match parameters in the distillation process. Experimental results show that the proposed method can outperform other SOTA dataset distillation methods on two benchmark datasets and have better performance in cross-architecture generalization. We also verify the effectiveness of our method in the real-world application on a COVID-19 CXR dataset. \newpage \bibliographystyle{IEEEbib}

efa3ca07983104e519d0f20786211bae7ab25228

\section{Introduction} Finite coverings of CW complexes are classical subject in algebraic topology. Recently, double coverings attract a lot of attentions for the topology of hyperplane arrangement complement \cite{Yos20,Suc22,ISY22}. A finite collection $\mathcal{A}$ of hyperplanes in $\mathbb{C}^n$ (or $\mathbb{CP}^n$) is called a complex affine (resp. projective) hyperplane arrangement. The topology of the hyperplane arrangement complement is very interesting. For instance, Dimca and Papadima \cite{DP03} and Randell \cite{Ran02} independently showed that the complement of hyperplane arrangement is homotopy equivalent to a minimal CW complex. A fundamental problem in the theory of hyperplane arrangements is to decide whether various topological invariants of the complement of $\mathcal{A}$ are determined by the combinatorial structure of $\mathcal{A}$. It is well known that Betti numbers and the cohomology ring of arrangement complements are combinatorially determined (e.g., see \cite{OT}). However, it is still an open question whether the Betti numbers of a finite abelian cover of a projective arrangement complement are combinatorially determined. This includes the Milnor fiber of a central hyperplane arrangement, see \cite{PS17} for recent progress in this direction and also \cite{Suc01} for an overview of the theory. Yoshinaga studied the mod-2 Betti numbers of double covering for hyperplane arrangement complement and showed that these are combinatorially determined \cite[Theorem 3.7]{Yos20}. As an application, he showed that the first integral homoloy group of the Milnor fiber of the icosidodecahedral arrangement has 2-torsion \cite[Theorem 1.2]{Yos20}. Ishibashi, Sugawara and Yoshinaga \cite{ISY22} further studied the 2-torsion part of the homology groups of the double coverings and gave a refinement of Papadima and Suciu's work \cite{PS10}. Based on computations, Ishibashi, Sugawara and Yoshinaga proposed a conjecture \cite[Conjecture 3.3]{ISY22} regarding the first integral homology group of double coverings and the first homology group of a rank one $\mathbb{Z}$-local system. In this note, we settle this conjecture for all degrees with the more general setting: the CW complex is homotopy equivalent to a minimal CW complex. As an application, we show that the integral homology groups of double coverings of hyperplane arrangement complement are combinatorially determined under certain conditions. \bigskip Let $X$ be a connected finite CW complex. Fix a non-zero element $\omega \in H^1(X,\mathbb{Z}_2)$, where $\mathbb{Z}_2=\mathbb{Z}/2\mathbb{Z}$. Then $\omega$ determines a surjective map $\pi_1(X) \to \mathbb{Z}_2\cong \{\pm 1\}$. This gives a double covering $ X^\omega \to X$. On the other hand, the group homomorphism $\pi_1(X)\to \{\pm 1\}=\mathbb{Z}^\times$ also gives a rank one $\mathbb{Z}$-local system which we denote by $\mathcal{L}_\omega$. What is the relation between $H_*(X^\omega,\mathbb{Z})$ and $H_*(X,\mathcal{L}_\omega)$? It is easy to see that $$H_i(X^\omega, \mathbb{C})\cong H_i(X,\mathbb{C})\oplus H_i(X, \mathcal{L}_\omega\otimes_\mathbb{Z} \mathbb{C}).$$ So the difficult part is about the torsions of the homology groups. We give a complete answer to this question when $X$ is homotopy equivalent to a minimal CW complex. \begin{defn} Let $X$ be a connected finite CW complex. We say that the CW-structure on X is minimal if the number of $i$-cells of $X$ coincides with the (rational) Betti number $b_i(X)$, for every $i\geq 0$. Equivalently, the boundary maps in the cellular chain complex $C_*(X, \mathbb{Z})$ are all the zero maps. \end{defn} \begin{thrm} \label{main} Let $X$ be a connected finite CW complex, which is homotopy equivalent to a minimal CW complex. Fix a non-zero element $\omega \in H^1(X,\mathbb{Z}_2)$. Then there exists a bounded complex $(E_*,\partial_*)$ of finitely generated free abelian groups $$\cdots \to E_{i+1} \overset{\alpha_{i}}{\to} E_i \overset{\alpha_{i-1}}{\to} E_{i-1} \to \cdots $$ such that $H_i(E_*,\alpha_*) \cong H_i(X,\mathcal{L}_\omega)$, $\mathrm{rank } E_i=b_i(X)$ and every entry in the boundary map $\alpha_i$ is divisible by 2. Then the complex $(E_*, \tfrac{1}{2}\alpha_*)$ is well-defined. Moreover, we have $$H_i(X^\omega,\mathbb{Z})\cong H_i(X,\mathbb{Z}) \oplus H _i(E_*, \tfrac{1}{2}\alpha_*).$$ \end{thrm} \begin{rmk} The minimal CW complex assumption is crucial for the above theorem. For example, consider the real projective space $\mathbb{RP}^n$ with $n>1$. Its double covering space is the sphere $S^n$. For any $0<i<n$, we have that $ H_i(S^n,\mathbb{Z})=0$, yet $$H_i(\mathbb{RP}^n,\mathbb{Z})=\begin{cases} \mathbb{Z}_2, & \text{ for } i \text{ odd },\\ 0, & \text{ for } i \text{ even }. \end{cases}$$ \end{rmk} \begin{cor} \label{cor1} With the same assumptions and notations as in Theorem \ref{main}, we have that $$ H_1(X^\omega,\mathbb{Z})\cong \mathbb{Z}^{b_1(X)+r} \oplus \mathbb{Z}/d_1 \mathbb{Z} \oplus \cdots \oplus \mathbb{Z}/d_k \mathbb{Z}$$ with $1<d_1 \mid \cdots \mid d_k$, if and only if, we have that $$ H_1(X,\mathcal{L}_\omega) \cong \mathbb{Z}^r \oplus \mathbb{Z}/2 d_1 \mathbb{Z} \oplus \cdots \oplus \mathbb{Z}/2 d_k \mathbb{Z} \oplus (\mathbb{Z}_2)^{b_1(X)-r-k-1} .$$ \end{cor} \begin{rmk} Since complement of hyperplane arrangement is homotopy equivalent to a minimal CW complex (see \cite{DP03} or \cite{Ran02}), this corollary gives a positive answer to a conjecture proposed by Ishibashi, Sugawara and Yoshinaga \cite[Conjecture 3.3]{ISY22}. In particular, this corollary is compatible with the computations in \cite[Example 3.1, Example 3.2]{ISY22}. \end{rmk} \begin{cor} \label{cor2} With the same assumptions and notations as in Theorem \ref{main}, the torsion part of $H_*(X, \mathcal{L}_\omega)$ is either 0 or a finite direct sum of $\mathbb{Z}_2$ for all degrees if and only if $H_*(X^\omega,\mathbb{Z})$ is torsion-free for all degrees. \end{cor} Both Corollary \ref{cor1} and \ref{cor2} follow directly from Theorem \ref{main} by using the Universal Coefficient Theorem. As an application of Corollary \ref{cor2}, we give a partial answer to a question asked by Yoshinaga \cite[Remark 3.8]{Yos20}. \begin{cor} \label{applcation} Let $X$ be a hyperplane arrangement complement in $\mathbb{C}^n$. Fix a non-zero element $\omega\in H^1(X,\mathbb{Z}_2)$. If $X$ and $\mathcal{L}_\omega$ satisfy one of the following conditions: \begin{itemize} \item[(i)] $\mathcal{L}_\omega$ satisfies the CDO-condition (see \cite[Definition 2.1]{Sug22}), \item[(ii)] the hyperplane arrangement is central and the total turn monodromy of $\mathcal{L}_\omega$ is $-1$ (see \cite[Lemma 2.1]{LMW22}), \end{itemize} then $H_*(X^\omega,\mathbb{Z})$ is torsion-free for all degrees. Moreover, $H_*(X^\omega, \mathbb{Z})$ is combinatorially determined. \end{cor} \textbf{Acknowledgments.} We thank Masahiko Yoshinaga for valuable comments. \section{Proofs} \begin{proof}[Proof of Theorem \ref{main}] Since the homology groups of local system are homotopy invariants, without loss of generality, we assume that $X$ is a minimal CW complex from now on. Note that $H_1(X,\mathbb{Z})$ is torsion free in this case. Then there exists a surjective group homomorphism $\nu \colon \pi_1(X)\to \mathbb{Z}$ with the following commutative diagram \begin{equation} \label{a} \xymatrix{ \pi_1(X)\ar@{->>}"1,3"^{\omega}\ar@{->>}[dr]_{\nu} & & \mathbb{Z}_2, \\ & \mathbb{Z}\ar@{->>}[ur] & }\end{equation} where $\mathbb{Z}\to \mathbb{Z}_2$ is the natural quotient map. Hence the representation of the rank one $\mathbb{Z}$-local system $\mathcal{L}_\omega$ factors through $\nu$ and it sends the generator $1_\mathbb{Z}$ of $\mathbb{Z}$ to $-1\in \mathbb{Z}^\times$. The group of covering transformations of the covering space $X^{\nu}$ is isomorphic to $\mathbb{Z}$ and acts on it. Consider the minimal CW-structure of $X$. By choosing fixed lifts of these cells of $X$ to $X^{\nu}$, we obtain a free basis for the cellular chain complex of $ X^\nu$ as $\mathbb{Z}[t^\pm]$-modules. So the cellular chain complex of $X^{\nu}$, $C_{*}(X^{\nu}, \mathbb{Z})$, is a bounded complex of finitely generated free $\mathbb{Z}[t^\pm]$-modules: \begin{equation} \label{chain compelx} \cdots \to C_{i+1}(X^\nu, \mathbb{Z}) \overset{\partial_{i}}{\to} C_i(X^\nu, \mathbb{Z}) \overset{\partial_{i-1}}{\to} C_{i-1}(X^\nu, \mathbb{Z}) \overset{\partial_{i-2}}{\to} \cdots \overset{\partial_0}{\to} C_0(X^\nu, \mathbb{Z}) \to 0 . \end{equation} With the above free basis for $C_*(X^\nu,\mathbb{Z})$, $\partial_i$ can be written down as a matrix with entries in $\mathbb{Z}[t^\pm]$. Since $X$ is a minimal CW complex, every entry in $\partial_i$ is divisible by $t-1$. Due to the commutative diagram (\ref{a}), $X^\nu$ is a covering space of $X^\omega$. In particular, $H_*(X^\omega,\mathbb{Z})$ can be computed by the chain complex $$C_*(X^\nu,\mathbb{Z})\otimes_{\mathbb{Z}[t^{\pm}]} \mathbb{Z}[t^{\pm}]/(t^2-1),$$ viewed as a complex of finitely generated free abelian groups. Consider the boundary map $\partial_i$. As mentioned above, we can write down $\partial_i$ as a matrix, say a $(m\times n)$-matrix $\big( f_{kj}\big)_{m\times n}$ with $f_{kj}\in \mathbb{Z}[t^{\pm}]$. We can choose $f_{kj}$ to be a integral valued polynomial with non-zero constant. Assume that \begin{center} $f_{kj} \equiv a_{kj}t+b_{kj} \mod (t^2-1)$ with $a_{kj},b_{kj}\in \mathbb{Z}$. \end{center} Since $f_{kj}$ takes the value of 0 at $t=1$, $b_{kj}=-a_{kj}$. Consider the direct sum $$\mathbb{Z}[t^{\pm}]/(t^2-1) \cong \mathbb{Z}\cdot 1 \oplus \mathbb{Z}\cdot t $$ as $\mathbb{Z}$-modules. Then by this choice of basis, $\big(f_{kj}\big)_{m\times n}$ becomes a $(2m\times 2n)$-matrix with entry $f_{kj}$ being replaced by $ \bigl( \begin{smallmatrix} -a_{kj}& a_{kj}\\ a_{kj}& -a_{kj} \end{smallmatrix} \bigr).$ By elementary row and column operations, this new matrix becomes $$ \begin{pmatrix} (-a_{kj})_{m\times n} & 0 \\ 0& 0 \end{pmatrix}_{2m\times 2n} .$$ On the other hand, the homology group of $\mathcal{L}_\omega$ can be computed by the following complex $$(E_*, \alpha_*) \coloneqq C_*(X^\nu,\mathbb{Z})\otimes_{\mathbb{Z}[t^{\pm}]} \mathbb{Z}[t^{\pm}]/(t+1).$$ Note that $f_{kj}$ takes the value $-2a_{kj}$ at $t=-1$. Hence the boundary map $\alpha_i$ can be written down as a matrix $\big(-2 a_{kj}\big)_{m\times n}$. Then the claim follows. \end{proof} \begin{proof}[Proof of Corollary \ref{applcation}] To give the proof, we need to show that the torsion part of $H_*(X, \mathcal{L}_\omega)$ is either 0 or a direct sum of $\mathbb{Z}_2$ for all degrees. In the first case, the cohomology version of this claim is proved in \cite[Theorem 1.3]{Sug22} and \cite[Theorem 1.2]{LMW22}; meanwhile, the cohomology version of the second case is proved in \cite[Lemma 2.1]{LMW22}. Then one gets the homology version by the Universal Coefficient Theorem. Corollary \ref{cor2} implies that $H_*(X^\omega,\mathbb{Z})$ is torsion-free for all degrees. Hence $H_*(X^\omega, \mathbb{Z})$ is combinatorially determined due to \cite[Theorem 2.4]{ISY22}. \end{proof} \begin{ex}[Double star arrangement] Consider the double star arrangement \cite[Figure 2]{ISY22} and we use the same notations as in \cite[Example 3.2]{ISY22}. Take $$\omega =e_8+e_9+e_{10}+e_{12}+e_{13}+e_{14}+e_{15}.$$ It is easy to check that the CDO-condition holds for $\mathcal{L}_\omega$ on the hyperplane $H_6$. Then we have $H_1(M(\mathcal{A}_{\mathcal{DS}}),\mathcal{L}_\omega)\cong \mathbb{Z}_2^9$ (see \cite[Theorem 1.3]{Sug22} and \cite[Theorem 1.2]{LMW22}), hence $H_1(M(\mathcal{A}_{\mathcal{DS}})^\omega,\mathbb{Z})\cong \mathbb{Z}^{10}.$ \end{ex}

42279eca62a150627a8493ca330ec8f5357c0c93

\section{Introduction} \label{sec:intro} Online recommendation is fundamental to functioning of the modern web. From online stores to streaming service engines and content discovery platforms, it enables users to discover relevant content faster, and product owners to make their products visible to a broader public. Recommender systems can be understood as a feedback loop between the recommendation engine and the target audience – based on provided recommendations, the interactions with the recommendations are stored and considered to improve a given recommender further~\cite{batmaz2019review}. Machine learning models underlie many online recommendation systems. The most common branch of algorithms considered, both due to their performance and lower complexity, are \textbf{factorization machines (FM)}. These algorithms require the modelers to provide both the space of features and their interactions (\emph{featurization}), as well as model hyperparameters~\cite{rendle2012factorization}. The latter is the main focus of this paper. Tuning hyperparameters can be considered as an optimization task, formulated as follows: $$\textrm{Solution} \approx \argmin_{\substack{\Theta \in \textrm{hyperParamSpace}}} \mathbb{E} \big [\textsc{Loss} \big ( \textrm{\emph{LearningEngine}},\Theta,\textrm{data}\big ) \big ],$$ \noindent where \emph{data} corresponds to the training data of choice, $\Theta$ to a given hyperparameter configuration, and \emph{LearningEngine} to the machine learning model considered. The minimization aims to identify minimal expected \emph{Loss} of choice (lower is better in this formulation). This formulation considers one \emph{LearningEngine} evaluation for each configuration. As the optimization progresses, configuration-target pairs are obtained and stored. \emph{Surrogate models} exploit this data to estimate which configuration the \emph{LearningEngine} will consider next. The problem with larger data sets is that the amount of engine evaluations is \emph{limited}, hence having \textbf{data-efficient surrogates} is a desired property that can substantially shorten research cycles. The contributions presented are multi-fold, and are stated next. First, we describe the process of integration of surrogate-based optimization into the \textbf{in-house AutoML framework} -- we considered both Bayesian and non-Bayesian (surrogate) models, reporting on their behaviour. We next present the implemented strategy of \textbf{Dynamic Surrogate Switching} (DSS) that builds on the recent ideas of automated model switching during optimization. The results of initial evaluations indicate this is a promising strategy to speed up hyperparameter optimization in real-life settings, where the number of learning engine evaluations is limited due to the scale of data considered. \section{Selected Related work} \label{sec:related} \begin{figure \centering \Description[Example 3D visualization of hyperparameter landscape.]{Hyperparameters visualized: two hyperparameters with corresponding RIG scores.} \includegraphics[width=1\linewidth]{hyperparamExample.png} \caption{Example two-hyperparameter non-convex (interpolated) landscape.} \label{fig:nonconvex} \end{figure} The vast majority of machine learning models require the specification of \textbf{hyperparameters} for their normal mode of operation. Early approaches mostly considered grid-based, random, or evolution-based search through the space of possible configurations~\cite{friedrichs2005evolutionary}. However, recent trends indicate that, albeit random search often represents a viable baseline, exploiting the information obtained during the optimization can be a better strategy. There are multiple ways these hyperparameter-target score data points can be utilized. The simplest examples include incorporation of their statistical properties into the search strategy, however, considering them as input data samples for the \emph{surrogate model} is also often considered~\cite{falkner2018bohb}. The ideas of \textbf{surrogate-based} optimization are tightly linked with the field of Bayesian optimization~\cite{malu2021bayesian}; effectively, if a surrogate is a probabilistic model capable of also outputting the (epistemic) uncertainty associated with a given prediction, both the prediction and the associated uncertainty can be used by the \emph{acquisition function} -- the procedure responsible for linking the surrogate's output(s) with parts of the search space that should be considered in the next round of learning engine evaluations. There exist a plethora of commonly considered acquisition functions -- examples include the probability of improvement, expected improvement and more~\cite{frazier2018tutorial}. If the surrogate is only capable of outputting the prediction, acquisition functions are simpler -- for example, viable configurations can be identified already by sorting the space of candidate configurations according to the surrogate's outputs. The \textbf{acquisition function} is thus responsible for using a trained surrogate to \emph{score} parts of the hyperparameter space -- these can be obtained via random sampling or more involved optimization schemes (such as for example, using L-BFGS~\cite{zhu1997algorithm} to minimize the acquisition function directly). Non-convex optimization landscapes can emerge already when considering as few as two hyperparameters in production environments (example shown in Figure~\ref{fig:nonconvex}). The presented methodology also builds on the recent ideas of Online \textbf{AutoML} (Automated Machine Learning)~\cite{he2021automl}. The goal of AutoML systems is to automate various data-related processes commonly manually performed by human modelers. In particular, we built on some of the ideas presented recently in~\cite{celik2022online}, where dynamic ensembles were used to perform online learning. The main finding of the aforementioned study is that \emph{switching} the models can help mitigate problems related to data quantity (at the initial stages) and concept drift (at the latter stages). Albeit the considered setting is being tested offline, we can interpret the search itself as a dynamic process -- commonly, a single surrogate type is considered throughout the search, however, is this the preferred strategy? We explored whether dynamic re-configuration of surrogate models \emph{during the optimization} is a sensible, sample-efficient hyperparameter optimization strategy suitable for \textbf{large-scale} model configuration search. \begin{figure \centering \Description[Workflow depicting DSS.]{A cyclic workflow showing how the models are iteratively refined/switched.} \includegraphics[width=0.95\linewidth]{scheme.png} \caption{Conceptual overview of \textbf{DSS}, part of the in-house AutoML.} \label{fig:scheme} \end{figure} \begin{figure}[t!] \centering \Description[Two figures showing DSS's performance]{Left image shows iterative model switching in action, right one shows benchmark performance against strong baselines.} \begin{subfigure}[b]{0.45\textwidth} \centering \includegraphics[width=\textwidth, height=4.3cm]{surrogateDynamic.png} \caption{Visualizing Dynamic Surrogate Switching -- the surrogates change as the optimization progresses.} \end{subfigure} \hfill \begin{subfigure}[b]{0.45\textwidth} \centering \includegraphics[width=\textwidth, height=4cm]{benchmark2.png} \caption{Benchmark results -- offline tests -- CTR task (Relative information gain~\cite{he2014practical}) with fixed stopping.} \end{subfigure} \caption{Visualization of DSS benchmark experiment and an ablation. The goal was to identify whether surrogate switching is a feasible strategy to obtain good model configurations. The benchmark results on a private data set consisting of hundreds of millions of instances (target was click-through rate -- CTR) are shown in Figure b). Results indicate DSS is a promising strategy candidate for sample-efficient hyperparameter optimization, requiring minimal human involvement during configuration.} \label{fig:three graphs} \end{figure} \section{Dynamic Surrogate Switching (DSS)} \label{sec:method} Surrogate models learn to estimate configuration quality during hyperparameter optimization. Changing surrogates during the search itself was previously shown to have a positive effect on the search's efficiency~\cite{Mehmani2018}. This work builds on similar ideas, extending them beyond common function optimization benchmark sets to operate as a part of the internal \emph{AutoML framework} that requires handling of \textbf{hundreds of millions} of instances during the search for suitable FM configurations. This section describes the key components of the presented method and the initial study of their behaviour. We refer to the proposed method as Dynamic Surrogate Switching (DSS). A conceptual overview of the proposed method is shown in Figure~\ref{fig:scheme}. The approach follows the paradigm of surrogate-based modeling extended by the idea of surrogate switching. We continue with \textbf{an overview of DSS}. As the first step, we will consider the evaluation of the learning engine (in our case, field-aware factorization machines (FFMs)~\cite{juan2016field}\footnote{Field aware factorization machines implementation available as \url{https://github.com/outbrain/fwumious_wabbit}.}. Initially, diverse parts of the search space are selected and considered by the learning engine to obtain the initial set of configuration-score tuples suitable for surrogate-based learning. The \textbf{evaluation data update} step is responsible for storing the configurations in a format suitable for learning. Further, this step also checks for anomalies in the evaluations (e.g., too many similar/same results). Note that the evaluation data set is constantly updated throughout the search, with all data considered each update (to maximize utilization of prior evaluations). The subsequent step of \textbf{surrogate selection} is what differentiates DSS from conventional surrogate-based learning the most -- the DSS is based on the previous work~\cite{Mehmani2018} and the ideas of OAML, aimed to enable dynamic surrogate re-configuration. We implemented it by considering a collection of different model types and their initial configurations. The considered models that can serve as a surrogate are, for example, Random Forests~\cite{breiman2001random}, Gaussian Processes~\cite{rasmussen2003gaussian} and Gradient Boosting Machines~\cite{friedman2001greedy}~\footnote{Implemented with components from~\cite{varoquaux2015scikit}.}. As the evaluation of such surrogates is inexpensive (and thus not the bottleneck of the whole search), the surrogate selection phase considers multiple parametrizations of the mentioned model types -- overall, hundreds of surrogates are evaluated each iteration. The next issue we tackled was how to \emph{score} the surrogates. We perform the selection based on the \emph{proportion of explained variance}. Thus, the surrogate model scoring (including ranking) can be formulated as: $\argsort_{s \in S} \frac{\textrm{Var} (y - s(D))}{\textrm{Var} (y)},$ \noindent where $S$ is the set of possible surrogate models and $D$ the current set of configuration-score tuples obtained by the learning engine during search. The highest-ranked surrogate is selected and used during the acquisition step for a given search iteration. An important component of surrogate-based optimization is \textbf{configuration generation}. Solution candidates are scored in mini-batches to tackle the memory overhead. We further augmented the random configuration search with a memory structure that discards parts of the space considered previously. Once the surrogate is selected, re-trained on the current data and the candidate space is selected, \textbf{solution scoring} takes place. Here, the surrogate is used to provide a score for each generated configuration. There are multiple possible ways of utilizing the obtained scores; e.g., in Bayesian search, probability-of-improvement or similar heuristics can be considered. As the considered DSS does not always offer probabilistic outputs, simpler acquisition functions are required. We exploit the fact that the AutoML system that implements DSS runs in \textbf{multi-threaded} mode, enabling us to devote some threads to top-ranked solutions obtained by the surrogate while leaving some threads to perform exploration based on more randomized configurations. Once selected, the learning engine considers the configurations, and the feedback loop continues. Example results are summarized in Figure~\ref{fig:three graphs}. Overall, the DSS was identified as a promising approach to tuning FFMs. \section{Conclusions} \label{sec:conclusions} The paper presents Dynamic Surrogate Switching (DSS), an approach aimed to facilitate hyperparameter search -- one of the components of the in-house AutoML utilized daily by many data scientists. The results indicate that dynamic surrogates indeed offer a promising alternative to existing strong baselines (e.g., Random Forest-based surrogates), are potentially more adaptive to particular problems considered, and require less human intervention during configuration. Further work includes testing the idea on different data sets and studying its behaviour when considering different learning engines. \section*{AUTHOR BIO} \label{sec:bio} Bla\v{z} \v{S}krlj is a machine learning researcher active in the areas of AutoML and representation learning. During his PhD at the Jo\v{z}ef Stefan International Postgraduate School he worked on low-resource AutoML and its applications to natural language processing, graph-based machine learning and compressibility of latent representations. Currently, Bla\v{z} is part of Outbrain’s AutoML team, where he is exploring the limits of AutoML for large-scale recommendation. \bibliographystyle{ACM-Reference-Format}

fe5cc626fbdb7e5c81bf6f9bbff2efe50d382df3

\section{Introduction} In this paper we provide a unified way for proving H\"older regularity for the gradient of viscosity solutions to fully nonlinear elliptic equations of the form \begin{align} \label{me} \begin{split} \Phi(x,|Du|) F(D^{2}u) = f(x) \quad\text{in}\quad B_{1}, \end{split} \end{align} where $B_{1}\equiv B_{1}(0)\subset \mathbb{R}^{n}$ with $n\geqslant 2$ is the unit ball, $F : \mathcal{S}(n)\rightarrow \mathbb{R}$ is a uniformly $(\lambda,\Lambda)$-elliptic operator in the sense of \ref{a1} and $\Phi : B_{1}\times [0,\infty)\rightarrow [0,\infty)$ is a continuous map featuring a degeneracy and singularity for the gradient described as in \ref{a2}. From a variational point of view, the fully nonlinear equation \eqref{me} is closely related to the energy functional \begin{align} \label{func} v\mapsto \int\limits_{B_{1}} \varphi(x,|Dv|)\,dx \end{align} for a integral density $\varphi : B_{1}\times [0,\infty)\rightarrow [0,\infty)$ in a way that the Euler-Lagrange equation corresposnding to the functional \eqref{func} forms an equation of type \eqref{me}. The functional in \eqref{func} is a highly general non-autonomous functional with Uhlenbeck structure including significant models such as $p-$, Orlicz-, $p(x)-$ and double phase- growth and so on. H\"older continuity for the gradient of local minima of the functional \eqref{func} under suitable optimal assumptions has been investigated in \cite{HO}, where fundamental assumptions on the integral density function $\varphi$ in \eqref{func} are that there exist constants $1<p,q$ such that the map $t\mapsto \frac{\varphi(x,t)}{t^{p}}$ is almost non-decreasing and the map $t\mapsto \frac{\varphi(x,t)}{t^{q}}$ is almost non-increasing, see \cite[Definition 3.1]{HO}. In this regard, our conditions on $\Phi$ in \eqref{me} to be introduced in \ref{a2} are absolutely reasonable. Let us present known regularity results for viscosity solutions of equations in the form of \eqref{me} as significant special cases of our problem. \begin{enumerate} \item[1.] For $\Phi(x,t) = t^{p}$ with $i(\Phi) = s(\Phi)=p>-1$ in condition \ref{a2}, fully nonlinear equations \eqref{me} with this type of $\Phi(x,t)$ have been studied in a series of papers. The authors of \cite{BD1} proved the comparison principle and Liouville type theorems in the singular case $(-1<p<0)$, and showed the regularity and uniqueness of the first eigenfunction in \cite{BD2}. Alexandrov-Bakelman-Pucci estimates and the Harnack inequality have been also obtained in \cite{DFQ1,DFQ2,Im1}. In particular, the authors of \cite{IS1} proved local H\"older continuity for the gradient of viscosity solutions of \eqref{me} in the degenerate case ($p\geqslant 0$). Moreover, the authors of \cite{ART1} proved the optimality of H\"older regularity for the gradient of viscosity solutions for the same problem in \cite{IS1} by showing that viscosity solutions are $C_{\loc}^{1,\beta}$ with $\beta=\min\left\{\bar{\alpha}, \frac{1}{p+1}\right\}$, where $\bar{\alpha}\in (0,1)$ is the H\"older exponent coming from the Krylov-Safonov regularity for the homogeneous equation $F(D^2h)=0$. \item[2.] For $\Phi(x,t) = t^{p} + a(x)t^{q}$ with $-1<p, q$ and $0\leqslant a(\cdot)\in C(B_{1})$, the constants in \ref{a2} can be determined as $i(\Phi)= \min\{p,q\} $ and $s(\Phi)= \max\{p,q\}$. The author of \cite{De1} proved the local $C^{1,\beta}-$regularity of viscosity solutions of \eqref{me} for $0\leqslant p \leqslant q$. Moreover, in this degenerate case, the sharpness of the local $C^{1,\beta}$-regularity estimates for bounded viscosity solutions is shown in \cite{SR1}. \item[3.] For $\Phi(x,t) = t^{p(x)}$ with $p(\cdot)\in C(B_{1})$, $i(\Phi)=\inf\limits_{x\in B_{1}}p(x)>-1$ and $s(\Phi) = \sup\limits_{x\in B_{1}}p(x)$ in \ref{a2}, $C^{1,\beta}$-regularity of viscosity solutions has been studied in \cite{BPRT1}. In this paper, we provide a novel way to prove H\"older continuity for the gradient of viscosity solutions of \eqref{me} for both degenerate/singular cases in the full generality. \item[4.] For $\Phi(x,t)= t^{p(x)} + a(x) t^{q(x)}$ with functions $0\leqslant a(\cdot)\in C(B_{1})$ and $-1< p(\cdot), q(\cdot)$ in $C(B_{1})$, the constants in \ref{a2} are $i(\Phi)=\inf\limits_{x\in B_{1}}\{p(x),q(x)\} $ and $ s(\Phi)= \sup\limits_{x\in B_{1}} \{p(x),q(x)\}$. In \cite{FRZ1}, local H\"older continuity for the gradient has been proved when $0\leqslant p(\cdot)\leqslant q(\cdot)$. \end{enumerate} For a variational point of these special cases we have discussed above, we refer to the recent survey paper \cite{MR1} presenting important results in problems with nonstandard growth conditions. We also point out the very recent paper \cite{Je1} dealing with viscosity solutions of an equation of the form \begin{align} \label{eq:va} |Du|^{\beta(x,u,Du)}F(D^2u) = f(x)\quad\text{in}\quad B_{1}, \end{align} where $\beta : B_{1}\times\mathbb{R}\times \mathbb{R}^{n}\rightarrow \mathbb{R}$ is a map satisfying $0<\beta _{m}\leqslant \beta(\cdot)\leqslant \beta_{M}$ for some positive constants $\beta_{m}$ and $\beta_{M}$. In \cite{Je1}, local H\"older continuity for the gradient of viscosity solutions of \eqref{eq:va} is obtained under general conditions on the exponent function $\beta(\cdot)$ for the degenerate case, while the singular case is not be treated due to the methods employed there and the equation \eqref{me} can not be represented as \eqref{eq:va} in general. Finally, let us recall a consequence of the classical Krylov-Safonov Harnack inequality, see \cite{CC1}, that viscosity solutions to the homogeneous equation \begin{align} \label{eq:hom} F(D^2h) = 0 \quad\text{in}\quad B_{1}, \end{align} under the assumption that $F : \mathcal{S}(n)\rightarrow \mathbb{R}$ satisfies \ref{a1}, are locally of class $C^{1,\bar{\alpha}}(B_{1})$ for a universal constant $\bar{\alpha}\equiv \bar{\alpha}(n,\lambda,\Lambda)\in (0,1)$ with the estimate \begin{align} \label{alpha} \norm{h}_{C^{1,\bar{\alpha}}(B_{1/2})} \leqslant c \norm{h}_{L^{\infty}(B_{1})} \end{align} for some constant $c\equiv c(n,\lambda,\Lambda)$. The main results of this paper read as follows. \begin{thm}[H\"older continuity of the gradient] \label{thm:mthm} Let $u\in C(B_{1})$ be a viscosity solution of \eqref{me} under the assumptions \ref{a1}-\ref{a3}. Then $u\in C^{1,\beta}_{\loc}(B_{1})$ for all $\beta>0$ satisfying \begin{align} \label{thm:mthm:1} \beta < \begin{cases} \min\left\{\bar{\alpha}, \frac{1}{1+ s(\Phi)}\right\} & \mbox{if } i(\Phi)\geqslant 0,\\ \min\left\{\bar{\alpha}, \frac{1}{1+ s(\Phi)-i(\Phi)}\right\} & \mbox{if } -1<i(\Phi)<0, \end{cases} \end{align} where $\bar{\alpha}$ is given in \eqref{alpha}. Moreover, for every $\beta$ in \eqref{thm:mthm:1}, there exists a constant $c\equiv c(n,\lambda,\Lambda,i(\Phi),L,\beta)$ such that \begin{align} \label{thm:mthm:2} \norm{u}_{L^{\infty}(B_{1/2})} + \sup\limits_{x\neq y\in B_{1/2}} \frac{|Du(x)-Du(y)|}{|x-y|^{\beta}} \leqslant c\left( 1 + \norm{u}_{L^{\infty}(B_{1})} + \norm{\frac{f}{\nu_0}}_{L^{\infty}(B_{1})}^{\frac{1}{1+ i(\Phi)}} \right). \end{align} \end{thm} The results of Theorem \ref{thm:mthm} are sharp in the view of an example given in \cite{IS1}. As we have discussed above, the results of Theorem \ref{thm:mthm} cover the main results of the papers \cite{BPRT1, FRZ1, De1, IS1} for both cases involving degenerate/singular terms in a unified way. Moreover, the results of Theorem \ref{thm:mthm} cover another important cases such as \begin{enumerate} \item[1.] $\Phi(x,t)=t^{p} + a(x)t^{p}\log(e+t)$ with $-1<p$ and $0\leqslant a(\cdot)\in C(B_{1})$, where the constants in \ref{a2} are given by $i(\Phi)=p$ and $s(\Phi)=p+\varepsilon$ for any $\varepsilon>0$, \item[2.] $\Phi(x,t) = \phi(t) + a(x)\psi(t) $ for suitable $N$-functions $\phi$, $\psi$ and $0\leqslant a(\cdot)\in C(B_{1})$. \end{enumerate} We again refer to \cite{MR1} for related variational problems. Finally, we outline the organization of the paper. In the next section we provide basic notations and assumptions to be used, and also smallness regime and basic regularity results. In Section \ref{sec3}, we prove basic regularity properties of viscosity solutions of \eqref{xi-eq} depending on $i(\Phi)$ and the size of the quantity $|\xi|$. Section \ref{sec4} is devoted to the approximation procedure for viscosity solutions of \eqref{me}. Finally, in last section we provide the proof of our main Theorem \ref{thm:mthm}. \section{Preliminaries} \label{sec2} \subsection{Notions and assumptions} \label{sec2-1} Throughout the paper, we denote by $B_{r}(x_0):= \{x\in \mathbb{R}^{n} : |x-x_0|< r\}$ the open ball of $\mathbb{R}^{n}$ with $n\geqslant 2$ centered at $x_0$ with positive radius $r$. If the center is clear in the context, we shall omit the center point by writing $B_{r}\equiv B_{r}(x_0)$. Also $B_{1}\equiv B_{1}(0)\subset \mathbb{R}^{n}$ denote the unit ball. We shall always denote by $c$ a generic positive constant, possible varying line to line, having dependecies on parameters using brackets, that is, for example $c\equiv c(n,i(\Phi),\nu_0)$ means that $c$ depends only on $n,i(\Phi)$ and $\nu_0$. For a measurable map $g : \mathcal{B}\subset B_{1}\rightarrow \mathbb{R}^{N}$ $(N\geqslant 1)$ with $\beta\in (0,1]$ being a given number, we shall use the notation \begin{align*} [g]_{C^{0,\beta}(\mathcal{B})}:= \sup\limits_{x\neq y\in \mathcal{B}} \frac{|g(x)-g(y)|}{|x-y|^{\beta}},\quad [g]_{C^{0,\beta}}:= [g]_{C^{0,\beta}(B_{1})}. \end{align*} Now we state the main assumptions in the paper. \begin{description} \item[(A1)\label{a1}] The operator $F : \mathcal{S}(n)\rightarrow \mathbb{R}$ in \eqref{me} is continuous and uniformly $(\lambda,\Lambda)$-elliptic in the sense that \begin{align*} \lambda \text{tr}(N) \leqslant F(M)-F(M+N) \leqslant \Lambda \text{tr}(N) \end{align*} holds with some constants $0<\lambda\leqslant \Lambda$, whenever $M,N\in \mathcal{S}(n)$ with $N\geqslant 0$, where we denote by $\mathcal{S}(n)$ to mean the set of $n\times n$ real symmetric matrices. \item[(A2)\label{a2}] $\Phi : B_{1}\times [0,\infty)\rightarrow [0,\infty) $ is a continuous map satisfying the following properties: \begin{enumerate} \item[1.] There exist constants $ s(\Phi)\geqslant i(\Phi)>-1$ such that the map $\displaystyle t\mapsto \frac{\Phi(x,t)}{t^{i(\Phi)}}$ is almost non-decreasing with constant $L\geqslant 1$ in $(0,\infty)$ and the map $\displaystyle t\mapsto \frac{\Phi(x,t)}{t^{s(\Phi)}}$ is almost non-increasing with constant $L\geqslant 1$ in $(0,\infty)$ for all $x\in B_{1}$. \item[2.] There exists constants $0<\nu_0\leqslant \nu_1$ such that $\displaystyle \nu_{0} \leqslant \Phi(x,1) \leqslant \nu_{1}$ for all $x\in B_{1}$. \end{enumerate} \item[(A3)\label{a3}] The term $f$ on the right hand side of \eqref{me} belongs to $C(B_{1})\cap L^{\infty}(B_{1})$. \end{description} The Pucci extremal operators $P_{\lambda,\Lambda}^{\pm} : \mathcal{S}(n)\rightarrow \mathbb{R}$ are defined as \begin{align*} P_{\lambda,\Lambda}^{+}(M):= -\lambda\sum\limits_{\lambda_{k}>0} \lambda_{k} - \Lambda\sum\limits_{\lambda_{k}<0}\lambda_{k} \end{align*} and \begin{align*} P_{\lambda,\Lambda}^{-}(M):= -\Lambda\sum\limits_{\lambda_{k}>0} \lambda_{k} -\lambda\sum\limits_{\lambda_{k}<0}\lambda_{k}, \end{align*} where $\{\lambda_{k}\}_{k=1}^{n}$ are the eigenvalues of the matrix $M$. The $(\lambda,\Lambda)$-ellipticity of the operator $F$ via the Pucci extremal operators can be formulated as \begin{align*} P_{\lambda,\Lambda}^{-}(N) \leqslant F(M+N)-F(M) \leqslant P_{\lambda,\Lambda}^{+}(N) \end{align*} for all $M,N\in \mathcal{S}(n)$. In what follows, for any vector $\xi\in\mathbb{R}^{n}$, we define a map $G_{\xi} : B_{1}\times \mathbb{R}^{n}\times \mathcal{S}(n)\rightarrow \mathbb{R}$ by \begin{align} \label{g-fun} G_{\xi}(x,p,M): = \Phi(x,|\xi+p|)F(M)-f(x) \end{align} under the assumptions prescribed in \ref{a1}-\ref{a3}. Then we shall focus on viscosity solutions of the equation \begin{align} \label{xi-eq} G_{\xi}(x,Du,D^2u)=0 \text{ in } B_{1}. \end{align} Now we give the definition of a viscosity solution $u$ of the equation \eqref{xi-eq} as follows. \begin{defn} A lower semicontinuous function $v$ is called a viscosity supersolution of \eqref{xi-eq} if for all $x_0\in B_{1}$ and $\varphi\in C^{2}(B_{1})$ such that $v-\varphi$ has a local minimum at $x_0$ and $D\varphi(x_0)\neq 0$, then \begin{align*} G_{\xi}(x_{0},D\varphi(x_0),D^{2}\varphi(x_0))\geqslant 0. \end{align*} An upper semicontinuous function $w$ is called is a viscosity subsolution of \eqref{xi-eq} if for all $x_0\in B_{1}$ and $\varphi\in C^{2}(B_{1})$ such that $w-\varphi$ has a local maximum at $x_0$ and $D\varphi(x_0)\neq 0$, there holds \begin{align*} G_{\xi}(x_{0},D\varphi(x_0),D^{2}\varphi(x_0))\leqslant 0. \end{align*} We say that $u\in C(B_{1})$ is a viscosity solution of \eqref{xi-eq} if $u$ is a viscosity supersolution and a subsolution simultaneously. \end{defn} Also we recall a concept of superjet and subjet introduced in \cite{CIL1}. \begin{defn} Let $v: B_{1}\rightarrow \mathbb{R}$ be an upper semicontinuous function and $w:B_{1}\rightarrow \mathbb{R}$ be a lower semicontinuous function. \begin{enumerate} \item[1.] A couple $(p,M)\in \mathbb{R}^{n}\times\mathcal{S}(n)$ is a superjet of $v$ at $x\in B_{1}$ if \begin{align*} v(x+y) \leqslant v(x) + \inner{p}{y} + \frac{1}{2}\inner{My}{y} + O(|y|^2). \end{align*} \item[2.] A couple $(p,M)\in \mathbb{R}^{n}\times \mathcal{S}(n)$ is a subjet of $w$ at $x\in B_{1}$ if \begin{align*} w(x+y)\geqslant w(x) + \inner{p}{y} + \frac{1}{2}\inner{My}{y} + O(|y|^2). \end{align*} \item[3.] A couple $(p,M)\in \mathbb{R}^{n}\times \mathcal{S}(n)$ is a limiting superjet of $v$ at $x\in B_{1}$ if there exists a sequence $\{x_k,p_{k},M_{k}\}\rightarrow \{x,p,M\}$ as $k\rightarrow \infty$ in a such way that $\{p_{k},M_{k}\}$ is a superjet of $v$ at $x_{k}$ and $\lim\limits_{k\to\infty}v(x_{k}) = v(x)$. \item[4.] A couple $(p,M)\in \mathbb{R}^{n}\times \mathcal{S}(n)$ is a limiting subjet of $w$ at $x\in B_{1}$ if there exists a sequence $\{x_k,p_{k},M_{k}\}\rightarrow \{x,p,M\}$ as $k\rightarrow \infty$ in such a way that $\{p_{k},M_{k}\}$ is a subjet of $v$ at $x_{k}$ and $\lim\limits_{k\to\infty}w(x_{k}) = w(x)$. \end{enumerate} \end{defn} \subsection{Small regime} \label{sec2-2} Here we verify that, for a viscosity solution $u$ of \eqref{xi-eq}, we are able to assume \begin{align} \label{small} \osc\limits_{B_{1}} u \leqslant 1 \quad\text{and}\quad \norm{f}_{L^{\infty}(B_{1})} \leqslant \varepsilon_{0} \end{align} for some constant $0<\varepsilon_{0}<1$ small enough, and also $\nu_0=\nu_1=1$ without loss of generality. In order to consider the problem in a small regime as in \eqref{small}, for a fixed ball $B_{R}(x_0)\subset B_{1}$, we define $\bar{u} : B_{1}\rightarrow \mathbb{R}$ by \begin{align} \label{sr:1} \bar{u}(x):= \frac{u(x_0+Rx)}{K} \end{align} for positive constants $K\geqslant 1\geqslant R$ to be determined later. It can be seen that $\bar{u}$ is a viscosity solution of \begin{align} \label{sr:2} \bar{G}_{\bar{\xi}}(x,D\bar{u},D^2\bar{u}):= \bar{\Phi}(x,|\bar{\xi}+D\bar{u}|)\bar{F}(D^{2}\bar{u})-\bar{f}(x) = 0, \end{align} where \begin{align*} \displaystyle \bar{\Phi}(x,t) &:= \frac{\Phi\left(x_0+Rx,\frac{K}{R}t\right)}{\Phi\left(x_0+Rx,\frac{K}{R}\right)}, \\ \bar{F}(M)&:= \frac{R^{2}}{K}F\left(\frac{K}{R^{2}}M \right), \\ \bar{f}(x) &:= \frac{R^{2}}{\Phi\left(x_0+Rx,\frac{K}{R}\right)K}f(x_0+Rx) \text{ and } \bar{\xi}:= \frac{R}{K}\xi. \end{align*} Note that $\bar{F}$ is still a uniformly $(\lambda,\Lambda)$-elliptic operator, the map $\displaystyle t\mapsto \frac{\bar{\Phi}(x,t)}{t^{i(\Phi)}}$ is almost non-decreasing and the map $\displaystyle t\mapsto \frac{\bar{\Phi}(x,t)}{t^{s(\Phi)}}$ is almost non-increasing with the same constants $L\geqslant 1$ and $s(\Phi)\geqslant i(\Phi)>-1$ as in \ref{a2}, and $\bar{\Phi}(x,1)=1$ for all $x\in B_{1}$. Moreover, the assumption \ref{a2} implies \begin{align*} \norm{\bar{f}}_{L^{\infty}(B_{1})} \leqslant \frac{L R^{2+i(\Phi)}}{\nu_{0} K^{1+i(\Phi)}}\norm{f}_{L^{\infty}(B_{1})} \leqslant \frac{L}{\nu_0}\norm{f}_{L^{\infty}(B_{1})}. \end{align*} By recalling $i(\Phi)> -1$ and setting \begin{align*} K:= 2\left(1+\norm{u}_{L^{\infty}(B_{1})} + \left[\frac{L}{\nu_0}\norm{f}_{L^{\infty}(B_{1})} \right]^{\frac{1}{1+i(\Phi)}}\right) \end{align*} and \begin{align*} R:= \varepsilon_{0}^{\frac{1}{2+i(\Phi)}}, \end{align*} we see that $\bar{u}$ solves the equation \eqref{sr:2} in the same class as \eqref{xi-eq} under the small regime in \eqref{small}. \subsection{Basic regularity results} \label{sec2-3} In this subsection, we state some basic regularity results for \eqref{xi-eq}. The first key tool to be employed later is the classical Ishii-Lions lemma, see \cite{CIL1}. \begin{lem}[Ishii-Lions Lemma] \label{lem_IL} Let $u$ be a viscosity solution of \eqref{xi-eq} with $\osc\limits_{B_{1}}u \leqslant 1$ and $\norm{f}_{L^{\infty}(B_{1})}\leqslant \varepsilon_{0}\ll 1$ under the assumptions \ref{a1}-\ref{a3}, where $\xi\in \mathbb{R}^{n}$ is any vector. Suppose that $\mathcal{B}\subset B_{1}$ is an open subset and $\psi \in C^{2}(\mathcal{B}\times\mathcal{B})$. Define a map $v : \mathcal{B}\times \mathcal{B}\rightarrow \mathbb{R}$ as \[ v(x,y):= u(x)-u(y). \] Suppose further $(\bar{x},\bar{y})\in \mathcal{B}\times \mathcal{B}$ is a local maximum point of $v-\psi$ in $\mathcal{B}\times\mathcal{B}$. Then, for each $\delta>0$, there exist matrices $X_{\delta},Y_{\delta}\in \mathcal{S}(n)$ such that \begin{align*} G_{\xi}(\bar{x},D_{x}\psi(\bar{x},\bar{y}),X_{\delta}) \leqslant 0 \leqslant G_{\xi}(\bar{y},-D_{y}\psi(\bar{x},\bar{y}),Y_{\delta}) \end{align*} and \begin{align*} -\left( \frac{1}{\delta} + \norm{A} \right)I \leqslant \begin{pmatrix} X_{\delta} & 0 \\ 0 & -Y_{\delta} \end{pmatrix} \leqslant A + \delta A^{2} \end{align*} with $A:= D^{2}\psi(\bar{x},\bar{y})$. \end{lem} Another important result to be applied afterwards is the results of \cite{IS2} in our settings. \begin{thm}[Imbert-Silvestre] \label{thm_IS} Let $u\in C(B_{1})$ be a viscosity solution to \eqref{xi-eq} for some $\xi\in \mathbb{R}^{n}$. Suppose there exists $\gamma>0$ such that \begin{enumerate} \item[1.] for all $(x,p)\in B_{1}\times \mathbb{R}^{n}$ with $|p|>\gamma$, it holds that \begin{align*} G_{\xi}(x,p,0) \leqslant c_{0}|p| \end{align*} for some constant $c_{0}>0$ and \item[2.] for any fixed $(x,p)\in B_{1}\times \mathbb{R}^{n}$ with $|p|>\gamma$, $G_{\xi}(x,p,M)$ is uniformly elliptic with respect to $M$. \end{enumerate} Then $u\in C_{\loc}^{0,\alpha}(B_{1})$ for some $\alpha\in (0,1)$. In particular, the following estimate \begin{align*} \norm{u}_{C^{0,\alpha}(B_{1/2})} \leqslant c\norm{u}_{L^{\infty}(B_{1})} \end{align*} holds true for some constant $c>0$. The constants $\alpha\in (0,1)$ and $c>0$ depending on $n$, the ellipticity constants and the parameter $\gamma>0$. \end{thm} \section{H\"older continuity} \label{sec3} In this section we provide H\"older regularity for solutions of \eqref{xi-eq}, where $\xi$ is any vector, under the small regime. \begin{lem}[H\"older continuity] \label{lem_dHC} Let $u$ be a viscosity solution of \eqref{xi-eq} under the assumptions \ref{a1}-\ref{a3} with $\osc\limits_{B_{1}}u \leqslant 1$, $\norm{f}_{L^{\infty}(B_{1})}\leqslant \varepsilon_{0}<1$ and $\nu_0 = \nu_1 =1$. Let $B_{R}\equiv B_{R}(x_0)\subset B_{1}$ be any ball. Then, we have the following regularity results: \begin{description} \item[(R1)\label{R1}] If $-1<i(\Phi)<0$ and $|\xi|=0$, then $u$ is Lipschitz continuous in $B_{R/2}$ with the estimate \begin{align} \label{lem_HC:1} [u]_{C^{0,1}(B_{R/2})} \leqslant C_{sl} \end{align} for some constant $C_{sl}\equiv C_{sl}(n,\lambda,\Lambda,i(\Phi),L,R)$. \item[(R2)\label{R2}] If $i(\Phi)\geqslant 0$ and $|\xi|> A_{0}$ with $A_{0}\equiv A_{0}(n,\lambda,\Lambda,i(\Phi),L,R)$, then $u$ is Lipschitz continuous in $B_{R/2}$ with the estimate \begin{align} \label{lem_HC:2} [u]_{C^{0,1}(B_{R/2})} \leqslant C_{dl} \end{align} for some constant $C_{dl}\equiv C_{dl}(n,\lambda,\Lambda,i(\Phi),L,R)$. \item[(R3)\label{R3}] If $i(\Phi)\geqslant 0$ and $|\xi|\leqslant A_0$, then $u\in C^{0,\beta}(B_{R/2})$ with the estimate \begin{align} \label{lem_HC:3} [u]_{C^{0,\beta}(B_{R/2})} \leqslant C_{ds}, \end{align} where $\beta\equiv \beta(n,\lambda,\Lambda, R, A_0)\in (0,1)$ and $C_{ds}\equiv C_{ds}(n,\lambda,\Lambda, R, A_0)$. \end{description} \end{lem} \begin{proof} For the proof of \ref{R1} and \ref{R2}, it suffices to show that there exist positive constants $L_1$ and $L_2$ such that \begin{align} \label{HC:1} \mathcal{L}:= \sup\limits_{x,y\in B_{R}} \left( u(x)-u(y)-L_1\omega(|x-y|) - L_{2}\left( |x-z_0|^2 + |y-z_0|^2 \right) \right)\leqslant 0 \end{align} for every $z_0\in B_{R/2}$, where \begin{align} \label{HC:omega} \omega(t) = \begin{cases} t-\omega_{0}t^{\frac{3}{2}} & \mbox{if } t \leqslant t_0:= \left(\frac{2}{3\omega_0}\right)^{2}, \\ \omega(t_0) & \mbox{if } t\geqslant t_0. \end{cases} \end{align} We choose $\omega_0\in (0,2/3)$ in such a way that $t_0\geqslant 1$. For instance, we take any constant $\omega_0 \leqslant 1/3$. By the contradiction, suppose that there are no such positive constants $L_1$ and $L_{2}$ satisfying \eqref{HC:1} for every $z_0\in B_{R/2}$. Then there exists a point $z_0\in B_{R/2}$ so that $\mathcal{L}>0$ for all numbers $L_1>0$ and $L_2>0$. Now we define two auxiliary functions $\phi, \psi : \overline{B_{R}}\times \overline{B_{R}}\rightarrow \mathbb{R}$ given by \begin{align} \label{HC:2} \psi(x,y):= L_1\omega(|x-y|) + L_{2}\left( |x-z_0|^2 + |y-z_0|^2 \right) \end{align} and \begin{align} \label{HC:3} \phi(x,y):= u(x)-u(y)-\psi(x,y). \end{align} Let $(\bar{x},\bar{y})\in \overline{B_{R}}\times \overline{B_{R}}$ be a maximum point for $\phi$. Then we have \begin{align*} \phi(\bar{x},\bar{y}) = \mathcal{L}>0 \end{align*} and \begin{align*} L_1\omega(|\bar{x}-\bar{y}|) + L_{2}\left( |\bar{x}-z_0|^2 + |\bar{y}-z_0|^2 \right) \leqslant \osc\limits_{B_{1}}u \leqslant 1. \end{align*} Now we select \begin{align*} L_2:= \frac{64}{R^2}. \end{align*} This choice of $L_2$ ensures \begin{align} \label{HC:7} |\bar{x}-z_0| + |\bar{y}-z_0| \leqslant \frac{R}{4} \quad\text{and}\quad |\bar{x}-\bar{y}|\leqslant \frac{R}{4}. \end{align} This means that the points $\bar{x}$ and $\bar{y}$ belong to the open ball $B_{R}$ and also we are able to assume that $\bar{x}\neq \bar{y}$; otherwise $\mathcal{L} \leqslant 0$ clearly. The rest of the proof is divided into several steps. \textbf{Step 1.} We are in a position to apply Lemma \ref{lem_IL} in order to ensure the existence of a limiting subjet $(\xi_{\bar{x}},X_{\delta})$ of $u$ at $\bar{x}$ and a limiting superjet $(\xi_{\bar{y}},Y_{\delta})$ of $u$ at $\bar{y}$, where \begin{align*} \xi_{\bar{x}}:= D_{x}\psi(\bar{x},\bar{y})= L_{1}\omega'(|\bar{x}-\bar{y}|)\frac{\bar{x}-\bar{y}}{|\bar{x}-\bar{y}|} + 2L_{2}(\bar{x}-z_0) \end{align*} and \begin{align*} \xi_{\bar{y}}:= -D_{y}\psi(\bar{x},\bar{y}) = L_{1}\omega'(|\bar{x}-\bar{y}|)\frac{\bar{x}-\bar{y}}{|\bar{x}-\bar{y}|} - 2L_{2}(\bar{y}-z_0), \end{align*} such that matrices $X_{\delta}$ and $Y_{\delta}$ satisfy the matrix inequality \begin{align} \label{HC:10} \begin{pmatrix} X_{\delta} & 0 \\ 0 & -Y_{\delta} \end{pmatrix} \leqslant \begin{pmatrix} Z & -Z \\ -Z & Z \end{pmatrix} + (2L_2 + \delta)I, \end{align} where \begin{align*} \begin{split} Z&:= L_1D^2(\omega(|\cdot|))(\bar{x}-\bar{y}) \\& = L_1\left[ \frac{\omega'(|\bar{x}-\bar{y}|)}{|\bar{x}-\bar{y}|}I + \left( \omega''(|\bar{x}-\bar{y}|)- \frac{\omega'(|\bar{x}-\bar{y}|)}{|\bar{x}-\bar{y}|} \right)\frac{(\bar{x}-\bar{y})\otimes (\bar{x}-\bar{y})}{|\bar{x}-\bar{y}|^{2}} \right] \end{split} \end{align*} and the constant $\delta>0$ only depends on the norm of $Z$, which can be selected sufficiently small. Applying the inequality \eqref{HC:10} for vectors of the form $(z,z)\in \mathbb{R}^{2n}$, we find \begin{align*} \inner{(X_{\delta}-Y_{\delta})z}{z} \leqslant (4L_2 + 2\delta)|z|^{2}. \end{align*} The last inequality yields that all the eigenvalues of the matrix $(X_{\delta}-Y_{\delta})$ are not larger than $4L_2 + 2\delta$. On the other hand, applying again \eqref{HC:10} for the vector $\bar{z}:= \left( \frac{\bar{x}-\bar{y}}{|\bar{x}-\bar{y}|}, \frac{\bar{y}-\bar{x}}{|\bar{x}-\bar{y}|} \right)$, we have \begin{align*} \begin{split} \inner{(X_{\delta}-Y_{\delta}) \frac{\bar{x}-\bar{y}}{|\bar{x}-\bar{y}|}}{ \frac{\bar{x}-\bar{y}}{|\bar{x}-\bar{y}|}} &\leqslant \left( 4L_2 + 2\delta + 4L_1\omega''(|\bar{x}-\bar{y}|)\right)\left|\frac{\bar{x}-\bar{y}}{|\bar{x}-\bar{y}|} \right|^2 \\& = \left( 4L_2 + 2\delta - \frac{6\omega_0 L_1}{|\bar{x}-\bar{y}|^{1/2}}\right)\left|\frac{\bar{x}-\bar{y}}{|\bar{x}-\bar{y}|} \right|^2 \\& \leqslant \left( 4L_2 + 2\delta - 6\omega_0 L_1\right)\left|\frac{\bar{x}-\bar{y}}{|\bar{x}-\bar{y}|} \right|^2, \end{split} \end{align*} where we have used the definition of $\omega$ in \eqref{HC:omega} together with $|\bar{x}-\bar{y}|\leqslant 1/4$ in \eqref{HC:7}. So at least one eigenvalue of $(X_{\delta}-Y_{\delta})$ is not larger than $4L_2 + 2\delta -6\omega_0 L_1$, where this quantity can be negative for large values of $L_{1}$. By the definition of the extremal Pucci operator, we see \begin{align*} \begin{split} P_{\lambda,\Lambda}^{-}(X_{\delta}-Y_{\delta}) &\geqslant -\lambda(4L_2 + 2\delta-6\omega_0 L_1)- \Lambda (n-1)(4L_2 + 2\delta) \\& \geqslant -(\lambda + (n-1)\Lambda)(4L_2 + 2\delta) + 6\omega_0\lambda L_1. \end{split} \end{align*} From two viscosity inequalities and the uniform ellipticity, we have \begin{align*} \Phi(\bar{x},|\xi + \xi_{\bar{x}}|)F(X_{\delta}) \leqslant f(\bar{x}),\quad \Phi(\bar{y},|\xi + \xi_{\bar{y}}|)F(Y_{\delta})\geqslant f(\bar{y}) \end{align*} and \begin{align*} F(X_{\delta})\geqslant F(Y_{\delta}) + P_{\lambda,\Lambda}^{-}(X_{\delta}-Y_{\delta}). \end{align*} Combining last three displays, we have \begin{align} \label{HC:18} \begin{split} 6\omega_0\lambda L_1 &\leqslant (\lambda + (n-1)\Lambda)(4L_2 + 2\delta) \\& \quad + \frac{f(\bar{x})}{\Phi(\bar{x},|\xi + \xi_{\bar{x}}|)} - \frac{f(\bar{y})}{\Phi(\bar{y},|\xi + \xi_{\bar{y}}|)}. \end{split} \end{align} At this stage, we shall separate it into several cases depending on the quantity of $|\xi|$ and the positiveness of $i(\Phi)$. \textbf{Step 2: Proof of \ref{R1}.} Suppose $-1<i(\Phi)<0$ and $\xi = 0$. By triangle inequality $\eqref{HC:7}_{2}$, we observe that \begin{align} \label{HC:19} |\xi_{\bar{x}}|\leqslant L_1(1+\frac{3}{2}\omega_0) + 2L_2 \leqslant \frac{7}{4} L_1 \end{align} and \begin{align} \label{HC:19_1} |\xi_{\bar{x}}| \geqslant L_{1}\left(1-\frac{3\omega_0}{2}|\bar{x}-\bar{y}|^{\frac{1}{2}}\right)-3L_2 \geqslant \frac{3L_1}{4}-3L_2 \geqslant 3L_2 \end{align} for all $L_1\geqslant 8L_2$. In the exactly same way, we see \begin{align} \label{HC:20} |\xi_{\bar{y}}|\leqslant \frac{7}{4}L_1 \quad\text{and}\quad |\xi_{\bar{y}}|\geqslant 2L_2 \end{align} for all $L_1\geqslant 8L_2$. Then we have \begin{align} \label{HC:20_1} \frac{f(\bar{x})}{\Phi(\bar{x},|\xi_{\bar{x}}|)} \leqslant c \frac{\norm{f}_{L^{\infty}(B_{1})}}{|\xi_{\bar{x}}|^{i(\Phi)}} \leqslant \frac{c}{L_1^{i(\Phi)}} \end{align} and \begin{align} \label{HC:20_2} \frac{-f(\bar{y})}{\Phi(\bar{y},|\xi_{\bar{y}}|)} \leqslant c \frac{\norm{f}_{L^{\infty}(B_{1})}}{|\xi_{\bar{y}}|^{i(\Phi)}} \leqslant \frac{c}{L_1^{i(\Phi)}} \end{align} for a constant $c\equiv c(i(\Phi),L)$. Using the last two displays in \eqref{HC:18}, we obtain \begin{align*} 6\omega_{0}\lambda L_1 \leqslant (\lambda + (n-1)\Lambda)(4L_2 + 2\delta) + \frac{c}{L_{1}^{i(\Phi)}} \end{align*} for a constant $c\equiv c(n,\lambda,\Lambda,i(\Phi),L,R)$. Recalling $-1<i(\Phi)<0$ and taking $L_1$ large enough, depending only on $n,\lambda,\Lambda,i(\Phi),L$ and $R$, we get a contradiction. Then the first part of the lemma is proved. \textbf{Step 3: Proof of \ref{R2}.} We suppose that $i(\Phi)\geqslant 0$ and $|\xi|>A_{0}$ for a constant $A_{0}$ to be determined in a moment. We set \begin{align} \label{HC:A0} A_0 := \frac{35 L_1}{2} \end{align} for $L_1>1$ to be selected soon. This choice of $A_0$ together with \eqref{HC:19} and \eqref{HC:20_1} leads to \begin{align*} |\xi + \xi_{\bar{x}}|\geqslant A_0 - \frac{A_0}{10} = \frac{9A_0}{10} \quad\text{and}\quad |\xi + \xi_{\bar{y}}|\geqslant \frac{9A_0}{10}. \end{align*} Therefore, we have \begin{align*} \frac{f(\bar{x})}{\Phi(\bar{x},|\xi + \xi_{\bar{x}}|)} \leqslant c \frac{\norm{f}_{L^{\infty}(B_{1})}}{|\xi + \xi_{\bar{x}}|^{i(\Phi)}} \leqslant \frac{c}{A_0^{i(\Phi)}} \end{align*} and \begin{align*} \frac{-f(\bar{y})}{\Phi(\bar{y},|\xi + \xi_{\bar{y}}|)} \leqslant c \frac{\norm{f}_{L^{\infty}(B_{1})}}{|\xi + \xi_{\bar{y}}|^{i(\Phi)}} \leqslant \frac{c}{A_{0}^{i(\Phi)}} \end{align*} for a constant $c\equiv c(i(\Phi),L)$. Again using the last two displays in \eqref{HC:18}, we obtain \begin{align*} 6\omega_{0}\lambda L_1 \leqslant (\lambda + (n-1)\Lambda)(4L_2 + 2\delta) + \frac{c}{L_{1}^{i(\Phi)}} \end{align*} for a constant $c\equiv c(n,\lambda,\Lambda,i(\Phi),L,R)$. By choosing $L_1$ large enough, depending only on $n,\lambda,\Lambda,i(\Phi),L$ and $R$, we have again a contradiction. Indeed, we have proved the second part of the lemma. \textbf{Step 4: Proof of \ref{R3}.} Finally, we shall focus on proving \ref{R3}. Suppose now $|\xi| \leqslant A_0$, where $A_0$ has been determined in \eqref{HC:A0}. We consider the operator \begin{align*} G_{\xi}(x,p,M):= \Phi(x,|\xi + p|)F(M)-f(x). \end{align*} In fact, $G_{\xi}(x,p,M)$ is uniformly elliptic, whenever $|p|>2A_0$. At this stage, we apply Theorem \ref{thm_IS} to conclude the last part of the Lemma. The proof is complete. \end{proof} \section{Approximation} \label{sec4} Now we prove a key approximation lemma, which plays a crucial role in later arguments. \begin{lem} \label{lem:AL} Let $u\in C(B_{1})$ be a viscosity solution of \eqref{xi-eq} with $\osc\limits_{B_{1}} \leqslant 1$, where $\xi\in \mathbb{R}^{n}$ is arbitrarily given. Suppose \ref{a1}-\ref{a3} hold true for $i(\Phi)\geqslant 0$ and $\nu_0=\nu_1=1$. Then, for any $\mu>0$, there exists a constant $\delta\equiv \delta(n,\lambda,\Lambda,i(\Phi),L,\mu)$ such that if \begin{align} \label{lem:AL:1} \norm{f}_{L^{\infty}(B_{1})} \leqslant \delta, \end{align} then one can find $h\in C^{1,\bar{\alpha}}(B_{3/4})$ with the estimate $\norm{h}_{C^{1,\bar{\alpha}}(B_{3/4})} \leqslant c\equiv c(n,\lambda,\Lambda)$, for some $0<\bar{\alpha}<1$, satisfying \begin{align} \label{lem:AL:2} \norm{u-h}_{L^{\infty}(B_{1/2})} \leqslant \mu. \end{align} \end{lem} \begin{proof} By contradiction, we suppose the conclusion of the lemma fails. Then there exist $\mu_0>0$ and sequences of $\{F_{k}\}_{k=1}^{\infty}$, $\{\Phi_{k}\}_{k=1}^{\infty}$, $\{f_{k}\}_{k=1}^{\infty}$, and $\{u_{k}\}_{k=1}^{\infty}$ and a sequence of vectors $\{\xi_{k}\}_{k=1}^{\infty}$ such that \begin{description} \item[(C1) \label{c1}] $F_{k} \in C(\mathcal{S}(n), \mathbb{R})$ is uniformly $(\lambda,\Lambda)$-elliptic, \item[(C2) \label{c2}] $\Phi_{k} \in C( B_{1}\times [0,\infty), [0,\infty))$ such that the map $t\mapsto \frac{\Phi_{k}(x,t)}{t^{i(\Phi)}}$ is almost non-decreasing and the map $t\mapsto \frac{\Phi(x,t)}{t^{s(\Phi)}}$ is almost non-increasing with constant $L\geqslant 1$, and $ \Phi_{k}(x,1) = 1$ for all $x\in B_{1}$, \item[(C3)\label{c3}] $f_{k}\in C(B_{1})$ with $\norm{f_{k}}_{L^{\infty}(B_{1})} \leqslant \frac{1}{k}$ and \item[(C4)\label{c4}] $u_{k}\in C(B_{1})$ with $ \osc_{B_{1}}u_k \leqslant 1$ solves the equation \begin{align} \label{AL:0} \Phi_{k}(x,|\xi_{k}+Du_{k}|)F_{k}(D^2u_k) = f_{k}(x), \end{align} \end{description} but \begin{align} \label{AL:1} \sup\limits_{x\in B_{1/2}}|u_{k}(x)-h(x)|> \mu_0 \end{align} for all $h\in C^{1,\bar{\alpha}}(B_{3/4})$ and every $0<\bar{\alpha}<1$. The condition \ref{c1} implies that $F_{k}$ converges to some uniformly $(\lambda,\Lambda)$-elliptic operator $F_{\infty}\in C(\mathcal{S}(n),\mathbb{R})$. Applying Lemma \ref{lem_dHC}, $u_{k}\in C^{0,\beta}_{\loc}(B_{1})\cap C(B_{1})$ for some $\beta\in (0,1)$. Using \eqref{lem_HC:2}, \eqref{lem_HC:3} and Arzela-Ascoli theorem, we have that the sequence $\{u_{k}\}_{k=1}^{\infty}$ converges to a function $u_{\infty}$ locally uniformly in $B_{1}$. In particular, there holds that \begin{align} \label{AL:2} u_{\infty}\in C(B_{1})\quad\text{and}\quad \osc\limits_{B_{1}} u_{\infty} \leqslant 1. \end{align} Now we prove that the limiting function $u_{\infty}$ is a viscosity solution of the homogeneous equation \begin{align} \label{AL:3} F_{\infty}(D^{2}u_{\infty}) = 0 \quad\text{in}\quad B_{3/4}. \end{align} For this, first we verify that $u_{\infty}$ is a viscosity supersolution. Let \begin{align*} p(x):= \frac{1}{2}\inner{M(x-y)}{x-y} + \inner{b}{x-y} + u_{\infty}(y) \end{align*} be a quadratic polynomial touching $u_{\infty}$ from below at a point $y\in B_{3/4}$. Without loss of generality, let us assume $|y|=u_{\infty}(y) = 0$. Then there exists a sequence $x_{k}\rightarrow 0$ as $k\rightarrow \infty$ such that $u_{k}-\varphi$ has a local minimum at $x_{k}$. Observe that $D\varphi(x_k)\rightarrow b$ and $D^{2}\varphi(x_k)\rightarrow M$. Since $u_{k}$ is a viscosity solution of \eqref{AL:0}, we have \begin{align} \label{AL:6} \Phi_{k}(x_k,|\xi_{k} + D\varphi(x_k)|)F_{k}(D^{2}\varphi(x_k)) \geqslant f_{k}(x_k). \end{align} For the ease of presentation, from now on we shall consider several cases depending on the boundedness of sequence $\{\xi_{k}\}_{k=1}^{\infty}$. \textbf{Case 1: Sequence $\{\xi_{k}\}_{k=1}^{\infty}$ is unbounded.} In this case, we can assume $|\xi_{k}|\rightarrow \infty$ (up to a subsequence). As a consequence, we can show (up to a subsequence) that \begin{align} \label{AL:7} |\xi_{k} + D\varphi(x_k)|\leqslant |\xi_{k}|-|D\varphi(x_k)| \geqslant |\xi_{k}| - (|b|+1)\geqslant 1, \end{align} which implies that \begin{align*} F_{\infty}(M) &= \lim\limits_{k\to\infty} F_{k}(D^2\varphi(x_k)) \geqslant \lim\limits_{k\to \infty} \frac{f_{k}(x_k)}{\Phi_{k}(x_k,|\xi_{k} +D\varphi(x_k)|)} \notag\\& \geqslant -\lim\limits_{k\to\infty} \frac{L}{k|\xi_k + D\varphi(x_k)|^{i(\Phi)}} = 0, \end{align*} where we have used \ref{c2} and \eqref{AL:6}. \textbf{Case 2: Sequence $\{\xi_{k}\}_{k=1}^{\infty}$ bounded } In the case we may assume $\xi_{k}\rightarrow \xi_{\infty}$ (up to a subsequence). Therefore, for the case $|\xi_{\infty}+b|\neq 0$, in the exactly same way as in \eqref{AL:7}, we infer that $F_{\infty}(M)\geqslant 0$. Then we focus on the case $|\xi_{\infty}+b|=0$. There are two possibilities as $|b|=|\xi_{\infty}| = 0$ or $b=-\xi_{\infty}$ with $|b|, |\xi_{\infty}|>0$. In those scenarios, we prove that $F_{\infty}(M)\geqslant 0$. By contradiction suppose \begin{align} \label{AL:9} F_{\infty}(M)<0. \end{align} From the uniformly ellipticity condition of $F_{\infty}$, the matrix $M$ has at least one positive eigenvalue. Let $\mathbb{R}^{n}= E\oplus Q$, where $E = \text{span}\{e_1,\ldots,e_{m}\}$ is the space consisting of those eigenvectors corresponding to positive eigenvalues of $M$. \textbf{Case 3: $b= -\xi_{\infty}$ with $|b|,|\xi_{\infty}|>0$.} Let $\gamma>0$ and set \begin{align*} p_{\gamma}(x):= p(x) + \gamma |P_{E}(x)| = \frac{1}{2}\inner{Mx}{x} + \inner{b}{x} + \gamma |P_{E}(x)|, \end{align*} where $P_{E}$ stands for the orthogonal projection on $E$. Since $u_{k}\rightarrow u_{\infty}$ locally uniformly in $B_{1}$ and $p(x)$ touches $u_{\infty}(x)$ from below at the origin, for $\gamma$ small enough, $p_{\gamma}(x)$ touches $u_{k}(x)$ from below at a point $x_{k}^{\gamma}\in B_{r}$ ($B_{r}$ is a small neighborhood of the origin). Moreover, there holds that $x_{k}^{\gamma}\rightarrow x_{\infty}^{\gamma}$ for some $x_{\infty}^{\gamma}$ as $k\rightarrow \infty$. At this point we consider two scenarios: $P_{E}(x_{k}^{\gamma})=0$ for all $k\in\mathbb{N}$ (up to a subsequence) or $P_{E}(x_{k}^{\gamma})\neq 0$ for all $k\in\mathbb{N}$ (up to a subsequence). \textit{Scenario 1: $P_{E}(x_{k}^{\gamma})=0$ for all $k\in\mathbb{N}$ (up to a subsequence).} In this scenario, first we note that \begin{align*} \bar{p}_{\gamma}(x):= \frac{1}{2}\inner{Mx}{x} + \inner{b}{x} + \gamma\inner{e}{P_{E}(x)} \end{align*} touches $u_{k}$ from below at $x_{k}^{\gamma}$ for every $e\in \mathbb{S}^{n-1}$. A straightforward computation gives us \begin{align*} D\bar{p}_{\gamma}(x_{k}^{\gamma}) = Mx_{k}^{\gamma} + b + \gamma P_{E}(e) \quad\text{and}\quad D^2\bar{p}_{\gamma}(x_k^{\gamma}) = M. \end{align*} Now we select $e\in E\cap \mathbb{S}^{n-1}$ such that $P_{E}(e)=e$. Therefore, by $u_k$ being a viscosity solution of \eqref{AL:0}, we see \begin{align*} \Phi_{k}(x_{k}^{\gamma}, |\xi_{k} + Mx_{k}^{\gamma} + b + \gamma e|)F_{k}(M) \geqslant f_{k}(x_{k}^{\gamma}). \end{align*} We also notice that if $Mx_{\infty}^{\gamma}=0$, then for $k$ enough large, we have \begin{align*} |\xi_{k} + Mx_{k}^{\gamma} + b| \leqslant \gamma/2 \quad\text{and}\quad 3\gamma/2 \geqslant |\xi_{k} + Mx_{k}^{\gamma} + b + \gamma e| \geqslant \gamma/2. \end{align*} Therefore, combining the last two displays and using \ref{c2} together with $\gamma\ll 1$, we have \begin{align*} \begin{split} F_{k}(M) & \geqslant \frac{f_{k}(x_{k}^{\gamma})}{\Phi_{k}(x_{k}^{\gamma},|\xi_{k} + Mx_{k}^{\gamma} + b + \gamma e| )} \\& \geqslant \frac{-L|f_{k}(x_{k}^{\gamma})|}{ |\xi_{k} + Mx_{k}^{\gamma} + b + \gamma e|^{s(\Phi)}} \geqslant -\frac{L}{k} \left(\frac{2}{\gamma} \right)^{s(\Phi)}. \end{split} \end{align*} Letting $k\rightarrow \infty$ in the last display, we obtain $F_{\infty}(M)\geqslant 0$. In the situation $|Mx_{\infty}^{\gamma}|>0$, we first look at the subcase $E\equiv \mathbb{R}^{n}$ and choose $e\in \mathbb{S}^{n-1}$ such that \begin{align*} |Mx_{\infty}^{\gamma} + \gamma P_{E}(e)| = |Mx_{\infty}^{\gamma} + \gamma e|>0. \end{align*} Therefore, for $k$ large enough, we have \begin{align} \label{AL:16} |Mx_{k}^{\gamma} + \gamma e|\geqslant \frac{1}{2}|Mx_{\infty}^{\gamma} + \gamma e| > 0 \quad\text{and}\quad |\xi_{k} + b| \leqslant \frac{1}{8}|Mx_{\infty}^{\gamma} + \gamma e|. \end{align} On the other hand, if $E\not\equiv \mathbb{R}^{n}$, then we can find $e\in \mathbb{S}^{n-1} \cap E^{\perp}$ so that \begin{align*} |Mx_{\infty}^{\gamma} + \gamma P_{E}(e)| = |Mx_{\infty}^{\gamma}|>0. \end{align*} Again for $k$ large enough, we have \begin{align} \label{AL:16_1} |Mx_{k}^{\gamma}|\geqslant \frac{1}{2}|Mx_{\infty}^{\gamma}| \quad\text{and}\quad |\xi_{k} + b| \leqslant \frac{1}{8}|Mx_{\infty}^{\gamma}|. \end{align} As a consequence, using either \eqref{AL:16} or \eqref{AL:16_1}, we see \begin{align*} |\xi_{k} + Mx_{k}^{\gamma} + b + \gamma P_{E}(e)| > \frac{1}{4}|Mx_{\infty}^{\gamma} + \gamma P_{E}(e)| > 0. \end{align*} Again applying \ref{c2} and taking into account the last display, we have \begin{align*} \begin{split} F_{k}(M) &\geqslant \frac{f_{k}(x_{k}^{\gamma})}{\Phi_{k}(x_{k}^{\gamma},|\xi_{k} + Mx_{k}^{\gamma} + b + \gamma P_{E}(e)| )} \\& \geqslant -\left(\frac{L}{ |\xi_{k} + Mx_{k}^{\gamma} + b + \gamma P_{E}(e)|^{i(\Phi)}} + \frac{L}{ |\xi_{k} + Mx_{k}^{\gamma} + b + \gamma P_{E}(e)|^{s(\Phi)}} \right)|f_{k}(x_{k}^{\gamma})| \\& \geqslant \frac{-L 4^{s(\Phi)}}{k}\left(\frac{1}{|Mx_{\infty}^{\gamma} + \gamma P_{E}(e)|^{i(\Phi)}} + \frac{1}{|Mx_{\infty}^{\gamma} + \gamma P_{E}(e)|^{s(\Phi)}} \right). \end{split} \end{align*} Again letting $k\rightarrow \infty$ in the last display, we again arrive at $F_{\infty}(M) \geqslant 0$. \textit{Scenario 2: $P_{E}(x_{k}^{\gamma})\neq 0$ for all $k\in \mathbb{N}$ (up to a subsequence).} In this scenario, we note that $P_{E}(x)$ is smooth and convex in a small neighborhood of $x_{k}^{\gamma}$. Let us denote \begin{align*} \zeta_{k}^{\gamma}:= \frac{P_{E}(x_{k}^{\gamma})}{|P_{E}(x_{k}^{\gamma})|}. \end{align*} A direct computation yields \begin{align*} D(|P_{E}(\cdot)|)(x_{k}^{\gamma}) = \zeta_{k}^{\gamma} \quad\text{and}\quad D^{2}(P_{E}(|\cdot|))(x_{k}^{\gamma}) = \frac{1}{|P_{E}(x_k^{\gamma})|}\left( I-\zeta_{k}^{\gamma}\otimes \zeta_{k}^{\gamma} \right). \end{align*} Hence, with $u_{k}$ being a viscosity solution of \eqref{AL:0}, we have the following viscosity inequality \begin{align*} \Phi_{k}(x_{k}^{\gamma}, |\xi_{k} + Mx_{k}^{\gamma} + b + \gamma \zeta_{k}^{\gamma}|) F_{k}\left(M + \frac{1}{|P_{E}(x_k^{\gamma})|}\left( I-\zeta_{k}^{\gamma}\otimes \zeta_{k}^{\gamma} \right)\right) \geqslant f_{k}(x_{k}^{\gamma}). \end{align*} Observing that $|\zeta_{k}^{\gamma}|=1$ and letting $e:= \zeta_{k}^{\gamma}$, we can perform the same procedure as in the first scenario of $P_{E}(x_{k}^{\gamma})=0$ by considering the cases of $Mx_{\infty}^{\gamma}=0$ and $Mx_{\infty}^{\gamma}\neq 0$. Finally, we conclude that $F_{\infty}(M)\geqslant 0$ when $b=-\xi_{\infty}\neq 0$, which contradicts to \eqref{AL:9}. \textbf{Case 4: $b=\xi_{\infty} = 0$.} In fact, this case is much easier to handle. Since $\frac{1}{2}\inner{Mx}{x}$ touches $u_{\infty}(x)$ from below at the origin and $u_{k}\rightarrow u_{\infty}$ locally uniformly, the function \begin{align*} \hat{p}_{\gamma}(x) := \frac{1}{2}\inner{Mx}{x} + \gamma |P_{E}(x)| \end{align*} touches $u_{k}$ from below at a point $\hat{x}_{k}^{\gamma}\in B_{r}$ ($B_{r}$ is a small neighborhood of the origin) for $\gamma > 0$ sufficiently small. Again the sequence $\{\hat{x}_{k}^{\gamma}\}$ is uniformly bounded. As in \textbf{Case 3}, we analyze those two scenarios $P_{E}(\hat{x}_{k}^{\gamma})=0$ for all $k\in \mathbb{N}$ (up to a subsequence) and $P_{E}(\hat{x}_{k}^{\gamma})\neq 0$ for all $k\in \mathbb{N}$ (up to a subsequence). All in all, we conclude $F_{\infty}(M)\geqslant 0$ in this case. Finally, taking into account all cases we have analyzed above, we have shown that $u_{\infty}$ is a viscosity supersolution of \eqref{AL:3}. In order to prove that $u_{\infty}$ is a viscosity subsolution of \eqref{AL:3}, we show that $-u_{\infty}$ is a viscosity supersolution of $\hat{F}_{\infty}(D^2h) = 0$, where $\hat{F}_{\infty}(M)= -F_{\infty}(-M)$ is uniformly $(\lambda,\Lambda)$-elliptic operator as well. Therefore, $u_{\infty}$ is a viscosity solution of \eqref{AL:3}. From the regularity results of \cite[Chap. 5]{CC1}, we see $u_{\infty}\in C^{1,\bar{\alpha}}_{\loc}(B_{3/4})$ for some $\bar{\alpha}\in (0,1)$. Moreover, $\norm{u_{\infty}}_{C^{1,\bar{\alpha}}(B_{1/2})} \leqslant c\equiv c(n,\lambda,\Lambda)$ via \eqref{AL:2}. So choosing $h:= u_{\infty}$ in \eqref{AL:1}, we have a contradiction. The proof is complete. \end{proof} \section{Proof of Theorem \ref{thm:mthm}} \label{sec5} Now we provide a proof of Theorem \ref{thm:mthm}. Let $u\in C(B_{1})$ be a viscosity solution with $\osc\limits_{B_{1}}u \leqslant 1$, $\norm{f}_{L^{\infty}(B_{1})}\leqslant \delta \ll 1$ for a constant $\delta\equiv \delta(n,\lambda,\Lambda,i(\Phi),L)$ to be determined in a moment and $\nu_0 = \nu_1 = 1$. The proof is divided into two main parts, where in the first part we shall deal with the case $i(\Phi)\geqslant 0$ and the remaining case $-1<i(\Phi)<0$ will be investigated in the second part. \textbf{Part 1: $i(\Phi)\geqslant 0$.} Let us first fix a point $y\in B_{1/2}$ and an exponent with \begin{align} \label{mthm:1} 0< \beta < \min\left\{\bar{\alpha}, \frac{1}{1+s(\Phi)} \right\}. \end{align} We prove that there exist universal constants $0<r \ll 1$, $C_{0}>1$ and a sequence of affine functions \begin{align} \label{mthm:2} l_{k}(x):= a_k + \inner{b_{k}}{x}, \end{align} where $\{a\}_{k=1}^{\infty}\subset \mathbb{R}$ and $\{b_{k}\}_{k=1}^{\infty}\subset \mathbb{R}^{n}$, such that for every $k\in \mathbb{N}$: \begin{description} \item[(E1)\label{e1}] $\sup\limits_{x\in B_{r^{k}}(y)} |u(x)-l_{k}(x)| \leqslant r^{k(1+\beta)}$, \item[(E2)\label{e2}] $|a_{k}-a_{k-1}| \leqslant C_{0}r^{(k-1)(1+\beta)}$ and \item[(E3)\label{e3}] $|b_{k}-b_{k-1}| \leqslant C_{0}r^{(k-1)\beta}$. \end{description} We show these estimates by mathematical induction. For the simplicity, we divide the proof into several steps. \textbf{Step 1. The basis of induction.} Without loss of generality we can assume $y=0$ by translating $x\mapsto y + \frac{1}{2}x$. Let us set \begin{align*} l_1(x):= h(0) + \inner{Dh(0)}{x}, \end{align*} where $h$ is the approximation function coming from Lemma \ref{lem:AL} for a certain constant $\mu>0$ to be determined in a few lines. Then there exists a constant $C_0\equiv C_0(n,\lambda,\Lambda)>1$ such that \begin{align*} \norm{h}_{C^{1,\bar{\alpha}}(B_{3/8})} \leqslant C_0 \quad\text{and}\quad \sup\limits_{x\in B_{r}} |h(x)-l_1(x)| \leqslant C_0r^{1+\bar{\alpha}} \end{align*} for every $r\leqslant 3/8$. The triangle inequality yields \begin{align*} \sup\limits_{x\in B_{r}} |u(x)-l_1(x)| \leqslant \mu + C_0r^{1+\bar{\alpha}}. \end{align*} We first select a universal constant $0<r\ll 1$ satisfying \begin{align} \label{mthm:6} r^{\beta} \leqslant \frac{1}{2},\quad C_{0}r^{1+\bar{\alpha}} \leqslant \frac{1}{2}r^{1+\beta} \quad\text{and}\quad r^{1-\beta(1+s(\Phi))} \leqslant 1, \end{align} which is possible by \eqref{mthm:1}. In a sequel, we select a constant $\mu>0$ as \begin{align} \label{mthm:7} \mu:= \frac{1}{2}r^{1+\beta}, \end{align} which fixes an arbitrary constant $\mu>0$ in Lemma \ref{lem:AL}. In turn, there exists a constant $\delta\equiv \delta(n,\lambda,\Lambda,i(\Phi),L,\beta)$ verifying the smallness assumption $\norm{f}_{L^{\infty}(B_{1})}\leqslant \delta$, but such a smallness assumption can be assumed without loss of generality. Therefore, to conclude this step we set \begin{align*} a_0:= 0,\quad a_1:= h(0),\quad b_0 = 0\quad\text{and}\quad b_1:= Dh(0). \end{align*} These choices with \eqref{mthm:6} and \eqref{mthm:7} verify that the estimates \ref{e1}-\ref{e3} are satisfied for $k=1$. \textbf{Step 2: Induction process.} Now we suppose that the hypotheses of the induction have been established for $k=1,2,\ldots,m$ for $m\geqslant 1$. We show that the estimates \ref{e1}-\ref{e3} hold true for $k=m+1$. For this, we introduce an auxiliary function as \begin{align*} w_m(x):= \frac{u(r^{m}x)- l_{m}(r^{m}x)}{r^{m(1+\beta)}}. \end{align*} We note that $w_{m}$ solves the following equation in the viscosity sense \begin{align*} \Phi_{m}(x, |r^{-m\beta}b_{m} + Dw_{m}|)F_{m}(D^{2}w_{m}) = f_{m}(x), \end{align*} where \begin{align*} F_{m}(M):= r^{m(1-\beta)}F(r^{(\beta-1)m}M), \end{align*} which is uniformly $(\lambda,\Lambda)$-operator, the function \begin{align*} \Phi_{m}(x,t):= \frac{\Phi(r^{m}x,r^{m\beta}t)}{\Phi(r^{m}x,r^{m\beta})}\quad (x\in B_{1}, t>0) \end{align*} still satisfies the properties that the map $t\mapsto \frac{\Phi_{m}(x,t)}{t^{i(\Phi)}}$ is almost non-decreasing, the map $t\mapsto \frac{\Phi_{m}(x,t)}{t^{s(\Phi)}}$ is almost non-increasing with the same constant $L\geqslant 1$ and $\Phi_{m}(x,1)=1$ for all $x\in B_{1}$, and \begin{align*} f_{m}(x):= \frac{r^{m(1-\beta)} f(r^{m}x)}{\Phi(r^{m}x,r^{m\beta})}. \end{align*} Using \ref{a2} and \eqref{mthm:1}, we notice that \begin{align*} \norm{f_{m}}_{L^{\infty}(B_{1})} \leqslant \frac{Lr^{m(1-\beta)}\norm{f}_{L^{\infty}(B_{1})}}{ r^{m\beta s(\Phi)}} \leqslant L\delta r^{m(1-(1+s(\Phi))\beta)} \leqslant L\delta. \end{align*} Therefore, we are in a position to apply Lemma \ref{lem:AL} to $w_{m}$. In turn, there exists a function $\bar{h}\in C^{1,\bar{\alpha}}(B_{3/4})$ such that \begin{align*} \sup\limits_{x\in B_{r}} |w_{m}(x)-\bar{h}(x)| \leqslant \mu. \end{align*} Arguing as in \textbf{Step 1}, we show that \begin{align*} \sup\limits_{x\in B_{r}} |w_{m}(x)-\bar{l}(x)| \leqslant r^{1+\beta}, \end{align*} where \begin{align*} \bar{l}(x):= \bar{a} + \inner{\bar{b}}{x}\quad \text{for some}\quad \bar{a}\in \mathbb{R} \text{ and } \bar{b}\in \mathbb{R}^{n}. \end{align*} Denoting \begin{align*} l_{m+1}:= l_{m}(x) + r^{m(1+\beta)}\bar{l}(r^{-m}x), \end{align*} we see \begin{align*} \sup\limits_{x\in B_{r^{m+1}}} |u(x)-l_{m+1}(x)|\leqslant r^{(m+1)(1+\beta)} \end{align*} and \begin{align*} |a_{m+1}-a_{m}| + r^{m}|b_{m+1}-b_{m}| \leqslant C_0r^{m(1+\beta)}. \end{align*} Therefore, the $(m+1)$-th step of the induction is complete. \textbf{Step 3: Conclusion.} Once we have the existence of universal constants $0<r\ll 1$, $C_0>1$ and a sequence of affine functions in \eqref{mthm:2} verifying the estimates \ref{e1}-\ref{e3}, the remaining part of the proof is very standard, see for instance \cite{IS1,De1}. Therefore, the proof of \eqref{thm:mthm:2} is complete when $i(\Phi)\geqslant 0$. \textbf{Part 2: $-1<i(\Phi)<0$.} Now we shall with the case of $-1< i(\Phi)<0$. Again we fix a point $y\in B_{1/2}$. Without loss of generality, we may assume $y=0$ by using the translation $x\mapsto y + \frac{1}{2}x$. Now we apply \ref{R1} of Lemma \ref{lem_dHC} in order to ensure that \begin{align} \label{mthm:21} [u]_{C^{0,1}(B_{3/4})} \leqslant C_{sl} \end{align} for a constant $C_{sl}\equiv C_{sl}(n,\lambda,\Lambda,i(\Phi),L)$. Therefore, it can be seen that $u$ is a viscosity solution of the equation \begin{align*} \tilde{\Phi}(x,|Dv|)F(D^2v) = \tilde{f}(x)\quad \text{in}\quad B_{3/4}, \end{align*} where \begin{align*} \tilde{\Phi}(x,t):= t^{-i(\Phi)}\Phi(x,t)\quad (x\in B_{1}, t>0), \end{align*} which satisfies the properties that the map $t\mapsto \tilde{\Phi}(x,t)$ is almost non-increasing, the map $t\mapsto \frac{\tilde{\Phi}(x,t)}{t^{s(\Phi)-i(\Phi)}}$ is almost non-increasing with constant $L\geqslant 1$, $\tilde{\Phi}(x,1)=1$ for all $x\in B_{1}$, and \begin{align*} \tilde{f}(x) = |Du(x)|^{-i(\Phi)}f(x). \end{align*} Using the estimate \eqref{mthm:21} together with $\norm{f}_{L^{\infty}B_{1}} \leqslant \delta \ll 1$, we see \begin{align*} \norm{\tilde{f}}_{L^{\infty}(B_{3/4})} \leqslant C_{sl}^{-i(\Phi)}\delta. \end{align*} So we are able to apply \textbf{Part 1} of the proof in order to have \ref{e1}-\ref{e3}. This means that we have the estimate \eqref{thm:mthm:2} for $-1<i(\Phi)<0$. The proof is complete. \vspace{0.2cm} \bibliographystyle{amsplain}

085645139fb67978533c9f8393a2fab32b34ee26

\section{Introduction} Offline reinforcement learning seeks to solve decision-making problems without interacting with the environment. This is compelling because online data collection can be dangerous or expensive in many realistic tasks. However, relying entirely on a static dataset imposes new challenges. One is that policy evaluation is hard because the mismatch between the behavior and the learned policy usually introduces extrapolation error \citep{bcq}. In most offline tasks, it is difficult or even impossible for the collected transitions to cover the whole state-action space. When evaluating the current policy via dynamic programming, leveraging actions that are not presented in the dataset (out-of-sample) may lead to highly unreliable results, and thus performance degrade. Consequently, in offline RL it is critical to stay close to the behavior policy during training. Recent advances in model-free offline methods mainly include two lines of work. The first is the adaptation of existing off-policy algorithms. These methods usually include value pessimism about unseen actions or regulations of feasible action space \citep{bcq, bear, cql}. The other line of work \citep{awr, crr, awac} is derived from constrained policy search and mainly trains a parameterized policy via weighted regression. Evaluations of every state-action pair in the dataset are used as regression weights. The main motivation behind weighted policy regression is that it helps prevent querying out-of-sample actions \citep{awac,iql}. However, we find that this argument is untenable in certain settings. Our key observation is that policy models in existing weighted policy regression methods are usually unimodal Gaussian models and thus lack distributional expressivity, while in the real world collected behaviors can be highly diverse. This distributional discrepancy might eventually lead to selecting unseen actions. For instance, given a bimodal target distribution, fitting it with a unimodal distribution unavoidably results in covering the low-density area between two peaks. In Section \ref{motivation}, we empirically show that lack of policy expressivity may lead to performance degrade. Ideally, this problem could be solved by switching to a more expressive distribution class. However, it is nontrivial in practice since weighted regression requires exact and derivable density calculation, which places restrictions on distribution classes that we can choose from. Especially, we may not know what the behavior or optimal policy looks like in advance. To overcome the limited expressivity problem, we propose to decouple the learned policy into two parts: an expressive generative behavior model and an action evaluation model. Such decoupling avoids explicitly learning a policy model whose target distribution is difficult to sample from, whereas learning a behavior model is much easier because sampling from the behavior policy is straightforward given the offline dataset collected by itself. Access to data samples from the target distribution is critical because it allows us to leverage existing advances in generative methods to model diverse behaviors. To sample from the learned policy, we use importance sampling to select actions from candidates proposed by the behavior model with the importance weights computed by the action evaluation model, which we refer to as \textbf{S}electing \textbf{f}rom \textbf{B}ehavior \textbf{C}andidates (\textbf{SfBC}). The fidelity of the learned behavior model is critical in our method because it directly determines the feasible action space. While covering any low-density area increases the possibility of selecting unseen actions during training, failing to cover all action modes in the dataset results in overly restricted action space. To fulfill this requirement, we propose to learn from diverse behaviors using diffusion probabilistic models \citep{diffusion}, which have recently achieved great success in modeling diverse image distributions, outperforming other existing generative models \citep{diffusion_beat_gan}. We also propose a planning-based operator for Q-learning, which performs implicit planning strictly within dataset trajectories based on the current policy, and is provably convergent. The planning scheme greatly reduces bootstrapping steps required for dynamic programming and thus can help to further reduce extrapolation error and increase computational efficiency. The main contributions of this paper are threefold: 1. We address the problem of limited policy expressivity in conventional methods by decoupling policy learning into behavior learning and action evaluation, which allows the policy to inherit distributional expressivity from a diffusion-based behavior model. 2. The learned policy is further combined with an implicit in-sample planning technique to suppress extrapolation error and assist dynamic programming over long horizons. 3. Extensive experiments demonstrate that our method achieves competitive or superior performance compared with state-of-the-art offline RL methods, especially in sparse-reward tasks such as AntMaze. \section{Background} \subsection{Constrained Policy Search in Offline RL} Consider a Markov Decision Process (MDP), described by a tuple $\langle\mathcal{S},\mathcal{A},P,r,\gamma\rangle$. $\mathcal{S}$ denotes the state space and $\mathcal{A}$ is the action space. $P({\bm{s}}'|{\bm{s}},{\bm{a}})$ and $r({\bm{s}}, {\bm{a}})$ respectively represent the transition and reward functions, and $\gamma \in (0,1]$ is the discount factor. Our goal is to maximize the expected discounted return $J(\pi) = \mathbb{E}_{{\bm{s}} \sim \rho_\pi({\bm{s}})}\mathbb{E}_{{\bm{a}} \sim \pi(\cdot|{\bm{s}})}\left[r({\bm{s}}, {\bm{a}})\right]$ of policy $\pi$, where $\rho_\pi({\bm{s}}) = \sum_{n=0}^\infty \gamma^n p_\pi({\bm{s}}_n = {\bm{s}})$ is the discounted state visitation frequencies induced by the policy $\pi$ \citep{rlbook}. According to the \textit{policy gradient theorem} \citep{PG}, given a parameterized policy $\pi_\theta$, and the policy's state-action function $Q^\pi$, the gradient of $J(\pi_\theta)$ can be derived as: \begin{equation} \label{Eq:objective} \nabla_\theta J(\pi_\theta) = \int_\mathcal{S} \rho_\pi({\bm{s}}) \int_\mathcal{A} \nabla_\theta \pi_\theta({\bm{a}} | {\bm{s}}) Q^\pi({\bm{s}}, {\bm{a}}). \end{equation} When online data collection from policy $\pi$ is not possible, it is difficult to estimate $\rho_\pi({\bm{s}})$ in \Eqref{Eq:objective}, and thus the expected value of the Q-function $\eta(\pi_\theta) := \int_\mathcal{S} \rho_\pi({\bm{s}}) \int_\mathcal{A} \pi_\theta({\bm{a}} | {\bm{s}}) Q^\pi({\bm{s}}, {\bm{a}})$. Given a static dataset $\mathcal{D}^\mu$ consisting of multiple trajectories $\{\left({\bm{s}}_n, {\bm{a}}_n, r_n \right)\}$ collected by a behavior policy $\mu({\bm{a}}|{\bm{s}})$, previous off-policy methods \citep{dpg, ddpg} estimate $\eta(\pi_\theta)$ with a surrogate objective $\hat{\eta}(\pi_\theta)$ by replacing $\rho_\pi({\bm{s}})$ with $\rho_\mu({\bm{s}})$. In offline settings, due to the importance of sticking with the behavior policy, prior works \citep{awr, awac} explicitly constrain the learned policy $\pi$ to be similar to $\mu$, while maximizing the expected value of the Q-functions: \begin{equation} \mathop{\mathrm{arg \ max}}_{\pi} \quad \int_\mathcal{S} \rho_\mu({\bm{s}}) \int_\mathcal{A} \pi({\bm{a}} | {\bm{s}}) Q_\phi({\bm{s}}, {\bm{a}}) \ d{\bm{a}} \ d{\bm{s}} - \frac{1}{\alpha}\int_\mathcal{S} \rho_\mu({\bm{s}}) D_{\mathrm{KL}} \left(\pi(\cdot |{\bm{s}}) || \mu(\cdot |{\bm{s}}) \right) d{\bm{s}}. \label{Eq:rl_main} \end{equation} The first term in \Eqref{Eq:rl_main} corresponds to the surrogate objective $\hat{\eta}(\pi_\theta)$, where $Q_\phi({\bm{s}}, {\bm{a}})$ is a learned Q-function of the current policy $\pi$. The second term is a regularization term to constrain the learned policy within support of the dataset $\mathcal{D}^\mu$ with $\alpha$ being the coefficient. \subsection{Policy Improvement via Weighted Regression} \label{Sec:weighted_regression} The optimal policy $\pi^*$ for \Eqref{Eq:rl_main} can be derived \citep{rwr, awr, awac} by use of Lagrange multiplier: \begin{align} \pi^*({\bm{a}}|{\bm{s}}) &= \frac{1}{Z({\bm{s}})} \ \mu({\bm{a}}|{\bm{s}}) \ \mathrm{exp}\left(\alpha Q_\phi({\bm{s}}, {\bm{a}}) \right), \label{Eq:pi_optimal} \end{align} where $Z({\bm{s}})$ is the partition function. \Eqref{Eq:pi_optimal} forms a policy improvement step. Directly sampling from $\pi^*$ requires explicitly modeling behavior $\mu$, which itself is challenging in continuous action-space domains since $\mu$ can be very diverse. Prior methods \citep{awr, crr, bail} bypass this issue by projecting $\pi^*$ onto a parameterized policy $\pi_\theta$: \begin{align} & \mathop{\mathrm{arg \ min}}_{\theta} \quad \mathbb{E}_{{\bm{s}} \sim \mathcal{D}^\mu} \left[ D_{\mathrm{KL}} \left(\pi^*(\cdot | {\bm{s}}) \middle|\middle| \pi_\theta(\cdot | {\bm{s}})\right) \right]\nonumber \\ = & \mathop{\mathrm{arg \ max}}_{\theta} \quad \mathbb{E}_{({\bm{s}}, {\bm{a}}) \sim \mathcal{D}^\mu} \left[ \mathrm{log} \ \pi_\theta({\bm{a}} | {\bm{s}}) \ \mathrm{exp}\left(\alpha Q_\phi({\bm{s}}, {\bm{a}}) \right) \right]. \label{Eq:wr} \end{align} Such method is usually referred to as weighted regression, with $\mathrm{exp}\left(\alpha Q_\phi({\bm{s}}, {\bm{a}})\right)$ being the regression weights. Although weighted regression avoids the need to explicitly model the behavior policy, it requires calculating the exact density function $\pi_\theta({\bm{a}} | {\bm{s}})$ as in \Eqref{Eq:wr}. This constrains the policy $\pi_\theta$ to distribution classes that have a tractable expression for the density function. We find this in practice limits the model expressivity and could be suboptimal in some cases (See Section \ref{motivation}). \subsection{Diffusion Probabilistic Model} \label{Sec:diffusion_bg} Diffusion models \citep{sohl2015deep,diffusion,sde} are generative models by firstly defining a forward process to gradually add noise to an unknown data distribution $p_0({\bm{x}}_0)$ and then learning to reverse it. The forward process $\{ {\bm{x}}(t) \}_{t\in [0, T]}$ is defined by a stochastic differential equation (SDE) $d{\bm{x}}_t = f({\bm{x}}_t, t) \mathrm{d}t + g(t) \mathrm{d} {\bm{w}}_t$, where ${\bm{w}}_t$ is a standard Brownian motion and $f(t)$, $g(t)$ are hand-crafted functions \citep{sde} such that the transition distribution $p_{t0}({\bm{x}}_t|{\bm{x}}_0)=\mathcal{N}({\bm{x}}_t|\alpha_t{\bm{x}}_0, \sigma_t^2\bm{I})$ for some $\alpha_t,\sigma_t>0$ and $p_T({\bm{x}}_T)\approx \mathcal{N}({\bm{x}}_T|0,\bm{I})$. To reverse the forward process, diffusion models define a scored-based model ${\bm{s}}_\theta$ and optimize the parameter $\theta$ by: \begin{equation} \mathop{\mathrm{arg \ min}}_{\theta} \quad \mathbb{E}_{t,{\bm{x}}_0,\bm{\epsilon}}[\| \sigma_t \mathbf{s}_\theta({\bm{x}}_t, t) + \bm{\epsilon} \|_2^2], \end{equation} where $t\sim\mathcal{U}(0,T)$, ${\bm{x}}_0\sim p_0({\bm{x}}_0)$, $\bm{\epsilon}\sim \mathcal{N}(0,\bm{I})$, ${\bm{x}}_t=\alpha_t{\bm{x}}_0+\sigma_t\bm{\epsilon}$. Sampling by diffusion models can be alternatively viewed as discretizing the diffusion ODEs~\citep{sde}, which are generally faster than discretizing the diffusion SDEs~\citep{song2020denoising,lu2022dpm}. Specifically, the sampling procedure needs to firstly sample a pure Gaussian ${\bm{x}}_T\sim\mathcal{N}(0,\bm{I})$, and then solve the following ODE from time $T$ to time $0$ by numerical ODE solvers: \begin{equation} d {\bm{x}}_t = \bigg[f(t){\bm{x}}_t - \frac{1}{2}g^2(t) {\bm{s}}_\theta({\bm{x}}_t,t)\bigg] \mathrm{d}t. \label{Eq:prob_ode} \end{equation} Then the final solution ${\bm{x}}_0$ at time $0$ is the sample from the diffusion models. \section{Method} \label{Method} We propose a Selecting-from-Behavior-Candidates (SfBC) approach to address the limited expressivity problem in offline RL. Below we first motivate our method by highlighting the importance of a distributionally expressive policy in learning from diverse behaviors. Then we derive a high-level solution to this problem from a generative modeling perspective. \subsection{Learning from Diverse Behaviors} \label{motivation} In this section, we show that the weighted regression broadly used in previous works might limit the distributional expressivity of the policy and lead to performance degrade. As described in Section \ref{Sec:weighted_regression}, conventional policy regression methods project the optimal policy $\pi^*$ in \Eqref{Eq:pi_optimal} onto a parameterized policy set. In continuous action-space domains, the projected policy is usually limited to a narrow range of unimodal distributions (e.g., squashed Gaussian), whereas the behavior policy could be highly diverse (e.g., multimodal). Lack of expressivity directly prevents the RL agent from exactly mimicking a diverse behavior policy. This could eventually lead to sampling undesirable out-of-sample actions during policy evaluation and thus large extrapolation error. Even if Q-values can be accurately estimated, an inappropriate unimodal assumption about the optimal policy might still lead to failure in extracting a policy that might have multiple similarly rewarding but distinctive action choices. \begin{wrapfigure}{r}{0.555\textwidth}{ \vskip -0.45cm \sbox{\measurebox}{% \begin{minipage}[b]{.27\textwidth} \centering \vskip 0.2cm \subfloat{\label{fig:figB}\hspace{-0.1cm}\includegraphics[width=0.95\textwidth]{pics/toydrawleft.drawio.pdf}} \vfill \vskip 0.4cm \subfloat{\label{fig:figC}\includegraphics[width=\textwidth]{pics/toypolicyillustration.pdf}} \end{minipage} } \usebox{\measurebox} \begin{minipage}[b][\ht\measurebox][s]{.27\textwidth} \vskip 0.1cm \centering \subfloat{\label{fig:figA}\includegraphics[width=\textwidth]{pics/toyresultcomparison.pdf}} \end{minipage} \caption{Illustration of the Bidirectional-Car task and comparison between SfBC and unimodal policies. See Section \ref{lfdb} for experimental details.} \label{fig:illustration} \vskip -0.3cm } \end{wrapfigure} We design a simple task named Bidirectional Car to better explain this point. Consider an environment where a car placed in the middle of two endpoints can go either side to gain the final reward. If an RL agent finds turning left and right similarly rewarding, by incorrectly assuming a unimodal distribution of the behavior policy, it ends up staying put instead of taking either one of the optimal actions (Figure \ref{fig:illustration}). As a result, unimodal policies fail to completely solve this task or loss diversity while a more distributionally expressive policy easily succeeds. We therefore deduce that distributional expressivity is a necessity to enable diverse behavior learning. To better model the complex behavior policy, we need more powerful generative modeling for the policy distribution, instead of the simple and unimodal Gaussians. \subsection{Selecting from Behavior Candidates} In this section, we provide a generative view of how to model a potentially diverse policy. Specifically, in order to model $\pi^*$ with powerful generative models, essentially we need to perform maximum likelihood estimation for the model policy $\pi_\theta$, which is equivalent to minimizing KL divergence between the optimal and model policy: \begin{equation} \mathop{\mathrm{arg \ max}}_{\theta} \quad \mathbb{E}_{{\bm{s}} \sim \mathcal{D}^\mu} \mathbb{E}_{a \sim \pi^*(\cdot | s)}\left[\log \pi_\theta(a|s) \right] \ \Leftrightarrow \ \mathop{\mathrm{arg \ min}}_{\theta} \quad \mathbb{E}_{{\bm{s}} \sim \mathcal{D}^\mu} \left[ D_{\mathrm{KL}} \left(\pi^*(\cdot | {\bm{s}}) \middle|\middle| \pi_{\theta}(\cdot | {\bm{s}})\right) \right]. \end{equation} However, drawing samples directly from $\pi^*$ is difficult, so previous methods \citep{awr, awac, crr} rely on the weighted regression as described in \Eqref{Eq:wr}. The main reason that limits the expressivity of $\pi_\theta$ is the need of calculating exact and derivable density function $\pi_\theta({\bm{a}} | {\bm{s}})$ in policy regression, which places restrictions on distribution classes that we can choose from. Also, we might not know what the behavior or optimal policy looks like previously. Our solution is based on a key observation that directly parameterizing the policy $\pi$ is not necessary. To better model a diverse policy, we propose to decouple the learning of $\pi$ into two parts. Specifically, we leverage \Eqref{Eq:pi_optimal} to form a policy improvement step: \begin{equation} \label{Eq:decouple} \pi({\bm{a}}|{\bm{s}}) \propto \mu_\theta({\bm{a}}|{\bm{s}}) \ \mathrm{exp}\left(\alpha Q_\phi({\bm{s}},{\bm{a}}) \right). \end{equation} One insight of the equation above is that minimizing KL divergence between $\mu$ and $\mu_\theta$ is much easier compared with directly learning $\pi_\theta$ because sampling from $\mu$ is straightforward given $D^\mu$. This allows to us to leverage most existing advances in generative modeling (Section \ref{dbc}). $Q_\phi(s,a)$ could be learned using the existing Q-learning framework (Section \ref{imp}). The inverse temperature parameter $\alpha$ in \Eqref{Eq:decouple} serves as a trade-off between conservative and greedy improvement. We can see that when $\alpha \to 0$, the learned policy falls back to the behavior policy, and when $\alpha \to +\infty$ the learned policy becomes a greedy policy. To sample actions from $\pi$, we use an importance sampling technique. Specifically, for any state ${\bm{s}}$, first we draw $M$ action samples from a learned behavior policy $\mu_\theta(\cdot|{\bm{s}})$ as candidates. Then we evaluate these action candidates with a learned critic $Q_\phi$. Finally an action is resampled from $M$ candidates with $\mathrm{exp}\left(\alpha Q_\phi({\bm{s}},{\bm{a}}) \right)$ being the sampling weights. We summarize this procedure as selecting from behavior candidates (SfBC), which could be understood as an analogue to rejection sampling. The main difference between our method and previous works is that we do not seek to fit a parameterized model $\pi_\theta$ to $\pi^*$. Although generative modeling of the behavior policy has been explored by several works \citep{bcq, bear}, it was mostly used to form an explicit distributional constraint for the policy model $\pi_\theta$. In contrast, we show directly leveraging the learned behavior model to generate actions is not only feasible but beneficial on the premise that high-fidelity behavior modeling can be achieved. We give a practical implementation in the next section. \section{Practical Implementation} In this section, we derive a practical implementation of SfBC, which includes diffusion-based behavior modeling and planning-based Q-learning. An algorithm overview is given in Appendix \ref{overview}. \subsection{Diffusion-based behavior modeling} \label{dbc} It is critical that the learned behavior model is of high fidelity because generating any out-of-sample actions would result in unwanted extrapolation error, while failing to cover all in-sample actions would restrict feasible action space for the policy. This requirement brings severe challenges to existing behavior modeling methods, which mainly include using Gaussians or VAEs. Gaussian models suffer from limited expressivity as we have discussed in Section \ref{motivation}. VAEs, on the other hand, need to introduce a variational posterior distribution to optimize the model distribution, which has a trade-off between the expressivity and the tractability~\citep{kingma2016improved, lucas2019understanding}. This still limits the expressivity of the model distribution. An empirical study is given in Section \ref{Sec:ablation}. To address this problem, we propose to learn from diverse behaviors using diffusion models \citep{diffusion}, which have recently achieved great success in modeling diverse image distributions \citep{dalle2, imagen}, outperforming other generative models \citep{diffusion_beat_gan}. Specifically, we follow \citet{sde} and learn a state-conditioned diffusion model $s_\theta$ to predict the time-dependent noise added to the action ${\bm{a}}$ sampled from the behavior policy $\mu(\cdot|{\bm{s}})$: \begin{equation} \theta = \mathop{\mathrm{arg \ min}}_{\theta} \quad \mathbb{E}_{({\bm{s}}, {\bm{a}}) \sim D^\mu,\bm{\epsilon}, t}[\| \sigma_t \mathbf{s}_\theta(\alpha_t{\bm{a}}+\sigma_t\bm{\epsilon}, {\bm{s}}, t) + \bm{\epsilon} \|_2^2], \end{equation} where $\bm{\epsilon}\sim \mathcal{N}(0,\bm{I})$, $t\sim\mathcal{U}(0,T)$. $\alpha_t$ and $\sigma_t$ are determined by the forward diffusion process. Intuitively $s_\theta$ is trained to denoise ${\bm{a}}_t := \alpha_t {\bm{a}} + \sigma_t \bm{\epsilon}$ into the unperturbed action ${\bm{a}}$ such that ${\bm{a}}_T \sim \mathcal{N}(0,\bm{I})$ can be transformed into ${\bm{a}} \sim \mu_\theta(\cdot|{\bm{s}})$ by solving an inverse ODE defined by $s_\theta$ (\Eqref{Eq:prob_ode}). \subsection{Q-learning via in-sample planning} \label{imp} Generally, Q-learning can be achieved via the Bellman expectation operator: \begin{equation} \label{Eq:one_step_bellman} \mathcal{T}^\pi Q({\bm{s}}, {\bm{a}}) = r({\bm{s}}, {\bm{a}}) + \gamma\mathbb{E}_{{\bm{s}}' \sim P(\cdot|{\bm{s}},{\bm{a}}), {\bm{a}}' \sim \pi(\cdot|{\bm{s}}')} Q({\bm{s}}', {\bm{a}}'). \end{equation} However, $\mathcal{T}^\pi$ is based on one-step bootstrapping, which has two drawbacks: First, this can be computationally inefficient due to its dependence on many steps of extrapolation. This drawback is exacerbated in diffusion settings since drawing actions from policy $\pi$ in \Eqref{Eq:one_step_bellman} is also time-consuming because of many iterations of Langevin-type sampling. Second, estimation errors may accumulate over long horizons. To address these problems, we take inspiration from episodic learning methods \citep{MFEC, vem} and propose a planning-based operator $\mathcal{T}^\pi_\mu$: \begin{equation} \label{Eq:planning_operator} \mathcal{T}^\pi_\mu Q({\bm{s}}, {\bm{a}}) := \max_{n \geq 0}\{(\mathcal{T}^{\mu})^{n}\mathcal{T}^{\pi}Q({\bm{s}}, {\bm{a}})\}, \end{equation} where $\mu$ is the behavior policy. $\mathcal{T}^\pi_\mu$ combines the strengths of both the n-step operator $(\mathcal{T}^{\mu})^{n}$, which enjoys a fast contraction property, and the operator $\mathcal{T}^\pi$, which has a more desirable fixed point. We prove in Appendix \ref{analysis} that $\mathcal{T}^\pi_\mu$ is also convergent, and its fixed point is bounded between $Q^\pi$ and $Q^*$. Practically, given a dataset $\mathcal{D}^\mu = \{(s_n, a_n, r_n)\}$ collected by behavior $\mu$, with $n$ being the timestep in a trajectory. We can rewrite \Eqref{Eq:planning_operator} in a recursive manner to calculate the Q-learning targets: \begin{align} \label{Eq:planning:1} &R_{n}^{(k)} = r_n + \gamma\max(R_{n+1}, V_{n+1}^{(k-1)}), \\ \label{Eq:planning:2} \text{where} \quad &V_{n}^{(k-1)} := \mathbb{E}_{{\bm{a}} \sim \pi(\cdot|{\bm{s}}_n)} Q_\phi({\bm{s}}_n, {\bm{a}}), \\ \text{and} \quad &\phi = \mathop{\mathrm{arg \ min}}_{\phi} \quad \mathbb{E}_{({\bm{s}}_n, {\bm{a}}_n) \sim \mathcal{D}^\mu} \| Q_\phi({\bm{s}}_n, {\bm{a}}_n) - R_n^{(k-1)} \|_2^2. \end{align} Above $k \in \{1, 2, \dots \}$ is the iteration number. We define $R_{n}^{(0)}$ as the vanilla return of trajectories. \Eqref{Eq:planning:1} offers an implicit planning scheme within dataset trajectories that mainly helps to avoid bootstrapping over unseen actions and to accelerate convergence. \Eqref{Eq:planning:2} enables the generalization of actions in similar states across different trajectories (stitching together subtrajectories). Note that we have omitted writing the iteration superscript of $\pi$ and $\mu$ for simplicity. During training, we alternate between calculating new Q-targets $R_n$ and fitting the action evaluation model $Q_\phi$. The operator $\mathcal{T}^\pi_\mu$ is similar to the multi-step estimation operator $\mathcal{T}_{\text{vem}}$ proposed by \citet{vem}. A notable difference between the two operators is that $\mathcal{T}_{\text{vem}}$ is combined with expectile regression, and thus can only apply to deterministic environments, while our method also applies to stochastic settings. However, unlike $\mathcal{T}_{\text{vem}}$, $\mathcal{T}^\pi_\mu$ does not share the same fixed point with $\mathcal{T}^\pi$. \begin{figure*} [t] \centering \includegraphics[width=1.00\textwidth]{pics/ant-medium-diverse-value.pdf} \caption{ Visualizations of the implicitly planned Q-targets $R_n^{(k)}$ sampled from the dataset of an AntMaze task in four subsequent value iterations. The red pentagram stands for the reward signal. Implicit planning helps to iteratively stitch together successful subtrajectories. } \label{fig:ant_stitch} \end{figure*} \section{Related Work} \label{Related} \textbf{Reducing extrapolation error in offline RL}. Offline RL typically requires careful trade-offs between maximizing expected returns and staying close to the behavior policy. Once the learned policy deviates from the behavior policy, extrapolation error will be introduced in dynamic programming, leading to performance degrade \citep{bcq}. Several works propose to address this issue by introducing either policy regularization on the distributional discrepancy with the behavior policy \citep{bcq, bear, brac, minimal}, or value pessimism about unseen actions \citep{cql, fisher}. Another line of research directly extracts policy from the dataset through weighted regression, hoping to avoid selecting unseen actions \citep{awr, awac, crr}. However, some recent works observe that the trade-off techniques described above are not sufficient to reduce extrapolation error, and propose to learn Q-functions through expectile regression without ever querying policy-generated actions \citep{iql, vem}. Unlike them, We find that limited policy expressivity is the main reason that introduces extrapolation error in previous weighted regression methods, and use an expressive policy model to help reduce extrapolation error. \textbf{Dynamic programming over long horizons}. Simply extracting policies from behavior Q-functions can yield good performance in many D4RL tasks because it avoids dynamic programming and therefore the accompanied extrapolation error \citep{awr, bail, noeval}. However, this method performs poorly in tasks that have sparse rewards and require stitching together successful subtrajectories (e.g., Maze-like environments). Such tasks are also challenging for methods based on one-step bootstrapping because they might require hundreds of steps to reach the reward signal, with the reward discounted and estimation error accumulated along the way. Episodic memory-based methods address this problem by storing labeled experience in the dataset, and plans strictly within the trajectory to update evaluations of every decision \citep{MFEC, gem, vem}. The in-sample planning scheme allows dynamic programming over long horizons to suppress the accumulation of extrapolation error, which inspires our method. \textbf{Generative models for behavior modeling}. Cloning diverse behaviors in a continuous action space requires powerful generative models. In offline RL, several works \citep{bcq, bear, brac} have tried using generative models such as Gaussians or VAEs \citep{vae} to model the behavior policy. However, the learned behavior model only serves as an explicit distributional constraint for another policy during training. In broader RL research, generative adversarial networks \citep{gan}, normalizing flows \citep{flow} and energy-based models \citep{EBM} have also been used for behavior modeling \citep{gail,Parrot, EBI}. Recently, diffusion models \citep{diffusion} have achieved great success in generating diverse and high-fidelity image samples \citep{diffusion_beat_gan}. However, exploration of its application in behavior modeling is still limited. \citet{diffuser} proposes to solve offline tasks by iteratively denoising trajectories, while our method uses diffusion models for single-step decision-making. \section{Experiments} In the following sections, we evaluate the performance of SfBC using several related or state-of-the-art offline RL methods as baselines. We additionally gain insight into SfBC by studying the following two questions: 1) How does SfBC benefit from an expressive generative policy in performing diverse behavior learning? 2) Which part of SfBC has a strong influence on the performance of the algorithm? \begin{table*}[t] \centering \small \resizebox{1.0\textwidth}{!}{% \begin{tabular}{llccccccccc} \toprule \multicolumn{1}{c}{\bf Dataset} & \multicolumn{1}{c}{\bf Environment} & \multicolumn{1}{c}{\bf SfBC (Ours)}& \multicolumn{1}{c}{\bf IQL} & \multicolumn{1}{c}{\bf VEM} & \multicolumn{1}{c}{\bf AWR} & \multicolumn{1}{c}{\bf BAIL} & \multicolumn{1}{c}{\bf BCQ} & \multicolumn{1}{c}{\bf CQL}& \multicolumn{1}{c}{\bf DT} & \multicolumn{1}{c}{\bf Diffuser} \\ \midrule Medium-Expert & HalfCheetah & $\bf{91.4 \pm 0.5}$ & $86.7 $& - & $52.7$ & $72.2$ & $64.7$ & $62.4$ & $\bf{86.8}$ & $ 79.8$ \\ Medium-Expert & Hopper & $\bf{110.4 \pm 0.9}$ & $ 91.5$& - & $27.1$ & $\bf{106.2}$ & $100.9$ & $98.7$ & $\bf{107.6}$ & $\bf{107.2}$ \\ Medium-Expert & Walker & $\bf{109.2 \pm 0.3}$ & $\bf{109.6}$& - & $53.8$ & $\bf{107.2}$ & $57.5$ & $\bf{111.0}$& $\bf{108.1}$& $\bf{108.4}$ \\ \midrule Medium & HalfCheetah & $ 42.4\pm 0.1$ & $\bf{47.4}$& $\bf{47.4}$& $37.4$ & $30.0$ & $40.7$ & $44.4$ & $42.6$ & $ 44.2$ \\ Medium & Hopper & $\bf{65.3 \pm 4.9}$ & $\bf{66.3}$& $56.6$ & $35.9$ & $62.2$ & $54.5$ &$ 58.0 $ & $\bf{67.6}$ & $58.5$ \\ Medium & Walker & $\bf{78.3\pm 1.0}$ & $\bf{78.3}$& $74.0$ & $17.4$ & $73.4$ & $53.1$ &$\bf{79.2}$ & $74.0$ & $\bf{79.7}$ \\ \midrule Medium-Replay & HalfCheetah & $ 38.1 \pm 2.1$& $\bf{44.2}$& - & $40.3$ & $40.3$ & $38.2$ &$\bf{46.2}$ & $36.6$ & $42.2$ \\ Medium-Replay & Hopper & $72.3 \pm 4.4$ & $\bf{94.7}$& - & $28.4$ & $\bf{94.7}$& $33.1$ & $48.6$ & $82.7$ & $\bf{96.8}$ \\ Medium-Replay & Walker & $\bf{71.9 \pm 4.2}$ & $\bf{73.9}$& - & $15.5$ & $58.8$ & $15.0$ & $26.7$ & $66.6$ & $61.2$ \\ \midrule \multicolumn{2}{c}{\bf Average (Locomotion)}&$\bf{75.5}$ & $\bf{76.9}$ & - & $34.3$ & $71.6$ & $51.9$ & $63.9$ & $\bf{74.7}$ & $\bf{75.3}$ \\ \specialrule{.05em}{.4ex}{.1ex} \specialrule{.05em}{.1ex}{.65ex} Default & AntMaze-umaze & $\bf{93.3 \pm 4.7}$ & $87.5$& $87.5$ & $56.0$ & $85.0$ & $78.9$ & $74.0$ & $59.2$ & - \\ Diverse & AntMaze-umaze & $\bf{86.7 \pm 4.7}$ & $62.2 $ & $78.0$ & $70.3$ & $76.7$ & $55.0$ &$\bf{84.0}$ & $53.0$ & - \\ \midrule Play & AntMaze-medium & $\bf{88.3 \pm 8.5}$ & $71.2$ & $78.0$ & $0.0$ & $15.0$ & $0.0$ & $61.2$ & $0.0$ & - \\ Diverse & AntMaze-medium & $\bf{90.0 \pm 4.1}$ & $70.0$ & $77.0$ & $0.0$ & $23.3$ & $0.0$ & $53.7$ & $0.0$ & - \\ \midrule Play & AntMaze-large & $\bf{63.3 \pm 2.4}$ & $39.6$ & $57.0$ & $0.0$ & $0.0$ & $6.7$ & $15.8$ & $0.0$ & - \\ Diverse & AntMaze-large & $41.7 \pm 7.1$ & $47.5$ & $\bf{58.0}$ & $0.0$ & $8.3$ & $2.2$ & $14.9$ & $0.0$ & - \\ \midrule \multicolumn{2}{c}{\bf Average (AntMaze)}& $\bf{77.2}$ & $63.0$ & $72.6$ & $21.0$ & $46.7$ & $23.8$ & $50.6$ & $18.7$ & - \\ \specialrule{.05em}{.4ex}{.1ex} \specialrule{.05em}{.1ex}{.65ex} \multicolumn{2}{c}{\bf Average (Maze2d)}& $74.0$ & $50.0$ & - & $10.8$ & - & $9.1$ & $7.7$ & - & $\bf{119.5}$ \\ \specialrule{.05em}{.4ex}{.1ex} \specialrule{.05em}{.1ex}{.65ex} \multicolumn{2}{c}{\bf Average (FrankaKitchen)}& $\bf{57.1}$ & $53.3$ & - & $8.7$ & - & $11.7$ & $48.2$ & - & - \\ \specialrule{.05em}{.4ex}{.1ex} \specialrule{.05em}{.1ex}{.65ex} Both-side & Bidirectional-Car&$\bf{100.0 \pm 0.0}$& $15.7$ & $0.0$ & $0.0$ & $52.0$ & $88.0$ & $42.3$ & $33.3$ & - \\ Single-side & Bidirectional-Car&$\bf{100.0 \pm 0.0}$& $\bf{100.0}$ & $\bf{100.0}$& $\bf{96.3}$ & $\bf{100.0}$ & $\bf{100.0}$ & $\bf{100.0}$ & $\bf{100.0}$ & - \\ \bottomrule \end{tabular} } \caption{ Evaluation numbers of SfBC. We report the mean and standard deviation over three seeds for SfBC. Scores are normalized according to \citet{d4rl}. Numbers within 5 percent of the maximum in every individual task are highlighted in boldface. Sources of referenced scores and experimental details are provided in Appendix \ref{details}. Note that Diffuser leverages the prior knowledge that Maze2d is a goal-based environment in ``trajectory inpainting'' while other algorithms don't.} \label{tbl:results} \end{table*} \subsection{Evaluations on D4RL Benchmarks} In Table \ref{tbl:results}, we compare the performance of SfBC to multiple offline RL methods in several D4RL \citep{d4rl} tasks. \texttt{MuJoCo locomotion} is a classic benchmark where policy-generated datasets only cover a narrow part of the state-action space, so avoiding querying out-of-sample actions is critical \citep{bcq, cql}. The Medium dataset of this benchmark is generated by a single agent, while the Medium-Expert and the Medium-Replay dataset are generated by a mixture of policies. \texttt{AntMaze} is about an ant robot navigating itself in a maze, which requires both low-level robot control and high-level navigation. Since the datasets consist of undirected trajectories, solving AntMaze typically requires the algorithm to have strong ``stitching'' ability \citep{d4rl}. Different environments contain mazes of different sizes, reflecting different complexity. \texttt{Maze2d} is very similar to AntMaze except that it's about a ball navigating in a maze instead of an ant robot. \texttt{FrankaKitchen} are robot-arm manipulation tasks. We only focus on the analysis of MuJoCo locomotion and AntMaze tasks due to the page limit. Our choices of referenced baselines are detailed in Appendix \ref{choice_baseline}. Overall, SfBC outperforms most existing methods by large margins in complex tasks with sparse rewards such as AntMaze. We notice that VEM also achieves good results in AntMaze tasks and both methods share an implicit in-sample planning scheme, indicating that episodic planning is effective in improving algorithms' stitching ability and thus beneficial in Maze-like environments. In easier locomotion tasks, SfBC provides highly competitive results compared to state-of-the-art algorithms. It can be clearly shown that performance gain is large in datasets generated by a mixture of distinctive policies (Medium-Expert) and is relatively small in datasets that are highly uniform (Medium). This is reasonable because SfBC is motivated to better model diverse behaviors. \subsection{Learning from Diverse Behaviors} \label{lfdb} \begin{figure} [t] \centering \includegraphics[width=1.00\textwidth]{pics/toy-action-visualize32.pdf} \caption{ Visualizations of actions taken by different RL agents in the Bidirectional-Car task. The ground truth corresponds to an agent which always takes the best actions, which is either 1.0 or -1.0. White space indicates suboptimal decisions. Green bounding boxes indicate possible initial states. } \label{fig:toyacion} \end{figure} In this section, we analyze the benefit of modeling behavior policy using highly expressive generative models. Although SfBC outperforms baselines in many D4RL tasks. The improvement is mainly incremental, but not decisive. We attribute this to the lack of multiple optimal solutions in existing benchmarks. To better demonstrate the necessity of introducing an expressive generative model, we design a simple task where a heterogeneous dataset is collected in an environment that allows two distinctive optimal policies. \textbf{Bidirectional-Car task}. As depicted in Figure \ref{fig:illustration}, we consider an environment where a car is placed in the middle of two endpoints. The car chooses an action in the range [-1,1] at each step, representing throttle, to influence the direction and speed of the car. The speed of the car will \textit{monotonically} increase based on the absolute value of throttle. The direction of the car is determined by the sign of the current throttle. Equal reward will be given on the arrival of either endpoint within the rated time. It can be inferred with ease that, in any state, the optimal decision should be either 1 or -1, which is not a unimodal distribution. The collected dataset also contains highly diverse behaviors, with an approximately equal number of trajectories ending at both endpoints. For the comparative study, we collect another dataset called ``Single-Side'' where the only difference from the original one is that we remove all trajectories ending at the left endpoint from the dataset. We test our method against several baselines, with the results given in Table \ref{tbl:results}. Among all referenced methods, SfBC is the only one that can always arrive at either endpoint within rated time in the Bidirectional-Car environment, whereas most methods successfully solve the ``Single-Side'' task. To gain some insight into why this happens, we illustrate the decisions made by an SfBC agent and other RL agents in the 2-dimensional state space. As is shown in Figure \ref{fig:toyacion}, the SfBC agent selects actions of high absolute values at nearly all states, while other unimodal actors fail to pick either one of the optimal actions when presented with two distinctive high-rewarding options. Therefore, we conclude that an expressive policy is necessary for performing diverse behavior learning. \subsection{Ablation Studies} \label{Sec:ablation} \begin{table*}[t] \centering \small \begin{tabular}{cccccc} \toprule \multicolumn{1}{c}{\bf Taks} & \multicolumn{1}{c}{\bf SfBC }& \multicolumn{1}{c}{\bf SfBC + Gaussian} & \multicolumn{1}{c}{\bf SfBC + VAE} & \multicolumn{1}{c}{\bf SfBC - Planning} \\ \midrule Medium-Expert & $\bf{103.7 \pm 0.6}$& $86.2 \pm 4.7$ & $95.5 \pm 4.8$ & $\bf{103.3 \pm 0.9} $ \\ Medium & $\bf{62.0 \pm 2.9}$ & $\bf{60.9} \pm 1.1$ & $\bf{62.7 \pm 2.5} $ & $\bf{60.9 \pm 2.5} $ \\ Medium-Replay & $\bf{60.8 \pm 3.6}$ & $56.6 \pm 4.6$ & $54.4 \pm 3.7 $ & $52.8 \pm 1.5 $ \\ \midrule \multicolumn{1}{c}{\bf Average (Locomotion)}& $\bf{75.5 \pm 2.3}$ & $67.9 \pm 3.3$ & $70.9 \pm 3.5 $ & $\bf{72.3 \pm 1.8} $ \\ \midrule AntMaze-umaze & $\bf{90.0 \pm 4.7}$& $\bf{90.8 \pm 2.4}$ & $85.0 \pm 3.7$ & $\bf{88.4 \pm 8.3}$ \\ AntMaze-medium & $\bf{89.2 \pm 6.7}$ & $82.5 \pm 5.8$ & $66.7 \pm 5.3 $ & $34.2 \pm 5.3 $ \\ AntMaze-large & $\bf{52.5 \pm 5.3}$ & $35.0 \pm 7.8$ & $27.5 \pm 5.8 $ & $7.5 \pm 6.9 $ \\ \midrule \multicolumn{1}{c}{\bf Average (AntMaze)}& $\bf{77.2 \pm 5.2}$ & $69.4 \pm 5.8$ & $59.7 \pm 4.8 $ & $43.3 \pm 5.5 $ \\ \bottomrule \end{tabular} \caption{Ablations of generative modeling methods and the implicit planning method. We report performance numbers averaged over three random seeds and multiple similar tasks. Detailed experimental results appear in Appendix \ref{missing}.} \label{tbl:ablation} \end{table*} We provide experimental results for several variants of SfBC in Table \ref{tbl:ablation}. \textbf{Diffusion vs. other generative models}. Our first ablation study aims to evaluate 3 variants of SfBC which are respectively based on diffusion models \citep{diffusion}, Gaussian probabilistic models and latent-based models (VAEs, \citet{vae}). The three variants use exactly the same training framework with the only difference being the behavior modeling method. The diffusion-based policy outperforms the other two variants by a clear margin in most experiments, especially in tasks with heterogeneous datasets (e.g., Medium-Expert), indicating that diffusion models are fit for ``high-fidelity'' behavior modeling. \textbf{Implicit in-sample planning}. To study the importance of implicit in-sample planning on the performance of SfBC, we first visualize the estimated state values learned at different iterations of Q-learning in an AntMaze environment (Figure \ref{fig:ant_stitch}). We can see that implicit planning helps to iteratively stitch together successful subtrajectories and provides optimistic action evaluations. Then we compare SfBC to a variant that removes the iterative planning scheme and instead learns the Q-function purely from vanilla returns. As shown in Table \ref{tbl:ablation}, implicit planning is beneficial in complex tasks like AntMaze-Medium and AntMaze-Large. However, it is less important in MuJoCo-locomotion tasks, except for Medium-Replay tasks in which many data trajectories suffer from an early-truncated problem and requires dynamic programming for more accurate evaluations. This finding is consistent with a prior work \citep{noeval}. \section{Conclusion} In this work, we address the problem of limited policy expressivity in previous weighted regression methods by decoupling the policy model into a behavior model and an action evaluation model. Such decoupling allows us to use a highly expressive diffusion model for high-fidelity behavior modeling, which is further combined with a planning-based operator to reduce extrapolation error. Our method enables learning from a heterogeneous dataset in a continuous action space while avoiding selecting out-of-sample actions. Experimental results on the D4RL benchmark show that our approach outperforms state-of-the-art algorithms in most tasks. With this work, we hope to draw attention to the application of high-capacity generative models in offline RL.

50447ea0572c5eda485ff2b85fcc8330428183d2

\section{Introduction} Index insurance is a promising tool to reduce the risk faced by smallholder farmers. By linking payouts to a regional index instead of individual losses, it reduces moral hazard, adverse selection and transaction costs. But delinking payouts from individual losses creates \emph{basis risk}, the possibility that a farmer experiences a loss yet does not receive any indemnity. Basis risk substantially reduces the benefit of index insurance, and if severe enough can make it worse than no insurance at all. Further, \cite{Clarke2016}, \cite{carter2017index}, and others argue that basis risk is likely among the most important barriers to index insurance adoption, and that the basis risk of some index insurance schemes has been be very high \citep{clarke2012weather,JensenBarrettEtAl2016}. Recent improvements in satellite remote sensing and machine learning show great potential to improve the accuracy of index and thus reduce basis risk. These technologies have triggered a very active literature extending far beyond the field of economics.\footnote{See for example the special issue of \textit{Remote Sensing} on ``Earth Observation for Index Insurance'' 2021, 13(5), or reviews by \citet{BenamiJinEtAl2021} and \citet{DeLeeuwVrielingEtAl2014}.} These technologies have led to three major shifts in the design of agricultural indices. First, satellite data has helped to design better weather-based indices: while early products were based on local weather stations, satellite data has increased the spatial resolution of weather indices, and facilitated the incorporation of new variables such as soil moisture. Second, it has led to a shift from input-based weather indices towards higher-accuracy output-based indices based on vegetation indices observed with optical satellite sensors. In these two first approaches, the satellite data is used primarily to obtain a better index, while the assessment of the quality of the index is conducted using traditionally-collected field-level yield data. In the third approach, satellite data is used directly to estimate farm or field-level crop yields. Although accurately predicting individual yields currently remains a challenge, rapid progress is being made and this approach shows great potential for deriving very accurate output-based indices. Even more importantly, it will provide a cost-effective way to assess the quality of a given index over a large number of fields, helping insurers to design better insurance zones and making it easier for governments, researchers, and others to reliably assess the quality of index insurance products. This paper identifies and studies a new challenge associated with assessing index insurance quality with more granular data: existing measures of quality are biased in large-$N$ (number of farms), small-$T$ (number of time periods) samples. We show that $R^2$-derived estimates of basis risk are systematically biased downward in the small-$T$ large-$N$ case, meaning practitioners who do not take this into account will generally overestimate the quality of index insurance products. Intuitively, this bias arises from the fact that the basis risk estimates are functions of the covariance matrix between fields. The covariance matrix between fields has $N\cdot(N-1)/2$ parameters, yet only $N\cdot T$ observations to estimate it. Having more fields $N$ than time periods $T$ is very typical of agricultural data, and is going to be exacerbated by developments of satellite data methods, which are particularly suited to extend the sample over space, but are unfortunately only available for a few recent years. The resulting bias has gone unnoticed in the existing literature and has important implications for both past and future estimates of basis risk. After documenting the bias in various measures of basis risk, we analyze it theoretically. We focus on linear measures of basis risk, which allows us to connect our problem with a rich literature in statistics. We start with a review of models to parameterize the inter-field covariance matrix, focusing on the \emph{spiked} model introduced by \citet{Johnstone2001}. We discuss then how the high-dimension, low-sample-size (HDLSS) framework introduced by \citet{HallMarronEtAl2005} can help us understand the bias of linear basis risk measures. In the HDLSS framework, the sample size $T$ is assumed fixed, while the number $N$ of variables is assumed to grow to infinity. This corresponds exactly to the situation we are facing in index insurance, where satellite data techniques increase sample sizes over space much faster than over time. Using results from the HDLSS literature, we provide a new theorem deriving the theoretical bias of our linear measure of basis risk. Going back to the simulations, we find that our theory predicts the empirical bias remarkably well. The bias we study in this paper is particularly pernicious for two reasons. First, we show the bias in basis risk measurement can actually be worsened by higher resolution data. Second, the bias is greatest when individual yields are poorly correlated, meaning it is likely to be particularly severe in smallholder systems in developing countries, which is precisely where new satellite data promise to have the largest positive impact. In light of these findings, understanding this bias is essential to realizing the promise of high resolution satellite data for index insurance. This paper's findings also provide a potential explanation for observed low uptake of existing index insurance products.\footnote{See \cite{carter2014index} and \cite{carter2017index} for discussions of the literature on low index insurance uptake} Farmers are experts on their yields and are aware of how they relate to their neighbors, and likely have an accurate understanding of how correlated their yields are to their neighbors'. Since the bias we study in this paper is essentially a result of inaccurate estimates of inter-field correlations due to a small number of observed time periods, farmers who have more accurate understandings of these correlations ought to buy insurance less often than biased estimates would suggest. \section{Measures of basis risk} Index insurance products are usually assessed following two broad approaches. In the first one, the interest is on basis risk, which is a measure of the frequency and size of errors in predicting individual yields and/or harvest losses. The second seeks to evaluate how the insurance product derived from this index performs. This entails typically specifying a indemnity and premium functions and assuming a utility function for the farmer. In this paper, we focus on the first approach. Basis risk is sometimes defined as the probability of a farmer experiencing a loss yet not receiving an indemnity. As such, it can be estimated as a simple conditional probability, the false negative probability. However, this measure fails to capture the severity of the prediction error, which has important consequences for farmer welfare. For this reason, \citet{ElabedBellemareEtAl2013} suggest focusing on the field-level share of the variation in yields not explained by the index. This is equivalent to using $1-R^2_i$, where $R^2_i$ is the coefficient of determination between the yields of field $i$ and the index. In this paper, we focus on output-based indices such as the zone average yield. This comes from our initial motivation to assess the benefits of third-generation datasets, which (will) allow estimates of yields at the field-level. That said, the same results hold for traditional area yield insurance contracts; the source of the field-level yield estimates does not matter. In recent work on output-based indices, \citet{StiglerLobellKenyaOptimal2021} discuss how to aggregate the field-specific $R_{i}^{2}$ and propose a zone-specific \emph{total} $\overline{\overline{R^{2}}}\equiv1-\sum_{i}SSR_{i}/\sum SST_{i}$. This $\overline{\overline{R^{2}}}$ is a generalization of the individual $R^{2}$ written as $R_{i}^{2}\equiv1-SSR_{i}/SST_{i}$, where SSR and SST stand respectively for sum of squared residuals and total sum of squares. This measure is simply a variance-weighted average of the individual $R_i^2$, meaning it puts more weight on farmers who are exposed to more risk. The $\overline{\overline{R^{2}}}$ measure can alternatively be obtained by running $N$ field-specific regressions of the index on individual yields and aggregating their $SSR_i$ and $SST_i$ into $\overline{\overline{R^{2}}}\equiv1-\sum_{i}SSR_{i}/\sum SST_{i}$. In the case of an output-based area-yield index, \citet{StiglerLobellKenyaOptimal2021} propose an alternate, numerically identical, formula: \begin{equation}\label{eq:R2_mean} \overline{\overline{R^{2}}}=tr\left(\Sigma\mathbf{1}(\mathbf{1}^{'}\Sigma\mathbf{1})^{-1}\mathbf{1}^{'}\Sigma\right)/tr(\Sigma) \end{equation} Here, $\Sigma$ is the covariance matrix between individual fields, and $\mathbf{1}$ denotes a vector of 1, of dimension $N$. While numerically equivalent to field-specific regression, formula \eqref{eq:R2_mean} has the advantage of establishing the connection between the basis risk and the covariance of fields. Intuitively, the strength of the index depends on the strength of the off-diagonal elements: a diagonal covariance matrix (uncorrelated fields) would result in higher basis risk than a covariance matrix with many positive off-diagonal elements (many correlated fields). \citet{StiglerLobellKenyaOptimal2021} show also that the formula \eqref{eq:R2_mean} can be generalized to a broader class of output-based indices, which use field-specific weights to form the index, $f_{t}=\sum_{i}w_{i}y_{it}$, or in matrix form, $f=Y\mathbf{w}$. The area-yield index is a special case in this class, with $\mathbf{w}=\mathbf{1}/N$. The formula becomes then: \begin{equation}\label{eq:R2_w} \overline{\overline{R^{2}(\mathbf{w})}}=tr\left(\Sigma\mathbf{w}(\mathbf{w}^{'}\Sigma\mathbf{w})^{-1}\mathbf{w}^{'}\Sigma\right)/tr(\Sigma) \end{equation} They show that this quantity is not maximized using the area-yield index $\mathbf{1}/N$, but instead taking the first principal component (PC) of the covariance matrix $\Sigma$, $\mathbf{w^\star}=PC_1(\Sigma)$. Evaluated at this optimal $\mathbf{w}^{\star}$, the objective function $\overline{\overline{R^{2}(\mathbf{w}^{\star})}}$ turns out to be equal to the share of the first eigenvalue of $\Sigma$, that is $\overline{\overline{R^{2}(\mathbf{w}^{\star})}}=\lambda_{1}/\sum\lambda$. The $\overline{\overline{R^{2}(\mathbf{w}^{\star})}}$ is an interesting measure that defines the upper-bound any index can achieve (according to the total $R^{2}$ criterion) for a given zone. In that sense, it can be interpreted as a measure of zone quality, a low $\overline{\overline{R^{2}(\mathbf{w}^{\star})}}$ for a given zone indicating that even the best index would not perform very well. The connection between $\overline{\overline{R^{2}(\mathbf{w}^{\star})}}$ and the eigenvalues of $\Sigma$ indicates that the $\overline{\overline{R^{2}(\mathbf{w}^{\star})}}$ is equivalent to the usual definition of $\lambda_{1}/\sum\lambda$ in terms of the \emph{percentage of total variance captured by the first principal component}. In addition, and of particular relevance for the current paper, the statistical properties of sample eigenvalues are a very well-studied problem in statistics. Admittedly, linear correlation measures have limitations in the context of index insurance. Arguably, it is more important for an index to accurately predict yield losses than to predict good harvests. Various approaches have been suggested to take this into account, ranging from quantile regression \citep{ConradtFingerEtAl2015} to more sophisticated left-tail dependence indices \citep{Bokusheva2018}. In the following, we include also a quantile version of our total $R^2$ measure, based on the quantile pseudo $R^2$ developed by \citet{KoenkerMachado1999}. \citet{KoenkerMachado1999} suggest a pseudo $R^2(\tau)=1-V(f,\tau)/V(const,\tau)$ at quantile $\tau$, where $V(\tau)$ is the quantile analogous to the SSR. In a similar way to our total $R^2$, we define the total quantile pseudo $R^2$ as $R^2_q=1 -\sum V_i(f,\tau)/\sum V_i(const,\tau)$, and use the value of $\tau=0.3$ following previous literature \citep{BucheliDalhausEtAl2020}. The bias we identify in this study may also apply to other nonlinear measures of basis risk, but we leave that for future studies. \section{Data and empirical simulations\label{sec:Data-and-empirical-simuls}} \subsection{Data} We use two state-of-the-art datasets of satellite-estimated yields in the USA and in Kenya to illustrate the potential bias in basis risk measures. Both datasets contain maize yield predictions produced with the Scalable Yield Mapper (SCYM) model initially developed by \citet{LobellEtAl2015}. The SCYM model is one of the most advanced yield prediction models available to date (see \citealp{JinAzzariEtAl2017, AzzariLobell2017, DeinesPatelEtAl2021OneMillionTruth} for the US and \citealp{BurkeLobell2016, JinAzzariEtAl2017SmallholderYieldHeterogeneity, JinAzzariEtAl2019, LobellAzzariEtAl2020AJAE} for Sub-Saharan Africa), and has been already used to analyze various questions such as the effect of cover crops, of conservation tillage or the dynamics of crop expansion \citep{SeifertAzzariEtAl2018, DeinesWangEtAl2019, Stigler2018}. The dataset has been used specifically for analyzing crop insurance in the US in \citet{StiglerLobellUSIns2020} and in Kenya in \citet{StiglerLobellKenyaOptimal2021}. While the US and Kenya datasets share a common methodology, they also differ in several respects. First, we have twenty years of data for the US: from 2000 to 2019, while for Kenya we have only four: from 2016 to 2019. This difference is due to the fact that fields are much smaller in Kenya which means that higher resolution satellite images are required (from Sentinel-2), and those images are only available for recent years. Another difference lies in the fact that maize is mainly cultivated in rotation together with soybeans in the US, while this is less common in Kenya. For the US, this means we observe a large number of missing maize values for those fields practicing rotation, which makes estimating accurate covariance matrices more difficult. We adopt a simple solution, and focus on the fields that only cultivate corn over the 2000-2019 period, and select counties that have at least 30 observations. Doing so, we end up with a sample of 37 \emph{zones} for the US. For Kenya, almost all fields cultivate maize every year, and we randomly sample 200 fields in each of 453 Kenyan sub-counties (Kenya's smallest administrative unit). A last aspect where both datasets differ is in the quality of the satellite predictions. Predictions are typically much better in the US, characterized by large uniform fields, than in Kenya, which has smaller fields and more heterogeneous cultivation practices. \citet{DeinesPatelEtAl2021OneMillionTruth} report that the SCYM yield estimates in the US have a $r^2$ of 0.45 when compared to a ground-truth dataset at the field level, raising to 0.69 when assessed against county-level means instead. On the other hand, \citet{JinAzzariEtAl2019} report that the yield estimates in Kenya have an agreement of about 50\% against district means. Clearly, the current accuracy of the satellite-based yield predictions is not yet perfect, and more research is still needed before using these datasets at an operational level for routine insurance assessment. Witnessing the rapid progress made in the field in the last ten years, there is good reason to believe that in a near future yield estimates will be much more accurate. In this paper, we focus on another source of error for insurance applications, that due to the small-$T$ large-$N$ setting, that has largely gone unnoticed in the literature. \subsection{Basis risk measures in SCYM data} Our analysis here proceeds in two steps. We first estimate various basis risk measures using both SCYM data sets. We then estimate the bias associated to these measures. To do so, we employ a Monte Carlo (MC) simulation experiment using pseudo true values calibrated to the data at hand. More precisely, we pretend that the covariance matrices and the basis risk measures estimated in the first step are the true ones, and simulate random samples of various $T$ sizes using these pseudo-true covariances. We then re-estimate the basis risk measures on the simulated samples, and infer the bias by comparing those simulated values to the pseudo-true basis risk measures used to simulate the data. Starting with the initial estimates of basis risk, we focus on three measures, 1) the total $R^2$ using the county mean as the index, 2) the total quantile pseudo $R^2$ using also the county mean, and 3) the total $R^2$ using the optimal index, which is equivalent to the share of the first eigenvalue. Figure~\ref{fig:Basis-risk-SCYM-empirial} shows the results for these three measures, highlighting a stark difference between Kenya and the USA. Remembering that basis risk is defined as $1-R^2$, we see that the basis risk is much lower in the USA than in Kenya according to the three measures.\footnote{Interestingly, these three measures of basis risk are highly correlated to each other, the lowest correlation being 92\% between the quantile pseudo $R^2$ and the first eigenvalue.} This result is consistent with the structure of agricultural production in the two settings, large-scale farms in the US use relatively similar production technologies, whereas smallholder farmers Kenya are very heterogeneous. It is therefore unsurprising that basis risk, interpreted as the lack of correlation between fields, is much higher in Kenya than in the USA. However, these estimates are based on small-T samples (T=4 for Kenya, and T=20 for the USA), and we show in the next section that they underestimate basis risk as predicted by the theory. \begin{figure} \caption{Basis risk measure on SCYM data\label{fig:Basis-risk-SCYM-empirial}} \includegraphics[width=0.95\columnwidth]{figures/SCYM_empirical/basisRisk_all} \end{figure} \subsection{SCYM-based simulations of the bias} To gauge the reliability of the estimates obtained above, we proceed to a Monte Carlo simulation. Using the covariance matrices and means implicitly estimated above and pretending those are pseudo-true covariance matrices, we simulate random samples assuming a normal distribution. We do this for three different T dimensions representative of dataset found in practice, T=4, 10 and 20, and for each of these run 500 simulations. For each simulated sample, we recompute the three measures of basis risk, average them over the 500 samples, and compute the resulting bias by comparing these averages to the pseudo-true values obtained from the pseudo-true covariance matrices.\footnote{For the linear basis risk metrics, the population value is directly obtained from on the covariance matrix based on the formulas derived above. For other measures such as the quantile pseudo $R^2$, we don't have analytical formulas for the population values, and hence obtain them by simulating with a large sample of 25000 observations.} Figure~\ref{fig:bias-SCYM-simul} shows the bias as estimated by simulation. Values on the x-axis denote the pseudo-true population value, and on the y-axis the average sample values (left column) and bias (right column). We note a very similar phenomenon over each basis risk metric: the bias is relatively high for low values of the pseudo-true basis risk, and tend to decrease for increasing values of basis risk, with a possible sign reversal for very high values. This indicates that an insurer assessing the basis risk of a zone will tend to be over-optimistic about the $R^2$ measure, and hence under-estimate the basis risk itself. This upward bias in the $R^2$ is relatively large, and even larger in percentage terms, with an upward bias of 150\% for low values of the quantile pseudo $R^2$ measure. The bias decreases for larger values of T, suggesting it would eventually disappear with a large enough sample over the T dimension. Taken together, the fact that the bias is higher for lower value and for lower T suggests that the bias in the initial estimates in Figure~\ref{fig:Basis-risk-SCYM-empirial} is relatively modest for the USA (T=20 and high $R^2$ values observed), but much more important in Kenya (T=4 and low $R^2$ values observed). \begin{figure} \caption{Bias from simulations based on SCYM\label{fig:bias-SCYM-simul}} \includegraphics[width=0.95\columnwidth]{figures/SCYM_simulated/R2measures_raw_vs_bias.png} \end{figure} The calibrated Monte Carlo exercise suggests that there is a substantive upwards bias in the total $R^2$ measure, resulting in downwards bias for the linear basis risk metrics considered. However, this approach of simulating pseudo-true covariance matrices faces several limitations. First, it rests on an initial estimate of the covariance matrix, which is itself potentially biased. Second, and most important, it is difficult to know whether these insights generalize to other cases, since the results were derived for specific covariance matrices. Generalizing these results calls for analyzing the bias theoretically, which is the subject of the next section. \section{Theory\label{sec:Theoretical-simulations}} The simulations in Section~\ref{sec:Data-and-empirical-simuls} show that there is potentially a substantive bias in various basis risk measures. These results were obtained by pretending that the sample covariance matrices we estimated are the population ones, and simulating from this population matrix. This is somewhat artificial, since the simulations show that our initial estimate could be biased. Escaping this conundrum calls for a more formal approach, assuming a-priori a specific covariance matrix to be used as data-generating process (DGP). This makes it possible to simulate data by controlling and varying the \emph{true} parameters, and assessing how sample estimates behave with respect to these known parameters. \subsection{Choice of the data generating process} To derive analytical results on the estimation of the basis risk metrics, we need to specify a data generating process. To do so, we adopt the so-called spiked covariance model introduced by \citet{Johnstone2001}. This model assumes that the eigenvalues of the covariance matrix are spiked, i.e. that a few eigenvalues clearly dominate the rest of the eigenvalues. This corresponds to the idea that a few common factors explain much of the variability of the data, and that the remaining variability is idiosyncratic. In fact, \citet{WangFan2017} show that a spiked structure can be generated from a factor model. This factor structure is exactly the assumption behind index insurance, which posits that a single common index will predict well individual yields. It is also implicitly assumed in existing literature on index insurance that models farm-level risk as a linear combination of covariate shocks and idiosyncratic farm-level shocks (See \citealt{Miranda1991}, \citealt{Mahul1999} and \citealt{ConradtFingerEtAl2015} for example). This suggests that the spiked model is a very natural starting point as generating process for agricultural data, noting that the main determinants of yield (weather, input and output prices) are highly correlated within a zone. As another benefit, the spiked model is the subject of a vast theoretical literature (see for reviews \citealp{JohnstonePaul2018}), providing us with important tools to understand the bias of our estimates. The idea behind the spiked model is to specify the eigenvalues of the covariance matrix, assuming that a few eigenvalues are much larger than the other ones, and furthermore grow with the dimensionality $N$ of the sample. In other words, letting $\lambda_{1}>\lambda_{2}>\ldots>\lambda_{M}>\lambda_{M_{1}}\geq\ldots\geq\lambda_{N}$ represent the ordered eigenvalues of $\Sigma$, one assumes that there are at most $M$ \emph{large} eigenvalues, and that the remaining $N-M$ are bounded. In the following, we will use the simplest specification, assuming that there is only one large spike $\lambda_{1}=aN^{\alpha}$, and that the remaining eigenvalues are all equal to a constant $b$. Denoting $\Lambda$ the diagonal matrix whose elements are the ordered eigenvalues, the population covariance matrix is then specified as $\Sigma=Q\Lambda Q^{'}$, where $Q$ is an orthogonal matrix ($QQ^{'}=I$). Using an orthogonal matrix guarantees that the eigenvalues of $\Sigma$ are the same as $\Lambda$, that is $[aN^{\alpha},b,\ldots,b]$. As discussed above, a measure of minimum zonal risk is given by $1-\overline{\overline{R^{2}(w^{\star})}}=1-\lambda_{1}/\sum\lambda\equiv1-\tilde{\lambda}_{1}$, which indicates the minimum $1-\overline{\overline{R^{2}(w)}}$ that can be reached within a zone by any linear index. From now on, we will focus on the population $\tilde{\lambda}_{1}\equiv\lambda_{1}/\sum\lambda$, and its sample counterpart, $\hat{\tilde{\lambda}}_{1}\equiv\hat{\lambda}_{1}/\sum\hat{\lambda}$. Starting with the population $\tilde{\lambda}_{1}$, in the spiked model it takes the value of $\tilde{\lambda}_{1}=aN^{\alpha}/(aN^{\alpha}+(N-1)b)$. The behavior of $\tilde{\lambda}_{1}$ with growing $N$ will depend on the value of $\alpha$, and we distinguish three cases: \begin{itemize} \item Vanishing spike: $\alpha<1$, and hence $\tilde{\lambda}_{1}\xrightarrow{N\to\infty}0$ \item Constant spike: $\alpha=1$, and hence $\tilde{\lambda}_{1}\xrightarrow{N\to\infty}a/(a+b)$ \item Expanding spike: $\alpha>1$, and hence $\tilde{\lambda}_{1}\xrightarrow{N\to\infty}1$ \end{itemize} Which spike structure should be considered is a complicated question. Typically, a researcher faces a dataset with given $T$ and $N$, and asking the thought experiment of what would happen if she had an infinite number of fields N is somewhat abstract. One could eventually argue that when the dimension $N$ is increased by extending the sample over space, adding fields further away would possibly reduce the value of $\tilde{\lambda}_{1}$, which would correspond to the vanishing spike case. Conversely, if the sample is extended by adding more pixels or fields within a zone, $\tilde{\lambda}_{1}$ could alternatively increase (expanding spike). At the theoretical level, the vanishing and expanding spike are, however, not very interesting since they only allow two extreme values of either 0 or 1. In the following, we focus hence on the constant spike, assuming that the value of $\tilde{\lambda}_{1}$ is constant, and equal to $a/(a+b)$. This will allow us to investigate the behavior of the basis risk measure for a variety of cases of $\tilde{\lambda}_{1}$, representing both low and high homogeneity zones that we encountered in Section~\ref{sec:Data-and-empirical-simuls}. \subsection{Analytical results} Turning now to the behavior of the sample eigenvalue under the moderate spike, we need now to make assumptions on the ratio between $T$ and $N$. There exist broadly three frameworks in statistics: 1) traditional asymptotics, with $N$ fixed and $T\to\infty$ so that $N/T\to0$, 2) random matrix theory $N/T\to c$ for a constant $c $ and 3) high-dimension low-sample size (HDLSS), where $N/T\to\infty$. The latter case, HDLSS, appears the most appropriate to describe third-generation datasets, where $T$ is considered fixed, and $N$ is allowed to grow very large. The HDLSS was introduced by \citet{HallMarronEtAl2005}, with notable contributions from \citet{AhnMarronEtAl2007,JungMarron2009}, see also \citet{AoshimaShenEtAl2018} for a review. While important results have been derived for the raw sample eigenvalue $\hat{\lambda}_{1}$ (see \citealp{AhnMarronEtAl2007}), there is, to the best of our knowledge, no result available on the relative eigenvalue $\hat{\tilde{\lambda}}_{1}\equiv\hat{\lambda}_{1}/\sum\hat{\lambda}$. To fill this gap, we derive a new result, describing the distribution and the bias of $\hat{\tilde{\lambda}}_{1}$, building on the seminal work by \citet{AhnMarronEtAl2007}. In what follows, $\stackrel{d}{\to}$ denotes convergence in distribution and $\stackrel{p}{\to}$ denotes convergence in probability. See the appendix for details on the notation. \begin{thm}[Distribution of the share of the first sample eigenvalue.] \label{thm:main-theorem-sample-eigenval} For $1\leq t\leq T$, set $Y_t=(y_{1t},...,y_{Nt})$ and assume that $ Y_t\stackrel{iid}{\sim}\mathcal N_N( \mu,\Sigma)$ where $\mathcal N_N( \mu,\Sigma)$ denotes the $N$-dimensional Normal distribution with mean vector $ \mu\in\mathbb R^N$ and covariance matrix $\Sigma\in\mathbb R^{N\times N}$, and $mu$ is unknown. Also assume a spiked model for $\Sigma$ i.e. the eigenvalues of $\Sigma$ satisfy $\lambda_1=aN^\alpha,$ and $\lambda_2=\cdots=\lambda_N=b$ for some positive constants $a,b$. We have the following results. \begin{enumerate} \item Vanishing spike: $\hat{\tilde{\lambda}}_1 \stackrel{p}{\to} \tfrac{1}{T-1} $, noting that $\tilde{\lambda}_1 \to 0 $. \item Constant spike: $\hat{\tilde{\lambda}}_1 \stackrel{d}{\to}\frac{aC^2+b}{aC^2+b(T-1)} $ where $C^2\sim\chi^2_{T-1}$, noting that $\tilde{\lambda}_1\to\dfrac{a}{a+b}$. \item Expanding spike: $\hat{\tilde{\lambda}}_1 \stackrel{p}{\to} 1 $ noting that $\tilde{\lambda}_1 \to 1 $. \end{enumerate} \end{thm} \begin{thm}[Bias of the share of the first sample eigenvalue.] \label{thm:theorem-bias} Assume the same setup as in Theorem \ref{thm:main-theorem-sample-eigenval}. \begin{enumerate} \item Vanishing spike: Asymptotic bias is $\tfrac{1}{T-1}$. \item Constant spike: Let $r=\tfrac{a}{a+b}$. Asymptotic bias is $\mathbb E\left(\tfrac{(1-r)(rC^2+1-r(T-1))}{rC^2+(1-r)(T-1)}\right)$ where $C^2\sim\chi^2_{T-1}$. \item Expanding spike: Asymptotic bias is 0. \end{enumerate} \end{thm} Figure~\ref{fig:distribution-lambda} shows the asymptotic distribution of $\hat{\tilde{\lambda}}_{1}$ in the constant spike case for four values of the true $\tilde{\lambda}_{1}$ and for various T dimensions. The asymptotic distribution explains the behavior of the bias that we observed in the calibrated simulations above. It is clear that the bias is large for low values of the true $\tilde{\lambda}_{1}$, and tends to decrease as the true $\tilde{\lambda}_{1}$ increases. This is especially problematic: the bias is greater when there is little shared risk between farmers. This is likely to be the case in places where farmers are more heterogeneous, such as smallholder systems in developing countries. It is also clear that the bias is a function of $T$: increasing $T$ reduces the bias, and for values as high as $T=100$, this bias appears negligible. Unfortunately, having 100 years of field-level data ($T=100$) is unrealistic, and climate change would limit the usefulness of such a long series even if it were available. Looking at a value of $T=20$ which is more realistic for agricultural data, bias is still present, in particular for lower values of $\tilde{\lambda}_{1}$. This suggests that even with high-quality data, an insurer will still be over-confident in her assessment of the quality of an insurance product, unless the true quality is very high. \begin{figure} \caption{Asymptotic distribution of $\hat{\tilde{\lambda}}_{1}$ \label{fig:distribution-lambda}} \begin{centering} \includegraphics[width=0.95\columnwidth]{figures/distribution_theo_sampleEigen_rJuYrPo.png} \end{centering} \end{figure} Ideally, deriving the bias would help construct a bias-corrected estimator. Unfortunately, the extent of the bias depends itself on the true yet unknown value of $\tilde{\lambda}_{1}$. This is a challenging statistical problem, which we leave for further research. However, the result we obtained can be used to derive an upper bound on the bias. Note indeed that the bias is maximum at $\tilde{\lambda}_{1}=0$, taking a value of $1/(T-1)$. This suggests a simple rule of thumb for practitioners to quantify the expected bias they can face in the worst case scenario. \subsection{Illustration of the theorem} To illustrate this result, we simulate population covariance matrices $\Sigma_N(\tilde{\lambda})$ according to the moderate spike model, varying $\tilde{\lambda}_{1}$ between 0 and 1. That is, $\Sigma_N(\tilde{\lambda})=Q_N\Lambda_N(\tilde{\lambda}) Q_N^{'}$, where $\Lambda_N(\tilde{\lambda})$ is a diagonal matrix with the eigenvalues $[aN=\tilde{\lambda}/(1-\tilde{\lambda})N, 1, \ldots,1]$ and $Q_N$ is a random orthogonal matrix. For each of the $\Sigma(\tilde{\lambda})$ covariances, we then simulate data with sample size of $T\in[4,20,100]$ and dimension $N\in[50,200,500,1000]$ assuming a multivariate normal distribution that is i.i.d. over time. The correlation metrics being insensitive to the values of the means vector $\mu_N$, we simply set it to 0. In other words, we now simulate data from: \begin{equation*}Y_{T,N}\stackrel{iid}{\sim} \mathcal{N} \left(0_N, \Sigma_N(\tilde{\lambda})=Q\Lambda_N(\tilde{\lambda}) Q^{'}\right) \end{equation*} We then estimate $\hat{\tilde{\lambda}}_{1}$ on the simulated data. Figure~\ref{fig:theo-simul-bias-T4} shows the resulting bias estimates on the y-axis, and the true population value $\tilde{\lambda}_{1}$ on the x-axis. The black line represent the bias of the values estimated on the simulated data from the spiked model, and the blue line represents the formula from Theorem~(\ref{thm:main-theorem-sample-eigenval}). The red dot corresponds to the worst bound $1/(T-1)$. Focusing first on the behaviour of the bias, we clearly see the phenomenon observed with the SCYM-calibrated simulations shown in section~\ref{sec:Data-and-empirical-simuls}. For low values of $\tilde{\lambda}$, we observe a very substantive upwards bias, leading to over-confidence in the quality of an insurance product. This bias decreases with increasing $\tilde{\lambda}$, and even reverses at very high values of $\tilde{\lambda}$. Comparing now the difference between the simulated bias (black line) and our asymptotic formula from (\ref{thm:main-theorem-sample-eigenval}), we see that the formula approximates remarkably well the empirical bias for dimensions as low as $N=200$. More interestingly, having a higher dimensionality $N$ is no longer a curse, but improves instead the validity of the bias formula! \begin{figure} \caption{Theoretical bias and bias from the spiked-model simulations\label{fig:theo-simul-bias-T4}} \begin{centering} \includegraphics[width=0.95\columnwidth]{figures/simuls_theo/bias_eigen1_facet_4N_T4_rKuOwPw.png} \par\end{centering} \end{figure} As a final test, we compare the empirical bias obtained in the calibrated simulations to the analytical bias formula. Remember that the calibrated simulations were generated based on the empirical covariance matrices estimated from the satellite data. As such, there is no guarantee that the asymptotic bias formula holds, since the latter is obtained assuming a constant spike model, which might not hold in the SCYM dataset. To further test the generality of the bias formula, we focus on the total $R^2$ measure based on the area-yield index, instead of focusing on $\hat{\tilde{\lambda}}_1=R^2(w^\star)$. In a constant spike model, one can show that these measures are very similar, but this is not necessarily the case in practice. Figure~\ref{fig:SCYMul-vs-theo} shows the empirical bias (straight line) together with the bias predicted by theorem 2 (dotted line). The analytical bias formula approximates the empirical bias well, with a difference that tends to vanish as the dimension $N$ increases. This result is very encouraging, suggesting that our choice of a constant spike model to represent the data might be relevant in practice. Furthermore, it indicates that the theory developed here provides a reliable tool for practitioners to assess the potential bias in their measures of basis risk. \begin{figure} \caption{Theoretical bias and bias from the SCYM-calibrated simulations\label{fig:SCYMul-vs-theo}} \begin{centering} \includegraphics[width=0.95\columnwidth]{figures/SCYM_simulated/SCYMmul_bias_vs_theo.png} \end{centering} \end{figure} \section{Conclusion} High resolution satellite images coupled with recent advances in machine learning are expected to yield significant progress in index insurance. However, the discussion to date has overlooked the fact that while data is becoming much richer in terms of number of fields observed, the number of years we observe remains low. As we have shown in this paper both theoretically and empirically through simulation, this introduces a downward bias in common measures of basis risk, which if left uncorrected is likely to yield overly optimistic assessments of insurance product quality. Our paper is the first to identify this bias in commonly used measures of index insurance quality, which is important because it is likely to lead to overly optimistic assessments of product quality. The theory we develop to explain the observed bias provides a useful set of building blocks to approximate and bound the bias in real-world situations. Academics as well as organizations and governments designing new index insurance products ought to take this into account, especially when products are being designed based on a small number of time periods; the simulation methods we develop provide a strategy for generating more accurate product quality estimates. We highlight two aspects of this bias that are especially problematic for developing country settings where satellite-based index insurance has the greatest potential benefit. First, the bias is generally larger when individual farm yields are less correlated, which is more true in rural smallholder systems than in large developed country farms. Second, the bias can actually increase as the number of fields in the data increases, meaning high resolution data may yield worse estimates of basis risk if the bias is uncorrected. For these reasons, it is critical that companies, governments, and non-governmental organizations designing index insurance products are aware of and take steps to correct the bias we study in this paper. Future work ought to focus on developing methods to correct the bias identified in this paper. More careful modeling of the covariance between fields is a promising avenue: methods that take into account spatial correlation and/or allow the covariance matrix between fields to vary over time are promising. For example, in the worst years, when poor weather conditions are the dominant factor affecting yields, it may be that field-level outcomes are more correlated than in more typical years, when individual non-weather shocks drive most of the variation. This paper showed evidence of bias in both linear and quantile-based measures of basis risk, yet confirmed analytically this bias only in the linear case. Focusing on linear measures of basis risk is convenient in that they are easy to understand and study analytically. At the same time, linear measures fail to capture the fact that the effect of basis risk is nonlinear: failures to accurately predict negative shocks are much more detrimental than failures in good years. Analyzing the theoretical bias of nonlinear measures of basis risk and of expected utility metrics remains a challenging task where future research will be needed.

7a9b09c62180b4b9f5c87983a808251ff6c58c0a

\section*{Acknowledgments} This work was supported by the funds of Research Project of the National Language Commission (No. ZDI135-131, No. ZDI145-24) and Fundamental Research Funds for the Central Universities in BLCU (No. 21PT04). We would like to thank all anonymous reviewers for their valuable comments and suggestions on this work. \bibliographystyle{ccl} \section{Introduction} Definition Generation (DG) is the task of describing the meaning that a word takes in a specific context. This task can help language learners by providing explanations for unfamiliar words. Recent researches \cite{Ishiwatari2019LearningTD,Zheng2021DecomposeFA} attempted to apply the task to the field of Intelligent Computer-Assisted Language Learning (ICALL), and have made a significant progress. Previous studies on DG mainly concentrate on generating different definitions for polysemous words \cite{Gadetsky2018ConditionalGO,Mickus-2019-mark,Reid-2020-vcdm}, or generating definitions with appropriate specificity \cite{Huang-2021-definition}. In these studies, researchers have faced various issues, such as the high complexity problem. High complexity definitions contain words that are more difficult than the defined word, and hence are labored for language learners to read and understand. Nevertheless, there have been few focuses on complexity controllable generation of definitions. A possible reason is that the complexities of definitions are not provided in currently existed datasets, which leads to the difficulty of automatic training and evaluation. Actually, the problems mentioned above are especially prominent in the language environment of Chinese. Definitions with suitable complexity are in urgent practical needs for Chinese as Foreign Language (CFL) learners. According to the Ministry of Education of China, by the end of 2020, more than 20 million foreign students are learning Chinese. But as \newcite{zhang-2011-duiwai} pointed out, since the difficulty of definitions is not considered, most existing dictionaries cannot meet CFL learner’s requirements. Besides, the existing Chinese learner dictionaries contain only a small number of words. For instance, the Contemporary Chinese Learner Dictionary (CCLD) only has about 6,400 words. In contrast, the Modern Chinese Dictionary (MCD), which is designed for native speakers, has about 69,000 words. Therefore, in this work, we focus on the task of generating definitions for CFL learners with appropriate complexities. At present, there are two datasets used for the Chinese definition generation task, but neither of them can meet the needs of this task. The most widely used CWN dataset \cite{Yang2020IncorporatingSI,Fan2020BERTChineseDM,Kong-2020-Toward} was built from the Chinese WordNet \cite{huang-2010-chinese}, which is a knowledge base of sense distinction\footnote{\url{http://lope.linguistics.ntu.edu.tw/cwn2}}. This dataset is limited in size with 8,221 words. \newcite{Zheng2021DecomposeFA} constructed a dataset from the 5th edition of MCD. But it only collects disyllabic nouns and verbs, and additional annotation of formation rules is required. Besides, both datasets didn't provide the complexity of definitions, which is essential information in the controllable generation. To enhance the study of this task, we propose to build a novel benchmark dataset named \textbf{COMPILING} (\textbf{C}hinese c\textbf{OMP}lex\textbf{I}ty contro\textbf{L}lable def\textbf{IN}ition \textbf{G}eneration). The dataset is large and of high quality, which contains 127,757 entries in total. Each entry consists of a word, an example, a definition, and two complexity measurements of this definition. More specifically, we build the dataset by using two Chinese dictionaries, namely the CCLD and the 7th edition of MCD. The former collects fewer words, but the definitions are simpler. The latter is the opposite. By combining these two dictionaries, we obtain a large amount of definitions that vary in different complexities. In order to quantitatively measure the \textit{complexity} of definitions, we refer to the graded vocabularies formulated by HSK (Chinese Proficiency Test). HSK is set up to test the proficiency of non-native speakers. It has nine levels from easy to hard, and each level corresponds to a vocabulary. The COMPILING dataset contains an average level and a maximum level of each definition. We find that both dictionaries tend to use phrases rather than complete sentences as examples in some cases. For instance, the word “\chinese{规模}” (scale) has two example phrases in MCD (Modern Chinese Dictionary), which are “\chinese{规模宏大}” (large scale) and “\chinese{初具规模}” (begin to take shape). We think that short phrases might be helpful for language learners to understand, but complete sentences can provide more context in the automatic definition generation. Thus, we design an algorithm to expand the phrases into sentences (Section \ref{section:expand-algo}). We believe that this dataset can further enhance the research on Chinese complexity controllable definition generation, which could not only benefit the language learners, but also low literacy readers, as well as people with aphasia or dyslexia. We also provide baselines of mainstream generation methods as references (Section \ref{section:experiments}). In summary, our contributions are listed below: \begin{itemize} \item{We propose a novel task of generating definitions for a word with appropriate complexity. The task is of great use in helping CFL learners to learn the vocabulary.} \item{ We propose the \textbf{COMPILING} dataset that is of large scale and high quality. This dataset could serve as the benchmark of the task we proposed.} \item {We perform several experiments on the COMPILING dataset and the results demonstrate it could assist models to achieve effective complexity controllable generations.} \end{itemize} \section{ Related Work} \subsection{Definition Generation} \newcite{Noraset2017DefinitionML} first proposed the definition modeling task and use word embeddings to generate definitions of the corresponding words. Referencing on the problem of word sense disambiguation, \newcite{Ishiwatari2019LearningTD} and \newcite{Gadetsky2018ConditionalGO} incorporated word contexts into definition modeling and demonstrated its effectiveness of distinguishing different meanings. Recent work \cite{Huang2021CDMCE} reformulates the task as generating descriptions using extracted knowledge. Research on Chinese definition modeling was first proposed by \newcite{Yang2020IncorporatingSI}, they adapted a transformer-based model and incorporated sememes into the model to provide more external semantic knowledge. \newcite{Fan2020BERTChineseDM} redefined the Chinese definition modeling as generating the corresponding definition for a target word and its context. \newcite{Zheng2021DecomposeFA} utilized the characteristics of Chinese by adding formation features to enhance definition modeling. Besides, there are also studies on multilingual definition generation \cite{Kong-2020-Toward} and combining extraction and generation for this task \cite{huang-2021-cdm}. Notably, \newcite{kong-etal-2022-multitasking} proposed to generate simple definitions employing a multitasking framework. Since the lack of a definition dataset with different complexities, they managed to generate both complex and simple definitions in an unsupervised way. Differently, we focus on building the benchmark dataset for different Chinese definition generation tasks and hope it could be beneficial for further research. \subsection{Controllable Generation} Controllable generation is widely adapted in kinds of language modeling tasks. For instance, data augmentation \cite{AminNejad2020ExploringTT}, dialog generation \cite{Firdaus2020EmoSenGS}, storytelling \cite{GoldfarbTarrant2020ContentPF}, and so on. And the objects controlled in different studies vary from each other. Specifically, considering the significance of sentiment in poetry definition, \newcite{Chen2019SentimentControllableCP} proposed a model to generate poetry with controllable emotions. \newcite{Gao2019DifficultyCG} first presented a framework to develop questions about specific answers that meet target difficulty levels. To attract more readers, \newcite{Jin2020HooksIT} introduced a headline generation model to produce enticing titles with target three styles. Likewise, in order to explore and release the practical value of definition generation, we propose the complexity controllable definition generation task committed to producing definitions satisfying users of all levels. Currently, the most controllable generation tasks are achieved based on pre-trained learning models. And \newcite{Zhang2022ASO} summarized the common methods as Finetuning, Retrain PLMs, and Post-Process. And we utilize the first method to control the complexity of the definition more efficiently. \subsection{Prompt Learning} In recent years, the pre-trained model with fine-tuning has gradually become the mainstream of natural language processing tasks. Due to the complex training objectives and large hyperparameter groups, large-scale pre-training models can effectively extract features from a large amount of supervised and unsupervised data. By storing the learned knowledge in parameters and fine-tuning the model for specific tasks, the same model can be applied to a series of downstream natural language processing tasks \cite{Han2021PreTrainedMP}. Prompt learning is a method of fully learning knowledge by adding additional text to the model’s input. Prompt can be divided into artificial and automatic construction according to the text attached to the input \cite{Han2021PreTrainedMP}. Among them, automatically constructed prompts are divided into discrete and continuous ones. A discrete prompt refers to the fact that the constructed prompt is composed of actual text symbols, and applicable tasks include text classification \cite{Han2021PTRPT}, text generation \cite{Zheng2021ExploringPF}, etc. Although the combination of pre-training and fine-tuning methods can be adapted to most NLP tasks, when it comes to each specific task, the number of parameters that need to be adjusted for are vast. By adopting prompt learning, the pre-training model can be applied to the required tasks by only modifying the part of the prompt for different downstream tasks. Therefore, the training process will become more efficient. \section{Problem Formulation} In this work, we aim to generate a definition $\bm d^c$ with appropriate complexity $c$, for a given word and example sentence $(w^*, \bm e)$. This task is feasible because the word and it's corresponding definition should be assumed to have the same semantics. A common solution is to predict tokens in the definition one by one, depending on the previous words and the other conditions, which can be formulated as: \begin{equation} P(\bm d^c | w^*, \bm e, c) = \prod_{t=1}^T P(\bm d^c_t | \bm d^c_{<t}, w^*, \bm e, c), \end{equation} where $d^c_t$ is the $t$-th token in the definition, and $T$ is the total length of definition. Each probability distribution can be approximated by the following equation: \begin{equation} P(\bm d^c_t | \bm d^c_{<t}, w^*, \bm e, c) \propto \exp(Wh_t/\tau), \end{equation} where $W$ is a matrix collecting word vectors, $h_t$ is a vector summarizing inputs at current time-step, and $\tau$ is a hyper-parameter for temperature, set to 1 in default. \section{Dataset Construction} The source corpora are extracted from the MCD and CCLD, both published by the Commercial Press. For corpus from MCD and CCLD, we process them separately with the same construction methods and finally put them together. The construction of the COMPILING dataset is divided into three stages: data structured annotation, example sentences expansion, and post-processing. First, we propose a strategy for building structured datasets due to the high complexity and compact construction of automatically extracted data. In this phase, we set up a platform. It not only helps annotators proofread and audit corpus data more efficiently but is also conducive for us to check and collect data. Besides, since the context of a targeted word in the dictionary is always a collocation instead of a complete sentence, we then conduct expanding context to enhance the overall abundance of language for our proposed datasets. Furthermore, to divide definitions into different complexity levels, we calculate the HSK level of each description. \subsection{Data Structured Annotation} In the beginning, we collect initial data and find they are disorganized and complex in structure, which is problematic to conduct automatic processing. Hence, we start up data structured annotation. To better manage and boost the whole process, we build up a platform before the formal annotation and deploy it on two servers, one for corpus from MCD, and the other for corpus from CCLD. This platform could not only serve specifically for this task, but it is also appropriate for the construction of any resource by replacing the data. Concentrating on tackling the problem of disorganized data, we suggest a series of rules for annotation. For a particular word, its attached contents include its spell, definition, example sentences of the usage of a specific definition, and so on. Hence, we propose to add labels before corresponding contents to distinguish different types of data, which is conducive for computers to extract this information based on their labels automatically. Both dictionaries have instructions illustrating the meta-information, such as the organization of entries, the style of definitions and examples, and basic usages. We invite a student who majors in linguistics to formulate the annotation guidelines based on the instructions, which will be the reference for annotators. By doing so, we hope annotators could restore that language information and the relationships between them to a large extent. Then, we invite 20 students majoring in linguistics to annotate the corpora on our platform regarding the guidelines. This phase lasted for two months. \begin{algorithm}[htb] \renewcommand{\algorithmicrequire}{\textbf{Input:}} \renewcommand{\algorithmicensure}{\textbf{Output:}} \caption{Example Sentences Expansion} \label{alg:1} \begin{algorithmic}[1] \Require phrase $p$, corpus $C$ \Ensure examples $E$ \State $D \gets \{\}, E \gets []$ \For{ $sentence$ \textbf{in} $C$ } \If {$p$ \textbf{in} $sentence$} \State $score \gets pplScore(sentence)$ \Comment Compute the PPL score for each sentence. \State $D[sentence] \gets score$ \EndIf \EndFor \State $sortedExamples \gets descSortByValue(D)$ \Comment Descendant sort by the scores. \For {$i=0 \to topN$} \Comment $topN$ is set to 5 in practice. \State $E.add(sortedExamples[i])$ \EndFor \end{algorithmic} \end{algorithm}\label{contextexpansion} \subsection{Example Sentences Expansion} \label{section:expand-algo} While the information extracted from dictionaries is large and abundant, the context attached to the targeted words given in dictionaries is too short to provide enough knowledge for the model to learn and generate descriptions. In the second stage of construction, considering the significance of sentences, we start up example sentence expansion. For contexts without sufficient length in the original corpus, we tend to find sentences with a longer length and higher quality in the new canon for replacement, and the specific process is as follows. We first screened each example sentence in the annotated texts. We set the length threshold to six, and if the length of the initial context is longer than the threshold, we will retain the sentences; otherwise, we will find longer sentences with more abundant information in the new corpus to cover the original ones. It is worth noting that if a term contains more than one sentence (collocation), for each sentence (collocation), we will replace it with new matching contexts. We design Algorithm \ref{alg:1} to match and gain new high-quality sentences. Given the ambiguity of most words, we utilize an allocation as the input of Algorithm \ref{alg:1} instead of a phrase to ensure the found sentences contain the corresponding usage of a specific definition. As shown in Algorithm \ref{alg:1}, we collect all the sentences that fit the requirements and grade them by utilizing Perplexity (PPL)\footnote{\url{https://huggingface.co/docs/transformers/perplexity}}, which is one of the most common metrics for evaluating language fluency. Eventually, the top five sentences in the rankings are designated to replace those original short contexts. \subsection{Post Processing} \label{hsk} \paragraph{Difficulty classification} The most crucial step of constructing a complexity-controlled dataset is integrating the difficulty level of definition into the dataset. We utilize the HSK metric to represent the complexity degree. HSK\footnote{\url{http://www.chinesetest.cn}}, called the Chinese Proficiency Test, set to evaluate the Chinese proficiency and application of non-native speakers. It is divided into nine levels, and the difficulty increases progressively from low to high. For convenience, we regard the seventh, eighth, and ninth levels as a whole. Finally, we set seven complexity levels of HSK, and each level corresponds to a vocabulary. For words that are not included in the first seven-level, we classify them as the highest level. \paragraph{Entry construction} Besides, For each definition, we first conduct word segmentation, then calculate the average and highest HSK level, and combine the HSK level into the dataset. Eventually, each entry of the COMPILING dataset consists of a target word, its definition, the average and highest HSK level, and the contexts of the corresponding usage of this description. \section{Dataset Analysis}\label{dataset} \begin{table}[htbp] \centering \setlength{\tabcolsep}{12pt} \caption{The main statistics of the COMPILING dataset.} \begin{tabular}{lcrrcrr} \toprule[1pt] \multirow{2}{*}{Datasets} & & \multicolumn{2}{c}{Count} & & \multicolumn{2}{c}{Average Length} \\ \cmidrule{3-4} \cmidrule{6-7} & & Words & Entries & & Definition & Context \\ \midrule MCD & & 67,801 & 101,314 & & 13.8 & 27.5 \\ CCLD & & 6,502 & 26,443 & & 13.4 & 20.4 \\ \bottomrule[1pt] \end{tabular}% \label{tableone}% \end{table}% As mentioned before, the smallest unit of the COMPILING dataset consists of five parts. In particular, if a word is polysemous or has numerous contexts, they are regarded as distinct entries. For instance, as shown in Table \ref{addlabe1}, the word “\chinese{收拾}” (clear up) has four different definitions, and each of them follows an example sentence. Hence there are four entries of “\chinese{收拾}” (clear up) in total. As shown in Table \ref{tableone}, we analyze statistics of data extracted from MCD and CCLD, respectively. Table \ref{tabletwo} shows the basic statistics of the COMPILING dataset and another dataset of Chinese definition modeling. For training, the given definitions of each entry are seen as the ground truth. \begin{table}[htbp] \centering \caption{Example entries of COMPILING dataset.} \resizebox{\textwidth}{!}{ \begin{tabular}{cccccc} \toprule Word & Definition & Average & Maximum & Sentence & Source \\ \midrule \makecell[c]{\chinese{收拾}\\clear up} & \makecell[c]{\chinese{使变干净整齐；整理} \\To make clean and tidy} & 2 & 3 & \makecell[c]{\chinese{东西都收拾好了，可以出门了。}\\ With everything packed up, \\we're ready to go.} & CCLD \\ \midrule \makecell[c]{\chinese{收拾}\\repair} & \makecell[c]{\chinese{使有毛病的东西功能正常；修理}\\To make something\\ defective function properly} & 2 & 4 & \makecell[c]{\chinese{我的手机坏了，得找厂家收拾一下。}\\My phone is out of order so I have to ask\\ manufacturer for help.} & CCLD \\ \midrule \makecell[c]{\chinese{收拾} \\settle} & \makecell[c]{\chinese{整理；整顿} \\Put in order} & 4 & 6 & \makecell[c]{\chinese{冬储夏衣，夏藏冬衣，收拾屋子，还要照看外孙女。}\\Store summer clothes in the winter, hide\\ winter clothes in the summer, clean \\the house, and look after her granddaughter.} & MCD \\ \midrule \makecell[c]{\chinese{收拾}\\kill} & \makecell[c]{\chinese{消灭；杀死} \\Eliminate} & 8 & 10 & \makecell[c]{\chinese{据点的敌人，全叫我们收拾了。} \\All the enemies in the stronghold \\have been eliminated.} & MCD \\ \bottomrule \end{tabular}}% \label{addlabe1}% \vspace{2.0em} \end{table}% \begin{table}[!htbp] \vspace*{-1cm} \centering \setlength{\tabcolsep}{12pt} \caption{Statistics of Chinese definition modeling datasets.} \footnotesize \begin{tabular}{lrrrrrr} \toprule[1pt] \multirow{2}{*}{Datasets} & & \multicolumn{2}{c}{Count} & & \multicolumn{2}{c}{Average Length} \\ \cmidrule{3-4} \cmidrule{6-7} & & Words & Entries & & Definition & Context \\ \midrule CWN & & 8,221 & 84,542 & & 9.07 & 21.57 \\ \textbf{COMPILING} & & 74,303 & 127,757 & & 13.60 & 23.95 \\ \bottomrule[1pt] \end{tabular}% \label{tabletwo}% \end{table}% To better highlight the complexity degree of the dataset, we set levels 1-3 in HSK as the simple grade, levels 3-7 as the medium grade, and levels 7-9 and 9+ as hard quality. We count the HSK level distribution of definitions of COMPILING, as shown in Figure \ref{distribution}. \begin{figure}[ht] \centering \includegraphics[width=.8\textwidth]{imgs/levels-crop.pdf} \caption{The distribution of average HSK level in CWN and COMPILING.} \label{distribution} \end{figure} The distribution of definitions in the COMPILING dataset in the three levels is closer than CWN. Given the particularity of the Complexity Controllable definition generation task, it is necessary to construct a dataset including entries covering all difficulty levels. In this way, the model can learn and distinguish the complexity of descriptions, hence generating a new definition of a word with a target complexity level. Hence, the COMPILING dataset could be applied to both general definition generation tasks and those which incorporate the complexity of definitions, demonstrating its value in being as a benchmark dataset. \section{Experiments}\label{section:experiments} \subsection{Baselines}\label{model} This section introduces several methods for common generation tasks, which can serve as baselines for our proposed task. \paragraph{LOG-CaD} LOG-CaD \cite{Ishiwatari2019LearningTD} is a model for generating descriptions for words and phrases. This model summarizes clues from the static, contextualized, as well as character-level embeddings of the given word, and then employs an LSTM-decoder for the generation. A gated attention mechanism is employed to capture and filter information from the embeddings during decoding. \paragraph{Transformser} We treat the task as a special type of single language translation and directly use the original transformer model proposed by \newcite{Vaswani-2017-attention}. We concatenate the word and example sentence as the input sequence and train the model to generate the definition. We use the same approach to deal with the input and output in BERT and BART models. All hyper-parameters are set according to the original paper for a fair comparison. \paragraph{BERT} Pretrained language models have been widely used in various NLP tasks in recent years. By obtaining prior knowledge during pretraining, the PLMs can encode the input sentence more effectively. Thus, we use the Chinese-bert-base \cite{Devlin-2019-bert} model to initialize all the parameters in a transformer encoder and employ a transformer decoder for the generation. Note that the decoder is trained from scratch without initialization. \paragraph{BART} Unlike BERT, BART \cite{Lewis-2019-bart} is a pretrained encoder-decoder language model, which is more suitable for generation tasks. Since the monolingual BART only support English, we use the multilingual version of BART and set both source and target language as Chinese for this task. \paragraph{MT5} T5 is one of the representative pre-training language models. It considers all NLP tasks as a uniform text-to-text paradigm. mT5 \cite{Xue2021mT5AM} is a multi-language variant of T5, and its performance on various benchmark tasks is generally outstanding. Therefore, we choose mT5 to perform the prompt learning method. \begin{table}[h] \centering \setlength{\tabcolsep}{15pt} \caption{Datasets divided by HSK level.} \begin{tabular}{lcc} \toprule[1pt] Complexity & HSK & Entries \\ \midrule Easy & 1-3 & 48,458 \\ Medium & 4-7 & 53,945 \\ Hard & 7+ & 25,354 \\ \bottomrule[1pt] \end{tabular}% \label{tablefour}% \end{table}% \subsection{Settings}\label{sett} As a benchmark dataset introduced to enhance the Chinese definition generation task, we set up the experiments to verify the effectiveness of the COMPILING dataset. \paragraph{Regardless of complexity levels} We first design the experiment to evaluate the overall performance of the baseline models on our dataset. In this setting, we train the models using the entire training set, despite of the different complexity levels. And the purpose of this setting is to provide a comparison standard for other experiments. We divide the dataset into training, development, and test sets according to 8:1:1. The training data are fine-tuned according to the input formats of different models. \paragraph{Complexity specific models} To evaluate the significant role of the COMPILING dataset in generating definitions across various difficulties, we set up an experiment to train the model on different complexity-level sub-datasets. First of all, we split the dataset into three subsets on basis of the average HSK level. As shown in Table \ref{tablefour}, the HSK levels of definitions in Easy Set are between 1 to 3, Medium Set corresponding to level 4-6, and Hard Set corresponding to level 7+. Then we split each subset into training, development, and test sets according to the ratio of 8:1:1. Finally, we fine-tune the BART model utilizing these three training sets, and hence getting three models. Each one could generate definitions with its corresponding complexity level. \paragraph{Unified model based on prompt learning} To assist the model to generate descriptions with different complexity of demand, we adopt the method of prompt learning. It allows the model to learn by adding tokens that represent difficulty information to the inputs, such as <extra\_id\_1> for level 1 (lowest), <extra\_id\_2> for level 2, and so on. The training set is formed by prefacing each definition of the COMPILING dataset with the corresponding special tokens. Each entry of the final dataset includes: <extra\_id\_x>, target word, its corresponding definitions and context. During the training phase, the model encodes both complexity and definition information. In the analysis stage, aiming to verify the effectiveness of this method, we select 10 entries from the test set of the COMPILING dataset. For each entry, only its difficulty token is modified with the other information keep remaining, so as to construct two copies of the entry. It is worth noting that the principle of constructing the new complexity tokens is, that the two new entries and the original one(a group of data) differ by at least 2 levels or more, which means they can represent easy, medium, and hard complexity respectively. For example, if the definition of the source entry is specified with the difficulty as 3, the complexities of the two copies of it need to be constructed as at least 1 and 5. Finally, a total of 30 entries are included in the new test set. Then, we perform the model on this new test set to observe whether the generated definitions are differentiated in line with their specified complexity. \subsection{Evaluation Metrics} In order to better analyze and quantify the experimental results, we select three evaluation metrics: BLEU \cite{Papineni2002BleuAM}, NIST \cite{Doddington2002AutomaticEO} and HSK, which are used to comprehensively evaluate the quality and complexity level of generated definitions. \paragraph{BLEU} BLEU (Bilingual Evaluation Understudy) \cite{Papineni2002BleuAM} was originally proposed for the evaluation of machine translation research. The core of BLEU is to separately calculate the N-gram in the generated and the reference sentence, and then compare them one by one to count the times that can be matched. The higher times illustrate higher accuracy. However, the shorter reference segment always leads to more co-occurrence times, which means the shorter generated definitions tend to get a higher BLEU score. \paragraph{NIST} On the basis of BLEU, NIST (National Institute of Standards and Technology) \cite{Doddington2002AutomaticEO} adds the calculation of the information weight of N-gram. While the BLEU simply sums up the number of N-grams, the NIST sums up the information weights and then divides it by the number of N-gram segments in the whole sentence. In this way, the weightage of those N-grams which appear less frequently will be heavier. \paragraph{HSK} As mentioned in section \ref{hsk}, HSK is a test set to evaluate the Chinese proficiency and application ability of non-native Chinese speakers. Based on the purpose of assisting CFL learners to understand Chinese well, we select HSK to measure the complexity level of definitions. Besides, we set seven difficulty levels (scores) of HSK and each of them corresponds to a vocabulary. The final level of a definition is determined by the average score of its segments. \subsection{Results and Analysis} \label{performance_on_model} \paragraph{Regardless of complexity levels} We report the experimental results on the entire COMPILING dataset in Table \ref{olala}. The results show that PLMs outperforms the other two methods in terms of the BLEU and NIST scores apparently. However, the results of BERT and BART models diverged on these two metrics. Since NIST assigns different weights to tokens, we believe it better reflects the model’s performance. We confirmed this by reading the generated samples. \begin{table}[htbp] \centering \caption{Experiment results on the COMPILING dataset.} \begin{tabular}{lccccccc} \toprule \multirow{2}{*}{ Models} & \multicolumn{3}{c}{Dev} & & \multicolumn{3}{c}{Test} \\ \cmidrule{2-4} \cmidrule{6-8} & BLEU & NIST & HSK & & BLEU & NIST & HSK \\ \midrule LOG-CaD & 27.66 & 25.55 & 3.74 & & 27.71 & 27.88 & 3.85 \\ Transformer & 28.61 & 25.85 & 3.92 & & 28.58 & 31.00 & 3.96 \\ BERT & \textbf{32.95 } & 29.66 & 4.05 & & \textbf{32.03} & 30.56 & 4.08 \\ BART & 29.49 & \textbf{36.90 } & \textbf{4.76} & & 30.63 & \textbf{42.79} & \textbf{4.80} \\ \bottomrule \end{tabular}% \label{olala}% \end{table}% We also notice that as the model performance improves, so does the average HSK level of the generated definitions. This phenomenon is because simpler words are used more frequently, and hence are more easily learned by models. As the modeling ability improves, the better-performing models learn to use more complex words. This can be challenging for future complexity controllable definition generation works, i.e., improving the performance and reducing the generation complexity at the same time. \paragraph{Complexity specific models} Table \ref{multi} illustrates experiment results on three different subsets. As listed in the table, we not only test on the subset in which the model is trained, but also on other subsets. Generally, all the models perform best on the subset it was trained, and poorly on other subsets. Moreover, the performance decays as the complexity level between the model and data increases. Definitions with different complexity have different lexical and syntax, resulting in poor cross-complexity generalization. Besides, we found that even on different test sets, definitions generated by the same model have similar complexity. \begin{table}[hbp] \centering \caption{Experiment results in terms of complexity controllable generation on three test sets.} \begin{tabular}{lccccccccccc} \toprule \multirow{2}{*}{Models} & \multicolumn{3}{c}{Easy Set} & & \multicolumn{3}{c}{Medium Set} & & \multicolumn{3}{c}{Hard Set} \\ \cmidrule{2-4}\cmidrule{6-8}\cmidrule{10-12} & BLEU & NIST & HSK & & BLEU & NIST & HSK & & BLEU & NIST & HSK \\ \midrule BART-Easy & \textbf{32.44 } & \textbf{64.40 } & 2.40 & & 21.56 & 27.61 & 2.73 & & 25.89 & 7.95 & 2.74 \\ BART-Medium & 22.92 & 24.59 & 4.70 & & \textbf{27.69 } & \textbf{40.68 } & 4.86 & & 29.37 & 16.09 & 5.01 \\ BART-Hard & 22.49 & 3.55 & \textbf{8.46} & & 23.70 & 7.04 & \textbf{8.45} & & \textbf{46.57 } & \textbf{18.22 } & \textbf{8.76} \\ \bottomrule \end{tabular}% \label{multi}% \end{table}% \paragraph{Unified model based on prompt learning} MT5-base \cite{Xue2021mT5AM} was selected as the benchmark model in this experiment. The best PPL obtained from the definitions generated on the validation set is 38.44. The BLEU and NIST of the model on the test set are 27.42 and 4.66, respectively. The model generates interpretations based on the new test mentioned in Section \ref{sett}. Table 7 lists two examples where it is fairly obvious that the resulting definitions are differentiated and conform to the expectations for their specified complexity levels. To evaluate the complexity of generating definitions more accurately, we adopt automatic evaluation, ranking the difficulty of each group\footnote{Each group of data refers to one original entry and its two copies, their specified complexity of definition is different and other information keep the same.}. The automatic evaluation is based on the Chinese Text Complexity Analysis Platform (CTAP)\footnote{\url{http://ctap.wenmind.net}} \cite{cui-2022-ctap}. We selected the features of word diversity and word density that reflect the difficulty of paraphrases and calculated the scores of definitions in each group based on the above features. Finally, the scatter distribution diagram is shown in Figure \ref{scatter}. \begin{figure*}[htbp] \centering \includegraphics[width=0.85\textwidth]{imgs/groups-crop.pdf} \caption{The automatic evaluation results. For example, the scatters of the Hard Group represent those definitions that are specified as the hardest, and the ordinate corresponds to the scores obtained by the automatic rating.} \label{scatter} \end{figure*} It can be seen that the complexity score of the Hard Group is mainly above 5, and the number of definitions with the highest score is the largest. The definition in the Easy Group scored the lowest overall score. This means the difficulty level of the model-generated interpretations obtained by automatic evaluation is roughly in line with expectations. The result proves the effectiveness of prompt learning on complexity controllable task, but since the difference in the overall distribution of scattered points in each group in the figure is not particularly obvious, it also reflects that there is room for exploration and improvement of this task in the future. \section{Conclusion} In this work, we propose a novel task of generating Chinese complexity controllable definitions for a given word and example sentence. This task is of great use in helping CFL learners and low literacy readers. Meanwhile, we introduce the COMPILING dataset, which is a benchmark adapting to kinds of definition generation tasks. We also provide several baselines for this task, among which the prompt learning method better assist models in generating definitions with specified complexity. Nevertheless, the experimental results also show that this task is challenging, and the performance needs further improvement. \section{Bibliographical References}

a491c2978ab73235572135a61177321b1dd4f543

\section{List of calibration lines} \label{sec:arc-lamp-lines-list} \begin{table}[h] \centering \begin{tabular}{lllllll} \input{arclinelist} \end{tabular} \caption{List of the arc lamp emission lines used for the wavelength calibration and PSF fit. Wavelength are in \AA\ in vacuum. Ambiguous lines have been discarded, but some blended lines have been kept as they are correctly handled during the PSF fit. Most faint lines have been ignored except in the wavelength regions where brighter lines are missing or when they are blended with other lines and could bias the fit.} \label{table:arc-lamp-lines-list} \end{table} \section{CCD read noise} \label{sec:ccdproperties} \input{spectrographs-table} \subsection{Co-addition} \label{sec:coaddition} For some of the delivered data products, the spectra from the same target are co-added, i.e. averaged, per camera. Co-added spectra from all of the past observations of a tile are computed as part of the daily processing and for data releases. We also provide co-added spectra of targets observed on multiple tiles (for instance Lyman-$\alpha$ QSOs), and group them per {\it Healpix} pixel \citep{Healpix}; see~\S\ref{sec:data-products} for more details about the data products. The fact that the wavelength grid is the same for all fibers in all exposures (and across cameras for their overlapping wavelength) simplifies greatly the co-addition over exposures: it is a simple weighted average per wavelength element, optionally with outlier rejection. The weights are the inverse variance of the calibrated spectra. They are the statistically optimal weights for non-variable objects. For instance exposures with poor sky transparencies are properly de-weighted as the inverse variance of the calibrated spectra scales with the square of the sky transparency. Co-added spectra still have uncorrelated fluxes because no re-sampling is involved in the process. The resolution matrix of the co-added spectra is the weighted mean of the resolution matrices of the input spectra, with the same weights as for the flux. This is obvious if one considers that the expected values of the extracted flux $F$ are the same for all input spectra (contrary to the recombined fluxes $\tilde F$ which depend on the resolution matrices of the input spectra, see \S\ref{sec:extraction}). $$ \langle \tilde F_{coadd} \rangle = \sum_i w_i \langle \tilde F_i \rangle = \left( \sum_i w_i R_i \right) \langle F \rangle \equiv R_{coadd} \langle F \rangle $$ For the co-addition of DESI main survey spectra, we use an outlier rejection threshold of 4 standard deviation as a last barrier to discard spurious spikes due to undetected cosmic ray hits. We do not co-add spectra across cameras before fitting redshifts to avoid averaging data of different resolution and losing information for the detection of sharp spectroscopic features. \section{Overview of the data products} \label{sec:data-products} The primary user-facing data products are sky-subtracted flux-calibrated spectra, coadds of those spectra across exposures, and tables of spectral identification (galaxy, quasar, star) and redshift measurements. The spectra are stored in multi-HDU FITS files\footnote{\url{https://fits.gsfc.nasa.gov}}, with separate extensions for each of the $b,r,z$ cameras for each of the quantities wavelength, flux, inverse variance, mask, and resolution data. A ``fibermap'' binary table tracks which targets are assigned to which fibers along with targeting photometry, shapes, target selection masks, and related metadata; and per-fiber per-observation data such as the focalplane ($x,y$) location of each target and its distance from the assigned position. An additional ``scores'' binary table records TSNR-squared values (see \S\ref{sec:efftime}) and the median and summed counts (or flux) at various steps of the processing: raw, fiber flat-fielded, sky subtracted, and fully calibrated spectra. The coadd files contain the same wavelength, flux, inverse variance, and resolution data per camera, coadded across multiple input exposures for the same target. In the coadd files, the fibermap is split into a table of information that is the same for every exposure of a target (e.g.~photometry), plus a separate table that contains the per-exposure information that contributed to the coadd (e.g.~exposure IDs and the per-exposure $(x,y)$ locations). A key feature of both the spectra and the coadds is that all targets use the same wavelength grid, and the individual wavelength bins are uncorrelated. See \S\ref{sec:extraction} for more details about the validity of the error (inverse variance) model and the uncorrelated nature of the spectra. See also Appendix~\ref{sec:resampleresolution} for the best use of the resolution matrix provided in the files. The redshift files contain a binary table with columns listed in Table~\ref{table:redshift-columns}. \begin{table}[h]\small \begin{tabular}{l|l} Z & redshift\\ ZERR & estimated error\\ ZWARN & warning mask (0=good)\\ SPECTYPE & spectral type\\ SUBTYPE & subtype for stellar templates\\ DELTACHI2 & $\chi^2$ difference between the best\\ & and the second-best model\\ NPIXELS & number of unmasked pixels\\ & contributing to the fit\\ NCOEFF & number of template coefficients\\ COEFF & PCA template coefficients\\ CHI2 & absolute $\chi^2$ \end{tabular} \caption{List of columns in the redshift files.} \label{table:redshift-columns} \end{table} Additional details including the $\chi^2$ vs.~$z$ scans and the best 3 fits for each SPECTYPE (GALAXY, QSO, STAR) are stored in a separate HDF5\footnote{\url{https://www.hdfgroup.org}} file. The spectra, coadd, and redshift files are grouped on disk either per-tile\footnote{A DESI ``tile'' is a single telescope pointing with a specific set of positioner locations.} or per-healpix \citep{Healpix} (default nested nside=64), with subgroups for each category. Tile-based spectra combine data across multiple exposures for a single tile, but not across tiles even if the same target was observed on multiple tiles (e.g.~Lyman-$\alpha$ QSOs). Healpix-based spectra combine data even across tiles, but not across surveys and programs (surveys being among {\it cmx}, for commissioning, {\it sv1}, {\it sv2}, {\it sv3}, and {\it main}; and programs being among {\it dark}, {\it bright}, {\it backup}, and {\it other}, see~\citealt{Myers22a} for more details). These provide all the data for a given (survey,program) for a given healpix on the sky without having to read multiple files covering spatially overlapping data. Tile-based spectra are provided per-exposure (in which case the coadd is the same data as the uncoadded single exposure spectra); per-night (combining all exposures on a single night; if only one exposure was observed it is the same as the per-exposure grouping); and cumulative (all exposures for the tile over all nights; if observed on only one night it is the same as the per-night group). The uncoadded cumulative spectra contain the exact same data as the combination of all of the per-night spectra, which in turn is the same as the combination of all of the per-exposure spectra. The coadds and redshifts, however, differ based on exactly which spectra were included in the coadd. Even though various spectra files are subsets of others, they are provided for end-user convenience to directly map which spectra were combined into which coadds for which redshifts. Healpix-based coadds combine data across tiles, including coadding spectra of the same target observed on different tiles using different fibers and different spectrographs. Healpix-based files are split by survey and program. e.g.~a subset of the SV targets also pass the main survey criteria and will be re-observed, but these two surveys are processed separately to prioritize the homogeneity of the individual surveys. Similarly, bright LRG targets (program=DARK) can also pass BGS cuts (program=BRIGHT), but are processed separately. We anticipate that the tile-based spectra/coadds/redshifts will be of primary interest for SV1 analyses where the tiles were planned and observed to be independently useful, while the healpix-based spectra/coadds/redshifts will be of primary use for the main survey and SV3 where overlapping groups of tiles are intended to be analyzed together. The Early Data Relase (EDR) and the matching ``Guadalupe'' run of the future Data Release 1 (DR1) will include all flavors of spectra/coadds/redshifts as described here. Other future data releases may provide only healpix and cumulative per-tile redshifts for main survey data, while still providing the full suite of groupings for SV data which were purposefully observed to greater depth than typical and thus make per-night and per-exposure comparisons to cumulative more useful than for the main survey data. In addition to the user-facing data described above, the pipeline includes outputs for per-night fiberflat and PSF fit calibrations, preprocessed CCD images, and intermediate files such as uncalibrated extracted spectra and fiberflatfielded sky-subtracted but not flux-calibrated spectra. These are useful for QA, debugging, and specialized analyses, but are not intended for most end-users. Full details of these data formats and directory structures are documented at \url{https://desidatamodel.readthedocs.io}. \subsection{Effective exposure time} \label{sec:efftime} One key ingredient of the survey optimization is the definition of a spectroscopic average signal to noise ratio (hereafter SNR) and its associated effective exposure time (proportional to the SNR squared). Once this quantity is defined and a minimal threshold is set, the role of the survey operations is to achieve this minimal SNR for all of the tiles (or pointings) that compose the survey in a minimum amount of time (see~\citealt{schlafly22a}). An ideal survey would have all the tiles observed exactly at the threshold; this would both minimize the survey duration and ensure its homogeneity in terms of redshift efficiency. Among an infinity of possible choices, we have resorted to consider for the SNR definition a quantity related to the inverse of the redshift measurement uncertainty for an ensemble of reference targets. We call this quantity TSNR (for Template SNR) and define it more precisely as a quadratic average for an ensemble of template spectra that follow a predefined redshift distribution. We have defined TSNR values for several target classes (LRG, ELG, QSO, BGS, Lyman-alpha forests) but we can use only one of them to decide on the exposure time. TSNR values are calculated for each fiber of each observation, and we use the average of the LRG TSNR values over all the fibers of a given exposure to define its effective exposure time. The TSNR values are measured per camera, fiber and observation as follows: $$ TSNR^2 = \sum_i T_i^2 \left< (\delta F)^2 \right>_i / \sigma_i^2 $$ where $i$ is a spectral wavelength index, $T_i$ the throughput times the exposure time, $\sigma_i$ the flux measurement uncertainty. $\left< (\delta F)^2 \right>$ is an average over an ensemble of templates that differs for each target class. $\delta F = F - med(F)$ is one template spectrum minus a median filtered version of the same spectrum (with a width of 100\,\AA). $\delta F$ is close but different from the derivative of $F$ with respect to the redshift; it was found to be a better indicator of the redshift success rate, which is defined by a minimal $\chi^2$ difference between the best and second best solution of the redshift fit (see~\S\ref{sec:redshift}). The quantity $\left< (\delta F)^2 \right>$ is shown as a function of wavelength for each target class on Figure~\ref{fig:tsnr-template-deltaf}. The redshift range considered for each target class is given in Table~\ref{table:tsnr-zrange}. \begin{table} \centering \begin{tabular}{ll} Target & Redshift range \\ \hline BGS & $0.13 - 0.37$ \\ LRG & $0.68 - 0.97$ \\ ELG & $0.80 - 1.40$ \\ QSO & $0.56 - 1.93$ \\ Lyman-$\alpha$ & $> 2$ \end{tabular} \caption{Redshift range used to compute the term $\left< (\delta F)^2 \right>$ entering in the definition of TSNR (Template Signal to Noise Ratio) for each target class.} \label{table:tsnr-zrange} \end{table} $T_i$ and $\sigma_i$ are derived from the observations. $T_i$ is the throughput term, it includes the fiber aperture losses (see \S\ref{sec:fiber-aperture-loss-correction}). For this calculation, a single reference angular size is considered for each target class\footnote{Half-light radii of 0.45\arcsec, 1\arcsec\ and 1.5\arcsec\ for ELG, LRG and BGS galaxies respectively.} but the specific plate scale and positioning error of each individual fiber, along with the specific atmospheric seeing of the observation is used. $T_i$ also contains the Galactic dust transmission. $\sigma_i$ comprises a read noise term, different for each amplifier of each CCD, with an effective number of CCD pixels derived from the PSF shape, and a Poisson noise term derived from the sky spectrum model (see \S\ref{sec:skymodelfit}), taking into account the specific flat field correction of each fiber (see \S\ref{sec:fiberflat}). \begin{figure} \centering \includegraphics[width=0.99\columnwidth,angle=0]{tsnr-deltaf.pdf} \caption{Templates $\sqrt{\left< (\delta F)^2 \right>}$ used to compute the TSNR values for LRG, ELG, QSO, Lyman-$\alpha$ and BGS targets (see Table~\ref{table:tsnr-zrange} for the redshift ranges).} \label{fig:tsnr-template-deltaf} \end{figure} The correlation between the median LRG TSNR$^2$ per exposure and the median $\Delta \chi^2$ from Redrock (see \S\ref{sec:redshift}) is shown on Figure~\ref{fig:tsnr2-vs-dchi2}. The median has been computed for the LRG targets only. A correlation coefficient of 0.97 is found, demonstrating that the TSNR values are a good indicator of the redshift success. A slightly lower correlation coefficient is obtained if the TSNR value for ELGs is used instead to the one for LRGs (the difference is due to the different weights as a function wavelength given by the values of $\left< (\delta F)^2 \right>$, the ELGs being more sensitive to the signal to noise in near infrared, see Figure~\ref{fig:tsnr-template-deltaf}). \begin{figure} \centering \includegraphics[width=0.99\columnwidth,angle=0]{deltachi2-vs-tsnr2.pdf} \caption{Comparison of the LRG TSNR$^2$ per exposure and the corresponding $\Delta \chi^2$ from Redrock. Each dot corresponds to one exposure; both the TSNR$^2$ and $\Delta \chi^2$ values are the median over the LRG target spectra observed during that exposure. The vertical dashed line indicates the TSNR$^2$ value which corresponds to an effective time of 1000 seconds.} \label{fig:tsnr2-vs-dchi2} \end{figure} As the signal to noise ratio increases as the square root of the exposure time if the noise is dominated by the Poisson noise from the sky background and not the read noise, it is natural to define an effective exposure time proportional to the TSNR$^2$ values. The normalization is such that the effective exposure time corresponds to an actual exposure time when observing in nominal conditions, for a dark sky, ideal transparency, at zenith, without Galactic dust extinction, and for a median seeing of 1.1\arcsec. Based on this, we define the following {\it spectroscopic effective time} \begin{equation} T_{\rm spec} = ( 12.15\,\sec )\times TSNR_{LRG}^2 \label{eq:efftime-spec} \end{equation} It is measured from the spectroscopic pipeline output for each observed tile, and used to verify the quality of the observations. It is interesting to compare $T_{\rm spec}$, obtained from the spectroscopic data only, with an estimate derived from other inputs, namely the transparency and fiber acceptance derived from the GFA images in r-band, along with a model of the atmospheric transmission with airmass, and the extinction. For this purpose we define another effective time associated with the r-band, \begin{eqnarray} T_{\rm r} & \equiv & T_{exp}\left( \frac{ft}{ft_{nom}}\right)^2 \frac{s_{nom}}{s} \nonumber \\ && \times 10^{\frac{-2}{2.5}( 0.114 (X-1) + 2.165 E(B-V))} \label{eq:efftime-r} \end{eqnarray} where $T_{exp}$ is the actual exposure time, $ft$ is the product of the fiber acceptance (which depends on seeing) and the sky transparency, and is obtained from the guide star images, $s$ is the sky flux measured in r-band (here we do use the spectroscopic sky measurement), $X$ the airmass, and $E(B-V)$ the reddening due to Galactic extinction. The subscript {\it nom} indicates the value in nominal conditions. $T_{\rm spec}$ and $T_{\rm r}$ values are compared in Figure~\ref{fig:teff}; they are tightly correlated. \begin{figure} \centering \includegraphics[width=0.99\columnwidth,angle=0]{efftime.pdf} \caption{Effective time derived from the spectroscopic TSNR values ($T_{\rm spec}$, see Eq.~\ref{eq:efftime-spec}) as a function of the effective time $T_{\rm r}$ from Eq.~\ref{eq:efftime-r}.} \label{fig:teff} \end{figure} We track both quantities $T_{\rm r}$, $T_{\rm spec}$ along with the real-time effective time derived from the exposure time calculator as part of the daily quality assessment of the observations. Another exposure is required in the rare cases where $T_{\rm spec}<850$~seconds for a tile in the main dark time program (or 153~seconds for the bright time program). Note that a tile can be composed of one or several exposures from which the spectra have been co-added (see \S\ref{sec:coaddition}). \subsection{Spectral extraction} \label{sec:extraction} \subsubsection{2D PSF spectral extraction} Extracting 1D spectra from the 2D CCD images follows the ``spectroperfectionism'' methodology described in \cite{BoltonSchlegel2010}. CCD pixels $p$ are modeled as a linear combination of the input spectral flux $f$ on a discrete wavelength grid such that \begin{equation} \label{eq:projection} p = A f + \text{noise} \nonumber \end{equation} The PSF model is encoded in the matrix $A$, {\it i.e.}~modeling how the input flux $f$ (interpreted as a series of $\delta$-functions) is distributed across the CCD pixels $p$, including the fact that an individual pixel can have contributions from multiple wavelengths and multiple fibers. In this notation the 2D array of pixels are flattened into a single 1D array of pixels $p$, and the 2D array of fluxes vs.~fiber and wavelength are flattened into a single 1D array $f$. The $A_{ij}$ element is the PSF for fiber wavelength $\lambda_j$ integrated over CCD pixel $i$. The optimal estimator of $f$ is obtained by maximizing the likelihood $L$ of $p$ or minimizing \begin{equation} \chi^2 = -2 \ln L + \mathrm{const} = (p-A f)^T W ( p-Af) \label{eq:extractionchi2} \end{equation} where $W$ is the inverse of the covariance of the pixels $p$. The pixel noise is uncorrelated to a good approximation so $W$ is diagonal and its elements are the inverse of the pixel variance (see \S\ref{sec:ccd-variance} for how the CCD pixel-level variance is estimated). Minimizing Eq.~\ref{eq:extractionchi2} with respect to $f$ consists of solving the following linear system \begin{equation} (A^T W A) f = A^T W p \label{eq:extract-f} \end{equation} The quantity $(A^T W A) = C^{-1}$ is the inverse covariance matrix of $f$, which in general has off-diagonal covariance, {\it i.e.} the elements of $f$ are correlated. Following \cite{BoltonSchlegel2010}, with a eigen decomposition $C^{-1} = P^T D^{-1} P$, one can define the square root matrix $Q = P^T D^{-1/2} P$, the diagonal matrix $S$ with $S_{i,i}= 1/\sum_j Q_{ij}$, and construct a {\it resolution matrix} $R$, \begin{equation} \label{eq:resolution-matrix} R = S Q = S P^T D^{-1/2} P \end{equation} This matrix can be used to recombine the correlated flux values into an uncorrelated array, \begin{equation} \tilde f \equiv R f \label{eq:ftilde} \end{equation} Indeed the covariance $\tilde C$ of $\tilde f$ is diagonal: \begin{eqnarray} \tilde C &\equiv& \left< (\delta \tilde f) (\delta \tilde f)^T \right> = R C R^T \nonumber \\ &=& S P^T D^{-1/2} P \left( P^T D P \right) P^T D^{-1/2} P S \nonumber \\ &=& S^{2} \nonumber \end{eqnarray} The final outputs are: the resolution convolved flux $\tilde f$ which is by construction uncorrelated between wavelengths; the variances $\sigma_\lambda^2$ which are the diagonals of $\tilde C$ (off-diagonal elements are 0); and the resolution matrix $R$ that was used to decorrelate the data. These can be combined to compare a high resolution model spectrum $m$ to the resolution convolved extracted spectra $\tilde f$ via $R m$. e.g.~as a $\chi^2$: \begin{equation} \label{eq:extract-chi2-1D} \chi^2_{\mathrm{1D}} = \sum_{\mathrm{wavelengths}~\lambda} \left( {\tilde{f} - (R m)_\lambda \over \sigma_\lambda}\right)^2 \end{equation} The fluxes of Eq.~\ref{eq:ftilde} obtained with the resolution matrix of Eq.~\ref{eq:resolution-matrix} are not only decorrelated across wavelength but also across fibers. What seems a good thing at first glance has the undesired effect of mixing the true underlying flux of neighboring fibers, and introducing fiber cross-talk. In order to avoid this, we have decided to compute the resolution matrix per fiber individually, such that the noise is decorrelated across wavelength but not fibers. This is obtained by considering independently the blocks of the covariance matrix $C$ that address each fiber spectrum when computing its corresponding resolution matrix. \subsubsection{Resolution matrix is not flux conserving} It is important to note that $R$ is not a flux-conserving convolution and its normalization vs.~wavelength depends upon the input noise model, i.e.~it is not simply a model of the instrument resolution. Appendix~\ref{sec:spectroperfconvolution} discusses the interpretation of the $R$ tranformation from $f$ to $\tilde f$ as a convolution, which is only valid in the limit of constant noise vs.~wavelength, which in general is not the case. When the input noise varies with wavelength, e.g.~near a sky emission line, then the normalization of $R$ can vary by up to 20\% for DESI spectra. As a consequence, any model fits to data should use a metric such as Equation~\ref{eq:extract-chi2-1D}, where the model is multiplied by the resolution matrix before comparing to data. Directly fitting a line amplitude to data can lead to wavelength dependent systematic biases as shown in Figure~\ref{fig:Rnorm-bias}. The top plot shows a 200\,\AA\ region of a raw extracted sky spectrum. The bottom plot blue line shows the resolution matrix $R$ normalization vs.~wavelength, exhibiting strong ringing where the input noise is varying due to the Poisson noise of the input sky. i.e.~if an input signal $f$ was a true $\delta$-function, the output extracted flux $\tilde f = Rf$ could vary by up to 20\% depending upon wavelength. Broader lines are less affected since they sample more of the $R$ normalization oscillations. e.g.~the Figure~\ref{fig:Rnorm-bias} bottom orange line shows the normalization bias of an [OIII]-like Lorentzian emission line with full-width half-max (FWHM) of 2.7\,\AA, and the green line shows the normalization bias of an [OII]-like double Gaussian with $\sigma=$2\,\AA. These broader lines have less normalization variation than the $\delta$-function, but still can vary by several percent. We emphasize, however, that this normalization ``bias'' disappears if one multiplies an input model by $R$ before comparing to data, as will be shown in \S\ref{sec:extraction-noise-model}. \begin{figure} \includegraphics[width=0.99\columnwidth,angle=0]{rnorm-bias.pdf} \caption{Measurement bias if fitting a model directly to data without using the resolution matrix. The upper plot shows a 200\,\AA\ region of a raw extracted sky spectrum. The bottom plot blue line shows the resolution matrix $R$ normalization vs.~wavelength. Orange and green lines show the normalization convolved with an input [OIII]-like Lorentzian or a [OII]-like double Gaussian.} \label{fig:Rnorm-bias} \end{figure} \subsubsection{Extraction priors and masking} Masking of CCD pixels for cosmic rays, hot columns, or defects can result in wavelength bins with few or no unmasked input pixels, thus requiring regularization to be added to Equation~\ref{eq:extractionchi2} to avoid a singular matrix or extracting ringing due to ill-constrained bins. At the same time, overly aggressive regularization can lead to biases in the extracted spectra, especially at the high flux limit where bright pixels have a relatively low weight due to the Poisson noise. For each wavelength bin, we sum the weights of the input pixels contributing to that bin (masked pixels have zero weight), and set a threshold of $10^{-4}$ of the maximum summed weight of any flux bin. Bins whose summed weight is below that threshold receive a regularization term towards 0-flux with that threshold weight. This approach was empirically tuned to avoid ringing from ill-constrained bins while minimizing bias on bright bins. Cosmic rays that are not completely masked can result in poor PSF fits which bias the extracted flux. To check for this, a CCD image model $Af$ is calculated and used to measure the $\chi^2$ of pixels contributing to each extracted flux bin. Bins with $\chi^2>100$ are flagged with a \textsc{bad2dfit} mask. Additionally, bins with $>50\%$ masked input pixels get a \textsc{somebadpix} mask, while those with completely masked inputs get a \textsc{allbadpix} mask bit. \subsubsection{Subdividing extractions for computational efficiency} Since the solution for $R$ requires an eigen-decomposition of $C^{-1}$, it is not practical to solve all wavelengths ($\sim$2500) for all fibers (500) simultaneously since this would require an ${\cal O}(n^3)$ calculation on a (1.25M $\times$ 1.25M) matrix per CCD per exposure. Instead, the problem is decomposed into multiple overlapping subregions which are extracted independently and then recombined, thus turning the calculation into a large number of small matrices to solve instead of a small number of very large matrices. Each subregion has a core range of wavelengths to extract, which could contribute photons to a contiguous region of CCD pixels. It is not sufficient to simply extract those wavelengths using those pixels, since those pixels also have photons from wavelengths outside of the core wavelengths of interest. The impact on the core wavelengths can be minimized by including additional wavelengths and additional pixels, solving the extraction for the full set, and then keeping the solution for only the core wavelengths of interest for each subregion. The core wavelengths extracted for each subregion are unique, but the CCD pixels used and the buffer wavelengths extracted do overlap. The PSF spot for a given (fiber,wavelength) is modeled over a $11\times17$ rectangular grid of CCD pixels. This size is large enough to include the peak of the PSF from the neighboring fibers, and extend in wavelength enough to represent $>99.95\%$ of the light. The amount of padding in pixels and wavelengths was empirically determined to minimize the extraction bias while maintaining pragmatic computational runtime. Considering the first fiber in the patch, the pixels are extended in $y$ (wavelength direction) by 8 pixels, and then the wavelength grid is extended to include any wavelengths whose centroids are within 2 pixels of that extended CCD patch. Analogous to padding in the wavelength direction, neighboring fibers for each subregion are also extracted and then discarded. This is shown in Figure~\ref{fig:extraction-ccd-patch} for a portion of CCD r1 fibers 25-29 for wavelengths 7300\,\AA\ to 7315.2\,\AA. Blue dots show the centroids of the PSFs on a 0.8\,\AA\ grid and the blue box shows the pixels receiving light from those fibers and wavelengths using a $11\times 17$ spot size. Red $\times$ marks and the red box indicate the extra wavelengths and pixels extracted to minimize the bias on the blue region. In this case, the leftmost fiber (25) is at the edge of a bundle of 25 fibers with a gap before the left-neighboring fiber, thus it isn't necessary to include a padding fiber on the left, but an extra padding fiber is included on the right. In practice, the extractions use a core region of 50 wavelengths (compared to 20 in Figure~\ref{fig:extraction-ccd-patch}). \begin{figure} \centering \includegraphics[width=0.9\columnwidth,angle=0]{extraction_ccd_patch.pdf} \caption{Example subregion of 2D extractions. See text for details.} \label{fig:extraction-ccd-patch} \end{figure} The impact of the biases from extracting overlapping subregions is shown in Figure~\ref{fig:extraction-bias} for fibers 25 and 29 (the left and right blue fibers in Figure~\ref{fig:extraction-ccd-patch}). On the top, the colored solid lines show the extractions from the subregions, while the overlayed dotted line shows the full solution without using subregions. The bottom figures show the difference between the two, normalized by the statistical error on the extractions. Although there is a residual systematic ringing due to the use of subregion extractions, it is less than 5\% of the statistical error. If needed in the future, this could be further reduced by extending the padding. \begin{figure} \centering \includegraphics[width=0.99\columnwidth,angle=0]{extraction_full_vs_patch_bias.pdf} \caption{Extraction bias from using overlapping subregions for fibers 25 and 29 of the r1 CCD. The top plot shows the subregion extracted flux (solid colored) overlayed with the full extraction (dotted black). Bottom show the difference normalized by the statistical error of the extractions. Although there is systematic ringing at the subregion boundaries, it is below 5\% of the statistical error. } \label{fig:extraction-bias} \end{figure} The mapping from wavelength to CCD row varies with fiber number, thus using a rectangular region of pixels to extract results in an asymmetric amount of padding from one fiber to another. This can be seen in Figure~\ref{fig:extraction-ccd-patch}, comparing the leftmost fiber with symmetric top/bottom pixel padding, vs. the rightmost fiber having more padding at higher (upper) wavelengths than lower. This results in slightly more bias at the subregion boundaries for the rightmost fiber, e.g.~Figure~\ref{fig:extraction-bias}, right. In practice this effect is still quite small compared to the statistical errors and we have not optimized the padding to be the same for every fiber, and have kept the use of rectangular CCD regions to simplify the code. \subsubsection{Wavelength grid} One of the key conveniences of the 2D extraction algorithm is that it allows one to choose a common grid of extraction wavelengths for every fiber, regardless of how those align with the CCD pixel grid. By construction the output wavelength grid is the same for every fiber and the flux bins are uncorrelated, unlike row-by-row extraction methods that either have a different wavelength grid for every fiber (tied to the CCD rows), or introduce correlations by resampling to a common wavelength grid. DESI extracts flux using a linear wavelength grid from 3600\,\AA--9824\,\AA\ in 0.8\,\AA\ wavelength steps. This choice of extraction resolution was chosen to be slightly larger than the native CCD pixel scale ($\sim$0.6\,\AA/row) to avoid numerical artifacts. $b$~cameras extract 3600--5800\,\AA, $r$-cameras use 5760--7620\,\AA, and $z$-cameras use 7520--9824\,\AA. Note that the extraction wavelengths for the individual cameras overlap, but are phased to be on a common 0.8\,\AA\ grid across all 3 cameras. \begin{figure} \centering \includegraphics[width=0.99\columnwidth,angle=0]{extraction_pull_top.pdf} \includegraphics[width=0.8\columnwidth,angle=0]{extraction_pull_bottom.pdf} \caption{2D flux extractions compared to resolution convolved input for 10000 noise realizations of simulated CCD images.} \label{fig:extraction-pull} \end{figure} \subsubsection{Validity of extraction noise model} \label{sec:extraction-noise-model} To study extraction bias and the validity of the variance model, we simulated 10000 CCD images of a spectrum with a smooth sloping continuum ranging from readnoise limited to photon shot noise limited, plus 3 emission lines of varying strengths. Each realization has the same truth spectrum but a different noise realization of the CCD pixels (Gaussian readnoise of 3 electrons/pixel plus Poisson signal variations), and then is processed with the 2D extraction code. The results are shown in Figure~\ref{fig:extraction-pull}. The extracted flux compared with the resolution-convolved input spectrum is very consistent with the reported noise model, with a pull=(flux-model)/$\sigma$ very consistent with a mean of 0 and standard deviation of 1 at all wavelengths. \begin{figure} \centering \includegraphics[width=0.99\columnwidth,angle=0]{extraction_covcorr_top.pdf} \includegraphics[width=0.99\columnwidth,angle=0]{extraction_covcorr_bottom.pdf} \caption{Variance and correlation of 2D extractions of 10000 noise realizations of CCD images of a spectrum. The upper plot shows that the wavelengths bins are uncorrelated, while the lower 3 plots show the excellent agreement with the reported variance. The second from the bottom plot shows the statistical uncertainty of the variance of 10000 simulations in thin gray lines. } \label{fig:extraction-covcorr} \end{figure} Figure~\ref{fig:extraction-covcorr} shows the variance and correlation of the wavelength bins, confirming that they are uncorrelated between wavelength bins and in agreement with the reported variance to within the statistical precision of the 10000 realizations. \subsubsection{Key features of the extraction} We conclude this section by emphasizing some of the features of the spectra extracted using a full 2D PSF model. On the positive side, spectra of every fiber use a common wavelength grid and are by construction uncorrelated across wavelengths. Use of the full 2D PSF model maximizes the information extracted from the raw data and has excellent validity of the resulting uncorrelated noise model. The resolution matrix $R$ models the per-fiber per-wavelength effective resolution of each spectrum with more fidelity than a simple Gaussian vs.~wavelength line-spread-function (LSF) model. On the negative side, the non-unity normalization of $R$ results in more complex model fitting since one can no longer directly fit a model $m$ to the spectrum $f$, but rather one must fit $Rm$ to $f$. However, the uncorrelated bins of $f$ on a common wavelength grid are a counterbalancing convenience for analyses. \subsection{Cross-talk correction} \label{sec:fibercrosstalk} The spectrograph PSF tails extend well beyond the separation between adjacent fibers and cause a measurable contamination among neighboring spectra. As explained in \S\ref{sec:psf}, these tails were not included in the PSF model used for the extraction. The result of this omission is the presence some residual fiber cross-talk among the extracted spectra. The PSF tails can asymptotically be described as a power law. We model them as a convolution with a kernel of the form $K(r)\propto\ r^2 \, (1+r^2)^{-(1+p/2)}$, where $r$ is the distance from the PSF center in units of pixels and $p \simeq 2.5$ the power law index. We estimate the contamination of neighboring spectra using the kernel $K(r)$ and the pixel coordinates of the spectral traces in the CCD. Corrections for nearest and second nearest fibers are computed. The normalization of the kernel has been fit as a function of the wavelength, fiber number, and fiber separation for each CCD camera independently. For the blue and red camera, the cross-talk is in the range 0.1-0.2\% for adjacent fibers. It varies slowly with wavelength and we have resorted to use the same value for each fiber of a given camera. In the NIR camera however, the cross-talk is rapidly increasing with wavelength for wavelength above 8900\,\AA. It reaches values as large as 1.5\% at the maximum wavelength of 9800\,\AA. It is also a function of the fiber location in the CCD. The cross-talk values for all cameras are shown in Figure~\ref{fig:fibercrosstalk}. The resulting contamination among fibers from the same camera is estimated from the extracted spectra and subtracted (we do not iterate the procedure because the correction is small). \begin{figure} \centering \includegraphics[width=0.9\columnwidth,angle=0]{fibercrosstalk_figure.pdf} \caption{Measured cross-talk between fibers $N$ and $N+1$ (top) or $N+2$ (bottom) for all 30 CCD cameras, as a function of wavelength, for a central fiber ($N \sim 250$).} \label{fig:fibercrosstalk} \end{figure} \subsection{Fiber flat fielding} \label{sec:fiberflat} The fiber flat fielding consists of determining a correction for the variations of throughput from fiber to fiber, as a function of wavelength. This correction is essential to homogenize the response of the fibers before the sky background subtraction (see~\S\ref{sec:skysubtraction}). It is also needed to propagate the spectro-photometric calibration obtained from the standard star fibers to the other fibers of the focal plane (see~\S\ref{sec:fluxcalibration}). This throughput variation is principally due the relative variation of fiber transmission, but it is also affected by the variation of transmission in the spectrograph optics (in particular the change of transmission of the dichroics with the incidence angle which is correlated with the position of the fibers on the slit), the change of solid angle seen by a fiber (due to the change of plate scale in the field of view), the vignetting (mostly the shadow of the cylindrical focal plane instrument), and the residual variations of CCD pixel efficiency that have not been perfectly corrected by the pixel flat field (see \S\ref{sec:preprocessing}). Because the primary goal is to homogenize the response of fibers for the sky background subtraction, one could think about using directly night sky observations for this purpose. We do not use this approach in practice because the sky spectrum is composed of bright emission lines that would leave an imprint on the estimated transmission, and spectral regions with a faint continuum that would require a prohibitive exposure time to acquire enough signal and reduce statistical fluctuations. The twilight sky is better in that respect but still not ideal; a dedicated calibration system with stable and smoother spectral features is preferred. We use for this purpose an array of LEDs lamps shining on a white screen in the dome as described in \S\ref{sec:calib-obs}. The drawback of this approach is that the phase space of the illumination is not exactly as the night sky. The flat-field correction is computed in several steps. We first compute a correction independently for each camera and exposure (\S\ref{sec:fiberflat1}). We then combine several exposures obtained with different lamps to get a more homogeneous illumination, closer to the night sky, and we combined the data from the cameras of all spectrographs to inter-calibrate them (\S\ref{sec:fiberflat2}). We evaluate the geometrical differences between the dome screen illumination and the night sky using twilight sky data in \S\ref{sec:twilight}, and finally we present the method used to correct for variation of the flat field with humidity in~\S\ref{sec:fiberflat_vs_humidity}. \subsubsection{Fiber flat fielding algorithm} \label{sec:fiberflat1} The algorithm consists in fitting simultaneously an average spectrum and the relative transmission of each fiber from a camera, for each wavelength of the spectral extraction grid. We take into account the different resolution of each fiber and take advantage of the fact that spectral bins are not correlated. We minimize the following quantity \begin{equation} \chi^2 = \sum_{f} \sum_{i} w_{f,i} \left( D_{f,i} - T_{f,i} \sum_j R_{f,i,j} S_j \right)^2 \label{eq:fiberflat} \end{equation} where $f$ is the fiber number, $i$ the wavelength index, $w_{f,i}$ the inverse variance, $D_{f,i}$ the data, $T_{f,i}$ the transmission we want to measure, $R_{f,i,j}$ an element of the resolution matrix defined in~\S\ref{sec:extraction} for the fiber $f$ and $S_j$ a flux value of the deconvolved spectrum (from an hypothetical average fiber). The only constraint is that the average transmission is one for all wavelength, $\left< T_{f,i} \right>_f=1$. We note we should in principle have considered the transmission before applying the resolution to the mean spectrum (with $T$ to the right of $R$ in Eq.~\ref{eq:fiberflat}), but in practice this is not important because we do not expect sharp variations of the relative fiber transmission over the typical scale of the spectral resolution\footnote{Figure 26 of \cite{desi-collaboration22a} shows the DESI throughput as a function of wavelength, which includes the fiber transmission. The only sharp feature in the blue channel at 4400\,\AA\ is caused by the collimator mirror reflectivity. The other absorption features at larger wavelength are due to the atmospheric transmission.}. The challenge in this fit are the cosmic ray hits. We first fit iteratively for a smooth mean spectrum and then a smooth fiber transmission, ignoring variations of resolution, while rejecting outlier spectral pixels. When all the outliers are removed, we fit for the exact solution. For both fits we use the Gauss-Newton algorithm. The fit is iterative because it is non-linear (we have to determine the product of $T$ and $S$). Examples of flat field corrections are shown in Figure~\ref{fig:fiberflat} for one 120\,sec exposure obtained with the LED array of one calibration lamp. \begin{figure} \centering \includegraphics[width=0.99\columnwidth,angle=0]{fiberflat-b8-00055314.pdf} \includegraphics[width=0.99\columnwidth,angle=0]{fiberflat-r8-00055314.pdf} \includegraphics[width=0.99\columnwidth,angle=0]{fiberflat-z8-00055314.pdf} \caption{Relative transmission of a sub-sample of 10 fibers from the spectrograph SM2 obtained with one exposure of 2\,min on the dome screen with LED lamps in the blue, red and NIR cameras. Each colored curve corresponds to a fiber. The measurement appear noisy but 12 such exposures are combined to determine the final correction. The spectral features in the blue camera at 3800 and 4400\,\AA\ are due to a dip in the reflectivity of the spectrograph collimator mirror at these wavelength, with a profile that varies with the location in the mirror and with the light incidence angle. The variations around 5750\,\AA\ and 7500\,\AA\ in the red and NIR camera spectra are due to the change of transmission and reflection of the dichroics again with the light incidence angle that varies with the fiber location in the pseudo-slit. Spectral regions affected by cosmic ray hits have been masked and are interpolated over in the figures.} \label{fig:fiberflat} \end{figure} \subsubsection{Combining the fiber flat fields from various lamps and cameras} \label{sec:fiberflat2} The fiber flat fields obtained with each of the four lamps placed on the upper ring of the telescope (see Figure~\ref{fig:mayall}) are combined to get a more homogeneous pattern. A numerical computation has shown that the illumination is azimuthally homogeneous to better than 0.1\% with four lamps, with an expected residual radial gradient of about 1\% when compared to the night sky illumination (see Figure~\ref{fig:lamp-illumination}). This residual difference is due to differences in the vignetting by the focal plane instrument. We combine the fiber flats from exposures obtained with each lamp one at a time instead of all the lamps together in order to mitigate the possible variations of lamp intensity. Indeed a variation of intensity of one lamp has no effect on the flat field correction obtained from an exposure with this individual lamp as it is absorbed in the mean spectrum term in Eq.~\ref{eq:fiberflat} while it would lead to a gradient in the flat field correction if all the lamps were used together. We also combine fiber flats from all cameras of the same type (blue, red, or NIR) while normalizing them to the same mean spectrum. This results in an inter-calibration of the spectrographs. We show these fiber flats in Figure~\ref{fig:fiberflatnight}. \begin{figure} \centering \includegraphics[width=0.49\columnwidth,angle=0]{lamp_0.pdf} \includegraphics[width=0.49\columnwidth,angle=0]{lamps0123.pdf} \caption{Left: illumination pattern on the focal plane for one calibration lamp. Right: illumination pattern obtained with four calibration lamps divided by the illumination from the sky. The non-homogeneity of the illumination ratio is purely radial, with a variation of about 1\% from the center to the edge of the focal plane.} \label{fig:lamp-illumination} \end{figure} \begin{figure} \centering \includegraphics[width=0.99\columnwidth,angle=0]{fiberflatnight-b8-20200313.pdf} \includegraphics[width=0.99\columnwidth,angle=0]{fiberflatnight-r8-20200313.pdf} \includegraphics[width=0.99\columnwidth,angle=0]{fiberflatnight-z8-20200313.pdf} \caption{Transmission of a sub-sample of fibers from spectrograph SM2 normalized to the average transmission of all the fibers from the 10 spectrographs. These curves are the result of the combination of 12 individual fiber flats as shown in Figure~\ref{fig:fiberflat}, but with a different normalization (average of 10 spectrographs instead of one).} \label{fig:fiberflatnight} \end{figure} \subsubsection{Validation with sky background measurements} \label{sec:twilight} The expected radial pattern in the ratio of dome flat to sky flat (Fig.~\ref{fig:lamp-illumination}) has been verified with sky background observations. We expect this residual anisotropy to be stable and achromatic as it is the combination of purely geometric terms (location of dome screen, lamps, shadows), and the reflectivity of the screen which is quasi-Lambertian over the whole wavelength range. Figure~\ref{fig:twilightflat} presents the focal plane views of the median flux in fibers for a twilight sky observation conducted on March 15, 2020, after flat fielding. We measure a gradient of a few percent along one axis and a quadratic term only along the other. More quantitatively, noting $x$ and $y$ the cartesian focal plane coordinates of fibers in units of 400\,mm, we have \begin{eqnarray} \mathrm{blue\ flat} & \propto & 0.991 -0.003 \, x +0.012 \, y -0.002 \, x^2 +0.041 \, y^2 \nonumber \\ \mathrm{red\ flat} & \propto & 0.988 +0.000 \, x +0.019 \, y +0.003 \, x^2 +0.041 \, y^2 \nonumber \\ \mathrm{NIR\ flat} & \propto & 0.987 -0.001 \, x +0.022 \, y +0.010 \, x^2 +0.039 \, y^2 \nonumber \end{eqnarray} As visually evident on the figure and the polynomial coefficients, the twilight flat is consistent in all three cameras, blue, red, and NIR, confirming it consists in a purely geometric correction term. The gradient is due to the variation of the twilight sky brightness with elevation (the parallactic angle was close to 90\,$\deg$). \begin{figure} \centering \includegraphics[width=0.99\columnwidth,angle=0]{focal_plane_twilight_brz.png} \caption{Focal plane view of the median flux in fibers from a twilight observation (exposure \#00055559 from 2020/03/15) after applying the nightly averaged fiber flat correction (see \S\ref{sec:fiberflat2}) and a normalization factor per camera (blue, red and NIR cameras from left to right). The color scales have been saturated to enhance the residual gradient. This figure highlights the agreement between cameras, demonstrating this twilight flat is very nearly achromatic.} \label{fig:twilightflat} \end{figure} This study is complemented with a measurement of the dark sky during science observations. The flat-fielded flux in the r-camera of all the sky fibers of 4 nights has been averaged as a function of the fiber coordinates in the focal plane. The normalized and average flux is shown as a function of the focal plane distance on Figure~\ref{fig:skyflat}. We measure a radial pattern with a variation of 2\% which is a bit larger than the expectations from simulations (see Figure~\ref{fig:lamp-illumination}, where the expected inverse ratio is shown). Note that this residual anisotropy is absorbed by a normalization coefficient estimated for each exposure and fiber (see~\S\ref{sec:skynormalization}). \begin{figure} \centering \includegraphics[width=0.9\columnwidth,angle=0]{flat-sky-profile.pdf} \caption{Flat-fielded night sky background variation in the r-camera as a function of the distance from the focal plane center. This figure has been obtained by averaging the sky spectra of all sky fibers from 4 nights of observations. The rms of 1 to 1.5\% comprises the measurement statistical errors, the residual contamination for stars, and systematic errors in the spectroscopic measurements. The radial variation is due to a difference in the vignetting between the sky and the calibrations with the dome screen.} \label{fig:skyflat} \end{figure} \subsubsection{Humidity correction} \label{sec:fiberflat_vs_humidity} There are absorption features at 3800 and 4400\,\AA\ in the collimator mirror reflectivity of most spectrographs. The wavelength of those features are a strong function of humidity. Figure~\ref{fig:fiberflat-vs-humidity} shows this variation for one fiber. It was obtained by averaging the flat field measurements obtained in 2021, in bins of humidity, from the minimum of about 8\% obtained in winter to the maximum of 50\% that is reached in summer during the moonsoon season (this maximum value is enforced with a dehumidifier controlled by the environmental system of the spectrographs enclosure). We use this variation as a model to correct for the change of humidity between the calibration runs in the afternoon and the observations during the night. \begin{figure} \centering \includegraphics[width=0.99\columnwidth,angle=0]{fiberflat-vs-humidity.pdf} \caption{Flat field variation of the central fiber of the SM10 blue camera as a function of the humidity in the spectrographs enclosure. Left: wavelength shift of the feature; right: fiber flat as a function of wavelength (normalized to the average of the 500 fibers for this camera).} \label{fig:fiberflat-vs-humidity} \end{figure} \subsection{Flux calibration} \label{sec:fluxcalibration} In this section is described the flux calibration of each exposure. It consists in converting counts in electrons per unit wavelength into spectral energy distributions. The default calibration that is applied to all spectra provides an estimate of the total flux for point sources. It is however only a lower limit on the total flux for extended sources. In order to obtain a more useful calibration for those, we also provide a coefficient (\texttt{PSF\_TO\_FIBER\_SPECFLUX}, see \S\ref{sec:fiber-aperture-loss-correction}) to convert the measurements to ``fiber fluxes'', corresponding to the flux one would record in a fiber of angular diameter 1.5$\arcsec$ observed for a seeing of $1\arcsec$ FWHM\footnote{Same fiber flux definition as in the Legacy Surveys Data Release 9 catalogs, see \url{https://www.legacysurvey.org/dr9/catalogs/}}. Fiber fluxes are designed to be valid both for point sources and extended sources. This flux calibration is performed as follows. \subsubsection{Selecting the stars} We first measure for each standard star from the 10 petals the ratio between the measured fiber flat-fielded flux in r-band and the expectation from the stellar model obtained in the previous section. The numerator is the average inverse variance weighted flux in the wavelength range [6000-7300]\,\AA\ and the denominator is the average model flux in the same range. The distribution of those ratios is used to exclude outliers at three standard deviation. Note that the fiber flat field includes a correction from spectrograph to spectrograph so that we can compare the fluxes from different spectrographs. Stars for which the difference between the model and observed $g-r$ color from the imaging catalogs exceed $0.1 + 0.2 \times E(B-V)$ are also excluded. The second term relaxes the color criterion in regions of large Galactic extinction because of uncertainties in the dust reddening law. \subsubsection{Average throughput per camera} Once the stars are selected, the calibration is performed for each spectrograph camera individually. The camera throughput is computed as the average ratio between the flat-fielded but otherwise uncalibrated measured spectra and the model of standard stars. Each stellar model is first resampled to the wavelength grid of the data. We then iteratively i) fit the mean throughput while accounting for the individual fiber flat field and resolution, ii) compute a scale factor for each star to correct for possible offsets in the fiber positioning, iii) reject outlier fluxes as a protection against residual spikes from unmasked cosmic ray hits. For step ii) we fit for a ``deconvolved'' mean throughput that we multiply by the model flux before convolving the product using the resolution matrix of each standard star fiber. At the end of this iterative procedure we use the resolution matrix of each of the 500 fibers of the camera to obtain a first estimate for the throughput of each fiber. \subsubsection{Corrections for fiber aperture losses} \label{sec:fiber-aperture-loss-correction} As described in \S\ref{sec:expo-data}, a table of fiber positioning offsets is part of the data product of each exposure. We use this information to improve the calibration while also considering the mean seeing (sometimes called image quality) of the exposure, the sources surface density profiles and their measured fiber fluxes from the imaging catalogs. We first estimate a ``flat to PSF'' correction coefficient valid for point sources only and apply it to the reported spectro-photometrically calibrated fluxes. We also determine another ``PSF to fiber'' coefficient to obtain the fiber fluxes that can be used for extended sources. The latter is not applied to the spectral fluxes but saved in the \texttt{FIBERMAP} table in the column \texttt{PSF\_TO\_FIBER\_SPECFLUX} (see \S\ref{sec:data-products}). The focal plane PSF has been precisely characterized using the GFA images. It follows to a good approximation a Moffat profile with a parameter $\beta=3.5$~\citep{Moffat1969,Meisner2020}. We use this property, the effective fiber angular radius $r_F$ (which depends on the fiber location in the focal plane as the plate scale varies), and the seeing (hereafter $s$, measured as the FWHM) to compute the fiber acceptance fraction for a point source that we call $F_{PSF}(s,r_F)$ in the following. It is the computed ratio of the flux entering a fiber to the total flux from a source on the focal surface. The ``flat to PSF'' correction is $$ c^{\ PSF}_{flat}(s,r) = \frac{\left< F_{PSF}(s) \right>}{F_{PSF}(s,r_F)} \frac{r_F^2}{\left< r_F^2 \right>} $$ where $\left< X \right>$ represents the average of $X$ over the focal plane. The second factor corrects for the variation of the fiber flat field caused by the change of solid angles of the fibers (i.e. the Jacobian of the transformation from angles to focal plane coordinates). The ``PSF to fiber'' correction depends on the fiber acceptance fraction for an extended source, which surface density is modeled with an exponential profile of half light radius $r_{1/2}$. We label it $F_{exp}(s,r_F,r_{1/2})$. Keeping in mind the definition of fiber flux given at the beginning of this section, this correction is $$ c^{\ fiber}_{PSF}(s,r_F,r_{1/2}) = \frac{F_{exp}(1\arcsec,1.5\arcsec/2,r_{1/2}) F_{PSF}(s,r_F)}{F_{exp}(s,r_F,r_{1/2})}$$ Figure~\ref{fig:imaging-spectro-fiberflux-r-ratio} shows the ratio between the spectroscopic flux in the DECam r-band and the same quantity from the imaging catalog. Both the total flux (\texttt{FLUX\_R}) and the fiber flux (\texttt{FIBERFLUX\_R}) are considered. For the spectroscopy, this consists in multiplying the integrated flux by the coefficient \texttt{PSF\_TO\_FIBER\_SPECFLUX}. The figure demonstrates that the fiber fluxes are consistent between spectroscopy and imaging for extended sources. On the bright end, the scatter in the ratio is of 6\%. \begin{figure} \centering \includegraphics[width=0.99\columnwidth,angle=0]{flux-ratio.pdf} \caption{Ratio of spectroscopic to imaging r-band flux for the targets from several main survey dark-time tiles. The top panels compare the total flux which is only valid for point sources for DESI. The lower panels compare the fiber fluxes which are consistent between spectroscopy and imaging for both point sources (in orange) and extended sources (in blue). The x axis is the imaging fiber magnitude on the left panels and the fraction of light in the fiber aperture on the right panels. } \label{fig:imaging-spectro-fiberflux-r-ratio} \end{figure} Figure~\ref{fig:dflux-vs-xy} shows color maps of the average ratio of spectroscopic to imaging fluxes for stars, in blue, red and NIR bands as a function of focal plane coordinates. While the distribution is mostly flat for the red some structures are found in the other bands. Those are due to chromatic distortions in the corrector. A variation of 10\% is typically caused by a offset of about 0.3$\arcsec$ or 20$\mu$m. This is irreducible in the sense that one has to choose which wavelength to optimize for the fiber positioning. We have chosen for DESI to optimize the throughput in the red. Note that this systematic calibration error is not corrected in the Early Data Release (EDR). \begin{figure} \centering \includegraphics[height=0.3\columnwidth,angle=0]{flux-ratio-vs-binned-xy-4500A-5500A.pdf} \includegraphics[height=0.3\columnwidth,angle=0]{flux-ratio-vs-binned-xy-6000A-7300A.pdf} \includegraphics[height=0.305\columnwidth,angle=0]{flux-ratio-vs-binned-xy-8500A-9800A.pdf} \caption{For left to right, average ratio of the spectroscopic to imaging flux of standard stars in the blue (4500-5500\,\AA), red (6000-7300\,\AA), and NIR (8500-9800\,\AA) as a function of focal plane coordinates. } \label{fig:dflux-vs-xy} \end{figure} \subsubsection{Flux calibration testing with white dwarfs} Given the relatively simple spectra of hydrogen-dominated atmosphere white dwarfs and their largely blue-dominated spectra both in terms of spectral features and flux (Fig.~\ref{fig:wd_fits}), they serve as good tests for flux calibration and as such were observed by DESI. White dwarfs were selected in DESI based on criteria described in~\cite{cooper22a}, which was itself adapted from the selection made in \cite{gentilefusilloetal19-1}. To test the flux calibration in DESI, we first cross-matched the observed DESI targets with the $Gaia$ DR2 catalog collated by \cite{gentilefusilloetal19-1}, which provides additional parameters beyond those in the $Gaia$ DR2 photometric catalog, including the probability, $P_{\textrm{WD}}$, that a white dwarf candidate is indeed a white dwarf. A white dwarf candidate is considered ``high-confidence’’ if $P_{\textrm{WD}} > 0.75$. This initial cross-match resulted in 4064 unique white dwarf candidates that were observed as part of the DESI EDR, with a total of 31734 exposures showing the many repeat spectra obtained. To use these white dwarf candidate spectra for comparing the DESI flux calibration, we limited our sample to the highest confidence white dwarf candidates selecting $P_{\textrm{WD}} > 0.95$, spectra with a signal-to-noise ratio in the $b$-arm $> 5$, and white dwarf candidates identified as hydrogen-dominated (i.e. DA white dwarfs) based on a random forest classifier. This resulted in 9854 spectra which were all fitted using the \textsc{WD} pipeline described in~\cite{cooper22a}, and examples of these fits to individual spectra are shown in Fig.\,\ref{fig:wd_fits}. These fits are applied to and scaled by a constant value to the $b$-arm DESI spectra, and do not use any information in the $r$-, or $z$-arm spectra. From these fits, we calculated a single weighted-average of the residuals from these 9854 fits which is shown for the DESI spectral range in Fig.\,\ref{fig:wd_calib_compare}. The \textsc{WD} fits in the $b$-arm show some structure but are largely constrained within $\pm 2$\% and to wavelengths below $\simeq$\,4500\,\AA. The largest discrepancy is seen below $\simeq$\,3700\,\AA\ where a rapid change in the residual results in a $\simeq$\,6\% increase that then remains constant to the blue-edge of the spectrum. The $r$- and $z$-arm comparisons show a slight offset of roughly $-1$\%, which is not unexpected when the spectra are fitted to the $b$-arm and normalised to a small range within that arm. There is evidence of small-scale features in the residuals that deviate from this offset, which appear to be largely constrained to H$\alpha$ in the Balmer series, and the Paschen series. Further work is needed to identify the origin of these features, as it is likely that some of these are generated from the \textsc{WD} fitting pipeline rather than the flux calibration itself. Additional improvements can also be made to the comparison by including E(B-V) values into these fits, which are currently not included. \begin{figure} \centering \includegraphics[width=\linewidth]{WD_model_fit_examples.pdf} \caption{White dwarf spectra observed by DESI showing the individual $b$- (blue), $r$- (orange), and $z$- (red) arm spectra and best fitted models from the WD pipeline~\citep{cooper22a} shown in gray, which have been scaled by a constant factor by normalizing to the region shaded in gray.} \label{fig:wd_fits} \end{figure} \begin{figure} \centering \includegraphics[width=\linewidth]{Fuji_calib_compare_paper.pdf} \caption{The average residuals (black) to white dwarf spectra observed by the three spectroscopic arms of DESI and fitted by the WD pipeline~\citep{cooper22a}. The gray shaded region indicates the wavelength range over which each model was scaled to the white dwarf $b$-arm spectrum. Orange tabs denote the wavelengths of the Balmer and Paschen lines, in particular highlighting the Balmer jump around where the largest discrepancy in the residuals are observed.} \label{fig:wd_calib_compare} \end{figure} \section{Future Development Work} \label{sec:future} This paper has described the DESI spectroscopic pipeline version used for the upcoming Early Data Release covering DESI Survey Validation data; it has additionally been validated on the first year of DESI survey data, to be released in Data Release 1 (DR1). The algorithms are mature and performing well, thus we do not anticipate major algorithmic changes in the future. At the same time, we highlight in this section a few areas that may have future development. If and when significant algorithmic changes are made, those changes will be documented in the corresponding data release papers for the first release including those updates. Redrock spectral classification and redshift estimation~(\S\ref{sec:redshift}) currently uses PCA templates to fit the data. Although this provides speed and flexibility, there is no constraint that the ``best fit'' answer is physically meaningful other than penalizing negative [OII] flux. This makes Redrock overly sensitive to false fits, especially when there are data problems upstream, e.g.~errors in the background subtraction or flux calibration between the spectrograph arms leading to steps in the spectra. Future work may introduce archetypes tuned for DESI spectra, non-negative matrix factorization (NMF) templates, or other priors focused on ensuring physically meaningful fits. The current Redrock galaxy and stellar templates were generated for DESI, but the QSO templates are still the same as those used by BOSS in idlspec2d\cite{Bolton2012}, trained on just a few hundred QSO spectra. Although there are many more QSO spectra available now that could be used to make improved QSO templates, initial work has found that improving the QSO performance comes at the cost of degrading the completeness of galaxy targets. Future releases may have new QSO templates tuned to improve QSO performance while preserving the performance of galaxy templates. The sky subtraction algorithm~(\S\ref{sec:skysubtraction}) is designed for spatially uniform sky, where fiber-to-fiber variations are primarily due to instrumental effects rather than the input sky spectrum actually being different for different fibers. This is sufficient for dark-time observations and most-bright time observations, though spatial sky gradients during the brightest observations lead to data quality problems. Explicitely detecting and fitting for non-uniform sky may be included in the future. The spectral resolution matrix~(\S\ref{sec:extraction}) is not a flux conserving convolution, which may lead to analysis challenges that we have not yet fully understood. As the DESI collaboration gains experience using these spectra for science analyses, we may learn better ways to model and use the spectral resolution, potentially leading to algorithmic and format updates. \section{Instrument} \label{sec:instrument} The DESI instrument is presented in detail in \cite{desi-collaboration22a} and references therein; we provide here only a brief overview. DESI is a multi object spectroscopic system installed at the Mayall 4-m telescope at Kitt Peak in Arizona. It features a prime focus instrument with a new corrector, a focal plane composed of 5000 fiber robots, and ten 3-arm spectrographs in the Coud\'e room connected to the focal plane with 50-m long fiber cables. The high throughput of the instrument, the fiber positioning accuracy, and the stability of the spectrograph optics allows us to achieve an excellent redshift efficiency. \subsection{Corrector} The DESI corrector is composed of six lenses of about 1-m diameter, converting the telescope focal ratio from f/2.8 to f/3.9 over a 3.2$^\circ$ diameter field of view, while achieving an image quality of 0.6$\arcsec$ FWHM for the best nights~\citep{miller22a}. The corrector incorporates an atmospheric dispersion corrector (ADC), which compensates for the wavelength-dependent atmospheric dispersion in the field of view up to 60\,$\deg$ from Zenith. The lenses are hold by a barrel that is connected to the rest of the telescope structure by an hexapod which provides 6 degrees of freedom to adjust the position and orientation of the corrector and the focal plane system. \subsection{Focal plane} The DESI focal plane is segmented in 10 petals, each of them composed of 500 fiber positioners, up to 12 fiducial sources of light, and one Guide/Focus/Alignment (GFA) detector system~\citep{silber22a}. Six of the ten GFAs are used for guiding and four are used for the auto-focus and optical alignment. The fibers can be back-illuminated from the spectrographs. A fiber view camera (FVC) placed in an opening at the center of the primary mirror is used as a feed-back system to obtain an accurate positioning of the fibers relative to the fiducials. Each fiber positioner is composed of two rotating 3-mm arms, the second being placed at the end of the first, providing a 12-mm diameter patrol area for each fiber. As the separation between adjacent positioners is of 10.4\,mm, most of the petal area is within the reach of a positioner. The positioning loop is composed of a first ``blind'' move, where a typical positioning precision of about 50\,$\mu$m is obtained, followed by a ``correction'' move, where a positioning accuracy better than 10\,$\mu$m is achieved after the analysis of a FVC image. This is much smaller that the 107\,$\mu$m diameter of a fiber, such that less than 3\% of the light is lost because of positioning errors (see more details in \S\ref{sec:fluxcalibration}). The complete positioning loop takes less than a minute. A last FVC image is taken to record the final position of the fibers. \subsection{Spectrographs} 50-m long fiber cables connect each petal in the focal plane to one spectrograph in the Coud\'e room. The fiber system and the spectrographs are described in detail in~\cite{poppett22a} and \cite{jelinsky22a}. On the spectrograph end of the cable, the 500 fibers are aligned to form a pseudoslit in a mechanical assembly called a slithead. The latter is inserted in the spectrograph with the pseudoslit precisely positioned in a slot in a NIR dichroic (see Figure~\ref{fig:spectrograph}). The light emitted from the fibers is reflected on a collimator mirror back to the NIR dichroic which is transparent to NIR wavelengths but reflects optical wavelengths at an angle. In front on the NIR dichroic is the main shutter which also contains a fiber back illumination system. A second shutter is located behind the dichroic to protect the NIR camera from this back illumination. The spectrograph is also equipped with a pair of Hartman doors placed in front of the collimator. Optical wavelengths are further split with a second dichroic, resulting in three arms for blue (3600-5930\,\AA), red (5600-7720\,\AA), and NIR (7470-9800\,\AA) wavelengths. Each arm utilizes a volume phase holographic (VPH) grating and a five lens camera, the last lens being the entrance window of a cryostat. Each cryostat hosts a CCD cooled down to 163 K for the blue channel, and 140 K for the red and NIR channels. \begin{figure} \centering \includegraphics[width=0.98\columnwidth,angle=0]{spectro-schematic.pdf} \caption{Schematic view of one of the 10 DESI spectrographs. See text for details.} \label{fig:spectrograph} \end{figure} \subsection{CCDs} \label{sec:ccdimage} The blue CCDs are $4096 \times 4096$ STA4150 CCDs from Semiconductor Technology Associates (STA), that were processed and packaged by the University of Arizona Imaging Technology Laboratory (ITL). The red and NIR CCDs are 250\,$\mu$m thick, fully-depleted p-channel CCDs of $4114 \times 4128$ pixels. Both types of CCDs are read with four amplifiers. Figure~\ref{fig:ccd-layout} shows the CCD layout as it appears in the files on disk, after some transformation performed by the instrument control system. The figure also shows the direction of the parallel and serial clocks, along with the location of over-scan and pre-scan regions. The measured CCD read noise (see~\S\ref{sec:preprocessing}) is in the range of 2.8 to 4.2 electrons r.m.s for the blue CCD amplifiers, with one outlier close to 5 electrons, and 2.2 to 3.5 electrons for red and NIR CCD amplifiers again with one outlier at 4.5 electrons. \begin{figure} \centering \includegraphics[width=0.9\columnwidth,angle=0]{ccdlayout.png} \caption{DESI CCD image layout with four amplifiers. The CCD images for all cameras have the same orientation after transformation of the data array by the instrument control system; the column index (AXIS1 in the {\it fits} files headers, often labeled $X$ in this paper) increases with the fiber number, and the row index (AXIS2 in {\it fits}, often label $Y$ in this paper) increases with increasing wavelength. The black arrows indicate the direction of readout, with the parallel clock in the wavelength direction ($Y$) and the serial clock along the fiber number ($X$). Also shown are the prescan and overscan regions used to measure and remove the bias level during pre-processing (see \S\ref{sec:preprocessing}).} \label{fig:ccd-layout} \end{figure} \subsection{Calibration system} \label{sec:calibration-system} The calibration system for DESI consists in a dome screen and a set of lamps installed on the upper ring of the telescope structure that can illuminate the screen when the telescope is pointing at it (see Figure~\ref{fig:mayall}). The screen is larger than the telescope mirror to account for the range of incidence angles and is covered with a coating of nearly Lambertian and achromatic reflectance. The lamps are placed in four boxes evenly placed along the upper ring. LED arrays are used for flat-fielding (see~\S\ref{sec:fiberflat}) and a combination of Mercury, Argon, Cadmium, Neon, Xenon, and Krypton lamps are used for the wavelength calibration and the point spread function modeling (see~\S\ref{sec:psf}). \begin{figure} \centering \includegraphics[width=0.8\columnwidth,angle=0]{DESI-calib-system-dome.png} \caption{Illustration of the DESI calibration system with the location of the screen and calibration lamps. The diameter of the dome screen is 5\,m.} \label{fig:mayall} \end{figure} \begin{figure} \centering \includegraphics[width=0.99\columnwidth,angle=0]{preproc-z1-00055643-annotated.png} \includegraphics[width=0.99\columnwidth,angle=0]{preproc-z1-00055643-zoom.png} \caption{Example CCD image after preprocessing. This is a NIR CCD image of spectrograph SM10 after a 900s exposure. On the top panel, one can see the 500 fibers organized in 20 blocks of 25 fibers each. The mostly horizontal curved lines are sky lines. Bright fibers (appearing as vertical dark bands in this negative color scale) are the spectral traces from standard star fibers. One can also note many cosmic ray hits. The lower panel is a zoom highlighting the clear separation of the fiber traces, the space between blocks, and the spectrograph resolution (see~\S\ref{sec:psf} for more details on the resolution).} \label{fig:ccd-image} \end{figure} \section{Introduction} \label{sec:Introduction} The Dark Energy Spectroscopic Instrument (DESI) is a project whose primary objective is to measure with unprecedented precision the expansion history of the universe and the growth rate of large scale structure~\citep{DESI2016a}. The goal is to constrain better or detect a deviation from the standard cosmological model which relies on a puzzling cosmological constant or an unknown source of ``dark energy'' to explain the recent acceleration of the expansion. DESI should achieve a measurement of cosmological distances with the baryon acoustic oscillation (BAO) technique with an aggregate precision\footnote{Precision on the BAO peak position from the 2-point statistics of a galaxy sample covering a large redshift range, which requires assuming a fiducial cosmology to convert angles on the sky and redshifts to distances.} better than 0.3\% for redshifts below 1.1, 0.4\% in the range $1.1 < z < 1.9$, and 1\% at higher redshifts. It will use a 3D catalog of Luminous Red Galaxies (LRG), Emission Line Galaxies (ELG) and quasars (or QSO) as tracers of the matter density field, along with Lyman-alpha forests in the spectra of quasars at redshifts $z>1.9$. This result will be obtained with the spectra and redshift of about 40 million galaxies and quasars covering a 14,000 square degree footprint. About 10 million stellar spectra will also be acquired during the survey. These improved statistics, about 10 times larger than the SDSS~\citep{SDSS-DR16}, are made possible by an instrument specifically optimized for redshift surveys, with a large multiplex factor with 5020 robotically actuated fibers in a 3.2-deg field of view, a large aperture 4-m telescope, and high-throughput spectrographs with a spectral resolution adapted to detect the [\ion{O}{2}]~$\lambda \lambda 3726,3729$ doublet of faint emission line galaxies. In this paper we present the DESI spectroscopic pipeline, which consists of converting the raw CCD images of the 30 cameras from the 10 spectrographs into 5000 wavelength and flux calibrated spectra, along with a spectroscopic identification and redshift for all the observed targets. This pipeline inherits from the experience gathered over the years in the processing of SDSS BOSS and eBOSS survey data~\citep{Bolton2012,Dawson2013,SDSS-DR16}. However, the code was entirely rewritten and includes several conceptual improvements. The main one is related to the spectral extraction technique which consists of a forward model of the CCD image~\citep{BoltonSchlegel2010}, using a precise model of the 2D point spread function instead of a projected 1D cross-dispersion profile. This approach is more complex, it requires a specific linear system solver to account for the large number of correlated parameters to extract, and a technique to decorrelate the output flux values. Its advantages are a minimal variance, a common wavelength grid for all fibers, and a well defined resolution matrix that improves the sky subtraction, the spectro-photometric calibration and the redshift estimation. Another notable difference with the SDSS pipeline is the treatment of the noise where we rely on a model of the CCD image to estimate the CCD pixel Poisson noise, which ensures a linearity of the fluxes. We also pay attention to the error propagation down to the final calibrated spectra. After an overview of the instrument and the observations in Sections~\ref{sec:instrument} and~\ref{sec:observations}, we provide a detailed description of the algorithms that were developed and a first evaluation on their performance in Section~\ref{sec:algorithms-and-performance}. We describe the collaborative software development methods we have used in Section~\ref{sec:methodology}. We then present in Section~\ref{sec:data-processing} the data processing, including the real-time analysis, the daily updates, which comprise a quality assurance to identify failures and monitor the survey progress, and the large reprocessing runs with an homogeneous software version for the astrophysical and cosmological analyses. Finally in Section~\ref{sec:data-products} we give an overview of the data products before concluding in Section~\ref{sec:summary} with prospects for further improvements both for the data quality and the efficiency of the processing. Machine-readable tables of the data shown in some of the figures of this paper are available at \url{https://doi.org/10.5281/zenodo.7087815}. This paper documents the version of the spectroscopic pipeline used for the ``Fuji'' data processing run to be included with the DESI Early Data Release (EDR), covering Survey Validation and ancillary observations from December 2020 through May 13 2021. The software codes used for this run are available on GitHub\footnote{\url{https://github.com/desihub}}, with the tags (or versions) desispec/0.51.13, specex/0.8.4, specter/0.10.0, redrock/0.15.4, redrock-templates/0.7.2, desimodel/0.17.0, desitarget/2.4.0, and desiutil/3.2.5 (see also \S\ref{sec:methodology}). Algorithmic updates for future data releases will be documented as part of those releases. \section{Software development methodology} \label{sec:methodology} The DESI data team developed the spectroscopic pipeline as open source from the beginning, hosted on GitHub\footnote{\url{https://github.com/desihub}} with a BSD 3-clause license\footnote{\url{https://opensource.org/licenses/BSD-3-Clause}}. This was originally motivated by the pragmatic reason that GitHub provided free hosting to open source projects but charged for closed source, and it provided a better user experience at less cost (free) than hosting and maintaining our own closed-source software repository. Additionally, some institutions of contributing authors had burdensome procedures for publicly releasing previously proprietary packages, and it wasn't even clear that they had compatible release policies, but contributing to an already open-source package was straightforward. The open source nature of the work also enabled us to share code with other collaborations, resulting in benefits to both. e.g.~eBOSS used Redrock as its redshift fitter for its final cosmology results~\citep{Ross2020,GilMarin2020,Raichoor2020}, providing testing with real-world data years before DESI observations began. eBOSS contributed improvements back to Redrock, which accelerated DESI's readiness for initial on-sky observations. Code contributions follow a workflow of creating a local git branch, making updates, pushing the branch to GitHub, then opening a ``pull request'' (PR) to merge the updates back into the main branch. In most cases the PR is reviewed by someone other than the original author before merging. We did not follow strict and extensive code review practices, but this workflow generally helped ensure that multiple people were knowledgeable about any given piece of code and helped maintain higher standards of code quality. Packages also include unit tests that are automatically run with each PR, and were required to pass before merging. Like the code reviews, these were also implemented at a pragmatic level, usually testing ``does the code still run without crashing?'' rather than deeply testing algorithmic correctness. Some unit tests required large input files that were not viable to host within the testing framework at GitHub, so they are only run if the tests are run at NERSC. A nightly cronjob at NERSC updates every package, runs all unit tests, and generates a report altering the Data Systems manager if any tests are failing. Unit tests were augmented with end-to-end functional tests run every night, including simulating spectra, target selection, fiber assignment, survey simulations, pixel-level data simulations, and the spectroscopic pipeline run from raw data through redshifts. This allowed simple bugs to be caught early, freeing up developer attention for more subtle algorithmic and data quality studies. Code tags are made after major updates and before production runs. Since the DESI code is split across multiple packages (desispec, desitarget, specter, redrock, desiutil, ...) we require that the main branch of all packages remain compatible with each other, and the latest tags of every package remain compatible with each other. e.g.~if a PR for package X requires a feature in a branch of package Y, that branch in Y must be reviewed and merged before the PR in X can be merged. Quarterly software releases define a combination of code tags that are more extensively tested and confirmed to work together. Python was selected as the primary development language to prioritize developer efficiency over raw computational efficiency. At the same time, the core algorithms heavily leverage compiled libraries wrapped by numpy\footnote{\url{https://numpy.org}} and scipy\footnote{\url{https://scipy.org}} (e.g. LAPACK, BLAS), and key functions use numba\footnote{\url{https://numba.pydata.org}} just-in-time compilation. This combination of libraries has enabled good computational performance while benefiting from the flexibility and non-expert accessibility of Python. Most code was developed by scientist-programmers within the DESI collaboration, who had a vested interest in the quality of the outcome, rather than by professional programmers with limited domain knowledge. Two postdocs at the National Energy Research Scientific Computing Center (NERSC) focused on improving the code efficiency of the algorithms (see \citealt{laurie_stephey-proc-scipy-2019}) and porting these to Graphics Processing Units (GPUs) for future machines (see \citealt{margala-gpu-2021}). The code architecture strictly separates algorithms from the pipeline workflow wrappers that run those algorithms in parallel on multiple compute nodes for multiple input exposures. Any individual step (e.g.~sky subtraction, flux calibration) can be run completely independently of a production workflow by reading input files, running an algorithm, and writing output files. The parallel pipeline workflow then calls the identical functions, also reading and writing files as the method of passing data from one pipeline step to another. This structure enabled focused algorithm development on laptops with a deployment to High Performance Computing supercomputers using tens of thousands of cores in parallel to process hundreds of terabytes of data. Although less efficient from an I/O perspective, this design has been key to developer efficiency by enabling focused work and iterative debugging on individual algorithms with minimal conceptual overhead on how to run those steps. The file-centric method of data passing also provides a natural ``checkpoint restart'' design, such that any later step of the pipeline can be run using files produced by earlier steps, without having to re-run those earlier steps. \section{Observations} \label{sec:observations} We present in this section the sequence of observations and the data set used by the spectroscopic pipeline, starting with the afternoon calibrations and then providing some information about a typical exposure of the main survey. \subsection{Spectroscopic calibration observations} \label{sec:calib-obs} The instrument control system allows us to operate all 30 cameras simultaneously and control the calibration lamps. Each action (an exposure or another operation) is registered in a queue. The queue can be filled with interactive text commands or with predefined scripts, the latter being preferentially used for the routine operations like the ones described here. The afternoon calibration sequence starts with a broadcast call to a specific stabilization routine applied to all the CCDs. This stabilization consists in running the CCDs with a series of non-standard configurations of the clocks voltages to efficiently remove charges and reach rapidly a stable state (for most CCDs, and as long as they are continually clocked). We then acquire 25 ``zero'' exposures (zero exposure time and shutters closed) that are used to build a nightly master bias frame. This sequence is ended with a 300 sec ``dark'' exposure (still with the spectrographs' shutter closed) which is used to monitor the quality of the pre-processing (see \S\ref{sec:preprocessing}). Calibration exposures using the dome screen are acquired later in the afternoon, when the dome lights can be switched off after the maintenance operations by the day crew. We first acquire 5 exposures of 5 seconds with all the arc lamps on (Hg, Ar, Cd, Ne, Xe, Kr), then 5 exposures of 30 seconds with only the Cd and Xe lamps in order to fill wavelength gaps with low S/N, and then we turn off the arc lamps, and we turn on each of the four LEDs arrays from the four calibration boxes one at a time, taking for each of them 3 exposures of 120 seconds. The first series of five exposures is used to fit the point spread function and determine the wavelength calibration (see arc lamp spectra in Figure~\ref{fig:psf1}). We use five exposures in order to be robust to cosmic rays that sometimes compromise the PSF fit by hitting regions of the CCD with useful emission lines (see \S\ref{sec:psf}). The set of LED lamps exposures is used for the fiber flat fielding. Three exposures per illumination configuration are needed to mitigate the effect of cosmic rays and four configurations with one lamp at a time in order to inter-calibrate the lamps (see \S\ref{sec:fiberflat} for details). \subsection{Standard exposure sequence} \label{sec:std-expo-seq} The standard exposure sequence is described in detail in \cite{desi-collaboration22a}. We summarize here the information we need for the spectroscopic pipeline. A survey tile with assigned fibers consists of the sky coordinates of the tile center, a field rotation angle, and the sky positions of the 5000 fibers. During the telescope slew towards the tile center and the hexapod moves to rotate the field, the first blind moves of the positioners are operated. After the telescope slew and the field rotation, a 15 sec acquisition exposure is obtained to precisely point the telescope and perform a residual field rotation. After this move (telescope + hexapod), guide stars are starting to be measured in 6 of the 10 GFAs to maintain the telescope pointing, the fiber back-illumination is turned on, and a FVC image is taken to adjust the location of positioners. After the positioners are adjusted, a last FVC image is acquired, the back-illumination turned off, and the spectrographs' shutters are opened. The hexapod is regularly rotated during the exposure to compensate for the residual field rotation. Also during the exposure a dynamic exposure time calculator estimates the remaining time needed to achieve a pre-set effective exposure time which is a function of the sky brightness, the sky transparency, the airmass and the image quality. The sky brightness is estimated from sky monitor fibers placed on the edge of the focal plane and read regularly with a dedicated CCD camera. The sky transparency and image quality are determined from the GFA images. \subsection{Exposure data set} \label{sec:expo-data} At the end of the exposure sequence, the shutters are closed, the CCD read, and the instrument control system collects the individual images and saves them in a single fits files with multiple HDUs, while adding many header keywords about the telescope and its pointing, environmental parameters in the dome, along with monitoring parameters for the CCDs, cryostat, readout electronics, and temperatures and humidity in the vicinity of each of the 30 cameras. All of the guide star images are saved for offline analysis. This is useful to model the fraction of light in the fiber aperture given the current image quality and the performance of the guiding. Those images also provide redundant information about the sky transparency. The FVC images and in particular the coordinates of the fiber tips are recorded along with their expected positions for a perfect alignment on the targets. This information is used for the flux calibration, the estimation of the expected $S/N$ in the spectra, and for general quality assurance. The data from the dynamic exposure time calculator is saved, along with the input fiber assignment table associated with the observed tile, and a table of fiber positioning offsets, determined from the analysis of the FVC control image. All of this data is saved in one directory per exposure. This directory is copied to NERSC within a few minutes, and additional copies are stored at NOIRLab and the NERSC archival tape backup system within one day. The fiber assignment table contains important information about the targets for the spectroscopic pipeline. The fibers pointing to blank sky coordinates are separately identified; they are used to model the sky spectrum (see \S\ref{sec:skysubtraction}). The table also contains the various target bitmasks among which standard stars can be found for the flux calibration (\S\ref{sec:starfit}). It provides useful information from the imaging catalogs. This includes in particular the total flux in the $g$, $r$, and $z$ DECam pass-bands\footnote{See \citealt{Dey2019} for a description of the surveys and the pass-bands, and \url{https://www.legacysurvey.org/dr9/} for the imaging catalogs.}, along with the fiber fluxes, which are the fluxes one would have collected in a 1.5$\arcsec$ diameter fiber for a seeing or image quality of 1$\arcsec$ FWHM for the target, given its surface density profile as determined from the imaging data. \subsection{Offline Data Pipeline Overview} The offline pipeline is responsible for managing the processing of DESI data using the algorithms described in \S\ref{sec:algorithms-and-performance}. It is designed to operate at NERSC using the Slurm\footnote{\url{https://slurm.schedmd.com}} job scheduler system on both Cori and Perlmutter machines. While currently specialized for Slurm jobs, the code was designed to be as machine agnostic as possible, with the scheduler and machine specifications localized to several calling functions that could be upgraded in the future to work on other systems. The pipeline is tasked with two objectives: nightly processing and full dataset reprocessing. The nightly processing is the near-realtime analysis of data taken throughout a night with the goal of having processed spectra and redshifts by the next morning to inform observations the following night. The full dataset reprocessing is intended to give a self-consistent set of data processed with a tagged version of the offline pipeline, which could be used for publishable scientific analyses. These full reprocessing runs are what become internal data releases and public data releases. Both types of data processing perform the same series of steps. Calibration data is processed first. Master bias frames are computed for a night based on available bias exposures (with null exposure time). These bias levels are then used to pre-process a dark exposure used to identify bad columns on each of the 30 cameras. Both the bias and badcolumn files are then accounted for when processing the arc lamp exposures for PSF fitting. The PSFs for each camera are then used in determining the flat fielding vectors of each fiber using the flat field exposures. Finally, the science exposures are processed using all of the calibration information. The pipeline assures that these are run in the proper order by leveraging Slurm job dependencies. For example, the individual arc lamp exposures can be processed in parallel, but they all depend on the bad columns being successfully completed prior to their execution. Individual science exposures of the same tile are treated as independent for the extracting of spectra from the images and for the sky subtraction, but are processed together to jointly fit the standard star models that are later used for flux calibration. The flux calibration is again done individually and in parallel. Finally, redshifts are fit using all exposures of a target coadded together. A full description of an older version of the workflow can be found in \citet{Kremin2020}, which still provides a relatively accurate description of the general pipeline even though the names and timing information have changed. \subsubsection{Nightly Data Pipeline} Figure~\ref{fig:pipeline-flowchart} shows the interplay between the various tools and data throughout a night of processing. The nightly pipeline is managed by a script (\emph{desi\_daily\_proc\_manager} in Fig.~\ref{fig:pipeline-flowchart}) which identifies when new data has arrived at NERSC and submits jobs to the Slurm queue to process this data. The jobs submitted to the queue call \emph{desi\_proc} and \emph{desi\_proc\_joint\_fit} to process single exposure and multi-exposure operations, respectively. Since the data appears chronologically, the pipeline has to be robust to hardware and software failures as well as differences in data acquisition strategy and data quality. Survey validation led to dramatic improvements in uniformity, and nightly processing is now reliably supplied by the following morning with many nights requiring no human intervention. Figure~\ref{fig:nightly-pipeline-timing} shows an example night, April 24th 2022, where we successfully observed 15 tiles in 26 science exposures. The orange lines indicated arc lamp jobs (for PSF fitting), the green lines fiber flat field jobs, and the blue lines science jobs. For a given tile the joint fitting of standard stars across all exposures necessitates the waiting of some jobs for later exposures of a tile to complete. The merging of jobs seen in the middle of the night are depicting this scenario. The large delay in the rightmost exposure is a consequence of the reactionary nature of the pipeline. It will not submit the joint standard star fitting until it receives new data from a different tile or the end of the night is reached. The delay in the rightmost job at the end of the night is the pipeline waiting for the end of the night when it knows no new data will arrive before submitting the final exposure for standard star fitting, flux calibration, and redshift fitting. \begin{figure} \centering \includegraphics[width=0.7\columnwidth,angle=0]{pipeline-flowchart.pdf} \caption{Flowchart depicting the path data takes from local storage on the mountaintop to processing and derived data products, and useful web services for monitoring each. Each exposure is transferred to NERSC, where a pipeline manager identifies it and submits several jobs to process the data to flux calibrated spectra and redshifts. A dashboard checks the output data and reports progress on an interactive html page for use in monitoring progress.} \label{fig:pipeline-flowchart} \end{figure} \begin{figure} \centering \includegraphics[width=1.0\columnwidth,angle=0]{nightly-pipeline-timing.pdf} \caption{The speed at which data is processed throughout an example night, April 24th 2022. Orange lines are arc lamp exposures, green lines are flat exposures, and blue lines are science exposures. Science exposures are processed in four stages, with the second and fourth stages being jointly performed with all exposures for a given tile of targets. The various merging events in Job 2 are due to this joint fitting forcing earlier exposures to wait on later exposures. Per-tile jobs are submitted once data for a different tile appears or the end of the night is reached. The last tile is delayed until an end of night signal is triggered, which is the cause of the delay seen on the right-most exposure. The plot shows that on this night we had fully calibrated spectra and redshifts of targets within 20-40 minutes of the data becoming available at NERSC.} \label{fig:nightly-pipeline-timing} \end{figure} A table of exposure properties for a single night, called an \emph{exposure\_table}, is automatically generated as the exposures are identified by the pipeline. These tables provide per-exposure information about the type of observation, cameras available, and columns designating whether special considerations need to be made due to noisy amplifiers, failed positioning on a petal, or whether that exposure failed to meet our quality criteria and should be ignored. During the night the information is auto-populated, but can be updated as new details about specific exposures are uncovered. Version control is used as a safeguard. In addition to the repository, copies of tables used to process a specific release are kept with the data for future reference. \subsubsection{Reprocessing Data Pipeline} The nightly processing is run using the latest version of the pipeline on that given day. That is beneficial for providing the best outputs possible given the current understanding of the instrument and data. The implication of this is that earlier daily data are processed with less optimized versions of the pipeline. The objective of the data reprocessing is to process all data with a tagged version of the code, and utilizing knowledge of all data derived that exists for a night via the exposure tables that is not known for the realtime processing. Figure \ref{fig:reprocessing-timing} shows the performance of science exposure processing for per-tile spectra and redshifts. The left column shows the distribution of times for individual jobs for that step. The number of resources used for each step are not the same, so the right column shows the percentage of the total per-tile science processing spent in that step, computed as a function of core-hours for more accurate resource comparisons. The majority of the computational time is spent extracting the individual fibers from the images and in doing the redshift fitting of the flux calibrated spectra. For the release in question three types of redshifts were generated, which nearly triples the number of redshift function calls and time spent computing redshifts. Note also that the percentages are of the total compute time and there are steps not shown in a row. \begin{figure} \centering \includegraphics[width=1.0\columnwidth,angle=0]{reprocessing-timing.pdf} \caption{Distribution of wall clock times for the key science exposure processing steps for a recent internal data release. On the left is the histogram of job times in minutes. On the right is the percentage of the total computational time devoted to that step, in core-hours used, for the per-tile based science processing. Note that not all steps are shown, so the sum of the righthand column is not 100\%. Note also that three types of redshifts were performed, which nearly triples the computational footprint of the redshifts compared to earlier stages. The timing shows that the spectral extractions is the most resource intensive step of the spectral reductions, while the redshifting is the largest overall consumer of compute time.} \label{fig:reprocessing-timing} \end{figure} \section{Offline Quality Assurance Example} \label{sec:example-qa} We illustrate in Figure~\ref{fig:qa-cte} a charge transfer efficiency (CTE) issue for the petal 3 for one night (May, 24, 2022), and its impact on redshifts (see \S\ref{sec:qa} on Quality Assurance for context). The top panel shows a CTE issue for the C amplifier of the z3-camera: it displays the median counts values over the 21 rows just above/below the amplifier boundary along the wavelength direction (see Figure~\ref{fig:ccd-layout} for the CCD image layout with the amplifiers). The electrons in the amplifier C are not properly transferred for the CCD columns 1670 to 2057 (last column read with this amplifier) which approximately correspond to fibers 1705 to 1749. This results in an offset of a few electrons in the extracted spectra from those fibers. As a consequence those $\sim$45 fibers read from the z3-camera amplifiers A and C have an artificial discontinuity at the wavelength corresponding to the A and C amplifier boundary ($\sim 8750$\,\AA), identified by the redshift pipeline as a Balmer break at redshift $\sim$1.2 for low signal-to-noise spectra. This is visible in the bottom left plot showing for that petal the redshifts as a function of the fiber number for the (sky-subtracted) sky fibers only, with all those fibers having a best-fit redshift of $\sim$1.2. Lastly, the bottom right plot shows an ELG spectrum drawn from one of those fibers, where one can see the artificial flux decrease redwards of $\sim 8750$\,\AA, interpreted as an (inverted) Balmer break by the redshift pipeline. For CTE issue, the obtained data usually cannot be fixed, hence the corresponding fibers are declared as "bad" if the effect is strong; however, the issue can be resolved for later observations, with adjusting the CCD serial clock voltages. \begin{figure} \begin{center} \begin{tabular}{cc} \multicolumn{2}{c}{\includegraphics[width=0.95\columnwidth]{qa-ctedet.pdf}}\\ \includegraphics[width=0.45\columnwidth]{qa-skyzfiber.pdf} & \includegraphics[width=0.45\columnwidth]{qa-elgspectrum.pdf}\\ \end{tabular} \caption{ Example of a CTE issue for the petal 3 for one night (May, 24, 2022) identified by the QA. The top panel shows our CTE diagnosis for the z3 camera, displaying the median counts over the 21 rows just above/below the CCD amplifier boundary along the wavelength direction for a flat field exposure. The electrons from columns 1670 to 2057 are not properly transferred. The bottom left plot shows the redshift vs. the fiber number for petal 3 (sky-subtracted) sky fibers for all dark tiles from that night; the black circle highlights the patch of $z \sim 1.2$ for the impacted fibers. The bottom right plot shows an ELG spectrum drawn from one of those fibers, where one can see the artificial flux decrease redwards of $\sim 8750$\,\AA, interpreted as an (inverted) Balmer break by the redshift pipeline; the colored lines are the spectrum for each camera, and the black line is the best-fit model at $z = 1.2369$, with some corresponding lines. } \label{fig:qa-cte} \end{center} \end{figure} \subsection{Offline Quality Assurance} \label{sec:qa} As described in \S\ref{sec:data-products}, DESI observations are done through tiles: we validate during daytime the bright and dark program tiles observed during the night, using a wide range of Quality Assurance (QA) checks. Once a tile is QA-validated, we archive it (i.e. freeze it), update the spectroscopic status of the observed targets, and enable the observation of overlapping tiles in subsequent nights \citep[see][]{schlafly22a}. We present in this subsection the principle of this QA approach, then list the performed QA. An example is provide in Appendix~\ref{sec:example-qa}. This procedure allows us to daily monitor from an end-user point-of-view the appearance of instrumental features or specific observing conditions which impact the quality of the redshift measurement, and reduces the probability to run several nights with a lingering, unidentified issue. Most of the time, the found issue can be solved with either an intervention on the instrument or a change in the spectroscopic pipeline, and re-running the pipeline. In few cases, the issue cannot be solved, and the affected observed fibers are flagged as bad, i.e. are discarded. This daily QA task has proved to be very efficient to quickly uncover issues in the first months of the DESI Main Survey. With the survey progressing, the operations becoming more stable, and the spectroscopic pipeline more robust, those have become less frequent. Lastly, this daily QA task is done by a member of the Survey Operations team (approximately ten members), with weekly rotations, allowing at the same time to avoid possible bias from having only a few people performing the QA, and to have a core team familiar with those aspects of the data. The QA are performed at two levels: for each tile, and for the overall night. For each tile, we currently look at: \begin{itemize} \item the per-tracer n(z), and compare it to the expected one; \item the redshift $z$ as a function of the fiber number: this is efficient to identify problematic regions (e.g. a petal, a block of fibers due to some CCD feature); \item the fiber positioning accuracy as a function of the position in focal plane: this is useful to monitor for instance that the turbulence correction\footnote{The apparent relative displacements of fixed fibers and fiducials in the fiber view camera images are used to compute a distortion map. It is attributed to air turbulence between the camera and the focal surface and is used to correct the measured fiber positions.} is effective; \item the per-fiber spectroscopic effective time as a function of the position in focal plane; \item various per-petal diagnoses: the read noise, the number of good positioners, the number of standard stars used by the pipeline, the r.m.s. of the standard stars r-band flux ratios, the throughput ratio to the overall tile throughput, the sky fibers throughput r.m.s. and their reduced $\chi^2$; \item the sky position of the tile in the DESI footprint: this allows one to quickly identify if the tile falls in a particular region (e.g., close to the Galactic Plane, in a Galactic dusty region, or in the Sagittarius Stream); \item a cutout sky image of the tile location, with an overlay of the petal geometry and the Galactic dust contours: in addition to showing possible very dusty small regions affecting some fibers, it allows one to identify if a bright star falls in the tile footprint. \end{itemize} And for each night, we currently monitor the following: \begin{itemize} \item all calibration files are present; \item 5min \textit{DARK} image display (10 petals, 3 cameras): this allows one to visualize possible appearing CCD features; \item CCD bad columns identification (10 petals, 3 cameras); \item charge transfer efficiency (CTE; 10 petals, 3 cameras): a subtle effect (of the order of a few electrons) is a lower CTE in one amplifier of a given camera, for some CCD columns, leading to a discontinuity in low signal-to-noise spectra. This effect is caused by a CCD defect localised the serial register. It has been found in several CCDs and evolves with time. It can be detected with the analysis of a one second flat field image nightly taken for that purpose, where we control for each fiber the median flux (over 21 rows) above and below the CCD amplifier boundary (see example in Appendix~\ref{sec:example-qa}); \item 2D-image of sky-subtracted sky-fibers only (1 image per exposure): the (sky-subtracted) sky fibers being supposed to mostly be noise, such images are efficient to visualize remaining systematics in the data; \item the redshift $z$ as a function of the fiber number for the sky fibers only: any systematic feature in this plot highlights possible remaining issues in the data; for instance a mis-calibration between two cameras for a petal could drive the redshift pipeline to systematically identify it as a Balmer break at a given redshift, hence creating a horizontal line in that plot; \item various per-petal diagnoses: per-tracer n(z), per-tracer redshift success rate, new Ly-$\alpha$ identification rate for dark tiles\footnote{For the dark program, any newly observed quasar target securely identified as a quasar at $z \ge 2.1$ will be assigned for re-observation, in order to increase the signal-to-noise ratio of the spectra used for the Ly-$\alpha$ forest analysis \citep[see][]{schlafly22a}}; \item new Ly-$\alpha$ identification for dark tiles: comparison of the number of new identifications to expectation, given the tile coverage from previous observations. \end{itemize} Though the list of diagnoses is mostly stabilized by now, we stress that, depending on the gained experience with the data, additional ones could be added in the future (e.g., a control for out-of-focus data).\\ \subsection{Overview} The spectroscopic pipeline deliverables are wavelength and flux calibrated spectra of the observed targets (with flux variance, bit mask, spectral resolution for each wavelength and fiber), and a redshift catalog with a spectroscopic classification of the targets, their redshift uncertainty and a confidence level (see \S\ref{sec:data-products} for a more complete description of the data products). We review here the data flow. The algorithms, fitting procedures and performances are presented in more details in the following sections. The data flow is graphically presented in Figure~\ref{fig:data-flow}. \tikzstyle{block} = [rectangle, draw, text centered] \tikzstyle{choice} = [ellipse, draw, text centered] \tikzstyle{line} = [draw, -latex'] \tikzstyle{cloud} = [rectangle, draw,fill=gray!20, text centered, rounded corners] \begin{figure} \centering \begin{tikzpicture}[node distance=0.5cm] \tiny \node [block] (raw) {raw CCD image}; \node [choice, below of=raw, node distance=0.55cm] (biaschoice) {bias?}; \node [cloud, left of=biaschoice, node distance=3.5cm] (biasfit) {bias averaging}; \node [block, below of=biasfit] (bias) {Night bias frame}; \node [cloud, below of=biaschoice, node distance=0.8cm] (preprocessing) {pre-processing (\S\ref{sec:preprocessing})}; \node [block, below of=preprocessing] (preproc) {pre-processed image}; \node [choice, below of=preproc, node distance=0.55cm] (arclampchoice) {arc lamps?}; \node [cloud, left of=arclampchoice, node distance=3.5cm] (psffit) {PSF fit and averaging (\S\ref{sec:psf})}; \node [block, below of=psffit] (psf) {Night PSF}; \node [cloud, below of=arclampchoice, node distance=0.8cm] (wavecalib) {trace and wavelength calibration (\S\ref{sec:wave-calib-trace-coords})}; \node [block, below of=wavecalib] (exposurepsf) {Exposure PSF}; \node [cloud, below of=exposurepsf] (extraction) {spectral extraction (\S\ref{sec:extraction})}; \node [block, below of=extraction] (frame) {uncalibrated spectral frame}; \node [choice, below of=frame, node distance=0.55cm] (fiberflatchoice) {LED lamps?}; \node [cloud, left of=fiberflatchoice, node distance=3.5cm] (computefiberflat) {compute flat-field (\S\ref{sec:fiberflat})}; \node [block, below of=computefiberflat] (fiberflat) {Night fiber flat-field}; \node [cloud, below of=fiberflatchoice, node distance=0.8cm] (applyfiberflat) {apply flat-field}; \node [block, below of=applyfiberflat] (fframe) {flat-fielded spectral frame}; \node [cloud, below of=fframe] (skysub) {sky subtraction (\S\ref{sec:skysubtraction})}; \node [block, below of=skysub] (sframe) {sky-subtracted spectral frame}; \node [cloud, below of=sframe] (starfit) {stellar model fit (\S\ref{sec:starfit})}; \node [block, left of=starfit, node distance=3.5cm] (startemplates) {stellar models}; \node [block, below of=starfit] (stdstar) {standard star calibrated spectra}; \node [cloud, below of=stdstar] (fluxcalib) {flux calibration (\S\ref{sec:fluxcalibration})}; \node [block, left of=fluxcalib, node distance=3.5cm] (throughputmodel) {throughput model}; \node [cloud, below of=fluxcalib] (xtalk) {fiber cross-talk correction (\S\ref{sec:fibercrosstalk})}; \node [block, below of=xtalk] (cframe) {calibrated spectral frame}; \node [cloud, below of=cframe] (coadd) {spectral grouping and co-addition (\S\ref{sec:coaddition})}; \node [block, below of=coadd] (spectra) {spectra group}; \node [cloud, below of=spectra] (redrock) {identification and redshift fitting (\S\ref{sec:redshift})}; \node [block, left of=redrock, node distance=3.5cm] (galaxytemplates) {spectral templates}; \node [block, below of=redrock] (catalog) {redshift catalog}; \path [line] (raw) -- (biaschoice); \path [line] (biaschoice) -- node[left] {no} (preprocessing); \path [line] (biaschoice) -- node[above] {yes} (biasfit); \path [line] (biasfit) -- (bias); \path [line] (bias) |- (preprocessing); \path [line] (preprocessing) -- (preproc); \path [line] (preprocessing) -- (preproc); \path [line] (preproc) -- (arclampchoice); \path [line] (arclampchoice) -- node[above] {yes} (psffit); \path [line] (psffit) -- (psf); \path [line] (psf) |- (wavecalib); \path [line] (arclampchoice) -- node[left] {no} (wavecalib); \path [line] (wavecalib) -- (exposurepsf); \path [line] (exposurepsf) -- (extraction); \path [line] (extraction) -- (frame); \path [line] (frame) -- (fiberflatchoice); \path [line] (fiberflatchoice) -- node[above] {yes} (computefiberflat); \path [line] (computefiberflat) -- (fiberflat); \path [line] (fiberflat) |- (applyfiberflat); \path [line] (fiberflatchoice) -- node[left] {no} (applyfiberflat); \path [line] (applyfiberflat) -- (fframe); \path [line] (fframe) -- (skysub); \path [line] (skysub) -- (sframe); \path [line] (sframe) -- (starfit); \path [line,dashed] (startemplates) -- (starfit); \path [line] (starfit) -- (stdstar); \path [line] (stdstar) -- (fluxcalib); \path [line,dashed] (throughputmodel) -- (fluxcalib); \path [line] (fluxcalib) -- (xtalk); \path [line] (xtalk) -- (cframe); \path [line] (cframe) -- (coadd); \path [line] (coadd) -- (spectra); \path [line] (spectra) -- (redrock); \path [line,dashed] (galaxytemplates) -- (redrock); \path [line] (redrock) -- (catalog); \end{tikzpicture} \caption{Spectroscopic pipeline data flow} \label{fig:data-flow} \end{figure} \begin{enumerate} \item A first step for all incoming images is the pre-processing (\S\ref{sec:preprocessing}) of the CCD images, where ADC counts are converted to electrons per pixel, with an estimate of their variance, a subtraction of dark current, a flat field correction, and a mask indicating bad pixels (due to CCD defects or cosmic rays); this process also incorporates other corrections of the imperfections of the hardware. \item Then arc lamp calibration images are analyzed to perform a precise wavelength calibration, fit the spectral traces CCD coordinates and measure the spectroscopic Point Spread Function (hereafter PSF) shape (\S\ref{sec:psf}). In general, an average PSF is computed from the calibration data acquired during an afternoon to process the following flat-fielding calibration exposures and then the scientific exposures from the night that follows. The pipeline can also use a PSF model derived from previous nights if calibration data is missing. \item The wavelength calibration and trace coordinates vary by a fraction of a CCD pixel during the night because of small variations of environmental parameters in the spectrographs. For this reason an adjustment is required for each science exposure (\S\ref{sec:wave-calib-trace-coords}). \item The next step is the spectral extraction (see \S\ref{sec:extraction}). It consists in fitting a full model of the CCD image, using the known 2D PSF. The fit is a least square minimization with inverse variance weights ($\chi^2$). It is very close to be the statistically optimal estimator because the pixel noise is nearly Gaussian and uncorrelated. The drawback of this method is the computation challenge and the need for a specific treatment of the high frequency noise resulting from the deconvolution. \item We then proceed with the calibration of these spectra. We first use the extracted spectra from calibration exposures on the white spot illuminated with LED lamps to determine a fiber flat-field correction (\S\ref{sec:fiberflat}). It also corrects for residual CCD non-uniformities and accounts for the difference of throughput from one spectrograph to another. \item Once the spectra are flat-fielded, we use fibers intentionally pointing to empty regions of the sky (hereafter called ``sky fibers'') to compute a sky spectrum model and subtract this sky model to all the target spectra (\S\ref{sec:skysubtraction}). It makes use of the precise estimate of the spectral resolution determined during the extraction. \item The flat-fielded and sky subtracted spectra of standard stars are used to fit stellar spectral models, normalized by the photometric fluxes from the input imaging catalog. The fit uses data from several exposures from the same target. It does not require a prior knowledge of the instrument throughput. We detail the fit algorithm and the stellar templates in \S\ref{sec:starfit}. \item The spectral models of the standard stars can then be used to derive the instrument throughput as a function of wavelength by comparing the measured counts in electrons with the expected incoming flux (\S\ref{sec:fluxcalibration}). We use for this purpose our past knowledge of the instrument response while accounting for the variable atmospheric conditions and the optical properties of the telescope (focal plane PSF, plate scale, vignetting). \item We apply a fiber cross-talk correction (\S\ref{sec:fibercrosstalk}) to remove the contamination of spectra from neighboring bright fiber traces in the CCD images. \item Calibrated spectra from several exposures and nights are grouped together according to their location on the sky, and optionally co-added (or averaged), while avoiding any resampling of the spectral data arrays (see \S\ref{sec:coaddition}). \item Finally, regrouped spectra are analyzed to classify targets and measure their redshifts. We give an overview of the methods used in~\S\ref{sec:redshift}. \end{enumerate} \section{Algorithms and performance} \label{sec:algorithms-and-performance} \input{overview} \input{preprocessing} \input{psf} \input{wavecalib} \input{extraction} \input{fiberflat} \input{skysubtraction} \input{starfit} \input{fluxcalibration} \input{fibercrosstalk} \input{coaddition} \input{redshifts} \input{effectivetime} \input{methodology} \section{Data processing} \label{sec:data-processing} \input{realtime-processing} \input{offline-processing} \input{offline-qa} \input{dataproducts} \input{summary} \include{acknowledgments} \bibliographystyle{aasjournal} \subsection{CCD calibration and pre-processing} \label{sec:preprocessing} Pre-processing is the task that consists in converting the 'raw' CCD images, consisting of an array of Analog Digital Units (hereafter ADUs) per pixel into a number of electrons per pixel, with a flat-field correction, an estimate of the variance in each pixel, and a mask to discard some pixels from the subsequent spectroscopic analysis. \subsubsection{Bias, dark and over-scan subtraction} \label{sec:bias-subtraction} The first step of the pre-processing is the subtraction of a bias frame to the raw image in ADUs. The bias frame is either a yearly master bias obtained independently from past observations or a nightly master bias. The latter is obtained from the series of 25 bias exposures taken during the afternoon as part of the daily calibration sequence (see \S\ref{sec:calib-obs}). This master bias is simply the clipped average per pixel of the raw CCD images after over-scan subtraction, including the pre-scan and over-scan columns and rows. The clipped average is an iterative evaluation of the mean and dispersion of the data with a 5-$\sigma$ outlier rejection. The initial values for the mean and dispersion are the median and the normalized median absolute deviation (NMAD) of the pixel values. The choice of using the yearly or nightly bias frame is based on the reduction of a 300 second dark exposure taken during the afternoon calibration. This dark frame is fully pre-processed (including dark current subtraction) with both bias frames, and the bias that gives the smallest residuals in the processed image is selected for the following night of data. This procedure allows us to account for night to night variations of the bias levels while protecting us from unsettled CCD biases that happen in some rare sequences when one or several CCDs had been turned off earlier during the day. After subtracting the reference bias frame we proceed with the over-scan subtraction. We use by default the clipped mean of all of the pixels in the over-scan columns, per CCD amplifier. We however apply a more complex treatment in two situations. First, some CCD quadrants have bias fluctuations with a correlation length exceeding the length of a CCD line. Some rare nights for some specific CCD amplifiers, the amplitude of the correlated fluctuations exceeds a threshold for which it becomes advantageous to perform a bias subtraction per row (i.e. using the clipped average of each row independently in the over-scan area). This decision is made automatically in the pre-processing with a measure of the amplitude of the variation from row to row (recorded under the {\it OSTEP\{A,B,C,D\}} keyword in the pre-processed CCD image headers). When doing so, we have to first detect and mask over-scan rows affected by residual trailing charges (positive or negative, see \S\ref{sec:darktrails}) from large cosmic ray deposits in the last columns of the CCD active region. For problematic CCD amplifiers (only one among 120 for the first DESI data release), we apply an additional correction derived from the average of the over-scan rows as a function of the column number. In addition to a standard dark current subtraction, proportional to the exposure time, we also correct for residual bias levels depending non-linearly on the exposure time. This extra bias level in the readout electronics builds up during the exposure (and so is not present in the master bias), and then decays as the CCD is read, leaving an exponential profile as function of the CCD row, with a maximum in the row that is read first for each amplifier (the top or bottom row on the CCD). This effect is largely mitigated by reading 300 'blank' lines in the CCD (without the row to row charge transfer with the parallel clocks) but a residual effect of a few electrons at maximum is left and corrected with calibration data obtained from the analysis of many dark exposures with various exposure times, from 1 second to 1200 seconds. The procedure we just described is not sufficient to obtain for all exposures a residual bias level below one electron per pixel for the blue CCDs. Such a residual bias can lead to spectral discontinuities of about 5 electrons per Angstrom (the effective width of the cross-dispersion is about 3 pixels, and a pixel corresponds to a wavelength variation of 0.6\,\AA\ along the dispersion axis). For this purpose we evaluate a residual background in the space between the fiber traces of adjacent fiber blocks in each science exposure, and then interpolate the values across the whole CCD. This background estimate is obtain with a median filter across CCD rows of the mean bias level in each row in the space between blocks. We have reprocessed a large number of science exposures to verify that this new approach improves the bias level estimation and the redshift performances. \subsubsection{Gains} Gains in electrons/ADU are applied to the bias subtracted images, per amplifier. The gain values have been obtained from photon transfer curves (i.e. variance as a function of mean counts) derived from exposures with the telescope pointed towards the dome screen, the latter being illuminated with LED lamps. We take advantage of the excellent temporal stability of the LEDs to estimate the variance from a comparison of spectra obtained in consecutive exposures. A boxcar extraction is used for this purpose to avoid potential biases introduced by variations in the position of spectral traces. Pairs of observations of different exposure times are used to cover the dynamic range from 0 to 5,000 electrons. The gains measured with this technique are reproducible and consistent with measurement obtained in very different conditions on a test bench at CEA, Saclay, France. We also do not detect discontinuities at the boundary of amplifiers after applying the gains. Figure~\ref{fig:gain} shows an example of such a transfer curve. \begin{figure} \centering \includegraphics[width=0.99\columnwidth,angle=0]{ptc-r2-C.pdf} \caption{Photon transfer curve for spectrograph SM5, camera r, amplifier C, obtained with pairs of exposures of 1, 2, 5, 10, 20 and 50 seconds with the LED lamps illuminating the dome screen. The units on the figure are ADU (or ADC counts). A gain of 1.509\,electron/ADU is determined from the inverse of the slope.} \label{fig:gain} \end{figure} \subsubsection{Electronic cross-talk correction} An electronic cross-talk between amplifiers of the same CCD has been detected for some chips. The measurement was done with a sparse fiber slithead where only a few isolated fiber traces were used to detect a flipped ghost trace on another amplifier. It is not possible to perform this study with the science slit as the true fiber traces and the ghosts from other amplifiers overlap. A few pairs of amplifiers have cross-talk amplitudes exceeding $10^{-4}$, the maximum value being $6 \times 10^{-4}$ with a contamination of amplifier A into B for the blue CCD of the spectrograph SM2, but the signal is much smaller or undetected for the majority of CCDs. Although the cross-talk effect is arguably negligible, we nevertheless correct for the effect in image pre-processing. \subsubsection{Cosmic ray identification} Cosmic ray hits have to be identified and masked with a very high efficiency to avoid spurious emission lines in the spectra that could lead to incorrect spectroscopic identification and redshifts for emission line galaxies. We are rejecting cosmic rays with the following steps during the pre-processing (see illustration on Figure~\ref{fig:cosmicrays}). 1) We first mask pixels using the spatial gradient of the intensity in the CCD. We mask a pixel if its value is significantly above its neighbors along at least two axes out of four, counting diagonals (Eq~\ref{eq:cosmicray_criterion1}), and if the intensity gradient relative to the peak value is significantly larger than the expectation given the spectrograph point spread function along one of the four axes (Eq~\ref{eq:cosmicray_criterion2}). \begin{subequations} \label{eq:cosmicray} \begin{equation} P > P_{neigh} + N \, \sigma(P_{neigh}) \label{eq:cosmicray_criterion1} \end{equation} \begin{equation} c_2 ( P - c \, \sigma(P) ) \delta_{PSF} > P_{neigh} + c \, \sigma(P_{neigh}) \label{eq:cosmicray_criterion2}\end{equation} \end{subequations} where $P$ is the value of the pixel being tested, $\sigma(P)$ its uncertainty, $P_{neigh}$ is the average of the two pixels neighboring the pixel of interest along one of the four axis, $\sigma(P_{neigh})$ its uncertainty, $\delta_{PSF}$ is the maximum PSF variation along this axis, and $N=6$, $c=3$, and $c_2=0.5$ coefficients that have been adjusted on the data. We follow here the algorithm implemented for SDSS imaging\footnote{See the Photo Pipelines documentation, \url{https://www.astro.princeton.edu/~rhl/photo-lite.pdf}}, with minor modifications. The gradients $\delta_{PSF}$ have been derived from the measured point spread function. There is one set of values for each camera arm (blue, red, and NIR). 2) After a first pass, we repeat the test with a lower significance threshold $N=3$, but only applying it to the pixels neighboring the ones that have already been masked. This last step is repeated as long as new pixels are masked. 3) We then broaden the masked regions in all four directions by 1 pixel. 4) Finally, to connect separated clusters of pixels belonging to individual cosmic-ray tracks, we perform a morphological binary closure \citep{Dougherty:2003}. The closure applies a sequence of seven ``structuring elements'' shown in Figure~\ref{fig:binaryclosure} to the cosmic-ray mask, with the elements chosen to align with possible track orientations. The closure operation produces seven masks which are combined with a bitwise OR. \begin{figure} \centering \includegraphics[width=0.99\columnwidth,angle=0]{cosmics-example.pdf} \caption{A $100\times100$-pixel region of an image from a red camera. The upper panels show the pixel values with the pixels identified as cosmic ray masked out. The bottom panels show the masked pixels at each step of the cosmic ray identification procedure.} \label{fig:cosmicrays} \end{figure} \begin{figure} \centering \setlength{\tabcolsep}{0pt} \renewcommand{\arraystretch}{0} \begin{tabular}{ccccccc} \includegraphics[width=0.142\columnwidth]{cosmics_lse_000.pdf} & \includegraphics[width=0.142\columnwidth]{cosmics_lse_020.pdf} & \includegraphics[width=0.142\columnwidth]{cosmics_lse_045.pdf} & \includegraphics[width=0.142\columnwidth]{cosmics_lse_070.pdf} & \includegraphics[width=0.142\columnwidth]{cosmics_lse_110.pdf} & \includegraphics[width=0.142\columnwidth]{cosmics_lse_135.pdf} & \includegraphics[width=0.142\columnwidth]{cosmics_lse_160.pdf} \end{tabular} \setlength{\tabcolsep}{6pt} \renewcommand{\arraystretch}{1} \caption{Seven $11\times11$-pixel binary ``structuring elements'' used to merge disconnected particle tracks with binary closure during the final step of cosmic-ray identification.} \label{fig:binaryclosure} \end{figure} The outcome of cosmic-ray masking, broadening, and binary closure is shown in Figure~\ref{fig:cosmicrays}. Note that the residual signal from cosmic rays is further detected in later steps of processing (during the spectral extraction, at each step of the spectral calibration, and finally when spectra from several exposures are co-added). \subsubsection{Negative trail correction} \label{sec:darktrails} Negative trails have been found for bright illuminations (typically cosmic rays) along the serial transfer direction for some amplifiers in the red and NIR CCDs. We have not been able to remove those with changes to the parameters of the CCD readout (clocks voltages and profiles, digital correlated double sampling parameters, namely the position and size of the readout time window and the number of digital samples). Fortunately this effect has been found to be linear, of small amplitude (0.1\% of the charge for the worst amplifier, see Table~\ref{table:negative-trails}), and easy to model with a simple spatial convolution of the CCD image. We have used images from bright illuminations with a continuum lamp to fit the parameters of an exponential convolution kernel along the direction of serial transfer (CCD rows). The parameters for the few affected amplifiers are given in Table~\ref{table:negative-trails}. We subtract to the original image the convolved image as part of pre-processing. This process should in principle be iterative but this is not necessary given the amplitude of the effect. It also introduces correlated noise between pixels but this is also fortunately a negligible effect. An example of an image before and after this correction is shown on Figure~\ref{fig:negative-trails}. \begin{table} \begin{tabular}{ccccc} Spectrograph, & $A$ & $L$ \\ Camera, Amplifier & & (pixels) \\ \hline SM1 r B & 0.00024 & 13 \\ SM2 r D & 0.00021 & 18 \\ SM2 z D & 0.00006 & 28 \\ SM3 r B & 0.00020 & 23 \\ SM6 r B & 0.00120 & 6 \\ SM6 z C & 0.00005 & 20 \\ SM6 z D & 0.00008 & 35 \\ SM7 r B & 0.00041 & 4 \\ \hline \end{tabular} \caption{Parameters of the negative trails convolution kernel $K(x) = -(A/L) \exp(-x/L)$ for the CCD amplifiers where this effect has been detected.} \label{table:negative-trails} \end{table} \begin{figure} \centering \includegraphics[width=0.9\columnwidth,angle=0]{negative-trails-sm6-r3.pdf} \caption{A 200x200 pixel region of a dark image from the red camera of spectrograph SM6, without and with the negative trail correction. Note the bright trails corresponding to negative values in this inverted gray scale on the left of cosmic ray hits on the left panel. The serial clock moves charges to the right in this region of the CCD read with amplifier B.} \label{fig:negative-trails} \end{figure} \subsubsection{CCD flat field} The CCD flat fielding has been obtained on site at Kitt Peak using a spectrograph slit head specifically designed for this purpose. During calibration runs, the science slit is removed from the spectrograph and the flat field slit inserted. It consists of a fiber, on the tip of which is located a lens that spread the light to a diffusing sheet in front of which is placed a thin slit. The thin slit is purposefully slightly offset from the spectrograph best focus position in order to smooth out the image on the spectrograph. A special purpose software is used to correct for the non-uniformities of the images due to the illumination pattern while preserving to the best possible extent the variation of efficiency of the CCD pixels. The algorithm consists in iteratively dividing the image by the same image convolved with a 1D Gaussian kernel, one axis at a time, while masking pixels identified as outliers at each iteration. An example is given on Figure~\ref{fig:pixel-flat-field}. \begin{figure} \centering \includegraphics[width=0.9\columnwidth,angle=0]{pixelflat.pdf} \caption{Left: a 300x300 pixel region showing the average of a series of images obtained with the flat field slit for the red camera of spectrograph SM8. Right: the pixel flat field correction derived from the image. The range of scales is $\pm$20\% in the left panel and only $\pm$5\% in the right panel. A lighter gray corresponds to a higher efficiency. The vertical strips on the left panel are caused by small variations in the slit width. The gradient from top to bottom is due to spectral variations in the illumination.} \label{fig:pixel-flat-field} \end{figure} \subsubsection{Scattered light} During the first few months of the spectrographs' commissioning, we have found a continuous increase of scattered light in the CCD images. This was caused by an accumulation of contaminants on the cryostat windows due to a polluted source of dry air. This has since then been corrected, the windows have been cleaned, and we no longer detect a significant amount of scattered light. In order to process this early data set, we have modeled this scattered light contribution as a convolution of the input image with a 2D kernel derived from the profile of scattered light around bright spots of arc lamp images. The model image is then modulated with a 2D smooth correction which coefficients are determined by comparing the model to the background level measured in the space between blocks of fiber traces. This is similar to the residual background subtraction presented in \S\ref{sec:bias-subtraction}. This algorithm was efficient when the amount of scattered light was small. It is disabled for the processing of recent data for which there is no detectable scattered light. \subsubsection{Bad pixel mask} A 2D map of bad pixels in each of the CCDs is provided as part of the pre-processing which we incorporate into a bitmask, each bit flagging a specific issue. The pixels flagged during pre-processing have large dark currents, are saturated, affected by cosmic rays, in regions of dead columns or large cosmetic defects, or have a low flat field value. A static bitmask marking some permanently flagged pixels is used as an input to the pre-processing step. This is obtained flagging outliers in the 2D distribution of the median and inter-quartile range (IQR) of pixels measured from a series of bias subtracted and gain applied set of dark exposures. The algorithm then uses a binary closing operation to flag CCD pixels completely surrounded by masked pixels. A standard set of thresholds for median and IQR work well for most CCDs but for some CCDs, it was necessary to manually adjust them based on a visual inspection of the mask and 2D distribution of the median and IQR values. These pixel masks need to be regularly updated to reflect any changes to the hardware. We do not provide here the list of bits and their meaning but invite the interested reader to find this information in the DESI data model documentation\footnote{See~\url{https://desidatamodel.readthedocs.io/en/latest/} for the current version, but note this documentation will evolve with data releases.} which will vary from one data release to the next. \subsubsection{Estimating the variance} \label{sec:ccd-variance} The read noise is determined from the dispersion in the over-scan regions. The values from each CCD and each amplifier are given in Table~\ref{table:spectrographs-ccd-rdnoise-table} of Appendix~\ref{sec:ccdproperties}. A method used in many spectroscopic pipelines to estimate the variance in pixels in the active region of the CCD has been to use the sum of the read noise variance and an estimate of the Poisson variance directly inferred from the individual pixel values, once converted in electrons. $$ \widetilde{\sigma}_P^2 = \sigma_R^2 + P \ \mathrm{if} \ P>0$$ As mentioned in~\citet{Horne1986}, this method introduces a correlation between the pixel value and its estimated variance that propagates to spectra and leads to biases in the estimation the sky level, the calibration, and the final calibrated spectra. For instance the weighted mean of $P$, with weights equal to the estimated inverse variance $\widetilde{\sigma}_P^{-2}$ is off by $-1$ at the leading order. The variance itself is also biased at low counts, because of the truncation to positive values. In order to circumvent this problem, we have resorted to use a full model of the CCD image to estimate the Poisson noise. We proceed as follows. i) We first extract the spectra with a fast boxcar extraction, after having adjusted the spectral traces (see \S\ref{sec:adjustcalib}). ii) After fiber flat fielding based on predetermined calibration data (see \S\ref{sec:fiberflat}), we compute an average sky model with a simple wavelength resampling and averaging of the sky fiber spectra. iii) We determine a smoothed version of the sky subtracted spectrum of each target, obtained with a convolution using a Gaussian kernel of parameter $\sigma=10$\,\AA, with outlier rejection (to limit the impact of residual cosmic rays), and after interpolating over values affected by masked pixels. iv) The final spectral model is obtained by adding back the sky model to the smoothed spectrum and applying the inverse of the fiber flat field correction v) We then project the spectra back on the CCD, assuming a Gaussian cross-dispersion profile. We keep the original CCD pixel values that are inconsistent with the model by more than 5 standard deviation and use the model for the others to derive the Poisson noise. An example of the inverse variance model image compared to the traditional estimate $\widetilde{\sigma}_P^{-2}$ is shown on Figure~\ref{fig:inverse_variance_model}. \begin{figure} \centering \includegraphics[width=0.9\columnwidth,angle=0]{inverse-variance-model.pdf} \caption{Left: a 150x150 pixel region showing the inverse variance derived from the individual pixel values for a 1200 second dark time exposure in a red camera CCD. Right: same pixel area with the inverse variance model used for the data processing. Spectral traces appear dark as the Poisson noise is larger and so the inverse of the variance smaller. Note the model has much lower noise. The residual spatial noise is coming from the pixel flat field. Note also that the inverse variance on cosmic rays and bright fibers is preserved.} \label{fig:inverse_variance_model} \end{figure} The variance model ensures the linearity of the spectroscopic fluxes. It is important for analyses like the Lyman-alpha forest \citep{Bautista2017}. The propagation of the pixel variance to the extracted spectral flux variance is described in \S\ref{sec:extraction}. This variance estimate is then combined with other sources of uncertainties and compared to the scatter in the sky background flux in \S\ref{sec:sky-residual-variance}. \subsubsection{Automated detection of bad columns} The daily calibration sequence includes a 300 second dark exposure that can be used to verify that the pre-processing accurately subtracts the bias level and dark current. As part of the automated daily data reduction, we pre-process this exposure, and measure the median value per column of each CCD amplifier. CCD columns with values exceeding a threshold of 0.005 electron per second per pixel, positive or negative, are recorded in a calibration file for each camera. This file is subsequently used for each exposure to flag fiber spectra affected by the bad columns. The spectral contamination is estimated using the distance between fiber spectral trace (see~\S\ref{sec:adjustcalib}) and the bad column, using a pre-determined cross-dispersion profile. Spectral fluxes with contamination larger than 0.005 electron per Angstrom per second are flagged. \subsection{Spectrograph point spread function} \label{sec:psf} We present in this section the point spread function (PSF) used for the spectral extraction (see \S\ref{sec:extraction}). We first describe the model and qualitative properties of the PSF, we then describe the method we use to fit this PSF model, and finally we discuss the stability of the PSF as a function of time, temperature, telescope pointing and fiber positioner moves. \subsubsection{Model} We consider here the PSF of the system composed of the fiber and spectrograph, such that the PSF is the convolution of the image of the fiber tip (near field) convolved with the optical blur from the spectrograph cameras. The diameter of the fiber image on the CCD is of 51\,$\mu$m, or 3.4~CCD pixels, after accounting for the camera demagnification of 0.48 (see \citealt{DESI2016b}). The PSF appears as a blured disk presenting a central plateau because the fiber image is partially resolved (see Fig.~\ref{fig:psf-r1}). We have considered an empirical model of the central part of the PSF consisting in a linear combination of Gauss-Hermite functions. \begin{eqnarray} \small PSF_H&&(x,y) = \sum_{i,j} a_{i,j} He_i \left( \frac{x-x_c}{\sigma_x} \right) He_j \left( \frac{y-y_c}{\sigma_y} \right) \nonumber \\ && \times \frac{1}{2 \pi \sigma_x \sigma_y} \exp \left( -\frac{(x-x_c)^2}{2 \sigma_x^2}-\frac{(y-y_c)^2}{2 \sigma_y^2} \right) \label{eq:psf} \end{eqnarray} where $He_i$ is a (probabilist's) Hermite polynomial of degree $i$ (from 0 to 6), $x_c$ and $y_c$ are the coordinates of the spectral traces in CCD pixels, and $\sigma_x$ and $\sigma_y$ the Gaussian parameters along both axis. The coefficients $(a_{i,j})$ and the parameters $x_c,y_c,\sigma_x,\sigma_y$ are allowed to vary continuously with wavelength and from fiber to fiber. We model them as Legendre polynomials of the wavelength, per fiber. The primary motivation for this modeling approach is its flexibility to describe most PSF shapes. Having the complex PSF shape described by a linear combination of components makes the fit convergence easier. Also the basis is orthogonal which means uncorrelated best fit $(a_{i,j})$ coefficients in the limit of small pixel size and constant pixel weights. Finally the integral of each component is analytic, so the integral of the PSF in the pixels is analytic as are all the derivatives of the model with respect to its parameters. This reduces considerably the computation time for the fit. This choice of basis is however not a perfect match to model a partially resolved disk: a Hermite degree as high as 6 is needed to model reasonably well the PSF shape. The spectrograph PSF is also composed of a faint extended tail due to the roughness of the optical surfaces and impurities that scatter a fraction of the light at large angles. A power law component has been implemented in the model to account for this effect. It is however not used in the current version of the pipeline as the Gauss-Hermite terms appear to be sufficient to describe the PSF tails up to a scale corresponding to the spacing between neighboring fiber traces (about 7 pixels). We will see however in \S\ref{sec:fibercrosstalk} that we still need to correct for a fiber cross-talk induced by those tails. We chose to ignore extended PSF tails in the spectral extraction (and PSF modeling) to mitigate the effect of unmasked cosmic ray hits that would contaminate many spectra across many fibers and wavelengths if extended PSFs were used for extraction. Reducing the PSF size also helps the computing performance, and finally, we have found that the cross-talk correction in post-processing was precise enough. Figure~\ref{fig:psf-r1} illustrates the PSF shape for a central fiber of a red camera CCD, at 7034.3\,\AA\ which is a bright Neon line. The cross-dispersion profile highlights the central plateau that shows the fiber image is nearly resolved. The profile also illustrates that there is very little overlap between the spectral traces of adjacent fibers. It is also made clear on the figure that the sharpness the PSF profile requires that we integrate precisely the PSF in the pixels instead of sampling its value at the pixel center. \begin{figure} \centering \includegraphics[width=\columnwidth,angle=0]{psf-r1.pdf} \caption{Example PSF of a central fiber (\#260) at 7034.3\,\AA\ in a red camera CCD. From top left to bottom right: 2D PSF model, PSF model integrated in pixels, arc lamp image where PSF from neighboring fibers is visible, and cross-dispersion profile of the PSF.} \label{fig:psf-r1} \end{figure} The average PSF FWHM (full width half max) is represented in Figure~\ref{fig:psf-fwhm} as a function of wavelength for the three cameras of a spectrograph. \begin{figure} \centering \includegraphics[width=\columnwidth,angle=0]{fwhm.pdf} \caption{Spectrograph PSF FWHM in the blue, red and NIR cameras, in units of CCD pixel (the pixel size is 15\,$\mu$m). This is the width of a slice of the PSF along the dispersion axis and it is not convolved with the pixel size, so this is different from the line spread function FWHM. This PSF width as to be compared with the diameter the fiber tip image of 3.4 pixel. The solid curves are the average PSF widths and the dotted curves the minimum and maximum PSF widths.} \label{fig:psf-fwhm} \end{figure} \subsubsection{Fit procedure} The PSF parameters are fit for each arc lamp exposure. The fit is performed with the {\it specex}\footnote{{\it specex:} \url{https://github.com/desihub/specex/}} C++ package that is part of the DESI software suite. The parameter estimation consists in a classic least square minimization of the model parameters using the Gauss-Newton algorithm. The inputs to the code are a pre-processed image (see~\S\ref{sec:preprocessing}) with the pixel counts, their variance and a mask, a first approximate solution of the trace coordinates, wavelength calibration, and PSF, and a list of arc lamp lines with their wavelength in vacuum (given in Table~\ref{table:arc-lamp-lines-list} in Appendix~\ref{sec:arc-lamp-lines-list}). The parameters are the Legendre coefficients of the parameters $(a_{i,j}),x_c,y_c,\sigma_x,\sigma_y$, and the intensity of all the arc lines from all the fibers. We note here that we use the arc lamp exposures to adjust the cross-dispersion coordinates of the fiber traces ($x_C$) when we could have used the continuum lamp exposures which contain more information. Our approach has the advantage to use in a consistent way the 2D PSF model for both $x$ and $y$ directions. We have verified that we have enough emission lines to get precise trace coordinates for all wavelength. As we will see in~\S\ref{sec:adjustcalib}, we also readjust on the actual data the fiber-trace coordinates with low order polynomial corrections. The fit actually determines the Legendre coefficients of two dimensional polynomials of the CCD coordinates ($x,y$), which allows us to account for spatially continuous changes of the resolution in the camera. The final output is presented in the form of Legendre polynomials of wavelength and fiber number which is more convenient for the spectral extraction. The fit is iterative because the model is a non-linear function of its parameters. Derivatives are analytic but the code is still CPU intensive because we integrate the PSF for all of the lines in all of the fibers at each minimization step. The balance between precision and computing time can be adjusted with varying the number of Gauss-Hermite terms, the PSF stamp size, and the number of emission lines entering in the fit. The fit is performed in several steps as follows: i) first the Legendre coefficients of trace coordinates and wavelength solution ($x_c,y_c$) polynomials are adjusted using the input PSF, ii) the Gaussian $(\sigma_x,\sigma_y)$ Legendre coefficient are fit, iii) the Legendre coefficients of the Gauss-Hermite terms $(a_{i,j})$ are fit. The line intensities are fit at the same time as the coefficients for all three steps. The code contains additional steps to optionally fit the Lorentzian tails, account for masked spots and dead fibers. The PSF fit is performed independently per group of 25 fibers, following the fiber pseudo-slit layout composed of 20 blocks of 25 tightly packed fibers each, with an extra spacing of about 2 inter fiber distance between the blocks. There is no constraint of non-negativity for the (pixel-deconvolved) PSF model during the PSF fit. However, for extractions, the PSF values integrated in CCD pixels are forced to be positive and the PSF renormalized accordingly. The agreement between the best fit PSF model and the spectroscopic data is presented in Figures~\ref{fig:psf1} and~\ref{fig:psf2}. In the first figure are shown the data and model spectra from a typical arc lamp calibration exposures with Mercury, Argon, Cadium, Neon, Krypton and Xenon lines. The quantity on the $y$-axis is the average pixel counts per row of a band covering the spectral traces of a central fiber block (fiber block with fibers 250 to 274). Both the average signal from the pre-processed image and the model image are computed and displayed. The spectra covering the whole CCD range are shown, along with zooms on some emission lines. There is an excellent agreement between the model and the data. The second figure (Fig.~\ref{fig:psf2}) represents cross-dispersion profiles from the CCD images of the same exposure and fiber block. The signal being displayed is now the average per column of an horizontal band centered on the brightest central emission line for each camera. Again a good agreement between the data and model is found. In particular, the logarithmic scaled figures show that the PSF tails are well described with the Gauss-Hermite model. As mentioned above, PSF Lorentzian tails are not part of the model used for extraction. \begin{figure} \centering \includegraphics[width=\columnwidth,angle=0]{arc-spectrum-b1.pdf} \includegraphics[width=\columnwidth,angle=0]{arc-spectrum-r1.pdf} \includegraphics[width=\columnwidth,angle=0]{arc-spectrum-z1.pdf} \caption{Spectra from arc-lamp exposures in the blue, red and NIR camera CCDs, from top to bottom. The y-axis gives the average number of electron per pixel, for each row of a vertical band covering fibers 250 to 274. The blue curve is the data and the orange dashed curve the best fit model from the PSF fit. The full wavelength range accessible for each camera is shown.} \label{fig:psf1} \end{figure} \begin{figure} \centering \includegraphics[width=\columnwidth,angle=0]{arc-xprof-b1.pdf} \includegraphics[width=\columnwidth,angle=0]{arc-xprof-r1.pdf} \includegraphics[width=\columnwidth,angle=0]{arc-xprof-z1.pdf} \caption{Cross-dispersion profile from arc-lamp exposures in the blue, red and NIR camera CCDs, from top to bottom. The y-axis gives the average number of electron per pixel, for each column of a narrow band centered on the brightest emission line in the central region of the CCD. The blue curve is the data and the orange dashed curve the best fit model from the PSF fit. Both linear and log-scale graphs are shown to highlight the agreement of the PSF model with data in the core and tails of the PSF.} \label{fig:psf2} \end{figure} \subsubsection{Stability} The stability of the spectroscopic PSF is critical for the quality of the spectroscopic data reduction. Because the PSF model is fit on the arc lamp calibration images in the afternoon preceeding the night of observation (see \S\ref{sec:calib-obs}), the spectroscopic PSF has to be stable for a duration of about 12 hours, and should not vary too much with the fiber flexure induced by the positioner moves and the change of telescope pointing. A quantitative criterion for the PSF stability has to be defined as there are several scalar numbers one could use: full width half maximum, second moments, flux measurement biases (for a continuum spectrum or emission lines), line fit bias. The most stringent scientific driver is the sky subtraction accuracy because it affects the false identification rate of emission line galaxies, which is a level 2 requirement of the project. This sky subtraction bias is affected by the relative error on the PSF shape from fiber to fiber, and not by a global systematic PSF error. Indeed any average PSF error which is shared among all fibers is compensated by a change of the average sky spectral shape and intensity during the sky subtraction procedure (see \S\ref{sec:skysubtraction}). In the following we will present the stability based on the most stringent criterion which is the emission line flux bias induced by a change of PSF shape. The effect on this multiplicative bias is maximum when the CCD pixel noise is dominated by readout noise and not Poisson noise. In this limit, assuming a perfect centering of the PSF, the same readout noise for all pixels, and identical pixel quantum efficiency and pixel size, this bias is \begin{equation} \delta F / F = \frac{ \sum_{i} P_i P^{truth}_i}{ \sum_j P_j^2 } - 1 \label{eq:flux-bias} \end{equation} where the sum on the numerator and denominator run on pixels, $P_i$ is the PSF model integrated in a pixel and $P^{truth}_i$ is the same quantity for the true PSF profile. In this formula it is assumed that both PSF profiles (model and truth) are normalized by their integral. We will also report the more standard variation of second moments. In the following we discuss the PSF stability with time, telescope move and positioner moves. {\bf Stability with time}: Figure~\ref{fig:psf-stability} shows the PSF stability over several days for the three cameras of a spectrograph. Each dot on the figure is the PSF model for one fiber and one specific wavelength. The fiber numbers and wavelength were chosen to sample evenly the CCDs. The primary conclusion is that the PSF shape is remarkably stable. Looking at the most sensitive quantity which is the flux bias induced by the change of PSF ($\delta F/F$), we find a maximum variation of about 0.6\% within several hours. We note that this is an upper limit because part of this number is due to the statistical noise of the PSF parameters (which is reduced when averaging the result of several calibration exposures for each night). The variation from night to night is also small. It however exceeds the requirement of 1\% in the blue camera for this particular set of observations. One can note a significant drift of the PSF center along the cross-dispersion axis ($X$ on the figure), in other words from fiber to fiber, which physically corresponds to a vertical axis as the pseudo slit is standing vertically on the spectrograph optical bench. We explain our approach to correct for this effect in $\S$\ref{sec:wave-calib-trace-coords}. During these three days the temperatures recorded on the spectrograph camera body have varied by less than 0.1 degree C thanks to the temperature control while the humidity level, which is monitored but not controlled, has varied from 40\% to 28\%. The atmospheric pressure was stable with maximum variation of 3\,mbar (the maximum being on March 14). \begin{figure} \centering \includegraphics[width=0.99\columnwidth,angle=0]{psf-stability.pdf} \caption{PSF stability measurement. Each dot on the figures is the variation of a property of the PSF from exposure to exposure (i.e. current value minus average over exposures) for a given exposure, fiber and wavelength. The properties are, from top to bottom, the variation of the PSF center along $X$ (fiber direction) or $Y$ (wavelength direction), the relative change of PSF width along $X$ and $Y$, and finally the variation of emission line flux bias (Eq~\ref{eq:flux-bias}). The fibers and wavelength were chosen to cover evenly the active region of the CCD, excluding regions at short and long wavelength for each camera where the statistical noise is too large. Fibers 10, 100, 200, 300, 400, 490 are shown, and the wavelength are respectively in the range 3860-5650\,\AA, 5900-7500\,\AA, and 7700-9500\,\AA, for the blue, red and NIR cameras. The PSF model of the spectrograph SM10 are shown. The calibration exposures are from the nights of March 13, 14 and 15, 2020. Note that the fiber positioners were moved during the nights, so the variations from night to night are a test of the stability with positioner moves. Also, the dome was moved between the calibration observations of March 13, so the stability of the PSF during this sequence is a test of the stability with the change of telescope pointing (because the calibration screen the telescope is pointing to is attached to the dome).} \label{fig:psf-stability} \end{figure} {\bf Stability with telescope moves}: The PSF calibration run of March 13, which is the first night of the 3 shown on Fig.~\ref{fig:psf-stability} is composed of 4 sequences with different azimuthal angles of the telescope dome: 253, 359, 107 and 180 degrees. This is the best possible test of the PSF stability with changes of the telescope orientation, because the telescope which needs to point to the white screen was also moved to follow the dome rotation. This is testing the possibility that the flexure of the fiber cables along the equatorial mount and in the cable chains affect the PSF. The figures demonstrate this is not a concern; we do not see a variation of the PSF shape larger than 0.6\%. {\bf Stability with positioner moves}: Figure~\ref{fig:psf-stability} also shows to some extent the PSF stability with positioner moves, because the fibers were positioned to targets during the nights following each afternoon calibration sequence. We see that the PSF is mostly unaffected by positioner moves. This has been confirmed with another data set where the positioners from petal 0 were moved between series of arc lamps exposures with the telescope pointing to the dome screen. The spectra have been obtained with a boxcar extraction and then combined per fiber over the exposures of each series (with a median to reduce the effect of cosmic rays). We then measured the intensity of each emission line in each fiber and computed their variation from one series to the next. In order to reduce further the effect of residual cosmic rays, we finally considered the median of the ratio of emission line flux for each fiber. Among the fibers from positioners that were moved, the maximum variation in flux of the emission lines was found to be of $0.015$, and the mean standard deviation of $0.0035$. This is marginally larger than the dispersion obtained from non-moving positioners. Considering the quadradic difference for the moving and non-moving positioners, the excess scatter caused by the moves was found to be only of $0.002$ (0.2\%). We note that at the time of this tests, positioners were moved in a restricted range to avoid collisions. More recent data based on the relative variation of sky line intensities show a larger variation of 0.009. This latter dispersion however includes other effects (anisotropy of the sky background, wavelength calibration or flat field systematic errors). \subsection{Real time processing} DESI exposures are automatically processed and displayed using a quality assurance tool called Nightwatch\footnote{\url{https://github.com/desihub/nightwatch}}. Its purpose is to provide observing scientists with exposure metadata and instrument diagnostics in real time. Nightwatch handles data in three steps: (1) exposure processing; (2) calculation of quality assurance metrics; and (3) creation of web-based plots and tables to allow visual inspection of exposures. These steps are described in detail below. {\bf Exposure processing}: real-time processing of exposures begins with a preprocessing step that extracts the raw CCD images from the $b$, $r$, and $z$ cameras and masks out saturated pixels, pixels affected by cosmic rays, and in rare cases entire CCD amplifiers affected by high read noise. For exposures without spectroscopic data, such as bias and dark frames, processing ends at this point, while calibration and science exposures continue to spectral extraction. A quick processing routine applies a simple boxcar extraction \citep{deBoer:1981} to estimate the spectra from fiber columns in the CCDs (Fig.\,\ref{fig:ccd-image}). While boxcar extraction provides sub-optimal S/N compared to spectroperfectionism (Section\,\ref{sec:extraction}), its speed and robustness make it well-suited for real-time analysis. After spectral extraction, a simplified sky subtraction is applied to science exposures to produce sky-subtracted, flux-calibrated spectra for all fibers. {\bf Data quality assurance (QA)}: following preprocessing and spectral extraction, several data quality metrics are computed and stored: \begin{itemize} \setlength{\itemsep}{-0.5em} \item CCD metrics: read noise, overscan bias level, and detected cosmic rays per minute are estimated independently for the four amplifiers in every CCD. \item Camera metrics: boxcar extraction spectral trace shifts in $x$ (fiber number) and $y$ (wavelength) are estimated independently for the three cameras in each spectrograph. \item Fiber/spectra metrics: integrated and median flux per fiber (raw and calibrated), as well as SNR (raw and calibrated). \end{itemize} \begin{figure} \centering \includegraphics[width=\columnwidth]{nightwatch_exptable.png} \caption{The random-access exposure table provided by the Nightwatch QA tool for one night of observing.} \label{fig:nwexp} \end{figure} \begin{figure} \centering \includegraphics[width=\columnwidth]{nightwatch_ccd.png} \caption{CCD images from the ten $b$ cameras during a dark-time exposure. An amplifier in the $b0$ camera has been masked out due to high overscan bias (yellow square).} \label{fig:nwccd} \end{figure} \begin{figure} \centering \includegraphics[width=\columnwidth]{nightwatch_focalplane.png} \caption{Focal plane SNR plots for fibers in the $b$, $r$, and $z$ cameras during a 1246~s dark-time exposure.} \label{fig:nwfp} \end{figure} {\bf QA Web interface}: after calculation of QA metrics, plots and tables are automatically generated and posted on websites hosted at NERSC and Kitt Peak. The Nightwatch main webpage provides a calendar that allows random access to every night in DESI commissioning, survey validation, and the main survey. For each night, a table of exposures lists exposure metadata: exposure ID, type, date, number of spectrographs in operation, and QA status (see Fig.\,\ref{fig:nwexp}). For each exposure, users can access preprocessed CCD images with pixel masks highlighted (Fig.\,\ref{fig:nwccd}), plots of the QA metrics, and raw and calibrated spectra. Additional focal plane metrics showing fibers on target, fiber SNR (Fig.\,\ref{fig:nwfp}), and the accuracy of fiber positioner moves are computed on the fly and plotted. The interface also allows interactive plotting of sky-subtracted, flux-calibrated spectra recorded in individual fibers. The web interface provides the first look at reduced data from DESI, and is the main tool used for human validation of exposures as they are recorded. During DESI operations, observing scientists inspect the QA plots, diagnose problems, and take appropriate action. For real-time operations, the QA is particularly useful for recognizing problems with the CCDs, identifying broken calibration lamps or stray light in the dome, and noticing systematic issues with positioners. \subsection{Classification and Redshift fitting} \label{sec:redshift} Spectral classifications and redshifts are measured using Redrock\footnote{\url{https://github.com/desihub/redrock}}, which performs a $\chi^2$ vs. redshift scan for a suite of templates, selecting the best classification (galaxy, quasar, star) and redshift from the solution with the best fit to the data. Details of the algorithm are provided in \cite{bailey_redrock} and summarized as follows. Redrock includes principal component analysis (PCA) templates for galaxies, quasars, and stars. The galaxy templates are generated from stellar population synthesis and emission-line modeling of galaxies at $0<z<1.5$. These galaxy templates utilized a grid of theoretical, high-resolution simple stellar population models from C. Conroy (2014, private communication; see also \citealt{Conroy2018}) which are described in more detail in~\cite{bailey_redrock}. The training sample consists of 20,000 realizations generated by the desisim package\footnote{\url{https://github.com/desihub/desisim}} covering restframe wavelengths 1602\,\AA\ to 11000\,\AA\ on a 0.1\,\AA\ grid and following the expected redshift distribution and target selection color cuts for DESI Emission Line Galaxies (10k realizations), Luminous Red Galaxies (5k), and Bright Galaxy Survey (5k). These 20k high resolution templates are processed into 10 principal component eigenvectors. The Weighted Expectation Maximization Principal Component Analysis code\footnote{\url{https://github.com/sbailey/empca}} was used for this purpose (see \citealt{Bailey2012}). Stellar templates are generated from theoretical spectral models of stars and white dwarfs, split by effective temperature $T_{\mathrm{eff}}$ to generate 5 PCA eigenvectors for each of stellar subtypes B,A,F,G,K,M, and White Dwarfs. Cataclysmic Variables (CV) templates of 3 PCA eigenvectors are generated from archetype CV spectra used by SDSS {\it idlspec2d}~\citep{Bolton2012}. QSO templates with 4 PCA eigenvectors are currently the same as those used by SDSS {\it idlspec2d}. For each target at each candidate redshift, the templates are resampled to the wavelength grid of the observed data and multiplied by the resolution matrix described in \S\ref{sec:extraction}, thus providing a template matched to the resolution of each spectrum without resampling the data. Galaxies scan redshifts $z$ on a logarithmic spacing from $0.005<z<1.7$; Quasars scan a logarithmic redshift grid from $0.05<z<6.0$, and stars are sampled on a linear redshift grid from $-0.002<z<0.002$ (i.e.~$\pm$600\,km\,s$^{-1}$). $\chi^2(z)$ is measured on a finer redshift spacing around the lowest three $\chi^2(z)$ minima for each template class. Redrock fits a parabola to these minima to derive the final $\chi^2$ and $z$ of each minimum. The template with the lowest $\chi^2$ is selected as the best, with warning flags indicating if this is within a $\Delta\chi^2<25$ of the next-best solution, or other potential problems such as being at the edge of the scanned redshift range or failed parabola fits. In early testing, a common failure mode of the galaxy PCA templates was fitting sky mis-subtraction or noise features with a negative [\ion{O}{2}]~$\lambda \lambda 3726,3729$ doublet. To limit this failure mode, if the model [\ion{O}{2}] flux is negative for the galaxy template fit at a particular redshift, a $\chi^2$ penalty is added proportional to the amount of negative flux to deweight the likely incorrect solution at this redshift. Although Redrock performs well on galaxies and stars, it misses $\sim$10-15\% of true quasars, in particular redder low redshift quasars that are not well represented by the training sample used to create the original SDSS QSO PCA templates. To recover these, the spectroscopic pipeline also runs QuasarNET \citep{Busca2018} and an \ion{Mg}{2}~$\lambda \lambda 2796,2803$ line fit and uses the combination of the three algorithms to make the final QSO selection as described in \cite{chaussidon22a}. Details of the performance of the pipeline and survey in terms of redshift success rate and purity are provided in a series of papers reporting on the Survey Validation (hereafter SV, see \citealt{sv22a} for an overview). \cite{hahn22a} present the Bright Galaxy Survey, for which a redshift efficiency of 98.5\% has been obtained for the faint sample with an effective exposure time of 180~sec (see \S\ref{sec:efftime} for a definition of the effective time), with a purity of 99.5\%, with even better results for the bright sample. A comparison with the GAMA survey (DR4, see \citealt{driver2022}) is also provided with 99.7\% of redshifts in common with $| \Delta z | <0.001 (1+z)$. \cite{zhou22a} present the Luminous Red Galaxy (LRG) sample. With a nominal effective exposure time of 1000~sec, and with quality selection cuts $\Delta \chi^2 > 15$ and $z<1.5$, they obtain a success rate of 98.9\% and a purity of 99.8\%. \cite{raichoor22a} study the Emission Line Galaxy (ELG) sample; they find that with a selection cut involving a combination of $\Delta \chi^2$ and [\ion{O}{2}] signal to noise ratio, they obtain a purity higher than 99.4\%. The ELG redshift success is however lower than for the other target classes because a significant fraction of the galaxy targets do not have bright enough emission lines or are not in the redshift range of interest. The QSO sample is studied in~\cite{chaussidon22a}. They find that the Redrock redshift code has to be complemented with a \ion{Mg}{2} line finder and a neural network classifier (QuasarNET, see~\citealt{Busca2018}) to increase the identification efficiency and avoid line confusion. A combination of the three methods give a redshift efficiency of $98.0 \pm 0.4$\% and a purity of $99.5 \pm 0.4$\%. Those figures have been derived from comparisons with truth tables obtained from deep observations and a large effort of visual inspection conducted by the collaboration \citep{lan22a,alexander22a}. Despite those good results, the spectral classification and redshift estimation could be further improved. Redrock currently uses PCA templates to fit the data; although this provides speed and flexibility, there is no constraint that the ``best fit'' answer is physically meaningful other than penalizing negative [\ion{O}{2}] flux. This makes Redrock overly sensitive to false fits, especially when there are data problems upstream, e.g.~errors in the background subtraction or flux calibration between the spectrograph arms leading to steps in the spectra. Future work may introduce updated templates based on deep DESI observations, archetypes~\citep{Cool2013,Hutchinson2016}, non-negative matrix factorization (NMF) templates~\citep{Blanton2007,Zhu2016}, or other priors focused on ensuring physically meaningful fits. Also, the current Redrock galaxy and stellar templates were generated for DESI, but the QSO templates are still the same as those used by BOSS, trained on just a few hundred QSO spectra. Although there are many more QSO spectra available now that could be used to make improved QSO templates, initial work has found that improving the QSO performance comes at the cost of degrading the completeness of galaxy targets. Future releases may have new QSO templates tuned to improve QSO performance while preserving the performance of galaxy templates. \section{Resampling the Resolution Matrix to alternate model wavelengths} \label{sec:resampleresolution} The spectral extraction examples in Section~\ref{sec:extraction} used an input model wavelength grid that was exactly aligned with the output extraction grid with a 0.8\,\AA\ wavelength step, thus simplifying the process of convolving the input model with the square resolution matrix $R$, which is tied to the extraction wavelength grid. In practice, input models may be on a higher resolution or differently aligned wavelenth grid than the output grid (default 0.8\,\AA\ for DESI). Resampling the model to the extraction grid before applying the resolution is incorrect; applying the resolution should happen before any model wavelength rebinning. It is worth noting that this issue is not unique to the spectro-perfectionism technique; it applies to any astronomical spectroscopy with unresolved components. \begin{figure} \includegraphics[width=0.55\columnwidth,angle=0]{exbias-offset-input.pdf} \centering \caption{Extraction model biases from an unresolved input model that is unaligned with the output spectral extraction wavelength grid. Top panel: input model, with the left-most input emission line aligned with the output wavelength grid, and other emission lines with offsets. Middle panel: input model resampled to output wavelength grid before multiplying by the resolution matrix $R$ (red) and 100 extractions with different noise realizations (black). Bottom panel: mean difference between extractions and resampled resolution convolved model. } \label{fig:exbias-offset-input} \end{figure} If the input model has features that are unresolved by the extraction wavelength grid, rebinning before resolution can lead to large data vs.~model biases as shown in Figure~\ref{fig:exbias-offset-input}. The leftmost input emission line is exactly aligned with the output wavelength grid, leading to no bias in the extractions. The other 3 lines have offsets of 33\%, 50\%, and 67\% of the 0.8\,\AA\ extraction wavelength step size. The extracted spectra have significant differences when compared to the (incorrectly) rebinned then resolution convolved model. \begin{figure} \includegraphics[width=0.55\columnwidth,angle=0]{exbias-resolved-input.pdf} \centering \caption{Like Figure~\ref{fig:exbias-offset-input}, except that the input model emission lines have a 0.8\,\AA\ Gaussian $\sigma$ dispersion and thus are nominally resolved by the extraction wavelength grid. Despite the offset centroids, the bias is much smaller than the unresolved case.} \label{fig:exbias-resolved-input} \end{figure} On the other hand, if the input model is resolved by the extraction wavelength grid even before considering instrumental resolution, these discrepancies are much smaller, as shown in Figure~\ref{fig:exbias-resolved-input}. In this case, the input model has a Gaussian $\sigma$ of 0.8\,\AA\ (one extraction pixel), equivalent to a velocity dispersion of 16\,km\,s$^{-1}$ for a $z=1$ object. Even though the centroids of the emission lines are still offset relative to the extraction wavelength grid, the data vs.~model differences are much smaller. For simplicity and algorithmic efficiency, the Redrock algorithm for classification and redshift estimation (Section~\ref{sec:redshift}) takes advantage of this and rebins its redshifted template spectra to the output wavelength grid prior to convolving with $R$ for comparing to data. Since the internal velocity dispersion of main survey quasar and galaxy targets are resolved by the 0.8\,\AA\ wavelength grid, the resulting biases are minimal when comparing the model to the data. \begin{figure} \includegraphics[width=0.55\columnwidth,angle=0]{exbias-resample-R.pdf} \centering \caption{Like Figure~\ref{fig:exbias-offset-input}, except that the resolution matrix $R$ columns are resampled to match the input wavelength grid, thus performing the resolution convolution and the rebinning in a single step, significantly reducing the extraction bias.} \label{fig:exbias-resample-R} \end{figure} However, if one needs to work with an input model that is unresolved by the default 0.8\,\AA\ wavelength grid (e.g.~stellar models with sharp absorption features), one needs to resample the resolution matrix to apply the resolution prior to rebinning to the output wavelength grid. i.e.~in the original definition of $R$ (equation~\ref{eq:resolution-matrix}), it is a square matrix that applies the resolution to an input model on the same wavelength grid as the output spectra, e.g.~in the model-to-data comparison $\chi^2_\mathrm{1D}$ in equation~\ref{eq:extract-chi2-1D}. To accomodate a higher resolution model, one can use a cubic spline to interpolate each row of $R$ into a higher resolution wavelength grid matching the input model, resulting in a non-square resolution matrix $R^\prime$ that takes a high-resolution input model $m$, applies the resolution, and rebins to the output wavelength grid. Figure~\ref{fig:exbias-resample-R} shows the improvement from using this resampled resolution matrix $R^\prime$. Although there are still some artifacts, it is much improved compared to Figure~\ref{fig:exbias-offset-input} which rebinned the model prior to applying the resolution. The DESI spectroscopic pipeline algorithms do not use this $R$ resampling procedure as part of the data processing, but we document this procedure here as an issue to consider for analyses comparing models to DESI spectra. \subsection{Sky subtraction} \label{sec:skysubtraction} The sky subtraction algorithm takes advantage of the spectral extraction method, which provides fiber spectra on the same wavelength grid, with uncorrelated noise (within a spectrum), along with a resolution matrix that provides the transform to apply from a high resolution spectral model to a lower resolution fiber spectrum. The sky subtraction is performed in two steps. First a sky spectrum model is fit to the sky fibers spectra for each camera independently and then this model is subtracted to each fiber of the camera. \subsubsection{Sky model fit} \label{sec:skymodelfit} We fit for a deconvolved sky spectrum by minimizing the following. \begin{equation} \chi^2 = \sum_{f} \sum_{i} w_{f,i} \left( D_{f,i} - T_{f,i} \sum_j R_{f,i,j} P(x_f,y_f,\lambda_j) S_j \right)^2 \label{eq:skymodel} \end{equation} where $f$ is a sky fiber number, $i$ the wavelength index, $w_{f,i}$ the inverse variance, $D_{f,i}$ the data, $T_{f,i}$ the fiber flat field correction (see~\S\ref{sec:fiberflat}), $R_{f,i,j}$ an element of the resolution matrix defined in~\S\ref{sec:extraction} for the fiber $f$, $P$ an optional polynomial correction and $S_j$ a flux value of the average sky model spectrum. Because of the resolution term, $S$ is a deconvolved spectrum. The polynomial correction is optional, it is intended to model angular variations of the sky spectrum across the focal plane. We can consider either an isotropic sky spectrum (in which case $P = 1$), an anisotropic sky intensity with a same spectral shape, or an anisotropic and color-varying sky spectrum. In that later case, the most complex, the polynomial is 3D, it is a function of the fiber coordinates in the focal plane (i.e. angular sky coordinates) and wavelength. We have not considered in the current version of the code the possible relative variations of the OH emission line intensities with respect to the sky continuum. We are presently developing corrections for variations across the field of view due to diffuse moonlight and the twilight sky. The fit is performed iteratively. Each iteration starts with a fit of the mean sky spectrum $S$ given the polynomial $P$ (simple linear system), then optionally $P$ given $S$, and then each iteration ends with an outlier rejection which aims at discarding flux bins that are affected by cosmic rays (and have not been flagged at the extraction level). The loop ends when the $\chi^2$ decrement between iterations falls below a given threshold and no more outliers are found. After the last iteration the covariance matrix of $S$ is computed (inverse of half of the Hessian of the $\chi^2$), assuming constant $P$, which is an approximation. We then compute the sky model for all fibers of a camera, by reconvolving the model using the resolution matrix of each fiber, and possibly applying the polynomial correction. We also save the sky model covariance assuming the average resolution and transmission of all fibers. \subsubsection{Normalization of the sky background} \label{sec:skynormalization} The sky model determined from the sky fibers is then subtracted to each flat-fielded target spectrum of the same camera. In this process, an achromatic scale factor is adjusted for each fiber and camera. This scale factor can absorb various calibration errors like a flux extraction bias due to a variation of PSF shape or more likely an imperfect centering of fiber traces, or an error in the flat field corrections like the percent variations revealed by the twilight flats (see \S\ref{sec:fiberflat}). It can also correct for a genuine variation of fiber throughput between the calibration exposures and the current one due to either fiber flexure, which changes the fiber focal ratio degradation\footnote{The Focal Ratio Degradation (FRD) characterizes the broadening of the light beam angular distribution between the fiber input and output. A larger FRD leads to a reduced fiber throughput because the light loss in the spectrograph}, or a change of vignetting with the fiber position for the fibers on the edge of the focal plane. This scale factor is estimated as follows. We first fit for the amplitude of each sky line in a predefined list along with a local background. The fit of the amplitude and background is performed simultaneously in a narrow spectral range of 9\,\AA\ centered on the line in order to account for the unknown target flux density. We use a least square minimization accounting for the flux variance and the known sky spectrum line profile from the model. We then combined the amplitudes of all the lines into a single scale factor, with inverse variance weights, and while rejecting 3 sigma outliers. Figure~\ref{fig:sky-throughput-correction} shows that the typical correction is of order 1\%. \begin{figure} \centering \includegraphics[width=0.99\columnwidth,angle=0]{sky-throughput-correction.pdf} \caption{Histogram of the multiplicative correction term applied in the sky subtraction per camera for a 900 second sky exposure. There is one entry per fiber. Measurements with statistical uncertainties ($\sigma$ in the legend) larger than 0.03 were discarded.} \label{fig:sky-throughput-correction} \end{figure} \subsubsection{Corrections with a principal component analysis} The model described above was successfully tested on simulations, however systematic sky residuals on bright sky lines were found when applied to real DESI data, in particular for the bright NIR lines at 9000\,\AA\ and above. The residuals can be described in first approximation as wavelength calibration errors of about 0.02\,\AA\ (about 3\% of the size of a CCD pixel), and errors on the Line Spread Function (hereafter LSF, which is the 1D spectral point spread function) width of a few 0.1\,\AA. Those errors are quite stable between exposures, but vary from fiber to fiber, with a correlation length of about 10-20 fibers (not fully correlated with the fiber bundles of 25 fibers). We have not understood the source of those systematic residuals. In particular, we could not relate them to PSF fit residuals. The working hypothesis is that they are the result of differences in the illumination pattern between the white spot in the dome and the night sky, causing a different distribution of light in the spectrographs and hence different PSFs. Our approach to correct those effects is purely empirical. For each exposure, each sky fiber and each bright sky line, we fit for 3 parameters: an amplitude correction, a wavelength calibration offset ($\Delta \lambda$), and a LSF width correction ($\Delta LSF$). The amplitude correction is ignored but the $\Delta \lambda$ and $\Delta LSF$ arrays are recorded. In a first calibration step, we computed those arrays for a large sample of sky observations conducted in 2020. For each camera/exposure, the $\Delta \lambda$ and $\Delta LSF$ values obtained for a finite set of sky lines in the sky fibers are linearly interpolated (or extrapolated) to the full 2D array that comprise the 500 fibers and the wavelength grid of the spectral extraction. The resulting 2D frames are then used to perform a principal component analysis (hereafter PCA, see also~\citealt{WildHewett2005} for another PCA approach to sky subtraction). The resulting few first components (also 2D frames) are saved as part of the calibration product for each of the 30 cameras of DESI spectrographs. Then, for each science exposure, the limited set of $\Delta \lambda$ and $\Delta LSF$ on the sky fibers are used to compute linear coefficients for the saved principal components, and the resulting linear combination is then applied to all of the fibers and wavelength from the frame. Presented in the form of an equation, the complete sky model for the flat-fielded spectrum in fiber $f$ for the wavelength index $i$ is : \begin{eqnarray} S_{f,i} &=& \alpha_f \sum_j R_{f,i,j} P(x_f,y_f,\lambda_j) S_j \nonumber \\ && + \sum_k X^{(\lambda)}_k (\partial_{\lambda} \tilde{S})_i C^{(\lambda)}_{k,f,i} \nonumber \\ && + \sum_p X^{(LSF)}_p (\partial_{LSF} \tilde{S})_i C^{(LSF)}_{p,f,i} \label{eq:fullskymodel} \end{eqnarray} Here, $\alpha_f$ is the scale factor described in \S\ref{sec:skynormalization}. $R_{f,i,j}$, $P(x_f,y_f,\lambda_j)$, and $S_j$ are the terms contributing to the initial sky model described in \S\ref{sec:skymodelfit}. $X^{(\lambda)}_k$ and $C^{(\lambda)}_{k,f,i}$ are respectively the PCA coefficients and components for the wavelength correction, and $\partial_{\lambda} \tilde{S}$ is the derivative of the reconvolved sky model ($\tilde{S} \equiv R S$) with respect to the wavelength $\lambda$. The terms in the last row are the equivalent for the LSF width correction. \subsubsection{Estimating the variance of the sky subtracted spectra} \label{sec:sky-residual-variance} The sky model covariance (hereafter $C_S$) is the inverse of the Fisher matrix (half of the Hessian of the $\chi^2$ given by Eq.~\ref{eq:skymodel}). It is not diagonal, the neighboring flux values in the vector $S$ have large anti-correlations as a result of the deconvolution process (the sky model $S$ is first multiplied by the resolution matrix $R$ before comparison with data). Conversely, the flux values of the sky model convolved with the resolution of a fiber ($\tilde{S_f} \equiv R_f S$) are uncorrelated to a good approximation. In consequence, rather than keeping the full covariance $C_S$, and evaluating the reconvolved model variance for each fiber, we estimate and save the diagonal of the covariance of the convolved sky model for the average resolution matrix $\bar{R}$ (average of all the valid fibers of the camera). We have the covariance $C_{\tilde{S}} \equiv \bar{R} C_S \bar{R}^T$, and we note $\sigma^2_{\tilde{S}}$ the diagonal terms of this covariance matrix in the following. Figure~\ref{fig:skymodel} shows a few model sky lines before and after reconvolution ($S$ and $\tilde{S}$). Also shown are the correlation coefficients between neighboring fluxes in the same spectra. The correlation between neighboring flux bins in the reconvolved spectrum does not exceed a few percent. \begin{figure} \centering \includegraphics[width=0.95\columnwidth,angle=0]{sky-convolved-deconvolved.pdf} \includegraphics[width=\columnwidth,angle=0]{sky-spectrum-correlation.pdf} \caption{Top: zoom on a few NIR lines of the sky model spectrum derived from a 900 second science exposure with 80 sky fibers. The blue curve is the deconvolved spectrum ($S$) and the orange the reconvolved spectrum $\tilde{S} = R S$. Bottom: correlation coefficients ($ \equiv C_{i,i+\Delta} / \sqrt{ C_{i,i} C_{i+\Delta,i+\Delta}}$ with $C$ the covariance matrix) of neighboring flux values in the same spectrum. The neighboring fluxes ($\Delta=1$) on the deconvolved spectrum are highly anti-correlated with a coefficient of about -0.8, whereas the correlations in the reconvolved spectrum never exceed a few percent.} \label{fig:skymodel} \end{figure} The variance of the target spectra after sky subtraction is modeled as \begin{equation} \sigma^2_{\mathrm{corr},f} = \sigma^2_{\mathrm{extract},f} \, T_f^{-2} + \sigma^2_{\tilde{S}} + \left( \sigma_{T} \, \tilde{S}_f \right)^2 + \left( \epsilon_{\lambda} \, \partial_{\lambda} \tilde{S}_f \right)^2 \label{eq:spectralvariance} \end{equation} where $\sigma^2_{\mathrm{extract},f}$ is the variance from the spectral extraction for the fiber $f$ (the diagonal of $\tilde{C}$ defined in \S\ref{sec:extraction}), $T_f^{-1}$ is the fiber flat field correction (inverse of the relative fiber throughput, see Eq.~\ref{eq:fiberflat}), $\sigma^2_{\tilde{S}}$ the sky model error, $\tilde{S_f}$ the resolution convolved sky for the fiber $f$ and $\partial_{\lambda} \tilde{S}_f $ its derivative with respect to wavelength. $\sigma_{T}$ is the flat field error and $\epsilon_{\lambda}$ is a term used to quantify the wavelength calibration error. The term $\epsilon_{\lambda}$ is adjusted for each sky line independently, but considering all sky fibers at once. Its value is chosen to get a reduced $\chi^2/ndf = 1$ in a spectral range of $\pm 3$\,\AA\ around each sky line ($ndf$ is the number of degrees of freedom of the fit). It is evaluated only on the sky fibers but is then applied to all the target fibers. Figure~\ref{fig:sky-epsilon-lambda} shows the median value $\epsilon_{\lambda}$ as a function of wavelength for several science exposures in NIR. Typical values are of 0.025\,\AA. \begin{figure} \centering \includegraphics[width=0.95\columnwidth,angle=0]{sky-epsilon-lambda.pdf} \caption{Wavelength calibration error term $\epsilon_{\lambda}$ determined from the sky line fit (one value per sky line, averaged over cameras and exposures, for typical dark time exposures of about 1000 sec).} \label{fig:sky-epsilon-lambda} \end{figure} \subsubsection{Results} We evaluate here the performance on the sky subtraction. Figure~\ref{fig:skyres} shows a comparison of the r.m.s. (root mean square) of the spectral residuals in the sky fibers after sky subtraction with the expected noise given by Eq.~\ref{eq:spectralvariance} (with $\epsilon_\lambda=0$), for the NIR cameras. Spectral regions where the rms of residuals exceeds the expected noise are highlighted in orange and red colors, when the extra scatter exceeds 1\% and 3\% of the sky level. One can see on the figure that the realized noise is consistent with the expectation in the continuum. This is an important validation of the estimation of the noise at the CCD level and its propagation to the spectra. On bright sky lines however, one can see that the noise is underestimated. The excess noise is typically between 1 and 3\% of the sky level on the brightest lines. This is of course dependent on the sky brightness as the relative contribution from electronic noise, Poisson noise, and systematics in the PSF model depend on the signal level. The data we are presenting here are from science exposures of 1000 $\pm$ 10 sec. The exposure time, the number of sky fibers used and the observation conditions are typical of the ones found for the main dark time survey of DESI. \begin{figure} \centering \includegraphics[width=\columnwidth,angle=0]{skyres-z.pdf} \caption{Dispersion of the spectral residuals in sky fibers as a function of wavelength. The value at each wavelength is the RMS of the residuals in all the sky fibers from all spectrographs, and for several exposures of 1000 $\pm$ 10 sec. The gray band is the expected noise RMS from Eq.~\ref{eq:spectralvariance}, in the ideal case with $\epsilon_\lambda=0$. The colored curves are the measured RMS, the colors indicating when the variance exceeds the expected noise. The top panel is the full spectral range of the NIR camera, whereas the lower panels focus on some narrower wavelength ranges. There is naturally a larger rms on the lines because of the Poisson noise, but the measurements also exhibit deviations due to systematic errors in the flat field, wavelength calibration or PSF shape.} \label{fig:skyres} \end{figure} Figure~\ref{fig:skyres-extra-variance} shows the same dispersion now compared to a variance model where the extra term $\epsilon_\lambda$ has been adjusted. The measured scatter is now consistent with the noise model for most wavelength. This is an important result for the optimization of the redshift fitting. \begin{figure} \centering \includegraphics[width=\columnwidth,angle=0]{skyres-z-with-extra-variance.pdf} \caption{Same as Figure~\ref{fig:skyres} with a variance model including $\sigma_T$ and $\epsilon_\lambda$ terms (see Eq.~\ref{eq:spectralvariance}).} \label{fig:skyres-extra-variance} \end{figure} \section{Spectro-perfectionism and convolution} \label{sec:spectroperfconvolution} We discuss in the section how the transformation $\tilde f = R f$ introduced in \S\ref{sec:extraction} can be interpreted as a convolution {\it if} the pixel noise is constant (which is not the case with real data, because of the contribution from the Poisson noise of the signal) and {\it if} the PSF is also constant. \subsection*{Interpretation of the covariance matrix of $f$} With the first assumption of a constant noise, Eq.~\ref{eq:extract-f} simplifies to $ C^{-1} f = A^T p$ , with $C^{-1} = A^T A$. Each row or column of $C^{-1}$ is the auto-correlation function of the $PSF$, centered on the diagonal, and sampled on a regular grid given by the position of the flux points. We call in the following $\Delta$ the period of this grid, in units of CCD pixels. We have \begin{equation} C^{-1}_{i,j} \equiv c_{|i-j|} = \left[ PSF \star PSF \right](\Delta \times |i-j|) \label{eq:2.1} \end{equation} The matrix $C^{-1}$ is a finite order (and symmetric) Toeplitz matrix (same coefficients of each transverse line of the matrix). Finite order means that $c_i=0$ for $i>k$ (we assume that the PSF has a finite extent). When the dimension of the matrix becomes large, which corresponds to the extraction of a very large set of flux points, this matrix can be approximated as a circulant matrix, with differences only on the top right and bottom left corners (corresponding to the edges of the spectrum we want to extract)\footnote{This is presented rigorously in {\it Toeplitz and Circulant Matrices: A review} by R. M. Gray, \url{http://www-isl.stanford.edu/~gray/toeplitz.pdf}}. The eigenvectors and eigenvalues of a circulant matrix are related to the discrete Fourier transform. $$ C^{-1} = F_n^{-1} diag(F_n c) F_n \ \ \mathrm{with} \ \ F_{n,(j,k)} = e^{-2 \pi i (jk)/n} $$ In other words, the eigenvalues of $C^{-1}$ is the Fourier transform of one of the rows of $C^{-1}$ that we labeled $c$ in Eq.~\ref{eq:2.1}, which is the auto-correlation of the PSF sampled on a grid of period $\Delta$. We label this quantity $\mathcal{F}\left( \left[ PSF \star PSF \right]_{|\Delta}\right)$. It is related to the power spectrum of the original PSF, but with aliasing. The matrix operation $C^{-1} f$ is asymptotically a ``circular convolution'' process with a kernel $c$. Far from the edges of the spectrum, it is to a good approximation a simple convolution, $ c * f$. \subsection*{Interpretation of $R$ as a convolution kernel} Since $C^{-1}$ is asymptotically a circulant matrix, it is the same for its square root matrix $Q$ and the resolution matrix\footnote{The resolution matrix $R$ is equal to $Q$ normalized row by row, which in our present case has no effect, as all rows have the same sum.} $R$. The eigenvalues of $R$ are the square root of those of $C^{-1}$. In addition, the operation $\widetilde{f} = R f$ is a convolution, with a kernel that we will call $r$, which is one row of $R$, and the inverse Fourier transform of the eigenvalues of $R$. In consequence we have \begin{equation} \mathcal{F}(r) = \sqrt{ \mathcal{F}\left( \left[ PSF \star PSF \right]_{|\Delta}\right) } \label{eq:Fr} \end{equation} This identification of the $R$ matrix of spectro-perfectionism as a convolution is quite accurate in the center of the extracted spectra and only becomes invalid on the edges. \subsection*{Interpretation of the extraction} The vector $A^Tp$ is the cross-correlation of the PSF with the pixel values sampled at $\Delta$. We can write it $(PSF \star p)_{|\Delta}$. Similarly, we have the coefficients of one row of $C^{-1}$ as $c=(PSF \star PSF)_{|\Delta}$. The full spectro-perfectionism algorithm, in Fourier space, {\it omitting the $\mathcal{F}$ labels in the following}, is the convolution - deconvolution process \begin{eqnarray} \tilde f &=& r \frac{( PSF^* p )_{|\Delta}}{c} = r \frac{( PSF^* p )_{|\Delta}}{( PSF^* PSF )_{|\Delta}} = \frac{( PSF^* p )_{|\Delta}}{\sqrt{(PSF^* PSF )_{|\Delta}}} \label{eq:ftilde2} \end{eqnarray} We invite the interested reader to also see~\cite{Bolton2012b} for more discussions on this topic. \subsection{Stellar model fit} \label{sec:starfit} The flux calibration relies on a comparison of the measured and expected spectra of standard stars observed during each science exposure. About ten fibers per petal are allocated to standard stars for the main survey, but we benefited from a larger number of standard stars per pointing during commissioning. The standard stars are selected from the imaging catalogs based on their $g$, $r$ and $z$ AB magnitudes from MzLS+BASS surveys in the north, and the DECaLS survey in the south (see~\citealt{Dey2019} for a detailed description of the surveys and the effective band-passes). The color cuts designed to select main-sequence F stars are $0 < g-r < 0.35$ and $r-z < 0.2$. An additional magnitude cut was used to select a bright ($15<g<18$) and a faint ($16<g<19$) sample designed for short and long observations. Finally, lower metallicity halo stars are preferentially selected with a cut on parallax ($<1$\,milliarcsecond, hereafter mas), and peculiar stars with cool kinematics are avoided with a lower limit on the proper motion ($>2$\,mas\,yr$^{-1}$). Both astrometric quantities are extracted from the GAIA DR2 catalog~\citep{GAIADR2}. A comprehensive description of the selection of standard star targets can be found in Section 4.2 of~\citet{Myers22a} (in particular the selection of stars based on GAIA data only in some parts of the sky). A stellar model is fit to each observed standard star. This fit relies on a grid of stellar models. We describe the templates, the fitting technique, and present early results in the subsequent sections. \subsubsection{Stellar templates} We use a theoretical stellar models grid from \citet{AllendePrieto2018}, spanning a wide range of stellar effective temperature ($T_{\rm eff}$), surface gravity ($\log_{10} g$), and iron abundance ([Fe/H]). The spectral coverage of the models ranges from 0.13-6.5\,$\mu$m at a resolution of $R= 10,000$, and the dispersion is constant in $\log \lambda$ with a pixel size of 9.9\,km\,s$^{-1}$. Figure~\ref{fig:startemplategrid} illustrates the stellar template grid, focusing on the surface temperature range of the standard stars. The temperature resolution of the grid is 500\,K, $[Fe/H]$ are in steps of 0.5 dex, from -5 to +1, and the surface gravity in steps of 1 dex from 1 to 10$^5$\,cm\,s$^{-2}$. One can see that the $g-r$ color cut selects stars in the temperature range from 5500 to 8000\,K. \begin{figure} \centering \includegraphics[width=0.95\columnwidth,angle=0]{template-grid.pdf} \caption{Projections of the tri-dimensional stellar parameter grid as a function of effective temperature ($T_{\rm eff}$), surface gravity ($\log_{10} g$), and iron abundance ([Fe/H]). Models satisfying the color selection cuts for the DESI standard stars are marked as orange dots while the others are in blue.} \label{fig:startemplategrid} \end{figure} \subsubsection{Fit procedure} The stellar model fit is based on a comparison of the uncalibrated observed spectrum of a star with a model, after both the data and model have been divided by a smooth continuum, effectively circumventing the need for a prior knowledge on the flux calibration. The smooth continuum is obtained with a sliding median filtering of the original spectrum with a filter width of 160\,\AA. This width was chosen empirically to efficiently erase the variations of throughput while retaining most of the information on the stellar absorption lines. We perform the fit a second time after a first calibration of the data in order to avoid any systematic bias on the selection of stars due to the absence of original calibration. The input data comprises the observed spectra of a star from the three cameras (blue, red and NIR) and from one or several exposures when available. The spectroscopic pipeline variance from Eq.~\ref{eq:spectralvariance} is used, with the proper normalization from the smooth continuum, and with an additional variance term of 0.1$^2$ introduced to improve the relative weights in the fit while accounting for the potential modeling errors. Bright sky lines and atmospheric (telluric) absorption lines are masked (inverse variance set to zero), along with the Ca H\&K absorption lines which receive contributions from the interstellar medium\footnote{We did not mask the Na D-lines which also vary with the interstellar gas column density but may consider this potential improvement for a future data release.}. The stellar model is a function of five parameters, a normalization based on the $r$-band magnitude from the targeting catalog (set independently and not affecting the fit), a radial velocity, and the three stellar parameters from the templates, $T_{\rm eff}$, $\log_{10} g$, and [Fe/H]. The radial velocity is first fit by cross-correlating the data with a canonical spectral model at $T_{\rm eff}=6000$\,K, $\log_{10} g=4$, and [Fe/H]=-1.5. We have found that cross-correlating the data with the final best fit model does not significantly improve the velocity measurement. For this radial velocity fit, the canonical model is first convolved with the resolution from the data (assuming no radial velocity), then both the model and data are resampled on a fine logarithmic wavelength grid where a change of velocity becomes a simple translation of indices, and both are divided by a smooth continuum before being compared. The scan is performed in the velocity range of $\pm 1500$\,km\,s$^{-1}$ with a step of $15$\,km\,s$^{-1}$. The best fit and uncertainty is given by a parabolic fit around the $\chi^2$ minimum in the scan. Once the radial velocity is known, the model templates are blue or red-shifted, resampled to the wavelength grid of the data, then multiplied by the corresponding resolution matrix, and finally divided by their smooth continuum. Because this procedure is CPU intensive, we first select the possible models from the template grid that have a $g-r$ color consistent with the measured color of the star (from the targeting catalog), within a color range of $\pm 0.2$. Prior to this selection, we first correct the measured color for the Milky-Way dust reddening, using \cite{SFD98} map with corrections from \cite{SchlaflyFinkbeiner2011}, and \cite{Fitzpatrick99} dust extinction law. Also, for practical interpolation purposes, we extend the selection to the models in the smallest rectangular region of the 3D parameter space that includes all of the color selected models. The model for a given set of parameters $T_{\rm eff}$, $\log_{10} g$, and [Fe/H] is a trilinear interpolation of spectra from the 8 closest nodes in the 3D parameter grid. The fit is a least square minimization with inverse variance weight. We start from the best fit node in the grid and refine the parameters with a linear combination of its neighbors in the grid. Once the best fit linear coefficients are found, the best fit spectrum is recomputed from the templates (with red-shifting, but without the resampling, resolution convolution and continuum normalization), dust extinction is applied to the model and it is then normalized to get the same $r$-band magnitude as the measured one from the photometric catalog. An example fit is shown on Figure~\ref{fig:starspectrum}. \begin{figure} \centering \includegraphics[width=\columnwidth,angle=0]{star-spectrum.pdf} \caption{A typical standard star spectrum from an exposure of 900s (colored curves) along with the best fit model (black curve). The upper panel shows the full spectrum while the lower panels present zooms in various wavelength range. For these, the model has been renormalized to better fit the data and highlight the agreement on the absorption lines.} \label{fig:starspectrum} \end{figure} An objective criterion to validate the quality of the temperature fit is to compare the measured $g-r$ color of the stars from the targeting catalog with the $g-r$ color computed from the best fit spectral model. These colors are compared in Figure~\ref{fig:stdstar-color-color} for one of the science exposures (a 900 sec exposure). The agreement between the model and measured color is excellent, with an r.m.s. of 0.023, corresponding approximately to a precision of approximately 100\,K on the surface temperature of the stars. \begin{figure} \centering \includegraphics[width=0.95\columnwidth,angle=0]{stdstar-color-color.pdf} \caption{Comparison of the $g-r$ colors from the targeting catalog with the colors inferred from the stellar spectrum fit. Each dot is a standard star from a typical 900sec exposure.} \label{fig:stdstar-color-color} \end{figure} \section{Summary} \label{sec:summary} We have presented the DESI spectroscopic pipeline, with all the algorithms developed to convert raw CCD images into calibrated spectra with their spectroscopic identification and redshifts. The pre-processing, spectroscopic extraction, PSF and wavelength calibration, flat-fielding, sky subtraction, and flux calibration have been described in detail. Some notable performance results are the stability of the spectrographs' PSF and relative throughput from fiber to fiber which is better than 1\% for most of the wavelength range during a night. The wavelength calibration has been tested against other surveys and the systematic error expressed in terms of radial velocity was found to be at maximum of 3\,km\,s$^{-1}$ (smaller than 0.1\,\AA), with an excess scatter of 0.8\,km\,s$^{-1}$ with respect to the expected statistical uncertainty. We are able to predict the $g-r$ colors of standard stars with a precision of 0.02 (or 2\% in flux ratio). The accuracy of the flux calibration was tested with white dwarf spectra, and the flux systematic error was found to be about 2\% in most of the wavelength range, possibly increasing to 6\% for wavelength shorter than 3700\,\AA. The scatter between the spectroscopic flux and the broadband photometric flux from the Legacy Survey was found to be of 6\% for bright objects. We also obtain consistent fiber fluxes for extended objects. We have paid great attention to the error propagation, starting with a model of the pixel variance in the CCD images, then using the spectroperfectionism technique to provide uncorrelated spectral fluxes and their variance. The error propagation was validated with an inspection of the residuals in the sky fibers. The residual scatter in the spectral continuum was found consistent with the expected uncertainties, whereas an excess scatter of a few percent of the total flux was found on the bright sky lines. This extra scatter is included in the reported flux uncertainties. While it was not the main purpose of this paper, we have given a overview of the redshift fitting technique used for DESI (see~\citealt{bailey_redrock} for more details). The resolution matrix retrieved from the extraction provides a well defined transform to apply to high resolution models before comparing with the data, while the uncorrelated noise in the spectra makes it easy to perform a maximum likelihood fit. The performance of the spectroscopic pipeline and survey are studied in a series of Survey Validation papers \citep{sv22a,hahn22a,zhou22a,raichoor22a,chaussidon22a,cooper22a,lan22a,alexander22a}. They demonstrate that the requirements in terms of redshift efficiency and purity are met or exceeded for all target classes. Some new developments have been started or are envisioned to further improve the data processing. We intend to improve the cosmic ray detection which still misses some pixels, the sky subtraction to accommodate for sky gradients in the focal plane caused by the presence of the moon or during twilight, the flux calibration as we have identified repeatable percent variations of throughput with positioner moves, the wavelength calibration with a better cross-correlation with sky lines, and finally we will improve the flux calibration guided by the systematic trends presented in the paper. The spectral resolution matrix is a new concept in spectroscopic surveys, and some of its properties are not trivial (it is not a flux conserving convolution). As the DESI collaboration gains experience using these spectra for science analyses, we may learn better ways to model and use the spectral resolution, potentially leading to algorithmic and format updates. We also expect to improve the Redrock spectral classification and redshift estimation with better templates and algorithms. In addition to the changes that will improve the data quality, we have also embarked on an effort to speed-up the pipeline, porting some of our software to GPUs in order to take advantage of the new computer system at NERSC. Some of those improvements will be used for the processing of the first year data set of the DESI survey. The algorithms presented in this paper were used to produce the {\it Fuji} data set that will be made public as part of the upcoming Early Data Release (EDR) covering DESI Survey Validation data. \subsection{Wavelength calibration and trace coordinates} \label{sec:wave-calib-trace-coords} Trace coordinates and wavelength calibration are determined at three different phases of the spectroscopic data analysis. The first one occurs when a new spectrograph is calibrated with no prior measurement of its optical properties. This procedure has to be robust but does not need to be very precise, it is described in~\S\ref{sec:bootcalib}. The second one is on the contrary a precise fit of the trace and wavelength calibration which requires a first solution. It is performed at the same time as the PSF model fit and is based on a forward model of the emission line spots in the CCD images. The model and fit procedure are described in detail in section~\ref{sec:psf} and we do not come back to it in this section. The third and final phase consists in tuning the trace coordinates and wavelength calibration for each scientific exposure to compensate for fine changes in the optics. We present our approach to this important adjustment in~\S\ref{sec:adjustcalib}. In all cases, the output are the coefficients of Legendre polynomials as a function of wavelength, per fiber, providing the $X$ and $Y$ coordinates of the PSF center. As described in \S\ref{sec:ccdimage}, $X$ is indexing CCD columns, and increases with increasing fiber number, and $Y$ is indexing CCD rows, and increases with increasing wavelength. \subsubsection{Initial wavelength calibration and trace coordinates fit} \label{sec:bootcalib} This initialisation is performed for a new hardware setup. It has been used for three different configurations of the spectrographs: first during the spectrograph tests at the vendor in France, second for the functional verification tests at Kitt Peak with a test slit and a dedicated illumination setup in the spectrographs enclosure, and finally when the science slits were inserted in the spectrographs, with calibration exposures obtained with the telescope pointing to the dome screen. Trace coordinates are first fit on preprocessed images obtained with continuum lamp exposures. The continuum light source is either a Tungsten lamp or an array of LEDs. A Tungsten lamp was used for tests at the vendor and for functional verification, whereas LED lamps shining on the dome screen were preferred for the initial calibration of the spectrographs with the science slits. First a median cross-dispersion profile is extracted from a central band of 100 pixel rows in the CCD. Isolated peaks are found and the $X$ coordinates of their maximum is recorded. This defines the fiber array. Then, starting from the coordinates at the center, the barycenter of the cross-dispersion profile of each CCD row and fiber is determined, walking up and then down the CCD rows, starting from the center and using as initial guess the fit from the previous row. Those coordinates are fit with a Legendre polynomial for each fiber. A Gaussian cross-dispersion profile is also fit per fiber and CCD row. This fast fit consists in an iterative weighted second moment computation where the weight at each iteration is a Gaussian profile which $\sigma$ is given by the previous iteration (with a $\sqrt{2}$ scaling factor due to the Gaussian weight). The wavelength calibration makes use of preprocessed images obtained with arc lamps exposures taken shortly after the continuum exposures. Not all lamps are used simultaneously to avoid confusion between neighboring lines. For the blue cameras, Mercury-Argon and Cadmium lamps are used together. For the red and near infrared (hereafter NIR) cameras, only the Neon lamp is used (see Table~\ref{table:arc-lamp-lines-list} in Appendix~\ref{sec:arc-lamp-lines-list} for the list of emission lines used for the wavelength calibration). The spectral extraction is a profile fit based on the trace $X$ coordinate and Gaussian $\sigma$ derived with the continuum lamp data. The resulting spectra are then analysed to find the most significant peaks. We then use the following algorithm to match the detected peaks with the known emission lines from the lamps, which is a challenge because of a poor initial knowledge of the wavelength calibration and the relative strength of the emission lines. For each of the two arrays to match (the known emission line wavelengths and the list of $Y$ coordinates from the detected peaks), we record all possible combinations of ordered triplets. We then compute all possible pairs of triplets from one array to the other and derive for each pair the parameters of a second order polynomial transformation from the wavelength to the $Y$ coordinates. A four-dimensional histogram is then filled with, for each triplet pair, the first and second degree terms of the transformation, the indices the longest wavelength in one triplet and the largest $Y$ coordinate in the other. The bin with the highest number of counts gives us a pair of matching lines and a good first guess of the transformation between wavelength and CCD coordinates. Then, we proceed to find additional matching pairs of lines and detected peaks and subsequently refine the transformation. This procedure is performed independently for each fiber. A last step consists in comparing the solutions among fibers to identify the set of good matches, enforcing this match among all fibers and refitting the transformation. This procedure works well for the combination of lamps given above and for a specific list of lines per lamp. One also has to predefine the bins of the 4D histogram which requires some prior on the first and second degree of the polynomial converting wavelength to CCD coordinates. \subsubsection{Adjustment of wavelength calibration and trace coordinates per exposure} \label{sec:adjustcalib} Despite a fine control of the temperature in the spectrographs enclosure, we have noticed drifts of trace coordinates with time as shown in Figure~\ref{fig:dxdy} for the spectrograph SM2. The other spectrographs have similar variations. These trace-coordinate variations have been correlated with environmental parameters, and a large correlation was found between the $X$ coordinates for the blue and red cameras with the humidity in the spectrographs' room. It is suspected that this correlation is due to a humidity dependent tilt of the NIR dichroic, possibly due to water absorption by the sealant that bonds the dichroics in their cells. This would explain why no correlation is found with the NIR camera data (as the dichroic is transmitting NIR light) while a correlation is found for the two other cameras (reflection on the dichroic). This would also explain the relative amplitude between the blue and red cameras. \begin{figure} \centering \includegraphics[width=\columnwidth,angle=0]{dxdy.pdf} \caption{Variations of the spectral traces coordinates with time for the three cameras of spectrograph SM2. Variation along $X$ (direction of fiber number, left panels) for the blue cameras are the largest with peak to peak variations of $ \pm 0.7$ pixel and an r.m.s of 0.3 pixel. The variations are smaller in the red and NIR cameras (0.15 and 0.03 pixels rms for $X$, respectively). The rms variation along the $Y$ axis (direction of wavelength dispersion, right panels) are about 0.05 pixel rms for all cameras. This would represent a wavelength calibration error of about 0.03\,\AA\ if not corrected.} \label{fig:dxdy} \end{figure} While further investigation may allow the development of a predictive model for the variations of trace coordinates, we have resorted to adjust the trace coordinates and wavelength calibration on the data itself. For each exposure, prior to the 2D extraction, we estimate the $\delta X$ offset of each fiber trace, for values of $Y$ coordinates, using cross-dispersion profiles averaged over several $Y$ rows. The $\delta X$ values across the whole CCD are then fit with a low order 2D polynomial of $X$ and $Y$, while making sure to have an efficient outlier rejection (because of cosmic rays, broken fibers, spectral regions with low signal). For the wavelength calibration adjustment ($\delta Y$), we rely on a cross-correlation of spectra. We first extract the spectra of all fibers with a fast algorithm (simple row-by-row sum of pixels around each trace) and then cross-correlate the spectra of all fibers with the average fiber spectrum to detect relative shifts of wavelength calibration. This simple approach works because of the strong sky lines that are present in all spectra. We also perform this cross-correlation for several wavelength intervals which makes it easier to discard erroneous measurements due to bright target spectra or cosmic rays. As for the other axis, a low order polynomial is fit across the CCD. Finally, we rely on an external sky spectrum to get the final wavelength adjustment for science exposures. This external sky spectrum is derived from a set of high-resolution spectra from~\citet{Hanuschik2003}, convolved to the resolution of the DESI spectrographs. The spectra were obtained at the Paranal Observatory with UVES, ESO's echelle spectrograph at the 8-m UT2 telescope of the Very Large Telescope (VLT). The performance of this method can be evaluated by studying the spectral residuals after the sky subtraction (see~\S\ref{sec:skysubtraction}), on sky lines to test the wavelength calibration, and in the continuum to test the stability of the flux calibration from fiber to fiber which is a function of the accuracy of the trace coordinates $\delta X$. \subsubsection{Solar system barycentric velocity correction} The wavelength are corrected for the shift due the relative velocity of the Earth with respect to the solar system barycenter. We treat this correction in two steps. First, we apply the same correction to all fibers of an exposure, using the velocity shift calculated for the center of the field of view. Our approach is to apply the opposite correction to the wavelength array used for the spectral extraction (see \S\ref{sec:extraction}), and then simply correct the reported wavelength. This avoids any resampling of the extracted spectra and in consequence does not introduce any extra correlation between the flux values. \subsubsection{Radial velocity measurements } \begin{figure} \centering \includegraphics[width=\columnwidth,angle=0]{rv_offsets1.pdf} \caption{Radial velocity offsets of stars with respect to the measurements in the SoS catalog \citep{Tsantaki2022} computed separately for each arm of the instrument. The left panel shows the histogram of radial velocity offsets in DESI Survey Validation data. The B (blue), R (red) and Z (for NIR) median offsets are of 0.97, 2.13, and 2.88\,km\,s$^{-1}$ respectively, and the median absolute deviations are of 4.2, 3.7 and 3.5\,km\,s$^{-1}$. The right panel shows the radial velocity offsets on the y-axis, in km\,s$^{-1}$, as a function of the MJD of observation.} \label{fig:rv_offset} \end{figure} As the Milky Way Survey (hereafter MWS) of DESI will be observing millions of stars in the Galaxy, we are interested in extracting stellar radial velocities that are accurate down to the photon-noise limits and are not biased with respect to other surveys. The detailed comparison of the radial velocities with external surveys is presented in the MWS overview paper~\citep{cooper22a}, while here we focus on issues related to the wavelength calibration and spectral reduction performance. The radial velocities of stars in the MWS are determined by chi-square fitting of the spectra by interpolated templates \citep{Koposov2011} from the PHOENIX stellar library \citep{Husser2013} using the rvspecfit\footnote{\url{https://github.com/segasai/rvspecfit}} code \citep{Koposov2019rvspecfit}. The radial velocity uncertainties are determined from the posterior on radial velocity conditional on best-fit stellar atmospheric parameters values. The accuracy and precision of radial velocities derived from DESI spectra was assessed through several tests described below. We looked at repeated observations of stars observed multiple times as these could be used to determine whether the observed scatter between radial velocities match the formal radial velocity uncertainties computed based on the pixel noise. The analysis of large number of repeats done during the science verification shows that the radial velocity scatter between different exposures can be approximated as $\sigma_{RV} \approx \sqrt{ \sigma_{RV,0}^2 + (0.8\,\rm{km}\,\rm{s}^{-1}) ^2}$ where the $\sigma_{RV,0}$ is the formal radial velocity error determined by rvspecfit. This implies that there are radial velocity systematics that are different from exposure to exposure and are likely associated with the inaccuracies of the wavelength calibration. We note however that for most but the bluest stars most of the radial velocity information comes from the NIR arm. To study the radial velocity systematics in further detail we look at the radial velocity measurements of all the stars observed during the DESI Survey Validation (hereafter SV, see~\citealt{sv22a}) that are part of the survey of surveys (SoS) catalog \citep{Tsantaki2022}. This catalog is based on stellar radial velocity measurements from multiple surveys such as APOGEE, SDSS, LAMOST, Gaia-ESO and {\it Gaia} that have been brought together to the same radial velocity zero-point and thus can be used as radial velocity reference catalog. To understand the wavelength calibration for different arms of the instrument separately, we separately measured the radial velocity from the blue, red and NIR arm spectra of each star from the SoS observed by DESI in SV. We then analyzed the radial velocity offsets with respect to measurements in the SoS catalog. The distribution of the offsets for stars with DESI RV errors smaller than 3\,km\,s$^{-1}$ is shown on Figure~\ref{fig:rv_offset}. It shows that the offset distribution is not the same for different arms, with the mean offset of $\sim 3$\,km\,s$^{-1}$ for the NIR arm and smaller offset for the blue arms. The blue arms show also larger spread of offsets than the red arm. This can be explained by some time dependence of those offsets, shown in the right-hand panels of Figure~\ref{fig:rv_offset}. These residuals seem to be also correlated with the wavelength corrections described in \S\ref{sec:adjustcalib} and the sky brightness. These correlations suggest that the likely cause of the radial velocity offsets in the blue arm is the wavelength adjustment based on sky lines. The offsets in the red and the NIR arm do not show much correlation with time of observation and observational conditions thus their cause remains to be determined.

6c10d5da3075639e41ab462a62f1212a1eabd4a8

\section{Introduction} Liquid Crystal (LC) technologies are starting to be studied at mm-wave bands in order to develop tunable devices that work properly at those frequency ranges. Because of its birefringence, by applying a low-frequency electric field to a cavity containing nematic LC, its electromagnetic properties can be varied and therefore the device response changes \cite{fundLC}. This phenomenon has been widely used in optics to develop LC displays and other devices such as spatial phase modulators \cite{spm}, but its potential use at mm-wave frequencies has only started to flourish \cite{employing}. This varying behaviour is related to a continuous change on the electric permittivity, which can be leveraged to vary the resonant frequency or beam of an antenna \cite{lcantenna, lcantenna2, lcantenna3, toso}, to sweep the shift in phase shifters \cite{LCphaseshifter, fast, phaseshift, reconf} or to tune the different elements in a reflectarray antenna \cite{lcreflectarray, reflectarray2,reflectarray3,folded,svhum}, to name a few. Very recently, the use of nematic LC as the key phase-shifting element of the upcoming Reconfigurable Intelligent Surfaces (RIS), planar devices able to manipulate electromagnetic propagation, has been proposed \cite{lcIRS,metasurf,lcmetasurface}, as it is one of the few tunable technologies with moderate cost capable of keeping the pace of the high frequencies ($>$100GHz) expected in future network generations beyond 5G. Moreover, since the LC fills an entire cavity and its behaviour can be locally modified, a pixel-wise active element implementation is avoided. This, together with the fact that these manufacturing procedures are widely common in optics, especially when developing electrically large planar devices with thousands of cells, make of this technology a very attractive solution for developing RIS panels. Alternative solutions to switch a beam in a metasurface, such as mechanical steering \cite{mech} or unit cells based on varactor diodes \cite{varactor}, although being commercially available are either of much higher cost or exhibit frequency limitations, as Table \ref{tab:techs} shows. Nevertheless, the relatively large losses and the slow switching times between states are the main weaknesses of such LC-based devices, as detailed in Table \ref{tab:techs}. Even though LC manufacturers are starting to develop novel composites specifically designed to present low losses at microwave and mm-wave frequencies \cite{employing}, current mm-wave LC devices provide reconfigurability times in the order of the seconds. However, to be fully implantable in future ultra-reliable low-latency communication networks, improving these times until they are at least comparable to the channel coherence time is of utmost importance due to the stringent dynamic requirements of upcoming communication protocols. \markboth{This paper has been accepted at IEEE Transactions on Antennas and Propagation. DOI: 10.1109/TAP.2022.3209734}% {} A number of strategies have been reported in optics to improve transition times, such as the use of polymerizable compounds \cite{pnlc,pnlc2} or dual-frequency LC \cite{dflc,dflc2,dflc3}, especially for decay time. Another strategy to reduce transition times is to employ sophisticated excitation signals by leveraging LC dynamics \cite{overdrive, overdrive_optics,overdrivelow}. In order to understand and completely control its dynamic behaviour during a state transition, which will ultimately impact on the switching time, its accurate temporal modelling is essential. In spite of this, few works tackle the temporal aspects of LC between arbitrary states \cite{smallangle,correlations,templc,comprehensive}, and all of them focus in the optical regime and not in RF. Moreover, given the challenge of achieving a proper phase shift range in mm-wave bands, multi-resonant cells (i.e. including several resonant elements in a single-band cell) must be used \cite{lcreflectarray,gerardorefl}. This makes modelling much more complex \cite{3dmodel} since resonators create phase shifts that can not be modelled with a medium constant, and they are also used to locally bias the LC. In \cite{accurate} an accurate LC modelling at mm-wave frequencies was proposed for cells in which it is necessary to include resonant elements. Using that model, it was shown that inhomogeneity and anisotropy can be considered only in the longitudinal direction of the cell, and the minimum number of layers in which the media should be stratified to obtain a precise phase prediction was also found. However, that model only considers a static LC regime, that is, when enough time has passed after an external excitation so that molecules lie in a stationary state after rotating. In the nematic LC characteristic equation, this translates in neglecting time-dependent terms. Consequently, there is no previous research in accurately characterizing the dynamics of LC-based mm-wave devices to reduce switching times. In this paper, we accurately model for the first time the LC dynamics between transition states (Fréedericksz transition) in RF past the known approximations, in order to obtain a temporal design control capability of the LC, representing a contribution to the model beyond the one reported in \cite{accurate}. First, we solve the LC dynamics equation applied to multi-resonant cells, analyzing its convergence as a function of the number of layers. This is done to investigate whether is it possible to compress the LC inhomogeneous molecular orientation in a single layer of an effective permittivity tensor, computed with the average molecule tilt across the cavity, with the aim of enabling a much more efficient electromagnetic cavity evaluation. The validation of the proposed model is carried out experimentally through different state transitions. As another novelty of this work given its applicability, the proposed and validated modelling is used to introduce an efficient design procedure of the LC bias signal, similar to the overdrive technique used in optical panels, capable of diminishing the switching time (especially rise transitions) of planar multi-resonant cells for phase control by a factor of 100X. Finally, both the model and the overdrive excitation technique are validated by comparing simulation and experimental data, obtained from reflective multi-resonant cells fed by a plane wave (reflectarray, metasurface or RIS), although the model is extendable to transmissive structures (transmitarrays) or other planar structures capable of controlling other parameters than phase. In the paper, we discuss the reconfigurability time improvement as a function of the used biasing signals, showing that the proposed method mitigates one of the most important challenges this technology must overcome to be widely used. The model results are validated at the cell (pixel) level for two different frequencies (97 GHz and 102 GHz), which in turn facilitates a tool for the analysis and synthesis of control signals at an arbitrary frequency and per each cell of the whole antenna, given that in a complete surface a plane wave with a different incident angle will arrive to each pixel. Therefore, this tool allows to synthesize overdriving control signals without depending on experimental measurements for each cell and angle of incidence in the array. \markboth{Journal of XXX,~Vol.~XX, No.~XX, September~XXX}% {} The paper is organized as follows. Section II introduces the dynamic modelling of LC multi-resonant cells and discusses analytical solutions to motivate the use of overdriving to reduce transition times. In Section III, the dynamic LC differential equation is solved exactly by combining a finite elements method (COMSOL) and an full-wave electromagnetic analysis tool (CST), the phase convergence of the cell is analyzed and the model is validated. Then, in Section IV, the overdrive design tool is presented and experimentally tested, and Section V concludes. \begin{table}[h] \centering \vspace{-0.2cm} \caption{Beam switching technologies comparison} \begin{tabular}{|c|c|c|c|c|c|} \hline \textbf{Strategy} & \textbf{Speed} & \textbf{Freq. range} & \textbf{Cost} & \textbf{Efficiency} & \textbf{Energy} \\ \hline \textbf{Mechanical} & $s$ & Wide & High & Low & High \\ \hline \textbf{Varactor} & $\mu s$ & Limited & Medium & High & Medium\\ \hline \textbf{LC} & $s$ & Wide & Low & Low & Very low \\ \hline \end{tabular} \vspace{0.1cm} \label{tab:techs} \vspace{-0.4cm} \end{table} \section{Dynamic Modelling of LC resonant cells} \subsection{Permittivity Tensor Calculation} \label{sec:Permittivity Tensor Calculation} \begin{figure}[!htb] \centering \vspace{-0.2cm} \includegraphics[width=1\columnwidth]{images/cavity_and_cellv3.pdf} \vspace{-0.5cm} \caption{a) LC cavity with applied electric field b) LC cavity without excitation c) Layered view of the reflectarray unit cell d) Top-view of the reflectarray unit cell. Dimensions (mm): $D_1 = 0.171$, $D_2 = 0.096$, $D_3 = 0.042$, $L_{y1} = 0.707$, $L_{y2} = 0.748$, $L_{y3} = 0.792$, $L_{x1} = 0.2$, $L_{x2} = 0.211$, $L_{x3} = 0.2$, $P_{x} = 1.145$, $P_y = 1.093$, $h_Q = 0.55$, $h_{LC} = 0.075$.} \label{fig:cavity} \vspace{-0.2cm} \end{figure} LC-based mm-wave devices leverage the tuning capability of this material, typically used as a substrate, to become dynamically reconfigurable. A typical LC cavity is shown in Fig. \ref{fig:cavity}a and \ref{fig:cavity}b with and without external excitation respectively, where the $z=0$ and $z=d$ planes are assumed to be indefinite electrodes. In this case, the rod-like molecules only present inhomogeneity along $z$ when biased. In the nematic state, due to the small degree of positional order of nematic LCs, a low-frequency (AC) electric field applied across the cavity will rotate its molecules. This, together with the LC anisotropy (i.e. large orientational order), allows the permittivity to be varied, as shown in Equation (\ref{eq:eptensor}). \begin{equation} \label{eq:eptensor} \overline{\overline{\varepsilon_r}}(\overline{r},t) = \varepsilon_{r\perp} \overline{\overline{I}} + \Delta\varepsilon_r \overline{\overline{N}}(\overline{r},t), \end{equation} where $\overline{\overline{I}}$ is the 3x3 identity matrix and $\overline{\overline{N}}(\overline{r},t) = \hat{n}(\overline{r},t) \otimes \hat{n}(\overline{r},t)$, being $\hat{n}(\overline{r},t)$ the macroscopic vector that defines the local orientation of the LC molecules at a certain point and time. This way, if an external electric field changes $\overline{\overline{N}}$, the macroscopic permittivity will be tuned. Given that the substrate under use is both inhomogeneous and anisotropic, its dielectric permittivity has to be expressed as a tensor, $\overline{\overline{\varepsilon_r}}$. The dielectric anisotropy of the material is defined as $\Delta\varepsilon_r = \varepsilon_{r||}-\varepsilon_{r\perp}$, being $\varepsilon_{r||}$ and $\varepsilon_{r\perp}$ the parallel and perpendicular dielectric constants with respect to $\hat{n}$, which respectively relate the parallel and perpendicular components of the RF electric field to the electric displacement field. A concurrent problem in LC-based mm-wave designs is that the birefringence of LC cells has been typically underestimated and modelled with a scalar permittivity value, ranging $\varepsilon_{\perp}>\varepsilon_r>\varepsilon_{||}$, instead of using its tensor form. This oversimplifies the problem by transforming it into an isotropic scenario, as by working with only the scalar permittivity the results are geometry-dependent and do not generalize, being insufficient for an accurate modelling. \begin{figure*}[b] \vspace{-0.4cm} \centering \includegraphics[width=1\textwidth]{images/angle_up_combined_v6.pdf} \vspace{-0.8cm} \caption{Tilt angle dynamics in a $75\mu m$ thick cavity filled with GT3-23001 LC. a) Step transition from 0V to 15V b) Step transition from 0V to 150V c) Transition from 150V to 10V after 19ms.} \label{fig:transition_up} \vspace{-0.2cm} \end{figure*} Both in reflectarray antennas and RIS, this dielectric anisotropy is used to perform a pixel-wise phase shift across an impinging electromagnetic wave. By applying a pre-computed bias voltage to each cell, the reflection coefficient phase is locally modified, thus obtaining a desired global phase distribution at the output, which will dictate the direction of the reflected wave. As shown in Fig. \ref{fig:cavity}c and \ref{fig:cavity}d, a unit cell of such reflectarray antennas consists of: a set of electrodes (typically dipoles, connected to the same potential within the cell) printed below a superstrate (e.g. quartz), which will act as a top plate of the LC cavity; the LC layer itself; a bottom conductive plate; and a substrate to support the structure. When carefully designed, the role of the dipoles is to produce an appropriate resonance in the reflection coefficient in RF, which will ultimately create a phase shift depending on the biasing. However, besides their role in RF, the dipoles also have a function in AC, since they are typically used to polarize the LC with the low-frequency electric field. This makes the tensorial permittivity to actually be inhomogeneous also in the transverse directions ($x$ and $y$), although in \cite{accurate} it was shown in the static case how its effects in the reflected phase are negligible as compared to the inhomogeneity and anisotropy effects in the longitudinal direction. In this paper, it will be shown how these effects also have little impact when modelling LC dynamics. Notwithstanding, in \cite{anibal,3dmodel} a more rigorous LC modelling is introduced for complex structures in which these effects are not negligible. It is worth noting that in literature there exist three general strategies to perform 2D addressing of the cells, in order to apply the proper voltage (and therefore, to obtain the proper phase) to each antenna pixel: direct, active and passive addressing \cite{fundLC}. In the cases of active or passive addressing, the sequential row sweep implies the need of synthesizing voltage sequences which must be properly computed, requiring a tool to analyse and synthesize control signals, like the one introduced in this paper. In the direct addressing case, the proposed model and tool are especially useful to reduce transition times between states, enabling overdrive techniques. The complex task of synthesizing control sequences could alternatively be done through measurement data instead of simulations. However, as long as sufficiently accurate models are used, the latter greatly simplifies the process given that the LC dynamic effects depend on the cell position within the array (different incidence angle), the dimensions of the resonators in each cell (cells could have different dimensions in each pixel), the operating frequency and the LC properties. \vspace{-0.2cm} \subsection{Dynamic Director Calculation} Assuming that the applied AC electric field is homogeneous across the cell, which is feasible if the effects of transverse inhomogeneity are negligible, the dynamic behaviour of the LC under a such external excitation is described by the Ericksen–Leslie equation \cite{ericksen,leslie, fundLC}: \vspace{-0.2cm} \begin{multline} \label{eq:ericksenleslie1} (k_{11}cos^2\theta + k_{33}sin^2\theta) \frac{\partial^2\theta}{\partial z^2} + (k_{33}-k_{11}) \cdot sin\theta \cdot cos\theta(\frac{\partial\theta}{\partial z})^2 + \\(\alpha_2 sin^2\theta - \alpha_3cos^2\theta)\frac{\partial v}{\partial z} + \varepsilon_0 E^2 \Delta\varepsilon_q\cdot sin\theta \cdot cos\theta\\ = \gamma_1\frac{\partial\theta}{\partial t} + I\frac{\partial^2\theta}{\partial t^2}, \end{multline} where $k_{11}$ and $k_{33}$ are the splay and bend elastic constants, $\alpha_i$ is the Leslie viscosity coefficient, $v$ is the flow velocity, $E = V_q/d$ is the applied quasi-static electric field, $\Delta\varepsilon_q$ is the low-frequency dielectric anisotropy, $\theta$ is the tilt angle of the director, $\gamma_1$ is the rotational viscosity and $I$ represents the inertia. Two modifications can be safely made to the previous equation. First, the inertia term is typically neglected as it has a very small weight \cite{chand}. Second, the Leslie coefficients terms can be disregarded, as they also play a minor role and obtaining them requires experimental measurements \cite{lesliecoef,fundLC}. Their impact will be seen later on in relation to backflow effects. Therefore, the previous equation reduces to \cite{smallangle,pretilt}: \begin{multline} \label{eq:ericksenleslie2} (k_{11}cos^2\theta + k_{33}sin^2\theta) \frac{\partial^2\theta}{\partial z^2} + (k_{33}-k_{11}) \cdot \\ sin\theta \cdot cos\theta(\frac{\partial\theta}{\partial z})^2 + \varepsilon_0 E^2 \Delta\varepsilon_q\cdot sin\theta \cdot cos\theta = \gamma_1\frac{\partial\theta}{\partial t}, \end{multline} which allows the time-varying director to be found. In order to solve Equation (\ref{eq:ericksenleslie2}), boundary conditions for $\theta_{(z=0)}$ and $\theta_{(z=d)}$ must be applied. In the ideal case, $\theta_{(z=0)}=\theta_{(z=d)}=\theta_p=0$ (zero pre-tilt), which makes Equation (\ref{eq:ericksenleslie2}) stable from a certain threshold voltage, $V_{th}$. Then, the LC molecules will not start reorienting until this threshold voltage is reached. Specifically, $V_{th}$ can be computed as: \begin{equation} \label{eq:vth} V_{th} = \pi\sqrt{\frac{k_{11}}{\Delta\varepsilon_q}} \end{equation} In a real implementation of a LC cell, the inner-most surface of the two conductive plates contains a rubbed polyimide orienting layer, which forces the pre-tilt boundary condition $\theta_p\neq0$ at $z=0$ and $z=d$. This ensures that the molecules are correctly oriented also in absence of excitation, as can be seen in Fig. \ref{fig:cavity}b. Under these conditions, $V_{th}$ is not a strict value anymore, as molecules are able rotate even below this threshold. However, it can still be used as a reference to identify a voltage point in which the LC starts reacting more energetically, since below it the molecular orientation is weak. Equation (\ref{eq:ericksenleslie2}) entails the dynamic behaviour of the LC under an excitation, but since it is too complex to be solved analytically, certain assumptions are typically made in order to obtain approximate analytical solutions. Specifically, it is usually assumed that $k_{11} = k_{33}$, and that the LC is excited with a low voltage source. This allows to model the $\theta(z)$ curve with a sinusoidal function, so that $sin(\theta)\sim\theta$, which greatly simplifies the previous expression, resulting in Equation (\ref{eq:simplifiedLes}). \begin{equation} \label{eq:simplifiedLes} k_{11}\frac{\partial^2\theta}{\partial z^2} + \varepsilon_0 E^2 \Delta\varepsilon_q\theta = \gamma_1\frac{\partial\theta}{\partial t} \end{equation} Then, the solution for the tilt angle along $z$ takes, in time, an exponential form with the following decay and rise time constants: \begin{equation} \tau_d = \gamma_1 \frac{d^2}{k_{11}\pi^2}, \label{eq:decay} \end{equation} \begin{equation} \tau_r = \frac{\tau_d}{\left|\left(\frac{V}{V_{th}}\right)^2-1\right|}, \label{eq:rise} \end{equation} However, these approximations oversimplify the dynamics problem as (i) they assume a pretilt angle equal to zero, which is not realistic \cite{correlations}; (ii) they are only valid for small voltage excitation where the tilt can be approximated with a sine, which is a rough estimation; (iii) they assume that the driving voltage is little above $V_{th}$, while we will later show that voltages much greater than that are needed in order to accelerate the LC response. Although more elaborated expressions have been introduced in \cite{pretilt} to include pretilt effects, small angle approximations and single elastic constants ($k = k_{11} = k_{33}$ or $k = (k_{11}+k_{33})/2$) are still assumed. In \cite{accurate}, the static voltage dependence of LC is studied and accurately predicted but the dynamics are not tackled. In \cite{fast}, the LC temporal behaviour is experimentally measured for different commercially available materials in $4\mu$m thick cells, but a model is not provided. \begin{figure}[!htb] \centering \includegraphics[width=1\columnwidth, height=4.5cm]{images/GT323001_75um_relative_rise_time_error_of_analytical_approximation_vs_COMSOL_nicev2.pdf} \vspace{-0.7cm} \caption{10\%-90\% rise time approximation error compared to the solver solution in a $75\mu m$ thick cavity. $Ta$ refers to the approximated time from the closed expressions and $Ts$ refers to the simulated time from Eq. (\ref{eq:ericksenleslie2}). It can be observed how the closed expressions fail for large voltages, as the sinusoidal approximation becomes invalid, as well as the Vth effect.} \label{fig:approx_error} \vspace{-0.3cm} \end{figure} Moreover, in optics, the phase change that occurs from this tilt dynamics can be well modelled, and the optical intensity change can be well predicted, since for such a short wavelength the LC is simply a medium in which several $\pi$-times phase changes happen \cite{correlations}. However, in mm-wave devices, extra resonant elements are needed to enlarge the phase range up to a few $\pi$-times, which makes the relationship between the tilt angle dynamics and phase changes much more complex to model. When including resonant elements such as printed dipoles, the electric field in the cavity contains significant components in all directions and anisotropy can not be overlooked. That is, since the phase change is generated through printed metallizations in the superstrate, the cavity can not be modelled anymore with a medium constant. Furthermore, in optics, the cavities are typically much narrower ($\sim10um$) than in microwaves ($\sim100um$), being such thick cavities quite unexplored and not modelled. Therefore, it is necessary to solve Equation (\ref{eq:ericksenleslie2}) for resonant cells without any of these limitations, so that an appropriate model is achieved considering the number of layers and the inhomogeneity to consider, and to be validated with measurement data. \begin{comment} \begin{figure}[!htb] \centering \includegraphics[width=0.7\columnwidth]{images/diagrama.pdf} \vspace{-0.6cm} \caption{Block diagram.} \label{fig:diagrama} \vspace{-0.6cm} \end{figure} \end{comment} \begin{figure}[!htb] \centering \vspace{-0.2cm} \includegraphics[width=1\columnwidth]{images/stratified2.pdf} \vspace{-0.6cm} \caption{Stratified and averaged LC cavity dynamic modelling strategies. In the stratified strategy, $\overline{\overline{\varepsilon}}_{ri}(\theta_i,t)$ is computed with Eq. (\ref{eq:eptensor}) and considering as $\theta_i$ the average tilt within the layer $i$. In the averaged strategy, $\overline{\overline{\varepsilon}}_{r,avg}(\theta_{avg},t)$ is computed considering as $\theta_i$ the average tilt across the entire cavity. \label{fig:stratified} \vspace{-0.4cm} \end{figure} \section{Model Results and Experimental Validation} \label{sec:Model Results and Experimental Validation} In order to accurately model the tilt angle dynamics of the LC when an excitation change occurs, we computationally solve Equation (\ref{eq:ericksenleslie2}) along $z$ and $t$ using finite elements method in COMSOL Multiphysics \cite{comsol}. This avoids several error sources as we specifically consider the pretilt effects and a more complete set of elastic constants, obtaining a more precise data for the $\theta(z)$ curves at any timestamp and LC driving voltage. Moreover, this allows to model any kind of excitation beyond step-like functions, although if the excitation signal has a frequency high enough so that its period is much smaller than the relaxation time of the LC, it can be substituted by its root mean square (RMS) value in Equation (\ref{eq:ericksenleslie2}). \insertWideFigure{images/layers_comparison_convergence_all_GT323001_75um_97GHz_v3}{Stratified simulation convergence study at 97 GHz a) from 0V to 5V b) from 0V to 45V c) from 0V to 150V.} Fig. \ref{fig:transition_up} shows the molecules tilt angle as a function of $z$ for different timestamps in three different excitation scenarios, in a $75\mu m$ thick cavity filled with GT3-23001 LC. It can be observed how large voltage excitations make transitions faster and much more homogeneous tilts. By using the proposed tool, Fig. \ref{fig:approx_error} shows the error on the computed rise time (10\% to 90\%) when the sine approximation and Equations (\ref{eq:decay}) and (\ref{eq:rise}) are used with respect to the real solution of Equation (\ref{eq:ericksenleslie2}) for the same cavity. As expected, it is necessary to model the cell rigorously especially in the high voltage regime, given that errors larger than 10\% could be made in predicting the LC temporal behaviour, which in turn induces large errors in the cell RF reflected phase as a function of time. \begin{figure*}[h!] \centering \includegraphics[width=1\textwidth]{images/setup_blocks.pdf} \vspace{-0.7cm} \caption{Measurement setup. a) Reflectarray picture b) Block diagram of the setup c) Quasi-optical bench picture} \label{fig:setup} \vspace{-0.5cm} \end{figure*} \begin{figure}[!htb] \centering \includegraphics[width=1\columnwidth, height=5.5cm]{images/comparison_phase_spectrum_gt323001_v2.pdf} \vspace{-0.6cm} \caption{Phase of the reflection coefficient. Dashed lines indicate simulation data and flat lines indicate experimental measurements.} \label{fig:phase_spectrum} \vspace{-0.5cm} \end{figure} Once the different tilt angles along $z$ are obtained for each timestamp of a certain transition, and therefore $\hat{n}(z,t)$ is known, it is necessary to compute the permittivity tensors in accordance with the formulation of section \ref{sec:Permittivity Tensor Calculation}. After the permittivity model has been calculated using COMSOL, this is used in an electromagnetic simulator (CST Studio \cite{cst}) with the aim of obtaining the electrical parameters of the periodical reflective cell in RF. As sketched in Fig. \ref{fig:stratified}, in order to model the LC inhomogeneity along $z$ two different strategies have been used and compared, similarly to \cite{accurate}. On one hand, a stratified media has been considered, deploying different substrate layers within the LC. In this way, we consider the inhomogeneity along $z$ by partitioning the LC in N uniform layers in which the average tilt angle is assumed. With a large number of layers, this method is the most accurate but highly inefficient, even more than in \cite{accurate} due to the multiple simulations needed for a dynamics study. On the other hand, an average of the tilt angle across the entire cavity has been considered in order to work with a single-layer material encompassing the permittivity inhomogeneity across z. The purpose of doing so is to find a trade-off between accuracy and efficiency given that a multi-layer electromagnetic analysis is very costly. Note that this more precise than directly averaging the permittivity across $z$. This introduces a certain error to the computation, but greatly increases the efficiency of the simulation. Specifically, the latter method will become more accurate with extreme bias voltages, as the average and the local tilt values are almost the same across all $z$, and the difference with the stratified method will be negligible. This can be noticed in the top curves of Fig. \ref{fig:transition_up}, where the 150V excitation makes almost all molecules along $z$ to be rotated 90º at $T=35ms$, in contrast with the 15V excitation at $T=4s$. Therefore, in medium-voltage excitations, the stratified procedure will work slightly better but its complexity will increase abruptly. Fig. \ref{fig:images/layers comparison convergence all GT323001 75um 97GHz v3} compares the convergence of such stratified models by computing phase transitions in a 1, 3, 5 and 20-layered unit cell, showing that the single layer method is really precise, especially at large voltage excitations, as opposed to the analytical approximations. As will be seen later in the paper, these high-voltage excitations are of special interest for the overdrive technique to reduce reconfigurability times. The error made in the homogeneous case with respect to the stratified model is found to be reduced (around 20º in the worst case) and admissible, given that the corresponding degradation of the radiation pattern for such phase deviation is negligible \cite{quantization,encinar}. Moreover, it can be appreciated how for increasing voltages the difference between both methods vanishes (20º and 0.5º difference in 5V and 150V excitation respectively). The error is similar to that found in \cite{accurate} for statics, but it is generalized here for dynamics. Therefore, single-layer modelling is carried out for the following experiments with the aim of increasing computation efficiency at a negligible accuracy loss, especially at high voltages. However, the stratified approach could be followed for a perfectionist modelling at higher computational cost. For mid-range problems, increasing the number of layers until the accuracy converges is a reasonable procedure. In order to validate the simulated dynamic results, the transient phase-curves at a certain frequency are compared against experimental data, captured from a reflectarray antenna whose unit cell is shown in Fig. \ref{fig:cavity}c, containing a LC cavity filled with GT3-23001 from Merck \cite{merck} ($k_{33}=34.5pN; k_{11}=24pN; \gamma_1=746mPas; \varepsilon_{r||}=3.27; \varepsilon_{r\perp}=2.47; \Delta\varepsilon_q=4.6$) sandwiched below a Quartz superstrate ($\varepsilon_r = 3.78, tan \delta = 0.002$). The complete antenna consists on 60 x 60 identical cells (see Fig. \ref{fig:setup}a), and its phase response is shown in Fig. \ref{fig:phase_spectrum} for extreme excitations (OFF state corresponds to $V=0$ and ON state corresponds to $V>>Vth$) in stationary regime. To experimentally acquire the dynamic cell phase curves at each timestamp for a specified frequency, incidence angle and bias signal, the setup shown in Fig. \ref{fig:setup}b and \ref{fig:setup}c has been implemented. An arbitrary waveform generator (Keysight 33611A), programmed to output different bias signals, drives the LC of the reflectarray antenna through a x15 voltage multiplier. In order to ensure a specular reflection, all unit cells of the antenna are short-circuited so that the LC is excited equally along the entire array. The waveform generator sends a SYNC signal to a vector network analyzer (VNA) to guarantee a timely capture of the transition. The VNA has been previously calibrated with a metallic plane reference and equipped with a pair of horn antennas, and captures the evolution of the transmission scattering parameters at 97 GHz and 102 GHz when the bias signal sequence starts. Those frequencies were selected since they show maximum phase range within the band of operation of the 360º cell, thus allowing the detection of the maximum phase errors. Both the experiments and the electromagnetic simulations have been carried out considering an impinging angle of 30º with respect to the normal of the reflectarray plane ($\phi_i=0$º, $\theta_i=30$º). Fig. \ref{fig:transition_modelo}b compares simulations and measurements of the transient reflected field phase at 102 GHz as a function of time for different voltage rise transitions starting from idle, verifying that the tool can predict relatively close the actual cell behaviour. The applied 1 kHz square signal allows to assume an amplitude equal to its RMS value in Equation (\ref{eq:ericksenleslie2}). The model has been validated with relaxation measurements as well. Fig. \ref{fig:transition_modelo}c shows simulation and measurement data of the reflected field phase evolution at 102 GHz for different decay transitions parting from varying voltages. The corresponding biasing signals are shown in Fig. \ref{fig:transition_modelo}a. As can be seen, the temporal model of the LC transitions matches the experimental data both in rise and decay time, as well as in phase range. It can be noticed how, as expected, while the rise transitions are highly dependent on the applied voltage, the decay transitions are quite similar regardless of the driving amplitude at $T=0$. Small discrepancies ($<$30º) between the expected and measured phase range in permanent regime can be explained by the phase curve in Fig. \ref{fig:phase_spectrum}, where the phase difference between both states slightly differs. It should be mentioned that this range of error becomes negligible when conforming a full radiation pattern, as it is equivalent to a 3-bit phase quantization, which generally suffices to synthesize a collimated beam and only deteriorates gain by 0.2 dB and SLL by 0.8 dB \cite{phd,quantization}. Errors in transient regime (up to 200º in the worst case of $V1=8V$) can be associated to different sources, including the 1-layer assumption and especially the pre-tilt angle estimation and LC RF characterization. This error could be minimized by choosing a denser layer stratification and assuming a lower computational efficiency, considering their trade-off. Additionally, the different error sources could be compensated in the model by performing, a posteriori, an effective parameter tuning for each voltage curve using measurement data. \begin{comment} \begin{figure}[!htb] \centering \includegraphics[width=1\columnwidth]{images/Comparison_Dynamic_LC_behaviour_0toX_GT323001_pretilt2_75um_102GHz_v8.pdf} \vspace{-0.7cm} \caption{GT3-23001 phase transition dynamics at 102 GHz when excited by a 1 kHz square signal with different amplitudes. The asterisk marker indicates simulation data and flat lines indicate experimental measurements.} \label{fig:transition_subida} \end{figure} \begin{figure}[!htb] \centering \vspace{-0.3cm} \includegraphics[width=1\columnwidth]{images/Comparison_Dynamic_LC_behaviour_Xto0_V_GT323001_75um_102GHz_v6.pdf} \vspace{-0.8cm} \caption{GT3-23001 phase transition dynamics at 102 GHz when relaxed to 0V from a 1 kHz square signal with different amplitudes. The asterisk marker indicates simulation data and flat lines indicate experimental measurements.} \label{fig:transition_bajada} \end{figure} \end{comment} \begin{figure*}[!htb] \centering \includegraphics[width=1\textwidth, height=6.5cm]{images/modelo_rise_decayv19.pdf} \vspace{-0.7cm} \caption{GT3-23001 phase transition dynamics at 102 GHz. a) 1 kHz biasing signal for excitation (top) and relaxation (bottom) dynamics b) Excitation transient phase for different V1 values c) Relaxation transient phase for different V2 values. The asterisk marker indicates simulation data and flat lines indicate experimental measurements.} \label{fig:transition_modelo} \end{figure*} \section{Biasing Synthesis Techniques} The dynamic modelling of LC multi-resonant cells, obtained and experimentally validated in Section \ref{sec:Model Results and Experimental Validation}, enables the development of bias voltage design techniques through simulations, which allow an improvement on the antenna reconfigurability times. These methods can be used to reduce both the relaxation times (underdrive) and the rising times (overdrive) although, as will be seen next, the main improvement occurs in the rising transitions. \begin{figure}[!htb] \centering \vspace{-0.3cm} \includegraphics[width=1\columnwidth]{images/overdrive_overviewv2.pdf} \includegraphics[width=1\columnwidth]{images/overdrive_overview_nominalv3.pdf} \vspace{-0.7cm} \caption{Overdrive technique (top) and nominal biasing (bottom) close-up look.} \label{fig:overdrive_overview} \vspace{-0.3cm} \end{figure} Specifically, by leveraging an accurate temporal control of the LC, a timely driving of the LC cells can be used to accelerate the transition times between phase states. That is, by overdriving it in the rise transitions, the LC orientation time can be accelerated when the electric field is increased (i.e. rotating the molecules towards parallel to $z$). This is achieved by using during a short period of time a larger voltage than the nominal biasing voltage (i.e. the voltage in which the cell presents the desired phase shift in permanent regime, after the molecules stopped rotating). This can be seen in Fig. \ref{fig:overdrive_overview}, where the overdrive LC driving signal amplitude is 150V until the 10V objective reflection coefficient phase is achieved, instant in which we switch the excitation to the nominal 10V signal amplitude. Similarly, by underdriving it (sometimes referred to as undershoot), the LC orientation time can be accelerated when the electric field is decreased (i.e. rotating the molecules towards perpendicular to $z$). In this case, the LC is briefly driven at a lower voltage than the nominal one. Additionally, by using dual frequency LC, the overdrive technique can be used to accelerate both transitions. The overdrive technique has been used in optic devices in the past \cite{overdrive1}, but by using approximations instead of an accurate profiling, and not in the mm-wave regime where the cell thickness and modelling become problematic. In the case of aperture antennas where the objective parameter is the pixel phase, the design procedure of the LC overdriving signal for quickly achieving the desired phase of an array cell, defined by its dimensions and incidence angle, is the following: \begin{enumerate} \item Identify the nominal voltage that achieves the desired phase shift in stationary state. \item Compute $\theta(z,t)$ for the rise transition towards the nominal voltage, by solving Equation (\ref{eq:ericksenleslie2}). \item Find the phase-time curve of such transition by solving the structure electromagnetically, for each timestamp, after finding $\overline{\overline{\varepsilon_r}}$ from Equation (\ref{eq:eptensor}). \item Repeat steps 2 and 3 for the rise transition towards the maximum voltage. \item Pick, from the rise transition towards the maximum voltage, the timestamp in which the instantaneous phase matches the converged phase of the nominal transition. \end{enumerate} \begin{figure}[t] \centering \vspace{-0.1cm} \includegraphics[width=1\columnwidth]{images/overdrive_0to10V_102Ghz_todo_v5.pdf} \vspace{-0.8cm} \caption{Measured and simulated phase of reflection coefficient at 102 GHz during a 0V to 10V transition. To quickly achieve the 10V state phase, we overdrive the LC to 75V during 75ms and then switch to the nominal 10V excitation.} \label{fig:overdrive 10V} \vspace{-0.3cm} \end{figure} \begin{figure*}[t] \vspace{-0.4cm} \centering \includegraphics[width=1\textwidth]{images/meas_and_sim_overdrive_underdrivev5.pdf} \vspace{-0.8cm} \caption{Phase transition between states using overdrive/underdrive and nominal excitations. Top row shows Simulations and middle row shows Measurements of a) 0V to 10V at 97 GHz, using a 150V overdrive for 19ms; b) 0V to 15V at 97 GHz, using a 150V overdrive for 21ms; c) 0V to 15V at 102 GHz, using a 75V overdrive for 90ms; d) 15V to 5V at 97 GHz, using 0V underdrive for 2.5s. Bottom row shows the applied overdrive/underdrive bias signal.} \label{fig:overdrive_underdrive_simsymeas} \vspace{-0.5cm} \end{figure*} \begin{table*}[h] \centering \begin{threeparttable}[b] \caption{Electrical beam-steering works comparison at both cell and antenna level} \begin{tabular}{|c|c|c|c|c|c|c|c|c|} \hline \textbf{Work} & \textbf{Technology} & \textbf{Freq. (GHz)} & \textbf{Losses} & \textbf{PR} & \textbf{Ton} & \textbf{Toff} & \textbf{Pol.} & \textbf{Complete antenna parameters}\\ \hline \cite{adualpol} & PIN & 7.45 & 12dB & 180º & $<$1ms\tnote{*} & $<$1ms\tnote{*} & DL & G: 21dB. SA: $\pm$40º 2D. BW: 0.85 (3dB)\\ \hline \cite{dualband} & Varactor & 11.3/14.7 & 2/3dB & 300º & $<$1ms\tnote{*} & $<$1ms\tnote{*} & C & G: 14/16.6dB. SA: $\pm$24º 2D. BW: 0.85/0.63 (3dB)\\ \hline \cite{realization} & Varactor & 23.5 & 3dB & 320º & $<$1ms\tnote{*} & $<$1ms\tnote{*} & SL & SA: $\pm$60º 1D\\ \hline \cite{vo2} & VO$_2$ & 32 & 1dB & 300º & 12ms & 2s & SL & -\\ \hline \cite{mems} & MEMS & 11.2 & 0.5dB & 180º & $<$1ms\tnote{*} & $<$1ms\tnote{*} & DL & -\\ \hline \cite{reflectarray2} & LC & 24.1 & 4dB & 360º & 5s\tnote{*} & 10s\tnote{*} & SL & G: 20.2dB. SA: $\pm$45º 2D. BW: 4 (3dB)\\ \hline \cite{folded} & LC & 78 & 12dB & 270º & 5s\tnote{*} & 10s\tnote{*} & SL & G: 25.1dB. SA: 17º 1D. BW: 3 (1dB)\\ \hline \cite{lcreflectarray}** & LC & 100 & 6dB & 360º & 5s & 10s & SL & G: 19.4dB. SA: 55º 1D. BW: 6 (3dB)\\ \hline This work** & LC + overdrive & 100 & 6dB & 360º & $<$50ms & 5s & SL & -\\ \hline \end{tabular} \vspace{0.1cm} \label{tab:works} \begin{tablenotes} \item[*] Estimated considering technology. ** Same cell design and LC material employed to facilitate the temporal comparison.\\ PR = Phase range; DL = Dual linear; SL = Single linear; C = Circular; G = Gain; SA = Scan angle; BW = Gain bandwidth. \end{tablenotes} \end{threeparttable} \vspace{-0.4cm} \end{table*} Then, the driving signal consists on modifying the amplitude of the nominal biasing signal to the maximum voltage between $t=0$ and the obtained timestamp. Regarding the underdriving signal design, the procedure is dual by using a drop transition towards a zero voltage. Given that implementing these strategies properly requires a very precise knowledge of the LC dynamics in the cells, and given that it is unfeasible to obtain such curves for each cell and incidence angle, it is necessary an accurate enough model like the one proposed in this work that provides such information through simulations. In order to validate such technique, different temporal driving signals have been computed so as to reduce the switching times for different state transitions, and experimental measures have been obtained using those excitations. In Fig. \ref{fig:overdrive 10V}, a comparison between simulated and measured data for both overdrive and nominal excitations is shown for a 0V to 10V transition. As can be seen, the overdrive technique applied to the reflectarray antenna allows to accelerate the rise time by a factor 100X. In Fig. \ref{fig:overdrive_underdrive_simsymeas}, measurements and simulations for both overdrive and underdrive techniques are compared against a normal operation for different phase transitions at two representative frequencies in the band of design, and the corresponding bias signals are shown. As can be observed, the overdrive strategy allows for a much quicker phase drop than the normal operation, reducing in several orders of magnitude the switching times. Although the underdrive technique also shows some time reduction (Fig. \ref{fig:overdrive_underdrive_simsymeas}d), its effect is not so pronounced, as the decay transition is less dependent on the excitation. Moreover, since the rise transitions are completely dependent on the driving amplitude, the voltage can theoretically be further increased to reduce the rising times (in practice, we will be restricted by equipment and the limited cavity impedance creating a short-circuit), but it can not be underdriven beyond 0V since it is the absolute magnitude of $E$ what makes LC molecules to rotate. Notwithstanding, the decay times can be further reduced by choosing a less viscous LC, or by employing dual-frequency LC, which can in turn benefit from this overdrive technique to decrease relaxation times. Additionally, both excitation and relaxation times may be drastically decremented by properly combining the overdrive techniques and the previously mentioned polymerizable materials, although future work on its characterization is required. It should be mentioned that even though the voltage is increased significantly during these transitions, the overall power consumption is minimum, as the LC cell hardly consumes any current. An interesting phenomenon that occurs in some of the captured phase transitions is a significant rebound of the phase right after the objective phase is initially achieved in the overdrive excitation, as shown in Fig. \ref{fig:overdrive 10V} and the middle row of Fig. \ref{fig:overdrive_underdrive_simsymeas}. This bounce is the manifestation of both the bias signal commutation (which causes the molecular reorientation of Fig. \ref{fig:transition_up}c) and the backflow effect \cite{highspeed,pretilt, kickback,bck}, which appears as a consequence of working in the high-voltage regime in thick cells and that our model did not completely capture. This is a well known phenomenon that, if one has access to the LC Leslie coefficients, either through manufacturer data or through experimental estimation \cite{backflow,lesliecoef}, could be included in the problem to make the model more accurate but more computationally costly. Even though this effect deteriorates the experimental measurements as the phase oscillates slightly ($<$65º in Fig. \ref{fig:overdrive 10V}) until reaching the final value, we can still approach the objective phase state much faster than under a nominal excitation, being the instantaneous phase during such transient effect $\pm$20\% deviated only. Overall, the predicted and measured reflectarray cells transitioned a maximum of 250X and an average of 100X faster between phase shift states when using overdrive techniques, as compared to using nominal excitations. On the other hand, by completely removing the biasing voltage temporarily, the underdrive excitations shortened in average a 2X time factor to achieve 90\% of the objective phase, as compared to the nominal excitations. To put this into perspective, Table \ref{tab:works} compares different works on electrical beam-steering phase-shift LC metasurfaces and unit cells, and its performance including transition times. Additionally, a comparison with other reconfigurable metasurface technologies is included. \section{Conclusion} LC-based reconfigurable metasurfaces are promising candidates for developing electrically large aperture antennas supporting the future generation of communications, given their easiness of manufacturing, low cost and wide operating frequency ranges. However, switching times between phase states must be reduced before they can be widely used in real-time applications. In this work, a dynamical model of LC transitions for different excitations beyond the known approximations is presented and validated in order to achieve a temporal control of the unit cell phase, useful for both reflective and transmissive multi-resonant cells. A further analysis on the LC stratified model is also provided by considering a different number of simulated layers, concluding that a reduced number of layers (N=20) is needed in the worst case, although using an effective tensor (N=1) will be enough to achieve reasonable accuracy most of the times, and useful to perform efficient electromagnetic simulations. Even though the effect of the different LC driving excitations on the phase change can be carried out through both measurements and simulations, a generalization in frequency, incident angle, cell designs and LC materials could be cumbersome to do by means of measures. Instead, a simulation tool like the one introduced in this work allows for a fast and accurate estimation of control signals to introduce the temporal parameter in the design space of electrically large antennas. In turn, this allowed to design and validate an overdriving technique capable of drastically reducing transition times by one or two orders of magnitude in a simple way. In the rising case, reductions are of a factor 100 while in the relaxation case improvements are less drastic. However, this strategy can be used together with others to enhance the whole LC dynamic behaviour and obtain reduced antenna scanning times. \vspace{-0.3cm}

6b704c17e1123871737b701776795f2de85cfb09

\section*{Acknowledgments} We are grateful to Yuni Iwamasa and Taihei Oki for initial discussions on the problem. The work was supported by the Lend\"ulet Programme of the Hungarian Academy of Sciences -- grant number LP2021-1/2021 and by the Hungarian National Research, Development and Innovation Office -- NKFIH, grant numbers FK128673 and TKP2020-NKA-06. Yutaro Yamaguchi was supported by JSPS KAKENHI Grant Numbers 20K19743 and 20H00605, and by Overseas Research Program in Graduate School of Information Science and Technology, Osaka University. Yu Yokoi was supported by JST PRESTO Grant Number JPMJPR212B. \section{Matroid Intersection under Common Independence Oracle} \label{sec:ci} As discussed in Section~\ref{sec:oracles}, the common independence oracle is strictly weaker than the minimum rank and the rank sum oracles. As weighted matroid intersection turned out to be tractable for the rank sum oracle, the complexity of the problem under the common independence oracle is especially interesting. In what follows, we present an algorithm for the unweighted matroid intersection problem when one of the matroids is a partition matroid, and an algorithm for the weighted matroid intersection problem when one of the matroids is an elementary split matroid. We also show that the common independence oracle, when complemented with the rank oracle, is strong enough to design an algorithm similar to that for the rank sum case. \subsection{Intersection with Partition Matroid} \label{sec:cipart} The aim of this section is to show that the unweighted matroid intersection problem is tractable under the common independence oracle when ${\mathbf{M}}_1$ is a \textbf{partition matroid} with all-one upper bound on the partition classes, that is, when ${\mathcal{I}}_1$ is represented as ${\mathcal{I}}_1 = \{\, I \subseteq E \mid |I \cap E_i| \leq 1\ \text{for $i=1,\dots,q$}\,\}$ for some partition $E=E_1\cup\dots\cup E_q$. We will provide an algorithm that emulates Algorithm~\ref{alg:1}, i.e., {{\sc Augment}$[E, {\mathcal{I}}_1, {\mathcal{I}}_2, I]$}, using only the common independence oracle. Take any $k = 0, 1, \dots, n - 1$ and let $I\in {\mathcal{I}}_1^k \cap {\mathcal{I}}_2^k$. To emulate Algorithm~\ref{alg:1}, we want to find a shortest $S_I$--$T_I$ path in the exchangeability graph $D[I] = (E \setminus I, I; A_1[I]\cup A_2[I])$. With only the common independence oracle, however, we cannot construct $D[I]$, and cannot determine even $S_I$ or $T_I$. Note that a shortest $S_I$--$T_I$ path in $D[I]$ never uses arcs entering sources or leaving sinks. Therefore, finding a shortest $S_I$--$T_I$ path in $D[I]$ is equivalent to finding it in $D'[I]$, where $D'[I]$ is the subgraph of $D[I]$ obtained by removing those arcs from $D[I]$ (it is used also in Section~\ref{sec:rank_sum}). We now provide a search procedure, described as Algorithm~\ref{alg:part2}, that will be used as a subroutine for our augmentating procedure. If a given element $s$ belongs to $S_I$, this search algorithm works like the breadth first search in $D'[I]$ rooted at $s$, and returns a shortest $s$--$T_I$ path or certifies the nonexistence of such a path. In Algorithm~\ref{alg:part2}, for each $y\in I$, a sequence $P_y$ of distinct elements is defined. In our analysis, $P_y$ will turn out to be a shortest $s$--$y$ path in $D'[I]$. We use the notation $P_y+x$ to denote the sequence obtained by appending an element $x$ to $P_y$. \begin{algorithm2e}[h!] \caption{{{\sc EmulatingBFS}$[E, {\mathcal{I}}_1 \cap {\mathcal{I}}_2, I, s]$}} \label{alg:part2} \SetAlgoLined \SetKwInOut{Input}{Input}\SetKwInOut{Output}{Output} \Input{Oracle access to ${\mathcal{I}}_1 \cap {\mathcal{I}}_2$ where ${\mathcal{I}}_1$ is the independent set family of a partition matroid, a common independent set $I \in {\mathcal{I}}^k_1 \cap {\mathcal{I}}^k_2$, and an element $s \in E \setminus I$.} \Output{A sequence $P$ with $I\triangle P\in {\mathcal{I}}_1 \cap {\mathcal{I}}_2$ and $|I\triangle P|=k+1$ if one exists, or a message \emph{``No''} otherwise. In particular, if $s\in S_I$ and $D'[I]$ has an $s$--$T_I$ path, then a shortest $s$--$T_I$ path is returned.} \BlankLine If $I+s\in {\mathcal{I}}_1 \cap {\mathcal{I}}_2$, halt with returning $s$. For each $y\in I$, set $P_y\gets sy$ if $I+s-y\in {\mathcal{I}}_1\cap {\mathcal{I}}_2$, and $P_y \gets {\mathsf{null}}$ otherwise. For $\ell=1,2,\dots$, do the following. \protect{ \begin{enumerate}[(i)] \item If there is no $y\in I$ with $|P_y|=2\ell$, halt with returning \emph{``No''}. \item If there exist $y'\in I$ and $x\in E\setminus I$ such that $|P_{y'}|=2\ell$, $x\not\in P_{y'}$, $\{y',x\}\not\in {\mathcal{I}}_1\cap{\mathcal{I}}_2$, and $I\triangle(P_{y'}+x)\in {\mathcal{I}}_1\cap {\mathcal{I}}_2$, then halt with returning $P_{y'}+x$. \item For each $y\in I$ with $P_y = {\mathsf{null}}$, if there exist $y'\in I$ and $x\in E\setminus I$ such that $|P_{y'}|=2\ell$, $x\not\in P_{y'}$, $\{y',x\}\not\in {\mathcal{I}}_1\cap {\mathcal{I}}_2$, and $I\triangle(P_{y'}+x+y)\in {\mathcal{I}}_1\cap {\mathcal{I}}_2$, then $P_y\gets P_{y'}+x+y$. \end{enumerate} } \end{algorithm2e} By the algorithm, it is clear that the output is either a sequence $P$ with $I\triangle P\in {\mathcal{I}}_1 \cap {\mathcal{I}}_2$ and $|I\triangle P|=k+1$ or a message \emph{``No''}. Also, if $s\in S_I\cap T_I$, we see that $s$ itself is a shortest $s$--$T_I$ path and is returned at Step~1. Therefore, we assume $s\in S_I\setminus T_I$ and show that a shortest $s$--$T_I$ path is returned if such a path exists. We denote by $\dist(s,T_I)$ the length (i.e., the number of vertices) of a shortest $s$--$T_I$ path in $D'[I]$ and by $\dist(s,y)$ the length of a shortest $s$--$y$ path in $D'[I]$ for each $y\in I$. Note that $\dist(s,T_I)$ is odd and $\dist(s,y)$ is even for any $y\in I$. \begin{lemma}\label{lem:part} Let $s\in S_I\setminus T_I$. The following hold for any $y\in I$ and any $\ell=1,2,\dots$. \begin{enumerate}[\rm (a)] \item $P_y$ is defined in Step~2 if and only if $\dist(s,y)=2$. If defined, it is a shortest $s$--$y$ path in $D'[I]$. \item A sequence is returned in the $\ell$th process of Step~3(ii) if and only if $\dist(s,T_I)=2\ell+1$. The returned sequence is a shortest $s$--$T_I$ path in $D'[I]$. \item $P_y$ is defined in the $\ell$th process of Step~3(iii) if and only if $\dist(s,y)=2\ell+2<\dist(s,T_I)$. If defined, it is a shortest $s$--$y$ path in $D'[I]$. \end{enumerate} \end{lemma} \begin{proof} For any $y\in I$, $\dist(s,y)=2$ means $(s,y)\in A'_2[I]$, which is equivalent to $I+s-y\in {\mathcal{I}}_2$ as $s\not\in T_I$. Since $s\in S_I$ implies $I+s-y\in {\mathcal{I}}_1$, then $I+s-y\in {\mathcal{I}}_1\cap {\mathcal{I}}_2$ holds if and only if $(s,y)\in A'_2[I]$. When $(s,y)\in A'_2[I]$, clearly $sy$ is a shortest $s$--$y$ path. Thus, (a) is shown. We show (b) and (c) by induction on $\ell$. Suppose that they hold for $1,2,\dots,\ell-1$ and we are at the beginning of the $\ell$th process of Step~3. Then $\dist(s,T_I)\geq 2\ell+1$ because otherwise the algorithm has halted before. Take any $y'\in I$ with $|P_{y'}|=2\ell$. Then \begin{itemize} \item $P_{y'}$ is a shortest $s$--$y'$ path in $D'[I]$ by (a) and induction for (c), \item $\mathrm{cl}_2(I)=\mathrm{cl}_2(I\triangle P_{y'})$ because $|I|=|I\triangle P_{y'}|$ and $I\triangle P_{y'}\in {\mathcal{I}}_2$ hold by the algorithm (Steps 2 and 3(iii)) and $P_{y'}\cap T_I=\emptyset$ follows from $\dist(s,T_I)\geq 2\ell+1$. \end{itemize} As $P_{y'}$ is a path in $D'[I]$, it uses arcs in $A'_2[I]$ and $A'_1[I]$ alternately. Let $N_1$ and $N_2$ be the sets of those arcs of $A'_1[I]$ and $A'_2[I]$, respectively. By $s\in E\setminus I$ and $y'\in I$, then $N_1$ and $N_2$ form matchings that cover $V(P_{y'})-s-y'$ and $V(P_{y'})$, respectively (recall that we denote by $V(P)$ the set of elements in a sequence $P$ for emphasizing that we focus on the set rather than the sequence). Take any $x\in E\setminus I$ with $x\not\in P_{y'}$. The following claim completes the proof of (b). \begin{claim}\label{claim:partition1} $(y',x)\in A'_1[I]$ and $x\in T_I$ if and only if $\{y',x\}\not\in {\mathcal{I}}_1\cap{\mathcal{I}}_2$ and $I\triangle(P_{y'}+x)\in {\mathcal{I}}_1\cap {\mathcal{I}}_2$. \end{claim} \begin{proof} For the ``only if'' part, suppose $(y',x)\in A'_1[I]$ and $x\in T_I$. As $\bf{M}_1$ is a partition matroid with all-one upper bounds, $(y',x)\in A'_1[I]$ means that $\{y',x\}$ is a circuit in $\bf{M}_1$, and hence $\{y',x\}\not\in {\mathcal{I}}_1\cap{\mathcal{I}}_2$. Since $s\in S_I$ and $N_1+(y',x)$ forms a matching that covers $V(P_{y'}+x)-s$, we have $I\triangle(P_{y'}+x)\in {\mathcal{I}}_1$. (In $I\triangle(P_{y'}+x)$, each element in $I\cap (P_{y'}+x-s)$ is replaced by another element in the same partition class and $s$ comes from a partition class whose element is not used in $I$.) Also, $\mathrm{cl}_2(I)=\mathrm{cl}_2(I\triangle P_{y'})$ and $x\in T_I$ imply $I\triangle (P_{y'}+x)=(I\triangle P_{y'})+x\in {\mathcal{I}}_2$. Thus, the ``only if'' part is shown. For the ``if'' part, suppose $\{y',x\}\not\in {\mathcal{I}}_1\cap{\mathcal{I}}_2$ and $I\triangle (P_{y'}+x)\in {\mathcal{I}}_1\cap{\mathcal{I}}_2$. As $\mathrm{cl}_2(I)=\mathrm{cl}_2(I\triangle P_{y'})$ holds, $I\triangle (P_{y'}+x)\in {\mathcal{I}}_2$ implies $x\in T_I$. Then $\{y',x\}\in {\mathcal{I}}_2$, and hence $\{y',x\}\not\in {\mathcal{I}}_1\cap{\mathcal{I}}_2$ implies that $\{y',x\}$ is a circuit in $\bf{M}_1$. Thus $(y',x)\in A'_1[I]$. \end{proof} Suppose that we are at the beginning of $\ell$th process of Step~3(iii). Take $y'$ and $x$ as before and take any $y\in I$ such that $P_y$ is undefined. Then $\dist(s,y)>2\ell$ by (a) and induction for (c). Also $(y',x)\in A'_1[I]$ implies $x\not\in T_I$ since otherwise the algorithm has halted at Step~3 (ii). The following claim completes the proof of (c). \begin{claim}\label{claim:partition2} Assume that $(y',x)\in A'_1[I]$ implies $x\not\in T_I$. Then $(y',x)\in A'_1[I]$ and $(x,y)\in A'_2[I]$ if and only if $\{y',x\}\not\in {\mathcal{I}}_1\cap{\mathcal{I}}_2$ and $I\triangle(P_{y'}+x+y)\in {\mathcal{I}}_1\cap {\mathcal{I}}_2$. \end{claim} \begin{proof} For the ``only if'' part, suppose $(y',x)\in A'_1[I]$ and $(x,y)\in A'_2[I]$. Similarly to the proof of Claim~\ref{claim:partition1}, $(y',x)\in A'_1[I]$ implies $\{y',x\}\not\in {\mathcal{I}}_1\cap{\mathcal{I}}_2$ and $I\triangle(P_{y'}+x)\in {\mathcal{I}}_1$, and hence $I\triangle(P_{y'}+x+y)=(I\triangle(P_{y'}+x))-y\in {\mathcal{I}}_1$. Suppose, to the contrary, $I\triangle(P_{y'}+x+y)\not\in {\mathcal{I}}_2$. Since $N_2+(x,y)\subseteq A'_2[I]$ forms a perfect matching on $V(P_{y'}+x+y)$, Lemma~\ref{lem:UPM} implies that there exists some other perfect matching $N'_2\subseteq A_2[I]$ on $V(P_{y'}+x+y)$. This $N'_2$ is included in $A'_2[I]$ because $(P_{y'}+x+y)\cap T_I=\emptyset$ follows from $P_{y'}\cap T_I=\emptyset$ and $x\not\in T_I$. Then, $D'[I]$ contains an $s$--$y$ path with arcs in $N_1\cup N'_2$ and length at most $2\ell$, which contradicts $\dist(s,y)>2\ell$. For the ``if'' part, suppose $\{y',x\}\not\in {\mathcal{I}}_1\cap{\mathcal{I}}_2$ and $I\triangle(P_{y'}+x+y)\in {\mathcal{I}}_1\cap {\mathcal{I}}_2$. By Lemma~\ref{lem:UPM-inv}, $I\triangle(P_{y'}+x+y)\in{\mathcal{I}}_2$ implies that $A_2[I]$ contains a perfect matching $N_2''$ on $V(P_{y'}+x+y)$. Conversely, suppose $(x,y)\not\in N''_2$. Then $(x^*,y)\in N_2''$ for some $x^*\in P_{y'}$, and $P_{y'}\cap T_I=\emptyset$ implies $(x^*,y)\in A'_2[I]$. Hence $D'[I]$ has an $s$--$y$ path with length at most $2\ell$, a contradiction. Thus, $(x,y)\in N''_2\subseteq A_2[I]$, which implies $x\in T_I$ or $(x,y)\in A'_2[I]$, where the latter implies $y\in C_2(I,x)\not\subseteq \{y',x\}$. Thus, in both cases, $\{y',x\}\in {\mathcal{I}}_2$. Therefore, $\{y',x\}\not\in {\mathcal{I}}_1\cap{\mathcal{I}}_2$ implies $\{y',x\}\not\in {\mathcal{I}}_1$, and hence $y'\in C_1(I,x)$, and $(y',x)\in A'_1[I]$ follows. By assumption, we then have $x\not\in T_I$, and hence $(x,y)\in A'_2[I]$. \end{proof} Thus, both (b) and (c) hold for $\ell$. \end{proof} Lemma~\ref{lem:part} completes the proof of correctness of Algorithm~\ref{alg:part2}. \begin{lemma}\label{lem:EmulatingBFS} The output of {\sc EmulatingBFS}$[E, {\mathcal{I}}_1 \cap {\mathcal{I}}_2, I,s]$ is always correct. \end{lemma} \begin{proof} If $\dist(s,T_I)=1$, i.e., if $s\in S_I\cap T_I$, then the algorithm correctly returns $s$ at Step 1. If $\dist(s,T_I)=2\ell+1>1$, then $s\in S_I\setminus T_I$, and hence Lemma~\ref{lem:part}(b) implies that a shortest $s$--$T_I$ path is returned in the $\ell$th process of Step 3(ii). \end{proof} Using {\sc EmulatingBFS} (Algorithm~\ref{alg:part2}) as a subroutine, we design a procedure that emulates {\sc Augment}$[E, {\mathcal{I}}_1, {\mathcal{I}}_2, I]$ with the common independence oracle. \begin{algorithm2e}[h!] \caption{{{\sc AugmentCommonIndependencePartition}$[E, {\mathcal{I}}_1 \cap {\mathcal{I}}_2, I]$}} \label{alg:part} \SetAlgoLined \SetKwInOut{Input}{Input}\SetKwInOut{Output}{Output} \Input{Oracle access to ${\mathcal{I}}_1 \cap {\mathcal{I}}_2$ where ${\mathcal{I}}_1$ is the independent set family of a partition matroid, and a common independent set $I \in {\mathcal{I}}^k_1 \cap {\mathcal{I}}^k_2$.} \Output{A common independent set $J \in {\mathcal{I}}^{k+1}_1 \cap {\mathcal{I}}^{k+1}_2$ if one exists, or a message \emph{``No''} otherwise.} \BlankLine If $I + x \in {\mathcal{I}}_1 \cap {\mathcal{I}}_2$ for some $x \in E \setminus I$, then halt with returning $J = I + x$. \label{st:part1} For each $s \in E \setminus I$, perform {\sc EmulatingBFS}$[E, {\mathcal{I}}_1 \cap {\mathcal{I}}_2, I, s]$. If some sequence $P$ is returned, then return $J=I\triangle P$. Otherwise return with the message \emph{``No''}. \label{st:part2} \end{algorithm2e} \begin{lemma}\label{lem:common_indep_partition} The output of {\sc AugmentCommonIndependencePartition}$[E, {\mathcal{I}}_1 \cap {\mathcal{I}}_2, I]$ (Algorithm~\ref{alg:part}) is always correct. \end{lemma} \begin{proof} The output in Step~\ref{st:part1} is clearly correct. As Step~\ref{st:part2} returns some sequence $P$ only if $I\triangle P$ is indeed a common independent set of size $k+1$, it suffices to show that if there exists a common independent set $J$ of size $k+1$, then {\sc EmulatingBFS}$[E, {\mathcal{I}}_1\cap{\mathcal{I}}_2,I, s]$ returns a sequence for some $s \in E \setminus I$. By the correctness of {\sc Augment}$[E, {\mathcal{I}}_1, {\mathcal{I}}_2, I]$, the existence of such $J$ implies that $D[I]$ has some $S_I$--$T_I$ path, and so does $D'[I]$. Then $D'[I]$ contains an $s$--$T_I$ path for some $s\in S_I$. For such $s$, {\sc EmulatingBFS}$[E, {\mathcal{I}}_1\cap{\mathcal{I}}_2,I, s]$ returns a shortest $s$--$T_I$ path in $D'[I]$. \end{proof} Lemma~\ref{lem:common_indep_partition} completes the proof of Theorem~\ref{thm:ci}. \begin{proof}[Proof of Theorem~\ref{thm:ci}] Starting from $I\coloneqq\emptyset$, the size of the common independent set can be gradually increased using {\sc AugmentCommonIndependencePartition}$[E, {\mathcal{I}}_1 \cap {\mathcal{I}}_2, I]$ until $I$ becomes a common independent set of maximum cardinality. The correctness of the algorithm follows by Lemma~\ref{lem:common_indep_partition}. \end{proof} \subsection{Intersection with Elementary Split Matroid} \label{sec:split} Motivated by the study of matroid polytopes from a tropical geometry point of view, Joswig and Schr\"oter~\cite{joswig2017matroids} introduced the notion of \textbf{split matroids}. This class does not only generalize paving matroids, but it is closed both under duality and taking minors. B\'erczi, Kir\'aly, Schwarcz, Yamaguchi and Yokoi~\cite{berczi2022hypergraph} later observed that every split matroid can be obtained as the direct sum of a so-called \textbf{elementary split matroid} and uniform matroids. Elementary split matroids capture all the nice properties of connected split matroids, and is closed not only under duality and taking minors but also truncation. Motivated by representations of paving matroids by hypergraphs, they provided a hypergraph characterization of elementary split matroids as follows. Let $E$ be a ground set of size at least $r$, ${\mathcal{H}}=\{H_1,\dots, H_q\}$ be a (possibly empty) collection of subsets of $E$, and $r, r_1, \dots, r_q$ be nonnegative integers satisfying \begin{align} |H_i \cap H_j| &\le r_i + r_j -r &&\text{for $1 \le i < j \le q$,}\tag*{(H1)}\label{eq:h1}\\ |E\setminus H_i| + r_i &\ge r &&\text{for $i=1,\dots, q$.} \tag*{(H2)}\label{eq:h2} \end{align} Then ${\mathcal{I}}=\{\, X\subseteq E\mid |X|\leq r,\ |X\cap H_i|\leq r_i\ \text{for $1\leq i \leq q$} \,\}$ forms the family of independent sets of a rank-$r$ matroid $M$ with rank function $r_M(Z)=\min\big\{r,|Z|,\min_{1\leq i\leq q}\{|Z\setminus H_i|+r_i\}\big\}$. Matroids that can be obtained in this form are called \textbf{elementary split matroids}. We call a set $F\subseteq E$ \textbf{$H_i$-tight} or \textbf{tight with respect to $H_i$} if $|F\cap H_i|=r_i$. The following lemma shows that an independent set of size less than $r$ cannot be tight with respect to two different hyperedges. \begin{lemma}\label{lem:tight} Let $M$ be an elementary split matroid with representation ${\mathcal{H}}=\{H_1,\dots,H_q\}$ and $r,r_1,\dots,r_q$, and let $F$ be a set of size less than $r$. Then $F$ is tight with respect to at most one of the hyperedges. \end{lemma} \begin{proof} Suppose to the contrary that $F$ is both $H_i$- and $H_j$-tight. Then we get \begin{align*} |H_i\cap H_j| {}&{}\geq |F\cap H_i\cap H_j| = |F\cap H_i|+|F\cap H_j|-|F\cap (H_i\cup H_j)|\\ {}&{}\geq r_i+r_j-|F| > r_i+r_j-r, \end{align*} contradicting~\ref{eq:h1}. \end{proof} Now we show that the weighted matroid intersection problem is tractable under the common independence oracle when ${\mathbf{M}}_1$ is an \textbf{elementary split matroid}, that is, when ${\mathcal{I}}_1$ can be represented as ${\mathcal{I}}_1=\{\, X\subseteq E\mid |X|\leq r,\ |X\cap H_i|\leq r_i\ \text{for $1\leq i \leq q$} \,\}$ for some (possibly empty) hypergraph ${\mathcal{H}}=\{H_1,\dots, H_q\}$ and nonnegative integers $r, r_1, \dots, r_q$ satisfying \ref{eq:h1} and \ref{eq:h2}. The proof is based on observing that the exchangeability graph has a special structure. \begin{proof}[Proof of Theorem~\ref{thm:split}] Suppose that we have oracle access to the common independent set family ${\mathcal{I}}_1\cap {\mathcal{I}}_2$ of two matroids on $E$, where ${\mathcal{I}}_1$ belongs to an elementary split matroid. Consider a $w$-maximal set $I \in {\mathcal{I}}_1^k \cap {\mathcal{I}}_2^k$ for some $k\in\{0, 1, \dots, n - 1\}$. According to Algorithm~\ref{alg:2} and Lemma~\ref{lem:shortest-cheapest-path}, a $w$-maximal set $J\in{\mathcal{I}}_1^{k+1}\cap{\mathcal{I}}_2^{k+1}$, if exists, can be obtained in the form $J=I\triangle P$ where $P$ is a shortest cheapest $S_I$--$T_I$ path in $D'[I]$. \begin{claim}\label{cl:short} If ${\mathcal{I}}_1^{k+1}\cap{\mathcal{I}}_2^{k+1}\neq\emptyset$, then there exists a shortest cheapest $S_I$--$T_I$ path in $D'[I]$ of length at most $3$. \end{claim} \begin{proof} As ${\mathcal{I}}_1^{k+1}\cap{\mathcal{I}}_2^{k+1}\neq\emptyset$, there necessarily exists an $S_I$--$T_I$ path in $D'[I]$; let $P=e_1e_2\cdots e_\ell$ be a shortest cheapest one. As $I$ is not a basis of ${\mathbf{M}}_1$, observe that $I+x\notin{\mathcal{I}}_1$ for some $x\in E\setminus I$ if and only if there exists a hyperedge $H_i$ such that $I$ is $H_i$-tight and $x\in H_i$. By Lemma~\ref{lem:tight}, $I$ is tight with respect to at most one of the hyperedges, hence we get that $E\setminus (S_I\cup I)\subseteq H_i$. This also implies that $A'_1[I]=\{\,(y,x)\mid x\in H_i\setminus I,~y\in H_i\cap I\,\}$. Assume that the length of the path is more than $3$. Then, by the above observation, both $(e_2, e_\ell)$ and $(e_{\ell-1}, e_3)$ exist in $D'[I]$. Therefore $C\coloneqq e_3e_4\cdots e_{\ell-1}e_3$ is a cycle in $D'[I]$, and $P'\coloneqq e_1e_2e_\ell$ is an $S_i$--$T_i$ path in $D'[I]$. By Lemma~\ref{lem:negative_cycle}, the cost of $C$ is nonnegative, and hence the cost of $P'$ is at most the cost of $P$, contradicting the choice of $P$. \end{proof} By Claim~\ref{cl:short}, a $w$-maximal member of ${\mathcal{I}}_1^{k+1}\cap{\mathcal{I}}_2^{k+1}$, if exists, can be found by checking every set $I'$ of size $k+1$ with $|I\triangle I'|\leq 2$. This concludes the proof of the theorem. \end{proof} \subsection{Complemented with Maximum Rank Oracle} \label{sec:cimax} When access is given to both a common independence and a maximum rank oracle, every step of Algorithms~\ref{alg:3} and \ref{alg:4}, i.e., {\sc EmulatingBellmanFord} and {\sc CheapestPathAugmentRankSum}, can be emulated and hence the weighted matroid intersection problem is solved as with Section~\ref{sec:rank_sum}. \begin{proof}[Proof of Theorem~\ref{thm:cimax}] Suppose that we have oracle access to the common independent set family ${\mathcal{I}}_1 \cap {\mathcal{I}}_2$ and the maximum rank function ${r_\mathrm{max}}$ of two matroids on $E$ instead of that to ${r_\mathrm{sum}}$. In {\sc EmulatingBellmanFord} and {\sc CheapestPathAugmentRankSum}, we ask the rank sum oracle the following types of questions: \begin{enumerate}[(a)] \item whether ${r_\mathrm{sum}}(I' + x) = 2|I'|$ or not for $I' \in {\mathcal{I}}_1 \cap {\mathcal{I}}_2$ and $x \in E \setminus I'$, \label{it:reda} \item whether ${r_\mathrm{sum}}(I' + x) = 2|I'| + 1$ or not for $I' \in {\mathcal{I}}_1 \cap {\mathcal{I}}_2$ and $x \in E \setminus I'$, \label{it:redb} \item whether ${r_\mathrm{sum}}(I' + x) = 2|I'|+2$ or not for $I' \in {\mathcal{I}}_1 \cap {\mathcal{I}}_2$ and $x \in E \setminus I'$, \label{it:redc} and \item whether ${r_\mathrm{sum}}(I') = 2|I'|$ or not for $I' \subseteq E$. \label{it:redd} \end{enumerate} These questions can be tested using the common independence and the maximum rank oracles together as follows. The answer to \eqref{it:reda} is \emph{``Yes''} if and only if $I' + x \notin {\mathcal{I}}_1 \cap {\mathcal{I}}_2$ and ${r_\mathrm{max}}(I' + x) = |I'|$, the answer to \eqref{it:redb} is \emph{``Yes''} if and only if $I' + x \not\in {\mathcal{I}}_1 \cap {\mathcal{I}}_2$ and ${r_\mathrm{max}}(I' + x) = |I'| + 1$, the answer to \eqref{it:redc} is \emph{``Yes''} if and only if $I' + x \in {\mathcal{I}}_1 \cap {\mathcal{I}}_2$, and the answer to \eqref{it:redd} is \emph{``Yes''} if and only if $I' \in {\mathcal{I}}_1 \cap {\mathcal{I}}_2$. \end{proof} \section{Introduction} \label{sec:introduction} A cornerstone of matroid theory is the efficient solvability of the matroid intersection problem introduced by Edmonds \cite{edmonds1970submodular}. Efficient algorithms for weighted matroid intersection were developed subsequently by Edmonds \cite{edmonds1979matroid}, by Lawler \cite{lawler1970optimal,lawler1976combinatorial}, and by Iri and Tomizawa \cite{iri1976algorithm}. The min-max duality theorem of Edmonds \cite{edmonds1970submodular} for the unweighted matroid intersection problem was generalized by Frank \cite{frank1981weighted} to the weighted case. These results do not only provide a well-established framework that includes various tractable combinatorial optimization problems such as bipartite matching and arborescence packing, but in certain cases they are unavoidable in solving natural optimization problems that seem to be unrelated to matroids. A beautiful example is the problem of computing a cheapest rooted $k$-connected spanning subgraph of a digraph \cite{frank2009rooted}. This is a pure graph optimization problem and yet the only known polynomial algorithm is based on the recognition that minimal rooted $k$-connected subgraphs of a digraph form the common bases of two matroids, and therefore a weighted matroid intersection algorithm can be applied. In order to design matroid algorithms and to analyze their complexity, it should be clarified how matroids are given. As the number of bases can be exponential in the size of the ground set, defining a matroid in an explicit form is inefficient. Rather than giving a matroid as an explicit input, it is usually assumed that one of the standard oracles is available, and the complexity of the algorithm is measured by the number of oracle calls and other elementary steps. Another way to define a matroid is to give an explicit linear representation, but this restricts the scope of the algorithm to linear matroids for which such an explicit representation is known. For both the unweighted and weighted problems, a variety of efficient algorithms have been developed; see e.g. \cite{edmonds1970submodular, edmonds1979matroid, lawler1970optimal, aigner1971matching, lawler1975matroid, iri1976algorithm, cunningham1986improved, frank1981weighted, brezovec1986two, huang2019exact}. A common feature of these algorithms, and also all previous studies on matroid intersection, is that they assume the availability of one of the standard oracles for both matroids; e.g., we can ask for the rank of a subset in each of them. Our main contribution is showing that this assumption is not necessary for the tractability of matroid intersection, not even in the weighted setting. One motivation for studying restricted oracles comes from polymatroid matching, a framework introduced by Lawler~\cite{lawler1976combinatorial} as a common generalization of matroid intersection and non-bipartite matching. In \cite{lovasz2009matching}, Edmonds' theorem was deduced from polymatroid matching using a sophisticated argument. The main point is that when the matroid intersection problem is formulated as a polymatroid matching problem, only the \emph{rank sum function} of the two matroids is used rather than the two rank functions. Although the polymatroid matching problem cannot be solved in polynomial time in general~\cite{lovasz1981matroid, jensen1982complexity}, the hardness was shown through special instances that seem to be far from matroid intersection. This suggests that matroid intersection might still be tractable when only the sum of the rank functions is available. We mention that another natural oracle to consider is the \emph{minimum rank function}, which answers the smaller value among the two ranks of a subset. It follows from the polyhedral results of Edmonds \cite{edmonds1970submodular} that this oracle suffices to describe the convex hull of common independent sets. In an unpublished manuscript, Bárász \cite{egresqp-06-03} gave a polynomial time algorithm for unweighted matroid intersection under the minimum rank oracle. We will present additional results about this oracle in a separate paper~\cite{inpreparation}. \paragraph{Our results} Our goal is to settle the tractability of the weighted matroid intersection problem under restricted oracles. In particular, we will focus on three different oracles: rank sum, common independence, and maximum rank oracles. The study of the rank sum oracle is motivated by the above discussed connection to polymatroid matching results. The difficulty of giving an efficient algorithm is that the usual augmenting path approach cannot be applied directly, since the exchangeability graphs are not determined by the rank sum oracle. Still, we are able to give a strongly polynomial time algorithm for the weighted matroid intersection problem by emulating the Bellman--Ford algorithm without explicitly knowing the underlying graph. \begin{theorem} \label{thm:ranksum} There exists a strongly polynomial time algorithm for the weighted matroid intersection problem in the rank sum oracle model. \end{theorem} It is not difficult to see that a common independence oracle can be constructed with the help of a rank sum oracle. Therefore, any algorithm that is based on the usage of a common independence oracle immediately translates into an algorithm that uses a rank sum oracle. Nevertheless, the reverse implication does not hold, hence the common independence oracle is strictly weaker. We show that unweighted matroid intersection remains tractable under the common independence oracle model when one of the matroids is a partition matroid. \begin{theorem} \label{thm:ci} There exists a strongly polynomial time algorithm for the unweighted matroid intersection problem in the common independence oracle model when one of the matroids is a partition matroid with all-one upper bound on the partition classes. \end{theorem} Although the complexity of the problem in general remains an intriguing open question even for the unweighted setting, this seemingly simple case already includes matchings in bipartite graphs and arborescences. Recently, Joswig and Schr\"oter~\cite{joswig2017matroids} introduced the notion of split matroids, a class with distinguished structural properties that generalizes paving matroids. B\'erczi, Kir\'aly, Schwarcz, Yamaguchi and Yokoi~\cite{berczi2022hypergraph} showed that every split matroid can be obtained as the direct sum of a so-called elementary split matroid and uniform matroids, and provided a hypergraph characterization of elementary split matroids. By relying on this characterization, we show that even weighted matroid intersection is tractable in the common independence oracle model when one of the matroids is from this class. \begin{theorem} \label{thm:split} There exists a strongly polynomial time algorithm for the weighted matroid intersection problem in the common independence oracle model when one of the matroids is an elementary split matroid. \end{theorem} We will see that the maximum rank oracle does not carry too much information on its own. However, when combined with the common independence oracle, they are strong enough to mimic our algorithm for the rank sum case. \begin{theorem} \label{thm:cimax} There exists a strongly polynomial time algorithm for the weighted matroid intersection problem when both the common independence and the maximum rank oracles are available. \end{theorem} \paragraph{Organization} The rest of the paper is organized as follows. Basic definitions and notation are introduced in Section~\ref{sec:preliminaries}, together with some fundamental results on matroid intersection. Section~\ref{sec:oracles} describes the relation between different oracles. We present our strongly polynomial algorithm under the rank sum oracle in Section~\ref{sec:rank_sum}. The common independence oracle case when one of the matroids is a partition matroid or an elementary split matroid, as well as the combination of the common independence and maximum rank oracles, is discussed in Section~\ref{sec:ci}. \section{Polynomial Reducibility of Oracles} \label{sec:oracles} Although there are many different types of oracles that are used in studies on matroids, many of these conventional oracles have the same computational power. More precisely, we say that an oracle ${\mathcal{O}}_1$ is \textbf{polynomially reducible} to another oracle ${\mathcal{O}}_2$ if ${\mathcal{O}}_1$ can be simulated by using a polynomial number of oracle calls to ${\mathcal{O}}_2$ measured in terms of the size of the ground set. Two oracles are \textbf{polynomially equivalent} if they are mutually polynomially reducible to each other. In this sense, the rank, independence, strong basis, circuit-finding, spanning, and port oracles are polynomially equivalent~\cite{robinson1980computational,hausmann1981algorithmic,coullard1996independence}. The aim of this section is to clarify the relation between the restricted oracles that we consider, that is, \textsc{Sum}, \textsc{Min}, \textsc{Max}, and \textsc{CI}. As it turns out, \textsc{Max} is not very useful on its own, but it provides a powerful tool when combined with any of the other three oracles. We denote by a `+' sign when we have access to two of the oracles, e.g., \textsc{Min+Max} means that for a set $X\subseteq E$ the oracle tells both ${r_\mathrm{min}}(X)$ and ${r_\mathrm{max}}(X)$. In what follows, we discuss the reducibility of the oracles one by one. In order to keep the presentation clear, some of the ideas appear multiple times. For an overview of the results, see Figure~\ref{fig:oracles}. Observe that, by ${r_\mathrm{sum}}(X)={r_\mathrm{min}}(X)+{r_\mathrm{max}}(X)$, any combination of at least two of \textsc{Min}, \textsc{Max}, and \textsc{Sum} is clearly equivalent. This immediately implies that each of {\sc Min}, {\sc Sum}, and {\sc Max} is reducible to each of {\sc Min+Max}, {\sc Min+Sum}, and {\sc Sum+Max}. \begin{figure}[t!] \centering \includegraphics[width=.5\linewidth]{reductions_new.pdf} \caption{Hierarchy of oracles, directed arcs representing polynomial reducibility. Grey boxes denote oracles for which strongly polynomial time algorithms are given in the present paper.} \label{fig:oracles} \end{figure} \begin{theorem}\label{thm:or1} \textsc{CI} is not polynomially reducible to \textsc{Max}, but it is polynomially reducible to \textsc{Min} and \textsc{Sum}. \end{theorem} \begin{proof} If ${\mathbf{M}}_1$ is the free matroid, then \textsc{Max} always answers $|X|$ independently from the choice of ${\mathbf{M}}_2$. Thus deciding if $X\subseteq E$ is a common independent set or not is impossible relying solely on \textsc{Max}. To see the second half, observe that for a set $X\subseteq E$, \textsc{CI} answers \emph{``Yes''} if and only if $X$ is a common independent set of the two matroids, that is, $r_1(X)=r_2(X)=|X|$. By the subcardinality of the rank functions, this is equivalent to ${r_\mathrm{min}}(X)=|X|$ and to ${r_\mathrm{sum}}(X)=2|X|$. As these conditions can be checked by \textsc{Min} and \textsc{Sum}, respectively, the theorem follows. \end{proof} \begin{theorem}\label{thm:or2} \textsc{Min} is not polynomially reducible to \textsc{Sum}, \textsc{CI}, \textsc{Max}, and \textsc{CI+Max}. \end{theorem} \begin{proof} We define two instances of the matroid intersection problem on the same ground set $E=\{a,b,c,d\}$ as follows. For $i=1,2$, let ${\mathbf{M}}_i$ be the graphic matroid of the graph $G_i$ on Figure~\ref{fig:m1m2a}, and let ${\mathbf{M}}'_i$ be the graphic matroid of the graph $G'_i$ on Figure~\ref{fig:m1m2b}. Consider the maximum-cardinality common independent set problem for ${\mathbf{M}}_1$ and ${\mathbf{M}}_2$, and for ${\mathbf{M}}'_1$ and ${\mathbf{M}}'_2$. For any subset $X$ of $E$, both \textsc{Sum} and \textsc{CI} give the same answer in the two instances, thus it is not possible to distinguish them from each other using one of these oracles. However, ${r_\mathrm{min}}(E)$ is $2$ in one of them while $3$ in the other one, showing that \textsc{Min} cannot be reduced to \textsc{Sum} or \textsc{CI}. Now take the $3$-truncation of these graphic matroids, and define ${\mathbf{N}}_1=({\mathbf{M}}_1)_3$, ${\mathbf{N}}_2=({\mathbf{M}}_2)_3$, ${\mathbf{N}}'_1=({\mathbf{M}}'_1)_3$, and ${\mathbf{N}}'_2=({\mathbf{M}}'_2)_3$. Consider the maximum-cardinality common independent set problem for ${\mathbf{N}}_1$ and ${\mathbf{N}}_2$, and for ${\mathbf{N}}'_1$ and ${\mathbf{N}}'_2$. By the slight change in the definitions, both \textsc{CI} and \textsc{Max} give the same answer in the two instances for any subset $X\subseteq E$, thus it is not possible to distinguish them from each other using a combination of these two oracles. However, ${r_\mathrm{min}}(E)$ is $2$ in one of them while $3$ in the other one, showing that \textsc{Min} cannot be reduced to \textsc{CI+Max}. The case when ${\mathbf{M}}_1$ is the free matroid shows again that ${r_\mathrm{min}}(X)$ cannot be determined relying solely on \textsc{Max}. \end{proof} \begin{figure}[t!] \centering \begin{subfigure}[t]{0.47\textwidth} \centering \includegraphics[width=.75\linewidth]{m1m2a.pdf} \caption{The graphs $G_1$ and $G_2$ defining ${\mathbf{M}}_1$ and ${\mathbf{M}}_2$.} \label{fig:m1m2a} \end{subfigure}\hfill \begin{subfigure}[t]{0.47\textwidth} \centering \includegraphics[width=.75\linewidth]{m1m2b.pdf} \caption{The graphs $G'_1$ and $G'_2$ defining ${\mathbf{M}}'_1$ and ${\mathbf{M}}'_2$.} \label{fig:m1m2b} \end{subfigure} \caption{Illustration of Theorems~\ref{thm:or2}, \ref{thm:or3}, and \ref{thm:or4}.} \label{fig:or2} \end{figure} \begin{theorem}\label{thm:or3} \textsc{Sum} is not polynomially reducible to \textsc{Min}, \textsc{CI}, \textsc{Max}, and \textsc{CI+Max}. \end{theorem} \begin{proof} If ${\mathbf{M}}_1$ is a uniform matroid of rank $1$, then ${r_\mathrm{min}}(X)=0$ if $X=\emptyset$ and $1$ otherwise, while \textsc{CI} answers \emph{``Yes''} if $|X|\leq 1$ and \emph{``No''} otherwise. These answers are independent from the choice of ${\mathbf{M}}_2$, therefore we cannot determine ${r_\mathrm{sum}}(X)$ with their help. The case when ${\mathbf{M}}_1$ is the free matroid shows again that ${r_\mathrm{sum}}(X)$ cannot be determined relying solely on \textsc{Max}. Consider the same two instances of the matroid intersection problem defined by the $3$-truncations of the graphic matroids on Figure~\ref{fig:or2} as in the proof of Theorem~\ref{thm:or2}. Recall that both \textsc{CI} and \textsc{Max} give the same answer in the two instances for any subset $X\subseteq E$, thus it is not possible to distinguish them from each other using a combination of these two oracles. However, ${r_\mathrm{sum}}(E)$ is $5$ in one of them while $6$ in the other one, showing that \textsc{Sum} cannot be reduced to \textsc{CI+Max}. \end{proof} \begin{theorem}\label{thm:or4} \textsc{Max} is not polynomially reducible to \textsc{Min}, \textsc{Sum}, and \textsc{CI}. \end{theorem} \begin{proof} The case when ${\mathbf{M}}_1$ is a uniform matroid of rank $1$ shows again that ${r_\mathrm{max}}(X)$ cannot be determined relying solely on \textsc{Min} or \textsc{CI}. Consider the same two instances of the matroid intersection problem defined by Figure~\ref{fig:or2} as in the proof of Theorem~\ref{thm:or2}. Recall that for any subset $X$ of $E$, \textsc{Sum} gives the same answer in both instances, thus it is not possible to distinguish them from each other using \textsc{Sum}. However, ${r_\mathrm{max}}(E)$ is $4$ in one of them while $3$ in the other one, showing that \textsc{Max} cannot be reduced to \textsc{Sum}. \end{proof} \section{Preliminaries} \label{sec:preliminaries} For the basics on matroids and the matroid intersection problem, we refer the reader to \cite{oxley2011matroid, schrijver2003combinatorial}. Throughout the paper, for $i=1,2$, let ${\mathbf{M}}_i=(E,{\mathcal{I}}_i)$ be loopless\footnote{The assumption that the matroids are loopless is not restrictive as loops can be easily detected by any of the oracles considered.} matroids on the same finite ground set $E$ of size $n$, whose \textbf{independent set families}, \textbf{rank functions}, and \textbf{closure operators} are denoted by ${\mathcal{I}}_i$, by $r_i$, and by $\mathrm{cl}_i$, respectively. For two sets $X,Y\subseteq E$, we denote their \textbf{symmetric difference} by $X\triangle Y=(X\setminus Y)\cup(Y\setminus X)$. The \textbf{$k$-truncation} of a matroid ${\mathbf{M}}=(S,{\mathcal{I}})$ is a matroid $({\mathbf{M}})_{k}=(S,{\mathcal{I}}^{\le k})$ with ${\mathcal{I}}^{\le k}=\{\, X\in{\mathcal{I}} \mid |X|\leq k \,\}$. For $I \in {\mathcal{I}}_i$ and $x \in \mathrm{cl}_i(I) \setminus I$, the \textbf{fundamental circuit} of $x$ with respect to $I$ in ${\mathbf{M}}_i$ is denoted by $C_i(I, x) = \{\, y \in I \mid I + x - y \in {\mathcal{I}}_i \,\}$. We consider four oracles for matroid intersection. Given a set $X\subseteq E$ as an input, a \textbf{rank sum oracle} (\textsc{\textsc{Sum}}) answers the sum ${r_\mathrm{sum}}(X)\coloneqq r_1(X)+r_2(X)$ of the ranks of $X$, a \textbf{minimum rank oracle} (\textsc{Min}) answers the minimum ${r_\mathrm{min}}(X)\coloneqq\min\left\{r_1(X),r_2(X)\right\}$ of the ranks of $X$, a \textbf{maximum rank oracle} (\textsc{Max}) answers the maximum ${r_\mathrm{max}}(X)\coloneqq\max\left\{r_1(X),r_2(X)\right\}$ of the ranks of $X$, and a \textbf{common independence oracle} (\textsc{CI}) answers \emph{``Yes''} if $X\in{\mathcal{I}}_1\cap{\mathcal{I}}_2$ and \emph{``No''} otherwise. Let us first overview some basic results on unweighted matroid intersection. In \cite{edmonds1970submodular}, Edmonds gave the following characterization for the maximum cardinality of a common independent set of two matroids. \begin{theorem}[Edmonds~\cite{edmonds1970submodular}]\label{thm:Edmonds} Given two matroids ${\mathbf{M}}_1=(E,{\mathcal{I}}_1)$ and ${\mathbf{M}}_2=(E,{\mathcal{I}}_2)$ on a common ground set $E$, the maximum cardinality of a common independent set of ${\mathbf{M}}_1$ and ${\mathbf{M}}_2$ is equal to \begin{align*} \min\left\{\, r_1(Z) + r_2(E \setminus Z) \mid Z \subseteq E \,\right\}. \end{align*} \end{theorem} The notion of exchangeability graphs plays a central role in any matroid intersection algorithm. \begin{definition}[Exchangeability Graphs]\label{def:exchange} Assume that $I\in{\mathcal{I}}_1\cap {\mathcal{I}}_2$. The \textbf{exchangeability graph} corresponding to $I$ is a bipartite digraph $D[I] = (E \setminus I, I; A[I])$ defined as follows. Set \begin{align*} S_I &\coloneqq \{\, s \in E \setminus I \mid I + s \in {\mathcal{I}}_1 \,\},\\ T_I &\coloneqq \{\, t \in E \setminus I \mid I + t \in {\mathcal{I}}_2 \,\}, \end{align*} where elements in $S_I$ and in $T_I$ are called \textbf{sources} and \textbf{sinks}, respectively. We then define the set $A[I] \coloneqq A_1[I] \cup A_2[I]$ of exchangeability arcs, where \begin{align*} A_1[I] \coloneqq&\ \{\, (y, x) \mid x \in E \setminus I,~y \in I,~I + x - y \in {\mathcal{I}}_1 \,\}\\ =&\ \{\, (y, s) \mid s \in S_I,~y \in I \,\} \cup \{\, (y, x) \mid x \in E \setminus (I \cup S_I),~y \in C_1(I, x) \,\},\\[1mm] A_2[I] \coloneqq&\ \{\, (x, y) \mid x \in E \setminus I,~y \in I,~I + x - y \in {\mathcal{I}}_2 \,\}\\ =&\ \{\, (t, y) \mid t \in T_I,~y \in I \,\} \cup \{\, (x, y) \mid x \in E \setminus (I \cup T_I),~y \in C_2(I, x) \,\}. \end{align*} Note that $S_I$ and $A_1[I]$ depend only on ${\mathcal{I}}_1$, and $T_I$ and $A_2[I]$ depend only on ${\mathcal{I}}_2$. \end{definition} Brualdi \cite{brualdi1969comments} observed that the set $A_i[I]$ satisfies the following property for $i=1,2$. \begin{lemma}\label{lem:UPM-inv} Let $I\in {\mathcal{I}}_i$ and let $Z\subseteq E$ satisfy $|I\triangle Z|=|I|$ and $I\triangle Z\in {\mathcal{I}}_i$. Then $A_i[I]$ contains a perfect matching on $Z$ (i.e., a set of vertex-disjoint arcs whose tails and heads constitute $Z$). \end{lemma} Krogdahl~\cite{krogdahl1974combinatorial,krogdahl1976combinatorial,krogdahl1977dependence} proved a partial converse to the above lemma. \begin{lemma}[Unique Perfect Matching Lemma]\label{lem:UPM} Let $I\in {\mathcal{I}}_i$ and let $Z\subseteq E$ satisfy $|I\triangle Z|=|I|$. If $A_i[I]$ contains a unique perfect matching on $Z$, then $I\triangle Z\in {\mathcal{I}}_i$. \end{lemma} Finally, let us recall that a standard algorithm for finding a maximum-cardinality common independent set is driven by the following subroutine (see \cite[$\S$~41.2]{schrijver2003combinatorial}). For any digraph $D=(E,A)$, a \textbf{path} in $D$ is a sequence $P=e_1e_2\cdots e_\ell$ of distinct vertices such that $(e_i,e_{i+1})\in A$ for each $i=1,2,\dots,\ell-1$; we call $P$ an \textbf{$e_1$--$e_\ell$ path} or \textbf{an $X$--$Y$ path} for sets $X \ni e_1$ and $Y \ni e_\ell$ to emphasize the end vertices, and define $\ell$ as the \textbf{length}. A \textbf{cycle} in $D$ is a sequence $e_1e_2\cdots e_\ell e_1$ such that $e_1e_2 \cdots e_\ell$ is a path and $(e_\ell, e_1)\in A$. We often identify a path or a cycle with its vertex set $\{e_1, e_2,\dots,e_\ell\}$. \begin{algorithm2e}[h!] \caption{{{\sc Augment}$[E, {\mathcal{I}}_1, {\mathcal{I}}_2, I]$}} \label{alg:1} \SetAlgoLined \SetKwInOut{Input}{Input}\SetKwInOut{Output}{Output} \Input{A finite set $E$, oracle access to ${\mathcal{I}}_1$ and ${\mathcal{I}}_2$, and a common independent set $I \in {\mathcal{I}}_1 \cap {\mathcal{I}}_2$.} \Output{A common independent set $J \in {\mathcal{I}}_1 \cap {\mathcal{I}}_2$ with $|J| = |I| + 1$ if one exists, or a subset $Z \subseteq E$ with $r_1(Z) + r_2(E \setminus Z) = |I|$ otherwise.} \BlankLine Construct the exchangeability graph $D[I]$ with source set $S_I$ and sink set $T_I$. If some $t \in T_I$ is reachable from some $s \in S_I$, then find a shortest $S_I$--$T_I$ path $P$ in $D[I]$, and return $J = I \triangle P$. Otherwise, return $Z = \{\, e \in E \mid e~\text{can reach some}~t \in T_I~\text{in}~D[I] \,\}$. \end{algorithm2e} Now we turn to the weighted setting. For a weight function $w \colon E \to {\mathbb{R}}$ and a subset $X \subseteq E$, define $w(X) \coloneqq \sum_{e \in X} w(e)$. For a family ${\mathcal{F}} \subseteq 2^E$, a subset $X \subseteq E$ is \textbf{$w$-maximal in ${\mathcal{F}}$} if $X \in \textrm{arg\,max}\left\{\, w(Y) \mid Y \in {\mathcal{F}} \,\right\}$. We define ${\mathcal{I}}_i^k \coloneqq \{\, X \in {\mathcal{I}}_i \mid |X| = k \,\}$ for $i = 1, 2$ and $k = 0, 1, \dots, n$. One approach to solve the weighted matroid intersection problem is via augmentation along cheapest paths in the exchangeability graph (see~\cite[$\S$~41.3]{schrijver2003combinatorial}), where the cost function $c \colon E \to {\mathbb{R}}$ is defined on the vertex set as follows: \begin{align}\label{eq:cost} c(e) &\coloneqq \begin{cases} w(e) & \text{if $e \in I$},\\ -w(e) & \text{if $e \in E \setminus I$}. \end{cases} \end{align} For each path (or cycle) $P$ in $D[I]$, we define the \textbf{cost} of $P$ as $c(P) \coloneqq \sum_{e \in P}c(e)$. \begin{algorithm2e}[h] \caption{{{\sc CheapestPathAugment}$[E, w, {\mathcal{I}}_1, {\mathcal{I}}_2, I]$}} \label{alg:2} \SetAlgoLined \SetKwInOut{Input}{Input}\SetKwInOut{Output}{Output} \Input{A finite set $E$, a weight function $w \colon E \to {\mathbb{R}}$, oracle access to ${\mathcal{I}}_1$ and ${\mathcal{I}}_2$, and a $w$-maximal set $I \in {\mathcal{I}}_1^k \cap {\mathcal{I}}_2^k$ for some $k\in\{0, 1, \dots, n - 1\}$.} \Output{A $w$-maximal set $J \in {\mathcal{I}}_1^{k+1} \cap {\mathcal{I}}_2^{k+1}$ if one exists, or a subset $Z \subseteq E$ with $r_1(Z) + r_2(E \setminus Z) = |I|$ otherwise.} \BlankLine Construct the exchangeability graph $D[I]$ with source set $S_I$ and sink set $T_I$. In addition, define the cost function $c \colon E \to {\mathbb{R}}$ by \eqref{eq:cost}. If some $t \in T_I$ is reachable from some $s \in S_I$, then find a shortest cheapest $S_I$--$T_I$ path $P$ in $D[I]$ (i.e., the cost $c(P)$ is minimum, and subject to this, the length of $P$ is minimum), and return $J = I \triangle P$. Otherwise, return $Z = \{\, e \in E \mid e~\text{can reach some}~t \in T_I~\text{in}~D[I] \,\}$. \end{algorithm2e} The next lemma characterizes $w$-maximal common independent sets in ${\mathcal{I}}^k_1\cap {\mathcal{I}}^k_2$. \begin{lemma}[cf.~{\cite[Theorem~41.5]{schrijver2003combinatorial}}]\label{lem:negative_cycle} A common independent set $I \in {\mathcal{I}}_1^k \cap {\mathcal{I}}_2^k$ is $w$-maximal if and only if $D[I]$ contains no negative-cost cycle with respect to the cost function $c$ defined as \eqref{eq:cost}. \end{lemma} \section{Matroid Intersection under Rank Sum Oracle} \label{sec:rank_sum} In Algorithm~\ref{alg:2}, we assume that we are given the independence oracles of matroids ${\mathbf{M}}_1$ and ${\mathbf{M}}_2$, which are polynomially equivalent to the oracles of rank functions $r_1$ and $r_2$, respectively. In this section, we consider the solvability of the weighted matroid intersection problem when only the \emph{rank sum function} ${r_\mathrm{sum}} \colon 2^E \to {\mathbb{Z}}_{\geq 0}$ is available, that is, for any set $X\subseteq E$ the oracle answers the value of ${r_\mathrm{sum}}(X) \coloneqq r_1(X)+r_2(X)$. Note that a subset $I \subseteq E$ belongs to ${\mathcal{I}}_1\cap {\mathcal{I}}_2$ if and only if ${r_\mathrm{sum}}(I) = 2|I|$. However, when ${r_\mathrm{sum}}(I)<2|I|$, we cannot decide whether $I\in {\mathcal{I}}_i$ or not for each $i=1,2$. The matroid intersection problem under this oracle model is exactly a special case of the \emph{polymatroid matching problem}, which is equivalent to the so-called \emph{matroid matching (or parity) problem}~\cite[\S~11.1]{lovasz2009matching}. While a max-min duality theorem was given by Lov\'asz~\cite{lovasz1980matroid} for the case when the matroids in question admit linear representations\footnote{For the matroid intersection considered as a special case, the two matroids are necessarily representable over the same field.}, such a good characterization is not known in general even in the matroid intersection case. In what follows, we provide an algorithm for the weighted matroid intersection problem with the rank sum oracle. We consider emulating Algorithm~\ref{alg:2}, i.e., {\sc CheapestPathAugment}$[E, w, {\mathcal{I}}_1, {\mathcal{I}}_2, I]$. \subsection{Searching a Shortest Cheapest Path} Take any $k \in \{0, 1, \dots, n - 1\}$ and let $I$ be a $w$-maximal set ${\mathcal{I}}_1^k \cap {\mathcal{I}}_2^k$. To emulate Algorithm~\ref{alg:2}, we want to find a shortest cheapest $S_I$--$T_I$ path in $D[I] = (E \setminus I, I; A_1[I]\cup A_2[I])$ with respect to vertex cost $c\colon E \to {\mathbb{R}}$ defined in Algorithm~\ref{alg:2}. With only the rank sum oracle, however, we cannot determine the sets $A_1[I]$, $A_2[I]$, $S_I$, and $T_I$, and hence cannot simply emulate Algorithm~\ref{alg:2}. We show that, despite this difficulty, we can compute a shortest cheapest $S_I$--$T_I$ path in $D[I]$. We first make some observations. Let $D'[I]=(E \setminus I, I; A'_1[I]\cup A'_2[I])$ be the subgraph of $D[I]$ obtained from $D[I]$ by removing arcs entering $S_I$ and leaving $T_I$, i.e., \begin{align*} A'_1[I] \coloneqq&\{\, (y, x) \mid x \in E \setminus (I \cup S_I),~y \in C_1(I, x) \,\},\\ A'_2[I] \coloneqq&\{\, (x, y) \mid x \in E \setminus (I \cup T_I),~y \in C_2(I, x) \,\}. \end{align*} Note that each element in $S_I\cap T_I$ is an isolated vertex in $D'[I]$. Recall that $D[I]$ has no negative cost cycle with respect to $c$ by Lemma~\ref{lem:negative_cycle}. This fact implies that finding a shortest cheapest $S_I$--$T_I$ path in $D[I]$ is equivalent to finding one in $D'[I]$. \begin{lemma}\label{lem:shortest-cheapest-path} Any shortest cheapest $S_I$--$T_I$ path in $D'[I]$ is a shortest cheapest $S_I$--$T_I$ path in $D[I]$. \end{lemma} \begin{proof} It is sufficient to show that any shortest cheapest $S_I$--$T_I$ path in $D[I]$ is contained in $D'[I]$. Suppose, to the contrary, that a shortest cheapest $S_I$--$T_I$ path $P$ uses some arc $(y,s^*)\in A_1[I]\setminus A'_1[I]$ or $(t^*,y)\in A_2[I]\setminus A'_2[I]$, and let $s \in S_I$ and $t \in T_I$ be its end vertices. If $P$ uses $(y,s^*)\in A_1[I]\setminus A'_1[I]$, let $P(s,y)$ and $P(s^*,t)$ be the subpaths of $P$ from $s$ to $y$ and from $s^*$ to $t$, respectively. Since $(y,s)\in A_1[I]$, the path $P(s,y)$ is extended to a cycle with the same vertex set in $D[I]$, and hence $c(P(s,y))$ is nonnegative by Lemma~\ref{lem:negative_cycle}. Then $c(P)=c(P(s,y))+c(P(s^*,t))\geq c(P(s^*,t))$ and $|P(s^*,t)|<|P|$, which contradicts that $P$ is a shortest cheapest $S_I$--$T_I$ path in $D[I]$. If $P$ uses $(t^*,y)\in A_2[I]\setminus A'_2[I]$, we can similarly show that $P(s,t^*)$ is an $S_I$--$T_I$ path that is at least as cheap as $P$ and shorter than $P$. \end{proof} While we cannot determine $S_I$ and $T_I$, we can determine $S_I\cup T_I$ as $S_I\cup T_I=\{\,s\in E\setminus I\mid {r_\mathrm{sum}}(I+s)\geq 2|I|+1\,\}$. We now provide a search algorithm with the rank sum oracle. Its description is given as Algorithm~\ref{alg:3}. For any $s\in S_I$ (resp., $s\in T_I$), it emulates the Bellman--Ford algorithm in $D'[I]$ (resp., in the inverse of $D'[I]$) rooted at $s$ without knowing $D'[I]$ explicitly. Since there is no negative cost cycle in $D'[I]$ by Lemma~\ref{lem:negative_cycle}, the algorithm finds shortest cheapest paths from $s$ to all reachable vertices, and it returns a shortest cheapest $s$--$T_I$ path (resp., $s$--$S_I$ path) if it exists. By applying this search algorithm for all $s\in S_I\cup T_I$, we can obtain a shortest cheapest $S_I$--$T_I$ path in $D'[I]$, which is also a shortest cheapest path in $D[I]$ by Lemma~\ref{lem:shortest-cheapest-path}. For each $e\in E$, the algorithm maintains a sequence $P_e$ of distinct elements of $E$. We will show that $P_e$ is an $s$--$e$ path in $D'[I]$. For a sequence $P_e$ and an element $e'\in E\setminus P_e$, we denote by $P_e+e'$ the sequence obtained by appending $e'$ to $P_e$. \begin{algorithm2e} \caption{{\sc EmulatingBellmanFord}$[E, c, {r_\mathrm{sum}}, I, s]$} \label{alg:3} \SetAlgoLined \SetKwInOut{Input}{Input}\SetKwInOut{Output}{Output} \Input{A finite set $E$, a weight function $w \colon E \to {\mathbb{R}}$, which defines $c \colon E \to {\mathbb{R}}$ by \eqref{eq:cost}, oracle access to ${r_\mathrm{sum}} \colon 2^E \to {\mathbb{Z}}_{\geq 0}$, a $w$-maximal set $I \in {\mathcal{I}}_1^k\cap {\mathcal{I}}_2^k$ for some $k \in\{ 0, 1, \dots, n - 1\}$, and $s\in S_I\cup T_I$.} \Output{For $s\in S_I$ (resp., $s\in T_I$), a shortest cheapest $s$--$T_I$ path (resp., $s$--$S_I$ path) in $D'[I]$ (resp, in the inverse of $D'[I]$) with respect to $c$ if one exists, or a message \emph{``No''} otherwise.} \BlankLine Set $P_s \gets s$, and $P_e \gets {\mathsf{null}}$ for each $e\in E-s$. For $\ell=1,2,\dots,n-1$, do the following. \protect{ \begin{description} \item[If $\ell$ is odd:] For each $y\in I$, do the following. \begin{itemize} \item Let $x\in E\setminus I$ minimize $c(P_x)$ subject to $P_x\neq {\mathsf{null}}$, $y\not\in P_x$, and \begin{align*} (*)\quad{r_\mathrm{sum}}(I\triangle P_x)=2|I|+1,\quad{r_\mathrm{sum}}(I\triangle(P_x+y))=2|I|.&& \end{align*} \item If $c(P_x+y)<c(P_y)$, update $P_y\gets P_x+y$. \end{itemize} \item[If $\ell$ is even:] For each $x\in E\setminus I$, do the following. \begin{itemize} \item Let $y\in I$ minimize $c(P_y)$ subject to $P_y\neq {\mathsf{null}}$, $x\not\in P_y$, and \begin{align*} (**)\quad&[~{r_\mathrm{sum}}(I+x)=2|I|,~{r_\mathrm{sum}}(I\triangle (P_y+x))=2|I|+1~]~\text{or}&&\\ &[~{r_\mathrm{sum}}(I+x)=2|I|+1,~ {r_\mathrm{sum}}(I\triangle (P_y+x))=2|I|+2~].&& \end{align*} \item If $c(P_y+x)<c(P_x)$, update $P_x\gets P_y+x$. \end{itemize} \end{description} } Let $t\in E\setminus I$ minimize $c(P_t)$ subject to ${r_\mathrm{sum}}(I\triangle P_t)=2|I|+2$. Return $P_t$ if $c(P_t)\neq\infty$, and otherwise return \emph{``No''}. \end{algorithm2e} \subsection{Correctness of the Search} The following lemma shows an important property of Algorithm~\ref{alg:3}, where we can assume that $s$ belongs to $S_I = \{\, s \in E \setminus I \mid I + s \in {\mathcal{I}}_1 \,\}$ by symmetry. (For $s\in T_I = \{\, s \in E \setminus I \mid I + s \in {\mathcal{I}}_2 \,\}$, replace $D'[I]$ with its inverse in the arguments.) \begin{lemma}\label{lem:invariants2} Let $s\in S_I$ and take any $\ell=1,2,\dots,n-1$. For any $e\in E$, just after the $\ell$th updating process of Step~2 of Algorithm~\ref{alg:3}, the sequence $P_e$ is a shortest cheapest $s$--$e$ path in $D'[I]$ subject to $|P_e|\leq \ell+1$, where $P_e={\mathsf{null}}$ means that there is no such path. \end{lemma} \begin{proof} We use induction on $\ell$. Note that the statement holds if $\ell=0$. We show the statement for any $\ell>0$ assuming that it holds for $\ell-1$. We use the following two claims. \begin{claim}\label{claim:equivalence1} Suppose that, for any $e$, $P_e$ is a shortest cheapest $s$--$e$ path in $D'[I]$ subject to $|P_e|\leq \ell$. Then, for any $y\in I$ and $x\in E\setminus I$ such that [$P_x\neq {\mathsf{null}}$, $y\not\in P_x$, and $c(P_x+y)<c(P_y)$], condition $(\ast)$ holds if and only if $(x,y)\in A'_2[I]$. \end{claim} \begin{claim}\label{claim:equivalence2} Suppose that, for any $e$, $P_e$ is a shortest cheapest $s$--$e$ path in $D'[I]$ subject to $|P_e|\leq \ell$. Then, for any $x\in E\setminus I$ and $y\in I$ such that [$P_y\neq{\mathsf{null}}$, $x\not\in P_y$, and $c(P_y+x)<c(P_x)$], condition $(\ast\ast)$ holds if and only if $(y,x)\in A'_1[I]$. \end{claim} We postpone the proofs of these claims and complete the proof of the lemma relying on them. For any $e\in E$, let $P^{\ell-1}_e$ and $P^{\ell}_e$ be the sequence $P_e$ just after the $(\ell-1)$th and $\ell$th process, respectively. By induction, $P^{\ell-1}_e$ is a shortest cheapest $s$--$e$ path in $D'[I]$ subject to $|P_e|\leq \ell$. Let $P^*_e$ be any shortest cheapest $s$--$e$ path in $D'[I]$ subject to $|P^*_e|\leq \ell+1$. By the ``only if'' parts of Claims~\ref{claim:equivalence1} and~\ref{claim:equivalence2}, $P^{\ell}_e$ is an $s$--$e$ path in $D'[I]$ with $|P^{\ell}_e|\leq \ell+1$, and hence $c(P^*_e)\leq c(P^{\ell}_e)$. Also, $c(P^{\ell}_e)\leq c(P^{\ell-1}_e)$ by the algorithm. If $c(P^*_e)=c(P^{\ell-1}_e)$, then $P^{\ell}_e=P^{\ell-1}_e$ and the statement immediately follows. Otherwise, $c(P^*_e)<c(P^{\ell-1}_e)$. This implies $|P^*_e|=\ell+1$. Then $e\in I$ if $\ell$ is odd and $e\in E\setminus I$ if $e$ is even (recall that $D'[I]$ is a bipartite digraph between $E\setminus I$ and $I$). Let $e'$ be the second last element in $P^*_e$ and let $P^*_{e'}:=P^*_{e}-e$ (i.e., delete $e$ from $P^*_e$). Then $P^*_{e'}$ is an $s$--$e'$ path with $|P_{e'}|=\ell$, and hence $c(P^{\ell-1}_{e'})\leq c(P^*_{e'})$. So $c(P^{\ell-1}_{e'}+e)\leq c(P^*_e)<c(P^{\ell-1}_e)$. If $\ell$ is odd, then $(e',e)\in A'_2[I]$, and hence ($\ast$) holds with $x:=e'$ and $y:=e$ by Claim~\ref{claim:equivalence1}. If $\ell$ is even, then $(e',e)\in A'_1[I]$, and hence ($\ast\ast$) holds with $y:=e'$ and $x:=e$ by Claim~\ref{claim:equivalence2}. In either case, we obtain $c(P^\ell_e)\leq c(P^*_{e'}+e)=c(P^*_e)$ and $|P^{\ell}_e|\leq \ell+1=|P^*_e|$. Hence $P^\ell_e$ is a shortest cheapest $s$--$e$ path subject to $|P_e|\leq \ell+1$. \end{proof} In what follows, we prove Claims~\ref{claim:equivalence1} and~\ref{claim:equivalence2}. First, we need the following lemma. \begin{lemma}\label{lem:invariants1} Let $s\in S_I$. At any moment of the algorithm, the following conditions hold. \begin{enumerate} \item[\rm (a)] Any $y\in I$ with $P_y\neq {\mathsf{null}}$ satisfies $|I\triangle P_y|=|I|$, $r_1(I\triangle P_y)=|I|$, and $r_2(I\triangle P_y)=|I|$. \item[\rm (b)] Any $x\in E\setminus I$ with $P_x\neq {\mathsf{null}}$ satisfies $|I\triangle P_x|=|I|+1$, $r_1(I\triangle P_x)=|I|+1$, and $r_2(I\triangle P_x)\geq|I|$. Moreover, it satisfies $r_2(I\triangle P_x)=|I|+1$ if and only if $x\in T_I$. \item[\rm (c)] For each $e\in E$ with $P_e \neq {\mathsf{null}}$, we have $(P_e-s)\cap S_I=\emptyset$ and $(P_e-e)\cap T_I=\emptyset$. \end{enumerate} \end{lemma} \begin{proof} By the algorithm, for each $e\in E$ with $P_e\neq {\mathsf{null}}$, the sequence $P_e$ starts with $s\in E\setminus I$ and uses elements in $E\setminus I$ and $I$ alternately. Then $|I\triangle P_y|=|I|$ for any $y\in I$ with $P_y\neq{\mathsf{null}}$ and $|I\triangle P_x|=|I|+1$ for any $x\in E\setminus I$ with $P_x\neq{\mathsf{null}}$. For any $y\in I$, after $P_y$ is updated, it satisfies ${r_\mathrm{sum}}(I\triangle P_y)=2|I|$ by the condition ($\ast$) for update. Then (a) follows. For any $e\in E\setminus I$ with $P_e \neq {\mathsf{null}}$, any $x'\in (P_e\setminus I)-e$ has some succeeding element $y'\in I$ in $P_e$ and ($\ast$) holds for $x'$ and $y'$. Hence ${r_\mathrm{sum}}(I\triangle P_{x'})=2|I|+1$. If $x'=s\in S_I$, it immediately implies $x'\not\in T_I$. If $x'\neq s$, then $x'$ has some preceding element $y''$ in $P_x$, and ($\ast\ast$) for $y''$ and $x'$ implies ${r_\mathrm{sum}}(I+x')=2|I|$, and hence $x'\not\in S_I\cup T_I$. Thus, $s\in S_I\setminus T_I$ and any $x'\in (P_e\setminus I)-s-e$ satisfies $x'\not\in S_I\cup T_I$. For any $x\in (E\setminus I)-s$ with $P_x\neq {\mathsf{null}}$, by ($\ast\ast$) for $x$ and its preceding element $y$, we have $[{r_\mathrm{sum}}(I+x)=2|I|, {r_\mathrm{sum}}(I\triangle P_x)=2|I|+1]$ or $[{r_\mathrm{sum}}(I+x)=2|I|+1, {r_\mathrm{sum}}(I\triangle P_x)=2|I|+2]$. In the former case, ${r_\mathrm{sum}}(I+x)=2|I|$ implies $x\not\in S_I\cup T_I$, and hence $P_x\cap T_I=\emptyset$. This implies $r_2(I\triangle P_x)\leq r_2(I \cup P_x) = |I|$, and then ${r_\mathrm{sum}}(I\triangle P_x)=2|I|+1$ implies $r_1(I\triangle P_x)=|I|+1$ and $r_2(I\triangle P_x)=|I|$. In the latter case, ${r_\mathrm{sum}}(I\triangle P_x)=2|I|+2$ implies $r_1(I\triangle P_x)=r_2(I\triangle P_x)=|I|+1$. Since any $x' \in (P_x \setminus I) - x$ satisfies $x' \not\in T_I$ (as seen in the previous paragraph), we must have $x\in T_I$, and then $x\not\in S_I$ follows from ${r_\mathrm{sum}}(I+x)=2|I|+1$. Thus, (b) and (c) are shown. \end{proof} Now we are ready to show Claims~\ref{claim:equivalence1} and~\ref{claim:equivalence2}. We sometimes denote by $V(P)$ the set of elements in a sequence $P$ for emphasizing that we focus on the set rather than the sequence. \begin{proof}[Proof of Claim~\ref{claim:equivalence1}] By Lemma~\ref{lem:invariants1}(b), ${r_\mathrm{sum}}(I\triangle P_x)=2|I|+1$ is equivalent to $x\not\in T_I$. Lemma~\ref{lem:invariants1}(b) also implies $r_1(I\triangle P_x)=|I|+1=|I\triangle P_x|$, and hence $r_1(I\triangle (P_x+y))=r_1((I\triangle P_x)-y)=|I|$. Therefore, ${r_\mathrm{sum}}(I\triangle (P_x+y))=2|I|$ is equivalent to $r_2(I\triangle (P_x+y))=|I|=|I\triangle (P_x+y)|$, i.e., $I\triangle (P_x+y)\in {\mathcal{I}}_2$. Then, the condition $(\ast)$ is equivalent to \begin{align*} (*)'\quad x \not\in T_I,\quad I \triangle (P_x + y) \in {\mathcal{I}}_2. \end{align*} We show that $(\ast)'$ holds if and only if $(x,y)\in A'_2[I]$. By induction hypothesis, $P_x$ is an $s$--$x$ path in $D'[I]$, and hence it uses arcs of $A'_2[I]$ and $A'_1[I]$ alternately. Let $N_1$ and $N_2$ be the sets of those arcs of $A'_1[I]$ and $A'_2[I]$, respectively. Since $x\in E\setminus I$, $N_1$ forms a matching that covers $V(P_x)-s$ and $N_2$ forms a matching that covers $V(P_x)-x$. To show the ``if'' part, suppose $(x,y)\in A'_2[I]$. Then $x\not\in T_I$ by the definition of $A'_2[I]$. Also, $N'_2:=N_2+(x,y)\subseteq A'_2[I]$ forms a perfect matching on $V(P_x+y)$. Suppose conversely that $I\triangle (P_x+y)\not\in {\mathcal{I}}_2$. Then Lemma~\ref{lem:UPM} implies that $A_2[I]$ contains some other perfect matching $N''_2$ on $V(P_x+y)$. We see that $N''_2\subseteq A'_2[I]$ because $(P_x+y)\cap T_I=\emptyset$ by Lemma~\ref{lem:invariants1}(c) and $x\not\in T_I$. Thus, $N'_2$, $N''_2$, and $N_1$ are all contained in $D'[I]$. Consider the digraph $D=(V(P_x+y), A)$ whose arc set $A$ consists of the arcs in $N'_2$, $N''_2$, and two copies of $N_1$, where we consider their multiplicity, i.e., each arc in $(N'_2 \cap N''_2) \cup N_1$ is taken twice (parallel). Since $N''_2\neq N'_2$, there exists an arc in $N''_2$ whose head precedes its tail on the path $P_x+y$. Then $D$ contains at least one directed cycle. The indegree and outdegree of each vertex in $D$ are given as $(d^{\rm in}(s),d^{\rm out}(s))=(0,2)$, $(d^{\rm in}(y),d^{\rm out}(y))=(2,0)$, and $(d^{\rm in}(e),d^{\rm out}(e))=(2,2)$ for all the other vertices $e$. Then $A$ can be decomposed into the arc sets of two $s$--$y$ paths and one or more cycles. Let $P_1$ and $P_2$ be those $s$--$y$ paths and $\mathcal{Q}$ be the set of those cycles (where paths and cycles are sequences of vertices). Then each vertex in $V(P_x+y)$ is used exactly twice in this decomposition, and hence \[\textstyle{c(P_1)+c(P_2)+\sum_{Q \in \mathcal{Q}}c(Q)=2c(P_x+y).}\] Since $D'[I]$ has no negative cycle, this implies $c(P_1)+c(P_2)\leq 2c(P_x+y)$. Also, $\mathcal{Q}\neq \emptyset$ implies $V(P_1)\subsetneq V(P_x+y)$ or $V(P_2)\subsetneq v(P_x+y)$. In case $V(P_1)=V(P_x+y)$, we have $V(P_2)\subsetneq V(P_x+y)$, which implies $|P_2|<|P_x+y|\leq\ell+1$ because $|P_x|\leq \ell$ holds by induction. Also, $V(P_1)=V(P_x+y)$ implies $c(P_2)\leq c(P_x+y)$, where $c(P_x+y)<c(P_y)$ by assumption. Thus, $P_2$ is an $s$--$y$ path in $D'[I]$ with $c(P_2)<c(P_y)$ and $|P_2|\leq \ell$, which contradicts the induction hypothesis that $P_y$ is a shortest cheapest $s$--$y$ path subject to $|P_y|\leq \ell$. The case $V(P_2)=V(P_x+y)$ is similar. In case $V(P_1)\subsetneq V(P_x+y)$ and $V(P_2)\subsetneq V(P_x+y)$, both $P_1$ and $P_2$ satisfy $|P_i|\leq \ell$ and at least one of them, say $P_i$, satisfies $c(P_i)\leq c(P_x+y)<c(P_y)$, which again contradicts the induction hypothesis on $P_y$. We next show the ``only if'' part. Let $(\ast)'$ hold. By $I\triangle (P_x+y)\in {\mathcal{I}}_2$, Lemma~\ref{lem:UPM-inv} implies that $A_2[I]$ contains a perfect matching $N'_2$ on $V(P_x+y)$. Also, $N'_2\subseteq A'_2[I]$ because $(P_x+y)\cap T_I=\emptyset$ by Lemma~\ref{lem:invariants1}(c) and $x\not\in T_I$. Thus, $N_2$, $N'_2$, and $N_1$ are all contained in $D'[I]$. Consider the digraph $D^*=(V(P_x+y), A^*)$ whose arc set $A^*$ consists of the arcs in $N_2$, $N'_2$, and two copies of $N_1$, where we consider their multiplicity as before. Conversely, suppose that $(x,y)\not\in A'_2[I]$. Then $(x,y)\not\in N'_2$. Since $N'_2$ covers $V(P_x+y)$, it has an arc whose tail is $x$ and whose head precedes $x$ in the path $P_x+y$. Then $D^*$ contains at least one directed cycle. Note that $(d^{\rm in}(s),d^{\rm out}(s))=(0,2)$, $(d^{\rm in}(x),d^{\rm out}(x))=(2,1)$, $(d^{\rm in}(y),d^{\rm out}(y))=(1,0)$, and $(d^{\rm in}(e),d^{\rm out}(e))=(2,2)$ for all the other vertices $e$ in $D^*$. Then $A^*$ can be decomposed into the arc sets of one $s$--$x$ path, one $s$--$y$ path, and one or more cycles. Let $R_x$, $R_y$, and $\mathcal{Q}'$ be that $s$--$x$ path, $s$--$y$ path, and the set of cycles, respectively. Then each vertex in $V(P_x+y)-y$ is used twice and $y$ is used once in this decomposition, and hence \[\textstyle{c(R_x)+c(R_y)+\sum_{Q \in \mathcal{Q}'}c(Q)=c(P_x)+c(P_x+y).}\] Since $D'[I]$ has no negative cost cycle, $c(R_x)+c(R_y)\leq c(P_x)+c(P_x+y)$. Also, $\mathcal{Q}\neq \emptyset$ implies $V(R_x)\subsetneq V(P_x)$ or $V(R_y)\subsetneq v(P_x+y)$. If $V(R_x)=V(P_x)$, then $V(R_y)\subsetneq V(P_x+y)$, which implies $|R_y|<|P_x+y|\leq \ell+1$. Also, $V(R_x)=V(P_x)$ implies $c(R_y)\leq c(P_x+y)<c(P_y)$. Thus, $R_y$ is an $s$--$y$ path with $c(R_y)<c(P_y)$ and $|R_y|\leq \ell$, which contradicts the induction hypothesis on $P_y$. In case $V(R_y)=V(P_x+y)$, we have $V(R_x)\subsetneq V(P_x)$, which implies $|R_x|<|P_x|$. Also, $V(R_y)=V(P_x+y)$ implies $c(R_x)\leq c(P_x)$. Thus, $R_x$ is an $s$--$x$ path with $c(R_x)\leq c(P_x)$ and $|R_x|<|P_x|$, which contradicts the induction hypothesis on $P_x$. In case $V(R_x)\subsetneq V(P_x)$ and $V(R_y)\subsetneq V(P_x+y)$, we have $|R_x|<|P_x|$ and $|R_y|\leq \ell$. Also, we have $c(R_x)\leq c(P_x)$ or $c(R_y)\leq c(P_x+y)<c(P_y)$, and each of them yields a contradiction. \end{proof} \begin{proof}[Proof of Claim~\ref{claim:equivalence2}] Since $I\triangle (P_y+x)=(I\triangle P_y)+x$, Lemma~\ref{lem:invariants1}(a) implies $|I\triangle (P_y+x)|=|I|+1$, $r_1(I\triangle (P_y+x))\geq |I|$, and $r_2(I\triangle (P_y+x))\geq |I|$. Also, by Lemma~\ref{lem:invariants1}(c), we have $P_y\cap T_I=\emptyset$ (which implies $\mathrm{cl}_2(I)=\mathrm{cl}_2(I\triangle P_y)$), and hence $r_2(I\triangle (P_y+x))=|I|+1$ holds if and only if $x\in T_I$. If $x\not\in T_I$ (resp., $x\in T_I$), then ${r_\mathrm{sum}}(I\triangle (P_y+x))=2|I|+1$ (resp., ${r_\mathrm{sum}}(I\triangle (P_y+x))=2|I|+2$) is equivalent to $r_1(I\triangle (P_y+x))=|I|+1$, and also ${r_\mathrm{sum}}(I+x)=2|I|$ (resp., ${r_\mathrm{sum}}(I+x)=2|I|+1$) is equivalent to $x\not\in S_I$. Therefore, $(\ast\ast)$ is equivalent to \begin{align*} (**)'\quad x\not\in S_I,\quad I\triangle (P_y+x)\in {\mathcal{I}}_1. \end{align*} We show that $(\ast\ast)'$ holds if and only if $(y,x)\in A'_1[I]$. By induction hypothesis, $P_y$ is an $s$--$y$ path in $D'[I]$, and hence it uses arcs of $A'_2[I]$ and $A'_1[I]$ alternately. Let $N_1$ and $N_2$ be the sets of those arcs of $A'_1[I]$ and $A'_2[I]$, respectively. Then $N_1$ forms a matching that covers $V(P_y)-y-s$ and $N_2$ forms a matching that covers $V(P_y)$. To show the ``if'' part, suppose $(y,x)\in A'_1[I]$. Then $x\not\in S_I$ by the definition of $A'_1[I]$. Also, $N'_1:=N_1+(y,x)\subseteq A'_1[I]$ forms a perfect matching on $V(P_y+x)-s$. Suppose conversely that $I\triangle (P_y+x)\not\in {\mathcal{I}}_1$. It implies $I\triangle(P_y+x-s)\not\in{\mathcal{I}}_1$ as follows. Lemma~\ref{lem:invariants1}(c) implies $(P_y+x)\cap S_I=\{s\}$, and hence $\mathrm{cl}_1(I\triangle (P_y+x-s))\subseteq \mathrm{cl}_1(I)$. If $\mathrm{cl}_1(I\triangle (P_y+x-s)) = \mathrm{cl}_1(I)$, then $\mathrm{cl}_1(I \triangle (P_y + x)) = \mathrm{cl}_1(I + s)$, and $I \triangle (P_y + x) \in {\mathcal{I}}_1$ as $I + s \in {\mathcal{I}}_1$, a contradiction. Thus, we have $\mathrm{cl}_1(I\triangle (P_y+x-s)) \subsetneq \mathrm{cl}_1(I)$, which implies $I \triangle (P_y + x - s) \not\in {\mathcal{I}}_1$. Then Lemma~\ref{lem:UPM} implies that $A_1[I]$ contains some other perfect matching $N''_1$ on $V(P_y+x)-s$. We see that $N''_1\subseteq A'_1[I]$ because $(P_y+x-s)\cap S_I=\emptyset$ by Lemma~\ref{lem:invariants1}(b) and $x\not\in S_I$. Thus, $N'_1$, $N''_1$, and $N_2$ are all contained in $D'[I]$. Consider the digraph $D=(V(P_x+y), A)$ whose arc set $A$ consists of the arcs in $N'_1$, $N''_1$, and two copies of $N_2$, where we consider their multiplicity as before. Similarly to the proof of Claim~\ref{claim:equivalence1}, we see that there exists an $s$--$y$ path $P$ in $D'[I]$ with $c(P)<c(P_x)$ and $|P|\leq \ell$, which contradicts the induction hypothesis on $P_x$. We next show the ``only if'' part. Let $(\ast\ast)'$ hold. By $I\triangle (P_y+x-s)\subseteq I\triangle (P_y+x) \in {\mathcal{I}}_1$, Lemma~\ref{lem:UPM-inv} implies that $A_1[I]$ contains a perfect matching $N'_1$ on $V(P_y+x)-s$. Also, $N'_1\subseteq A'_1[I]$ because $(P_y+x-s)\cap S_I=\emptyset$ by Lemma~\ref{lem:invariants1}(c) and $x\not\in S_I$. Thus, $N_1$, $N'_1$, and $N_2$ are all contained in $D'[I]$. Consider the digraph $D^*=(V(P_x+y), A^*)$ whose arc set $A^*$ consists of the arcs in $N_1$, $N'_1$, and two copies of $N_2$, where we consider their multiplicity as before. Then, similarly to the proof of Claim~\ref{claim:equivalence1}, we see that there exists an $s$--$y$ or $s$--$x$ path whose property contradicts the induction hypothesis on $P_y$ or $P_x$. \end{proof} \begin{lemma}\label{lem:BellmanFord} The output of {\sc EmulatingBellmanFord}$[E, c, {r_\mathrm{sum}}, I, s]$ is always correct. \end{lemma} \begin{proof} For any $e\in E$, a shortest cheapest $s$--$e$ path $P$ satisfies $|P|\leq n$. Then, after the $(n-1)$th updating process, $P_e$ is indeed a shortest cheapest $s$--$e$ path by Lemma~\ref{lem:invariants2}. Also, when some path is returned, it is a shortest cheapest $s$--$T_I$ path by Lemma~\ref{lem:invariants1}(b). \end{proof} \subsection{Matroid Intersection Algorithm under Rank Sum Oracle} Using Algorithm~\ref{alg:3} as a subroutine, we can emulate {{\sc CheapestPathAugment}$[E, w, {\mathcal{I}}_1, {\mathcal{I}}_2, I]$} as Algorithm~\ref{alg:4}. \begin{algorithm2e} \caption{{{\sc CheapestPathAugmentRankSum}$[E, w, {\mathcal{I}}_1, {\mathcal{I}}_2, I]$}} \label{alg:4} \SetAlgoLined \SetKwInOut{Input}{Input}\SetKwInOut{Output}{Output} \Input{A finite set $E$, a weight function $w \colon E \to {\mathbb{R}}$, oracle access to ${r_\mathrm{sum}} \colon 2^E \to {\mathbb{Z}}_{\geq 0}$, and a $w$-maximal set $I \in {\mathcal{I}}_1^k \cap {\mathcal{I}}_2^k$ for some $k = 0, 1, \dots, n - 1$.} \Output{A $w$-maximal set $J \in {\mathcal{I}}_1^{k+1} \cap {\mathcal{I}}_2^{k+1}$ if one exists, or a message \emph{``No''}.} \BlankLine Determine the set $S_I\cup T_I=\{\,s\in E\setminus I\mid {r_\mathrm{sum}}(I+s)\geq 2|I|+1\,\}$. Define a cost function $c \colon E \to {\mathbb{R}}$ by \eqref{eq:cost}. For each $s\in S_I\cup T_I$, apply {\sc EmulatingBellmanFord}$[E, c, {r_\mathrm{sum}}, I, s]$. If some path is returned in Step 2, then let $P$ be a shortest cheapest one among all returned paths, and return $J = I \triangle P$. Otherwise, return a message \emph{``No''}. \end{algorithm2e} The following theorem concludes the section. \begin{lemma}\label{lem:cpars} The output of {\sc CheapestPathAugmentRankSum}$[E, w, {\mathcal{I}}_1, {\mathcal{I}}_2, I]$ is always correct. \end{lemma} \begin{proof} The correctness of Algorithm~\ref{alg:4} immediately follows from Lemmas~\ref{lem:shortest-cheapest-path} and~\ref{lem:BellmanFord}. \end{proof} Lemma~\ref{lem:cpars} completes the proof of Theorem~\ref{thm:ranksum}. \begin{proof}[Proof of Theorem~\ref{thm:ranksum}] Starting from $I=\emptyset$, the size of the common independent set can be gradually increased using {\sc CheapestPathAugmentRankSum}$[E, w, {\mathcal{I}}_1, {\mathcal{I}}_2, I]$ until $I$ becomes a common independent set of maximum cardinality. The correctness of the algorithm follows from Lemma~\ref{lem:cpars}. Note that, if we are asked to find a maximum-weight common independent set, then it suffices to output one with maximum weight among the obtained $w$-maximal common independent sets. \end{proof}

b5591bc18f640bff9ebb077b72ba7bd206354f43

\section{Introduction } \label{intro} Everywhere irregular signals are ubiquitous in nature: Classical examples are supplied by natural phenomena (hydrodynamic turbulence \cite{H18}, geophysics, natural textures \cite{Johnson2014}), physiological data (medical imaging \cite{Arneodo2003a}, heartbeat intervals \cite{ABRY:2010:A}, E.E.G \cite{EMBC19_Brain}); they are also present in human activity and technology (finance \cite{Bacry2010b}, internet traffic \cite{MandMemor}, repartition of population \cite{frankhauser1998fractal,SemecurbeHuman} , text analysis \cite{Leonarduzzi2017ICASSP}, art \cite{Wendt2013a}). The analysis of such phenomena requires the modelling by everywhere irregular functions, and it is therefore natural to use mathematical regularity parameters in order to classify such data, and to study mathematical models which would fit their behavior. Constructing and understanding the properties of such functions has been a major challenge in mathematical analysis for a long time: Shortly after Cauchy gave the proper definition of a continuous function, the question of determining if a continuous function is necessarily differentiable at some points was a major issue for a large part of the 19th century; though a first counterexample was found by Bolzano, his construction remained unknown from the mathematical community, and it was only in 1872, with the famous Weierstrass functions \begin{equation} \label{weier1} { \cal W}_{a, \omega} (x) = \sum_{n =0}^{+ \infty} \frac{\sin ( a^n x)}{a^{\omega n}} \hspace{15mm} \mbox{for} \hspace{6mm} a >1 \hspace{6mm} \mbox{and } \hspace{6mm} \omega \in (0,1), \end{equation} that the problem was settled. However, such constructions were considered as weird counterexamples, and not representative of what is commonly met, both in mathematics and in applications. In 1893, Charles Hermite wrote to Thomas Stieltjes: { \em I turn my back with fright and horror to this lamentable plague: continuous functions without derivative}. The first statement that smooth or piecewise smooth functions were not adequate for modelling natural phenomena but were rather exceptional came from physicists, see e.g. the introduction of the famous book of Jean Perrin ``Les atomes'', published in 1913. On the mathematical side, the evolution was slow: In 1931, Mazurkiewicz and Banach showed that most continuous functions are nowhere differentiable (``most'' meaning here that such functions form a residual set in the sense of Baire categories). This spectacular result changed the perspective: Functions which were considered as exceptional and rather pathological actually were the common rule, and smooth functions turn out to be exceptional. A first purpose of multifractal analysis is to supply mathematical notions which allow to quantify the irregularity of functions, and therefore yield quantitative tools that can be applied to real life data in order to determine if they fit a given model, and, if it is the case, to determine the correct parameters of the model. One can also be more ambitious and wonder which ``types'' of singularities are present in the data, which may yield an important information of the nature of the signal; a typical example is supplied by { \em chirps} which are singularities which behave like \begin{equation} \label{chirp} g (x) = | x-x_0|^\alpha \cos \left(\displaystyle\frac{1}{| x-x_0|^\beta }\right) , \end{equation} displaying fast oscillations near the singularity at $x_0$. Such singularities are e.g. predicted by some models of turbulence and therefore determining if they can be found in the recorded data in wind tunnels is an important issue in the understanding of the physical nature of turbulence. A first step in this program was performed by A. Kolmogorov in 1941 \cite{Kol41}. Let $f: {\mathbb R}^d \rightarrow {\mathbb R}$. The { \em Kolmogorov scaling function} of $f$ is the function $\eta_f (p)$ implicitly defined by \begin{equation} \label{kolmo} \forall p >0, \hspace{8mm} \int |f(x+h) - f(x)|^p dx \quad\sim\quad |h|^{\eta_f(p)}, \end{equation} the symbol $\sim$ meaning that \begin{equation} \label{scalKolmo} \hspace{6mm} \eta_f (p) =\liminf_{|h | \rightarrow 0} \frac{\log \left(\displaystyle\int |f(x+h) - f(x)|^p dx\right) }{\log |h| } . \end{equation} Note that, if $f$ is smooth, then one has to use differences of order 2 or more in order to define correctly the scaling function. Kolmogorov proposed to use this tool as a way to determine if some simple stochastic processes are fitted to model the velocity of turbulent fluids at small scales, and a first success of this approach was that fractional Brownian motions (see Section \ref{Univariate2}) do not yield correct models (their scaling functions are linear, whereas the one measured on turbulent flows are significatively concave \cite{arneodoetal9600}). An important interpretation of the Kolmogorov scaling function can be given in terms of { \em global smoothness} indices in families of functions spaces: the spaces $\mbox{Lip} ( s , L^p ({\mathbb R}^d)) $ defined as follows. Let $s \in (0,1)$, and $p \in [1, \infty]$; $ f \in \mbox{Lip} ( s , L^p ({\mathbb R}^d)) $ if $ f \in L^p ({\mathbb R}^d)$ and \begin{equation} \label{nicol1} \exists C>0, \;\; \forall h >0, \hspace{6mm} \int | f(x+h) -f(x) |^p dx \leq C |h|^{sp} \end{equation} (here also, larger smoothness indices $s$ are reached by replacing the first-order difference $| f(x+h) -f(x) |$ by higher order differences). It follows from (\ref{kolmo}) and (\ref{nicol1}) that, \begin{equation} \label{nicol} \forall p \geq 1, \hspace{6mm} \eta_f (p) = p \cdot \sup \{ s: f \in \mbox{Lip} ( s , L^p ({\mathbb R}^d)) \} . \end{equation} An alternative formulation of the scaling function can be given in terms of global regularity indices supplied by Sobolev spaces, the definition of which we now recall. \begin{defi} \label{sobspac} Let $s \in {\mathbb R} $ and $p \geq 1$. A function $f$ belongs to the { Sobolev space } $L^{p,s} ({\mathbb R}^d)$ if $\left( Id -\Delta \right)^{s/2} f \in L^p$, where $ g= (Id -\Delta )^{s/2}f$ is defined through its Fourier transform as \[ \hat{g}(\xi ) = ( 1 + | \xi |^2)^{s/2}\hat{f}(\xi ). \] \end{defi} This definition amounts to state that the fractional derivative of $f$ of order $s$ belongs to $L^p$. The classical embeddings between the Sobolev and the $\mbox{Lip} ( s , L^p ) $ spaces imply that \begin{equation} \label{nicol2} \forall p \geq 1, \hspace{6mm} \eta_f (p) = p \cdot \sup \{ s: f \in L^{p,s} ({\mathbb R}^d) \} . \end{equation} In other words, the scaling function tells, for each $p$, the order of (fractional) derivation of $f$ up to which $f^{(s)}$ belongs to $L^p$. A limitation of the use of the Kolmogorov scaling function for classification purposes is that many models display almost identical scaling functions (a typical example is supplied by the velocity of fully developed turbulence, see e.g. \cite{muzyetal91,LASHERMES:2008:A}); the next challenge therefore is to construct alternative scaling functions which would allow to draw distinctions between such models. A major advance in this direction was reached in 1985 when Uriel Frisch and Giorgio Parisi proposed another interpretation of the scaling function in terms of { \em pointwise singularities} of the data \cite{ParFri85}. In order to state their assertion, we first need the recall the most commonly used notion of pointwise regularity. \begin{defi} \label{defholdponc} Let $f: \; {\mathbb R}^d \rightarrow {\mathbb R} $ be a { locally bounded function}, $x_0 \in {\mathbb R}^d$ and let $\gamma \geq 0$; $f $ belongs to $ C^\gamma (x_0)$ if there exist $C>0$, $R >0$ and a polynomial $P$ of degree less than $\gamma$ such that \[ \mbox{ if} \;\; |x-x_0| \leq R, \;\; \mbox{ then} \hspace{8mm} |f(x) -P(x-x_0) | \leq C | x-x_0|^\gamma . \] The H\"older exponent of $f$ at $x_0 $ is \begin{equation} \label{holexpo} h_f (x_0) =\sup \left\{ \gamma : \;\; f \;\; \mbox{ is } \;\; C^{\gamma } (x_0) \right\} . \end{equation} \end{defi} Some functions have a very simple H\"older exponent. For instance, the H\"older exponent of the Weierstrass functions ${ \cal W}_{a, \omega}$ is constant and equal to $\omega$ at every point (such functions are referred to as { \em monoh\"older functions}); since $\omega <1$ we thus recover the fact that ${ \cal W}_{a, \omega}$ is nowhere differentiable. However, the H\"older exponent of other functions turn out to be extremely irregular, and U. Frisch and G. Parisi introduced the { \em \hspace{-2mm} multifractal spectrum } ${ \cal D}_f $ as a new quantity which allows to quantify some of its properties: ${ \cal D}_f (H)$ denotes the fractional dimension of the { \em \hspace{-2mm} isoregularity sets}, i.e. the sets \begin{equation} \label{isoh} \{ x : \quad h_f (x) = H\} . \end{equation} Based on statistical physics arguments, they proposed the following relationship between the scaling function and ${ \cal D}_f (H)$: \begin{equation} \label{formul} { \cal D}_f (H) = \inf_{p } \left( d+Hp - \eta_f (p) \right), \end{equation} which is referred to as the { \em \hspace{-2mm} multifractal formalism}, see \cite{ParFri85} (we will discuss in Section \ref{Univariate1} the ``right'' notion of fractional dimension needed here). Though the remarkable intuition which lies behind this formula proved extremely fruitful, it needs to be improved in order to be completely effective; indeed many natural processes used in signal or image modelling do not follow this formula if one tries to extend it to negative values of $p$, see \cite{AJL05}; additionally, the only mathematical result relating the spectrum of singularities and the Kolmogorov scaling function in all generality is very partial, see \cite{jmf2,Jaf7}. In Section \ref{Univariate2} we will discuss (\ref{formul}), and see how it needs to be reformulated in terms of wavelet expansions in order to reach a fairly general level of validity. In Section \ref{Poitexp} we will discuss the relevance of the H\"older exponent \eqref{holexpo} and introduce alternative exponents which are better fitted to the analysis of large classes of real-life data. Their characterization requires the introduction of orthonormal wavelet bases. This tool and its relevance for global regularity is recalled in Section \ref{Univariate3} and the characterizations of pointwise regularity which they allow are performed in Section \ref{subsec:plead}. This leads to a classification of pointwise singularities which yields a precise description of the oscillations of the function in the neighbourhood of its singularities which is developed in Section \ref{classif}. This implications of this classification on the different formulations of the multifractal formalism are developed in Section \ref{mathresmf}. The tools thus developed are applied to marathon runners physiological data (heart rate, acceleration, cadence, i.e. number of steps per minute) in Section \ref{Marathon1}; thus showing that they lead to a sharper analysis of the physiological modifications during the race. The numerical results derived on real-life data have been obtained using the Wavelet $p$-Leader and Bootstrap based MultiFractal analysis (PLBMF) toolbox available on-line at \\ $https://www.irit.fr/~Herwig.Wendt/software.html$ The explosion of data sciences recently made available collections of signals the singularities of which are expected to be related in some way; typical examples are supplied by EEG collected at different areas of the brain, or by collections of stock exchange prizes. The purpose of Section \ref{Multivariate} is to address the extension of multifractal analysis to the multivariate setting, i.e. to several functions. In such situations, a pointwise regularity exponent $h_i(x)$ is associated with each signal $f_i (x)$ and the challenge is to recover the { \em joint multivariate spectrum} of the $f_i $ which is defined as the fractional dimension of the sets of points $x$ where each of the exponents $h_i(x)$ takes a given value: If $m$ signals are available, we define \begin{equation} E_{f_1, \dots, f_m} (H_1, \dots, H_m) = \{ x : \quad h_{1} (x) = H_1, \dots, h_{m} (x) = H_m \} , \end{equation} and the {\em joint multifractal spectrum} is \begin{equation} D_{f_1, \dots, f_m} ( H_1, \dots, H_m) = \dim (E_{f_1, \dots, f_m} (H_1, \dots, H_m) ) . \end{equation} These notions were introduced by C.~Meneveau \emph{et al.}~in the seminal paper \cite{Meneveau90} which addressed the joint analysis of the dissipation rate of kinetic energy and passive scalar fluctuations for fully developed turbulence, and a general abstract setting was proposed by J.~Peyri\`ere in \cite{pey2004}; In Section \ref{MultivariateSpec}, we introduce the mathematical concepts which are relevant to this study. In Section \ref{MultivariateWSF} we give a probabilistic interpretation of the scaling functions introduced in Section \ref{Univariate}, and we show how they naturally lead to a 2-variable extension in terms of correlations. The initial formulation of the multifractal formalisms based on extensions of the Kolmogorov scaling function suffers from the same drawbacks as in the univariate case. This leads naturally to a reformulation of the multifractal formalism which is examined in Section \ref{MultivariateMF}, where we also investigate the additional advantagess supplied by multivariate multifractal analysis for singularity classifications. In order to investigate its relevance, we study a toy-example which is supplied by Brownian motions in multifractal time in Section \ref{BrwonianMT}. In Section \ref{MultivariateMar}, we illustrate the mathematical results thus collected by applications to the joint analysis of heartbeat, cadence and acceleration of marathon runners. \section{Univariate multifractal analysis} \label{Univariate} \subsection{ The multifractal spectrum} \label{Univariate1} In order to illustrate the motivations of multifractal analysis, let us come back to the initial problem we mentioned: How badly can a continuous function behave? We mentioned the surprising result of Mazurkiewicz and Banach stating that a generic continuous function is nowhere differentiable, and the Weierstrass functions yield examples of continuous functions which may have an arbitrarily small (and constant) H\"older exponent. This can actually be improved: A generic continuous function satisfies \begin{equation} \label{genercont} \forall x \in {\mathbb R}, \qquad h_f (x) = 0, \end{equation} see \cite{HSY}: At every point the H\"older exponent of $f$ is as bad as possible. An example of such a continuous function is supplied by a slight variant of Weierstrass functions: \[ f(x) = \sum_{j=1}^\infty \frac{1}{j^2} \sin (2^j x). \] Let us now consider a different functional setting: Let $f : [0,1] \longrightarrow [0, 1]$ be an increasing function. At any given point $x\in [0,1]$ $f$ can have a discontinuity at $x$, in which case $h_f (x) = 0$. Nonetheless, this worse possible behavior cannot be met everywhere: An important theorem of Lebesgue states that $f$ is almost everywhere differentiable and therefore satisfies \[ \mbox{ for almost every } x \in [0, 1], \qquad h_f (x) \geq 1. \] The global regularity assumption (the fact that $f$ is increasing implies that its derivative in the sense of distributions is a bounded Radon measure) implies that, in sharp contradistinction with generic continuous functions, the set of points such that $h_f (x) <1$ is ``small'' (its Lebesgue measure vanishes). On other hand, the set of points where it is discontinuous can be an arbitrary countable set (but one easily checks that it cannot be larger). What can we say about the size the sets of points with intermediate regularity (i.e. having H\"older exponents between 0 and 1), beyond the fact that they have a vanishing Lebesgue measure? Answering this problem requires to use some appropriate notion of ``size'' which allows to draw differences between sets of vanishing Lebesgue measure. The right mathematical notion fitted to this problem can be guessed using the following argument. Let \[ E^\al_f = \{ x: f \notin C^\alpha (x) \} . \] Clearly, if $x \in E^\al_f$, then there exists a sequence of dyadic intervals \begin{equation} \label{dyadint} \lambda_{j,k} = \left[\frac{k}{2^j},\frac{k+1}{2^j} \right] \end{equation} such that \begin{itemize} \item $x$ belongs either to $\lambda_{j,k}$ or to one of its two closest neighbours of the same width, \item the increment of $f$ on $\lambda_{j,k}$ is larger than $2^{-\alpha j} = |\lambda_{j,k}|^\alpha$ (where $|A|$ stands for the diameter of the set $A$). \end{itemize} Let $\varepsilon >0$, and consider the { \em maximal} dyadic intervals of this type of width less than $\varepsilon /3$, for all possible $x \in E^\al_f$, and denote this set by $\Lambda^\varepsilon_\alpha$. These intervals are disjoint (indeed two dyadic intervals are either disjoint or one is included in the other); and, since $f$ is increasing, the increment of $f$ on $[0,1]$ is bounded by the sum of the increments on these intervals. Therefore \[ \sum_{ \lambda \in \Lambda^\varepsilon_\alpha} | \lambda |^\alpha \leq f(1)-f(0). \] The intervals $3 \lambda$ (which consists in the dyadic interval $\lambda$ and its two closest neighbours of the same length) for $ \lambda \in \Lambda^\varepsilon_\alpha$ form an { \em $\varepsilon$-covering }\ of $E^\al_f$ (i.e. a covering by intervals of length at most $\varepsilon$), and this $\varepsilon$-covering satisfies \[ \sum_{ \lambda \in \Lambda^\varepsilon_\alpha} | 3\lambda |^\alpha = 3^\alpha \sum_{ \lambda \in \Lambda^\varepsilon_\alpha} | \lambda |^\alpha \leq f(1)-f(0). \] This property can be interpreted as stating that the { \em $\alpha$-dimensional Hausdorff measure} of $E^\al_f$ is finite; we now give a precise definition of this notion. \begin{defi} \label{defmeshaus} Let $A$ be a subset of $ {\mathbb R}^d$. If $\varepsilon>0$ and $\delta \in [0,d]$, let \[ M^\delta_{\varepsilon} =\inf_R \; \left( \sum_{ i} | A_i |^\delta \right) ,\] where $R$ is an $\varepsilon$-covering of $A$, i.e. a covering of $A$ by bounded sets $\{ A_i\}_{i \in {\mathbb N}}$ of diameters $| A_i | \leq \varepsilon$ (the infimum is therefore taken on all $\varepsilon$-coverings). For any $\delta \in [0,d]$, the $\delta$-dimensional Hausdorff measure of $A$ is \[ mes_\delta (A) = \displaystyle\lim_{\varepsilon\rightarrow 0} M^\delta_{\varepsilon}. \] \end{defi} One can show that there exists $\delta_0 \in [0,d]$ such that \[ \left\{ \begin{array}{l} \forall \delta < \delta_0, \;\;\; mes_\delta (A) = + \infty \\ \forall \delta > \delta_0, \;\;\; mes_\delta (A) = 0 . \end{array}\right. \] This critical $\delta_0$ is called the { \em Hausdorff dimension} of $A$, and is denoted by $\dim (A)$ (and an important convention is that, if $A$ is empty, then $\dim \,(\emptyset )= -\infty $). The example we just worked out shows that a global regularity information on a function yields information on the Hausdorff dimensions of its sets of H\"older singularities. This indicates that the Hausdorff dimension is the natural choice in (\ref{formul}), and motivates the following definition. \begin{defi} Let $f: {\mathbb R}^d \rightarrow {\mathbb R}$ be a locally bounded function. The { multifractal H\"older spectrum } of $f$ is the function \[ { \cal D}_f (H)= \dim (\{ x: \hspace{3mm} h_f(x) = H\} ), \] where $\dim$ denotes the Hausdorff dimension. \end{defi} This definition justifies the denomination of { \em multifractal functions}: One typically considers functions $f$ that have non-empty isoregularity sets \eqref{isoh} for $H$ taking all values in an interval of positive length, and therefore one deals with an infinite number of fractal sets $E_f (H) $. The result we obtained thus implies that, if $f$ is an increasing function, then \begin{equation} \label{majspecinc} { \cal D}_f (H) \leq H. \end{equation} This can be reformulated in a function space setting which puts in light the sharp contrast with (\ref{genercont}): Indeed, recall that any function of bounded variation is the difference of an increasing and a decreasing function; we have thus obtained the following result. \begin{prop} Let $f: {\mathbb R} \rightarrow {\mathbb R}$ be a function of bounded variation. Then its multifractal spectrum satisfies \[ \forall H, \qquad { \cal D}_f (H) \leq H . \] \end{prop} { \bf Remark:} This result does not extend to several variables functions of bounded variation which, in general, are not locally bounded, in which case their H\"older exponent is not even well defined. \subsection{ Alternative formulations of the multifractal formalism} \label{Univariate2} We mentioned that (\ref{formul}) yields a poor estimate of the multifractal spectrum. A typical example is supplied by sample paths of { \em fractional Brownian motion} (referred to as fBm), a family of stochastic processes introduced by Kolmogorov \cite{Kolmo40}, the importance of which was put in light for modeling by Mandelbrot and Van Ness \cite{mvn68}. This family is indexed by a parameter $\alpha\in (0,1)$, and generalizes Brownian motion (which corresponds to the case $\alpha= 1/2$); fBm of index $\alpha $ is the only centered Gaussian random process $B^\alpha $ defined on ${\mathbb R}^+$ which satisfies \[ \forall x, y \geq 0\qquad {\mathbb E} ( | B^\alpha (x)-B^\alpha (y)|^2) = |x-y|^{2\alpha} .\] FBm plays an important role in signal processing because it supplies the most simple one parameter family of stochastic processes with stationary increments. Its sample paths are monoh\"older and satisfy \[ \mbox{ a.s. } \forall x , \qquad h_{B^\alpha} (x) =\alpha , \] (see \cite{KahCamb} (and \cite{EssLoos} for a recent sharp analysis of the pointwise regularity of their sample paths) so that their multifractal spectrum is \[ \mbox{ a.s. } \quad \forall H, \qquad \left\{\begin{array}{rll} { \cal D}_{B^\alpha} (H) = & 1 & \quad\mbox{if } H = \alpha\\[2mm] = & -\infty & \quad\mbox{else. } \end{array}\right. \] However, the right hand-side of (\ref{formul}) yields a different value for $H \in (\alpha, \alpha +1]$: It coincides almost surely with the function defined by \[ \left\{\begin{array}{rll} { \cal L}_{B^\alpha} (H) = & \alpha +1 - H & \quad\mbox{if } H \in [\alpha, \alpha +1] \\[2mm] = & -\infty & \quad\mbox{else, } \end{array}\right. \] see \cite{JAFFARD:2010:A,JAFFARD:2006:A,MandMemor}. This is due to the fact that the decreasing part of the spectrum is recovered from negative values of $p$ in (\ref{formul}), and the corresponding integral is not well defined for negative $p$s, and may even diverge. It follows that sharper estimates of the multifractal spectrum require a renormalization procedure which would yield a numerically robust output for negative $p$s. Several methods have been proposed to solve this deadlock. They are all based on a modification of the Kolmogorov scaling function in order to incorporate the underlying intuition that it should include some pointwise regularity information. A consequence will be that they provide an extension of the scaling function to negative $p$s. This extra range of parameters plays a crucial role in several applications where it is required for classifications, see e.g. \cite{muzyetal91,LashermesEuropeanPhysicalJournalB2008} where the validation of turbulence models is considered, and for which the key values of the scaling function which are needed to draw significative differences between these models are obtained for $p<0$. A first method is based on the { \em continuous wavelet transform}, which is defined as follows. Let $\psi$ be a { \em wavelet}, i.e. a well localized, smooth function with, at least, one vanishing moment. The continuous wavelet transform of a one-variable function $f$ is \begin{equation} \label{dec} C_{a,b} (f) = \frac{1}{a}\int_{\mathbb R} f(t) \psi \left(\frac{t-b}{a}\right) dt \qquad ( a >0, \quad b \in {\mathbb R} );\end{equation} Alain Arneodo, Emmanuel Bacry and Jean-François Muzy proposed to replace, in the integral (\ref{kolmo}), the increments $ | f (x+\delta) -f (x) |$ at scale $\delta$ by the continuous wavelet transform $C_{a,b} (f) $ for $a = \delta $ and $b = x$. This choice follows the heuristic that the continuous wavelet transform satisfies $ |C_{a,b} (f)| \sim a^{h_f (x)} $ when $a$ is small enough and $| b-x| \sim a $. Note that it is not valid in all generality, but typically fails for { \em oscillating singularities}, such as the chirps (\ref{chirp}). Nonetheless Yves Meyer showed that this heuristic actually characterizes another pointwise regularity exponent, the { \em weak scaling exponent}, see \cite{MeyWVS}. Assuming that the data do not include oscillating singularities, the integral (\ref{kolmo}) is discretized and replaced by the more meaningful values of the continuous wavelet transform i.e. at its local maxima \cite{ABM95}; if we denote by $b_k $ the points where these extrema are reached at the scale $a$, the integral (\ref{kolmo}) is thus replaced by the sum \begin{equation} \label{foncechbis} \sum_{b_k} | C_{a,b_k} (f)|^p \sim a^{ \zeta_f (p)} \quad \mbox{ when } a \rightarrow 0 , \end{equation} This reformulations using the { \em multiresolution quantities } $| C_{a,b_k} (f)|$ yields better numerical results than when using the increments $ | f (x+\delta) -f (x) |$; above all, the restriction to the local suprema is a way to bypass the small values of the increments which were the cause of the divergence of the integral (\ref{kolmo}) when $p$ is negative. Numerical experiments consistently show that the multifractal formalism based on these quantities yields the correct spectrum for the fBm, and also for large collections of mutifractal models, see \cite{AADMV}. Another way to obtain a numerically robust procedure in order to perform multifractal analysis is supplied by { \em \hspace{-2mm} Detrended Fluctuation Analysis} (DFA) : From the definition of the H\"older exponent, Kantelhardt et al. \cite{kantelhardt2002multifractal} proposed the following multiresolution quantity based on the following local $L^2$ norms \begin{equation} \label{equ:mradfa} {T}_{mfd}(a,k) = \left( \frac{1}{a} \sum_{i=1}^{a} |f(ak+i) - P_{k,a, N_P}(i) |^2 \right)^{\frac{1}{2}}, k =1, \ldots, n/a, \end{equation} where $n $ denotes the number of available samples and $P_{t,a,N_P}$ is a polynomial of degree $N_P$ obtained by {local fit to $f$} on portions of length proportional to $a$. The integral (\ref{kolmo}) is now replaced by $$ S_{mfd}(a,q)= \frac{a}{n} \sum_k^{n/a} {T}_{mfd}(a,k)^q \sim a^{\zeta_{mfd}(q) }, $$ and the multifractal spectrum is obtained as usual through a Legendre transform of this new scaling function $\zeta_{mfd}$, thus yielding the \emph{multifractal detrended fluctuation analysis} (MFDFA). Note that, here again, we cannot expect the multifractal formalism based on such a formula to be fitted to the H\"older exponent: The choice of an $L^2$ norm in (\ref{equ:mradfa}) is rather adapted to an alternative pointwise exponent, the 2-exponent, which is defined through local $L^2$-norms, see Def. \ref{def:pexp} (and \cite{PART2} for an explanation of this interpretation). The MFDFA formalism performs satisfactorily and is commonly used in applications (cf., e.g., \cite{galaska2008comparison,Wang2012}). The methods we mentioned meet the following limitations: They cannot be taylored to a particular pointwise exponent: We saw that the WTMM is fitted to the weak-scaling exponent, and the MFDFA to the 2-exponent. They lack of theoretical foundation, and therefore the estimates that they yield on the multifractal spectrum are not backed by mathematical results. In practice, they are difficult to extend to data in two or more variables (for MFDFA, the computation of local best fit polynomials is an intricate issue). The obtention of an alternative formulation of the multifractal formalism which brings an answer to these two problems requires a detour through the notions of pointwise exponents, and their characterizations. \subsection{Pointwise exponents} \label{Poitexp} At this point we need to discuss the different notions of pointwise regularity. One of the reasons is that, though H\"older regularity is by far the one which is most used in mathematics and in applications, it suffers a major limitation: Definition \ref{defholdponc} requires $f$ to be locally bounded. In applications, this limitation makes the H\"older exponent unfitted in many settings where modelling data by locally bounded functions is inadequate; in Section \ref{Univariate3} we will give a numerically simple criterium which allows to verify if this assumption is valid, and we will see that the physiological data we analyse are typical examples for which it is not satisfied. On the mathematical side too, this notion often is not relevant. A typical example is supplied by the Riemann series defined as \begin{equation} \label{riem} \forall x \in {\mathbb R}, \qquad {\mathcal R}_s (x) = \sum_{n=1}^\infty \frac{\sin(n^2 x) }{n^s} ,\end{equation} which, for $s > 1$, are locally bounded and turn out to be multifractal (in which case their multifractal analysis can be performed using the H\"older exponent \cite{Vindas,JaffRie}), but it is no more the case if $s <1$, in which case an alternative analysis is developed in \cite{SeurUb} (using the $p$-exponent for $p =2$, see Def. \ref{def:pexp} below). There exist two ways to deal with such situations. The first one consists in first regularizing the data, and then analyzing the new data thus obtained. Mathematically, this means that a { \em fractional integral} is performed on the data. Recall that, if $f$ is a tempered distribution defined on ${\mathbb R}$, then the fractional integral of order $t$ of $f$, denoted by $ f^{(-t )}$ is defined as follows: Let $(Id-\Delta )^{-t /2}$ be the convolution operator which amounts to multiplying the Fourier transform of $f$ with $(1+|\xi |^2)^{-t /2}$. The fractional integral of order $t$ of $f$ is the function \[ f^{(-t )} =(Id-\Delta )^{-t /2} ( f) . \] If $f$ is large enough, then $f^{(-t )}$ is a locally bounded function, and one can consider the H\"older exponent of $t$ (the exact condition under which this is true is that $t$ has to be larger than the exponent ${h^{min}_f}$ defined below by \eqref{hmin1} or equivalently by \eqref{caracbeswav3hol}). This procedure presents the obvious disadvantage of not yielding a direct analysis of the data but of a smoothed version of them. The other alternative available in order to characterize the pointwise regularity of non-locally bounded functions consists in using a weaker notion of pointwise regularity, the $p$-exponent, which we now recall. We define $B(x_0, r)$ as the ball of center $x_0$ and radius $r$. \begin{defi} \label{def:pexp} Let $p \geq 1$ and assume that $f\in L^p_{loc} ({\mathbb R}^d)$. Let $\alpha \in {\mathbb R}$; $ f $ belongs to $T^p_\alpha (x_0)$ if there exists a constant $ C$ and a polynomial $P_{x_0}$ of degree less than $\alpha $ such that, for $r$ small enough, \begin{equation} \label{pexpa} \left( \frac{1}{r^d} \int_{B(x_0, r} | f(x) -P_{x_0}(x)|^p dx \right)^{1/p} \leq C r^\alpha. \end{equation} The { \bf $p$-exponent} of $f$ at $x_0$ is \begin{equation} \label{equ-pexp} h_f^p(x_0) = \sup \{ \alpha : f \in T^p_\alpha (x_0)\} \end{equation} (the case $p= + \infty$ corresponds to the H\"older exponent). \end{defi} This definition was introduced by Calder\'on and Zygmund in 1961 in order to obtain pointwise regularity results for the solutions of elliptic PDEs, see \cite{CalZyg}. For our concern, it has the important property of being well defined under the assumption that $f \in L^p_{loc}$. For instance, in the case of the Riemann series (\ref{riem}), an immediate computation yields that they belong to $L^2$ if $s >1/2$ so that, if $ 1/2 < s <1$, $p$-exponents with $p \leq 2$ are relevant to study their regularity, in contradistinction with the H\"older exponent which won't be defined. Another example of multifractal function which is not locally bounded is supplied by Brjuno's function, which plays an important role in holomorphic dynamical systems, see \cite{MaMoYo}. Though its is nowhere locally bounded, it belongs to all $L^p$ spaces and its multifractal analysis using $p$-exponents has been performed in \cite{JaMar}. Note that $p$-exponents can take values down to $-d/p$, see \cite{PART1}. Therefore, they allow the use of { \em negative regularity exponents}, such as singularities of the form $f(x) = 1/|x-x_0|^\alpha$ for $\alpha<d/p$. The general framework supplied by multifractal analysis now is ubiquitous in mathematical analysis and has been successfully used in a large variety of mathematical situations, using diverse notion of pointwise exponents such as pointwise regularity of probability measures \cite{Bmp92}, rates of convergence or divergence of series of functions (either trigonometric \cite{Aubry06,BaHeu} or wavelet \cite{Aubry06,EssJaff}) order of magnitude of ergodic averages \cite{Ling,Ling2}, to mention but a few. \subsection{ Orthonormal wavelet decompositions} \label{Univariate3} Methods based on the use of orthonormal wavelet bases follow the same motivations we previously developed, namely to construct alternative scaling functions based on multiresolution quantities which ``incorporate'' some pointwise regularity information. However, we will see that they allow to turn some of the limitations met by the previously listed methods, and they enjoy the following additional properties: \begin{itemize} \item numerical simplicity, \item explicit links with pointwise exponents (which, as we saw, may differ from the H\"older exponent), \item no need to construct local polynomial approximations (which is the case for DFA methods now in use), \item mathematical results hold concerning either the validity of the multifractal formalism supplied by (\ref{formul}) or of some appropriate extensions; such results can be valid for all functions, or for ``generic'' functions, in the sense of Baire categories, or for other notions of genericity. \end{itemize} Let us however mention an alternative technique which was proposed in \cite{Abel} where multiresolution quantities based on local oscillations, such as \[ d_\lambda = \sup_{ 3 \lambda} f(x) -\inf_{3\lambda} f(x) , \] or higher order differences such as \[ d_\lambda = \sup_{x,y \in 3 \lambda} \left| f(x) + f(y) -2 f\left(\frac{x + y}{2}\right)\right| , \] and which wouldn't present the third problem that we mention. However, as far as we know, they haven't been tested numerically. One of the reasons for these remarkable properties is that (in contradistinction with other expansions, such as e.g. Fourier series) wavelet analysis allows to characterize both global and pointwise regularity by simple conditions on the moduli of the wavelet coefficients; as already mentioned, the multifractal formalism raises the question of how global and pointwise regularity are interconnected; wavelet analysis therefore is a natural tool in order to investigate this question and this explains why it was at the origin of major advances in multifractal analysis both in theory and applications. We now recall the definition of orthonormal wavelet bases. For the sake of notational simplicity, we assume in all the remaining of Section \ref{Univariate} that $d=1$, i.e. the functions we consider are defined on ${\mathbb R}$, extensions in several variables being straightforward. Let $\varphi(x) $ denote an smooth function with fast decay, and good joint time-frequency localization, referred to as the {\em \hspace{-2mm} scaling function}, and let $\psi(x) $ denote an oscillating function (with $N$ first vanishing moments), with fast decay, and good joint time-frequency localization, referred to as the {\em \hspace{-2mm} wavelet}. These functions can be chosen such that the \begin{equation} \label{wavbasbad1} \varphi (x-k), \hspace{5mm} \mbox{ for} , \hspace{3mm} k\in {\mathbb Z} \end{equation} and \begin{equation} \label{wavbasbad} 2^{j/2}\psi (2^{j}x-k), \hspace{5mm} \mbox{ for} , \hspace{3mm} j \geq 0, \;\; k\in {\mathbb Z} \end{equation} form an orthonormal basis of $L^2 ({\mathbb R})$ \cite{Mey90I}. The wavelet coefficients of a function $f$ are defined as \begin{equation} \label{eq-wc} c_{ k} = \displaystyle\int_{{\mathbb R}} f(x) \; \varphi (x-k) \, dx \qquad \mbox {and} \qquad c_{ j,k} = 2^{j} \displaystyle\int_{{\mathbb R}} f(x) \; \psi (2^{j}x-k) \, dx \end{equation} Note the use of an $L^1$ normalization for the wavelet coefficients that better fits local regularity analysis. As stated above, the H\"older exponent can be used as a measurement of pointwise regularity in the locally bounded functions setting only, see \cite{JAFFARD:2010:A}. Whether empirical data can be well-modelled by locally bounded functions or not can be determined numerically through the computation of the {\em uniform H\"older exponent} ${h^{min}_f}$, which, as for the scaling function, enjoys a function space characterization \begin{equation} \label{hmin1} {h^{min}_f} = \sup\{ \alpha: f \in C^\alpha ({\mathbb R} )\}, \end{equation} where $C^\alpha ({\mathbb R} )$ denotes the usual H\"older spaces. Assuming that $\varphi$ and $\psi$ are smooth enough and that $\psi$ has enough vanishing moments, then the exponent ${h^{min}_f}$ has the following simple wavelet characterization: \begin{equation} \label{caracbeswav3hol} {h^{min}_f} =\liminf_{j \rightarrow + \infty} \;\; \;\; \frac{ \log \left( \displaystyle \sup_{ k} | c_{ j,k} | \right) }{\log (2^{-j})}. \end{equation} It follows that, if { ${h^{min}_f} > 0$}, then $f$ is a continuous function, whereas, if { ${h^{min}_f} < 0$}, then $f$ is not a locally bounded function, see \cite{MandMemor,Bergou}. In numerous real world applications the restriction ${h^{min}_f} > 0$ constitutes a severe limitation; we will meet such examples in the case of physiological data (see also \cite{MandMemor} for other examples). From a pratical point of view, the regularity of the wavelets should be larger than ${h^{min}_f}$ in order to compute the estimation of ${h^{min}_f}$. In the applications that we will see later, we took Daubechies compactly supported wavelets of increasing regularity and we stopped as soon as we found a threshold beyond which there is no more modification of the results. In our case, we stopped at order 3. In applications, the role of ${h^{min}_f}$ is twofold: It can be used as a classification parameter and it tells whether a multifractal analysis based on the H\"older exponent is licit. Unlike other multifractality parameters that will be introduced in the following, its computation does not require a priori assumptions: It can be defined in the widest possible setting of tempered distributions. \\ We represent these two types of data on Fig. \ref{hmin} for a marathon runner. The race is composed of several stages including a warm-up at the beginning, a recovery at the end of the marathon, and several moments of small breaks during the marathon. The signal was cleaned by removing the data that did not correspond to the actual race period (warm-ups, recoveries and breaks) and by making continuous connections to keep only the homogeneous parts. This type of connection is suitable for regularities exponents lower than 1 as in the case of our applications. \begin{figure} \centering \includegraphics[scale=0.3]{data.png} \caption{Representation of data: heart rate (left) in beats per minute, cadence (middle) in steps per minute and acceleration (top) in meters per second squared. The time scale is in 0.1s} \label{data} \end{figure} \begin{figure} \centering \includegraphics[scale=0.3]{h_min_FC.png} \quad \includegraphics[scale=0.3]{h_min_Acc.png} \caption{Estimation by log-log regression of the $h_{min}$ of a heart rate (left) and an acceleration (right). The points of the regression line up successfully along a close to straight line thus showing that the values of $h_{min}$, are precisely estimated and are negative. It follows that a multifractal analysis based on H\"older exponent cannot be performed on these data.} \label{hmin} \end{figure} If ${h^{min}_f} <0$, then a multifractal analysis based on the H\"older exponent cannot be developed, and the question whether a multifractal analysis based on the $p$-exponent can be raised. Wavelet coefficients can also be used to determine whether $f$ locally belongs to $L^p$ or not (which is the a priori requirement needed in order to use the corresponding $p$-exponent), see \cite{Bergou,Abel,MandMemor}: Indeed, a simple wavelet criterium can be applied to check this assumption, through the computation of the { \em wavelet structure function}. Let \begin{equation} \label{equ-WSF} S_c(j,p) = 2^{-j} \displaystyle\sum_{ k } | c_{ j,k} |^p. \end{equation} The \emph{wavelet scaling function} is defined as \begin{equation} \label{defscalond} \forall p >0, \hspace{6mm} \eta_f (p) = \displaystyle \liminf_{j \rightarrow + \infty} \;\; \frac{\log \left( S_c(j,p) \right) }{\log (2^{-j})}; \end{equation} one can show that it coincides with the Kolmogorov scaling function if $p >1$, see \cite{jmf2}. The following simple criterion can be applied in order to check if data locally belong to $L^p$ \cite{PART1}: \begin{equation} \label{equ:etap} \left. \begin{array}{rl} \mbox{ if} \; \eta_f(p) >0 & \mbox{then} \; f \in L^p_{loc}, \\ \mbox{ if} \; \eta_f (p) < 0 & \mbox{then} \; f \notin L^p_{loc} . \\ \end{array}\right\} \end{equation} { \bf Remarks:} The wavelet scaling function enjoys the same property as ${h^{min}_f}$: Its computation does not require some a priori assumptions on the data, and it can be defined in the general setting of tempered distributions. Note that it is also defined for $p \in (0,1]$; in that case the Sobolev space interpretation of the scaling function has to be slightly modified: In Def. \ref{sobspac} the Lebesgue space $L^p$ has to be replaced by the real Hardy spaces $H^p$, see \cite{Mey90I} for the notion of Hardy spaces and their wavelet characterization. Note that these function space interpretations imply that the wavelet scaling function does not depend on the specific (smooth enough) wavelet basis which is used; it also implies that it is unaltered by the addition of a smooth function, or by a smooth change of variables, see \cite{Abel} and ref. therein. For the same reasons, these properties also hold for the exponent ${h^{min}_f}$; they are required in order to derive intrinsic parameters for signal or image classification. In the following, we shall refer to them as { \em robustness properties}. In applications \eqref{defscalond} can be used only if $\eta_f(p)$ can be determined by a log-log plot regression, i.e. when the limit actually is a limit, see e.g. Fig. \ref{scaling function for p=1}. This means that the structure functions \eqref{equ-WSF} satisfy $S_c (j,p) \sim 2^{-\eta_f (p)j} $ in the limit of small scales, a phenomenon coined { \em scale invariance}. The practical relevance of the wavelet scaling function (and other multifractal parameters that we will meet later), comes from the fact that it can be used for classification of signals and images without assuming that the data follow an a priori model. \begin{figure} \centering \includegraphics[scale=0.3]{fonc_echelle_FC_Dylan.png} \quad \includegraphics[scale=0.3]{fonc_echelle_Cadence_Dylan.png} \caption{Wavelet scaling function of heart rate (left) and cadence (right) of a marathon runner. It allows to determine the values of $p$ such that $\eta_f (p)>0$. We conclude that a multifractal analysis based on $p$-exponents is directly possible for heart rate data, but not for the cadence, where the analysis will have to be carried out on a fractional integral of the data} \label{scaling function} \end{figure} \begin{figure} \centering \includegraphics[scale=0.3]{Fonc_echelle_FC_p=1_Dylan.png} \quad \includegraphics[scale=0.3]{Fonc_echelle_Cadence_p=1_Dylan.png} \caption{Estimation by log-log regression of the wavelet scaling function of heart rate (left) and cadence (right) for $p=1$. The slope of the regression is positive for heart rate and negative for cadence. These regressions, estimated for a suffiently large number of values of $p$ allow to plot the wavelet scaling functions, as shown in Fig. \ref{scaling function}} \label{scaling function for p=1} \end{figure} \subsection{Wavelet pointwise regularity characterizations} \label{subsec:plead} One advantage of orthonormal wavelet based methods is that they allow to construct a multifractal analysis which is taylored for a given $p$-exponent, which is not the case of the alternative methods we mentioned. We shall see in Sections \ref{classif} and \ref{Marathon1} the benefits of this extra flexibility. For this purpose, we have to construct multiresolution quantities (i.e., in this context, a non-negative function defined on the collection of dyadic cubes) which are fitted to $p$-exponents. We start by introducing more adapted notations for wavelets and wavelet coefficients; instead of the two indices $(j,k)$, we will use dyadic intervals \eqref{dyadint} and, accordingly, $c_{\lambda} = c_{ j,k} $, and $\psi_{\lambda} = \psi_{ j,k}$. The wavelet characterization of $p$-exponents requires the definition of {$p$-leaders}. If $f \in L^p_{loc} ({\mathbb R})$, the wavelet $p$-{\em leaders} of $f$ are defined as \begin{equation} \label{pleaders} \ell^{(p)}_{j,k} \equiv \ell^{(p)}_\lambda= \left( \sum_{ \lambda' \subset 3 \lambda} | c_{\lambda'} |^p \, \, 2^{j-j'} \right)^{1/p}, \end{equation} where $j'\geq j$ is the scale associated with the sub-cube $\lambda'$ included in $3 \lambda$ (i.e. $\lambda'$ has width $2^{-j'}$). Note that, when $p = +\infty $ (and thus $f \in L^{\infty}_{loc} ({\mathbb R})$), {$p$-leaders} boil down to { \em wavelet leaders} \[ \ell_\lambda = \sup_{ \lambda ' \subset 3 \lambda} | c_{\lambda'}|, \] \cite{Jaffard2004,Wendt:2007:E}. Let us indicate where such quantities come from. They are motivated by constructing quantities based on simple conditions on wavelet coefficients and which well approximate the local $L^p$ norm of Definition \ref{def:pexp}. For that purpose we use the wavelet characterization of the Besov space $B^{0,p}_p$ which is ``close'' to $L^p$ (indeed the classical embeddings between Besov and $L^p$ spaces imply that $B^{0,1}_p \hookrightarrow L^p \hookrightarrow B^{0,\infty}_p $); with the normalization we chose for wavelet coefficients, the wavelet characterization of $B^{0,p}_p$ is given by \[ f \in B^{0,p}_p \quad \mbox{ if} \qquad \sum_k | c_k |^p < \infty \qquad \mbox{and} \qquad \sum_{j,k} 2^{(sp-1)j}| c_{j,k} |^p < \infty , \] see \cite{Mey90I} and, because of the localization of the wavelets, the restriction of the second sum to the dyadic cubes $\lambda' \subset 3 \lambda$ yields an approximation of the local $L^p$ norm of $f-P$ around the interval $\lambda$ (the substraction of the polynomial $P$ comes from the fact that the wavelets have vanishing moments so that $P$ is reconstructed by the first sum in \eqref{wavbasbad1}, and the wavelet coefficients $c_{j,k}$ of $f$ and $f-P$ coincide). Actually, the uniform regularity assumption $\eta_f (p) >0$ (which we will make) implies that the quantities \eqref{pleaders} are finite. Denote by $\lambda_{j,k} (x)$ the unique dyadic interval of length $2^{-j}$ which includes $x$; a key result is that both the H\"older exponent and the $p$-exponent can be recovered from, respectively, wavelet leaders and $p$-leaders, according to the following formula. \begin{defi} \label{derived} Let $h(x)$ be a pointwise exponent and $(d_\lambda)$ a multiresolution quantity indexed by the dyadic cubes. The exponent $h$ is derived from the $(d_\lambda)$ if \begin{equation} \label{carachqf} \forall x, \qquad h(x) = \liminf_{j \rightarrow + \infty} \frac{ \log \left( d_{\lambda_{j,k} (x)} \right) }{\log (2^{-j})}. \end{equation} \end{defi} It is proved in \cite{JaffToul, Bergou,JaffMel} that if $\eta_f (p) >0$, then the $p$-exponent is derived from $p$-leaders, and, if ${h^{min}_f} >0$, then the H\"older exponent is derived from wavelet leaders. Note that the notion of $p$-exponent can be extended to values of $p$ smaller that 1, see \cite{JaffCies}; this extension requires the use of ``good'' substitutes of the $L^p$ spaces for $p <1$ which are supplied by the real Hardy spaces $H^p$. The important practical result is that the $p$-leaders associated with this notion also are given by \eqref{pleaders}. In applications, one first computes the exponent ${h^{min}_f} $ and the function $\eta_f (p)$. If ${h^{min}_f} >0$, then one has the choice of using either $p$-leaders or wavelet leaders as multiresolution quantities. Though leaders are often preferred because of the simple interpretation that they yield in terms of the most commonly used (H\"older) exponent, it has been remarked that $p$-leaders constitute a quantity which displays better statistical properties, because it is based on averages of wavelet coefficients, instead of a supremum, i.e. a unique extremal value, see \cite{CRAS2019} and ref. therein. If both ${h^{min}_f} <0$ and $\eta_f (p) <0$ for all $p$s, then one cannot use directly these techniques and one performs a (fractional) integration on the data first. If one wants to use wavelet leaders, the order of integration $s$ has to satisfy $s > -{h^{min}_f}$ since $h^{min}_{f^{(-s)}} = {h^{min}_f} +s$. Similarly, in the case of $p$-leaders it follows immediately from the Sobolev interpretation \eqref{nicol2} of the wavelet scaling function that \[ \eta_{f^{(-s)}}(p) = ps + \eta_f(p). \] Thus, if $\eta_f(p) <0$, then an analysis based on $p$-leaders will be valid if the order of fractional integration $s$ applied to $f$ satisfies $s > -\eta_f(p)/p$. In practice, one does not perform a fractional integration on the data, but one simply replaces the wavelet coefficients $c_{j,k}$ by $2^{-sj}c_{j,k}$, which leads to the same scaling functions \cite{MandMemor}, and has the advantage of being performed at no extra computational cost. \subsection{Towards a classification of pointwise singularities} \label{classif} In Section \ref{Poitexp} we motivated the introduction of alternative pointwise regularity exponents by the requirement of having a tool available for non locally bounded functions, which allows to deal directly with the data without having recourse to a smoothing procedure first; but this variety of exponents can also serve another purpose: By comparing them, one can draw differences between several types of singularities. This answers an important challenge in several areas of science; for example, in fully developed turbulence, some models predict the existence of extremely oscillating structures such as \eqref{chirp} and the key signal processing problem for the detection of gravitational waves also involves the detection of pointwise singularities similar to \eqref{chirp} in extremely noisy data \cite{Fland}. Let us start with a simple example: Among the functions which satisfy $h_f (x_0) = \alpha $, the most simple pointwise singularities are supplied by { \em cusps} singularities, i.e. by functions which ``behave'' like \begin{equation} \label{cusp} \mathcal{C}_{\alpha}(x)= | x-x_0|^\alpha \qquad \mbox{ (if $ \alpha >0$ and $\alpha \notin 2 {\mathbb N}$)} . \end{equation} How can we ``model'' such a behavior? A simple answer consists in remarking that the primitive of \eqref{cusp} is of the same form, and so on if we iterate integrations. Since the mapping $t \rightarrow h_{f^{(-t)}}(0)$ is concave \cite{ABJM}, it follows that \eqref{cusp} satisfies \[ \forall t >0, \qquad h_{\mathcal{C}^{(-t)}_{\alpha}} (t_0) = \alpha +t . \] For cusp singularities, the pointwise H\"older exponent is exactly shifted by the order of integration. This is in sharp contrast with the chirps \eqref{chirp}, for which a simple integration by parts yields that the H\"older exponent of its $n$-th iterated primitive is \[ \forall n \in {\mathbb N} , \qquad h_{\mathcal{C}^{(-n)} _{\alpha, \beta }} (t_0) = \alpha + (1 + \beta ) n , \] from which it easily follows that the fractional primitives of the chirp satisfy \[ \forall t >0, \qquad h_{\mathcal{C}_{\alpha, \beta}^{(-t)}} (t_0) = \alpha + (1 + \beta ) t , \] \cite{ABJM}. We conclude from these two typical examples that inspecting simultaneously the H\"older exponents of $f$ and its primitives, or its fractional integrals, allows to put in light that oscillating behaviour of $f$ in the neighbourhood of its singularities which is typical of \eqref{chirp} (see \cite{LVS} for an in-depth study of the information revealed by the mapping $t \rightarrow h_{f^{(-t)}} (t_0)$). To that end, the following definition was proposed, which encapsulates the relevant ``oscillatory'' information contained in this function, using a single parameter. \begin{defi}\label{def.2} Let $f: {\mathbb R}^d \rightarrow {\mathbb R} $ be such that $f \in L^p_{loc} $. If $h^p_f(x_0) \neq +\infty$, then the oscillation exponent of $f$ at $x_0$ is \begin{equation} \label{eq:oscill} \mathcal{O}_f (x_0)= \left( \frac{\partial}{\partial t} h^p_{f^{(-t)}} (x_0) \right)_{t=0^+} -1 . \end{equation} \end{defi} { \bf Remark:} In theory, a dependency in $p$ should appear in the notation since $f$ belongs to several $L^p$ spaces. However, in practice, a given $p$ is fixed, and this inaccuracy does not pose problems. The choice of taking the derivative at $t=0^+$ is motivated by a robustness argument: The exponent should not be perturbed when adding to $f$ a smoother term, i.e. a term that would be a $O( |x-x_0|^{h})$ for an $h > h_f(x_0) $; it is a consequence of the following lemma, which we state in the setting of H\"older exponents (i.e. we take $p = + \infty$ in Definition \ref{def.2}). \begin{lem} Let $f$ be such that $h_f (x_0) < +\infty$ and $\mathcal{O}_f (x_0) < +\infty $; let $g\in C^\alpha (x_0) $ for an $\alpha > h_f (x_0)$. Then, for $s$ small enough, the H\"older exponents of $(f+g)^{(-s)} $ and of $ f^{(-s}) $ coincide. \end{lem} { \bf Proof:} By the concavity of the mapping $s \rightarrow h_{f^{(-s)}} (x_0)$, see \cite{Bandt2015,Porqu2017}, it follows that \[ h_{f^{(-s)}} (x_0) \leq h_{f} (x_0) + (1+ \mathcal{O}_f (x_0))s; \] but one also has $ h_{g^{(-s)}} (x_0) \leq \alpha +s$; so that, for $s$ small enough, $ h_{g^{(-s)}} (x_0) > h_{f^{(-s)}} (x_0) $, and it follows that $h_{(f+g)^{(-s)}} (x_0) = h_{f^{(-s)}} (x_0) $. \vspace{0.5cm} The oscillation exponent takes the value $\beta$ for a chirp; it is the first of { \em second generation exponents } that do not measure a regularity, but yield additional information, paving the way to a richer description of singularities. In order to go further in this direction, we consider another example: { \em Lacunary combs}, which were first considered in \cite{Bandt2015,Porqu2017} (we actually deal here with a slight variant). Let $\phi =\mathbbm{1}_{[0, 1]} $. \begin{defi} Let $\alpha \in {\mathbb R} $ and $\gamma >\omega >0 $. The { lacunary comb } $F^\alpha_{ \omega, \gamma}$, is \begin{equation} \label{defaog} F^\alpha_{ \omega, \gamma} (x) = \sum_{j=1}^\infty 2^{-\alpha j} \phi \left( 2^{\gamma j} (x-2^{-\omega j}) \right) . \end{equation} \end{defi} We consider its behaviour near the singularity at $x_0 =0$: if $\alpha > -\gamma$, then $F^\alpha_{ \omega, \gamma} \in L^1 ({\mathbb R}) $ and it is locally bounded if and only if $\alpha \geq 0$. In that case, one easily checks that \begin{equation} h_{F^\alpha_{ \omega, \gamma}}(0) = \frac{\alpha}{\omega}, \quad \mbox{and} \quad h_{{F^\alpha_{ \omega, \gamma}}^{(-1)}}(0) =\frac{\alpha + \gamma}{\omega} \end{equation} and one obtains (see \cite{Porqu2017}) that $\mathcal{O}_{F^\alpha_{ \omega, \gamma}}(0) =\frac{ \gamma}{\omega} -1 . $ We conclude that chirps and lacunary combs are two examples of oscillating singularities. They are, however, of different nature: In the second case, oscillation is due to the fact that this function vanishes on larger and larger proportions of small balls centered at the origin (this is detailed in \cite{Bandt2015}, where this phenomenon is precisely quantified through the use of { \em accessibility exponent } of a set at a point). On the other hand, chirps are oscillating singularities for a different reason: It is due to very fast oscillations, and compensations of signs. This can be checked by verifying that the oscillation exponent of $| \mathcal{C}_{\alpha,\beta}| $ at $0$ vanishes. We will now see that this difference can be put in evidence by considering the variations of the $p$-exponent. Comparing the $p$-exponents of chirps and lacunary combs allows to draw a distinction between their singularities; indeed, for $p\geq 1$, see \cite{PART1}, \begin{equation}\label{eq:pFalphagamma} h_{F^\alpha_{ \omega, \gamma}}^{p}(0)=\alpha+\frac{1}{p}\left(\frac{\gamma}{\omega}-1\right) \end{equation} whereas a straightforward computation yields that \[ \forall p , \qquad h^{p}_{\mathcal{C}_{\alpha,\beta}}(0)=\alpha . \] We conclude that the $p$-exponent of $F^\alpha_{ \omega, \gamma}$ varies with $p$, whereas the one of $\mathcal{C}_{\alpha,\beta}$ does not. We will introduce another pointwise exponent which captures the lacunarity of the combs; it requires first the following notion: If $f \in L^p_{loc}$ in a neighborhood of $x_0$ for $p >1$, the { \em critical Lebesgue index} of $f$ at $x_0 $ is \begin{equation} \label{critleb} p_f(x_0) = \sup \{ p: f \in L^{p}_{loc} ({\mathbb R}) \mbox{ in a neighborhood of $x_0$} \} . \end{equation} The $p$-exponent at $x_0$ is defined on the interval $[1, p_f(x_0) ]$ or $[1, p_f(x_0) )$. We denote: $ q_f(x_0) ={1}/{p_f(x_0)}.$ Note that $p_f(x_0)$ can take the value $+\infty$. An additional pointwise exponent, which, in the case of lacunary combs, quantifies the sparsity of the ``teeth'' of the comb, can be defined as follows see \cite{Bandt2015}. Its advantage is that it quantifies the ``lacunarity information'' using a single parameter instead of the whole function $p\rightarrow {h}^{ (p)}_f (x_0)$. \begin{defi} \label{deflac} Let $f \in L^p_{loc}$ in a neighborhood of $x_0$ for a $p >1$. The { lacunarity exponent } of $f$ at $x_0$ is \begin{equation} \label{eq:lac_exp} { L}_f (x_0) = \frac{\partial }{\partial q} \left( {h}^{ (1/q)}_f (x_0) \right)_{ q={q_f(x_0)}^+}. \end{equation} \end{defi} This quantity may have to be understood as a limit when $q \to q_f(x_0)$, since $h_f^{1/q} (x_0)$ is not necessarily defined for $q=q_f(x_0) $. This limit always exists as a consequence of the concavity of the mapping $q \rightarrow h_f^{1/q} (x_0)$, and it is nonnegative (because this mapping is increasing). The lacunarity exponent of $F^\alpha_{ \omega, \gamma}$ at $0$ is $\frac{\gamma}{\omega}-1$, which puts into light the fact that this exponent allows to measure how $F^\alpha_{ \omega, \gamma}$ vanishes on "large sets" in the neighborhood of $0$ (see \cite{Bandt2015} for a precise statement). Furthermore the { oscillation exponent } of $F^\alpha_{ \omega, \gamma}$ at $0$ is $\frac{\gamma}{\omega} - 1$, so that it coincides with the lacunarity exponent. The { oscillation exponent } is always larger than the lacunarity exponent. A way to distinguish between the effect due to lacunarity and the one due to cancellations is to introduce a third exponent, the { \em cancellation exponent} \[ {\mathcal C}_f (x_0)= {\mathcal O}_f (x_0) - {L}_f (x_0). \] The lacunarity and the cancellation exponents lead to the following classification of pointwise singularities see \cite{Porqu2017}. \begin{defi} Let $f$ be a tempered distribution on ${\mathbb R}$: \begin{itemize} \item $f$ has a { \bf canonical singularity} at $x_0$ if ${\cal O}_f(x_0)= 0 $. \item $f$ has a { \bf balanced singularity} at $x_0$ if ${L}_f(x_0)= 0 $ and $\mathcal{C}_f(x_0)\neq 0 . $ \item $f$ has a { \bf lacunary singularity} at $x_0$ if $\mathcal{C}_f(x_0)= 0 $ and ${L}_f(x_0) \neq 0 $. \end{itemize} \end{defi} Cusps are typical examples of canonical singularities, chirps are typical examples of balanced singularities and lacunary combs are typical examples of lacunary singularities. Many probabilistic models display lacunary singularities: It is the case e.g. for random wavelet series \cite{Bandt2015,Porqu2017}, some L\'evy processes, see \cite{Balanca} or fractal sums of pulses \cite{Saes}. Note that our comprehension of this phenomenon is very partial: For instance, in the case of L\'evy processes, the precise determination of the conditions that a L\'evy measure should satisfy in order to guarantee the existence of lacunary singularities has not been worked out: in \cite{Balanca}, P. Balanca proved that some self-similar L\'evy processes with even L\'evy measure display oscillating singularities, which actually turn out to be lacunary singularities and also that L\'evy processes which have only positive jumps do not display such singularities; and, even in these cases, only a lower bound on their Hausdorff dimensions has been obtained. In other words, for L\'evy processes, a joint multifractal analysis of the H\"older and the lacunarity exponent remains to be worked out. Note also that there exists much less examples of functions with balanced singularities: In a deterministic setting it is the case for the Riemam function \cite{JaffMey1} at certain rational points. However, to our knowledge, stochastic processes with balanced singularities have not been met up to now. Another important question is to find numerically robust ways to determine if a signal has points where it displays balanced or lacunary singularities. This question is important in several areas of physics; for instance, in hydrodynamic turbulence, proving the presence of oscillating singularities would validate certain vortex stretching mechanisms which have been proposed, see \cite{Frisch1995}. Another motivation is methodological: if a signal only has canonical singularities, then its $p$-multifractal spectrum does not depend on $p$ and its singularity spectrum is translated by $t$ after a fractional integral of order, so that all methods that can be used to estimate its multifractal spectrum yield the same result (up to a known shift in the case of a fractional integration). An important questions related with the multifractal formalism is to determine if some of its variants allow to throw some light on these problems. Motivated by applications to physiological data, we shall come back to this question in Sections \ref{Marathon1} and \ref{intfrac}. Note that the choice of three exponents to characterize the ``behaviour'' of a function in the neighbourhhood of one of its singularities may seem arbitrary; indeed, one could use the very complete information supplied by the following two variables function: If $f$ is a tempered distribution, then the {\em fractional exponent} of $f$ at $x_0$ is the two variable function \[ { \mathcal H}_{f, x_0} (q,t ) = h^{1/q}_{f^{(-t)}} (x_0) -t, \] see \cite{Porqu2017} where this notion is introduced and its properties are investigated. However, storing the pointwise regularity behaviour through the use of a two-variables function defined at every point is unrealistic, hence the choice to store only the information supplied by the three parameters we described. This choice is motivated by two conflicting requirements: On one hand, one wishes to introduce mathematical tools which are sophisticated enough to describe several ``natural'' behaviours that can show up in the data, such as those supplied by cusps, chirps, and lacunary combs. On other hand, at the end, classification has to bear on as little parameters as possible in order to be of practical use in applications; the goal here is to introduce a multivariate multifractal analysis based on a single function $f$, but applied to several pointwise exponents associated with $f$ (say two or three among a regularity, a lacunarity and a cancellation exponent). Our theoretical comprehension of which functions can be pointwise exponents is extremly partial, see \cite{SeuretPres} for a survey on this topic: It has been known for a long time that a pointwise H\"older exponent $h_f (x)$ can be any nonnegative function of $x$ which can be written as a liminf of a sequence of continuous functions, see \cite{Jaffard1995,AyaJaff,Daoud}, but the same question for $p$-exponents is open (at least in the case where it takes negative values). Similarly, which couples of functions $(h(x), O(x))$ can be the joint H\"older and oscillation exponents of a function also is an open question (see \cite{JaffReV} for partial results), and it is the same if we just consider the oscillation exponent, or couples including the lacunarity exponent. One meets similar limitations for multifractal spectra: In the univariate setting supplied by the multifractal H\"older spectrum, the general form of functions which can be multifractal spectra is still open; nonetheless a partial result is available: functions which can be written as infima of a sequence of continuous functions are multifractal spectra \cite{JaffFrac}; additionally, as soon as two exponents are involved, extremly few results are available. For instance, if $f$ is a locally bounded function, define its { \em bivariate oscillation spectrum} as \[ { \cal D}_f (H, \beta ) = \dim \{ h: h_f (x) = H \quad \mbox{and} \quad {\cal O}_f (x) = \beta \} . \] Which functions of two variables $D (H, \beta)$ can be bivariate oscillation spectra is a completely open problem. \subsection{Mathematical results concerning the multifractal formalism} \label{mathresmf} We now consider a general setting where $h: {\mathbb R} \rightarrow {\mathbb R}$ is a pointwise exponent derived from a multiresolution quantity $ d_{\lambda} (= d_{j,k} )$ according to Def. \ref{derived}, and defined in space dimension $d$. The associated multifractal spectrum ${\cal D} $ is \[ {\cal D}(H) = \dim( \{ x: \quad h(x) = H\} ). \] The { \em support} of the spectrum is the image of the mapping $x\rightarrow h(x)$, i.e. the collection of values of $H$ such that \[ \{ x \in {\mathbb R}: h(x) = H \} \neq \emptyset \] (note that this denomination, though commonly used, is misleading, since it may not coincide with the mathematical notion of support of a function). The { \em leader scaling function} associated with the multiresolution quantities $(d_{j,k})$ is \begin{equation} \label{defscalond2} \forall q \in {\mathbb R}, \hspace{6mm} \zeta_f (q) = \displaystyle \liminf_{j \rightarrow + \infty} \;\; \frac{\log \left( 2^{-j} \displaystyle\sum_{k } | d_{j,k} |^q. \right) }{\log (2^{-j})}. \end{equation} Note that, in contradistinction with the wavelet scaling function, it is also defined for $p <0$. Referring to ``leaders'' in the name of the scaling function does not mean that the $d_{j,k}$ are necessarily obtained as wavelet leaders or wavelet $p$-leaders, but only to prevent any confusion with the wavelet scaling function. The { \em Legendre spectrum} is \begin{equation} \label{lspec} {\cal L} (H) : = \inf_{q \in {\mathbb R}} (1+qH - \zeta_f(q)) . \end{equation} As soon as relationships such as (\ref{carachqf}) hold, then the following upper bound is valid \begin{equation} \label{equ-tl} \forall H, \qquad {\cal D}(H) \leq {\cal L} (H) \end{equation} (see \cite{Jaffard2004} for particular occurrences of this statement, and \cite{Abel} for the general setting). However, for a number of synthetic processes with known ${\cal D}(H) $ (and for a proper choice of the multiresolution quantity), this inequality turns out to be an equality, in which case, we will say that the { \em multifractal formalism holds}. The leader scaling functions obtained using wavelet leaders or $p$-leaders can be shown to enjoy the same robustness properties as listed at the end of Section \ref{Univariate3}, see \cite{Abel} (it is therefore also the case for the Legendre spectrum). It follows from their mathematical and numerical properties that wavelet leader based techniques form the state of the art for real-life signals multifractal analysis. \begin{figure} \centering \includegraphics[scale=0.3]{univ_FC_PASCAL.PNG} \qquad \includegraphics[scale=0.3]{univ_Cad_PASCAL.PNG} \caption{Representation of scale function and the univariate H\"older Legendre spectra of the primitives of heart beat frequency (left) and cadence (right) of one marathon runner during the entire race. the multiresolution quantities used in these derivation are the wavelet leaders of the primitive of the data} \label{bilan_pl3} \end{figure} In applications, one cannot have access to the regularity exponent at every point in a numerically stable way, and thus ${\cal D}(H) $ is unaccessible; this explains why, in practice, ${\cal L} (H) $ is the only computationally available spectrum, and it is used as such in applications. However, information on the pointwise exponent may be inferred from the Legendre spectrum. Such results are collected in the following theorem, where they are stated in decreasing order of generality. \begin{Theo} \label{theo1} Let $h: {\mathbb R} \rightarrow {\mathbb R})$ be a pointwise exponent, and assume that it is derived from multiresolution quantities $d_{j,k}$ according to Def. \ref{derived}. The following results on $h$ hold: \begin{itemize} \item Let \begin{equation} \label{caracbeswav3hol2} h^{min} =\liminf_{j \rightarrow + \infty} \;\; \;\; \frac{ \log \left( \displaystyle \sup_{ k} d_{j,k} \right) }{\log (2^{-j})} \quad \mbox{ and } \quad h^{max} =\liminf_{j \rightarrow + \infty} \;\; \;\; \quad \frac{ \log \left( \displaystyle \inf_{ k} d_{j,k} \right) }{\log (2^{-j})} \end{equation} then \begin{equation} \label{encadrhol} \forall x \in {\mathbb R} \qquad h^{min} \leq h(x) \leq h^{max} . \end{equation} \item If the Legendre spectrum has a unique maximum for $H= c_1$, then \begin{equation} \label{almoster} \mbox{ for almost every } x, \qquad h (x) = c_1;\end{equation} \item If the leader scaling function \eqref{defscalond2} associated with the $d_{j,k}$ is affine, then $f$ is a monoh\"older function, i.e. \[ \exists H_0: \quad \forall x, \qquad h (x) =H_0,\] where $H_0$ is the slope of the leader scaling function. \end{itemize} \end{Theo} {\bf Remark:} The last statement asserts that, if $h$ is a pointwise exponent associated with a function $f$, then $f$ is a monoh\"older function. This result has important implications in modeling since it yields a numerically simple test, based on global quantities associated with the signal, and which yields the pointwise exponent everywhere. This is in strong contradistinction with the standard pointwise regularity estimators, see e.g. \cite{Bardet} and ref. therein, which are based on local estimates, and therefore on few data thus showing strong statistical variabilities, and additionally often assume that the data follow some a priori models. \vspace{0.5cm} {\bf Proof:} We first prove the upper bound in (\ref{encadrhol}). Let $\alpha > h^{max}$; there exists a sequence $j_n \rightarrow + \infty $ such that \[ \log \left(\inf_{ k } d_{j_n,k} \right) \geq \log (2^{-\alpha j_n}), \] so that at the scales $j_n$ all $d_\lambda$ are larger than $2^{-\alpha j_n}$. It follows from (\ref{carachqf}) that \[ \forall x, \qquad h (x) \leq \alpha , \] and the upper bound follows. The proof of the lower bound is similar (see e. g. \cite{Zuhai}). The second statement is direct consequence of the following upper bounds for the dimensions of the sets \begin{equation} \label{uppersets} E_{H}^+ = \{ h (x) \geq H \} \quad \mbox{ and } \quad E_{H}^- = \{ h (x) \leq H \} \end{equation} which are a slight improvement of (\ref{equ-tl}), see \cite{Zuhai}: \begin{prop} Let $h$ be a pointwise exponent derived from the multiresolution quantity $( d_{j,k})$. Then the following bounds hold: \begin{equation} \label{majdim3}\dim (E_{H}^-) \leq \inf_{q >0} (1+qH - \zeta_f(q)) \quad \mbox{ and } \quad \dim (E_{H}^+) \leq \inf_{q<0} (1+qH - \zeta_f(q)) \end{equation} \end{prop} Let us check how \eqref{almoster} follows from this result. Note that the first (partial) Legendre transform yields the increasing part of ${\cal L}(H)$ for $H \leq c_1$ and the second one yields the decreasing part for $H \geq c_1$. If ${\cal L}$ has a unique maximum for $H= c_1$, it follows from (\ref{majdim3}) that \[ \forall n , \qquad \dim (E_{c_1-1/n}^-) <1 \quad \mbox{ and } \quad \dim (E_{c_1+1/n}^-) <1. \] All of these sets therefore have a vanishing Lebesgue measure, which is also the case of their union. But this union is $\{ x: h(x) \neq c_1\} $. It follows that almost every $x$ satisfies $h(x) = c_1$. Finally, if the leader scaling function is affine, then its Legendre transform is supported by a point $H_0$ and takes the value $-\infty$ elsewhere. The upper bound (\ref{equ-tl}) implies that, if $H \neq H_0$ the corresponding isoregularity set is empty. In other words, $H_0$ is the only value taken by the pontwise exponent, and $f$ is a monoh\"older function. \vspace{0.5cm} { \bf Remarks:} If $h^{min}= h^{max}$, the conclusion of the first and last statement are the same. However, one can check that the condition $h^{min}= h^{max}$ is slightly less restrictive than requiring the leader scaling function to be affine (the two conditions are equivalent if, additionally, the $\liminf$ in (\ref{caracbeswav3hol2}) is a limit). The parameter $c_1$ defined in Theorem \ref{theo1} can be directly estimated using log-log plot (see \cite{MandMemor} and ref. therein), and, in practice it plays an important role in classification as we will see in the next section. When the multiresolution quantity used is the $p$-leaders of a function $f$, the associated exponent $c_1$ may depend on $p$, and we will mention this dependency and denote this parameter by $c_1(p, f)$. This is in contradistinction with the exponent $h^{min}$ defined by (\ref{caracbeswav3hol2}), which, in the case of functions with some uniform H\"older regularity, coincides with the exponent ${h^{min}_f}$ defined by (\ref{caracbeswav3hol}) for leaders and $p$-leaders, as shown by the following lemma; note that it is actually preferable to compute it using (\ref{caracbeswav3hol}), which has the advantages of being well defined without any a priori assumption on $f$. \begin{lem} Let $f: {\mathbb R} \rightarrow {\mathbb R} $ be such that ${h^{min}_f} >0$. Then the $h^{min} $ parameter computed using $p$-leaders all coincide with the ${h^{min}_f} $ computed using wavelet coefficients. \end{lem} Let us sketch the poor of this result. Suppose that ${h^{min}_f} >0$ and let $\alpha>0$ be such that $\alpha < {h^{min}_f}$. Then, the wavelet coefficients of $f$ satisfy \[ \exists C, \quad \forall j,k \qquad | c_{j,k} | \leq C 2^{-\alpha j} . \] Therefore the $p$-leaders of $f$ satisfy \[ \ell^{(p)}_\lambda \leq \left( \sum_{ \lambda' \subset 3 \lambda} (2^{-\alpha j'})^p \, \, 2^{j-j'} \right)^{1/p} \] \[ \leq \left( \sum_{ j' \geq j } 2^{-\alpha pj'} \, \, 2^{j-j'} \right)^{1/p} \leq C 2^{-\alpha j} ;\] it follows that the corresponding $p$-leader is smaller that $|c_{\lambda_n}|$ so that the $h^{min} $ computed using $p$-leaders is smaller that the one computed using wavelet coefficients. Conversely, by definition of ${h^{min}_f} $, there exists a sequence of dyadic intervals $c_{\lambda_n}$ of width decreasing to 0, and such that \[ |c_{\lambda_n}| \sim 2^{-{h^{min}_f} j_n} , \] and the corresponding $p$-leader is larger that $|c_{\lambda_n}|$ so that the $h^{min} $ computed using $p$-leaders is smaller that the one computed using wavelet coefficients. \vspace{0.5cm} The following result yields an important a priori bound on the dimensions of the singularity sets corresponding to negative regularity exponents, see \cite{PART1}. \begin{prop} Let $p > 0$, and let $f: {\mathbb R} \rightarrow {\mathbb R}$ be a function such that $\eta_f (p) >0$. Then its $p$-spectrum satisfies \[ \forall h , \qquad { \mathcal D}_p (H) \leq 1 + Hp\] \end{prop} Let us elaborate on the information supplied by the exponent $c_1 (p,f)$: A direct consequence of \eqref{almoster} is that, if a signal $f$ satisfies that the exponent $c_1 (p,f)$ takes the same value for $p_1 < p_2$, then this implies that the $p$-exponent satisfies that \[ \mbox{ for almost every } x, \qquad h^{p_1}_f (x) = h^{p_2}_f (x) ,\] which implies that the mapping $p \rightarrow h^{p_1}_f (x)$ is constant for $p \in [p_1, p_2]$; but, since the mapping $p\rightarrow h^{1/p}_f (x_0) $ is concave and increasing, see \cite{Bandt2015,Porqu2017}, it follows that this mapping is constant for $p$ small enough; as a consequence, the lacunarity exponent vanishes at $x$. Similarly, if, for a given $p$, $c_1 (p,f^{(-1)}) -c_1 (p,f) =1$, this implies that \[ \mbox{ for almost every } x, \qquad h^p_{f^{(-1)}} (x) = h^p_{f} (x) +1,\] and the same argument as above, see \cite{Bandt2015,Porqu2017}, yields the absence of oscillating singularities for almost every point. In other words, the computation of $c_1(p)$ yields a key information on the nature of the singularities a.e. of the signal, which we sumarize in the following statement, which will have implications in the next section for the analysis of marathon runners data. \begin{prop} \label{propae} Let $f: {\mathbb R} \rightarrow {\mathbb R}$ be a function in $L^p$. If \[ \exists q > p : \qquad c_1 (p, f) = c_1 (q, f),\] then for almost every $x$, $f$ has no lacunary singularity at $x$. If $f$ satisfies \[ \exists p : \qquad c_1 (p,f^{(-1)}) -c_1 (p,f) =1, \] then, for almost every $x$, $f$ has a canonical singularity at $x$. \end{prop} These two results are characteristic of signals that only contain canonical singularities, see Section \ref{classif}, and they also demonstrate that $c_1(p, f)$, which, in general, depends on the value of $p$ is intrinsic for such data (see a contrario \cite{Bandt2015} where the exponent $c_1(p, f)$ of lacunary wavelet series is shown to depend on the value of $p$, and \cite{Saes} where the same result is shown for random sums of pulses). Note that such results are available in the discrete wavelet approach only; they would not be possible using the WTMM or the MFDFA approaches, which do not allow to draw differences between various pointwise regularity exponents and therefore do not yield spectra fitted to different values of the $p$-exponent. To summarize, the advantages of the $p$-leader based multifractal analysis framework are: the capability to estimate negative regularity exponents, better estimation performances, and a refined characterization of the nature of pointwise regularities. One important argument in favor of multifractal analysis is that it supplies robust classification parameters, in contradistinction with pointwise regularity which can be extremely erratic. Consider for instance the example of a sample path of a L\'evy process without Brownian component (we choose this example because such processes now play a key role in statistical modeling): Its H\"older exponent is a random, everywhere discontinuous, function which cannot be numerically estimated or even drawn \cite{JaffLev}: In any arbitrary small interval $[ a, b]$ it takes all possible values $H\in [0, H^{max}] $. On the opposite, the multifractal spectrum (which coincides with the Legendre spectrum) is extremly simple and robust to estimate numerically: It is a deterministic linear function on the interval $[0, H^{max}] $ (with $D(H^{max}) = 1$). This example is by no means accidental: though one can simply construct stochastic processes with a random multifractal spectrum (consider for instance a Poisson process restricted to an interval of finite length), large classes of classical processes have simple deterministic multifractal spectra (and Legendre spectra), though no simple assumption which would guarantee this results is known. The determination of a kind of ``0-1 law'' for multifractal spectra, which would guarantee that, under fairly general assumptions, the spectrum almost surely is a deterministic function, is a completely open problem, and its resolution would greatly improve our understanding of the subject. Even in the case of Gaussian processes, though it is known that such processes can have a random H\"older exponent \cite{AyaGauss}, the possibility of having a random multifractal spectrum still is a open issue. \subsection{Generic results} \label{gener} Let us come back to the problem raised in Section \ref{Univariate1} of estimating the size of the H\"older singularity sets of increasing functions which led us to the key idea that the Hausdorff dimension is the natural way to estimate this size. One can wonder if the estimate \eqref{majspecinc} that we found for the multifractal spectrum is optimal. In 1999, Z. Buczolich and J. Nagy answered this question in a very strong way, showing that it is sharp for a { \em residual} set of continuous increasing functions, see \cite{BucNag}. What does this statement precisely mean? Let $E$ be the set of continuous increasing functions $f: \; {\mathbb R} \rightarrow {\mathbb R}$, endowed with the natural distance supplied by the $\sup$ norm. Then equality in \eqref{majspecinc} holds (at least) on a residual set in the sense of Baire categories, i.e. on a countable intersection of open dense sets. This first breakthrough opened the way to genericity results in multifractal analysis. They were the consequence of the important remark that scaling functions for $p >0$ can be interpreted as stating that $f$ belongs to an intersection of Sobolev spaces $E_\eta$ (in the case of the Kolmogorov scaling function) or of a variant of these spaces, the { \em oscillation spaces} in the case of the leader scaling function \cite{Jaf05}. One easily checks that $E_\eta$ is a complete metric space, and the Baire property therefore is valid (i.e. a countable intersection of open dense sets is dense). The question formulated by Parisi and Frisch in \cite{ParFri85}, can be reformulated in this setting: If equality in \eqref{equ-tl} cannot hold for { \em every} function in $E_\eta$ (since e.g. because it contains $C^\infty$ functions), nonetheless it holds on a residual set \cite{JafBaire}. This result found many extensions: The first one consists in replacing the genericity notion supplied by Baire's theorem by the more natural notion supplied by { \em prevalence}, which is an extension, in infinite dimensional function spaces of the notion of ``Lebesgue almost everywhere'', see \cite{Christ,HSY} for the defintion of this notion and its main properties, and \cite{FJ} for its use in the setting of multifractal analysis. The conclusions drawn in the Baire setting also hold in the prevalence setting, and raise the question of the determination of a stronger notion of genericity, which would imply both Baire and prevalence genericity, and which would be the ``right '' setting for the validity of the multifractal formalism. A natural candidate is supplied by the notion of { \em \hspace{-2mm} porosity }, see \cite{Linden}, but the very few results concerning multifractal analysis in this setting do not allow to answer this question yet. Note also that Baire and prevalence results have been extended to the $p$-exponent setting \cite{Fra}, which allows to deal with spaces of functions that are not locally bounded. Another key problem concerning the generic validity of the multifractal formalism concerns the question of taking into account the information supplied by negative values of $p$ in the scaling function \eqref{defscalond2}. The main difficulty here is that the scaling function does not define a function space any longer, and the ``right'' notion of genericity which should be picked is competely open: Though Baire and prevalence do not really require the setting supplied by a (linear) function space, nonetheless these notions are not fitted to the setting supplied by a given scaling function which includes negative values of $p$. In \cite{BaSeurGen} J. Barral and S. Seuret developed an alternative point of view which is less ``data driven'': They reinterpreted the question in the following way: Given a certain scaling function $\eta (p)$, they considered the problem of constructing an ad hoc function space which is taylored so that generically (for the Baire setting), functions in such a space satisfy the multifractal formalism for the corresponding scaling function, including its values for $p <0$ (and Legendre spectrum). Another limitation of the mathematical results of genericity at hand is that they are not able to take into account { \em selfsimilarity} information: In \eqref{defscalond}, in order to introduce a quantity which is always well-defined, and corresponds to a function space regularity index, the scaling function is defined by a $\liminf$. But, most of the time, what is actually observed on the data (and what is really needed in order to obtain a numerically robust estimate) is that this $\liminf$ actually is a true limit, which means that the $L^p$ averages of the data display exact power-law behaviours at small scales. Up to now, one has not been able to incorporate this type of information in the function space modeling developed. \subsection{Implications on the analysis of marathon runners data} \label{Marathon1} The increasing popularity of marathons today among all ages and levels is inherited from the human capacity to run long distances using the aerobic metabolism \cite{lieberman07a}, which led to a rising number of amateur marathon runners who end the 42,195 km between 2h40min and 4h40min. Therefore, even if nowadays, marathon running becomes “commonplace”, compared with ultra-distance races, this mythic Olympic race is considered to be the acme of duration and intensity \cite{Maron}. Running a marathon remains scary and complex due to the famous “hitting the wall” phenomenon, which is the most iconic feature of the marathon \cite{Berndsen}. This phenomenon was previously evaluated in a large-scale data analysis of late-race pacing collapse in the marathon \cite{Smyth21}; \cite{Smyth18} presented an analysis of 1.7 million recreational runners, focusing on pacing at the start and end of the marathon, two particularly important race stages. They showed how starting or finishing too quickly could result in poorer finish-times, because fast starts tend to be very fast, leading to endurance problems later, while fast finishes suggest overly cautious pacing earlier in the race \cite{Smyth18}. Hence, the definition of a single marathon pace is based on the paradigm that a constant pace would be the ideal one. However, in \cite{Billat19}, a 3 years study shows that large speed and pace variations are the best way to optimize performance. Marathon performance depends on pacing oscillations between non symmetric extreme values \cite{pycke2022marathon}. Heart rate (HR) monitoring, which reflects exercise intensity and environmental factors, is often used for running strategies in marathons. However, it is difficult to obtain appropriate feedback for only the HR value since, as we saw above, the cardiovascular drift occurs during prolonged exercise. Therefore, now we have still to investigate whether this pace (speed) variation has a fractal behavior and if so, whether this is the case for the runners's heart rate which remains a pacer for the runners who aim to keep their heart rate in a submaximal zone (60-80 $\%$ of the maximal heart rate) \cite{Maron}. Here, we hypothesized that marathonians acceleration (speed variation), cadence (number of steps per minute) and heart rate time series follow a multifractal formalism and could be described by a self similar functions. Starting in the 1990s, many authors demonstrated the fractal behavior of physiological data such as heart rate, arterial blood pressure, and breath frequency of human beings, see e.g. \cite{ABRY:2010:A,ivanov1999}. In 2005, using the Wavelet Transform Maxima Method, E. Wesfreid, V. L. Billat and Y. Meyer \cite{wesfreid05a} performed the first multifractal analysis of marathonians heartbeats. This study was complemented in 2009 using the DFA (Detrended Fluctuation Analysis) and wavelet leaders applied on a primitive of the signal \cite{wesfreid09a}. Comparing the outputs of these analyses is hasardous; indeed, as already mentioned, these methods are not based on the same regularity exponents: WTMM is adapted to the {\sl weak scaling exponent} \cite{MeyWVS}, DFA to the $p$-exponent for $p =2$ \cite{PART1,PART2}, and wavelet leaders to the Hölder exponent \cite{Jaffard2004}. In the following, we will propose a method of digital multifractal analysis of signals based on $p$-leaders, which, in some cases, can avoid performing fractional integrations (or primitives) and thus transform the signal. In \cite{wesfreid09a}, it was put in evidence that multifractal parameters associated with heart beat intervals evolve during the race when the runner starts to be deprived of glycogen (which is the major cause of the speed diminution at the end of the race. This study also revealed that fatigue decreases the running speed and affects the regularity properties of the signal which can be related with the feelings of the runner measured by the Rate of Perception of Exhaustion (RPE), according to the psychophysiological scale of Borg (mainly felt through the breathing frequency). In addition, there is a consistent decrease in the relationship between speed, step rate, cardiorespiratory responses (respiratory rate, heart rate, volume of oxygen consumed), and the level of Rate of Perception of Exhaustion (RPE), as measured by Borg's psychophysiological scale. The runner does not feel the drift of his heart rate, in contradistinction with his respiratory rate. These physiological data are not widely available and only heart rate and stride rate are the measures available to all runners for economic reasons. Moreover, these data are generated heartbeat by heartbeat and step by step. Our purpose in this section is to complement these studies by showing that a direct analysis on the data is possible if using $p$-leaders (previous studies using the WTMM or the standard leaders had to be applied to a primitive of the signal), and that they lead to a sharper analysis of the physiological modifications during the race. We complement the previous analyses in order to demonstrate the modifications of multifractal parameters during the race, and put in evidence the physiological impact of the intense effort after the 20th Km. For that purpose, we will perform a multifractal analysis based on $p$-leaders. We analyzed the heartbeat frequency of 8 marathon runners (men in the same age area). Fig.\ref{hmin} shows the determination of exponents ${h^{min}_f}$ for heartbeat frequency and cadence through a log-log regression; the regression is always performed between the scales j = 8 and j = 11 (i.e. between 26s and 3mn 25s), which have been identified as the pertinent scales for such physiological data, see \cite{ABRY:2010:A}. For most marathon runners, ${h^{min}_f}$ is negative, see Table \ref{tab1}, which justifies the use of $p$-leaders. We then compute the wavelet scaling function in order to determine a common value of $p$ for which all runners satisfy $\eta (p) >0$, see Fig. \ref{scaling function} where examples of wavelet scaling function are supplied for heartbeat frequency and cadence. In the case of heartbeat frequency, the computation of the 8 wavelet scaling functions yields that $p = 1$ and $p= 1.4$ can be picked. The corresponding $p$-leaders multifractal analysis is performed for these two values of $p$, leading to values of $c_1 (p)$ which are also collected in Table \ref{tab1}. \begin{table}[htb] \caption{ Multifractal Analysis of heartbeat frequency of marathon runners (Pr. : primitive)}\label{tab1} \begin{center} \resizebox{100mm}{!}{ \begin{tabular}{||c||*{10}{m{2cm}|}|} \hline\hline $ $ & $H_{min}$ & $H_{min}$ of the Pr. & $c_1$ for $p=1$ & $c_1$ for $p=1.4$ & $c_1$ of the Pr. for $p=1$ & $c_1$ of the Pr. for $p=1.4$ \\ \hline R1 & $-0,2768$ & $0,7232$ & $0,8099$ & $0,8064$ & $1,8242$ & $1,8213$ \\ \hline R2 & $-0,0063$ & $0,9937$ & $0,4564$ & $0,4043$ & $1,3926$ & $1,3509$ \\ \hline R3 & $-0,0039$ & $0,9961$ & $0,6856$ & $0,6625$ & $1,6942$ & $1,6351$ \\ \hline R4 & $-0,1633$ & $0,8367$ & $0,6938$ & $0,6785$ & $1,6653$ & $1,6636$ \\ \hline R5 & $-0,2434$ & $0,7566$ & $0,5835$ & $0,5689$ & $1,5401$ & $1,5224$ \\ \hline R6 & $-0,3296$ & $0,6704$ & $0,5809$ & $0,5636$ & $1,5644$ & $1,5500$ \\ \hline R7 & $0,1099$ & $1,1099$ & $0,5652$ & $0,5483$ & $1,4754$ & $1,4379$ \\ \hline R8 & $-0,5380$ & $0,4620$ & $0,3382$ & $0,2977$ & $1,2588$ & $1,2086$ \\ \hline\hline \end{tabular}} \end{center} \end{table} \begin{figure} \begin{center} \resizebox{80mm}{!}{ \includegraphics{runbilfc.png}} \end{center} \caption{Representation of the pair $(H_{min}, c_1(p))$ with $p=1$ deduced from the 1-spectrum of heart rate and computed for the entire race; $H_{min}$ appears as the most relevant classification parameter. The isolated point on the left corresponds to R8, the most trained runner.}\label{Fig6} \end{figure} In Fig. \ref{Fig6}, the value of the couple $({h^{min}_f}, c_1 (p) )$ is plotted (where we denote by $c_1 (p)$ the value of $H$ for which the maximum of the $p$-spectrum is reached). The values of $c_1 (p) $ are very close to $0.4$ whereas the values of ${h^{min}_f} $ notably differ, and are clearly related with the level of practice of the runners. Thus M8 is the only trail runner and improved his personal record on that occasion; he practices more and developed a very uneven way of running. Table \ref{tab1} shows that the values of $c_1 (p) $ do not notably differ for different values of $p$ and, when computed on a primitive of the signal, are shifted by 1. We are in the situation described in Prop. \ref{propae} and we conclude in the absence of oscillating singularities at almost every point. This result also shows that $c_1 (p)$, which may depend on the value of $p$ (see \cite{Bandt2015} where it is shown that it is the case for lacunary wavelet series), is intrinsic for such data. We will see in Section \ref{MultivariateMar} that a bivariate analysis allows to investigate further in the nature of the pointwise singularities of the data. \begin{figure} \centering \includegraphics[scale=0.12]{regdebFC.png}\qquad \includegraphics[scale=0.12]{regfinFC.png} \caption{Estimation of ${h^{min}_f}$ by log-log regression for the heart rate of a marathon runner at the beginning (50\% first part of the race) on the left and the end (25\% last part of the race) on the right. The clear difference of the values obtained shows that the exponent ${h^{min}_f}$ is well fitted to characterize the evolution of physiological rythms during the race. These data, together with the evolution of the parameter $c_1 (p)$, are collected in Fig.\ref{fig:my_label} with $p=1$.} \label{hmindebfin} \end{figure} \begin{figure} \centering \includegraphics[scale =0.35]{rundebfin.png} \caption{Evolution of the couple $(H_{min}, c_1(p))$ with $p=1$ deduced from the 1-spectrum of the heart rate between the beginning (in blue) and the end (in red) of the marathon: the evolutions are similar except for three runners: R3 and R6 who had great difficulties and R7 who is the least experienced runner with a much longer running time.} \label{fig:my_label} \end{figure} We now consider the evolution of the multifractality parameters during a marathon: at about the 25th Km (circa 60 \% of the race) runners feel an increased penibility on the RPE Borg scale. Therefore we expect to find two regimes with different parameters before and after this moment. This is put in evidence by Fig. \ref{fig:my_label} which shows the evolution of the multifractality parameters during the first half and the last fourth of the marathon thus putting in evidence the different physiological reactions at about the 28th Km. From the evolution of the multifractal parameters between the beginning and the end of the marathon race, we can distinguish between the less experimented marathon runners, whichever their level of fitness, and those who know how to self pace their race. Indeed, according to the evolution of the couple $({h^{min}_f}, c_1 (p))$, the less experimented (R 7) loosed the regularity of his heart rate variation. This shows that the mararathon running experience allows to feel how to modulate the speed for a conservative heart rate variability. From the evolution of the multifractal parameters between the beginning and the end of the marathon race, we can distinguish between the less experimented marathon runners, whichever their level of fitness and those who know how to self pace their race. In \cite{pycke2022marathon} its was shown that the best marathon performance was achieved with a speed variation between extreme values. Furthermore, a phsyiological steady state (heart rate and other cardiorespiratory variables), are obtained with pace variation \cite{BillatPet}. This conclusion is in opposition with the less experimented runners beliefs that the constant pace is the best, following the mainstream non scientific basis recommendations currently available on internet. In Section \ref{MultivariateMar} we will investigate the additional information which is revealed by the joint analysis of several physiological data. \section{Multivariate multifractal analysis} \label{Multivariate} Up to now, in most applications, multifractal analysis was performed in univariate settings, (see a contrario \cite{Lux}), which was mostly due to a lack of theoretical foundations and practical analysis tools. Our purpose in this section is to provide a comprehensive survey of the recent works that started to provide these foundations, and to emphasize the mathematical questions which they open. In particular, multivariate spectra also encode on specific data construction mechanisms. Multivariate multifractal analysis deals with the joint multifractal analysis of several functions. For notational simplicity, we assume in the following that we deal with two functions $f_1$ and $f_2$ defined on ${\mathbb R}^d$ and that, to each function is associated a pointwise regularity exponent $h_1 (x) $ and $h_2 (x)$ (which need not be the same). \subsection{Multivariate spectrum} \label{MultivariateSpec} On the mathematical side, the main issue is to understand how the isoregularity sets \[ E_{f_1} (H_1) = \{ x: h_1(x) = H_1 \} \quad \mbox{ and } \quad E_{f_2} (H_2) = \{ x: h_2(x) = H_2 \} \] of each function are ``related''. A natural way to translate this loose question into a precise mathematical problem is to ask for the determination of the { \em multivariate multifractal spectrum} defined as the two-variables function \begin{equation} { \cal D}_{(f_1,f_2)} (H_1, H_2) = \dim (\{ x: h_1(x) = H_1 \mbox{ and } h_2(x) = H_2 \} ). \end{equation} this means that we want to determine the dimension of the intersection of the two isoregularity sets $E_{f_1} (H_1) $ and $E_{f_2} (H_2) $. The determination of the dimension of the intersection of two fractal sets usually is a difficult mathematical question, with no general results available, and it follows that few multivariate spectra have been determined mathematically, see e.g. \cite{Barr,Schmel} for a joint analysis of invariant measures of dynamical systems. One can also mention correlated and anticorrelated binomial cascades, see Section \ref{BrwonianMT} for the definition of these cascades, and \cite{PRSA} for the determination of bivariate spectra when two of these cascades are considered jointly. On the mathematical side, two types of results often show up. A first category follows from the intuition supplied by intersections of smooth manifolds: In general, two surfaces in ${\mathbb R}^3$ intersect along a curve and, more generally, in ${\mathbb R}^d$, manifolds intersect \emph{generically} according to the \emph{sum of codimensions rule}: \[ \dim (A \cap B) = \min (\dim A + \dim B -d, -\infty) \] (i.e. the ``codimensions'' $d-\dim A $ and $d- \dim B$ add up except if the output is negative, in which case we obtain the emptyset). This formula is actually valid for numerous examples of fractal sets, in particular when the Hausdorff and Packing dimensions of one of the sets $A$ or $B$ coincide (e.g. for general Cantor sets) \cite{Matti}; in that case ``generically'' has to be understood in the following sense: For a subset of positive measure among all rigid motions $\sigma$, $ \dim (A \cap \sigma (B) ) = \min (\dim A + \dim B -d, -\infty)$. However the coincidence of Hausdorff and Packing dimensions needs not be satisfied by isoregularity sets, so that such results cannot be directly applied for many mathematical models. The only result that holds in all generality is the following: if $A$ and $B$ are two Borel subsets of ${\mathbb R}^d$, then, for a generic set of rigid motions $\sigma$, $ \dim(A \cap \sigma (B) ) \geq \dim A + \dim B -d$. This leads to a first rule of thumb for multivariate multifractal spectra: When two functions are randomly shifted, then their singularity sets will be in ``generic'' position with respect to each other, yielding \[ { \cal D}_{(f_1,f_2)} (H_1, H_2) \geq { \cal D}_{f_1} (H_1) + { \cal D}_{f_2} (H_2) -d. \] In practice, this result suffers from two limitations: the first one is that, usually, one is not interested in randomly shifted signals but on the opposite for particular configurations where we expect the conjunction of singularity sets to carry relevant information. Additionally, for large classes of fractal sets, the {\em sets with large intersection}, the codimension formula is not optimal as they satisfy \[ \dim (A \cap B) = \min ( \dim A , \; \dim B ). \] While this alternative formula may seem counterintuitive, general frameworks where it holds were uncovered, cf. e.g., \cite{Fal93,Dur,BaSeurBraz} and references therein. This is notably commonly met by { \em limsup sets}, obtained as follows: There exists a collection of sets $A_n$ such that $A$ is the set of points that belong to an infinite number of the $A_n$. This is particularly relevant for multifractal analysis where the singularity sets $E_{H}^-$ defined in (\ref {uppersets}) often turn out to be of this type: It is the case for L\'evy processes or random wavelet series, see e.g. \cite{Jaf6,AJ02,{JaffLev}}). For multivariate multifractal spectra, this leads to an alternative formula \begin{equation} \label{largint2} { \cal D}_{(f_1,f_2)} = \min ( { \cal D}_{f_1} (H_1), { \cal D}_{f_2} (H_2)) \end{equation} expected to hold in competition with the codimension formula, at least for the sets $E_{H}^-$. The existence of two well motivated formulas in competition makes it hard to expect that general mathematical results could hold under fairly reasonable assumptions. Therefore, we now turn towards the construction of multifractal formalisms adapted to a multivariate setting, first in order to inspect if this approach can yield more intuition on the determination of multivariate spectra and, second, in order to derive new multifractality parameters which could be used for model selection and identification, and also in order to get some understanding on the ways that singularity sets of several functions are correlated. In order to get some intuition in that direction, it is useful to start with a probabilistic interpretation of the multifractal quantities that were introduced in the univariate setting. \subsection{Probabilistic interpretation of scaling functions} \label{MultivariateWSF} We consider the following probabilistic toy-model: We assume that, for a given $j$, the wavelet coefficients $(c_{ j,k})_{k\in {\mathbb Z}}$ of the signal considered share a common law $X_j$ and display short range memory, i.e. become quickly decorrelated when the wavelets $\psi_{j,k}$ and $\psi_{j,k'}$ are located far away (i.e. when $k-k'$ gets large); then, the wavelet structure functions (\ref{equ-WSF}) can be interpreted as an empirical estimation of ${\mathbb E} ( |X_j |^p)$, i.e. the moments of the random variables $X_j $, and the wavelet scaling function characterizes the power law behaviour of these moments (as a function of the scale $2^{-j}$). This interpretation is classically acknowledged for signals which display some stationarity, and the vanishing moments of the wavelets reinforce this decorrelation even if the initial process displays long range correlations, see e.g. the studies performed on classical models such as fBm (\cite{agf95} and ref. therein). We will not discuss the relevance of this model; we just note that his interpretation has the advantage of pointing towards probabilistic tools when one shifts from one to several signals, and these tools will allow to introduce natural classification parameters which can then be used even when the probabilistic assumptions which led to their introduction have no reason to hold. From now on, we consider two signals $f_1$ and $f_2$ defined on ${\mathbb R}$ (each one satisfying the above assumptions) with wavelet coefficients respectively $c^1_{ j,k}$ and $c^2_{ j,k}$. The ``covariance'' of the wavelet coefficients at scale $j$ is estimated by the empirical { correlations } \begin{equation} \label{crosscorr} \mbox{ for } m,n = 1,2, \qquad S_{m,n} (j) = 2^{-j} \displaystyle\sum_{ k } c^m_{ j,k} c^n_{ j,k}. \end{equation} Log-log regressions of these quantities (as a function of $\log (2^{-j})$ allow to determine if some power-law behaviour of these auto-correlations (if $m =n$) and cross-correlations (if $m \neq n$) can be put in evidence: When these correlations are found to be significantly non-negative, one defines the { \em scaling exponents} $H_{m,n}$ implicitly by \[ S_{m,n} (j)\sim 2^{-H_{m,n} j} \] in the limit of small scales. Note that, if $m=n$, the exponent associated with the auto-correlation simply is $\eta_f (2)$ and is referred to as the { \em Hurst exponent} of the data. Additionally, the { \em wavelet coherence function} is defined as \[ C_{1,2} (j) = \frac{S_{1,2} (j)}{\sqrt{S_{1,1} (j) S_{2,2} (j) }}. \] It ranges within the interval $[-1, 1]$ and quantifies, as a scale-dependent correlation coefficient, which scales are involved in the correlation of the two signals, see \cite{CRAS2019, Whitc}. Note that probabilistic denominations such as ``auto-correlation'', ``cross-correlations'' and ``coherence function'' are used even if no probabilistic model is assumed, and used in order to derive scaling parameters obtained by log-log plot regression which can prove powerful as classification tools. As an illustration, we estimated these crosscorrelations concerning the following couples of data recorded on marathon runners: heart-beat frequency vs. cadence, and cadence vs. acceleration, see Fig. \ref{CrossCor}. In both cases, no correlation between the wavelet coefficients at a given scale is put in evidence. Therefore, this is a situation where the additional bonus brought by measuring multifractal correlations is needed. Indeed, if the cross-correlations of the signals do not carry substantial information, this does not imply that the singularity sets of each signal are not related (as shown by the example supplied by { \em Brownian motions in multifractal time}, see below in Section \ref{BrwonianMT}). In that case, a natural idea is to look for correlations that would be revealed by the multiscale quantities associated with pointwise exponents rather than by wavelet coefficients. \begin{figure} \centering \includegraphics[scale=0.3]{Correlation_Cadence_FC_Pascal.png} \quad \includegraphics[scale=0.3]{Correlation_Acc_Cadence_Pascal.png} \caption{ Wavelet coherence between heart-beat frequency and cadence (left) and between acceleration and cadence (right).} \label{CrossCor} \end{figure} \subsection{Multivariate multifractal formalism} \label{MultivariateMF} The idea that leads to a multivariate multifractal formalism is quite similar as the one which led us from wavelet scaling functions to leaders and $p$-leaders scaling functions: One should incorporate in the cross-correlations the multiscale quantities which allow to characterize pointwise regularity, i.e. replace wavelet coefficients by wavelet leaders in \eqref{crosscorr}. Suppose that two pointwise regularity exponents $h_1 $ and $ h_2 $ defined on ${\mathbb R}$ are given. We assume that each of these exponents can be derived from corresponding multiresolution quantities $d^1_{j,k}$, and $d^2_{j,k}$ according to \eqref{carachqf}. A {\em grandcanonical multifractal formalism} allows to estimate the {joint spectrum} ${ \mathcal D}(H_1, H_2)$ of the couple of exponents $(h_1 , h_2) $ as proposed in \cite{Meneveau90}. In the general setting provided by multiresolution quantities, it is derived as follows: The {\em multivariate structure functions} associated with the couple $(d^1_{j,k}, d^2_{j,k})$ are defined by \begin{equation} \label{equ-WSF2} \forall r= (r_1, \ r_2) \in {\mathbb R}^2, \qquad S (r,j) = 2^{-j} \displaystyle\sum_{ k } ( d^1_{j,k} )^{r_1} ( d^2_{j,k} )^{r_2}, \end{equation} see \cite{Porqu2017,Benslim} for the seminal idea of proposing such multivariate multiresolution quantities as building blocks of a { \em grandcanonical formalism}. Note that they are defined as a cross-correlation, which would be based on the quantities $ d^1_{j,k} $ and $ d^2_{j,k} $, with the extra flexibility of raising them to arbitrary powers, as is the case for univariate structure functions. The corresponding {\em bivariate scaling function} is \begin{equation} \label{defscalond3} \zeta (r) = \displaystyle \liminf_{j \rightarrow + \infty} \;\; \frac{\log \left( S (r,j) \right) }{\log (2^{-j})}. \end{equation} The {\em bivariate Legendre spectrum} is obtained through a 2-variable Legendre transform \begin{equation} \label{formult1} \forall H = (H_1, H_2) \in {\mathbb R}^2, \qquad { \mathcal L} (H) = \inf_{r\in {\mathbb R}^2 } (1-\zeta (r) + H \cdot r), \end{equation} where $H\cdot r$ denotes the usual scalar product in ${\mathbb R}^2$. Apart from \cite{Meneveau90}, this formalism has been investigated in a wavelet framework for joint H\"older and oscillation exponents in \cite{ABJM}, in an abstract general framework in \cite{pey2004}, and on wavelet leader and $p$-leader based quantities in \cite{Bandt2015,Porqu2017}. \vspace{0.5cm} { \bf Remark:} The setting supplied by orthonormal wavelet bases is well fitted to be extended to the multivariate setting, because the multiresolution quantities $d_\lambda$ are defined on a preexisting (dyadic) grid, which is shared by both quantities. Note that this is not the case for the WTMM, where the multiresolution quantities are defined at the local maxima of the continuous wavelet transform (see \eqref{foncechbis}), and these local maxima differ for different signals; thus, defining multivariate structure functions in this setting would lead to the complicated questions of matching these local maxima correctly in order to construct bivariate structure functions similar to \eqref{equ-WSF2}. \vspace{0.5cm} The multivariate multifractal formalism is backed by only few mathematical results. A first reason is that, as already mentioned, the Legendre spectrum does not yield in general an upper bound for the multifractal spectrum, and this property is of key importance in the univariate setting. Another drawback is that, in constradistinction with the univariate case, the scaling function \eqref{defscalond3} has no function space interpretation. It follows that there exists no proper setting for genericity results except if one defines a priori this function space setting (as in \cite{Benslim,Moez} where generic results are obtained in couples of function spaces endowed with the natural norm on a product space). We meet here once again the problem of finding a ``proper'' genericity setting that would be fitted to the quantities supplied by scaling functions. We now list several positive results concerning multivariate Legendre spectra. The following result of \cite{ACHA2018} shows how to recover the univariate Legendre spectra from the bivariate one. \begin{prop} \label{propbivleg} Let $d^1_{j,k}$, and $d^2_{j,k}$ be two multiresolution quantities associated with two pointwise exponents $h_1 (x) $ and $ h_2 (x) $. The associated uni- and bi-variate Legendre spectra are related as follows: \[ {\mathcal L}_1 (H_1) = \sup_{H_2} \; {\mathcal L} ( H_1, H_2) \quad \mbox{and} \quad {\mathcal L}_2 (H_2) = \sup_{H_1}\; {\mathcal L} ( H_1, H_2). \] \end{prop} This property implies that results similar to Theorem \ref{theo1} hold in the multivariate setting. \begin{coro} \label{coro1} Let $d^1_{j,k}$ and $d^2_{j,k}$ be two multiresolution quantities associated with two pointwise exponents $h_1 (x) $ and $ h_2 (x) $. The following results on the couple $(h_1(x), h_2(x)) $ hold: \begin{itemize} \item If the bivariate Legendre spectrum has a unique maximum for $(H_1, H_2)= (c_1,c_2)$, then \begin{equation} \label{almoster2} \mbox{ for almost every } x, \qquad h_1 (x) = c_1 \quad \mbox{ and } \quad h_2 (x) = c_2 . \end{equation} \item If the leader scaling function is affine then \[ \exists (c_1, c_2), \quad \forall x, \qquad h_1 (x) = c_1 \quad \mbox{ and } \quad h_2 (x) = c_2.\] \end{itemize} \end{coro} Note that the fact that the leader scaling function is affine is equivalent to the fact that the bivariate Legendre spectrum is supported by a point. In that case, if the exponents $h_1$ and $h_2$ are associated with the functions $f_1$ and $f_2$, then they are monoh\"older functions. \vspace{0.5cm} { \bf Proof:} The first point holds because, if the bivariate Legendre spectrum has a unique maximum, then, its projections on the $H_1$ and the $H_2$ axes also have a unique maximum at respectively $H_1 = c_1$ and $ H_2= c_2$ and Proposition \ref{propbivleg} together with Theorem \ref{theo1} imply (\ref{almoster2}). As regards the second statement, one can use Proposition \ref{propbivleg}: If the bivariate scaling function is affine, then ${\mathcal L} ( H_1, H_2)$ is supported by a point, so that Proposition \ref{propbivleg} implies that it is also the case for univariate spectra ${\mathcal L} ( H_1)$ and ${\mathcal L} (H_2)$, and Theorem \ref{theo1} then implies that $h_1$ is constant and the same holds for $h_2$. \vspace{0.5cm} Recall that, in general, the bivariate Legendre spectrum does not yield an upper bound for the multifractal spectrum (in contradistinction with the univariate case), see \cite{PRSA} where a counterexample is constructed; this limitation raises many open questions: Is there another way to construct a Legendre spectrum which would yield an upper bound for ${ \cal D} (H_1, H_2)$? which information can actually be derived from the Legendre spectrum? A first positive result was put in light in \cite{PRSA}, where a notion of ``compatibility'' between exponents is put in light and is shown to hold for several models: When this property holds, then the upper bound property is satisfied. It is not clear that there exists a general way to check directly on the data if it is satisfied; however, an important case where it is the case is when the exponents derived are the H\"older exponent and one of the ``second generation exponents'' that we mentioned, see \cite{Bandt2015,Porqu2017}. In that case, the upper bound property holds, and it allows to conclude that the signal does not display e.g. oscillating singularities, an important issue both theoretical and practical. Let us mention a situation where this question shows up: In \cite{Balanca}, P. Balanca showed the existence of oscillating singularities in the sample of some L\'evy processes and also showed that they are absent in others (depending on the L\'evy measure which is picked in the construction); however, he only worked out several examples, and settling the general case is an important issue; numerical estimations of such bivariate spectra could help to make the right conjectures in this case. The general results listed in Corollary \ref{coro1} did not require assumptions on correlations between the exponents $h_1$ and $h_2$. We now investigate the implications of such correlations on the joint Legendre spectrum. For that purpose, let us come back to the probabilistic interpretation of the structure functions (\ref{equ-WSF2}) in terms of cross-correlation of the $ ( d^1_{j,k} )^{r_1}$ and $ ( d^2_{j,k} )^{r_2} $. As in the univariate case, if we assume that, for a given $j$, the multiresolution quantities $d^1_{j,k} $ and $d^2_{j,k} $ respectively share common laws $X^1_j$ and $X^2_j$ and display short range memory, then (\ref{equ-WSF2}) can be interpreted as an empirical estimation of ${\mathbb E} ( |X^1_j |^{r_1} |X^2_j |^{r_2})$. If we additionally assume that the $(d^1_{j,k}) $ and $(d^2_{j,k} )$ are independent, then we obtain \[ S (r,j) = {\mathbb E} ( |X^1_j |^{r_1} |X^2_j |^{r_2})= {\mathbb E} ( |X^1_j |^{r_1} ) \cdot {\mathbb E} ( |X^2_j |^{r_2}) , \] which can be written \begin{equation} \label{strcubiz} S (r_1, r_2,j) = S^1 (r_1,j) S^2 (r_2, j) . \end{equation} Assuming that $\liminf $ in (\ref{defscalond3}) actually is a limit, we obtain $S (r_1, r_2,j) \sim 2^{ -(\zeta^1(r_1)+\zeta^2(r_2)) j}$ yielding $ \zeta( r_1, r_2) = \zeta^1 (r_1) + \zeta^2 (r_2). $ Applying (\ref{formult1}), we get \[ {\mathcal L} (H_1, H_2 ) = \inf_{(r_1, r_2)\in {\mathbb R}^2 } (1-\zeta^1( r_1) + \zeta^2( r_2) + H_1 r_1 + H_2 r_2) \] \[ = \inf_{r_1 } (1-\zeta^1( r_1) + H_1 r_1 ) + \inf_{r_2 } (1- \zeta^2( r_2) + H_2 r_2) -1, \] which leads to \begin{equation} \label{spectindep} {\mathcal L} (H_1, H_2) = {\mathcal L} (H_1) + {\mathcal L} (H_2) -1. \end{equation} Thus, under stationarity and independence, the codimension rule applies for the multivariate Legendre spectrum. In practice, this means that any departure of the Legendre spectrum from (\ref{spectindep}), which can be checked on real-life data, indicates that one of the assumptions required to yield (\ref{spectindep}) (either stationarity or independence) does not hold. \begin{figure} \centering \includegraphics[scale=0.3]{bivarie_FC_Cad_PASCAL.png} \quad \includegraphics[scale=0.3]{bivarie_Cad_Acc_PASCAL.png} \caption{On the left, the bivariate multifractal spectrum between heart-beat frequency primitive and cadence primitive are shown, and, on the right, the bivariate multifractal spectrum between acceleration and cadence with fractional integral of order 1.5 are shown. This demonstrates the strong correlation between the pointwise singularities of the two data: indeed the bivariate spectra are almost carried by a segment, and a bivariate spectrum carried by a line $H_2 = a H_1 + b$ indicates a perfect match between the pointwise exponents accrording to the same relationship: $\forall x$, $ h_2 (x) = a h_1(x) + b$ } \label{SpectBivar} \end{figure} \label{intfrac} As a byproduct, we now show that multivariate multifractal analysis can give information on the nature of the singularities of { \em one} signal, thus complementing results such as Proposition \ref{propae} which yielded almost everywhere information of this type. Let us consider the joint multifractal spectrum of a function $f$ and its fractional integral of order $s$, denoted by $f^{(-s)}$. If $f$ only has canonical singularities, then the H\"older exponent of $f^{(-s)}$ satisfies $ \forall x_0$, $h_{f^{(-s)}}(x_0) = h_{f}(x_0) +s $, so that the joint Legendre spectrum is supported by the line $H_2 = H_1 +s$. In that case, the synchronicity assumption is satisfied and one can conclude that the joint multifractal spectrum is supported by the same segment; a contrario, a joint Legendre spectrum which is not supported by this line is interpreted as the signature of { \em oscillating singularities} in the data, as shown by the discussion above concerning the cases where the upper bound for bivariate spectra holds. Figs. \ref{BivarFC}, \ref{BivarCad} and \ref{BivarAcc} illustrate this use of bivariate multifractal analysis: In each case, a signal and its primitive are jointly analyzed: The three signals are collected on the same runner and the whole race is analyzed. Fig. \ref{BivarFC} shows the analysis of heartbeat, Fig. \ref{BivarCad} shows the cadence and Fig. \ref{BivarAcc} shows the acceleration. In the first case, the analysis is performed directly on the data using a $p$-exponent with $p= 1$, whereas, for the two last ones, the analysis is performed on a fractional integral of order $1/2$. In each case, the results yield a bivariate Legendre spectrum supported by the segment $H_2 = H_1 + s$, which confirms the almost everywhere results obtained in Section \ref{Marathon1}: The data only contain canonical singularities. \begin{figure} \centering \includegraphics[scale = 0.1]{CEDRIC_FC_spect_p1.png} \caption{Bivariate $1$-spectrum of heartbeat frequency and its primitive: the bivariate spectrum lines up perfectly along the line $H_2 = H_1 +1$. } \label{BivarFC} \end{figure} \begin{figure} \centering \includegraphics[scale = 0.1]{CEDRIC_Cad_spect_05int.png} \caption{Bivariate H\"older spectrum of fractional integrals order $1/2$ and $3/2$ of cadence: the bivariate spectrum lines up perfectly along the line $H_2 = H_1 +1$.} \label{BivarCad} \end{figure} \begin{figure} \centering \includegraphics[scale = 0.1]{CEDRIC_Acc_spect_05int.png} \caption{Bivariate H\"older spectrum of fractional integrals of order $1/2$ and $3/2$ of acceleration: the bivariate spectrum lines up perfectly along the line $H_2 = H_1 +1$.} \label{BivarAcc} \end{figure} \subsection{Fractional Brownian motions in multifractal time} \label{BrwonianMT} In order to put in light the additional information between wavelet correlations and bivariate scaling functions (and the associated Legendre spectrum), we consider the model supplied by Brownian motion in multifractal time, which has been proposed by B. Mandelbrot \cite{Mandelbrot1997,MandCal97} as a simple model for financial time series: Instead of the classical Brownian model $B(t)$, he introduced a time change (sometimes referred to as a { \em \hspace{-2mm} subordinator}) \[ B(f(t))= (B \circ f)(t) \] where the irregularities of $f$ model the fluctuations of the intrinsic ``economic time'', and typically is a multifractal function. In order to be a ``reasonable'' time change, the function $f$ has to be continuous and strictly increasing; such functions usually are obtained as distribution functions of probability measures $d\mu$ supported on ${\mathbb R}$ (or on an interval), and which have no atoms (i.e. $\forall a \in {\mathbb R}$, $\mu (a) =0$); typical examples are supplied by deterministic or random cascades, and this is the kind of models that were advocated by B. Mandelbrot in \cite{Mandelbrot1997}. Such examples will allow to illustrate the different notions that we introduced, and the additional information which is put into light by the bivariate Legendre spectrum and is absent from wavelet correlations. Let us consider the slightly more general setting of one fBm of Hurst exponent $\alpha$ (the cas of Brownian motion corresponds to $\alpha = 1/2$) modified by a time change $f$. In order to simplify its theoretical multifractal analysis, we take for pointwise regularity exponent the H\"older exponent and we make the following assumptions of $f$: We assume that it has only canonical singularities and that, if they exist, the non-constant terms of the Taylor polynomial of $f$ vanish at every point even if the H\"older exponent at some points is larger than 1 (this is typically the case for primitives of singular measures). In that case, classical uniform estimates on increments of fBm, see \cite{KahCamb} imply that \begin{equation} \label{holponcbrown} \mbox{ a.s. } \quad \forall t, \qquad h_{B\circ f} (t) = \alpha h_f (t), \end{equation} so that \[ \mbox{ a.s. } \quad \forall H, \qquad { \cal D}_{B\circ f} (H) = { \cal D}_{f} (H/ \alpha ) ; \] Note that the simple conclusion \eqref{holponcbrown} may fail if the Taylor polynomial is not constant at every point, as shown by the simple example supplied by $f (x) = x$ on the interval $[0,1]$. We now consider $B_1\circ f$ and $B_2\circ f$: two independent fBm modified by the { \em same deterministic time change} $f$ (with the same assumptions as above). It follows from (\ref{holponcbrown}) that, with probability 1, the H\"older exponents of $B_1\circ f$ and $B_2\circ f$ coincide everywhere, leading to the following multifractal spectrum, which holds almost surely: \begin{equation} \label{specbivarbrown} \left\{ \begin{array}{rl} \mbox{ if } H_1 = H_2, & { \cal D}_{(B_1\circ f,B_2\circ f )} (H_1, H_2 ) = { \cal D}_{f} \left(\displaystyle\frac{H_1}{\alpha}\right) \\ & \\ \mbox{ if } H_1 \neq H_2, & { \cal D}_{(B_1\circ f,B_2\circ f )} (H_1, H_2 ) = -\infty . \end{array} \right. \end{equation} Fig. \ref{fbmSpect} gives a numerical backing of this result: The Legendre spectrum numerically obtained corresponds to the theoretical multifractal spectrum. Let us give a non-rigourous argument which backs this result: The absence of oscillating singularities in the data implies that the maxima in the wavelet leaders are attained for a $\lambda'$ close to $\lambda$, so that the wavelet leaders of a given magnitude will be close to coincide for both processes, and therefore the bivariate structure functions (\ref{equ-WSF2}) satisfy \[ S_f (r,j) = 2^{-j} \displaystyle{\sum}_{ \lambda \in \Lambda_j } ( d^1_\lambda )^{r_1} ( d^2_\lambda )^{r_2} \sim 2^{-dj} \displaystyle{\sum}_{ \lambda \in \Lambda_j } ( d^1_\lambda )^{r_1 + r_2} \] so that \[ \mbox{ a.s. }, \quad \forall r_1, r_2, \qquad \tilde{\zeta} (r_1, r_2) = \zeta (r_1 + r_2). \] where $\tilde{\zeta}$ is the bivariate scaling function of the couple $(B_1\circ f,B_2\circ f )$ and ${\zeta}$ is the univariate scaling function of $B_1\circ f$. Taking a Legendre transform yields that the bivariate Legendre spectrum $ { \cal L} (H_1, H_2)$ also satisfies a similar formula as (\ref{specbivarbrown}), i.e. \begin{equation} \label{specbivarbrown2} \mbox{ a.s. }, \quad \forall H_1, H_2, \left\{ \begin{array}{rl} \mbox{ if } H_1 = H_2, & { \cal L}_{(B_1\circ f,B_2\circ f )} (H_1, H_2 ) = { \cal L}_{f} \left(\displaystyle\frac{H_1}{\alpha}\right) \\ & \\ \mbox{ if } H_1 \neq H_2, & { \cal L}_{f} \left(H_1, H_2\right) = -\infty . \end{array} \right. \end{equation} Let us now estimate the wavelet cross correlations. Since $f$ is deterministic, the processes $B_1\circ f $ and $B_2\circ f $ are two independent centered Gaussian processes. Their wavelet coefficients $c^1_{ j,k} $ and $ c^2_{ j,k}$ therefore are independent centered Gaussians, and, at scale $j$ the quantity \[ \qquad S_{m,n} (j) = 2^{-j} \displaystyle\sum_{ k } c^1_{ j,k} c^2_{ j,k} \] is an empirical estimation of their covariance, and therefore vanishes (up to small statistical fluctuation). In contradistinction with the bivariate spectrum, the { wavelet cross correlations } reveal the decorrelation of the processes but does not yield information of the correlation of the singularity sets. In order to illustrate these results, we will use for time change the distributuon function of a binomial cascade $\mu_p$ carried on $[0,1]$. Let $p \in (0,1)$; $\mu_p$ is the only probability measure on $[0, 1]$ defined by recursion as follows: Let $\lambda \subset [0, 1]$ be a dyadic interval of length $2^{-j}$; we denote by $\lambda^+$ and $\lambda^-$ respectively its two ``children'' of length $2^{-j-1}$, $\lambda^+$ being on the left and $\lambda^-$ being on the right. Then, $\mu_{p}$ is the only probability measure carried by $[0, 1]$ and satisfying \[ \mu_p (\lambda^{+}) = p \cdot\mu_p (\lambda) \quad \mbox{ and } \quad \mu_p (\lambda^{-}) = (1-p) \cdot \mu_p (\lambda) . \] Then the corresponding time change is the function \[ \forall x \in [0, 1] \qquad f_{\mu_p} (x) = \mu_p ([0, x]). \] \begin{figure} \centering \includegraphics[scale = 0.15]{binom.png} \caption{Binomial measure with $p=1/4$ (left) and its repartition function (right) which is used as the time change in Fig. \ref{fbmCT}.} \label{binom} \end{figure} In Fig. \ref{binom}, we show the binomial cascade $\mu_{1/4}$ and its distribution function, and in Fig. \ref{fbmCT} we use this time change composed with a fBm of Hurst exponent $\alpha=0.3$. \begin{figure} \centering \includegraphics[scale = 0.25]{fbmCTcross.png} \caption{Cross-correlation of the wavelet coefficients of two independent fBm with the same time change : the distribution function of the binomial measure $\mu_p$ with $p=1/4$. The Cross-correlation reflects the independence of the two processes.} \label{fbmCross} \end{figure} \begin{figure} \centering \includegraphics[scale = 0.13]{fbmCT.png} \caption{fBm with $H=0.3$ subordinated by the multifractal time change supplied by the distribution function of the binomial measure $\mu_p$ with $p=1/4$.} \label{fbmCT} \end{figure} { \bf Remarks:} The fact that the same time change is performed does not play a particular role for the estimation of the wavelet cross-correlations; the same result would follow for two processes $B_1\circ f $ and $B_2\circ g $ with $B_1$ and $B_2$ independent, and where $f$ and $g$ are two deterministic time changes. Similarly, $B_1$ and $B_2$ can be replaced by two (possibly different) centered Gaussian processes. \begin{figure} \centering \includegraphics[scale = 0.13]{fbmCTSpect.png} \caption{Bivariate multifractal spectrum of two independent fBm with the same time change: the distribution function of a binomial measure with $p=1/4$; in contradistinction with the cross-correlation of wavelet coefficients, the wavelet leaders are strongly correlated, leading to a bivariate Legendre spectrum theoretically supported by the line $H_1 = H_2$, which is close to be the case numerically. } \label{fbmSpect} \end{figure} Let us mention at this point that the mathematical problem of understanding what is the multifractal spectrum of the composition $f\circ g$ of two multifractal functions $f$ and $g$, where $g$ is a { \em time subordinator} i.e. an increasing function, is a largely open problem (and is posed here in too much generality to find a general answer). This problem was initially raised by B. Mandelbrot and also investigated R. Riedi \cite{Riedi2003} who worked out several important subcases; see also the article by S. Seuret \cite{Seur3}, who determined a criterium under which a function can be written as the composition of a time subordinator and a monoh\"older function, and \cite{BaSe3} where J. Barral and S. Seuret studied the multifractal spectrum of a L\'evy process, under a time subordinator given by the repartition function of a multifractal cascade. \subsection{Multivariate analysis of marathon physiological data} \label{MultivariateMar} Let us consider one of the marathon runners, and denote his heart beat frequency by $f_f$ and his cadence by $f_c$ and by $f_f^{(-1)}$ and $f_c^{(-1)}$ their primitives. We performed the computation of the bivariate scaling function $\zeta_{f_f^{(-1)},f_c^{(-1)}}$ (using wavelet leaders) and we show its Legendre transform ${ \cal L}_{f_f^{(-1)},f_c^{(-1)}}$ on Fig. \ref{bilan_pl2}. This spectrum is widely spread, in strong contradistinction with the bivariate spectra obtained in the previous section; this indicates that no clear correlations between the H\"older singularities of the primitives can be put in evidence. Fig. \ref{bilan_pl3} shows the two corresponding univariate spectra (which can be either computed directly, or obtained as projections of the bivariate spectrum). \begin{figure} \centering \includegraphics[scale=0.25]{PLbilan2.png} \caption{Representation of the bivariate H\"older Legendre spectrum of the primitives of heart beat frequency and cadence: this bivariate spectrum is derived from the same data that were used to derive the two univariate spectra shown in Fig \ref{bilan_pl3}. } \label{bilan_pl2} \end{figure} In order to test possible relationships between the bivariate spectrum and the two corresponding univariate spectra, we compute the difference \[ {\mathcal L} (H_1, H_2) - {\mathcal L} (H_1) - {\mathcal L} (H_2) +1, \] which allows to test the validity of \eqref{spectindep} and \[ { \cal L}_{(f_1,f_2)} - \min ( { \cal L}_{f_1} (H_1), { \cal L}_{f_2} (H_2)), \] which allows to test the validity of \eqref{largint2}, they are shown in Fig. \ref{bilan_biv1}. This comparison suggests that the large intersection formula is more appropriate than the codimension formula in this case. Keeping in mind the conclusions of Section \ref{MultivariateSpec}, these results indicate that an hypothesis of both stationarity and independence for each signals is inapropriate (indeed this would lead to the validity of the codimension formula), and on the opposite, these results are compatible with a pointwise regularity yielded by a limsup set procedure, as explained in Sec \ref{MultivariateSpec}. \begin{figure} \centering \hbox{\hspace{-4em}\includegraphics[scale=0.35]{PLbilan4.png}} \caption{Representation of the difference of the bivariate spectrum and the two formulas proposed in \eqref{spectindep} and \eqref{largint2}. The graph on the left is closer to zero, which suggests that the large intersection formula seems more appropriate in this case.} \label{bilan_biv1} \end{figure} \section{Conclusion} Let us give a summary of the conclusions that can be drawn from a bivariate multifractal analysis of data based on the Legendre transform method. This analysis goes beyond the (now standard) technique of estimating correlations of wavelet coefficients; indeed here wavelet coefficients are replaced by wavelet leaders, which leads to new scaling parameters on which classification can be performed. On the mathematical side, even if the relationship between the Legendre and the multifractal spectra is not as clear as in the univariate case, nonetheless, situations have been identified where this technique can either yield information on the nature of the singularities (e.g. the absence of oscillating singularities), or on the type of processes that can be used to model the data (either of additive or of multiplicative type). In the particular case of marathon runners, the present study shows a bivariate spectra between heart rate and cadence are related by the large intersection formula. In a recent study \cite{BillatPet} a multivariate analysis revealed that, for all runners, RPE and respiratory frequency measured on the same runners during the marathon were close (their angle is acute on correlation circle of a principal component analysis) while the speed was closer to the cadence and to the Tidal respiratory volume at each inspiration and expiration). The sampling frequency of the respiratory parameters did not allow to apply the multifractal analysis which here reveals that the cadence and heart rate could be an additive process such as, possibly a generalization of a Lévy process. Heart rate and cadence are under the autonomic nervous system control and Human beings optimize their cadence according his speed for minimizing his energy cost of running. Therefore, we can conclude that is not recommended to voluntarily change the cadence and this bivariate multifractal analysis mathematically shows that the cadence and heart rate are not only correlated but we can conjecture that they can be modeled by an additive process until the end of the marathon. \label{conclu} \bibliographystyle{plain}

e4dc8e2a321fb206bea1716ce512d8d8dbefd50a

\section{Introduction}\label{Section1} With the rapid development of the Internet of Things (IoT) and $5$th Generation Mobile Communication ($5$G) technologies, the power supply and high-rate communication for electronic devices (e.g., sensors, cameras) have become well-known issues \cite{wu2014cognitive, sun2019cooperative}. Recently, to support novel applications in the IoT and 5G wireless networks, some studies have been conducted to implement self-sustainable communication systems with energy harvesting (EH) techniques \cite{perera2018simultaneous}. \begin{figure}[!t] \centering \includegraphics[scale=0.55]{Fig1.pdf} \caption{The comparison of the beam transmission in the RBS without TIM and TIM-RBS.} \label{Fig1} \end{figure} Compared with the currently available batteries and charged super-capacitors, EH can capture and convert wasted or important energy into electricity, which makes it possible to integrate energy sources into wireless networks \cite{hou2016preliminary, niyato2007wireless}. However, since the irregular and unforeseeable nature of ambient sources, the application of EH faces many limitations and applies only to specific environments \cite{krikidis2014simultaneous}. The wireless power transfer (WPT) technology is one of the novel EH technologies that can overcome the above limitations \cite{lu2015wireless}. Recent WPT technologies that have been widely used include near-field and far-field energy transmission \cite{garnica2013wireless}. For near-field WPT, such as magnetic induction and magnetic resonance, watt-level power can be transferred within multi-centimeter-level distances using centimeter-sized transceivers \cite{jawad2017opportunities}. For far-field WPT, radio-frequency (RF) or laser can transfer milliwatt-level energy over meter-level distances under safety regulations \cite{xia2015efficiency}. The prospect of integrating WPT into communication networks creates a demand for technology that can simultaneously transmit information and power to electronic devices \cite{krikidis2014simultaneous}. Thus, the concept of simultaneous wireless information and power transfer (SWIPT) was first introduced in \cite{varshney2008transporting} with theoretical analysis. To realize long-range, high-power WPT and high-rate communication, the existing near-field and far-field SWIPT face the challenge of balancing distance, power, spectral efficiency, and safety \cite{ulukus2015energy, chen2016secrecy, ku2015advances}. The latest SWIPT, resonant beam system (RBS), is firstly proposed in \cite{fang2021end}, where the end-to-end power transfer was theoretically and numerically studied, and the $5\rm{m}$ distance was exhibited. The resonant beam (i.e. the intra-cavity laser) is used to transmit energy between the spatially separated transmitter and receiver \cite{liu2016dlc, zhang2018distributed2}. However, due to the resonant beam spot diffuses with the increase of transmission distance, the transmission loss grows with the extension of cavity length between the transmitter and the receiver. Afterwards, if the transmission distance keeps extending, the resonance condition in the RBS will be no longer satisfied (i.e. losses are greater than gain), so the energy transfer will be suspended, which will result in limited transmission distances \cite{wang2018channel}. In this paper, we introduce an RBS with a telescope internal modulator (TIM), which can be called TIM-RBS. As shown in Fig.~\ref{Fig1}(b), the telescope contains a concave mirror and a convex mirror, which has been used in the laser resonator for obtaining reliable operation of an Nd:YAG laser with a large-volume TEM$_{00}$ mode in \cite{hanna1981telescopic}. Then, due to the telescope can reduce the size of the beam to increase the diffraction per unit length, Sarkies \cite{sarkies1979stable} reported using a telescope in an Nd:YAG resonator allowing easily controllable adjustment to compensate thermal lensing under varied pumping conditions. In Fig.~\ref{Fig1}(b), the resonant beam carrying energy is transmitted from the transmitter to the receiver through input reflector, gain medium, TIM, and output reflector in order while it passes through input reflector, gain medium, and output reflector in the RBS without TIM depicted in Fig.~\ref{Fig1}(a). From \cite{hanna1981telescopic, sarkies1979stable} and Fig.~\ref{Fig1}, the main contribution of TIM is the compression of resonant beam spot, which brings small transmission losses and allows more beam energy to enter the gain medium for gain amplification, releasing more energy to be transmitted to the receiver for charging and communication. Thus, the transmission distance will increase with a higher power in the TIM-RBS. \begin{figure}[!t] \centering \includegraphics[scale=0.54]{Fig2.pdf} \caption{The structure and E$2$E power transmission model of the TIM-RBS.} \label{Fig2} \end{figure} The contributions of this paper include: \begin{enumerate} \item [\bf c1)] We propose an accurate transmission model for the TIM-RBS, which combines the electromagnetic field propagation and E$2$E power transfer model and allows to obtain precise energy distribution on the optical plane, transmission efficiencies, and output power. \item [\bf c2)] Using the proposed accurate transmission model, we visualise the process of beam spot compression and verify that the TIM-RBS can achieve longer range and higher power compared with the RBS without TIM. \item [\bf c3)] The numerical evaluation shows that the TIM-RBS can realize SWIPT at a distance of two times longer than the RBS without TIM, and about $34\rm{m}$ transmission distance, $4\rm{W}$ electric power, and $12\rm{bps/Hz}$ spectral efficiency can be achieved in the TIM-RBS with $200\rm{W}$ input power. \end{enumerate} In the rest of this paper, the structure of the TIM-RBS, the effect of the TIM, and the E$2$E power transfer will be illustrated in Section II, the accurate transmission model including electromagnetic field propagation, E$2$E power transfer model, power and information reception will be proposed in Section III. Then, Section IV will depict the performance comparison of TIM-RBS and RBS without TIM and the power transfer evaluation of the TIM-RBS. Finally, the conclusion with future research of the TIM-RBS will be presented in Section V. \section{System Overview}\label{Section2} To realize long-range power and information transfer using resonant beam, the telescope internal modulator (TIM) is adopted in the RBS compressing the beam spot for decreasing transmission loss. In this section, we will illustrate the structure of the TIM-RBS and propose the power transfer model. \subsection{TIM-RBS System}\label{} The TIM-RBS shown in Fig.~\ref{Fig2} incorporates spatially separated transmitter and receiver structures, with the lasing gain medium situated in the transmitter and the transmitter's output directed into the receiver. The implementation of a TIM-RBS comprises: i) a transmitter containing: a) a pumping source providing the input power for the system; b) an input reflector R$1$ having high reflectivity, usually 100\rm{\%}; c) a gain medium offering sites for photon emission; d) a TIM compressing the resonant beam for decreasing the diffraction loss. ii) a receiver including: a) an output reflector R$2$ with partial reflectivity; b) a power splitter (PS) dividing the received power into two power streams based on a certain PS ratio; c) a photovoltaic (PV) panel converting the received beam power into electric power for charging the battery; d) an avalanche photodiode (APD) converting the received beam power to electronic signals for communication. \begin{figure}[!t] \centering \includegraphics[scale=0.58]{Fig3.pdf} \caption{The diffraction loss due to optical apertures.} \label{Fig_diffractionloss} \end{figure} For a TIM-RBS, the transmitter can be located in a fixed position and properly sealed from dust and other contaminants, while the receiver is embedded in electronic devices for charging and communication. The resonant beam carrying the power transmits in the free space between the transmitter and the receiver. During the power transmission in the system, the resonant beam will pass through a series of optical apertures, e.g. input/output reflectors, gain medium. Since the limitations of optical aperture size, the transmission undergoes a series of diffraction losses. That is, in a resonant beam system, some part of the resonant beam will be lost either by spillover at the reflectors or by limiting apertures, such as the lateral boundaries of the gain medium. As shown in Fig.~\ref{Fig_diffractionloss}, these losses depend on the diameter of the beam in the plane of the aperture and the aperture radius, the diffracted beam energy outside the effective aperture of the optical components (reflectors or gain medium) will be lost, thus generating diffraction loss \cite{li1965diffraction}. \begin{figure}[!t] \centering \includegraphics[scale=0.58]{Fig4.pdf} \caption{The beam transmission in the TIM.} \label{Fig_TIM} \end{figure} Thus, to decrease the diffraction loss during the resonant beam transmission, the diameter of resonant beam should be compressed. The role of the TIM in the system lies in compressing the resonant beam, reducing transmission losses, and improving transmission efficiency including extending transmission distance and improving output power. As displayed in Fig.~\ref{Fig_TIM}, the TIM in the TIM-RBS comprises i) a concave lens with a focal length $f_{\rm{d1}}$, and ii) a convex lens with a focal length $f_{\rm{d2}}$, wherein the two lenses are placed parallel to each other and the focal points of them coincide \cite{van1977invention}. Then, the effective aperture of the gain medium is placed at the focal position, which allows the resonant beam with the smallest diameter to pass through the gain medium for amplification. \subsection{End-to-End Power Transfer}\label{} In the TIM-RBS, the energy is transformed into electricity and data through a series of conversions: i) power pumping, ii) power transfer including beam compression through the TIM and free-space transmission, iii) power output, and iv) beam power converting into electric power and communication resources. For power pumping, the electric power from the pumping source simulates the gain medium to excite the resonant beam. Firstly, the input electric power is converted into optical power. Then, the optical power is absorbed by the gain medium for realizing the population inversion. Finally, the inverted photons are spilled by the gain medium to form resonant beam. After the resonant beam is excited out of the gain medium, the TIM will shape its spot. When the resonant beam is transmitted forward (i.e. from the transmitter to the receiver) entering the TIM, the concave lens will perform the first phase conversion on it, so that the beam is transmitted to the focal point of the lens. Then, the beam passes through the convex lens and undergoes the second phase change. Conversely, if the resonant beam is transmitted from the receiver to the transmitter, the resonant beam will pass through the convex lens and the concave lens in the TIM successively, which causes the phase of the beam to be reversely adjusted, and the resonant beam is emitted with a smaller spot in parallel from the concave lens. Then, the resonant beam is transmitted from the transmitter to the receiver over the air. During the free-space transmission, the resonant beam will suffer a series of losses, such as dust and other contaminants, which will affect the end-to-end transmission efficiency. Afterwards, at the receiver, a portion of the resonant beam passes through the output reflector R$2$ converting into the output beam power for charging and communication. Another part of the resonant beam is reflected to the transmitter for amplification in the gain medium. Overall, the energy carried by the resonant beam is transmitted from the transmitter to the receiver by the above transfer stages. \section{Transmission Model Theoretical Analysis}\label{Section3} The charging power and the communication resource at the receiver are decided by the input power and the transmission efficiency of each transfer stage. The linear equation between the input power $P_{\rm{in}}$ and output power $P_{\rm{out}}$ is \cite{koechner2013solid} \begin{equation}\label{E2Epower} \begin{aligned} P_{\rm{out}} = (P_{\rm{in}} - P_{\rm{th}})\eta_{\rm{g}}\eta_{\rm{s}}\eta_{\rm{m}}\eta_{\rm{t}}(1-R_2), \end{aligned} \end{equation} where $P_{\rm{th}}$ is the input power threshold of enabling beam oscillation, $\eta_{\rm{g}}$ and $\eta_{\rm{s}}$ denote the stimulation efficiency and the transfer factor of gain medium, $\eta_{\rm{m}}$ represents the transmission efficiency in the telescope while $\eta_{\rm{t}}$ is the resonant beam transmission efficiency over the free-space. Besides, $R_2$ is the reflectivity of the output reflector. Among them, $P_{\rm{th}}$, $\eta_{\rm{g}}$ and $\eta_{\rm{s}}$ are the parameters relying on the hardware, such as the gain medium, the cavity, while $\eta_{\rm{m}}$ and $\eta_{\rm{t}}$ depend on the power transfer process through each optical element including TIM, input and output reflector. To explore the transmission efficiency and the output power of the TIM-RBS, we adopt the electromagnetic field propagation inside the resonant cavity to obtain the steady distribution on the optical plane. \subsection{Field Propagation inside the Cavity}\label{} The finite size of the transfer aperture causes diffraction loss, and diffraction plays a major role in determining the spatial distribution of the oscillation energy \cite{li1965diffraction}. In the TIM-RBS, the resonant beam will be diffracted if it passes through a series of transfer apertures, e.g. reflectors, gain medium, and TIM. Therefore, the amplitude and phase would change, and there exist the diffraction loss during power transfer. \begin{figure}[!t] \centering \includegraphics[scale=0.52]{Fig5.pdf} \caption{The parameters of optical elements and the transfer functions during a round-trip transmission in the TIM-RBS.} \label{Fig_Transfunction} \end{figure} As Fig.~\ref{Fig_Transfunction}(a) illustrated, we set the center point of the gain medium as the coordinate origin, that is $z=0$, and in which the $(x,y)$-plane is perpendicular to the $z$-axis. Then, $U(x,y,z)$ is the field mode distribution on the plane $(x,y)$ in $z$. A round-trip field propagation in the resonant cavity can be depicted in an eigenfunction as \cite{asoubar2016simulation} \begin{equation}\label{eq:transmissionengie} \gamma \mathbf{U} = \mathbf{T} \mathbf{U}, \end{equation} where the eigenvalue $\gamma$ is a complex constant related to the change in amplitude and phase. $\mathbf{U}$ represents the field distribution on a plane $U(x,y)$. $\mathbf{T}$ is the transfer function of a round-trip transmission, which can be formulated as $\mathbf{T}=\mathbf{T}_{\rm g}\mathbf{T}_{\rm m}\mathbf{T}_{\rm t}\mathbf{T}_{\rm r}\mathbf{T}_{\rm t}\mathbf{T}_{\rm m}\mathbf{T}_{\rm g}\mathbf{T}_{\rm r}$ as illustrated in Fig.~\ref{Fig_Transfunction}(b). Among them, $\mathbf{T}_{\rm g}$, $\mathbf{T}_{\rm m}$, $\mathbf{T}_{\rm t}$ and $\mathbf{T}_{\rm r}$ are the transfer function through gain medium, TIM, free space between transmitter and receiver, and reflector. Besides, $\mathbf{T}_{\rm m} = \mathbf{T}_{\rm ce}\mathbf{T}_{\rm cx}$ since the TIM comprises a concave lens and a convex lens (Fig.~\ref{Fig_Transfunction}(b)). Using electromagnetic wave propagation theory, if the field distribution on the source wave plane $U_1(x_1, y_1, z_1)$ is known, the field distribution on the observing plane $U_2(x_2, y_2, z_2)$ can be calculated as \cite{kortz1981diffraction} \begin{equation}\label{eq:planefield} U_2(x_2, y_2, z_2) = \iint_s U_1(x_1, y_1, z_1) T_{12}(x_1, y_1, x_2, y_2) ds, \end{equation} where $s$ is the area of the source wave plane. $T_{12}$ is the transfer function between the two planes, and it can be depicted as \begin{equation}\label{eq:transfunction} T_{12} = \frac{i\exp{(-ikL)}}{\lambda L}\exp \left\{\frac{-ik}{2 L}\left[\left(x_{2}-x_{1}\right)^{2}+\left(y_{2}-y_{1}\right)^{2}\right]\right\}, \end{equation} where $i$ is the imaginary unit, $L$ is the distance between the two planes, $\lambda$ is the wavelength of resonant beam, and $k = 2\pi/\lambda$ represents the wavenumber. Afterwards, if the beam passes through a lens, the phase and amplitude of beam field distribution are affected and changed. Here, we treat a lens as a diffraction screen, the complex amplitude transfer function of each lens can be given by \begin{equation}\label{eq:transplane} \begin{aligned} t(x, y) &= A(x, y) P(x, y) \\ &= A(x,y) \exp{\left[-\frac{ik}{2f}(x^2 + y^2)\right]}, \end{aligned} \end{equation} where $P(x, y)$ is the phase factor causing the phase changes, and $f$ is the focal length of the lens. $A(x, y)$ denotes the optical pupil function corresponding amplitude changes in the beam field, which is defined as $1$ with the area within the aperture and $0$ as the area is outside the aperture, i.e., \begin{equation}\label{eq:pupilfunction} A(x, y) = \left\{ \begin{aligned} & 1,\ x^2 + y^2 \leq r^2 \\ & 0,\ x^2 + y^2 > r^2 \end{aligned} \right. \end{equation} where $r$ is the radius of the aperture. \begin{figure}[!t] \centering \includegraphics[scale=0.60]{Fig6.pdf} \caption{The field propagation through a lens from the front surface to the rear surface and the formation of self-reproductive model by the Fox-Li iteration algorithm.} \label{Fig_lens_fox} \end{figure} Then, given the transfer function of the lens, the relationship between amplitude distribution on the front surface of lens $U(x, y)$ and on the rear surface $U'(x, y)$ (Fig.~\ref{Fig_lens_fox}(a)) is \begin{equation}\label{eq:frontandrear} U'(x, y) = U(x, y) t(x, y). \end{equation} For a round-trip power transfer in the TIM-RBS, the resonant beam passes through gain medium, TIM, output reflector, TIM, gain medium, and input reflector in sequence. To obtain the field distribution on the plane in the TIM-RBS, we adopt Fast-Fourier-Transform (FFT) method for a round-trip field propagation calculation \cite{sziklas1975mode, shen2006fast}. Based on \eqref{eq:transmissionengie}, \eqref{eq:planefield}, \eqref{eq:transfunction}, \eqref{eq:transplane}, and \eqref{eq:frontandrear}, the field propagation from the front surface $s_1$ of an optical element $E_1$ to the front surface $s_2$ of another optical element $E_2$ can be expressed as \cite{shen2006fast} \begin{equation}\label{eq:S1s2} \begin{aligned} U_2(x,y) &= \mathbf{T}_{12} \mathbf{U} \\ &= \mathscr{F}^{-1}\left\{\mathscr{F}\{U_1(x,y) t_{1}(x,y)\} \mathscr{F}\{T_{12}(x,y)\}\right\}, \end{aligned} \end{equation} where $U_1(x,y)$ and $U_2(x,y)$ are the field distribution on the front surface of the optical element $E_1$ and $E_2$, $t_{1}(x,y)$ is the field transfer function from the front surface to the rear surface of $E_1$, and $T_{12}$ denotes the transfer function from $E_1$ to $E_2$ and can be depicted as \begin{equation}\label{eq:TransfunCal2} \begin{aligned} T_{12}(x,y) = \frac{j \exp{(-j k L)}}{\lambda L} \exp{ \left[\frac{-j k}{2 L}\left(x^{2}+y^{2}\right)\right]}, \end{aligned} \end{equation} with the distance $L$ between $E_1$ and $E_2$. Therefore, the field distribution on the input reflector after a round-trip transmission can be calculated by $\mathbf{T}_{\rm g}\mathbf{U}$, $\mathbf{T}_{\rm{ce}}\mathbf{U}$, $\mathbf{T}_{\rm{cx}}\mathbf{U}$, $\mathbf{T}_{\rm{t}}\mathbf{U}$, $\mathbf{T}_{\rm{r}}\mathbf{U}$, $\mathbf{T}_{\rm{t}}\mathbf{U}$, $\mathbf{T}_{\rm{cx}}\mathbf{U}$, $\mathbf{T}_{\rm{ce}}\mathbf{U}$, $\mathbf{T}_{\rm g}\mathbf{U}$, and $\mathbf{T}_{\rm{r}}\mathbf{U}$ in order. Combining electromagnetic field propagation theory and FFT method, the field distribution on any plane during a round-trip transfer can be obtained. As shown in Fig.~\ref{Fig_lens_fox} (b), after several round-trip transmissions, the field distribution on a plane can achieve a stable mode, i.e., self-reproductive mode, whose amplitude and phase will not vary after the next transition. To solve \eqref{eq:transmissionengie} and obtain the self-reproductive mode, we adopt the ``Fox-Li" iteration algorithm to simulate the multiple round trips \cite{fox1961resonant}. A round-trip intra-cavity power transfer can be considered as the resonant beam passing through a series of apertures (reflector, gain medium, TIM) (Fig.~\ref{Fig_Transfunction}(a)), and one iterative process in ``Fox-Li" is equivalent to a round-trip power transfer. If a certain number of iterations are implemented, the field distribution on a plane no longer changes during transmission, the stable mode is achieved, and eventually, the self-reproductive mode is obtained (Fig.~\ref{Fig_lens_fox}(b)). After the self-reproductive mode achieved in the TIM-RBS, the transmission loss between the two planes $s_1$, $s_2$ of the optical element $E_1$ and $E_2$ can be calculated as \cite{gordon1964equivalence} \begin{equation}\label{eq:transmissionloss} \delta_{12} = \frac{\iint_{s_1}|U_1(x_1, y_1)|^2ds_1-\iint_{s_2}|U_{2}(x_{2}, y_{2})|^2ds_{2}}{\iint_{s_1}|U_1(x_1, y_1|^2ds_1}, \end{equation} where $|U|^2$ denotes the power intensity on the plane, while the power on the plane can be obtained by double integration of power intensity $\iint|U|^2$. Afterwards, the transmission efficiency between the two planes can be depicted as \begin{equation}\label{eq:transmissioneff} \begin{aligned} \eta_{12} & = 1 - \delta_{12} \\ & = \frac{\iint_{s_2}|U_{2}(x_{2}, y_{2})|^2ds_{2}}{\iint_{s_1}|U_1(x_1, y_1)|^2ds_1}. \end{aligned} \end{equation} Hence, the transmission efficiency between any two planes can be obtained based on the electromagnetic field propagation. \begin{figure}[!t] \centering \includegraphics[scale=0.6]{Fig7.pdf} \caption{The transmission efficiency during a round-trip transmission.} \label{Fig_transeff} \end{figure} \subsection{End-to-End Power Transfer}\label{} According to the output power of stable resonators, during a round-trip forward and backward traveling, the beam intensity is decreased due to diffraction losses (transfer factors $\eta_1$ - $\eta_4$), absorption inside the gain medium (transfer factor $\eta_{\rm s}$), and by output coupling (the reflectivity of the output reflector $R_2$) \cite{hodgson2005laser}. In the steady-state operation, these losses are compensated by the amplification process in the gain medium ($\eta_{\rm g}$) characterized by the small-signal gain coefficient $g_0$. If the transfer factors (i.e. transfer efficiencies) and the small-signal gain coefficient $g_0$ are known, and the cross-sectional area of the beam is $A_{\rm b}$, the output beam power is given by \cite{hodgson2005laser} \begin{equation}\label{eq:outputpower} \begin{aligned} P_{\text {out }} &= A_{\rm b} I (1-R_2) \eta_{\rm m} \eta_2 \\ &= A_{\rm b} I_{\rm s} \frac{(1-R_2) \eta_{\rm m} \eta_{2}}{1-R_2 \eta_{\rm m}^2 \eta_{2} \eta_{3}+\sqrt{R_2 \eta_{\rm m}^2 \eta_{\rm t}}\left[1 /\left(\eta_{1} \eta_{\rm{s}} \eta_{\rm m} \eta_{2} \right)-\eta_{\rm{s}}\right]} \\ &\ \ \ \ \left[ g_{0} \ell-\left| \ln \sqrt{R_2 \eta_{\rm{s}}^{2} \eta_{\rm m}^2 \eta_{\rm t}} \right| \right] \end{aligned} \end{equation} where $I_{\rm s}$ is the saturated beam intensity of gain medium. As shown in Fig.~\ref{Fig_transeff}, $\eta_{\rm m}$ is the transfer efficiency of the TIM, which can be calculated by multiplying the transfer efficiencies of the gain medium to the concave lens and the concave lens to the convex lens as $\eta_{\rm m} = \eta_{\rm ce} \eta_{\rm cx}$. The transmission efficiency $\eta_{\rm t} = \eta_1 \eta_2 \eta_3 \eta_4$ with $\eta_1$, $\eta_2$, $\eta_3$ and $\eta_4$ representing the forward and backward transfer efficiency between the input reflector and the gain medium, the TIM and the output reflector, respectively. Moreover, all the transmission efficiency can be calculated through \eqref{eq:transmissioneff} with the field distribution on the two planes accurately. Additionally, based on \eqref{E2Epower} and \eqref{eq:outputpower}, to pump the resonant beam, the input power $P_{\rm{in}}$ is firstly converted into the stored power in the gain medium for stimulating the resonant beam with the efficiency $\eta_{\rm g}$, which can be expressed as \cite{hodgson2005laser} \begin{equation}\label{eq:gaineff} \begin{aligned} \eta_{\rm{g}} =\frac{AI_{\rm s}g_0\ell}{P_{\rm{in}}}, \end{aligned} \end{equation} where $A$ is the cross-sectional area of gain medium. Besides, to obtain the beam oscillation, the laser threshold condition should be satisfied. That is, the input power $P_{\rm{in}}$ needs to be greater than the pumping power threshold $P_{\rm{th}}$ \begin{equation}\label{eq:powerthreshold} \begin{aligned} P_{\rm{th}} = \frac{A I_{\rm s} \left| \ln \sqrt{R_2 \eta_{\rm{s}}^{2} \eta_{\rm m}^2 \eta_{\rm t}} \right|}{\eta_{\rm{g}}}. \end{aligned} \end{equation} Thus, in the TIM-RBS, the transmission efficiency calculation and the E$2$E power transfer analysis enable the receiver to obtain the output power with different parameters. \subsection{Power Harvesting and Information Receiving}\label{} At the receiver, the power splitter (PS) divides the received power into two power streams of different levels with a certain PS ratio. Afterwards, both power streams are transmitted to an energy harvester and an information decoder to achieve simultaneous power harvesting and information decoding. The information rate and the harvested energy can be balanced and optimized according to the system requirements by varying the PS ratio. Let $\theta$ denote the PS ratio for the receiver. The PV panel is responsible for converting beam power received by the receiver into electrical power, while the APD converts the light beam into electronic signals for communication. \begin{figure}[!t] \centering \includegraphics[scale=0.6]{Fig8.pdf} \caption{The equivalent circuit of a PV panel.} \label{Fig_pv} \end{figure} For power harvesting, the equivalent circuit of the single-diode model for PV panel is shown in Fig.~\ref{Fig_pv}. The meaning of the five parameters for the PV model is illustrated as follows: $I_{\rm{ph}}$ represents the photo-generated current, which depends on the input beam power of the PV panel and can be depicted as \begin{equation}\label{current} \begin{aligned} I_{\rm{ph}} = \theta \eta_{\rm{pv}} P_{\rm{out}}, \end{aligned} \end{equation} where $\eta_{\rm{pv}}$ is the conversion responsivity of the PV panel. $I_{\rm o}$ is the dark saturation current, $R_{\rm s}$ is the panel resistance, $R_{\rm{sh}}$ is panel parallel (shunt) resistance, and $D$ is the diode quality (ideality) factor. Afterwards, the general current-voltage characteristic of a PV panel based on the single exponential model can be expressed as \cite{sera2007pv} \begin{equation}\label{PV} \begin{aligned} i_{\rm{pv}} = I_{\rm{ph}} - I_{\rm o} \left\{\exp{\left[\frac{(v_{\rm{pv}} + i_{\rm{pv}} R_{\rm s})q}{n_{\rm s} D K T}\right]} - 1\right\} - \frac{v_{\rm{pv}} + i_{\rm{pv}} R_{\rm s}}{R_{\rm{sh}}}, \end{aligned} \end{equation} where $i_{\rm{pv}}$ and $v_{\rm{pv}}$ are the output current and voltage of the PV panel, and the output electric power through the PV panel can be calculated as \begin{equation}\label{outelepower} \begin{aligned} P_{\rm e} &= i_{\rm{pv}} v_{\rm{pv}} = i_{\rm{pv}}^2 R_{\rm{pv}}, \end{aligned} \end{equation} where $R_{\rm{pv}}$ denotes the load resistance of the PV panel. Additionally, the constants in \eqref{PV} include Boltzmann’s constant $K$, the charge of electron $q$, the number of cells in the panel connected in series $n_{\rm s}$, and the temperature in Kelvin $T$. For information receiving, the power stream received by APD is \begin{equation}\label{powerapd} \begin{aligned} P_{\rm{apd}} = (1 - \theta) P_{\rm{out}}. \end{aligned} \end{equation} Then, the spectral efficiency, also known as information transfer rate, can be expressed as \cite{cvijetic2008performance} \begin{equation}\label{SE} \begin{aligned} C = \frac{1}{2}\log \left( 1+ \frac{S}{N}\frac{e}{2 \pi} \right), \end{aligned} \end{equation} where $S$ is the signal power while $N$ is the noise power. Thereby, $S/N$ denotes the signal-to-noise ratio (SNR). Additionally, $S$ relies on the output beam power \begin{equation}\label{SP} \begin{aligned} S = \left(P_{\rm{apd}}\eta_{\rm{apd}}\right)^2, \end{aligned} \end{equation} where $\eta_{\rm{apd}}$ is the photo-electric conversion efficiency of APD. $N$ is the additive Gaussian white noise (AGWN) power and can be calculated as \begin{equation}\label{NP} \begin{aligned} N & = 2q \left(P_{\rm{apd}}\eta_{\rm{apd}} + I_{\rm{bc}} \right) B_{\rm n} + \frac{4KTB_{\rm n}}{R_{\rm{apd}}}, \end{aligned} \end{equation} where $I_{\rm{bc}}$ is the background current, $B_{\rm n}$ is the noise bandwidth, and $R_{\rm{apd}}$ is the load resistor of the APD. The meanings of $q$, $K$, and $T$ are similar to them in \eqref{PV}. \begin{table}[!t] \setlength{\abovecaptionskip}{0pt} \setlength{\belowcaptionskip}{-3pt} \centering \caption{Parameters related to the input/output reflectors, the TIM, and the gain medium \cite{hodgson2005laser}.} \begin{tabular}{ccc} \hline \textbf{Symbol} & \textbf{Parameter} & \textbf{Value} \\ \hline \bfseries{$r_{\rm{in}}$} & {Radius of the input reflector} & {$2.5\rm{mm}$} \\ \bfseries{$r_{\rm{out}}$} & {Radius of the output reflector} & {$2.5\rm{mm}$} \\ \bfseries{$R_1$} & {Input retro-reflector reflectivity} & {$100\rm{\%}$} \\ \bfseries{$R_2$} & {Output retro-reflector reflectivity} & {$70\rm{\%}$} \\ \bfseries{$L_1$} & {Length from input reflector to gain medium} & {$0.2\rm{m}$} \\ \bfseries{$L_2$} & {Length from gain medium to TIM} & {$0.05\rm{m}$} \\ \bfseries{$f_{\rm{d1}}$} & {Focal length of concave lens} & {$0.05\rm{m}$} \\ \bfseries{$f_{\rm{d2}}$} & {Focal length of convex lens} & {$0.1\rm{m}$} \\ \bfseries{$I_{\rm s}$} & {Saturation intensity} & {$1260\rm{W/cm^2}$} \\ \bfseries{$\eta_{\rm s}$} & {Transfer coefficient of gain medium} & {$99\rm{\%}$} \\ \bfseries{$\eta_{\rm{g}}$} & {Excitation efficiency} & {$72\rm{\%}$} \\ \hline \label{table1} \end{tabular} \end{table} In summary, based on the output beam power in the TIM-RBS, the electric power and communication rate in the receiver can be obtained through \eqref{outelepower} and \eqref{SE}. \section{Numerical Evaluation}\label{Section4} To demonstrate the validity of the proposed transfer model in the TIM-RBS including the field propagation and the E$2$E power transfer, we will simulate the field distribution on the optical planes, the beam radius, the transmission efficiency, and the output power in this section. \subsection{Parameters Setting}\label{} The parameters of the TIM-RBS system are depicted in Table \ref{table1}. Among them, the input and output reflectors are plane mirrors. The Nd:YVO$_4$ is adopted as the gain medium, whose radius is less than or equal to that of the reflectors \cite{hodgson2005laser}. \begin{table}[!t] \setlength{\abovecaptionskip}{0pt} \setlength{\belowcaptionskip}{-3pt} \centering \caption{Parameters of energy harvesting and information receiving \cite{perales2016characterization}}. \begin{tabular}{ccc} \hline \textbf{Symbol} & \textbf{Parameter} & \textbf{Value} \\ \hline \bfseries{$R_{\rm{pv}}$} & {Load resistance of PV} & {$100\rm{\Omega}$} \\ \bfseries{$\eta_{\rm{pv}}$} & {Conversion responsivity of PV} & {$0.0161\rm{A/W}$} \\ \bfseries{$I_{\rm o}$} & {Dark saturation current} & {$9.89\times10^{-9}\rm{A}$}\\ \bfseries{$R_{\rm s}$} & {Panel series resistance} & {$0.93\rm{\Omega}$} \\ \bfseries{$R_{\rm{sh}}$} & {Panel parallel resistance} & {$52.6\rm{k\Omega}$} \\ \bfseries{$D$} & {Diode quality factor} & {$1.105$} \\ \bfseries{$n_{\rm s}$} & {Number of PV cell} & {$40$} \\ \bfseries{$\eta_{\rm{apd}}$} & {Conversion efficiency of APD} & {$0.6\rm{A/W}$} \cite{demir2017handover} \\ \bfseries{$q$} & {Quantity of electric charge} & {$1.6\times10^{-19}$} \\ \bfseries{$I_{\rm{bc}}$} & {Background current} & {$5100\rm{\mu A}$} \cite{moreira1997optical} \\ \bfseries{$B_{\rm n}$} & {Noise bandwidth} & {$811.7\rm{MHz}$} \cite{quintana2017high} \\ \bfseries{$R_{\rm{apd}}$} & {Load resistor} & {$10\rm{k\Omega}$} \cite{xu2011impact} \\ \bfseries{$K$} & {Boltzmann constant} & {$1.38\times10^{-23}$} \\ \bfseries{$T$} & {Kelvin Temperature} & {$300{\rm K}$} \\ \hline \label{table2} \end{tabular} \end{table} By using the PV equivalent circuit fitting measured data of a real PV cell product, we obtain the parameters of the vertical multi-junction PV cell, which are shown in Table \ref{table2} \cite{perales2016characterization}. Additionally, the parameter of the APD $\eta_{\rm{apd}}$ is from the experimental results using APD for information decoding \cite{demir2017handover}. Furthermore, other parameters related to the information receiving including $B_{\rm n}$, $I_{\rm{bc}}$, and $R_{\rm{apd}}$ are extracted from the literature \cite{quintana2017high, moreira1997optical, xu2011impact}. \subsection{Comparison of Field Distribution}\label{} \begin{figure*}[htbp] \centering \subfigure[Gain medium]{ \begin{minipage}[t]{0.23\linewidth} \centering \includegraphics[width=1.8in]{Fig9a.pdf} \end{minipage}% }% \subfigure[Output reflector]{ \begin{minipage}[t]{0.23\linewidth} \centering \includegraphics[width=1.8in]{Fig9b.pdf} \end{minipage}% }% \subfigure[Gain medium]{ \begin{minipage}[t]{0.23\linewidth} \centering \includegraphics[width=1.8in]{Fig9c.pdf} \end{minipage} }% \subfigure[Input reflector]{ \begin{minipage}[t]{0.23\linewidth} \centering \includegraphics[width=1.8in]{Fig9d.pdf} \end{minipage} }% \centering \caption{The normalized beam intensity distribution on the gain medium, output reflector, gain medium, and input reflector during a round-trip transmission in the TIM-RBS.} \label{Fig_field_TIM} \end{figure*} \begin{figure*}[htbp] \centering \subfigure[Gain medium]{ \begin{minipage}[t]{0.23\linewidth} \centering \includegraphics[width=1.8in]{Fig10a.pdf} \end{minipage}% }% \subfigure[Output reflector]{ \begin{minipage}[t]{0.23\linewidth} \centering \includegraphics[width=1.8in]{Fig10b.pdf} \end{minipage}% }% \subfigure[Gain medium]{ \begin{minipage}[t]{0.23\linewidth} \centering \includegraphics[width=1.8in]{Fig10c.pdf} \end{minipage} }% \subfigure[Input reflector]{ \begin{minipage}[t]{0.23\linewidth} \centering \includegraphics[width=1.8in]{Fig10d.pdf} \end{minipage} }% \centering \caption{The normalized beam intensity distribution on the gain medium, output reflector, gain medium, and input reflector during a round-trip transmission in the RBS without TIM.} \label{Fig_Field_no} \end{figure*} Based on \eqref{eq:planefield}-\eqref{eq:TransfunCal2}, the field distribution $U$ on the planes of the optical elements can be accurately obtained. Then, the beam intensity on the planes can be calculated by $|U|^2$. To observe the change of field distribution during a round-trip transmission and compare the compression effect of beam spot with and without TIM, we illustrate the normalized beam intensity distribution of TIM-RBS and RBS without TIM under the same radii of the input/output reflectors and gain medium with $10\rm{m}$ transmission distance in Figs.~\ref{Fig_field_TIM} and~\ref{Fig_Field_no} respectively. During a round trip, the resonant beam passes through the gain medium, TIM, output reflector, TIM, gain medium and input reflector in turn in the TIM-RBS, while it is transmitted through the gain medium, output reflector, gain medium, and input reflector in the RBS without TIM. After the resonance in the system has achieved a steady state, that is the self-reproductive mode of resonant beam has been obtained, we study the normalized beam intensity distribution in the TIM-RBS and the RBS without TIM. Fig.~\ref{Fig_field_TIM}(a)-(d) shows the beam intensity distribution on the gain medium, output reflector, gain medium, and the input reflector in the TIM-RBS. From the gain medium to the output reflector (Fig.~\ref{Fig_field_TIM}(a)-(b)), the cross-sectional area of resonant beam increases because the resonant beam spreads when it travels in free space. Then, since the spot compression effect of the TIM, the cross-sectional area of the resonant beam reduces from the output reflector to the gain medium (Fig.~\ref{Fig_field_TIM}(b)-(c)). Finally, the resonant beam is focused onto the input reflector due to the function of the gain medium with convex lens properties (Fig.~\ref{Fig_field_TIM}(c)-(d)). Similarly, Fig.~\ref{Fig_Field_no}(a)-(d) depicts the normalized beam intensity distribution on the gain medium, output reflector, gain medium, and input reflector during a round-trip transmission in the RBS without TIM. The variation of the resonant beam spot transmitting from the gain medium to the output reflector (Fig.~\ref{Fig_Field_no}(a)-(b)) and from the gain medium to the input reflector (Fig.~\ref{Fig_Field_no}(c)-(d)) is identical to that in the TIM-RBS. Differently, since there isn't the TIM between the transmitter and the receiver, the transmission from the output reflector to the gain medium without spot compression, the spot of resonant beam transmitted from the output reflector to the gain medium remains almost unchanged (Fig.~\ref{Fig_Field_no}(b)-(c)). Therefore, the TIM in the TIM-RBS can compress the resonant beam spot during the transmission. Thus, the transmission loss in the TIM-RBS can be reduced, and finally, the transmission distance can be extended while the transmission power is able to be improved. \subsection{Comparison of Power Transfer performance}\label{} The TIM can suppress the transmission loss by compressing the resonant beam spot in the TIM-RBS. To illustrate the advantages of the TIM-RBS, we compare the radius of resonant beam on the gain medium (i.e. the radius at which the amplitude on the gain medium is $20\%$ of its peak value), the transmission efficiency, and the output beam power in the TIM-RBS and the RBS without TIM. Here, we analyze the power transfer performance in the TIM-RBS and RBS without TIM with $200\rm{W}$ input power $P_{\rm{in}}$ and $1.5\rm{mm}$ gain medium radius for comparison. Under the maximum transmission distance of $20\rm{m}$, the changes of resonant beam radius on gain medium $r_{\rm b}$ and transmission efficiency $\eta$ with the transmission distance $D_{\rm t}$ in the TIM-RBS and the RBS without TIM are shown in Fig.~\ref{Fig_TIM_no_eff_radius}. \subsubsection{For the resonant beam radius} Firstly, due to the diffusion effect of the resonant beam, the resonant beam radius increases as the transmission distance $D_{\rm t}$ extends in the two systems. Besides, under the same transmission distance $D_{\rm t}$, the resonant beam radius on the gain medium in the TIM-RBS is less than that in the RBS without TIM. Moreover, the radius of resonant beam in RBS without TIM can increase to the same size as the gain medium if the transmission distance increases to $4\rm{m}$. For example, if $D_{\rm t}$ is $10\rm{m}$, $r_{\rm b}$ in the TIM-RBS is about $1.3\rm{mm}$ while it is about $1.5\rm{mm}$ in the RBS without TIM. \subsubsection{For the transmission efficiency} Since the resonant beam radius increases with the extension of the transmission, the diffraction loss increases as the transmission increases. Thus, the transmission efficiency $\eta$ during a round trip (i.e., $\eta=\eta_{\rm t}\eta_{\rm m}^2$) decreases if the transmission distance grows. Moreover, the transmission efficiency in the TIM-RBS is greater than that in the RBS without TIM regardless of the transmission distance. If $D_{\rm t}$ is $10\rm{m}$, the transmission efficiency is $44\rm{\%}$ in the TIM-RBS and $14\rm{\%}$ in the RBS without TIM, respectively. \begin{figure}[!t] \centering \includegraphics[scale=0.55]{Fig11.pdf} \caption{Comparisons of the resonant beam radius on the gain medium and the transmission efficiency in the TIM-RBS and RBS without TIM.} \label{Fig_TIM_no_eff_radius} \end{figure} \begin{figure}[!t] \centering \includegraphics[scale=0.55]{Fig12.pdf} \caption{Comparison of the output beam power in the TIM-RBS and RBS without TIM with $200\rm{W}$ input power.} \label{Fig_power_TIM_no} \end{figure} Additionally, the variation of output beam power $P_{\rm{out}}$ of the TIM-RBS and the RBS without TIM with $200\rm{W}$ input power $P_{\rm{in}}$ is depicted in Fig.~\ref{Fig_power_TIM_no}. From Fig.~\ref{Fig_power_TIM_no}, the increase in transmission distance causes a decrease in output beam power due to the decrease of transmission efficiency. Besides, the largest transmission distance of the TIM-RBS is $35\rm{m}$ while it is $20\rm{m}$ of the RBS without TIM. Moreover, if the transmission distance is $1\rm{m}$, the output power in the TIM-RBS is less than that in the RBS without TIM due to the large beam spot ($A_{\rm b}=\pi r_{\rm b}^2$) and the similar transmission efficiency. Conversely, the output beam power in the TIM-RBS is greater than that in the RBS without TIM as the transmission distance is larger than $1\rm{m}$ and equal. For instance, if $D_{\rm t}$ is $10\rm{m}$, the output power is $15.30\rm{W}$ and $8.07\rm{W}$ in the TIM-RBS and the RBS without TIM, respectively. \begin{figure}[!t] \centering \includegraphics[scale=0.55]{Fig13.pdf} \caption{The changes of output beam power with different transmission distances and input power (e.g. $100\rm{W}$, $200\rm{W}$, and $300\rm{W}$) in the TIM-RBS.} \label{Fig_outpower_diffin} \end{figure} \begin{figure}[!t] \centering \includegraphics[scale=0.55]{Fig14.pdf} \caption{The changes of output electric power and spectral efficiency with different transmission distances and PS ratios under $200\rm{W}$ input power.} \label{Fig_comm_charging} \end{figure} In general, since the beam spot compression of the TIM, the diffraction loss is suppressed in the TIM-RBS. Thus, the power transfer performance in the TIM-RBS is superior to the RBS without TIM, which can be illustrated in higher output power and longer transmission distance. \subsection{Power Transfer performance in TIM-RBS }\label{} For the TIM-RBS, the output beam power can be obtained based on \eqref{eq:outputpower} with the determined input power $P_{\rm{in}}$. To analyze the power transfer performance of the TIM-RBS, we firstly explore the output beam power with different input power, and then study the output electric power and spectral efficiency with various PS ratios. Firstly, if the input power $P_{\rm{in}}$ is $100\rm{W}$, $200\rm{W}$ and $300\rm{W}$, the output beam power $P_{\rm{out}}$ obtained from \eqref{eq:outputpower} with the parameters in Table \ref{table1} is drawn in Fig.~\ref{Fig_outpower_diffin}. From Fig.~\ref{Fig_outpower_diffin}, the growth of the transmission distance $D_{\rm t}$ brings the decrease of the output beam power $P_{\rm{out}}$. If the input power is $300\rm{W}$, the output beam power is $23.87\rm{W}$, $8.91\rm{W}$, and $3.36\rm{W}$ with $10\rm{m}$, $20\rm{m}$, and $30\rm{m}$ transmission distance. Then, the output beam power increases as the input power increases under the same transmission distance. For example, $P_{\rm{out}}$ is $2.62\rm{W}$, $13.25\rm{W}$, and $23.87\rm{W}$ if $D_{\rm t}$ is $10\rm{m}$ and $P_{\rm{in}}$ is $100\rm{W}$, $200\rm{W}$, and $300\rm{W}$. Moreover, the maximum transmission distance varies if the input power is different. That is, the maximum transmission distance extends as the input power increases. The maximum transmission distances are $15\rm{m}$, $34\rm{m}$ and $58\rm{m}$ with the $100\rm{W}$, $200\rm{W}$ and $300\rm{W}$ input power. Additionally, the output electric power and the communication spectral efficiency (i.e. data rate) can be calculated by \eqref{outelepower} and \eqref{SE} with the known output beam power. If the input power $P_{\rm{in}}$ is $200\rm{W}$, under different transmission distance $D_{\rm t}$, the output electric power $P_{\rm e}$ and the spectral efficiency $C$ with various PS ratio $\theta$ are depicted in Fig.~\ref{Fig_comm_charging}. Since the output beam power decreases with the extension of the transmission distance, and $P_{\rm e}$ and $C$ are proportional to the output beam power, the output electric power and spectral efficiency decrease as the transmission distance increases regardless of the PS ratio. Besides, the output electric power increases with the growth of PS ratio, while the spectral efficiency decreases as the PS ratio increases under the same transmission distance. This is because that the PS ratio determines the distribution proportion of the output beam power, i.e., if $\theta$ is $0.5$, $50\rm{\%}$ of the output beam power is converted into the output electric power and the rest is transformed into the communication resources. Under the $10\rm{m}$ transmission distance, the output electric power is $0.41\rm{W}$ and the spectral efficiency is $11.47\rm{bps/Hz}$ with $0.3$ PS ratio while they are $2.10\rm{W}$ and $11.05\rm{bps/Hz}$ as the PS ratio is $0.7$. Moreover, the maximum electric power and spectral efficiency are about $4\rm{W}$ and $12\rm{bps/Hz}$ if the input power is $200\rm{W}$. In summary, based on the electromagnetic field propagation and power transfer model in the TIM-RBS, the output beam power, electric power, and spectral efficiency can be calculated accurately, and they all decrease as the transmission distance extends. Moreover, the maximum transmission distance (i.e. cavity length) increases with the growth of input power. \section{Conclusions}\label{Section5} In this paper, we proposed an accurate transmission model based on the electromagnetic field propagation and E$2$E power transfer for TIM-RBS. Then, we evaluated the power transfer performance of the TIM-RBS by comparing it with the RBS without TIM and analyzing the output beam/electric power and spectral efficiency. Comparing the normalized beam intensity distribution and the beam radius in the TIM-RBS and RBS without TIM, we concluded that the TIM plays an important role in compressing the beam spot. Then, the comparisons of transmission efficiency and output beam power shown that the power transfer performance in TIM-RBS is better than that in the RBS without TIM. Additionally, the power transfer performance of the TIM-RBS is related to the input power and transmission distance. The maximum transmission distance is $59\rm{m}$ with $300\rm{W}$ input power, and maximum output electric power and spectral efficiency are about $4\rm{W}$ and $12\rm{bps/Hz}$ if the input power is $200\rm{W}$. Hence, the TIM-RBS can realize longer-range SWIPT. However, some problems in the TIM-RBS need to be solved in future research, including i) proof of safety in the system, ii) power transfer performance in the TIM-RBS with different radii of input/output reflector and gain medium, and iii) experimental validation of the TIM-RBS. \bibliographystyle{IEEEtran}

333e1bd61f70ec5277bc4dc81890b5ee4b944c99